fs/btrfs/inode.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/kernel.h>
  20 #include <linux/bio.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/time.h>
  27 #include <linux/init.h>
  28 #include <linux/string.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/mpage.h>
  31 #include <linux/swap.h>
  32 #include <linux/writeback.h>
  33 #include <linux/statfs.h>
  34 #include <linux/compat.h>
  35 #include <linux/bit_spinlock.h>
  36 #include <linux/xattr.h>
  37 #include <linux/posix_acl.h>
  38 #include <linux/falloc.h>
  39 #include <linux/slab.h>
  40 #include <linux/ratelimit.h>
  41 #include <linux/mount.h>
  42 #include <linux/btrfs.h>
  43 #include <linux/blkdev.h>
  44 #include <linux/posix_acl_xattr.h>
  45 #include <linux/uio.h>
  46 #include "ctree.h"
  47 #include "disk-io.h"
  48 #include "transaction.h"
  49 #include "btrfs_inode.h"
  50 #include "print-tree.h"
  51 #include "ordered-data.h"
  52 #include "xattr.h"
  53 #include "tree-log.h"
  54 #include "volumes.h"
  55 #include "compression.h"
  56 #include "locking.h"
  57 #include "free-space-cache.h"
  58 #include "inode-map.h"
  59 #include "backref.h"
  60 #include "hash.h"
  61 #include "props.h"
  62 #include "qgroup.h"
  63 #include "dedupe.h"
  64
  65 struct btrfs_iget_args {
  66         struct btrfs_key *location;
  67         struct btrfs_root *root;
  68 };
  69
  70 struct btrfs_dio_data {
  71         u64 outstanding_extents;
  72         u64 reserve;
  73         u64 unsubmitted_oe_range_start;
  74         u64 unsubmitted_oe_range_end;
  75 };
  76
  77 static const struct inode_operations btrfs_dir_inode_operations;
  78 static const struct inode_operations btrfs_symlink_inode_operations;
  79 static const struct inode_operations btrfs_dir_ro_inode_operations;
  80 static const struct inode_operations btrfs_special_inode_operations;
  81 static const struct inode_operations btrfs_file_inode_operations;
  82 static const struct address_space_operations btrfs_aops;
  83 static const struct address_space_operations btrfs_symlink_aops;
  84 static const struct file_operations btrfs_dir_file_operations;
  85 static const struct extent_io_ops btrfs_extent_io_ops;
  86
  87 static struct kmem_cache *btrfs_inode_cachep;
  88 struct kmem_cache *btrfs_trans_handle_cachep;
  89 struct kmem_cache *btrfs_transaction_cachep;
  90 struct kmem_cache *btrfs_path_cachep;
  91 struct kmem_cache *btrfs_free_space_cachep;
  92
  93 #define S_SHIFT 12
  94 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  95         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  96         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  97         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  98         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  99         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
 100         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 101         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 102 };
 103
 104 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
 105 static int btrfs_truncate(struct inode *inode);
 106 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
 107 static noinline int cow_file_range(struct inode *inode,
 108                                    struct page *locked_page,
 109                                    u64 start, u64 end, u64 delalloc_end,
 110                                    int *page_started, unsigned long *nr_written,
 111                                    int unlock, struct btrfs_dedupe_hash *hash);
 112 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
 113                                            u64 len, u64 orig_start,
 114                                            u64 block_start, u64 block_len,
 115                                            u64 orig_block_len, u64 ram_bytes,
 116                                            int type);
 117
 118 static int btrfs_dirty_inode(struct inode *inode);
 119
 120 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 121 void btrfs_test_inode_set_ops(struct inode *inode)
 122 {
 123         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 124 }
 125 #endif
 126
 127 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 128                                      struct inode *inode,  struct inode *dir,
 129                                      const struct qstr *qstr)
 130 {
 131         int err;
 132
 133         err = btrfs_init_acl(trans, inode, dir);
 134         if (!err)
 135                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 136         return err;
 137 }
 138
 139 /*
 140  * this does all the hard work for inserting an inline extent into
 141  * the btree.  The caller should have done a btrfs_drop_extents so that
 142  * no overlapping inline items exist in the btree
 143  */
 144 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 145                                 struct btrfs_path *path, int extent_inserted,
 146                                 struct btrfs_root *root, struct inode *inode,
 147                                 u64 start, size_t size, size_t compressed_size,
 148                                 int compress_type,
 149                                 struct page **compressed_pages)
 150 {
 151         struct extent_buffer *leaf;
 152         struct page *page = NULL;
 153         char *kaddr;
 154         unsigned long ptr;
 155         struct btrfs_file_extent_item *ei;
 156         int err = 0;
 157         int ret;
 158         size_t cur_size = size;
 159         unsigned long offset;
 160
 161         if (compressed_size && compressed_pages)
 162                 cur_size = compressed_size;
 163
 164         inode_add_bytes(inode, size);
 165
 166         if (!extent_inserted) {
 167                 struct btrfs_key key;
 168                 size_t datasize;
 169
 170                 key.objectid = btrfs_ino(inode);
 171                 key.offset = start;
 172                 key.type = BTRFS_EXTENT_DATA_KEY;
 173
 174                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 175                 path->leave_spinning = 1;
 176                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 177                                               datasize);
 178                 if (ret) {
 179                         err = ret;
 180                         goto fail;
 181                 }
 182         }
 183         leaf = path->nodes[0];
 184         ei = btrfs_item_ptr(leaf, path->slots[0],
 185                             struct btrfs_file_extent_item);
 186         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 187         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 188         btrfs_set_file_extent_encryption(leaf, ei, 0);
 189         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 190         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 191         ptr = btrfs_file_extent_inline_start(ei);
 192
 193         if (compress_type != BTRFS_COMPRESS_NONE) {
 194                 struct page *cpage;
 195                 int i = 0;
 196                 while (compressed_size > 0) {
 197                         cpage = compressed_pages[i];
 198                         cur_size = min_t(unsigned long, compressed_size,
 199                                        PAGE_SIZE);
 200
 201                         kaddr = kmap_atomic(cpage);
 202                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 203                         kunmap_atomic(kaddr);
 204
 205                         i++;
 206                         ptr += cur_size;
 207                         compressed_size -= cur_size;
 208                 }
 209                 btrfs_set_file_extent_compression(leaf, ei,
 210                                                   compress_type);
 211         } else {
 212                 page = find_get_page(inode->i_mapping,
 213                                      start >> PAGE_SHIFT);
 214                 btrfs_set_file_extent_compression(leaf, ei, 0);
 215                 kaddr = kmap_atomic(page);
 216                 offset = start & (PAGE_SIZE - 1);
 217                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 218                 kunmap_atomic(kaddr);
 219                 put_page(page);
 220         }
 221         btrfs_mark_buffer_dirty(leaf);
 222         btrfs_release_path(path);
 223
 224         /*
 225          * we're an inline extent, so nobody can
 226          * extend the file past i_size without locking
 227          * a page we already have locked.
 228          *
 229          * We must do any isize and inode updates
 230          * before we unlock the pages.  Otherwise we
 231          * could end up racing with unlink.
 232          */
 233         BTRFS_I(inode)->disk_i_size = inode->i_size;
 234         ret = btrfs_update_inode(trans, root, inode);
 235
 236         return ret;
 237 fail:
 238         return err;
 239 }
 240
 241
 242 /*
 243  * conditionally insert an inline extent into the file.  This
 244  * does the checks required to make sure the data is small enough
 245  * to fit as an inline extent.
 246  */
 247 static noinline int cow_file_range_inline(struct btrfs_root *root,
 248                                           struct inode *inode, u64 start,
 249                                           u64 end, size_t compressed_size,
 250                                           int compress_type,
 251                                           struct page **compressed_pages)
 252 {
 253         struct btrfs_trans_handle *trans;
 254         u64 isize = i_size_read(inode);
 255         u64 actual_end = min(end + 1, isize);
 256         u64 inline_len = actual_end - start;
 257         u64 aligned_end = ALIGN(end, root->sectorsize);
 258         u64 data_len = inline_len;
 259         int ret;
 260         struct btrfs_path *path;
 261         int extent_inserted = 0;
 262         u32 extent_item_size;
 263
 264         if (compressed_size)
 265                 data_len = compressed_size;
 266
 267         if (start > 0 ||
 268             actual_end > root->sectorsize ||
 269             data_len > BTRFS_MAX_INLINE_DATA_SIZE(root) ||
 270             (!compressed_size &&
 271             (actual_end & (root->sectorsize - 1)) == 0) ||
 272             end + 1 < isize ||
 273             data_len > root->fs_info->max_inline) {
 274                 return 1;
 275         }
 276
 277         path = btrfs_alloc_path();
 278         if (!path)
 279                 return -ENOMEM;
 280
 281         trans = btrfs_join_transaction(root);
 282         if (IS_ERR(trans)) {
 283                 btrfs_free_path(path);
 284                 return PTR_ERR(trans);
 285         }
 286         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
 287
 288         if (compressed_size && compressed_pages)
 289                 extent_item_size = btrfs_file_extent_calc_inline_size(
 290                    compressed_size);
 291         else
 292                 extent_item_size = btrfs_file_extent_calc_inline_size(
 293                     inline_len);
 294
 295         ret = __btrfs_drop_extents(trans, root, inode, path,
 296                                    start, aligned_end, NULL,
 297                                    1, 1, extent_item_size, &extent_inserted);
 298         if (ret) {
 299                 btrfs_abort_transaction(trans, ret);
 300                 goto out;
 301         }
 302
 303         if (isize > actual_end)
 304                 inline_len = min_t(u64, isize, actual_end);
 305         ret = insert_inline_extent(trans, path, extent_inserted,
 306                                    root, inode, start,
 307                                    inline_len, compressed_size,
 308                                    compress_type, compressed_pages);
 309         if (ret && ret != -ENOSPC) {
 310                 btrfs_abort_transaction(trans, ret);
 311                 goto out;
 312         } else if (ret == -ENOSPC) {
 313                 ret = 1;
 314                 goto out;
 315         }
 316
 317         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 318         btrfs_delalloc_release_metadata(inode, end + 1 - start);
 319         btrfs_drop_extent_cache(inode, start, aligned_end - 1, 0);
 320 out:
 321         /*
 322          * Don't forget to free the reserved space, as for inlined extent
 323          * it won't count as data extent, free them directly here.
 324          * And at reserve time, it's always aligned to page size, so
 325          * just free one page here.
 326          */
 327         btrfs_qgroup_free_data(inode, 0, PAGE_SIZE);
 328         btrfs_free_path(path);
 329         btrfs_end_transaction(trans, root);
 330         return ret;
 331 }
 332
 333 struct async_extent {
 334         u64 start;
 335         u64 ram_size;
 336         u64 compressed_size;
 337         struct page **pages;
 338         unsigned long nr_pages;
 339         int compress_type;
 340         struct list_head list;
 341 };
 342
 343 struct async_cow {
 344         struct inode *inode;
 345         struct btrfs_root *root;
 346         struct page *locked_page;
 347         u64 start;
 348         u64 end;
 349         struct list_head extents;
 350         struct btrfs_work work;
 351 };
 352
 353 static noinline int add_async_extent(struct async_cow *cow,
 354                                      u64 start, u64 ram_size,
 355                                      u64 compressed_size,
 356                                      struct page **pages,
 357                                      unsigned long nr_pages,
 358                                      int compress_type)
 359 {
 360         struct async_extent *async_extent;
 361
 362         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 363         BUG_ON(!async_extent); /* -ENOMEM */
 364         async_extent->start = start;
 365         async_extent->ram_size = ram_size;
 366         async_extent->compressed_size = compressed_size;
 367         async_extent->pages = pages;
 368         async_extent->nr_pages = nr_pages;
 369         async_extent->compress_type = compress_type;
 370         list_add_tail(&async_extent->list, &cow->extents);
 371         return 0;
 372 }
 373
 374 static inline int inode_need_compress(struct inode *inode)
 375 {
 376         struct btrfs_root *root = BTRFS_I(inode)->root;
 377
 378         /* force compress */
 379         if (btrfs_test_opt(root->fs_info, FORCE_COMPRESS))
 380                 return 1;
 381         /* bad compression ratios */
 382         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 383                 return 0;
 384         if (btrfs_test_opt(root->fs_info, COMPRESS) ||
 385             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 386             BTRFS_I(inode)->force_compress)
 387                 return 1;
 388         return 0;
 389 }
 390
 391 /*
 392  * we create compressed extents in two phases.  The first
 393  * phase compresses a range of pages that have already been
 394  * locked (both pages and state bits are locked).
 395  *
 396  * This is done inside an ordered work queue, and the compression
 397  * is spread across many cpus.  The actual IO submission is step
 398  * two, and the ordered work queue takes care of making sure that
 399  * happens in the same order things were put onto the queue by
 400  * writepages and friends.
 401  *
 402  * If this code finds it can't get good compression, it puts an
 403  * entry onto the work queue to write the uncompressed bytes.  This
 404  * makes sure that both compressed inodes and uncompressed inodes
 405  * are written in the same order that the flusher thread sent them
 406  * down.
 407  */
 408 static noinline void compress_file_range(struct inode *inode,
 409                                         struct page *locked_page,
 410                                         u64 start, u64 end,
 411                                         struct async_cow *async_cow,
 412                                         int *num_added)
 413 {
 414         struct btrfs_root *root = BTRFS_I(inode)->root;
 415         u64 num_bytes;
 416         u64 blocksize = root->sectorsize;
 417         u64 actual_end;
 418         u64 isize = i_size_read(inode);
 419         int ret = 0;
 420         struct page **pages = NULL;
 421         unsigned long nr_pages;
 422         unsigned long nr_pages_ret = 0;
 423         unsigned long total_compressed = 0;
 424         unsigned long total_in = 0;
 425         unsigned long max_compressed = SZ_128K;
 426         unsigned long max_uncompressed = SZ_128K;
 427         int i;
 428         int will_compress;
 429         int compress_type = root->fs_info->compress_type;
 430         int redirty = 0;
 431
 432         /* if this is a small write inside eof, kick off a defrag */
 433         if ((end - start + 1) < SZ_16K &&
 434             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 435                 btrfs_add_inode_defrag(NULL, inode);
 436
 437         actual_end = min_t(u64, isize, end + 1);
 438 again:
 439         will_compress = 0;
 440         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 441         nr_pages = min_t(unsigned long, nr_pages, SZ_128K / PAGE_SIZE);
 442
 443         /*
 444          * we don't want to send crud past the end of i_size through
 445          * compression, that's just a waste of CPU time.  So, if the
 446          * end of the file is before the start of our current
 447          * requested range of bytes, we bail out to the uncompressed
 448          * cleanup code that can deal with all of this.
 449          *
 450          * It isn't really the fastest way to fix things, but this is a
 451          * very uncommon corner.
 452          */
 453         if (actual_end <= start)
 454                 goto cleanup_and_bail_uncompressed;
 455
 456         total_compressed = actual_end - start;
 457
 458         /*
 459          * skip compression for a small file range(<=blocksize) that
 460          * isn't an inline extent, since it doesn't save disk space at all.
 461          */
 462         if (total_compressed <= blocksize &&
 463            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 464                 goto cleanup_and_bail_uncompressed;
 465
 466         /* we want to make sure that amount of ram required to uncompress
 467          * an extent is reasonable, so we limit the total size in ram
 468          * of a compressed extent to 128k.  This is a crucial number
 469          * because it also controls how easily we can spread reads across
 470          * cpus for decompression.
 471          *
 472          * We also want to make sure the amount of IO required to do
 473          * a random read is reasonably small, so we limit the size of
 474          * a compressed extent to 128k.
 475          */
 476         total_compressed = min(total_compressed, max_uncompressed);
 477         num_bytes = ALIGN(end - start + 1, blocksize);
 478         num_bytes = max(blocksize,  num_bytes);
 479         total_in = 0;
 480         ret = 0;
 481
 482         /*
 483          * we do compression for mount -o compress and when the
 484          * inode has not been flagged as nocompress.  This flag can
 485          * change at any time if we discover bad compression ratios.
 486          */
 487         if (inode_need_compress(inode)) {
 488                 WARN_ON(pages);
 489                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 490                 if (!pages) {
 491                         /* just bail out to the uncompressed code */
 492                         goto cont;
 493                 }
 494
 495                 if (BTRFS_I(inode)->force_compress)
 496                         compress_type = BTRFS_I(inode)->force_compress;
 497
 498                 /*
 499                  * we need to call clear_page_dirty_for_io on each
 500                  * page in the range.  Otherwise applications with the file
 501                  * mmap'd can wander in and change the page contents while
 502                  * we are compressing them.
 503                  *
 504                  * If the compression fails for any reason, we set the pages
 505                  * dirty again later on.
 506                  */
 507                 extent_range_clear_dirty_for_io(inode, start, end);
 508                 redirty = 1;
 509                 ret = btrfs_compress_pages(compress_type,
 510                                            inode->i_mapping, start,
 511                                            total_compressed, pages,
 512                                            nr_pages, &nr_pages_ret,
 513                                            &total_in,
 514                                            &total_compressed,
 515                                            max_compressed);
 516
 517                 if (!ret) {
 518                         unsigned long offset = total_compressed &
 519                                 (PAGE_SIZE - 1);
 520                         struct page *page = pages[nr_pages_ret - 1];
 521                         char *kaddr;
 522
 523                         /* zero the tail end of the last page, we might be
 524                          * sending it down to disk
 525                          */
 526                         if (offset) {
 527                                 kaddr = kmap_atomic(page);
 528                                 memset(kaddr + offset, 0,
 529                                        PAGE_SIZE - offset);
 530                                 kunmap_atomic(kaddr);
 531                         }
 532                         will_compress = 1;
 533                 }
 534         }
 535 cont:
 536         if (start == 0) {
 537                 /* lets try to make an inline extent */
 538                 if (ret || total_in < (actual_end - start)) {
 539                         /* we didn't compress the entire range, try
 540                          * to make an uncompressed inline extent.
 541                          */
 542                         ret = cow_file_range_inline(root, inode, start, end,
 543                                                     0, 0, NULL);
 544                 } else {
 545                         /* try making a compressed inline extent */
 546                         ret = cow_file_range_inline(root, inode, start, end,
 547                                                     total_compressed,
 548                                                     compress_type, pages);
 549                 }
 550                 if (ret <= 0) {
 551                         unsigned long clear_flags = EXTENT_DELALLOC |
 552                                 EXTENT_DEFRAG;
 553                         unsigned long page_error_op;
 554
 555                         clear_flags |= (ret < 0) ? EXTENT_DO_ACCOUNTING : 0;
 556                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 557
 558                         /*
 559                          * inline extent creation worked or returned error,
 560                          * we don't need to create any more async work items.
 561                          * Unlock and free up our temp pages.
 562                          */
 563                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 564                                                      clear_flags, PAGE_UNLOCK |
 565                                                      PAGE_CLEAR_DIRTY |
 566                                                      PAGE_SET_WRITEBACK |
 567                                                      page_error_op |
 568                                                      PAGE_END_WRITEBACK);
 569                         btrfs_free_reserved_data_space_noquota(inode, start,
 570                                                 end - start + 1);
 571                         goto free_pages_out;
 572                 }
 573         }
 574
 575         if (will_compress) {
 576                 /*
 577                  * we aren't doing an inline extent round the compressed size
 578                  * up to a block size boundary so the allocator does sane
 579                  * things
 580                  */
 581                 total_compressed = ALIGN(total_compressed, blocksize);
 582
 583                 /*
 584                  * one last check to make sure the compression is really a
 585                  * win, compare the page count read with the blocks on disk
 586                  */
 587                 total_in = ALIGN(total_in, PAGE_SIZE);
 588                 if (total_compressed >= total_in) {
 589                         will_compress = 0;
 590                 } else {
 591                         num_bytes = total_in;
 592                         *num_added += 1;
 593
 594                         /*
 595                          * The async work queues will take care of doing actual
 596                          * allocation on disk for these compressed pages, and
 597                          * will submit them to the elevator.
 598                          */
 599                         add_async_extent(async_cow, start, num_bytes,
 600                                         total_compressed, pages, nr_pages_ret,
 601                                         compress_type);
 602
 603                         if (start + num_bytes < end) {
 604                                 start += num_bytes;
 605                                 pages = NULL;
 606                                 cond_resched();
 607                                 goto again;
 608                         }
 609                         return;
 610                 }
 611         }
 612         if (pages) {
 613                 /*
 614                  * the compression code ran but failed to make things smaller,
 615                  * free any pages it allocated and our page pointer array
 616                  */
 617                 for (i = 0; i < nr_pages_ret; i++) {
 618                         WARN_ON(pages[i]->mapping);
 619                         put_page(pages[i]);
 620                 }
 621                 kfree(pages);
 622                 pages = NULL;
 623                 total_compressed = 0;
 624                 nr_pages_ret = 0;
 625
 626                 /* flag the file so we don't compress in the future */
 627                 if (!btrfs_test_opt(root->fs_info, FORCE_COMPRESS) &&
 628                     !(BTRFS_I(inode)->force_compress)) {
 629                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 630                 }
 631         }
 632 cleanup_and_bail_uncompressed:
 633         /*
 634          * No compression, but we still need to write the pages in the file
 635          * we've been given so far.  redirty the locked page if it corresponds
 636          * to our extent and set things up for the async work queue to run
 637          * cow_file_range to do the normal delalloc dance.
 638          */
 639         if (page_offset(locked_page) >= start &&
 640             page_offset(locked_page) <= end)
 641                 __set_page_dirty_nobuffers(locked_page);
 642                 /* unlocked later on in the async handlers */
 643
 644         if (redirty)
 645                 extent_range_redirty_for_io(inode, start, end);
 646         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
 647                          BTRFS_COMPRESS_NONE);
 648         *num_added += 1;
 649
 650         return;
 651
 652 free_pages_out:
 653         for (i = 0; i < nr_pages_ret; i++) {
 654                 WARN_ON(pages[i]->mapping);
 655                 put_page(pages[i]);
 656         }
 657         kfree(pages);
 658 }
 659
 660 static void free_async_extent_pages(struct async_extent *async_extent)
 661 {
 662         int i;
 663
 664         if (!async_extent->pages)
 665                 return;
 666
 667         for (i = 0; i < async_extent->nr_pages; i++) {
 668                 WARN_ON(async_extent->pages[i]->mapping);
 669                 put_page(async_extent->pages[i]);
 670         }
 671         kfree(async_extent->pages);
 672         async_extent->nr_pages = 0;
 673         async_extent->pages = NULL;
 674 }
 675
 676 /*
 677  * phase two of compressed writeback.  This is the ordered portion
 678  * of the code, which only gets called in the order the work was
 679  * queued.  We walk all the async extents created by compress_file_range
 680  * and send them down to the disk.
 681  */
 682 static noinline void submit_compressed_extents(struct inode *inode,
 683                                               struct async_cow *async_cow)
 684 {
 685         struct async_extent *async_extent;
 686         u64 alloc_hint = 0;
 687         struct btrfs_key ins;
 688         struct extent_map *em;
 689         struct btrfs_root *root = BTRFS_I(inode)->root;
 690         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 691         struct extent_io_tree *io_tree;
 692         int ret = 0;
 693
 694 again:
 695         while (!list_empty(&async_cow->extents)) {
 696                 async_extent = list_entry(async_cow->extents.next,
 697                                           struct async_extent, list);
 698                 list_del(&async_extent->list);
 699
 700                 io_tree = &BTRFS_I(inode)->io_tree;
 701
 702 retry:
 703                 /* did the compression code fall back to uncompressed IO? */
 704                 if (!async_extent->pages) {
 705                         int page_started = 0;
 706                         unsigned long nr_written = 0;
 707
 708                         lock_extent(io_tree, async_extent->start,
 709                                          async_extent->start +
 710                                          async_extent->ram_size - 1);
 711
 712                         /* allocate blocks */
 713                         ret = cow_file_range(inode, async_cow->locked_page,
 714                                              async_extent->start,
 715                                              async_extent->start +
 716                                              async_extent->ram_size - 1,
 717                                              async_extent->start +
 718                                              async_extent->ram_size - 1,
 719                                              &page_started, &nr_written, 0,
 720                                              NULL);
 721
 722                         /* JDM XXX */
 723
 724                         /*
 725                          * if page_started, cow_file_range inserted an
 726                          * inline extent and took care of all the unlocking
 727                          * and IO for us.  Otherwise, we need to submit
 728                          * all those pages down to the drive.
 729                          */
 730                         if (!page_started && !ret)
 731                                 extent_write_locked_range(io_tree,
 732                                                   inode, async_extent->start,
 733                                                   async_extent->start +
 734                                                   async_extent->ram_size - 1,
 735                                                   btrfs_get_extent,
 736                                                   WB_SYNC_ALL);
 737                         else if (ret)
 738                                 unlock_page(async_cow->locked_page);
 739                         kfree(async_extent);
 740                         cond_resched();
 741                         continue;
 742                 }
 743
 744                 lock_extent(io_tree, async_extent->start,
 745                             async_extent->start + async_extent->ram_size - 1);
 746
 747                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 748                                            async_extent->compressed_size,
 749                                            async_extent->compressed_size,
 750                                            0, alloc_hint, &ins, 1, 1);
 751                 if (ret) {
 752                         free_async_extent_pages(async_extent);
 753
 754                         if (ret == -ENOSPC) {
 755                                 unlock_extent(io_tree, async_extent->start,
 756                                               async_extent->start +
 757                                               async_extent->ram_size - 1);
 758
 759                                 /*
 760                                  * we need to redirty the pages if we decide to
 761                                  * fallback to uncompressed IO, otherwise we
 762                                  * will not submit these pages down to lower
 763                                  * layers.
 764                                  */
 765                                 extent_range_redirty_for_io(inode,
 766                                                 async_extent->start,
 767                                                 async_extent->start +
 768                                                 async_extent->ram_size - 1);
 769
 770                                 goto retry;
 771                         }
 772                         goto out_free;
 773                 }
 774                 /*
 775                  * here we're doing allocation and writeback of the
 776                  * compressed pages
 777                  */
 778                 btrfs_drop_extent_cache(inode, async_extent->start,
 779                                         async_extent->start +
 780                                         async_extent->ram_size - 1, 0);
 781
 782                 em = alloc_extent_map();
 783                 if (!em) {
 784                         ret = -ENOMEM;
 785                         goto out_free_reserve;
 786                 }
 787                 em->start = async_extent->start;
 788                 em->len = async_extent->ram_size;
 789                 em->orig_start = em->start;
 790                 em->mod_start = em->start;
 791                 em->mod_len = em->len;
 792
 793                 em->block_start = ins.objectid;
 794                 em->block_len = ins.offset;
 795                 em->orig_block_len = ins.offset;
 796                 em->ram_bytes = async_extent->ram_size;
 797                 em->bdev = root->fs_info->fs_devices->latest_bdev;
 798                 em->compress_type = async_extent->compress_type;
 799                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
 800                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 801                 em->generation = -1;
 802
 803                 while (1) {
 804                         write_lock(&em_tree->lock);
 805                         ret = add_extent_mapping(em_tree, em, 1);
 806                         write_unlock(&em_tree->lock);
 807                         if (ret != -EEXIST) {
 808                                 free_extent_map(em);
 809                                 break;
 810                         }
 811                         btrfs_drop_extent_cache(inode, async_extent->start,
 812                                                 async_extent->start +
 813                                                 async_extent->ram_size - 1, 0);
 814                 }
 815
 816                 if (ret)
 817                         goto out_free_reserve;
 818
 819                 ret = btrfs_add_ordered_extent_compress(inode,
 820                                                 async_extent->start,
 821                                                 ins.objectid,
 822                                                 async_extent->ram_size,
 823                                                 ins.offset,
 824                                                 BTRFS_ORDERED_COMPRESSED,
 825                                                 async_extent->compress_type);
 826                 if (ret) {
 827                         btrfs_drop_extent_cache(inode, async_extent->start,
 828                                                 async_extent->start +
 829                                                 async_extent->ram_size - 1, 0);
 830                         goto out_free_reserve;
 831                 }
 832                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
 833
 834                 /*
 835                  * clear dirty, set writeback and unlock the pages.
 836                  */
 837                 extent_clear_unlock_delalloc(inode, async_extent->start,
 838                                 async_extent->start +
 839                                 async_extent->ram_size - 1,
 840                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 841                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 842                                 PAGE_SET_WRITEBACK);
 843                 ret = btrfs_submit_compressed_write(inode,
 844                                     async_extent->start,
 845                                     async_extent->ram_size,
 846                                     ins.objectid,
 847                                     ins.offset, async_extent->pages,
 848                                     async_extent->nr_pages);
 849                 if (ret) {
 850                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 851                         struct page *p = async_extent->pages[0];
 852                         const u64 start = async_extent->start;
 853                         const u64 end = start + async_extent->ram_size - 1;
 854
 855                         p->mapping = inode->i_mapping;
 856                         tree->ops->writepage_end_io_hook(p, start, end,
 857                                                          NULL, 0);
 858                         p->mapping = NULL;
 859                         extent_clear_unlock_delalloc(inode, start, end, NULL, 0,
 860                                                      PAGE_END_WRITEBACK |
 861                                                      PAGE_SET_ERROR);
 862                         free_async_extent_pages(async_extent);
 863                 }
 864                 alloc_hint = ins.objectid + ins.offset;
 865                 kfree(async_extent);
 866                 cond_resched();
 867         }
 868         return;
 869 out_free_reserve:
 870         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
 871         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
 872 out_free:
 873         extent_clear_unlock_delalloc(inode, async_extent->start,
 874                                      async_extent->start +
 875                                      async_extent->ram_size - 1,
 876                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 877                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 878                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 879                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 880                                      PAGE_SET_ERROR);
 881         free_async_extent_pages(async_extent);
 882         kfree(async_extent);
 883         goto again;
 884 }
 885
 886 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 887                                       u64 num_bytes)
 888 {
 889         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 890         struct extent_map *em;
 891         u64 alloc_hint = 0;
 892
 893         read_lock(&em_tree->lock);
 894         em = search_extent_mapping(em_tree, start, num_bytes);
 895         if (em) {
 896                 /*
 897                  * if block start isn't an actual block number then find the
 898                  * first block in this inode and use that as a hint.  If that
 899                  * block is also bogus then just don't worry about it.
 900                  */
 901                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 902                         free_extent_map(em);
 903                         em = search_extent_mapping(em_tree, 0, 0);
 904                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 905                                 alloc_hint = em->block_start;
 906                         if (em)
 907                                 free_extent_map(em);
 908                 } else {
 909                         alloc_hint = em->block_start;
 910                         free_extent_map(em);
 911                 }
 912         }
 913         read_unlock(&em_tree->lock);
 914
 915         return alloc_hint;
 916 }
 917
 918 /*
 919  * when extent_io.c finds a delayed allocation range in the file,
 920  * the call backs end up in this code.  The basic idea is to
 921  * allocate extents on disk for the range, and create ordered data structs
 922  * in ram to track those extents.
 923  *
 924  * locked_page is the page that writepage had locked already.  We use
 925  * it to make sure we don't do extra locks or unlocks.
 926  *
 927  * *page_started is set to one if we unlock locked_page and do everything
 928  * required to start IO on it.  It may be clean and already done with
 929  * IO when we return.
 930  */
 931 static noinline int cow_file_range(struct inode *inode,
 932                                    struct page *locked_page,
 933                                    u64 start, u64 end, u64 delalloc_end,
 934                                    int *page_started, unsigned long *nr_written,
 935                                    int unlock, struct btrfs_dedupe_hash *hash)
 936 {
 937         struct btrfs_root *root = BTRFS_I(inode)->root;
 938         u64 alloc_hint = 0;
 939         u64 num_bytes;
 940         unsigned long ram_size;
 941         u64 disk_num_bytes;
 942         u64 cur_alloc_size;
 943         u64 blocksize = root->sectorsize;
 944         struct btrfs_key ins;
 945         struct extent_map *em;
 946         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 947         int ret = 0;
 948
 949         if (btrfs_is_free_space_inode(inode)) {
 950                 WARN_ON_ONCE(1);
 951                 ret = -EINVAL;
 952                 goto out_unlock;
 953         }
 954
 955         num_bytes = ALIGN(end - start + 1, blocksize);
 956         num_bytes = max(blocksize,  num_bytes);
 957         disk_num_bytes = num_bytes;
 958
 959         /* if this is a small write inside eof, kick off defrag */
 960         if (num_bytes < SZ_64K &&
 961             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 962                 btrfs_add_inode_defrag(NULL, inode);
 963
 964         if (start == 0) {
 965                 /* lets try to make an inline extent */
 966                 ret = cow_file_range_inline(root, inode, start, end, 0, 0,
 967                                             NULL);
 968                 if (ret == 0) {
 969                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 970                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 971                                      EXTENT_DEFRAG, PAGE_UNLOCK |
 972                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 973                                      PAGE_END_WRITEBACK);
 974                         btrfs_free_reserved_data_space_noquota(inode, start,
 975                                                 end - start + 1);
 976                         *nr_written = *nr_written +
 977                              (end - start + PAGE_SIZE) / PAGE_SIZE;
 978                         *page_started = 1;
 979                         goto out;
 980                 } else if (ret < 0) {
 981                         goto out_unlock;
 982                 }
 983         }
 984
 985         BUG_ON(disk_num_bytes >
 986                btrfs_super_total_bytes(root->fs_info->super_copy));
 987
 988         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 989         btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
 990
 991         while (disk_num_bytes > 0) {
 992                 unsigned long op;
 993
 994                 cur_alloc_size = disk_num_bytes;
 995                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
 996                                            root->sectorsize, 0, alloc_hint,
 997                                            &ins, 1, 1);
 998                 if (ret < 0)
 999                         goto out_unlock;
1000
1001                 em = alloc_extent_map();
1002                 if (!em) {
1003                         ret = -ENOMEM;
1004                         goto out_reserve;
1005                 }
1006                 em->start = start;
1007                 em->orig_start = em->start;
1008                 ram_size = ins.offset;
1009                 em->len = ins.offset;
1010                 em->mod_start = em->start;
1011                 em->mod_len = em->len;
1012
1013                 em->block_start = ins.objectid;
1014                 em->block_len = ins.offset;
1015                 em->orig_block_len = ins.offset;
1016                 em->ram_bytes = ram_size;
1017                 em->bdev = root->fs_info->fs_devices->latest_bdev;
1018                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
1019                 em->generation = -1;
1020
1021                 while (1) {
1022                         write_lock(&em_tree->lock);
1023                         ret = add_extent_mapping(em_tree, em, 1);
1024                         write_unlock(&em_tree->lock);
1025                         if (ret != -EEXIST) {
1026                                 free_extent_map(em);
1027                                 break;
1028                         }
1029                         btrfs_drop_extent_cache(inode, start,
1030                                                 start + ram_size - 1, 0);
1031                 }
1032                 if (ret)
1033                         goto out_reserve;
1034
1035                 cur_alloc_size = ins.offset;
1036                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1037                                                ram_size, cur_alloc_size, 0);
1038                 if (ret)
1039                         goto out_drop_extent_cache;
1040
1041                 if (root->root_key.objectid ==
1042                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1043                         ret = btrfs_reloc_clone_csums(inode, start,
1044                                                       cur_alloc_size);
1045                         if (ret)
1046                                 goto out_drop_extent_cache;
1047                 }
1048
1049                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
1050
1051                 if (disk_num_bytes < cur_alloc_size)
1052                         break;
1053
1054                 /* we're not doing compressed IO, don't unlock the first
1055                  * page (which the caller expects to stay locked), don't
1056                  * clear any dirty bits and don't set any writeback bits
1057                  *
1058                  * Do set the Private2 bit so we know this page was properly
1059                  * setup for writepage
1060                  */
1061                 op = unlock ? PAGE_UNLOCK : 0;
1062                 op |= PAGE_SET_PRIVATE2;
1063
1064                 extent_clear_unlock_delalloc(inode, start,
1065                                              start + ram_size - 1, locked_page,
1066                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1067                                              op);
1068                 disk_num_bytes -= cur_alloc_size;
1069                 num_bytes -= cur_alloc_size;
1070                 alloc_hint = ins.objectid + ins.offset;
1071                 start += cur_alloc_size;
1072         }
1073 out:
1074         return ret;
1075
1076 out_drop_extent_cache:
1077         btrfs_drop_extent_cache(inode, start, start + ram_size - 1, 0);
1078 out_reserve:
1079         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
1080         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
1081 out_unlock:
1082         extent_clear_unlock_delalloc(inode, start, end, locked_page,
1083                                      EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
1084                                      EXTENT_DELALLOC | EXTENT_DEFRAG,
1085                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1086                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK);
1087         goto out;
1088 }
1089
1090 /*
1091  * work queue call back to started compression on a file and pages
1092  */
1093 static noinline void async_cow_start(struct btrfs_work *work)
1094 {
1095         struct async_cow *async_cow;
1096         int num_added = 0;
1097         async_cow = container_of(work, struct async_cow, work);
1098
1099         compress_file_range(async_cow->inode, async_cow->locked_page,
1100                             async_cow->start, async_cow->end, async_cow,
1101                             &num_added);
1102         if (num_added == 0) {
1103                 btrfs_add_delayed_iput(async_cow->inode);
1104                 async_cow->inode = NULL;
1105         }
1106 }
1107
1108 /*
1109  * work queue call back to submit previously compressed pages
1110  */
1111 static noinline void async_cow_submit(struct btrfs_work *work)
1112 {
1113         struct async_cow *async_cow;
1114         struct btrfs_root *root;
1115         unsigned long nr_pages;
1116
1117         async_cow = container_of(work, struct async_cow, work);
1118
1119         root = async_cow->root;
1120         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1121                 PAGE_SHIFT;
1122
1123         /*
1124          * atomic_sub_return implies a barrier for waitqueue_active
1125          */
1126         if (atomic_sub_return(nr_pages, &root->fs_info->async_delalloc_pages) <
1127             5 * SZ_1M &&
1128             waitqueue_active(&root->fs_info->async_submit_wait))
1129                 wake_up(&root->fs_info->async_submit_wait);
1130
1131         if (async_cow->inode)
1132                 submit_compressed_extents(async_cow->inode, async_cow);
1133 }
1134
1135 static noinline void async_cow_free(struct btrfs_work *work)
1136 {
1137         struct async_cow *async_cow;
1138         async_cow = container_of(work, struct async_cow, work);
1139         if (async_cow->inode)
1140                 btrfs_add_delayed_iput(async_cow->inode);
1141         kfree(async_cow);
1142 }
1143
1144 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1145                                 u64 start, u64 end, int *page_started,
1146                                 unsigned long *nr_written)
1147 {
1148         struct async_cow *async_cow;
1149         struct btrfs_root *root = BTRFS_I(inode)->root;
1150         unsigned long nr_pages;
1151         u64 cur_end;
1152         int limit = 10 * SZ_1M;
1153
1154         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1155                          1, 0, NULL, GFP_NOFS);
1156         while (start < end) {
1157                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1158                 BUG_ON(!async_cow); /* -ENOMEM */
1159                 async_cow->inode = igrab(inode);
1160                 async_cow->root = root;
1161                 async_cow->locked_page = locked_page;
1162                 async_cow->start = start;
1163
1164                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1165                     !btrfs_test_opt(root->fs_info, FORCE_COMPRESS))
1166                         cur_end = end;
1167                 else
1168                         cur_end = min(end, start + SZ_512K - 1);
1169
1170                 async_cow->end = cur_end;
1171                 INIT_LIST_HEAD(&async_cow->extents);
1172
1173                 btrfs_init_work(&async_cow->work,
1174                                 btrfs_delalloc_helper,
1175                                 async_cow_start, async_cow_submit,
1176                                 async_cow_free);
1177
1178                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1179                         PAGE_SHIFT;
1180                 atomic_add(nr_pages, &root->fs_info->async_delalloc_pages);
1181
1182                 btrfs_queue_work(root->fs_info->delalloc_workers,
1183                                  &async_cow->work);
1184
1185                 if (atomic_read(&root->fs_info->async_delalloc_pages) > limit) {
1186                         wait_event(root->fs_info->async_submit_wait,
1187                            (atomic_read(&root->fs_info->async_delalloc_pages) <
1188                             limit));
1189                 }
1190
1191                 while (atomic_read(&root->fs_info->async_submit_draining) &&
1192                       atomic_read(&root->fs_info->async_delalloc_pages)) {
1193                         wait_event(root->fs_info->async_submit_wait,
1194                           (atomic_read(&root->fs_info->async_delalloc_pages) ==
1195                            0));
1196                 }
1197
1198                 *nr_written += nr_pages;
1199                 start = cur_end + 1;
1200         }
1201         *page_started = 1;
1202         return 0;
1203 }
1204
1205 static noinline int csum_exist_in_range(struct btrfs_root *root,
1206                                         u64 bytenr, u64 num_bytes)
1207 {
1208         int ret;
1209         struct btrfs_ordered_sum *sums;
1210         LIST_HEAD(list);
1211
1212         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, bytenr,
1213                                        bytenr + num_bytes - 1, &list, 0);
1214         if (ret == 0 && list_empty(&list))
1215                 return 0;
1216
1217         while (!list_empty(&list)) {
1218                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1219                 list_del(&sums->list);
1220                 kfree(sums);
1221         }
1222         return 1;
1223 }
1224
1225 /*
1226  * when nowcow writeback call back.  This checks for snapshots or COW copies
1227  * of the extents that exist in the file, and COWs the file as required.
1228  *
1229  * If no cow copies or snapshots exist, we write directly to the existing
1230  * blocks on disk
1231  */
1232 static noinline int run_delalloc_nocow(struct inode *inode,
1233                                        struct page *locked_page,
1234                               u64 start, u64 end, int *page_started, int force,
1235                               unsigned long *nr_written)
1236 {
1237         struct btrfs_root *root = BTRFS_I(inode)->root;
1238         struct btrfs_trans_handle *trans;
1239         struct extent_buffer *leaf;
1240         struct btrfs_path *path;
1241         struct btrfs_file_extent_item *fi;
1242         struct btrfs_key found_key;
1243         u64 cow_start;
1244         u64 cur_offset;
1245         u64 extent_end;
1246         u64 extent_offset;
1247         u64 disk_bytenr;
1248         u64 num_bytes;
1249         u64 disk_num_bytes;
1250         u64 ram_bytes;
1251         int extent_type;
1252         int ret, err;
1253         int type;
1254         int nocow;
1255         int check_prev = 1;
1256         bool nolock;
1257         u64 ino = btrfs_ino(inode);
1258
1259         path = btrfs_alloc_path();
1260         if (!path) {
1261                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1262                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1263                                              EXTENT_DO_ACCOUNTING |
1264                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1265                                              PAGE_CLEAR_DIRTY |
1266                                              PAGE_SET_WRITEBACK |
1267                                              PAGE_END_WRITEBACK);
1268                 return -ENOMEM;
1269         }
1270
1271         nolock = btrfs_is_free_space_inode(inode);
1272
1273         if (nolock)
1274                 trans = btrfs_join_transaction_nolock(root);
1275         else
1276                 trans = btrfs_join_transaction(root);
1277
1278         if (IS_ERR(trans)) {
1279                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1280                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1281                                              EXTENT_DO_ACCOUNTING |
1282                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1283                                              PAGE_CLEAR_DIRTY |
1284                                              PAGE_SET_WRITEBACK |
1285                                              PAGE_END_WRITEBACK);
1286                 btrfs_free_path(path);
1287                 return PTR_ERR(trans);
1288         }
1289
1290         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1291
1292         cow_start = (u64)-1;
1293         cur_offset = start;
1294         while (1) {
1295                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
1296                                                cur_offset, 0);
1297                 if (ret < 0)
1298                         goto error;
1299                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1300                         leaf = path->nodes[0];
1301                         btrfs_item_key_to_cpu(leaf, &found_key,
1302                                               path->slots[0] - 1);
1303                         if (found_key.objectid == ino &&
1304                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1305                                 path->slots[0]--;
1306                 }
1307                 check_prev = 0;
1308 next_slot:
1309                 leaf = path->nodes[0];
1310                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1311                         ret = btrfs_next_leaf(root, path);
1312                         if (ret < 0)
1313                                 goto error;
1314                         if (ret > 0)
1315                                 break;
1316                         leaf = path->nodes[0];
1317                 }
1318
1319                 nocow = 0;
1320                 disk_bytenr = 0;
1321                 num_bytes = 0;
1322                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1323
1324                 if (found_key.objectid > ino)
1325                         break;
1326                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1327                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1328                         path->slots[0]++;
1329                         goto next_slot;
1330                 }
1331                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1332                     found_key.offset > end)
1333                         break;
1334
1335                 if (found_key.offset > cur_offset) {
1336                         extent_end = found_key.offset;
1337                         extent_type = 0;
1338                         goto out_check;
1339                 }
1340
1341                 fi = btrfs_item_ptr(leaf, path->slots[0],
1342                                     struct btrfs_file_extent_item);
1343                 extent_type = btrfs_file_extent_type(leaf, fi);
1344
1345                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1346                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1347                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1348                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1349                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1350                         extent_end = found_key.offset +
1351                                 btrfs_file_extent_num_bytes(leaf, fi);
1352                         disk_num_bytes =
1353                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1354                         if (extent_end <= start) {
1355                                 path->slots[0]++;
1356                                 goto next_slot;
1357                         }
1358                         if (disk_bytenr == 0)
1359                                 goto out_check;
1360                         if (btrfs_file_extent_compression(leaf, fi) ||
1361                             btrfs_file_extent_encryption(leaf, fi) ||
1362                             btrfs_file_extent_other_encoding(leaf, fi))
1363                                 goto out_check;
1364                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1365                                 goto out_check;
1366                         if (btrfs_extent_readonly(root, disk_bytenr))
1367                                 goto out_check;
1368                         if (btrfs_cross_ref_exist(trans, root, ino,
1369                                                   found_key.offset -
1370                                                   extent_offset, disk_bytenr))
1371                                 goto out_check;
1372                         disk_bytenr += extent_offset;
1373                         disk_bytenr += cur_offset - found_key.offset;
1374                         num_bytes = min(end + 1, extent_end) - cur_offset;
1375                         /*
1376                          * if there are pending snapshots for this root,
1377                          * we fall into common COW way.
1378                          */
1379                         if (!nolock) {
1380                                 err = btrfs_start_write_no_snapshoting(root);
1381                                 if (!err)
1382                                         goto out_check;
1383                         }
1384                         /*
1385                          * force cow if csum exists in the range.
1386                          * this ensure that csum for a given extent are
1387                          * either valid or do not exist.
1388                          */
1389                         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
1390                                 goto out_check;
1391                         if (!btrfs_inc_nocow_writers(root->fs_info,
1392                                                      disk_bytenr))
1393                                 goto out_check;
1394                         nocow = 1;
1395                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1396                         extent_end = found_key.offset +
1397                                 btrfs_file_extent_inline_len(leaf,
1398                                                      path->slots[0], fi);
1399                         extent_end = ALIGN(extent_end, root->sectorsize);
1400                 } else {
1401                         BUG_ON(1);
1402                 }
1403 out_check:
1404                 if (extent_end <= start) {
1405                         path->slots[0]++;
1406                         if (!nolock && nocow)
1407                                 btrfs_end_write_no_snapshoting(root);
1408                         if (nocow)
1409                                 btrfs_dec_nocow_writers(root->fs_info,
1410                                                         disk_bytenr);
1411                         goto next_slot;
1412                 }
1413                 if (!nocow) {
1414                         if (cow_start == (u64)-1)
1415                                 cow_start = cur_offset;
1416                         cur_offset = extent_end;
1417                         if (cur_offset > end)
1418                                 break;
1419                         path->slots[0]++;
1420                         goto next_slot;
1421                 }
1422
1423                 btrfs_release_path(path);
1424                 if (cow_start != (u64)-1) {
1425                         ret = cow_file_range(inode, locked_page,
1426                                              cow_start, found_key.offset - 1,
1427                                              end, page_started, nr_written, 1,
1428                                              NULL);
1429                         if (ret) {
1430                                 if (!nolock && nocow)
1431                                         btrfs_end_write_no_snapshoting(root);
1432                                 if (nocow)
1433                                         btrfs_dec_nocow_writers(root->fs_info,
1434                                                                 disk_bytenr);
1435                                 goto error;
1436                         }
1437                         cow_start = (u64)-1;
1438                 }
1439
1440                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1441                         struct extent_map *em;
1442                         struct extent_map_tree *em_tree;
1443                         em_tree = &BTRFS_I(inode)->extent_tree;
1444                         em = alloc_extent_map();
1445                         BUG_ON(!em); /* -ENOMEM */
1446                         em->start = cur_offset;
1447                         em->orig_start = found_key.offset - extent_offset;
1448                         em->len = num_bytes;
1449                         em->block_len = num_bytes;
1450                         em->block_start = disk_bytenr;
1451                         em->orig_block_len = disk_num_bytes;
1452                         em->ram_bytes = ram_bytes;
1453                         em->bdev = root->fs_info->fs_devices->latest_bdev;
1454                         em->mod_start = em->start;
1455                         em->mod_len = em->len;
1456                         set_bit(EXTENT_FLAG_PINNED, &em->flags);
1457                         set_bit(EXTENT_FLAG_FILLING, &em->flags);
1458                         em->generation = -1;
1459                         while (1) {
1460                                 write_lock(&em_tree->lock);
1461                                 ret = add_extent_mapping(em_tree, em, 1);
1462                                 write_unlock(&em_tree->lock);
1463                                 if (ret != -EEXIST) {
1464                                         free_extent_map(em);
1465                                         break;
1466                                 }
1467                                 btrfs_drop_extent_cache(inode, em->start,
1468                                                 em->start + em->len - 1, 0);
1469                         }
1470                         type = BTRFS_ORDERED_PREALLOC;
1471                 } else {
1472                         type = BTRFS_ORDERED_NOCOW;
1473                 }
1474
1475                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1476                                                num_bytes, num_bytes, type);
1477                 if (nocow)
1478                         btrfs_dec_nocow_writers(root->fs_info, disk_bytenr);
1479                 BUG_ON(ret); /* -ENOMEM */
1480
1481                 if (root->root_key.objectid ==
1482                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1483                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1484                                                       num_bytes);
1485                         if (ret) {
1486                                 if (!nolock && nocow)
1487                                         btrfs_end_write_no_snapshoting(root);
1488                                 goto error;
1489                         }
1490                 }
1491
1492                 extent_clear_unlock_delalloc(inode, cur_offset,
1493                                              cur_offset + num_bytes - 1,
1494                                              locked_page, EXTENT_LOCKED |
1495                                              EXTENT_DELALLOC |
1496                                              EXTENT_CLEAR_DATA_RESV,
1497                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1498
1499                 if (!nolock && nocow)
1500                         btrfs_end_write_no_snapshoting(root);
1501                 cur_offset = extent_end;
1502                 if (cur_offset > end)
1503                         break;
1504         }
1505         btrfs_release_path(path);
1506
1507         if (cur_offset <= end && cow_start == (u64)-1) {
1508                 cow_start = cur_offset;
1509                 cur_offset = end;
1510         }
1511
1512         if (cow_start != (u64)-1) {
1513                 ret = cow_file_range(inode, locked_page, cow_start, end, end,
1514                                      page_started, nr_written, 1, NULL);
1515                 if (ret)
1516                         goto error;
1517         }
1518
1519 error:
1520         err = btrfs_end_transaction(trans, root);
1521         if (!ret)
1522                 ret = err;
1523
1524         if (ret && cur_offset < end)
1525                 extent_clear_unlock_delalloc(inode, cur_offset, end,
1526                                              locked_page, EXTENT_LOCKED |
1527                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1528                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1529                                              PAGE_CLEAR_DIRTY |
1530                                              PAGE_SET_WRITEBACK |
1531                                              PAGE_END_WRITEBACK);
1532         btrfs_free_path(path);
1533         return ret;
1534 }
1535
1536 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1537 {
1538
1539         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1540             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1541                 return 0;
1542
1543         /*
1544          * @defrag_bytes is a hint value, no spinlock held here,
1545          * if is not zero, it means the file is defragging.
1546          * Force cow if given extent needs to be defragged.
1547          */
1548         if (BTRFS_I(inode)->defrag_bytes &&
1549             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1550                            EXTENT_DEFRAG, 0, NULL))
1551                 return 1;
1552
1553         return 0;
1554 }
1555
1556 /*
1557  * extent_io.c call back to do delayed allocation processing
1558  */
1559 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1560                               u64 start, u64 end, int *page_started,
1561                               unsigned long *nr_written)
1562 {
1563         int ret;
1564         int force_cow = need_force_cow(inode, start, end);
1565
1566         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1567                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1568                                          page_started, 1, nr_written);
1569         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1570                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1571                                          page_started, 0, nr_written);
1572         } else if (!inode_need_compress(inode)) {
1573                 ret = cow_file_range(inode, locked_page, start, end, end,
1574                                       page_started, nr_written, 1, NULL);
1575         } else {
1576                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1577                         &BTRFS_I(inode)->runtime_flags);
1578                 ret = cow_file_range_async(inode, locked_page, start, end,
1579                                            page_started, nr_written);
1580         }
1581         return ret;
1582 }
1583
1584 static void btrfs_split_extent_hook(struct inode *inode,
1585                                     struct extent_state *orig, u64 split)
1586 {
1587         u64 size;
1588
1589         /* not delalloc, ignore it */
1590         if (!(orig->state & EXTENT_DELALLOC))
1591                 return;
1592
1593         size = orig->end - orig->start + 1;
1594         if (size > BTRFS_MAX_EXTENT_SIZE) {
1595                 u64 num_extents;
1596                 u64 new_size;
1597
1598                 /*
1599                  * See the explanation in btrfs_merge_extent_hook, the same
1600                  * applies here, just in reverse.
1601                  */
1602                 new_size = orig->end - split + 1;
1603                 num_extents = div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1604                                         BTRFS_MAX_EXTENT_SIZE);
1605                 new_size = split - orig->start;
1606                 num_extents += div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1607                                         BTRFS_MAX_EXTENT_SIZE);
1608                 if (div64_u64(size + BTRFS_MAX_EXTENT_SIZE - 1,
1609                               BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1610                         return;
1611         }
1612
1613         spin_lock(&BTRFS_I(inode)->lock);
1614         BTRFS_I(inode)->outstanding_extents++;
1615         spin_unlock(&BTRFS_I(inode)->lock);
1616 }
1617
1618 /*
1619  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1620  * extents so we can keep track of new extents that are just merged onto old
1621  * extents, such as when we are doing sequential writes, so we can properly
1622  * account for the metadata space we'll need.
1623  */
1624 static void btrfs_merge_extent_hook(struct inode *inode,
1625                                     struct extent_state *new,
1626                                     struct extent_state *other)
1627 {
1628         u64 new_size, old_size;
1629         u64 num_extents;
1630
1631         /* not delalloc, ignore it */
1632         if (!(other->state & EXTENT_DELALLOC))
1633                 return;
1634
1635         if (new->start > other->start)
1636                 new_size = new->end - other->start + 1;
1637         else
1638                 new_size = other->end - new->start + 1;
1639
1640         /* we're not bigger than the max, unreserve the space and go */
1641         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1642                 spin_lock(&BTRFS_I(inode)->lock);
1643                 BTRFS_I(inode)->outstanding_extents--;
1644                 spin_unlock(&BTRFS_I(inode)->lock);
1645                 return;
1646         }
1647
1648         /*
1649          * We have to add up either side to figure out how many extents were
1650          * accounted for before we merged into one big extent.  If the number of
1651          * extents we accounted for is <= the amount we need for the new range
1652          * then we can return, otherwise drop.  Think of it like this
1653          *
1654          * [ 4k][MAX_SIZE]
1655          *
1656          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1657          * need 2 outstanding extents, on one side we have 1 and the other side
1658          * we have 1 so they are == and we can return.  But in this case
1659          *
1660          * [MAX_SIZE+4k][MAX_SIZE+4k]
1661          *
1662          * Each range on their own accounts for 2 extents, but merged together
1663          * they are only 3 extents worth of accounting, so we need to drop in
1664          * this case.
1665          */
1666         old_size = other->end - other->start + 1;
1667         num_extents = div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1668                                 BTRFS_MAX_EXTENT_SIZE);
1669         old_size = new->end - new->start + 1;
1670         num_extents += div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1671                                  BTRFS_MAX_EXTENT_SIZE);
1672
1673         if (div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1674                       BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1675                 return;
1676
1677         spin_lock(&BTRFS_I(inode)->lock);
1678         BTRFS_I(inode)->outstanding_extents--;
1679         spin_unlock(&BTRFS_I(inode)->lock);
1680 }
1681
1682 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1683                                       struct inode *inode)
1684 {
1685         spin_lock(&root->delalloc_lock);
1686         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1687                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1688                               &root->delalloc_inodes);
1689                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1690                         &BTRFS_I(inode)->runtime_flags);
1691                 root->nr_delalloc_inodes++;
1692                 if (root->nr_delalloc_inodes == 1) {
1693                         spin_lock(&root->fs_info->delalloc_root_lock);
1694                         BUG_ON(!list_empty(&root->delalloc_root));
1695                         list_add_tail(&root->delalloc_root,
1696                                       &root->fs_info->delalloc_roots);
1697                         spin_unlock(&root->fs_info->delalloc_root_lock);
1698                 }
1699         }
1700         spin_unlock(&root->delalloc_lock);
1701 }
1702
1703 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1704                                      struct inode *inode)
1705 {
1706         spin_lock(&root->delalloc_lock);
1707         if (!list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1708                 list_del_init(&BTRFS_I(inode)->delalloc_inodes);
1709                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1710                           &BTRFS_I(inode)->runtime_flags);
1711                 root->nr_delalloc_inodes--;
1712                 if (!root->nr_delalloc_inodes) {
1713                         spin_lock(&root->fs_info->delalloc_root_lock);
1714                         BUG_ON(list_empty(&root->delalloc_root));
1715                         list_del_init(&root->delalloc_root);
1716                         spin_unlock(&root->fs_info->delalloc_root_lock);
1717                 }
1718         }
1719         spin_unlock(&root->delalloc_lock);
1720 }
1721
1722 /*
1723  * extent_io.c set_bit_hook, used to track delayed allocation
1724  * bytes in this file, and to maintain the list of inodes that
1725  * have pending delalloc work to be done.
1726  */
1727 static void btrfs_set_bit_hook(struct inode *inode,
1728                                struct extent_state *state, unsigned *bits)
1729 {
1730
1731         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1732                 WARN_ON(1);
1733         /*
1734          * set_bit and clear bit hooks normally require _irqsave/restore
1735          * but in this case, we are only testing for the DELALLOC
1736          * bit, which is only set or cleared with irqs on
1737          */
1738         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1739                 struct btrfs_root *root = BTRFS_I(inode)->root;
1740                 u64 len = state->end + 1 - state->start;
1741                 bool do_list = !btrfs_is_free_space_inode(inode);
1742
1743                 if (*bits & EXTENT_FIRST_DELALLOC) {
1744                         *bits &= ~EXTENT_FIRST_DELALLOC;
1745                 } else {
1746                         spin_lock(&BTRFS_I(inode)->lock);
1747                         BTRFS_I(inode)->outstanding_extents++;
1748                         spin_unlock(&BTRFS_I(inode)->lock);
1749                 }
1750
1751                 /* For sanity tests */
1752                 if (btrfs_is_testing(root->fs_info))
1753                         return;
1754
1755                 __percpu_counter_add(&root->fs_info->delalloc_bytes, len,
1756                                      root->fs_info->delalloc_batch);
1757                 spin_lock(&BTRFS_I(inode)->lock);
1758                 BTRFS_I(inode)->delalloc_bytes += len;
1759                 if (*bits & EXTENT_DEFRAG)
1760                         BTRFS_I(inode)->defrag_bytes += len;
1761                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1762                                          &BTRFS_I(inode)->runtime_flags))
1763                         btrfs_add_delalloc_inodes(root, inode);
1764                 spin_unlock(&BTRFS_I(inode)->lock);
1765         }
1766 }
1767
1768 /*
1769  * extent_io.c clear_bit_hook, see set_bit_hook for why
1770  */
1771 static void btrfs_clear_bit_hook(struct inode *inode,
1772                                  struct extent_state *state,
1773                                  unsigned *bits)
1774 {
1775         u64 len = state->end + 1 - state->start;
1776         u64 num_extents = div64_u64(len + BTRFS_MAX_EXTENT_SIZE -1,
1777                                     BTRFS_MAX_EXTENT_SIZE);
1778
1779         spin_lock(&BTRFS_I(inode)->lock);
1780         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG))
1781                 BTRFS_I(inode)->defrag_bytes -= len;
1782         spin_unlock(&BTRFS_I(inode)->lock);
1783
1784         /*
1785          * set_bit and clear bit hooks normally require _irqsave/restore
1786          * but in this case, we are only testing for the DELALLOC
1787          * bit, which is only set or cleared with irqs on
1788          */
1789         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1790                 struct btrfs_root *root = BTRFS_I(inode)->root;
1791                 bool do_list = !btrfs_is_free_space_inode(inode);
1792
1793                 if (*bits & EXTENT_FIRST_DELALLOC) {
1794                         *bits &= ~EXTENT_FIRST_DELALLOC;
1795                 } else if (!(*bits & EXTENT_DO_ACCOUNTING)) {
1796                         spin_lock(&BTRFS_I(inode)->lock);
1797                         BTRFS_I(inode)->outstanding_extents -= num_extents;
1798                         spin_unlock(&BTRFS_I(inode)->lock);
1799                 }
1800
1801                 /*
1802                  * We don't reserve metadata space for space cache inodes so we
1803                  * don't need to call dellalloc_release_metadata if there is an
1804                  * error.
1805                  */
1806                 if (*bits & EXTENT_DO_ACCOUNTING &&
1807                     root != root->fs_info->tree_root)
1808                         btrfs_delalloc_release_metadata(inode, len);
1809
1810                 /* For sanity tests. */
1811                 if (btrfs_is_testing(root->fs_info))
1812                         return;
1813
1814                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
1815                     && do_list && !(state->state & EXTENT_NORESERVE)
1816                     && (*bits & (EXTENT_DO_ACCOUNTING |
1817                     EXTENT_CLEAR_DATA_RESV)))
1818                         btrfs_free_reserved_data_space_noquota(inode,
1819                                         state->start, len);
1820
1821                 __percpu_counter_add(&root->fs_info->delalloc_bytes, -len,
1822                                      root->fs_info->delalloc_batch);
1823                 spin_lock(&BTRFS_I(inode)->lock);
1824                 BTRFS_I(inode)->delalloc_bytes -= len;
1825                 if (do_list && BTRFS_I(inode)->delalloc_bytes == 0 &&
1826                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1827                              &BTRFS_I(inode)->runtime_flags))
1828                         btrfs_del_delalloc_inode(root, inode);
1829                 spin_unlock(&BTRFS_I(inode)->lock);
1830         }
1831 }
1832
1833 /*
1834  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1835  * we don't create bios that span stripes or chunks
1836  *
1837  * return 1 if page cannot be merged to bio
1838  * return 0 if page can be merged to bio
1839  * return error otherwise
1840  */
1841 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1842                          size_t size, struct bio *bio,
1843                          unsigned long bio_flags)
1844 {
1845         struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
1846         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1847         u64 length = 0;
1848         u64 map_length;
1849         int ret;
1850
1851         if (bio_flags & EXTENT_BIO_COMPRESSED)
1852                 return 0;
1853
1854         length = bio->bi_iter.bi_size;
1855         map_length = length;
1856         ret = btrfs_map_block(root->fs_info, bio_op(bio), logical,
1857                               &map_length, NULL, 0);
1858         if (ret < 0)
1859                 return ret;
1860         if (map_length < length + size)
1861                 return 1;
1862         return 0;
1863 }
1864
1865 /*
1866  * in order to insert checksums into the metadata in large chunks,
1867  * we wait until bio submission time.   All the pages in the bio are
1868  * checksummed and sums are attached onto the ordered extent record.
1869  *
1870  * At IO completion time the cums attached on the ordered extent record
1871  * are inserted into the btree
1872  */
1873 static int __btrfs_submit_bio_start(struct inode *inode, struct bio *bio,
1874                                     int mirror_num, unsigned long bio_flags,
1875                                     u64 bio_offset)
1876 {
1877         struct btrfs_root *root = BTRFS_I(inode)->root;
1878         int ret = 0;
1879
1880         ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1881         BUG_ON(ret); /* -ENOMEM */
1882         return 0;
1883 }
1884
1885 /*
1886  * in order to insert checksums into the metadata in large chunks,
1887  * we wait until bio submission time.   All the pages in the bio are
1888  * checksummed and sums are attached onto the ordered extent record.
1889  *
1890  * At IO completion time the cums attached on the ordered extent record
1891  * are inserted into the btree
1892  */
1893 static int __btrfs_submit_bio_done(struct inode *inode, struct bio *bio,
1894                           int mirror_num, unsigned long bio_flags,
1895                           u64 bio_offset)
1896 {
1897         struct btrfs_root *root = BTRFS_I(inode)->root;
1898         int ret;
1899
1900         ret = btrfs_map_bio(root, bio, mirror_num, 1);
1901         if (ret) {
1902                 bio->bi_error = ret;
1903                 bio_endio(bio);
1904         }
1905         return ret;
1906 }
1907
1908 /*
1909  * extent_io.c submission hook. This does the right thing for csum calculation
1910  * on write, or reading the csums from the tree before a read
1911  */
1912 static int btrfs_submit_bio_hook(struct inode *inode, struct bio *bio,
1913                           int mirror_num, unsigned long bio_flags,
1914                           u64 bio_offset)
1915 {
1916         struct btrfs_root *root = BTRFS_I(inode)->root;
1917         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1918         int ret = 0;
1919         int skip_sum;
1920         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1921
1922         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1923
1924         if (btrfs_is_free_space_inode(inode))
1925                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1926
1927         if (bio_op(bio) != REQ_OP_WRITE) {
1928                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, metadata);
1929                 if (ret)
1930                         goto out;
1931
1932                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1933                         ret = btrfs_submit_compressed_read(inode, bio,
1934                                                            mirror_num,
1935                                                            bio_flags);
1936                         goto out;
1937                 } else if (!skip_sum) {
1938                         ret = btrfs_lookup_bio_sums(root, inode, bio, NULL);
1939                         if (ret)
1940                                 goto out;
1941                 }
1942                 goto mapit;
1943         } else if (async && !skip_sum) {
1944                 /* csum items have already been cloned */
1945                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1946                         goto mapit;
1947                 /* we're doing a write, do the async checksumming */
1948                 ret = btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info,
1949                                    inode, bio, mirror_num,
1950                                    bio_flags, bio_offset,
1951                                    __btrfs_submit_bio_start,
1952                                    __btrfs_submit_bio_done);
1953                 goto out;
1954         } else if (!skip_sum) {
1955                 ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1956                 if (ret)
1957                         goto out;
1958         }
1959
1960 mapit:
1961         ret = btrfs_map_bio(root, bio, mirror_num, 0);
1962
1963 out:
1964         if (ret < 0) {
1965                 bio->bi_error = ret;
1966                 bio_endio(bio);
1967         }
1968         return ret;
1969 }
1970
1971 /*
1972  * given a list of ordered sums record them in the inode.  This happens
1973  * at IO completion time based on sums calculated at bio submission time.
1974  */
1975 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
1976                              struct inode *inode, u64 file_offset,
1977                              struct list_head *list)
1978 {
1979         struct btrfs_ordered_sum *sum;
1980
1981         list_for_each_entry(sum, list, list) {
1982                 trans->adding_csums = 1;
1983                 btrfs_csum_file_blocks(trans,
1984                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
1985                 trans->adding_csums = 0;
1986         }
1987         return 0;
1988 }
1989
1990 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
1991                               struct extent_state **cached_state)
1992 {
1993         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
1994         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
1995                                    cached_state);
1996 }
1997
1998 /* see btrfs_writepage_start_hook for details on why this is required */
1999 struct btrfs_writepage_fixup {
2000         struct page *page;
2001         struct btrfs_work work;
2002 };
2003
2004 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2005 {
2006         struct btrfs_writepage_fixup *fixup;
2007         struct btrfs_ordered_extent *ordered;
2008         struct extent_state *cached_state = NULL;
2009         struct page *page;
2010         struct inode *inode;
2011         u64 page_start;
2012         u64 page_end;
2013         int ret;
2014
2015         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2016         page = fixup->page;
2017 again:
2018         lock_page(page);
2019         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2020                 ClearPageChecked(page);
2021                 goto out_page;
2022         }
2023
2024         inode = page->mapping->host;
2025         page_start = page_offset(page);
2026         page_end = page_offset(page) + PAGE_SIZE - 1;
2027
2028         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2029                          &cached_state);
2030
2031         /* already ordered? We're done */
2032         if (PagePrivate2(page))
2033                 goto out;
2034
2035         ordered = btrfs_lookup_ordered_range(inode, page_start,
2036                                         PAGE_SIZE);
2037         if (ordered) {
2038                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2039                                      page_end, &cached_state, GFP_NOFS);
2040                 unlock_page(page);
2041                 btrfs_start_ordered_extent(inode, ordered, 1);
2042                 btrfs_put_ordered_extent(ordered);
2043                 goto again;
2044         }
2045
2046         ret = btrfs_delalloc_reserve_space(inode, page_start,
2047                                            PAGE_SIZE);
2048         if (ret) {
2049                 mapping_set_error(page->mapping, ret);
2050                 end_extent_writepage(page, ret, page_start, page_end);
2051                 ClearPageChecked(page);
2052                 goto out;
2053          }
2054
2055         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state);
2056         ClearPageChecked(page);
2057         set_page_dirty(page);
2058 out:
2059         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2060                              &cached_state, GFP_NOFS);
2061 out_page:
2062         unlock_page(page);
2063         put_page(page);
2064         kfree(fixup);
2065 }
2066
2067 /*
2068  * There are a few paths in the higher layers of the kernel that directly
2069  * set the page dirty bit without asking the filesystem if it is a
2070  * good idea.  This causes problems because we want to make sure COW
2071  * properly happens and the data=ordered rules are followed.
2072  *
2073  * In our case any range that doesn't have the ORDERED bit set
2074  * hasn't been properly setup for IO.  We kick off an async process
2075  * to fix it up.  The async helper will wait for ordered extents, set
2076  * the delalloc bit and make it safe to write the page.
2077  */
2078 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2079 {
2080         struct inode *inode = page->mapping->host;
2081         struct btrfs_writepage_fixup *fixup;
2082         struct btrfs_root *root = BTRFS_I(inode)->root;
2083
2084         /* this page is properly in the ordered list */
2085         if (TestClearPagePrivate2(page))
2086                 return 0;
2087
2088         if (PageChecked(page))
2089                 return -EAGAIN;
2090
2091         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2092         if (!fixup)
2093                 return -EAGAIN;
2094
2095         SetPageChecked(page);
2096         get_page(page);
2097         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2098                         btrfs_writepage_fixup_worker, NULL, NULL);
2099         fixup->page = page;
2100         btrfs_queue_work(root->fs_info->fixup_workers, &fixup->work);
2101         return -EBUSY;
2102 }
2103
2104 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2105                                        struct inode *inode, u64 file_pos,
2106                                        u64 disk_bytenr, u64 disk_num_bytes,
2107                                        u64 num_bytes, u64 ram_bytes,
2108                                        u8 compression, u8 encryption,
2109                                        u16 other_encoding, int extent_type)
2110 {
2111         struct btrfs_root *root = BTRFS_I(inode)->root;
2112         struct btrfs_file_extent_item *fi;
2113         struct btrfs_path *path;
2114         struct extent_buffer *leaf;
2115         struct btrfs_key ins;
2116         int extent_inserted = 0;
2117         int ret;
2118
2119         path = btrfs_alloc_path();
2120         if (!path)
2121                 return -ENOMEM;
2122
2123         /*
2124          * we may be replacing one extent in the tree with another.
2125          * The new extent is pinned in the extent map, and we don't want
2126          * to drop it from the cache until it is completely in the btree.
2127          *
2128          * So, tell btrfs_drop_extents to leave this extent in the cache.
2129          * the caller is expected to unpin it and allow it to be merged
2130          * with the others.
2131          */
2132         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2133                                    file_pos + num_bytes, NULL, 0,
2134                                    1, sizeof(*fi), &extent_inserted);
2135         if (ret)
2136                 goto out;
2137
2138         if (!extent_inserted) {
2139                 ins.objectid = btrfs_ino(inode);
2140                 ins.offset = file_pos;
2141                 ins.type = BTRFS_EXTENT_DATA_KEY;
2142
2143                 path->leave_spinning = 1;
2144                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2145                                               sizeof(*fi));
2146                 if (ret)
2147                         goto out;
2148         }
2149         leaf = path->nodes[0];
2150         fi = btrfs_item_ptr(leaf, path->slots[0],
2151                             struct btrfs_file_extent_item);
2152         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2153         btrfs_set_file_extent_type(leaf, fi, extent_type);
2154         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2155         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2156         btrfs_set_file_extent_offset(leaf, fi, 0);
2157         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2158         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2159         btrfs_set_file_extent_compression(leaf, fi, compression);
2160         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2161         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2162
2163         btrfs_mark_buffer_dirty(leaf);
2164         btrfs_release_path(path);
2165
2166         inode_add_bytes(inode, num_bytes);
2167
2168         ins.objectid = disk_bytenr;
2169         ins.offset = disk_num_bytes;
2170         ins.type = BTRFS_EXTENT_ITEM_KEY;
2171         ret = btrfs_alloc_reserved_file_extent(trans, root,
2172                                         root->root_key.objectid,
2173                                         btrfs_ino(inode), file_pos,
2174                                         ram_bytes, &ins);
2175         /*
2176          * Release the reserved range from inode dirty range map, as it is
2177          * already moved into delayed_ref_head
2178          */
2179         btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2180 out:
2181         btrfs_free_path(path);
2182
2183         return ret;
2184 }
2185
2186 /* snapshot-aware defrag */
2187 struct sa_defrag_extent_backref {
2188         struct rb_node node;
2189         struct old_sa_defrag_extent *old;
2190         u64 root_id;
2191         u64 inum;
2192         u64 file_pos;
2193         u64 extent_offset;
2194         u64 num_bytes;
2195         u64 generation;
2196 };
2197
2198 struct old_sa_defrag_extent {
2199         struct list_head list;
2200         struct new_sa_defrag_extent *new;
2201
2202         u64 extent_offset;
2203         u64 bytenr;
2204         u64 offset;
2205         u64 len;
2206         int count;
2207 };
2208
2209 struct new_sa_defrag_extent {
2210         struct rb_root root;
2211         struct list_head head;
2212         struct btrfs_path *path;
2213         struct inode *inode;
2214         u64 file_pos;
2215         u64 len;
2216         u64 bytenr;
2217         u64 disk_len;
2218         u8 compress_type;
2219 };
2220
2221 static int backref_comp(struct sa_defrag_extent_backref *b1,
2222                         struct sa_defrag_extent_backref *b2)
2223 {
2224         if (b1->root_id < b2->root_id)
2225                 return -1;
2226         else if (b1->root_id > b2->root_id)
2227                 return 1;
2228
2229         if (b1->inum < b2->inum)
2230                 return -1;
2231         else if (b1->inum > b2->inum)
2232                 return 1;
2233
2234         if (b1->file_pos < b2->file_pos)
2235                 return -1;
2236         else if (b1->file_pos > b2->file_pos)
2237                 return 1;
2238
2239         /*
2240          * [------------------------------] ===> (a range of space)
2241          *     |<--->|   |<---->| =============> (fs/file tree A)
2242          * |<---------------------------->| ===> (fs/file tree B)
2243          *
2244          * A range of space can refer to two file extents in one tree while
2245          * refer to only one file extent in another tree.
2246          *
2247          * So we may process a disk offset more than one time(two extents in A)
2248          * and locate at the same extent(one extent in B), then insert two same
2249          * backrefs(both refer to the extent in B).
2250          */
2251         return 0;
2252 }
2253
2254 static void backref_insert(struct rb_root *root,
2255                            struct sa_defrag_extent_backref *backref)
2256 {
2257         struct rb_node **p = &root->rb_node;
2258         struct rb_node *parent = NULL;
2259         struct sa_defrag_extent_backref *entry;
2260         int ret;
2261
2262         while (*p) {
2263                 parent = *p;
2264                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2265
2266                 ret = backref_comp(backref, entry);
2267                 if (ret < 0)
2268                         p = &(*p)->rb_left;
2269                 else
2270                         p = &(*p)->rb_right;
2271         }
2272
2273         rb_link_node(&backref->node, parent, p);
2274         rb_insert_color(&backref->node, root);
2275 }
2276
2277 /*
2278  * Note the backref might has changed, and in this case we just return 0.
2279  */
2280 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2281                                        void *ctx)
2282 {
2283         struct btrfs_file_extent_item *extent;
2284         struct btrfs_fs_info *fs_info;
2285         struct old_sa_defrag_extent *old = ctx;
2286         struct new_sa_defrag_extent *new = old->new;
2287         struct btrfs_path *path = new->path;
2288         struct btrfs_key key;
2289         struct btrfs_root *root;
2290         struct sa_defrag_extent_backref *backref;
2291         struct extent_buffer *leaf;
2292         struct inode *inode = new->inode;
2293         int slot;
2294         int ret;
2295         u64 extent_offset;
2296         u64 num_bytes;
2297
2298         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2299             inum == btrfs_ino(inode))
2300                 return 0;
2301
2302         key.objectid = root_id;
2303         key.type = BTRFS_ROOT_ITEM_KEY;
2304         key.offset = (u64)-1;
2305
2306         fs_info = BTRFS_I(inode)->root->fs_info;
2307         root = btrfs_read_fs_root_no_name(fs_info, &key);
2308         if (IS_ERR(root)) {
2309                 if (PTR_ERR(root) == -ENOENT)
2310                         return 0;
2311                 WARN_ON(1);
2312                 pr_debug("inum=%llu, offset=%llu, root_id=%llu\n",
2313                          inum, offset, root_id);
2314                 return PTR_ERR(root);
2315         }
2316
2317         key.objectid = inum;
2318         key.type = BTRFS_EXTENT_DATA_KEY;
2319         if (offset > (u64)-1 << 32)
2320                 key.offset = 0;
2321         else
2322                 key.offset = offset;
2323
2324         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2325         if (WARN_ON(ret < 0))
2326                 return ret;
2327         ret = 0;
2328
2329         while (1) {
2330                 cond_resched();
2331
2332                 leaf = path->nodes[0];
2333                 slot = path->slots[0];
2334
2335                 if (slot >= btrfs_header_nritems(leaf)) {
2336                         ret = btrfs_next_leaf(root, path);
2337                         if (ret < 0) {
2338                                 goto out;
2339                         } else if (ret > 0) {
2340                                 ret = 0;
2341                                 goto out;
2342                         }
2343                         continue;
2344                 }
2345
2346                 path->slots[0]++;
2347
2348                 btrfs_item_key_to_cpu(leaf, &key, slot);
2349
2350                 if (key.objectid > inum)
2351                         goto out;
2352
2353                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2354                         continue;
2355
2356                 extent = btrfs_item_ptr(leaf, slot,
2357                                         struct btrfs_file_extent_item);
2358
2359                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2360                         continue;
2361
2362                 /*
2363                  * 'offset' refers to the exact key.offset,
2364                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2365                  * (key.offset - extent_offset).
2366                  */
2367                 if (key.offset != offset)
2368                         continue;
2369
2370                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2371                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2372
2373                 if (extent_offset >= old->extent_offset + old->offset +
2374                     old->len || extent_offset + num_bytes <=
2375                     old->extent_offset + old->offset)
2376                         continue;
2377                 break;
2378         }
2379
2380         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2381         if (!backref) {
2382                 ret = -ENOENT;
2383                 goto out;
2384         }
2385
2386         backref->root_id = root_id;
2387         backref->inum = inum;
2388         backref->file_pos = offset;
2389         backref->num_bytes = num_bytes;
2390         backref->extent_offset = extent_offset;
2391         backref->generation = btrfs_file_extent_generation(leaf, extent);
2392         backref->old = old;
2393         backref_insert(&new->root, backref);
2394         old->count++;
2395 out:
2396         btrfs_release_path(path);
2397         WARN_ON(ret);
2398         return ret;
2399 }
2400
2401 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2402                                    struct new_sa_defrag_extent *new)
2403 {
2404         struct btrfs_fs_info *fs_info = BTRFS_I(new->inode)->root->fs_info;
2405         struct old_sa_defrag_extent *old, *tmp;
2406         int ret;
2407
2408         new->path = path;
2409
2410         list_for_each_entry_safe(old, tmp, &new->head, list) {
2411                 ret = iterate_inodes_from_logical(old->bytenr +
2412                                                   old->extent_offset, fs_info,
2413                                                   path, record_one_backref,
2414                                                   old);
2415                 if (ret < 0 && ret != -ENOENT)
2416                         return false;
2417
2418                 /* no backref to be processed for this extent */
2419                 if (!old->count) {
2420                         list_del(&old->list);
2421                         kfree(old);
2422                 }
2423         }
2424
2425         if (list_empty(&new->head))
2426                 return false;
2427
2428         return true;
2429 }
2430
2431 static int relink_is_mergable(struct extent_buffer *leaf,
2432                               struct btrfs_file_extent_item *fi,
2433                               struct new_sa_defrag_extent *new)
2434 {
2435         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2436                 return 0;
2437
2438         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2439                 return 0;
2440
2441         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2442                 return 0;
2443
2444         if (btrfs_file_extent_encryption(leaf, fi) ||
2445             btrfs_file_extent_other_encoding(leaf, fi))
2446                 return 0;
2447
2448         return 1;
2449 }
2450
2451 /*
2452  * Note the backref might has changed, and in this case we just return 0.
2453  */
2454 static noinline int relink_extent_backref(struct btrfs_path *path,
2455                                  struct sa_defrag_extent_backref *prev,
2456                                  struct sa_defrag_extent_backref *backref)
2457 {
2458         struct btrfs_file_extent_item *extent;
2459         struct btrfs_file_extent_item *item;
2460         struct btrfs_ordered_extent *ordered;
2461         struct btrfs_trans_handle *trans;
2462         struct btrfs_fs_info *fs_info;
2463         struct btrfs_root *root;
2464         struct btrfs_key key;
2465         struct extent_buffer *leaf;
2466         struct old_sa_defrag_extent *old = backref->old;
2467         struct new_sa_defrag_extent *new = old->new;
2468         struct inode *src_inode = new->inode;
2469         struct inode *inode;
2470         struct extent_state *cached = NULL;
2471         int ret = 0;
2472         u64 start;
2473         u64 len;
2474         u64 lock_start;
2475         u64 lock_end;
2476         bool merge = false;
2477         int index;
2478
2479         if (prev && prev->root_id == backref->root_id &&
2480             prev->inum == backref->inum &&
2481             prev->file_pos + prev->num_bytes == backref->file_pos)
2482                 merge = true;
2483
2484         /* step 1: get root */
2485         key.objectid = backref->root_id;
2486         key.type = BTRFS_ROOT_ITEM_KEY;
2487         key.offset = (u64)-1;
2488
2489         fs_info = BTRFS_I(src_inode)->root->fs_info;
2490         index = srcu_read_lock(&fs_info->subvol_srcu);
2491
2492         root = btrfs_read_fs_root_no_name(fs_info, &key);
2493         if (IS_ERR(root)) {
2494                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2495                 if (PTR_ERR(root) == -ENOENT)
2496                         return 0;
2497                 return PTR_ERR(root);
2498         }
2499
2500         if (btrfs_root_readonly(root)) {
2501                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2502                 return 0;
2503         }
2504
2505         /* step 2: get inode */
2506         key.objectid = backref->inum;
2507         key.type = BTRFS_INODE_ITEM_KEY;
2508         key.offset = 0;
2509
2510         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2511         if (IS_ERR(inode)) {
2512                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2513                 return 0;
2514         }
2515
2516         srcu_read_unlock(&fs_info->subvol_srcu, index);
2517
2518         /* step 3: relink backref */
2519         lock_start = backref->file_pos;
2520         lock_end = backref->file_pos + backref->num_bytes - 1;
2521         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2522                          &cached);
2523
2524         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2525         if (ordered) {
2526                 btrfs_put_ordered_extent(ordered);
2527                 goto out_unlock;
2528         }
2529
2530         trans = btrfs_join_transaction(root);
2531         if (IS_ERR(trans)) {
2532                 ret = PTR_ERR(trans);
2533                 goto out_unlock;
2534         }
2535
2536         key.objectid = backref->inum;
2537         key.type = BTRFS_EXTENT_DATA_KEY;
2538         key.offset = backref->file_pos;
2539
2540         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2541         if (ret < 0) {
2542                 goto out_free_path;
2543         } else if (ret > 0) {
2544                 ret = 0;
2545                 goto out_free_path;
2546         }
2547
2548         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2549                                 struct btrfs_file_extent_item);
2550
2551         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2552             backref->generation)
2553                 goto out_free_path;
2554
2555         btrfs_release_path(path);
2556
2557         start = backref->file_pos;
2558         if (backref->extent_offset < old->extent_offset + old->offset)
2559                 start += old->extent_offset + old->offset -
2560                          backref->extent_offset;
2561
2562         len = min(backref->extent_offset + backref->num_bytes,
2563                   old->extent_offset + old->offset + old->len);
2564         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2565
2566         ret = btrfs_drop_extents(trans, root, inode, start,
2567                                  start + len, 1);
2568         if (ret)
2569                 goto out_free_path;
2570 again:
2571         key.objectid = btrfs_ino(inode);
2572         key.type = BTRFS_EXTENT_DATA_KEY;
2573         key.offset = start;
2574
2575         path->leave_spinning = 1;
2576         if (merge) {
2577                 struct btrfs_file_extent_item *fi;
2578                 u64 extent_len;
2579                 struct btrfs_key found_key;
2580
2581                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2582                 if (ret < 0)
2583                         goto out_free_path;
2584
2585                 path->slots[0]--;
2586                 leaf = path->nodes[0];
2587                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2588
2589                 fi = btrfs_item_ptr(leaf, path->slots[0],
2590                                     struct btrfs_file_extent_item);
2591                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2592
2593                 if (extent_len + found_key.offset == start &&
2594                     relink_is_mergable(leaf, fi, new)) {
2595                         btrfs_set_file_extent_num_bytes(leaf, fi,
2596                                                         extent_len + len);
2597                         btrfs_mark_buffer_dirty(leaf);
2598                         inode_add_bytes(inode, len);
2599
2600                         ret = 1;
2601                         goto out_free_path;
2602                 } else {
2603                         merge = false;
2604                         btrfs_release_path(path);
2605                         goto again;
2606                 }
2607         }
2608
2609         ret = btrfs_insert_empty_item(trans, root, path, &key,
2610                                         sizeof(*extent));
2611         if (ret) {
2612                 btrfs_abort_transaction(trans, ret);
2613                 goto out_free_path;
2614         }
2615
2616         leaf = path->nodes[0];
2617         item = btrfs_item_ptr(leaf, path->slots[0],
2618                                 struct btrfs_file_extent_item);
2619         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2620         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2621         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2622         btrfs_set_file_extent_num_bytes(leaf, item, len);
2623         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2624         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2625         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2626         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2627         btrfs_set_file_extent_encryption(leaf, item, 0);
2628         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2629
2630         btrfs_mark_buffer_dirty(leaf);
2631         inode_add_bytes(inode, len);
2632         btrfs_release_path(path);
2633
2634         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2635                         new->disk_len, 0,
2636                         backref->root_id, backref->inum,
2637                         new->file_pos); /* start - extent_offset */
2638         if (ret) {
2639                 btrfs_abort_transaction(trans, ret);
2640                 goto out_free_path;
2641         }
2642
2643         ret = 1;
2644 out_free_path:
2645         btrfs_release_path(path);
2646         path->leave_spinning = 0;
2647         btrfs_end_transaction(trans, root);
2648 out_unlock:
2649         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2650                              &cached, GFP_NOFS);
2651         iput(inode);
2652         return ret;
2653 }
2654
2655 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2656 {
2657         struct old_sa_defrag_extent *old, *tmp;
2658
2659         if (!new)
2660                 return;
2661
2662         list_for_each_entry_safe(old, tmp, &new->head, list) {
2663                 kfree(old);
2664         }
2665         kfree(new);
2666 }
2667
2668 static void relink_file_extents(struct new_sa_defrag_extent *new)
2669 {
2670         struct btrfs_path *path;
2671         struct sa_defrag_extent_backref *backref;
2672         struct sa_defrag_extent_backref *prev = NULL;
2673         struct inode *inode;
2674         struct btrfs_root *root;
2675         struct rb_node *node;
2676         int ret;
2677
2678         inode = new->inode;
2679         root = BTRFS_I(inode)->root;
2680
2681         path = btrfs_alloc_path();
2682         if (!path)
2683                 return;
2684
2685         if (!record_extent_backrefs(path, new)) {
2686                 btrfs_free_path(path);
2687                 goto out;
2688         }
2689         btrfs_release_path(path);
2690
2691         while (1) {
2692                 node = rb_first(&new->root);
2693                 if (!node)
2694                         break;
2695                 rb_erase(node, &new->root);
2696
2697                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2698
2699                 ret = relink_extent_backref(path, prev, backref);
2700                 WARN_ON(ret < 0);
2701
2702                 kfree(prev);
2703
2704                 if (ret == 1)
2705                         prev = backref;
2706                 else
2707                         prev = NULL;
2708                 cond_resched();
2709         }
2710         kfree(prev);
2711
2712         btrfs_free_path(path);
2713 out:
2714         free_sa_defrag_extent(new);
2715
2716         atomic_dec(&root->fs_info->defrag_running);
2717         wake_up(&root->fs_info->transaction_wait);
2718 }
2719
2720 static struct new_sa_defrag_extent *
2721 record_old_file_extents(struct inode *inode,
2722                         struct btrfs_ordered_extent *ordered)
2723 {
2724         struct btrfs_root *root = BTRFS_I(inode)->root;
2725         struct btrfs_path *path;
2726         struct btrfs_key key;
2727         struct old_sa_defrag_extent *old;
2728         struct new_sa_defrag_extent *new;
2729         int ret;
2730
2731         new = kmalloc(sizeof(*new), GFP_NOFS);
2732         if (!new)
2733                 return NULL;
2734
2735         new->inode = inode;
2736         new->file_pos = ordered->file_offset;
2737         new->len = ordered->len;
2738         new->bytenr = ordered->start;
2739         new->disk_len = ordered->disk_len;
2740         new->compress_type = ordered->compress_type;
2741         new->root = RB_ROOT;
2742         INIT_LIST_HEAD(&new->head);
2743
2744         path = btrfs_alloc_path();
2745         if (!path)
2746                 goto out_kfree;
2747
2748         key.objectid = btrfs_ino(inode);
2749         key.type = BTRFS_EXTENT_DATA_KEY;
2750         key.offset = new->file_pos;
2751
2752         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2753         if (ret < 0)
2754                 goto out_free_path;
2755         if (ret > 0 && path->slots[0] > 0)
2756                 path->slots[0]--;
2757
2758         /* find out all the old extents for the file range */
2759         while (1) {
2760                 struct btrfs_file_extent_item *extent;
2761                 struct extent_buffer *l;
2762                 int slot;
2763                 u64 num_bytes;
2764                 u64 offset;
2765                 u64 end;
2766                 u64 disk_bytenr;
2767                 u64 extent_offset;
2768
2769                 l = path->nodes[0];
2770                 slot = path->slots[0];
2771
2772                 if (slot >= btrfs_header_nritems(l)) {
2773                         ret = btrfs_next_leaf(root, path);
2774                         if (ret < 0)
2775                                 goto out_free_path;
2776                         else if (ret > 0)
2777                                 break;
2778                         continue;
2779                 }
2780
2781                 btrfs_item_key_to_cpu(l, &key, slot);
2782
2783                 if (key.objectid != btrfs_ino(inode))
2784                         break;
2785                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2786                         break;
2787                 if (key.offset >= new->file_pos + new->len)
2788                         break;
2789
2790                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2791
2792                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2793                 if (key.offset + num_bytes < new->file_pos)
2794                         goto next;
2795
2796                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2797                 if (!disk_bytenr)
2798                         goto next;
2799
2800                 extent_offset = btrfs_file_extent_offset(l, extent);
2801
2802                 old = kmalloc(sizeof(*old), GFP_NOFS);
2803                 if (!old)
2804                         goto out_free_path;
2805
2806                 offset = max(new->file_pos, key.offset);
2807                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2808
2809                 old->bytenr = disk_bytenr;
2810                 old->extent_offset = extent_offset;
2811                 old->offset = offset - key.offset;
2812                 old->len = end - offset;
2813                 old->new = new;
2814                 old->count = 0;
2815                 list_add_tail(&old->list, &new->head);
2816 next:
2817                 path->slots[0]++;
2818                 cond_resched();
2819         }
2820
2821         btrfs_free_path(path);
2822         atomic_inc(&root->fs_info->defrag_running);
2823
2824         return new;
2825
2826 out_free_path:
2827         btrfs_free_path(path);
2828 out_kfree:
2829         free_sa_defrag_extent(new);
2830         return NULL;
2831 }
2832
2833 static void btrfs_release_delalloc_bytes(struct btrfs_root *root,
2834                                          u64 start, u64 len)
2835 {
2836         struct btrfs_block_group_cache *cache;
2837
2838         cache = btrfs_lookup_block_group(root->fs_info, start);
2839         ASSERT(cache);
2840
2841         spin_lock(&cache->lock);
2842         cache->delalloc_bytes -= len;
2843         spin_unlock(&cache->lock);
2844
2845         btrfs_put_block_group(cache);
2846 }
2847
2848 /* as ordered data IO finishes, this gets called so we can finish
2849  * an ordered extent if the range of bytes in the file it covers are
2850  * fully written.
2851  */
2852 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2853 {
2854         struct inode *inode = ordered_extent->inode;
2855         struct btrfs_root *root = BTRFS_I(inode)->root;
2856         struct btrfs_trans_handle *trans = NULL;
2857         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2858         struct extent_state *cached_state = NULL;
2859         struct new_sa_defrag_extent *new = NULL;
2860         int compress_type = 0;
2861         int ret = 0;
2862         u64 logical_len = ordered_extent->len;
2863         bool nolock;
2864         bool truncated = false;
2865
2866         nolock = btrfs_is_free_space_inode(inode);
2867
2868         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2869                 ret = -EIO;
2870                 goto out;
2871         }
2872
2873         btrfs_free_io_failure_record(inode, ordered_extent->file_offset,
2874                                      ordered_extent->file_offset +
2875                                      ordered_extent->len - 1);
2876
2877         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2878                 truncated = true;
2879                 logical_len = ordered_extent->truncated_len;
2880                 /* Truncated the entire extent, don't bother adding */
2881                 if (!logical_len)
2882                         goto out;
2883         }
2884
2885         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2886                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2887
2888                 /*
2889                  * For mwrite(mmap + memset to write) case, we still reserve
2890                  * space for NOCOW range.
2891                  * As NOCOW won't cause a new delayed ref, just free the space
2892                  */
2893                 btrfs_qgroup_free_data(inode, ordered_extent->file_offset,
2894                                        ordered_extent->len);
2895                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2896                 if (nolock)
2897                         trans = btrfs_join_transaction_nolock(root);
2898                 else
2899                         trans = btrfs_join_transaction(root);
2900                 if (IS_ERR(trans)) {
2901                         ret = PTR_ERR(trans);
2902                         trans = NULL;
2903                         goto out;
2904                 }
2905                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2906                 ret = btrfs_update_inode_fallback(trans, root, inode);
2907                 if (ret) /* -ENOMEM or corruption */
2908                         btrfs_abort_transaction(trans, ret);
2909                 goto out;
2910         }
2911
2912         lock_extent_bits(io_tree, ordered_extent->file_offset,
2913                          ordered_extent->file_offset + ordered_extent->len - 1,
2914                          &cached_state);
2915
2916         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2917                         ordered_extent->file_offset + ordered_extent->len - 1,
2918                         EXTENT_DEFRAG, 1, cached_state);
2919         if (ret) {
2920                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2921                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
2922                         /* the inode is shared */
2923                         new = record_old_file_extents(inode, ordered_extent);
2924
2925                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2926                         ordered_extent->file_offset + ordered_extent->len - 1,
2927                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2928         }
2929
2930         if (nolock)
2931                 trans = btrfs_join_transaction_nolock(root);
2932         else
2933                 trans = btrfs_join_transaction(root);
2934         if (IS_ERR(trans)) {
2935                 ret = PTR_ERR(trans);
2936                 trans = NULL;
2937                 goto out_unlock;
2938         }
2939
2940         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2941
2942         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2943                 compress_type = ordered_extent->compress_type;
2944         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2945                 BUG_ON(compress_type);
2946                 ret = btrfs_mark_extent_written(trans, inode,
2947                                                 ordered_extent->file_offset,
2948                                                 ordered_extent->file_offset +
2949                                                 logical_len);
2950         } else {
2951                 BUG_ON(root == root->fs_info->tree_root);
2952                 ret = insert_reserved_file_extent(trans, inode,
2953                                                 ordered_extent->file_offset,
2954                                                 ordered_extent->start,
2955                                                 ordered_extent->disk_len,
2956                                                 logical_len, logical_len,
2957                                                 compress_type, 0, 0,
2958                                                 BTRFS_FILE_EXTENT_REG);
2959                 if (!ret)
2960                         btrfs_release_delalloc_bytes(root,
2961                                                      ordered_extent->start,
2962                                                      ordered_extent->disk_len);
2963         }
2964         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
2965                            ordered_extent->file_offset, ordered_extent->len,
2966                            trans->transid);
2967         if (ret < 0) {
2968                 btrfs_abort_transaction(trans, ret);
2969                 goto out_unlock;
2970         }
2971
2972         add_pending_csums(trans, inode, ordered_extent->file_offset,
2973                           &ordered_extent->list);
2974
2975         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2976         ret = btrfs_update_inode_fallback(trans, root, inode);
2977         if (ret) { /* -ENOMEM or corruption */
2978                 btrfs_abort_transaction(trans, ret);
2979                 goto out_unlock;
2980         }
2981         ret = 0;
2982 out_unlock:
2983         unlock_extent_cached(io_tree, ordered_extent->file_offset,
2984                              ordered_extent->file_offset +
2985                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
2986 out:
2987         if (root != root->fs_info->tree_root)
2988                 btrfs_delalloc_release_metadata(inode, ordered_extent->len);
2989         if (trans)
2990                 btrfs_end_transaction(trans, root);
2991
2992         if (ret || truncated) {
2993                 u64 start, end;
2994
2995                 if (truncated)
2996                         start = ordered_extent->file_offset + logical_len;
2997                 else
2998                         start = ordered_extent->file_offset;
2999                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3000                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
3001
3002                 /* Drop the cache for the part of the extent we didn't write. */
3003                 btrfs_drop_extent_cache(inode, start, end, 0);
3004
3005                 /*
3006                  * If the ordered extent had an IOERR or something else went
3007                  * wrong we need to return the space for this ordered extent
3008                  * back to the allocator.  We only free the extent in the
3009                  * truncated case if we didn't write out the extent at all.
3010                  */
3011                 if ((ret || !logical_len) &&
3012                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3013                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3014                         btrfs_free_reserved_extent(root, ordered_extent->start,
3015                                                    ordered_extent->disk_len, 1);
3016         }
3017
3018
3019         /*
3020          * This needs to be done to make sure anybody waiting knows we are done
3021          * updating everything for this ordered extent.
3022          */
3023         btrfs_remove_ordered_extent(inode, ordered_extent);
3024
3025         /* for snapshot-aware defrag */
3026         if (new) {
3027                 if (ret) {
3028                         free_sa_defrag_extent(new);
3029                         atomic_dec(&root->fs_info->defrag_running);
3030                 } else {
3031                         relink_file_extents(new);
3032                 }
3033         }
3034
3035         /* once for us */
3036         btrfs_put_ordered_extent(ordered_extent);
3037         /* once for the tree */
3038         btrfs_put_ordered_extent(ordered_extent);
3039
3040         return ret;
3041 }
3042
3043 static void finish_ordered_fn(struct btrfs_work *work)
3044 {
3045         struct btrfs_ordered_extent *ordered_extent;
3046         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3047         btrfs_finish_ordered_io(ordered_extent);
3048 }
3049
3050 static int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3051                                 struct extent_state *state, int uptodate)
3052 {
3053         struct inode *inode = page->mapping->host;
3054         struct btrfs_root *root = BTRFS_I(inode)->root;
3055         struct btrfs_ordered_extent *ordered_extent = NULL;
3056         struct btrfs_workqueue *wq;
3057         btrfs_work_func_t func;
3058
3059         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3060
3061         ClearPagePrivate2(page);
3062         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3063                                             end - start + 1, uptodate))
3064                 return 0;
3065
3066         if (btrfs_is_free_space_inode(inode)) {
3067                 wq = root->fs_info->endio_freespace_worker;
3068                 func = btrfs_freespace_write_helper;
3069         } else {
3070                 wq = root->fs_info->endio_write_workers;
3071                 func = btrfs_endio_write_helper;
3072         }
3073
3074         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3075                         NULL);
3076         btrfs_queue_work(wq, &ordered_extent->work);
3077
3078         return 0;
3079 }
3080
3081 static int __readpage_endio_check(struct inode *inode,
3082                                   struct btrfs_io_bio *io_bio,
3083                                   int icsum, struct page *page,
3084                                   int pgoff, u64 start, size_t len)
3085 {
3086         char *kaddr;
3087         u32 csum_expected;
3088         u32 csum = ~(u32)0;
3089
3090         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3091
3092         kaddr = kmap_atomic(page);
3093         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3094         btrfs_csum_final(csum, (char *)&csum);
3095         if (csum != csum_expected)
3096                 goto zeroit;
3097
3098         kunmap_atomic(kaddr);
3099         return 0;
3100 zeroit:
3101         btrfs_warn_rl(BTRFS_I(inode)->root->fs_info,
3102                 "csum failed ino %llu off %llu csum %u expected csum %u",
3103                            btrfs_ino(inode), start, csum, csum_expected);
3104         memset(kaddr + pgoff, 1, len);
3105         flush_dcache_page(page);
3106         kunmap_atomic(kaddr);
3107         if (csum_expected == 0)
3108                 return 0;
3109         return -EIO;
3110 }
3111
3112 /*
3113  * when reads are done, we need to check csums to verify the data is correct
3114  * if there's a match, we allow the bio to finish.  If not, the code in
3115  * extent_io.c will try to find good copies for us.
3116  */
3117 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3118                                       u64 phy_offset, struct page *page,
3119                                       u64 start, u64 end, int mirror)
3120 {
3121         size_t offset = start - page_offset(page);
3122         struct inode *inode = page->mapping->host;
3123         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3124         struct btrfs_root *root = BTRFS_I(inode)->root;
3125
3126         if (PageChecked(page)) {
3127                 ClearPageChecked(page);
3128                 return 0;
3129         }
3130
3131         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3132                 return 0;
3133
3134         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3135             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3136                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3137                 return 0;
3138         }
3139
3140         phy_offset >>= inode->i_sb->s_blocksize_bits;
3141         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3142                                       start, (size_t)(end - start + 1));
3143 }
3144
3145 void btrfs_add_delayed_iput(struct inode *inode)
3146 {
3147         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3148         struct btrfs_inode *binode = BTRFS_I(inode);
3149
3150         if (atomic_add_unless(&inode->i_count, -1, 1))
3151                 return;
3152
3153         spin_lock(&fs_info->delayed_iput_lock);
3154         if (binode->delayed_iput_count == 0) {
3155                 ASSERT(list_empty(&binode->delayed_iput));
3156                 list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3157         } else {
3158                 binode->delayed_iput_count++;
3159         }
3160         spin_unlock(&fs_info->delayed_iput_lock);
3161 }
3162
3163 void btrfs_run_delayed_iputs(struct btrfs_root *root)
3164 {
3165         struct btrfs_fs_info *fs_info = root->fs_info;
3166
3167         spin_lock(&fs_info->delayed_iput_lock);
3168         while (!list_empty(&fs_info->delayed_iputs)) {
3169                 struct btrfs_inode *inode;
3170
3171                 inode = list_first_entry(&fs_info->delayed_iputs,
3172                                 struct btrfs_inode, delayed_iput);
3173                 if (inode->delayed_iput_count) {
3174                         inode->delayed_iput_count--;
3175                         list_move_tail(&inode->delayed_iput,
3176                                         &fs_info->delayed_iputs);
3177                 } else {
3178                         list_del_init(&inode->delayed_iput);
3179                 }
3180                 spin_unlock(&fs_info->delayed_iput_lock);
3181                 iput(&inode->vfs_inode);
3182                 spin_lock(&fs_info->delayed_iput_lock);
3183         }
3184         spin_unlock(&fs_info->delayed_iput_lock);
3185 }
3186
3187 /*
3188  * This is called in transaction commit time. If there are no orphan
3189  * files in the subvolume, it removes orphan item and frees block_rsv
3190  * structure.
3191  */
3192 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3193                               struct btrfs_root *root)
3194 {
3195         struct btrfs_block_rsv *block_rsv;
3196         int ret;
3197
3198         if (atomic_read(&root->orphan_inodes) ||
3199             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3200                 return;
3201
3202         spin_lock(&root->orphan_lock);
3203         if (atomic_read(&root->orphan_inodes)) {
3204                 spin_unlock(&root->orphan_lock);
3205                 return;
3206         }
3207
3208         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3209                 spin_unlock(&root->orphan_lock);
3210                 return;
3211         }
3212
3213         block_rsv = root->orphan_block_rsv;
3214         root->orphan_block_rsv = NULL;
3215         spin_unlock(&root->orphan_lock);
3216
3217         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3218             btrfs_root_refs(&root->root_item) > 0) {
3219                 ret = btrfs_del_orphan_item(trans, root->fs_info->tree_root,
3220                                             root->root_key.objectid);
3221                 if (ret)
3222                         btrfs_abort_transaction(trans, ret);
3223                 else
3224                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3225                                   &root->state);
3226         }
3227
3228         if (block_rsv) {
3229                 WARN_ON(block_rsv->size > 0);
3230                 btrfs_free_block_rsv(root, block_rsv);
3231         }
3232 }
3233
3234 /*
3235  * This creates an orphan entry for the given inode in case something goes
3236  * wrong in the middle of an unlink/truncate.
3237  *
3238  * NOTE: caller of this function should reserve 5 units of metadata for
3239  *       this function.
3240  */
3241 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
3242 {
3243         struct btrfs_root *root = BTRFS_I(inode)->root;
3244         struct btrfs_block_rsv *block_rsv = NULL;
3245         int reserve = 0;
3246         int insert = 0;
3247         int ret;
3248
3249         if (!root->orphan_block_rsv) {
3250                 block_rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
3251                 if (!block_rsv)
3252                         return -ENOMEM;
3253         }
3254
3255         spin_lock(&root->orphan_lock);
3256         if (!root->orphan_block_rsv) {
3257                 root->orphan_block_rsv = block_rsv;
3258         } else if (block_rsv) {
3259                 btrfs_free_block_rsv(root, block_rsv);
3260                 block_rsv = NULL;
3261         }
3262
3263         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3264                               &BTRFS_I(inode)->runtime_flags)) {
3265 #if 0
3266                 /*
3267                  * For proper ENOSPC handling, we should do orphan
3268                  * cleanup when mounting. But this introduces backward
3269                  * compatibility issue.
3270                  */
3271                 if (!xchg(&root->orphan_item_inserted, 1))
3272                         insert = 2;
3273                 else
3274                         insert = 1;
3275 #endif
3276                 insert = 1;
3277                 atomic_inc(&root->orphan_inodes);
3278         }
3279
3280         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3281                               &BTRFS_I(inode)->runtime_flags))
3282                 reserve = 1;
3283         spin_unlock(&root->orphan_lock);
3284
3285         /* grab metadata reservation from transaction handle */
3286         if (reserve) {
3287                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3288                 ASSERT(!ret);
3289                 if (ret) {
3290                         atomic_dec(&root->orphan_inodes);
3291                         clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3292                                   &BTRFS_I(inode)->runtime_flags);
3293                         if (insert)
3294                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3295                                           &BTRFS_I(inode)->runtime_flags);
3296                         return ret;
3297                 }
3298         }
3299
3300         /* insert an orphan item to track this unlinked/truncated file */
3301         if (insert >= 1) {
3302                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3303                 if (ret) {
3304                         atomic_dec(&root->orphan_inodes);
3305                         if (reserve) {
3306                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3307                                           &BTRFS_I(inode)->runtime_flags);
3308                                 btrfs_orphan_release_metadata(inode);
3309                         }
3310                         if (ret != -EEXIST) {
3311                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3312                                           &BTRFS_I(inode)->runtime_flags);
3313                                 btrfs_abort_transaction(trans, ret);
3314                                 return ret;
3315                         }
3316                 }
3317                 ret = 0;
3318         }
3319
3320         /* insert an orphan item to track subvolume contains orphan files */
3321         if (insert >= 2) {
3322                 ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
3323                                                root->root_key.objectid);
3324                 if (ret && ret != -EEXIST) {
3325                         btrfs_abort_transaction(trans, ret);
3326                         return ret;
3327                 }
3328         }
3329         return 0;
3330 }
3331
3332 /*
3333  * We have done the truncate/delete so we can go ahead and remove the orphan
3334  * item for this particular inode.
3335  */
3336 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3337                             struct inode *inode)
3338 {
3339         struct btrfs_root *root = BTRFS_I(inode)->root;
3340         int delete_item = 0;
3341         int release_rsv = 0;
3342         int ret = 0;
3343
3344         spin_lock(&root->orphan_lock);
3345         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3346                                &BTRFS_I(inode)->runtime_flags))
3347                 delete_item = 1;
3348
3349         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3350                                &BTRFS_I(inode)->runtime_flags))
3351                 release_rsv = 1;
3352         spin_unlock(&root->orphan_lock);
3353
3354         if (delete_item) {
3355                 atomic_dec(&root->orphan_inodes);
3356                 if (trans)
3357                         ret = btrfs_del_orphan_item(trans, root,
3358                                                     btrfs_ino(inode));
3359         }
3360
3361         if (release_rsv)
3362                 btrfs_orphan_release_metadata(inode);
3363
3364         return ret;
3365 }
3366
3367 /*
3368  * this cleans up any orphans that may be left on the list from the last use
3369  * of this root.
3370  */
3371 int btrfs_orphan_cleanup(struct btrfs_root *root)
3372 {
3373         struct btrfs_path *path;
3374         struct extent_buffer *leaf;
3375         struct btrfs_key key, found_key;
3376         struct btrfs_trans_handle *trans;
3377         struct inode *inode;
3378         u64 last_objectid = 0;
3379         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3380
3381         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3382                 return 0;
3383
3384         path = btrfs_alloc_path();
3385         if (!path) {
3386                 ret = -ENOMEM;
3387                 goto out;
3388         }
3389         path->reada = READA_BACK;
3390
3391         key.objectid = BTRFS_ORPHAN_OBJECTID;
3392         key.type = BTRFS_ORPHAN_ITEM_KEY;
3393         key.offset = (u64)-1;
3394
3395         while (1) {
3396                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3397                 if (ret < 0)
3398                         goto out;
3399
3400                 /*
3401                  * if ret == 0 means we found what we were searching for, which
3402                  * is weird, but possible, so only screw with path if we didn't
3403                  * find the key and see if we have stuff that matches
3404                  */
3405                 if (ret > 0) {
3406                         ret = 0;
3407                         if (path->slots[0] == 0)
3408                                 break;
3409                         path->slots[0]--;
3410                 }
3411
3412                 /* pull out the item */
3413                 leaf = path->nodes[0];
3414                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3415
3416                 /* make sure the item matches what we want */
3417                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3418                         break;
3419                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3420                         break;
3421
3422                 /* release the path since we're done with it */
3423                 btrfs_release_path(path);
3424
3425                 /*
3426                  * this is where we are basically btrfs_lookup, without the
3427                  * crossing root thing.  we store the inode number in the
3428                  * offset of the orphan item.
3429                  */
3430
3431                 if (found_key.offset == last_objectid) {
3432                         btrfs_err(root->fs_info,
3433                                 "Error removing orphan entry, stopping orphan cleanup");
3434                         ret = -EINVAL;
3435                         goto out;
3436                 }
3437
3438                 last_objectid = found_key.offset;
3439
3440                 found_key.objectid = found_key.offset;
3441                 found_key.type = BTRFS_INODE_ITEM_KEY;
3442                 found_key.offset = 0;
3443                 inode = btrfs_iget(root->fs_info->sb, &found_key, root, NULL);
3444                 ret = PTR_ERR_OR_ZERO(inode);
3445                 if (ret && ret != -ENOENT)
3446                         goto out;
3447
3448                 if (ret == -ENOENT && root == root->fs_info->tree_root) {
3449                         struct btrfs_root *dead_root;
3450                         struct btrfs_fs_info *fs_info = root->fs_info;
3451                         int is_dead_root = 0;
3452
3453                         /*
3454                          * this is an orphan in the tree root. Currently these
3455                          * could come from 2 sources:
3456                          *  a) a snapshot deletion in progress
3457                          *  b) a free space cache inode
3458                          * We need to distinguish those two, as the snapshot
3459                          * orphan must not get deleted.
3460                          * find_dead_roots already ran before us, so if this
3461                          * is a snapshot deletion, we should find the root
3462                          * in the dead_roots list
3463                          */
3464                         spin_lock(&fs_info->trans_lock);
3465                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3466                                             root_list) {
3467                                 if (dead_root->root_key.objectid ==
3468                                     found_key.objectid) {
3469                                         is_dead_root = 1;
3470                                         break;
3471                                 }
3472                         }
3473                         spin_unlock(&fs_info->trans_lock);
3474                         if (is_dead_root) {
3475                                 /* prevent this orphan from being found again */
3476                                 key.offset = found_key.objectid - 1;
3477                                 continue;
3478                         }
3479                 }
3480                 /*
3481                  * Inode is already gone but the orphan item is still there,
3482                  * kill the orphan item.
3483                  */
3484                 if (ret == -ENOENT) {
3485                         trans = btrfs_start_transaction(root, 1);
3486                         if (IS_ERR(trans)) {
3487                                 ret = PTR_ERR(trans);
3488                                 goto out;
3489                         }
3490                         btrfs_debug(root->fs_info, "auto deleting %Lu",
3491                                 found_key.objectid);
3492                         ret = btrfs_del_orphan_item(trans, root,
3493                                                     found_key.objectid);
3494                         btrfs_end_transaction(trans, root);
3495                         if (ret)
3496                                 goto out;
3497                         continue;
3498                 }
3499
3500                 /*
3501                  * add this inode to the orphan list so btrfs_orphan_del does
3502                  * the proper thing when we hit it
3503                  */
3504                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3505                         &BTRFS_I(inode)->runtime_flags);
3506                 atomic_inc(&root->orphan_inodes);
3507
3508                 /* if we have links, this was a truncate, lets do that */
3509                 if (inode->i_nlink) {
3510                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3511                                 iput(inode);
3512                                 continue;
3513                         }
3514                         nr_truncate++;
3515
3516                         /* 1 for the orphan item deletion. */
3517                         trans = btrfs_start_transaction(root, 1);
3518                         if (IS_ERR(trans)) {
3519                                 iput(inode);
3520                                 ret = PTR_ERR(trans);
3521                                 goto out;
3522                         }
3523                         ret = btrfs_orphan_add(trans, inode);
3524                         btrfs_end_transaction(trans, root);
3525                         if (ret) {
3526                                 iput(inode);
3527                                 goto out;
3528                         }
3529
3530                         ret = btrfs_truncate(inode);
3531                         if (ret)
3532                                 btrfs_orphan_del(NULL, inode);
3533                 } else {
3534                         nr_unlink++;
3535                 }
3536
3537                 /* this will do delete_inode and everything for us */
3538                 iput(inode);
3539                 if (ret)
3540                         goto out;
3541         }
3542         /* release the path since we're done with it */
3543         btrfs_release_path(path);
3544
3545         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3546
3547         if (root->orphan_block_rsv)
3548                 btrfs_block_rsv_release(root, root->orphan_block_rsv,
3549                                         (u64)-1);
3550
3551         if (root->orphan_block_rsv ||
3552             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3553                 trans = btrfs_join_transaction(root);
3554                 if (!IS_ERR(trans))
3555                         btrfs_end_transaction(trans, root);
3556         }
3557
3558         if (nr_unlink)
3559                 btrfs_debug(root->fs_info, "unlinked %d orphans", nr_unlink);
3560         if (nr_truncate)
3561                 btrfs_debug(root->fs_info, "truncated %d orphans", nr_truncate);
3562
3563 out:
3564         if (ret)
3565                 btrfs_err(root->fs_info,
3566                         "could not do orphan cleanup %d", ret);
3567         btrfs_free_path(path);
3568         return ret;
3569 }
3570
3571 /*
3572  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3573  * don't find any xattrs, we know there can't be any acls.
3574  *
3575  * slot is the slot the inode is in, objectid is the objectid of the inode
3576  */
3577 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3578                                           int slot, u64 objectid,
3579                                           int *first_xattr_slot)
3580 {
3581         u32 nritems = btrfs_header_nritems(leaf);
3582         struct btrfs_key found_key;
3583         static u64 xattr_access = 0;
3584         static u64 xattr_default = 0;
3585         int scanned = 0;
3586
3587         if (!xattr_access) {
3588                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3589                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3590                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3591                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3592         }
3593
3594         slot++;
3595         *first_xattr_slot = -1;
3596         while (slot < nritems) {
3597                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3598
3599                 /* we found a different objectid, there must not be acls */
3600                 if (found_key.objectid != objectid)
3601                         return 0;
3602
3603                 /* we found an xattr, assume we've got an acl */
3604                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3605                         if (*first_xattr_slot == -1)
3606                                 *first_xattr_slot = slot;
3607                         if (found_key.offset == xattr_access ||
3608                             found_key.offset == xattr_default)
3609                                 return 1;
3610                 }
3611
3612                 /*
3613                  * we found a key greater than an xattr key, there can't
3614                  * be any acls later on
3615                  */
3616                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3617                         return 0;
3618
3619                 slot++;
3620                 scanned++;
3621
3622                 /*
3623                  * it goes inode, inode backrefs, xattrs, extents,
3624                  * so if there are a ton of hard links to an inode there can
3625                  * be a lot of backrefs.  Don't waste time searching too hard,
3626                  * this is just an optimization
3627                  */
3628                 if (scanned >= 8)
3629                         break;
3630         }
3631         /* we hit the end of the leaf before we found an xattr or
3632          * something larger than an xattr.  We have to assume the inode
3633          * has acls
3634          */
3635         if (*first_xattr_slot == -1)
3636                 *first_xattr_slot = slot;
3637         return 1;
3638 }
3639
3640 /*
3641  * read an inode from the btree into the in-memory inode
3642  */
3643 static int btrfs_read_locked_inode(struct inode *inode)
3644 {
3645         struct btrfs_path *path;
3646         struct extent_buffer *leaf;
3647         struct btrfs_inode_item *inode_item;
3648         struct btrfs_root *root = BTRFS_I(inode)->root;
3649         struct btrfs_key location;
3650         unsigned long ptr;
3651         int maybe_acls;
3652         u32 rdev;
3653         int ret;
3654         bool filled = false;
3655         int first_xattr_slot;
3656
3657         ret = btrfs_fill_inode(inode, &rdev);
3658         if (!ret)
3659                 filled = true;
3660
3661         path = btrfs_alloc_path();
3662         if (!path) {
3663                 ret = -ENOMEM;
3664                 goto make_bad;
3665         }
3666
3667         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3668
3669         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3670         if (ret) {
3671                 if (ret > 0)
3672                         ret = -ENOENT;
3673                 goto make_bad;
3674         }
3675
3676         leaf = path->nodes[0];
3677
3678         if (filled)
3679                 goto cache_index;
3680
3681         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3682                                     struct btrfs_inode_item);
3683         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3684         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3685         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3686         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3687         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
3688
3689         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3690         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3691
3692         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3693         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3694
3695         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3696         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3697
3698         BTRFS_I(inode)->i_otime.tv_sec =
3699                 btrfs_timespec_sec(leaf, &inode_item->otime);
3700         BTRFS_I(inode)->i_otime.tv_nsec =
3701                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3702
3703         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3704         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3705         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3706
3707         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3708         inode->i_generation = BTRFS_I(inode)->generation;
3709         inode->i_rdev = 0;
3710         rdev = btrfs_inode_rdev(leaf, inode_item);
3711
3712         BTRFS_I(inode)->index_cnt = (u64)-1;
3713         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3714
3715 cache_index:
3716         /*
3717          * If we were modified in the current generation and evicted from memory
3718          * and then re-read we need to do a full sync since we don't have any
3719          * idea about which extents were modified before we were evicted from
3720          * cache.
3721          *
3722          * This is required for both inode re-read from disk and delayed inode
3723          * in delayed_nodes_tree.
3724          */
3725         if (BTRFS_I(inode)->last_trans == root->fs_info->generation)
3726                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3727                         &BTRFS_I(inode)->runtime_flags);
3728
3729         /*
3730          * We don't persist the id of the transaction where an unlink operation
3731          * against the inode was last made. So here we assume the inode might
3732          * have been evicted, and therefore the exact value of last_unlink_trans
3733          * lost, and set it to last_trans to avoid metadata inconsistencies
3734          * between the inode and its parent if the inode is fsync'ed and the log
3735          * replayed. For example, in the scenario:
3736          *
3737          * touch mydir/foo
3738          * ln mydir/foo mydir/bar
3739          * sync
3740          * unlink mydir/bar
3741          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3742          * xfs_io -c fsync mydir/foo
3743          * <power failure>
3744          * mount fs, triggers fsync log replay
3745          *
3746          * We must make sure that when we fsync our inode foo we also log its
3747          * parent inode, otherwise after log replay the parent still has the
3748          * dentry with the "bar" name but our inode foo has a link count of 1
3749          * and doesn't have an inode ref with the name "bar" anymore.
3750          *
3751          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3752          * but it guarantees correctness at the expense of occasional full
3753          * transaction commits on fsync if our inode is a directory, or if our
3754          * inode is not a directory, logging its parent unnecessarily.
3755          */
3756         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3757
3758         path->slots[0]++;
3759         if (inode->i_nlink != 1 ||
3760             path->slots[0] >= btrfs_header_nritems(leaf))
3761                 goto cache_acl;
3762
3763         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3764         if (location.objectid != btrfs_ino(inode))
3765                 goto cache_acl;
3766
3767         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3768         if (location.type == BTRFS_INODE_REF_KEY) {
3769                 struct btrfs_inode_ref *ref;
3770
3771                 ref = (struct btrfs_inode_ref *)ptr;
3772                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3773         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3774                 struct btrfs_inode_extref *extref;
3775
3776                 extref = (struct btrfs_inode_extref *)ptr;
3777                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3778                                                                      extref);
3779         }
3780 cache_acl:
3781         /*
3782          * try to precache a NULL acl entry for files that don't have
3783          * any xattrs or acls
3784          */
3785         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3786                                            btrfs_ino(inode), &first_xattr_slot);
3787         if (first_xattr_slot != -1) {
3788                 path->slots[0] = first_xattr_slot;
3789                 ret = btrfs_load_inode_props(inode, path);
3790                 if (ret)
3791                         btrfs_err(root->fs_info,
3792                                   "error loading props for ino %llu (root %llu): %d",
3793                                   btrfs_ino(inode),
3794                                   root->root_key.objectid, ret);
3795         }
3796         btrfs_free_path(path);
3797
3798         if (!maybe_acls)
3799                 cache_no_acl(inode);
3800
3801         switch (inode->i_mode & S_IFMT) {
3802         case S_IFREG:
3803                 inode->i_mapping->a_ops = &btrfs_aops;
3804                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3805                 inode->i_fop = &btrfs_file_operations;
3806                 inode->i_op = &btrfs_file_inode_operations;
3807                 break;
3808         case S_IFDIR:
3809                 inode->i_fop = &btrfs_dir_file_operations;
3810                 if (root == root->fs_info->tree_root)
3811                         inode->i_op = &btrfs_dir_ro_inode_operations;
3812                 else
3813                         inode->i_op = &btrfs_dir_inode_operations;
3814                 break;
3815         case S_IFLNK:
3816                 inode->i_op = &btrfs_symlink_inode_operations;
3817                 inode_nohighmem(inode);
3818                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3819                 break;
3820         default:
3821                 inode->i_op = &btrfs_special_inode_operations;
3822                 init_special_inode(inode, inode->i_mode, rdev);
3823                 break;
3824         }
3825
3826         btrfs_update_iflags(inode);
3827         return 0;
3828
3829 make_bad:
3830         btrfs_free_path(path);
3831         make_bad_inode(inode);
3832         return ret;
3833 }
3834
3835 /*
3836  * given a leaf and an inode, copy the inode fields into the leaf
3837  */
3838 static void fill_inode_item(struct btrfs_trans_handle *trans,
3839                             struct extent_buffer *leaf,
3840                             struct btrfs_inode_item *item,
3841                             struct inode *inode)
3842 {
3843         struct btrfs_map_token token;
3844
3845         btrfs_init_map_token(&token);
3846
3847         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3848         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3849         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3850                                    &token);
3851         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3852         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3853
3854         btrfs_set_token_timespec_sec(leaf, &item->atime,
3855                                      inode->i_atime.tv_sec, &token);
3856         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3857                                       inode->i_atime.tv_nsec, &token);
3858
3859         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3860                                      inode->i_mtime.tv_sec, &token);
3861         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3862                                       inode->i_mtime.tv_nsec, &token);
3863
3864         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3865                                      inode->i_ctime.tv_sec, &token);
3866         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3867                                       inode->i_ctime.tv_nsec, &token);
3868
3869         btrfs_set_token_timespec_sec(leaf, &item->otime,
3870                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3871         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3872                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3873
3874         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3875                                      &token);
3876         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3877                                          &token);
3878         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3879         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3880         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3881         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3882         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3883 }
3884
3885 /*
3886  * copy everything in the in-memory inode into the btree.
3887  */
3888 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3889                                 struct btrfs_root *root, struct inode *inode)
3890 {
3891         struct btrfs_inode_item *inode_item;
3892         struct btrfs_path *path;
3893         struct extent_buffer *leaf;
3894         int ret;
3895
3896         path = btrfs_alloc_path();
3897         if (!path)
3898                 return -ENOMEM;
3899
3900         path->leave_spinning = 1;
3901         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3902                                  1);
3903         if (ret) {
3904                 if (ret > 0)
3905                         ret = -ENOENT;
3906                 goto failed;
3907         }
3908
3909         leaf = path->nodes[0];
3910         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3911                                     struct btrfs_inode_item);
3912
3913         fill_inode_item(trans, leaf, inode_item, inode);
3914         btrfs_mark_buffer_dirty(leaf);
3915         btrfs_set_inode_last_trans(trans, inode);
3916         ret = 0;
3917 failed:
3918         btrfs_free_path(path);
3919         return ret;
3920 }
3921
3922 /*
3923  * copy everything in the in-memory inode into the btree.
3924  */
3925 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3926                                 struct btrfs_root *root, struct inode *inode)
3927 {
3928         int ret;
3929
3930         /*
3931          * If the inode is a free space inode, we can deadlock during commit
3932          * if we put it into the delayed code.
3933          *
3934          * The data relocation inode should also be directly updated
3935          * without delay
3936          */
3937         if (!btrfs_is_free_space_inode(inode)
3938             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3939             && !root->fs_info->log_root_recovering) {
3940                 btrfs_update_root_times(trans, root);
3941
3942                 ret = btrfs_delayed_update_inode(trans, root, inode);
3943                 if (!ret)
3944                         btrfs_set_inode_last_trans(trans, inode);
3945                 return ret;
3946         }
3947
3948         return btrfs_update_inode_item(trans, root, inode);
3949 }
3950
3951 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3952                                          struct btrfs_root *root,
3953                                          struct inode *inode)
3954 {
3955         int ret;
3956
3957         ret = btrfs_update_inode(trans, root, inode);
3958         if (ret == -ENOSPC)
3959                 return btrfs_update_inode_item(trans, root, inode);
3960         return ret;
3961 }
3962
3963 /*
3964  * unlink helper that gets used here in inode.c and in the tree logging
3965  * recovery code.  It remove a link in a directory with a given name, and
3966  * also drops the back refs in the inode to the directory
3967  */
3968 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3969                                 struct btrfs_root *root,
3970                                 struct inode *dir, struct inode *inode,
3971                                 const char *name, int name_len)
3972 {
3973         struct btrfs_path *path;
3974         int ret = 0;
3975         struct extent_buffer *leaf;
3976         struct btrfs_dir_item *di;
3977         struct btrfs_key key;
3978         u64 index;
3979         u64 ino = btrfs_ino(inode);
3980         u64 dir_ino = btrfs_ino(dir);
3981
3982         path = btrfs_alloc_path();
3983         if (!path) {
3984                 ret = -ENOMEM;
3985                 goto out;
3986         }
3987
3988         path->leave_spinning = 1;
3989         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3990                                     name, name_len, -1);
3991         if (IS_ERR(di)) {
3992                 ret = PTR_ERR(di);
3993                 goto err;
3994         }
3995         if (!di) {
3996                 ret = -ENOENT;
3997                 goto err;
3998         }
3999         leaf = path->nodes[0];
4000         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4001         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4002         if (ret)
4003                 goto err;
4004         btrfs_release_path(path);
4005
4006         /*
4007          * If we don't have dir index, we have to get it by looking up
4008          * the inode ref, since we get the inode ref, remove it directly,
4009          * it is unnecessary to do delayed deletion.
4010          *
4011          * But if we have dir index, needn't search inode ref to get it.
4012          * Since the inode ref is close to the inode item, it is better
4013          * that we delay to delete it, and just do this deletion when
4014          * we update the inode item.
4015          */
4016         if (BTRFS_I(inode)->dir_index) {
4017                 ret = btrfs_delayed_delete_inode_ref(inode);
4018                 if (!ret) {
4019                         index = BTRFS_I(inode)->dir_index;
4020                         goto skip_backref;
4021                 }
4022         }
4023
4024         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4025                                   dir_ino, &index);
4026         if (ret) {
4027                 btrfs_info(root->fs_info,
4028                         "failed to delete reference to %.*s, inode %llu parent %llu",
4029                         name_len, name, ino, dir_ino);
4030                 btrfs_abort_transaction(trans, ret);
4031                 goto err;
4032         }
4033 skip_backref:
4034         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4035         if (ret) {
4036                 btrfs_abort_transaction(trans, ret);
4037                 goto err;
4038         }
4039
4040         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
4041                                          inode, dir_ino);
4042         if (ret != 0 && ret != -ENOENT) {
4043                 btrfs_abort_transaction(trans, ret);
4044                 goto err;
4045         }
4046
4047         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
4048                                            dir, index);
4049         if (ret == -ENOENT)
4050                 ret = 0;
4051         else if (ret)
4052                 btrfs_abort_transaction(trans, ret);
4053 err:
4054         btrfs_free_path(path);
4055         if (ret)
4056                 goto out;
4057
4058         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4059         inode_inc_iversion(inode);
4060         inode_inc_iversion(dir);
4061         inode->i_ctime = dir->i_mtime =
4062                 dir->i_ctime = current_fs_time(inode->i_sb);
4063         ret = btrfs_update_inode(trans, root, dir);
4064 out:
4065         return ret;
4066 }
4067
4068 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4069                        struct btrfs_root *root,
4070                        struct inode *dir, struct inode *inode,
4071                        const char *name, int name_len)
4072 {
4073         int ret;
4074         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4075         if (!ret) {
4076                 drop_nlink(inode);
4077                 ret = btrfs_update_inode(trans, root, inode);
4078         }
4079         return ret;
4080 }
4081
4082 /*
4083  * helper to start transaction for unlink and rmdir.
4084  *
4085  * unlink and rmdir are special in btrfs, they do not always free space, so
4086  * if we cannot make our reservations the normal way try and see if there is
4087  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4088  * allow the unlink to occur.
4089  */
4090 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4091 {
4092         struct btrfs_root *root = BTRFS_I(dir)->root;
4093
4094         /*
4095          * 1 for the possible orphan item
4096          * 1 for the dir item
4097          * 1 for the dir index
4098          * 1 for the inode ref
4099          * 1 for the inode
4100          */
4101         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4102 }
4103
4104 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4105 {
4106         struct btrfs_root *root = BTRFS_I(dir)->root;
4107         struct btrfs_trans_handle *trans;
4108         struct inode *inode = d_inode(dentry);
4109         int ret;
4110
4111         trans = __unlink_start_trans(dir);
4112         if (IS_ERR(trans))
4113                 return PTR_ERR(trans);
4114
4115         btrfs_record_unlink_dir(trans, dir, d_inode(dentry), 0);
4116
4117         ret = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4118                                  dentry->d_name.name, dentry->d_name.len);
4119         if (ret)
4120                 goto out;
4121
4122         if (inode->i_nlink == 0) {
4123                 ret = btrfs_orphan_add(trans, inode);
4124                 if (ret)
4125                         goto out;
4126         }
4127
4128 out:
4129         btrfs_end_transaction(trans, root);
4130         btrfs_btree_balance_dirty(root);
4131         return ret;
4132 }
4133
4134 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4135                         struct btrfs_root *root,
4136                         struct inode *dir, u64 objectid,
4137                         const char *name, int name_len)
4138 {
4139         struct btrfs_path *path;
4140         struct extent_buffer *leaf;
4141         struct btrfs_dir_item *di;
4142         struct btrfs_key key;
4143         u64 index;
4144         int ret;
4145         u64 dir_ino = btrfs_ino(dir);
4146
4147         path = btrfs_alloc_path();
4148         if (!path)
4149                 return -ENOMEM;
4150
4151         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4152                                    name, name_len, -1);
4153         if (IS_ERR_OR_NULL(di)) {
4154                 if (!di)
4155                         ret = -ENOENT;
4156                 else
4157                         ret = PTR_ERR(di);
4158                 goto out;
4159         }
4160
4161         leaf = path->nodes[0];
4162         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4163         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4164         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4165         if (ret) {
4166                 btrfs_abort_transaction(trans, ret);
4167                 goto out;
4168         }
4169         btrfs_release_path(path);
4170
4171         ret = btrfs_del_root_ref(trans, root->fs_info->tree_root,
4172                                  objectid, root->root_key.objectid,
4173                                  dir_ino, &index, name, name_len);
4174         if (ret < 0) {
4175                 if (ret != -ENOENT) {
4176                         btrfs_abort_transaction(trans, ret);
4177                         goto out;
4178                 }
4179                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4180                                                  name, name_len);
4181                 if (IS_ERR_OR_NULL(di)) {
4182                         if (!di)
4183                                 ret = -ENOENT;
4184                         else
4185                                 ret = PTR_ERR(di);
4186                         btrfs_abort_transaction(trans, ret);
4187                         goto out;
4188                 }
4189
4190                 leaf = path->nodes[0];
4191                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4192                 btrfs_release_path(path);
4193                 index = key.offset;
4194         }
4195         btrfs_release_path(path);
4196
4197         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4198         if (ret) {
4199                 btrfs_abort_transaction(trans, ret);
4200                 goto out;
4201         }
4202
4203         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4204         inode_inc_iversion(dir);
4205         dir->i_mtime = dir->i_ctime = current_fs_time(dir->i_sb);
4206         ret = btrfs_update_inode_fallback(trans, root, dir);
4207         if (ret)
4208                 btrfs_abort_transaction(trans, ret);
4209 out:
4210         btrfs_free_path(path);
4211         return ret;
4212 }
4213
4214 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4215 {
4216         struct inode *inode = d_inode(dentry);
4217         int err = 0;
4218         struct btrfs_root *root = BTRFS_I(dir)->root;
4219         struct btrfs_trans_handle *trans;
4220         u64 last_unlink_trans;
4221
4222         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4223                 return -ENOTEMPTY;
4224         if (btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID)
4225                 return -EPERM;
4226
4227         trans = __unlink_start_trans(dir);
4228         if (IS_ERR(trans))
4229                 return PTR_ERR(trans);
4230
4231         if (unlikely(btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4232                 err = btrfs_unlink_subvol(trans, root, dir,
4233                                           BTRFS_I(inode)->location.objectid,
4234                                           dentry->d_name.name,
4235                                           dentry->d_name.len);
4236                 goto out;
4237         }
4238
4239         err = btrfs_orphan_add(trans, inode);
4240         if (err)
4241                 goto out;
4242
4243         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4244
4245         /* now the directory is empty */
4246         err = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4247                                  dentry->d_name.name, dentry->d_name.len);
4248         if (!err) {
4249                 btrfs_i_size_write(inode, 0);
4250                 /*
4251                  * Propagate the last_unlink_trans value of the deleted dir to
4252                  * its parent directory. This is to prevent an unrecoverable
4253                  * log tree in the case we do something like this:
4254                  * 1) create dir foo
4255                  * 2) create snapshot under dir foo
4256                  * 3) delete the snapshot
4257                  * 4) rmdir foo
4258                  * 5) mkdir foo
4259                  * 6) fsync foo or some file inside foo
4260                  */
4261                 if (last_unlink_trans >= trans->transid)
4262                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4263         }
4264 out:
4265         btrfs_end_transaction(trans, root);
4266         btrfs_btree_balance_dirty(root);
4267
4268         return err;
4269 }
4270
4271 static int truncate_space_check(struct btrfs_trans_handle *trans,
4272                                 struct btrfs_root *root,
4273                                 u64 bytes_deleted)
4274 {
4275         int ret;
4276
4277         /*
4278          * This is only used to apply pressure to the enospc system, we don't
4279          * intend to use this reservation at all.
4280          */
4281         bytes_deleted = btrfs_csum_bytes_to_leaves(root, bytes_deleted);
4282         bytes_deleted *= root->nodesize;
4283         ret = btrfs_block_rsv_add(root, &root->fs_info->trans_block_rsv,
4284                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4285         if (!ret) {
4286                 trace_btrfs_space_reservation(root->fs_info, "transaction",
4287                                               trans->transid,
4288                                               bytes_deleted, 1);
4289                 trans->bytes_reserved += bytes_deleted;
4290         }
4291         return ret;
4292
4293 }
4294
4295 static int truncate_inline_extent(struct inode *inode,
4296                                   struct btrfs_path *path,
4297                                   struct btrfs_key *found_key,
4298                                   const u64 item_end,
4299                                   const u64 new_size)
4300 {
4301         struct extent_buffer *leaf = path->nodes[0];
4302         int slot = path->slots[0];
4303         struct btrfs_file_extent_item *fi;
4304         u32 size = (u32)(new_size - found_key->offset);
4305         struct btrfs_root *root = BTRFS_I(inode)->root;
4306
4307         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4308
4309         if (btrfs_file_extent_compression(leaf, fi) != BTRFS_COMPRESS_NONE) {
4310                 loff_t offset = new_size;
4311                 loff_t page_end = ALIGN(offset, PAGE_SIZE);
4312
4313                 /*
4314                  * Zero out the remaining of the last page of our inline extent,
4315                  * instead of directly truncating our inline extent here - that
4316                  * would be much more complex (decompressing all the data, then
4317                  * compressing the truncated data, which might be bigger than
4318                  * the size of the inline extent, resize the extent, etc).
4319                  * We release the path because to get the page we might need to
4320                  * read the extent item from disk (data not in the page cache).
4321                  */
4322                 btrfs_release_path(path);
4323                 return btrfs_truncate_block(inode, offset, page_end - offset,
4324                                         0);
4325         }
4326
4327         btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4328         size = btrfs_file_extent_calc_inline_size(size);
4329         btrfs_truncate_item(root, path, size, 1);
4330
4331         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4332                 inode_sub_bytes(inode, item_end + 1 - new_size);
4333
4334         return 0;
4335 }
4336
4337 /*
4338  * this can truncate away extent items, csum items and directory items.
4339  * It starts at a high offset and removes keys until it can't find
4340  * any higher than new_size
4341  *
4342  * csum items that cross the new i_size are truncated to the new size
4343  * as well.
4344  *
4345  * min_type is the minimum key type to truncate down to.  If set to 0, this
4346  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4347  */
4348 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4349                                struct btrfs_root *root,
4350                                struct inode *inode,
4351                                u64 new_size, u32 min_type)
4352 {
4353         struct btrfs_path *path;
4354         struct extent_buffer *leaf;
4355         struct btrfs_file_extent_item *fi;
4356         struct btrfs_key key;
4357         struct btrfs_key found_key;
4358         u64 extent_start = 0;
4359         u64 extent_num_bytes = 0;
4360         u64 extent_offset = 0;
4361         u64 item_end = 0;
4362         u64 last_size = new_size;
4363         u32 found_type = (u8)-1;
4364         int found_extent;
4365         int del_item;
4366         int pending_del_nr = 0;
4367         int pending_del_slot = 0;
4368         int extent_type = -1;
4369         int ret;
4370         int err = 0;
4371         u64 ino = btrfs_ino(inode);
4372         u64 bytes_deleted = 0;
4373         bool be_nice = 0;
4374         bool should_throttle = 0;
4375         bool should_end = 0;
4376
4377         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4378
4379         /*
4380          * for non-free space inodes and ref cows, we want to back off from
4381          * time to time
4382          */
4383         if (!btrfs_is_free_space_inode(inode) &&
4384             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4385                 be_nice = 1;
4386
4387         path = btrfs_alloc_path();
4388         if (!path)
4389                 return -ENOMEM;
4390         path->reada = READA_BACK;
4391
4392         /*
4393          * We want to drop from the next block forward in case this new size is
4394          * not block aligned since we will be keeping the last block of the
4395          * extent just the way it is.
4396          */
4397         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4398             root == root->fs_info->tree_root)
4399                 btrfs_drop_extent_cache(inode, ALIGN(new_size,
4400                                         root->sectorsize), (u64)-1, 0);
4401
4402         /*
4403          * This function is also used to drop the items in the log tree before
4404          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4405          * it is used to drop the loged items. So we shouldn't kill the delayed
4406          * items.
4407          */
4408         if (min_type == 0 && root == BTRFS_I(inode)->root)
4409                 btrfs_kill_delayed_inode_items(inode);
4410
4411         key.objectid = ino;
4412         key.offset = (u64)-1;
4413         key.type = (u8)-1;
4414
4415 search_again:
4416         /*
4417          * with a 16K leaf size and 128MB extents, you can actually queue
4418          * up a huge file in a single leaf.  Most of the time that
4419          * bytes_deleted is > 0, it will be huge by the time we get here
4420          */
4421         if (be_nice && bytes_deleted > SZ_32M) {
4422                 if (btrfs_should_end_transaction(trans, root)) {
4423                         err = -EAGAIN;
4424                         goto error;
4425                 }
4426         }
4427
4428
4429         path->leave_spinning = 1;
4430         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4431         if (ret < 0) {
4432                 err = ret;
4433                 goto out;
4434         }
4435
4436         if (ret > 0) {
4437                 /* there are no items in the tree for us to truncate, we're
4438                  * done
4439                  */
4440                 if (path->slots[0] == 0)
4441                         goto out;
4442                 path->slots[0]--;
4443         }
4444
4445         while (1) {
4446                 fi = NULL;
4447                 leaf = path->nodes[0];
4448                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4449                 found_type = found_key.type;
4450
4451                 if (found_key.objectid != ino)
4452                         break;
4453
4454                 if (found_type < min_type)
4455                         break;
4456
4457                 item_end = found_key.offset;
4458                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4459                         fi = btrfs_item_ptr(leaf, path->slots[0],
4460                                             struct btrfs_file_extent_item);
4461                         extent_type = btrfs_file_extent_type(leaf, fi);
4462                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4463                                 item_end +=
4464                                     btrfs_file_extent_num_bytes(leaf, fi);
4465                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4466                                 item_end += btrfs_file_extent_inline_len(leaf,
4467                                                          path->slots[0], fi);
4468                         }
4469                         item_end--;
4470                 }
4471                 if (found_type > min_type) {
4472                         del_item = 1;
4473                 } else {
4474                         if (item_end < new_size)
4475                                 break;
4476                         if (found_key.offset >= new_size)
4477                                 del_item = 1;
4478                         else
4479                                 del_item = 0;
4480                 }
4481                 found_extent = 0;
4482                 /* FIXME, shrink the extent if the ref count is only 1 */
4483                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4484                         goto delete;
4485
4486                 if (del_item)
4487                         last_size = found_key.offset;
4488                 else
4489                         last_size = new_size;
4490
4491                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4492                         u64 num_dec;
4493                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4494                         if (!del_item) {
4495                                 u64 orig_num_bytes =
4496                                         btrfs_file_extent_num_bytes(leaf, fi);
4497                                 extent_num_bytes = ALIGN(new_size -
4498                                                 found_key.offset,
4499                                                 root->sectorsize);
4500                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4501                                                          extent_num_bytes);
4502                                 num_dec = (orig_num_bytes -
4503                                            extent_num_bytes);
4504                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4505                                              &root->state) &&
4506                                     extent_start != 0)
4507                                         inode_sub_bytes(inode, num_dec);
4508                                 btrfs_mark_buffer_dirty(leaf);
4509                         } else {
4510                                 extent_num_bytes =
4511                                         btrfs_file_extent_disk_num_bytes(leaf,
4512                                                                          fi);
4513                                 extent_offset = found_key.offset -
4514                                         btrfs_file_extent_offset(leaf, fi);
4515
4516                                 /* FIXME blocksize != 4096 */
4517                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4518                                 if (extent_start != 0) {
4519                                         found_extent = 1;
4520                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4521                                                      &root->state))
4522                                                 inode_sub_bytes(inode, num_dec);
4523                                 }
4524                         }
4525                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4526                         /*
4527                          * we can't truncate inline items that have had
4528                          * special encodings
4529                          */
4530                         if (!del_item &&
4531                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4532                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4533
4534                                 /*
4535                                  * Need to release path in order to truncate a
4536                                  * compressed extent. So delete any accumulated
4537                                  * extent items so far.
4538                                  */
4539                                 if (btrfs_file_extent_compression(leaf, fi) !=
4540                                     BTRFS_COMPRESS_NONE && pending_del_nr) {
4541                                         err = btrfs_del_items(trans, root, path,
4542                                                               pending_del_slot,
4543                                                               pending_del_nr);
4544                                         if (err) {
4545                                                 btrfs_abort_transaction(trans,
4546                                                                         err);
4547                                                 goto error;
4548                                         }
4549                                         pending_del_nr = 0;
4550                                 }
4551
4552                                 err = truncate_inline_extent(inode, path,
4553                                                              &found_key,
4554                                                              item_end,
4555                                                              new_size);
4556                                 if (err) {
4557                                         btrfs_abort_transaction(trans, err);
4558                                         goto error;
4559                                 }
4560                         } else if (test_bit(BTRFS_ROOT_REF_COWS,
4561                                             &root->state)) {
4562                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4563                         }
4564                 }
4565 delete:
4566                 if (del_item) {
4567                         if (!pending_del_nr) {
4568                                 /* no pending yet, add ourselves */
4569                                 pending_del_slot = path->slots[0];
4570                                 pending_del_nr = 1;
4571                         } else if (pending_del_nr &&
4572                                    path->slots[0] + 1 == pending_del_slot) {
4573                                 /* hop on the pending chunk */
4574                                 pending_del_nr++;
4575                                 pending_del_slot = path->slots[0];
4576                         } else {
4577                                 BUG();
4578                         }
4579                 } else {
4580                         break;
4581                 }
4582                 should_throttle = 0;
4583
4584                 if (found_extent &&
4585                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4586                      root == root->fs_info->tree_root)) {
4587                         btrfs_set_path_blocking(path);
4588                         bytes_deleted += extent_num_bytes;
4589                         ret = btrfs_free_extent(trans, root, extent_start,
4590                                                 extent_num_bytes, 0,
4591                                                 btrfs_header_owner(leaf),
4592                                                 ino, extent_offset);
4593                         BUG_ON(ret);
4594                         if (btrfs_should_throttle_delayed_refs(trans, root))
4595                                 btrfs_async_run_delayed_refs(root,
4596                                                              trans->transid,
4597                                         trans->delayed_ref_updates * 2, 0);
4598                         if (be_nice) {
4599                                 if (truncate_space_check(trans, root,
4600                                                          extent_num_bytes)) {
4601                                         should_end = 1;
4602                                 }
4603                                 if (btrfs_should_throttle_delayed_refs(trans,
4604                                                                        root)) {
4605                                         should_throttle = 1;
4606                                 }
4607                         }
4608                 }
4609
4610                 if (found_type == BTRFS_INODE_ITEM_KEY)
4611                         break;
4612
4613                 if (path->slots[0] == 0 ||
4614                     path->slots[0] != pending_del_slot ||
4615                     should_throttle || should_end) {
4616                         if (pending_del_nr) {
4617                                 ret = btrfs_del_items(trans, root, path,
4618                                                 pending_del_slot,
4619                                                 pending_del_nr);
4620                                 if (ret) {
4621                                         btrfs_abort_transaction(trans, ret);
4622                                         goto error;
4623                                 }
4624                                 pending_del_nr = 0;
4625                         }
4626                         btrfs_release_path(path);
4627                         if (should_throttle) {
4628                                 unsigned long updates = trans->delayed_ref_updates;
4629                                 if (updates) {
4630                                         trans->delayed_ref_updates = 0;
4631                                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4632                                         if (ret && !err)
4633                                                 err = ret;
4634                                 }
4635                         }
4636                         /*
4637                          * if we failed to refill our space rsv, bail out
4638                          * and let the transaction restart
4639                          */
4640                         if (should_end) {
4641                                 err = -EAGAIN;
4642                                 goto error;
4643                         }
4644                         goto search_again;
4645                 } else {
4646                         path->slots[0]--;
4647                 }
4648         }
4649 out:
4650         if (pending_del_nr) {
4651                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4652                                       pending_del_nr);
4653                 if (ret)
4654                         btrfs_abort_transaction(trans, ret);
4655         }
4656 error:
4657         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4658                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4659
4660         btrfs_free_path(path);
4661
4662         if (be_nice && bytes_deleted > SZ_32M) {
4663                 unsigned long updates = trans->delayed_ref_updates;
4664                 if (updates) {
4665                         trans->delayed_ref_updates = 0;
4666                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4667                         if (ret && !err)
4668                                 err = ret;
4669                 }
4670         }
4671         return err;
4672 }
4673
4674 /*
4675  * btrfs_truncate_block - read, zero a chunk and write a block
4676  * @inode - inode that we're zeroing
4677  * @from - the offset to start zeroing
4678  * @len - the length to zero, 0 to zero the entire range respective to the
4679  *      offset
4680  * @front - zero up to the offset instead of from the offset on
4681  *
4682  * This will find the block for the "from" offset and cow the block and zero the
4683  * part we want to zero.  This is used with truncate and hole punching.
4684  */
4685 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4686                         int front)
4687 {
4688         struct address_space *mapping = inode->i_mapping;
4689         struct btrfs_root *root = BTRFS_I(inode)->root;
4690         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4691         struct btrfs_ordered_extent *ordered;
4692         struct extent_state *cached_state = NULL;
4693         char *kaddr;
4694         u32 blocksize = root->sectorsize;
4695         pgoff_t index = from >> PAGE_SHIFT;
4696         unsigned offset = from & (blocksize - 1);
4697         struct page *page;
4698         gfp_t mask = btrfs_alloc_write_mask(mapping);
4699         int ret = 0;
4700         u64 block_start;
4701         u64 block_end;
4702
4703         if ((offset & (blocksize - 1)) == 0 &&
4704             (!len || ((len & (blocksize - 1)) == 0)))
4705                 goto out;
4706
4707         ret = btrfs_delalloc_reserve_space(inode,
4708                         round_down(from, blocksize), blocksize);
4709         if (ret)
4710                 goto out;
4711
4712 again:
4713         page = find_or_create_page(mapping, index, mask);
4714         if (!page) {
4715                 btrfs_delalloc_release_space(inode,
4716                                 round_down(from, blocksize),
4717                                 blocksize);
4718                 ret = -ENOMEM;
4719                 goto out;
4720         }
4721
4722         block_start = round_down(from, blocksize);
4723         block_end = block_start + blocksize - 1;
4724
4725         if (!PageUptodate(page)) {
4726                 ret = btrfs_readpage(NULL, page);
4727                 lock_page(page);
4728                 if (page->mapping != mapping) {
4729                         unlock_page(page);
4730                         put_page(page);
4731                         goto again;
4732                 }
4733                 if (!PageUptodate(page)) {
4734                         ret = -EIO;
4735                         goto out_unlock;
4736                 }
4737         }
4738         wait_on_page_writeback(page);
4739
4740         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4741         set_page_extent_mapped(page);
4742
4743         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4744         if (ordered) {
4745                 unlock_extent_cached(io_tree, block_start, block_end,
4746                                      &cached_state, GFP_NOFS);
4747                 unlock_page(page);
4748                 put_page(page);
4749                 btrfs_start_ordered_extent(inode, ordered, 1);
4750                 btrfs_put_ordered_extent(ordered);
4751                 goto again;
4752         }
4753
4754         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4755                           EXTENT_DIRTY | EXTENT_DELALLOC |
4756                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4757                           0, 0, &cached_state, GFP_NOFS);
4758
4759         ret = btrfs_set_extent_delalloc(inode, block_start, block_end,
4760                                         &cached_state);
4761         if (ret) {
4762                 unlock_extent_cached(io_tree, block_start, block_end,
4763                                      &cached_state, GFP_NOFS);
4764                 goto out_unlock;
4765         }
4766
4767         if (offset != blocksize) {
4768                 if (!len)
4769                         len = blocksize - offset;
4770                 kaddr = kmap(page);
4771                 if (front)
4772                         memset(kaddr + (block_start - page_offset(page)),
4773                                 0, offset);
4774                 else
4775                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4776                                 0, len);
4777                 flush_dcache_page(page);
4778                 kunmap(page);
4779         }
4780         ClearPageChecked(page);
4781         set_page_dirty(page);
4782         unlock_extent_cached(io_tree, block_start, block_end, &cached_state,
4783                              GFP_NOFS);
4784
4785 out_unlock:
4786         if (ret)
4787                 btrfs_delalloc_release_space(inode, block_start,
4788                                              blocksize);
4789         unlock_page(page);
4790         put_page(page);
4791 out:
4792         return ret;
4793 }
4794
4795 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4796                              u64 offset, u64 len)
4797 {
4798         struct btrfs_trans_handle *trans;
4799         int ret;
4800
4801         /*
4802          * Still need to make sure the inode looks like it's been updated so
4803          * that any holes get logged if we fsync.
4804          */
4805         if (btrfs_fs_incompat(root->fs_info, NO_HOLES)) {
4806                 BTRFS_I(inode)->last_trans = root->fs_info->generation;
4807                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4808                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4809                 return 0;
4810         }
4811
4812         /*
4813          * 1 - for the one we're dropping
4814          * 1 - for the one we're adding
4815          * 1 - for updating the inode.
4816          */
4817         trans = btrfs_start_transaction(root, 3);
4818         if (IS_ERR(trans))
4819                 return PTR_ERR(trans);
4820
4821         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4822         if (ret) {
4823                 btrfs_abort_transaction(trans, ret);
4824                 btrfs_end_transaction(trans, root);
4825                 return ret;
4826         }
4827
4828         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
4829                                        0, 0, len, 0, len, 0, 0, 0);
4830         if (ret)
4831                 btrfs_abort_transaction(trans, ret);
4832         else
4833                 btrfs_update_inode(trans, root, inode);
4834         btrfs_end_transaction(trans, root);
4835         return ret;
4836 }
4837
4838 /*
4839  * This function puts in dummy file extents for the area we're creating a hole
4840  * for.  So if we are truncating this file to a larger size we need to insert
4841  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4842  * the range between oldsize and size
4843  */
4844 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4845 {
4846         struct btrfs_root *root = BTRFS_I(inode)->root;
4847         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4848         struct extent_map *em = NULL;
4849         struct extent_state *cached_state = NULL;
4850         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4851         u64 hole_start = ALIGN(oldsize, root->sectorsize);
4852         u64 block_end = ALIGN(size, root->sectorsize);
4853         u64 last_byte;
4854         u64 cur_offset;
4855         u64 hole_size;
4856         int err = 0;
4857
4858         /*
4859          * If our size started in the middle of a block we need to zero out the
4860          * rest of the block before we expand the i_size, otherwise we could
4861          * expose stale data.
4862          */
4863         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4864         if (err)
4865                 return err;
4866
4867         if (size <= hole_start)
4868                 return 0;
4869
4870         while (1) {
4871                 struct btrfs_ordered_extent *ordered;
4872
4873                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4874                                  &cached_state);
4875                 ordered = btrfs_lookup_ordered_range(inode, hole_start,
4876                                                      block_end - hole_start);
4877                 if (!ordered)
4878                         break;
4879                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4880                                      &cached_state, GFP_NOFS);
4881                 btrfs_start_ordered_extent(inode, ordered, 1);
4882                 btrfs_put_ordered_extent(ordered);
4883         }
4884
4885         cur_offset = hole_start;
4886         while (1) {
4887                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
4888                                 block_end - cur_offset, 0);
4889                 if (IS_ERR(em)) {
4890                         err = PTR_ERR(em);
4891                         em = NULL;
4892                         break;
4893                 }
4894                 last_byte = min(extent_map_end(em), block_end);
4895                 last_byte = ALIGN(last_byte , root->sectorsize);
4896                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4897                         struct extent_map *hole_em;
4898                         hole_size = last_byte - cur_offset;
4899
4900                         err = maybe_insert_hole(root, inode, cur_offset,
4901                                                 hole_size);
4902                         if (err)
4903                                 break;
4904                         btrfs_drop_extent_cache(inode, cur_offset,
4905                                                 cur_offset + hole_size - 1, 0);
4906                         hole_em = alloc_extent_map();
4907                         if (!hole_em) {
4908                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4909                                         &BTRFS_I(inode)->runtime_flags);
4910                                 goto next;
4911                         }
4912                         hole_em->start = cur_offset;
4913                         hole_em->len = hole_size;
4914                         hole_em->orig_start = cur_offset;
4915
4916                         hole_em->block_start = EXTENT_MAP_HOLE;
4917                         hole_em->block_len = 0;
4918                         hole_em->orig_block_len = 0;
4919                         hole_em->ram_bytes = hole_size;
4920                         hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
4921                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4922                         hole_em->generation = root->fs_info->generation;
4923
4924                         while (1) {
4925                                 write_lock(&em_tree->lock);
4926                                 err = add_extent_mapping(em_tree, hole_em, 1);
4927                                 write_unlock(&em_tree->lock);
4928                                 if (err != -EEXIST)
4929                                         break;
4930                                 btrfs_drop_extent_cache(inode, cur_offset,
4931                                                         cur_offset +
4932                                                         hole_size - 1, 0);
4933                         }
4934                         free_extent_map(hole_em);
4935                 }
4936 next:
4937                 free_extent_map(em);
4938                 em = NULL;
4939                 cur_offset = last_byte;
4940                 if (cur_offset >= block_end)
4941                         break;
4942         }
4943         free_extent_map(em);
4944         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4945                              GFP_NOFS);
4946         return err;
4947 }
4948
4949 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4950 {
4951         struct btrfs_root *root = BTRFS_I(inode)->root;
4952         struct btrfs_trans_handle *trans;
4953         loff_t oldsize = i_size_read(inode);
4954         loff_t newsize = attr->ia_size;
4955         int mask = attr->ia_valid;
4956         int ret;
4957
4958         /*
4959          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
4960          * special case where we need to update the times despite not having
4961          * these flags set.  For all other operations the VFS set these flags
4962          * explicitly if it wants a timestamp update.
4963          */
4964         if (newsize != oldsize) {
4965                 inode_inc_iversion(inode);
4966                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
4967                         inode->i_ctime = inode->i_mtime =
4968                                 current_fs_time(inode->i_sb);
4969         }
4970
4971         if (newsize > oldsize) {
4972                 /*
4973                  * Don't do an expanding truncate while snapshoting is ongoing.
4974                  * This is to ensure the snapshot captures a fully consistent
4975                  * state of this file - if the snapshot captures this expanding
4976                  * truncation, it must capture all writes that happened before
4977                  * this truncation.
4978                  */
4979                 btrfs_wait_for_snapshot_creation(root);
4980                 ret = btrfs_cont_expand(inode, oldsize, newsize);
4981                 if (ret) {
4982                         btrfs_end_write_no_snapshoting(root);
4983                         return ret;
4984                 }
4985
4986                 trans = btrfs_start_transaction(root, 1);
4987                 if (IS_ERR(trans)) {
4988                         btrfs_end_write_no_snapshoting(root);
4989                         return PTR_ERR(trans);
4990                 }
4991
4992                 i_size_write(inode, newsize);
4993                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
4994                 pagecache_isize_extended(inode, oldsize, newsize);
4995                 ret = btrfs_update_inode(trans, root, inode);
4996                 btrfs_end_write_no_snapshoting(root);
4997                 btrfs_end_transaction(trans, root);
4998         } else {
4999
5000                 /*
5001                  * We're truncating a file that used to have good data down to
5002                  * zero. Make sure it gets into the ordered flush list so that
5003                  * any new writes get down to disk quickly.
5004                  */
5005                 if (newsize == 0)
5006                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5007                                 &BTRFS_I(inode)->runtime_flags);
5008
5009                 /*
5010                  * 1 for the orphan item we're going to add
5011                  * 1 for the orphan item deletion.
5012                  */
5013                 trans = btrfs_start_transaction(root, 2);
5014                 if (IS_ERR(trans))
5015                         return PTR_ERR(trans);
5016
5017                 /*
5018                  * We need to do this in case we fail at _any_ point during the
5019                  * actual truncate.  Once we do the truncate_setsize we could
5020                  * invalidate pages which forces any outstanding ordered io to
5021                  * be instantly completed which will give us extents that need
5022                  * to be truncated.  If we fail to get an orphan inode down we
5023                  * could have left over extents that were never meant to live,
5024                  * so we need to guarantee from this point on that everything
5025                  * will be consistent.
5026                  */
5027                 ret = btrfs_orphan_add(trans, inode);
5028                 btrfs_end_transaction(trans, root);
5029                 if (ret)
5030                         return ret;
5031
5032                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
5033                 truncate_setsize(inode, newsize);
5034
5035                 /* Disable nonlocked read DIO to avoid the end less truncate */
5036                 btrfs_inode_block_unlocked_dio(inode);
5037                 inode_dio_wait(inode);
5038                 btrfs_inode_resume_unlocked_dio(inode);
5039
5040                 ret = btrfs_truncate(inode);
5041                 if (ret && inode->i_nlink) {
5042                         int err;
5043
5044                         /*
5045                          * failed to truncate, disk_i_size is only adjusted down
5046                          * as we remove extents, so it should represent the true
5047                          * size of the inode, so reset the in memory size and
5048                          * delete our orphan entry.
5049                          */
5050                         trans = btrfs_join_transaction(root);
5051                         if (IS_ERR(trans)) {
5052                                 btrfs_orphan_del(NULL, inode);
5053                                 return ret;
5054                         }
5055                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5056                         err = btrfs_orphan_del(trans, inode);
5057                         if (err)
5058                                 btrfs_abort_transaction(trans, err);
5059                         btrfs_end_transaction(trans, root);
5060                 }
5061         }
5062
5063         return ret;
5064 }
5065
5066 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5067 {
5068         struct inode *inode = d_inode(dentry);
5069         struct btrfs_root *root = BTRFS_I(inode)->root;
5070         int err;
5071
5072         if (btrfs_root_readonly(root))
5073                 return -EROFS;
5074
5075         err = inode_change_ok(inode, attr);
5076         if (err)
5077                 return err;
5078
5079         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5080                 err = btrfs_setsize(inode, attr);
5081                 if (err)
5082                         return err;
5083         }
5084
5085         if (attr->ia_valid) {
5086                 setattr_copy(inode, attr);
5087                 inode_inc_iversion(inode);
5088                 err = btrfs_dirty_inode(inode);
5089
5090                 if (!err && attr->ia_valid & ATTR_MODE)
5091                         err = posix_acl_chmod(inode, inode->i_mode);
5092         }
5093
5094         return err;
5095 }
5096
5097 /*
5098  * While truncating the inode pages during eviction, we get the VFS calling
5099  * btrfs_invalidatepage() against each page of the inode. This is slow because
5100  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5101  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5102  * extent_state structures over and over, wasting lots of time.
5103  *
5104  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5105  * those expensive operations on a per page basis and do only the ordered io
5106  * finishing, while we release here the extent_map and extent_state structures,
5107  * without the excessive merging and splitting.
5108  */
5109 static void evict_inode_truncate_pages(struct inode *inode)
5110 {
5111         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5112         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5113         struct rb_node *node;
5114
5115         ASSERT(inode->i_state & I_FREEING);
5116         truncate_inode_pages_final(&inode->i_data);
5117
5118         write_lock(&map_tree->lock);
5119         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5120                 struct extent_map *em;
5121
5122                 node = rb_first(&map_tree->map);
5123                 em = rb_entry(node, struct extent_map, rb_node);
5124                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5125                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5126                 remove_extent_mapping(map_tree, em);
5127                 free_extent_map(em);
5128                 if (need_resched()) {
5129                         write_unlock(&map_tree->lock);
5130                         cond_resched();
5131                         write_lock(&map_tree->lock);
5132                 }
5133         }
5134         write_unlock(&map_tree->lock);
5135
5136         /*
5137          * Keep looping until we have no more ranges in the io tree.
5138          * We can have ongoing bios started by readpages (called from readahead)
5139          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5140          * still in progress (unlocked the pages in the bio but did not yet
5141          * unlocked the ranges in the io tree). Therefore this means some
5142          * ranges can still be locked and eviction started because before
5143          * submitting those bios, which are executed by a separate task (work
5144          * queue kthread), inode references (inode->i_count) were not taken
5145          * (which would be dropped in the end io callback of each bio).
5146          * Therefore here we effectively end up waiting for those bios and
5147          * anyone else holding locked ranges without having bumped the inode's
5148          * reference count - if we don't do it, when they access the inode's
5149          * io_tree to unlock a range it may be too late, leading to an
5150          * use-after-free issue.
5151          */
5152         spin_lock(&io_tree->lock);
5153         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5154                 struct extent_state *state;
5155                 struct extent_state *cached_state = NULL;
5156                 u64 start;
5157                 u64 end;
5158
5159                 node = rb_first(&io_tree->state);
5160                 state = rb_entry(node, struct extent_state, rb_node);
5161                 start = state->start;
5162                 end = state->end;
5163                 spin_unlock(&io_tree->lock);
5164
5165                 lock_extent_bits(io_tree, start, end, &cached_state);
5166
5167                 /*
5168                  * If still has DELALLOC flag, the extent didn't reach disk,
5169                  * and its reserved space won't be freed by delayed_ref.
5170                  * So we need to free its reserved space here.
5171                  * (Refer to comment in btrfs_invalidatepage, case 2)
5172                  *
5173                  * Note, end is the bytenr of last byte, so we need + 1 here.
5174                  */
5175                 if (state->state & EXTENT_DELALLOC)
5176                         btrfs_qgroup_free_data(inode, start, end - start + 1);
5177
5178                 clear_extent_bit(io_tree, start, end,
5179                                  EXTENT_LOCKED | EXTENT_DIRTY |
5180                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5181                                  EXTENT_DEFRAG, 1, 1,
5182                                  &cached_state, GFP_NOFS);
5183
5184                 cond_resched();
5185                 spin_lock(&io_tree->lock);
5186         }
5187         spin_unlock(&io_tree->lock);
5188 }
5189
5190 void btrfs_evict_inode(struct inode *inode)
5191 {
5192         struct btrfs_trans_handle *trans;
5193         struct btrfs_root *root = BTRFS_I(inode)->root;
5194         struct btrfs_block_rsv *rsv, *global_rsv;
5195         int steal_from_global = 0;
5196         u64 min_size;
5197         int ret;
5198
5199         trace_btrfs_inode_evict(inode);
5200
5201         if (!root) {
5202                 kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
5203                 return;
5204         }
5205
5206         min_size = btrfs_calc_trunc_metadata_size(root, 1);
5207
5208         evict_inode_truncate_pages(inode);
5209
5210         if (inode->i_nlink &&
5211             ((btrfs_root_refs(&root->root_item) != 0 &&
5212               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5213              btrfs_is_free_space_inode(inode)))
5214                 goto no_delete;
5215
5216         if (is_bad_inode(inode)) {
5217                 btrfs_orphan_del(NULL, inode);
5218                 goto no_delete;
5219         }
5220         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5221         if (!special_file(inode->i_mode))
5222                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5223
5224         btrfs_free_io_failure_record(inode, 0, (u64)-1);
5225
5226         if (root->fs_info->log_root_recovering) {
5227                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5228                                  &BTRFS_I(inode)->runtime_flags));
5229                 goto no_delete;
5230         }
5231
5232         if (inode->i_nlink > 0) {
5233                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5234                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5235                 goto no_delete;
5236         }
5237
5238         ret = btrfs_commit_inode_delayed_inode(inode);
5239         if (ret) {
5240                 btrfs_orphan_del(NULL, inode);
5241                 goto no_delete;
5242         }
5243
5244         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
5245         if (!rsv) {
5246                 btrfs_orphan_del(NULL, inode);
5247                 goto no_delete;
5248         }
5249         rsv->size = min_size;
5250         rsv->failfast = 1;
5251         global_rsv = &root->fs_info->global_block_rsv;
5252
5253         btrfs_i_size_write(inode, 0);
5254
5255         /*
5256          * This is a bit simpler than btrfs_truncate since we've already
5257          * reserved our space for our orphan item in the unlink, so we just
5258          * need to reserve some slack space in case we add bytes and update
5259          * inode item when doing the truncate.
5260          */
5261         while (1) {
5262                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5263                                              BTRFS_RESERVE_FLUSH_LIMIT);
5264
5265                 /*
5266                  * Try and steal from the global reserve since we will
5267                  * likely not use this space anyway, we want to try as
5268                  * hard as possible to get this to work.
5269                  */
5270                 if (ret)
5271                         steal_from_global++;
5272                 else
5273                         steal_from_global = 0;
5274                 ret = 0;
5275
5276                 /*
5277                  * steal_from_global == 0: we reserved stuff, hooray!
5278                  * steal_from_global == 1: we didn't reserve stuff, boo!
5279                  * steal_from_global == 2: we've committed, still not a lot of
5280                  * room but maybe we'll have room in the global reserve this
5281                  * time.
5282                  * steal_from_global == 3: abandon all hope!
5283                  */
5284                 if (steal_from_global > 2) {
5285                         btrfs_warn(root->fs_info,
5286                                 "Could not get space for a delete, will truncate on mount %d",
5287                                 ret);
5288                         btrfs_orphan_del(NULL, inode);
5289                         btrfs_free_block_rsv(root, rsv);
5290                         goto no_delete;
5291                 }
5292
5293                 trans = btrfs_join_transaction(root);
5294                 if (IS_ERR(trans)) {
5295                         btrfs_orphan_del(NULL, inode);
5296                         btrfs_free_block_rsv(root, rsv);
5297                         goto no_delete;
5298                 }
5299
5300                 /*
5301                  * We can't just steal from the global reserve, we need to make
5302                  * sure there is room to do it, if not we need to commit and try
5303                  * again.
5304                  */
5305                 if (steal_from_global) {
5306                         if (!btrfs_check_space_for_delayed_refs(trans, root))
5307                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5308                                                               min_size, 0);
5309                         else
5310                                 ret = -ENOSPC;
5311                 }
5312
5313                 /*
5314                  * Couldn't steal from the global reserve, we have too much
5315                  * pending stuff built up, commit the transaction and try it
5316                  * again.
5317                  */
5318                 if (ret) {
5319                         ret = btrfs_commit_transaction(trans, root);
5320                         if (ret) {
5321                                 btrfs_orphan_del(NULL, inode);
5322                                 btrfs_free_block_rsv(root, rsv);
5323                                 goto no_delete;
5324                         }
5325                         continue;
5326                 } else {
5327                         steal_from_global = 0;
5328                 }
5329
5330                 trans->block_rsv = rsv;
5331
5332                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5333                 if (ret != -ENOSPC && ret != -EAGAIN)
5334                         break;
5335
5336                 trans->block_rsv = &root->fs_info->trans_block_rsv;
5337                 btrfs_end_transaction(trans, root);
5338                 trans = NULL;
5339                 btrfs_btree_balance_dirty(root);
5340         }
5341
5342         btrfs_free_block_rsv(root, rsv);
5343
5344         /*
5345          * Errors here aren't a big deal, it just means we leave orphan items
5346          * in the tree.  They will be cleaned up on the next mount.
5347          */
5348         if (ret == 0) {
5349                 trans->block_rsv = root->orphan_block_rsv;
5350                 btrfs_orphan_del(trans, inode);
5351         } else {
5352                 btrfs_orphan_del(NULL, inode);
5353         }
5354
5355         trans->block_rsv = &root->fs_info->trans_block_rsv;
5356         if (!(root == root->fs_info->tree_root ||
5357               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5358                 btrfs_return_ino(root, btrfs_ino(inode));
5359
5360         btrfs_end_transaction(trans, root);
5361         btrfs_btree_balance_dirty(root);
5362 no_delete:
5363         btrfs_remove_delayed_node(inode);
5364         clear_inode(inode);
5365 }
5366
5367 /*
5368  * this returns the key found in the dir entry in the location pointer.
5369  * If no dir entries were found, location->objectid is 0.
5370  */
5371 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5372                                struct btrfs_key *location)
5373 {
5374         const char *name = dentry->d_name.name;
5375         int namelen = dentry->d_name.len;
5376         struct btrfs_dir_item *di;
5377         struct btrfs_path *path;
5378         struct btrfs_root *root = BTRFS_I(dir)->root;
5379         int ret = 0;
5380
5381         path = btrfs_alloc_path();
5382         if (!path)
5383                 return -ENOMEM;
5384
5385         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(dir), name,
5386                                     namelen, 0);
5387         if (IS_ERR(di))
5388                 ret = PTR_ERR(di);
5389
5390         if (IS_ERR_OR_NULL(di))
5391                 goto out_err;
5392
5393         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5394 out:
5395         btrfs_free_path(path);
5396         return ret;
5397 out_err:
5398         location->objectid = 0;
5399         goto out;
5400 }
5401
5402 /*
5403  * when we hit a tree root in a directory, the btrfs part of the inode
5404  * needs to be changed to reflect the root directory of the tree root.  This
5405  * is kind of like crossing a mount point.
5406  */
5407 static int fixup_tree_root_location(struct btrfs_root *root,
5408                                     struct inode *dir,
5409                                     struct dentry *dentry,
5410                                     struct btrfs_key *location,
5411                                     struct btrfs_root **sub_root)
5412 {
5413         struct btrfs_path *path;
5414         struct btrfs_root *new_root;
5415         struct btrfs_root_ref *ref;
5416         struct extent_buffer *leaf;
5417         struct btrfs_key key;
5418         int ret;
5419         int err = 0;
5420
5421         path = btrfs_alloc_path();
5422         if (!path) {
5423                 err = -ENOMEM;
5424                 goto out;
5425         }
5426
5427         err = -ENOENT;
5428         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5429         key.type = BTRFS_ROOT_REF_KEY;
5430         key.offset = location->objectid;
5431
5432         ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, path,
5433                                 0, 0);
5434         if (ret) {
5435                 if (ret < 0)
5436                         err = ret;
5437                 goto out;
5438         }
5439
5440         leaf = path->nodes[0];
5441         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5442         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(dir) ||
5443             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5444                 goto out;
5445
5446         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5447                                    (unsigned long)(ref + 1),
5448                                    dentry->d_name.len);
5449         if (ret)
5450                 goto out;
5451
5452         btrfs_release_path(path);
5453
5454         new_root = btrfs_read_fs_root_no_name(root->fs_info, location);
5455         if (IS_ERR(new_root)) {
5456                 err = PTR_ERR(new_root);
5457                 goto out;
5458         }
5459
5460         *sub_root = new_root;
5461         location->objectid = btrfs_root_dirid(&new_root->root_item);
5462         location->type = BTRFS_INODE_ITEM_KEY;
5463         location->offset = 0;
5464         err = 0;
5465 out:
5466         btrfs_free_path(path);
5467         return err;
5468 }
5469
5470 static void inode_tree_add(struct inode *inode)
5471 {
5472         struct btrfs_root *root = BTRFS_I(inode)->root;
5473         struct btrfs_inode *entry;
5474         struct rb_node **p;
5475         struct rb_node *parent;
5476         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5477         u64 ino = btrfs_ino(inode);
5478
5479         if (inode_unhashed(inode))
5480                 return;
5481         parent = NULL;
5482         spin_lock(&root->inode_lock);
5483         p = &root->inode_tree.rb_node;
5484         while (*p) {
5485                 parent = *p;
5486                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5487
5488                 if (ino < btrfs_ino(&entry->vfs_inode))
5489                         p = &parent->rb_left;
5490                 else if (ino > btrfs_ino(&entry->vfs_inode))
5491                         p = &parent->rb_right;
5492                 else {
5493                         WARN_ON(!(entry->vfs_inode.i_state &
5494                                   (I_WILL_FREE | I_FREEING)));
5495                         rb_replace_node(parent, new, &root->inode_tree);
5496                         RB_CLEAR_NODE(parent);
5497                         spin_unlock(&root->inode_lock);
5498                         return;
5499                 }
5500         }
5501         rb_link_node(new, parent, p);
5502         rb_insert_color(new, &root->inode_tree);
5503         spin_unlock(&root->inode_lock);
5504 }
5505
5506 static void inode_tree_del(struct inode *inode)
5507 {
5508         struct btrfs_root *root = BTRFS_I(inode)->root;
5509         int empty = 0;
5510
5511         spin_lock(&root->inode_lock);
5512         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5513                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5514                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5515                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5516         }
5517         spin_unlock(&root->inode_lock);
5518
5519         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5520                 synchronize_srcu(&root->fs_info->subvol_srcu);
5521                 spin_lock(&root->inode_lock);
5522                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5523                 spin_unlock(&root->inode_lock);
5524                 if (empty)
5525                         btrfs_add_dead_root(root);
5526         }
5527 }
5528
5529 void btrfs_invalidate_inodes(struct btrfs_root *root)
5530 {
5531         struct rb_node *node;
5532         struct rb_node *prev;
5533         struct btrfs_inode *entry;
5534         struct inode *inode;
5535         u64 objectid = 0;
5536
5537         if (!test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
5538                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5539
5540         spin_lock(&root->inode_lock);
5541 again:
5542         node = root->inode_tree.rb_node;
5543         prev = NULL;
5544         while (node) {
5545                 prev = node;
5546                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5547
5548                 if (objectid < btrfs_ino(&entry->vfs_inode))
5549                         node = node->rb_left;
5550                 else if (objectid > btrfs_ino(&entry->vfs_inode))
5551                         node = node->rb_right;
5552                 else
5553                         break;
5554         }
5555         if (!node) {
5556                 while (prev) {
5557                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5558                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
5559                                 node = prev;
5560                                 break;
5561                         }
5562                         prev = rb_next(prev);
5563                 }
5564         }
5565         while (node) {
5566                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5567                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
5568                 inode = igrab(&entry->vfs_inode);
5569                 if (inode) {
5570                         spin_unlock(&root->inode_lock);
5571                         if (atomic_read(&inode->i_count) > 1)
5572                                 d_prune_aliases(inode);
5573                         /*
5574                          * btrfs_drop_inode will have it removed from
5575                          * the inode cache when its usage count
5576                          * hits zero.
5577                          */
5578                         iput(inode);
5579                         cond_resched();
5580                         spin_lock(&root->inode_lock);
5581                         goto again;
5582                 }
5583
5584                 if (cond_resched_lock(&root->inode_lock))
5585                         goto again;
5586
5587                 node = rb_next(node);
5588         }
5589         spin_unlock(&root->inode_lock);
5590 }
5591
5592 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5593 {
5594         struct btrfs_iget_args *args = p;
5595         inode->i_ino = args->location->objectid;
5596         memcpy(&BTRFS_I(inode)->location, args->location,
5597                sizeof(*args->location));
5598         BTRFS_I(inode)->root = args->root;
5599         return 0;
5600 }
5601
5602 static int btrfs_find_actor(struct inode *inode, void *opaque)
5603 {
5604         struct btrfs_iget_args *args = opaque;
5605         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5606                 args->root == BTRFS_I(inode)->root;
5607 }
5608
5609 static struct inode *btrfs_iget_locked(struct super_block *s,
5610                                        struct btrfs_key *location,
5611                                        struct btrfs_root *root)
5612 {
5613         struct inode *inode;
5614         struct btrfs_iget_args args;
5615         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5616
5617         args.location = location;
5618         args.root = root;
5619
5620         inode = iget5_locked(s, hashval, btrfs_find_actor,
5621                              btrfs_init_locked_inode,
5622                              (void *)&args);
5623         return inode;
5624 }
5625
5626 /* Get an inode object given its location and corresponding root.
5627  * Returns in *is_new if the inode was read from disk
5628  */
5629 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5630                          struct btrfs_root *root, int *new)
5631 {
5632         struct inode *inode;
5633
5634         inode = btrfs_iget_locked(s, location, root);
5635         if (!inode)
5636                 return ERR_PTR(-ENOMEM);
5637
5638         if (inode->i_state & I_NEW) {
5639                 int ret;
5640
5641                 ret = btrfs_read_locked_inode(inode);
5642                 if (!is_bad_inode(inode)) {
5643                         inode_tree_add(inode);
5644                         unlock_new_inode(inode);
5645                         if (new)
5646                                 *new = 1;
5647                 } else {
5648                         unlock_new_inode(inode);
5649                         iput(inode);
5650                         ASSERT(ret < 0);
5651                         inode = ERR_PTR(ret < 0 ? ret : -ESTALE);
5652                 }
5653         }
5654
5655         return inode;
5656 }
5657
5658 static struct inode *new_simple_dir(struct super_block *s,
5659                                     struct btrfs_key *key,
5660                                     struct btrfs_root *root)
5661 {
5662         struct inode *inode = new_inode(s);
5663
5664         if (!inode)
5665                 return ERR_PTR(-ENOMEM);
5666
5667         BTRFS_I(inode)->root = root;
5668         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5669         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5670
5671         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5672         inode->i_op = &btrfs_dir_ro_inode_operations;
5673         inode->i_fop = &simple_dir_operations;
5674         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5675         inode->i_mtime = current_fs_time(inode->i_sb);
5676         inode->i_atime = inode->i_mtime;
5677         inode->i_ctime = inode->i_mtime;
5678         BTRFS_I(inode)->i_otime = inode->i_mtime;
5679
5680         return inode;
5681 }
5682
5683 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5684 {
5685         struct inode *inode;
5686         struct btrfs_root *root = BTRFS_I(dir)->root;
5687         struct btrfs_root *sub_root = root;
5688         struct btrfs_key location;
5689         int index;
5690         int ret = 0;
5691
5692         if (dentry->d_name.len > BTRFS_NAME_LEN)
5693                 return ERR_PTR(-ENAMETOOLONG);
5694
5695         ret = btrfs_inode_by_name(dir, dentry, &location);
5696         if (ret < 0)
5697                 return ERR_PTR(ret);
5698
5699         if (location.objectid == 0)
5700                 return ERR_PTR(-ENOENT);
5701
5702         if (location.type == BTRFS_INODE_ITEM_KEY) {
5703                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5704                 return inode;
5705         }
5706
5707         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
5708
5709         index = srcu_read_lock(&root->fs_info->subvol_srcu);
5710         ret = fixup_tree_root_location(root, dir, dentry,
5711                                        &location, &sub_root);
5712         if (ret < 0) {
5713                 if (ret != -ENOENT)
5714                         inode = ERR_PTR(ret);
5715                 else
5716                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5717         } else {
5718                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5719         }
5720         srcu_read_unlock(&root->fs_info->subvol_srcu, index);
5721
5722         if (!IS_ERR(inode) && root != sub_root) {
5723                 down_read(&root->fs_info->cleanup_work_sem);
5724                 if (!(inode->i_sb->s_flags & MS_RDONLY))
5725                         ret = btrfs_orphan_cleanup(sub_root);
5726                 up_read(&root->fs_info->cleanup_work_sem);
5727                 if (ret) {
5728                         iput(inode);
5729                         inode = ERR_PTR(ret);
5730                 }
5731         }
5732
5733         return inode;
5734 }
5735
5736 static int btrfs_dentry_delete(const struct dentry *dentry)
5737 {
5738         struct btrfs_root *root;
5739         struct inode *inode = d_inode(dentry);
5740
5741         if (!inode && !IS_ROOT(dentry))
5742                 inode = d_inode(dentry->d_parent);
5743
5744         if (inode) {
5745                 root = BTRFS_I(inode)->root;
5746                 if (btrfs_root_refs(&root->root_item) == 0)
5747                         return 1;
5748
5749                 if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5750                         return 1;
5751         }
5752         return 0;
5753 }
5754
5755 static void btrfs_dentry_release(struct dentry *dentry)
5756 {
5757         kfree(dentry->d_fsdata);
5758 }
5759
5760 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5761                                    unsigned int flags)
5762 {
5763         struct inode *inode;
5764
5765         inode = btrfs_lookup_dentry(dir, dentry);
5766         if (IS_ERR(inode)) {
5767                 if (PTR_ERR(inode) == -ENOENT)
5768                         inode = NULL;
5769                 else
5770                         return ERR_CAST(inode);
5771         }
5772
5773         return d_splice_alias(inode, dentry);
5774 }
5775
5776 unsigned char btrfs_filetype_table[] = {
5777         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5778 };
5779
5780 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5781 {
5782         struct inode *inode = file_inode(file);
5783         struct btrfs_root *root = BTRFS_I(inode)->root;
5784         struct btrfs_item *item;
5785         struct btrfs_dir_item *di;
5786         struct btrfs_key key;
5787         struct btrfs_key found_key;
5788         struct btrfs_path *path;
5789         struct list_head ins_list;
5790         struct list_head del_list;
5791         int ret;
5792         struct extent_buffer *leaf;
5793         int slot;
5794         unsigned char d_type;
5795         int over = 0;
5796         u32 di_cur;
5797         u32 di_total;
5798         u32 di_len;
5799         int key_type = BTRFS_DIR_INDEX_KEY;
5800         char tmp_name[32];
5801         char *name_ptr;
5802         int name_len;
5803         int is_curr = 0;        /* ctx->pos points to the current index? */
5804         bool emitted;
5805         bool put = false;
5806
5807         /* FIXME, use a real flag for deciding about the key type */
5808         if (root->fs_info->tree_root == root)
5809                 key_type = BTRFS_DIR_ITEM_KEY;
5810
5811         if (!dir_emit_dots(file, ctx))
5812                 return 0;
5813
5814         path = btrfs_alloc_path();
5815         if (!path)
5816                 return -ENOMEM;
5817
5818         path->reada = READA_FORWARD;
5819
5820         if (key_type == BTRFS_DIR_INDEX_KEY) {
5821                 INIT_LIST_HEAD(&ins_list);
5822                 INIT_LIST_HEAD(&del_list);
5823                 put = btrfs_readdir_get_delayed_items(inode, &ins_list,
5824                                                       &del_list);
5825         }
5826
5827         key.type = key_type;
5828         key.offset = ctx->pos;
5829         key.objectid = btrfs_ino(inode);
5830
5831         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5832         if (ret < 0)
5833                 goto err;
5834
5835         emitted = false;
5836         while (1) {
5837                 leaf = path->nodes[0];
5838                 slot = path->slots[0];
5839                 if (slot >= btrfs_header_nritems(leaf)) {
5840                         ret = btrfs_next_leaf(root, path);
5841                         if (ret < 0)
5842                                 goto err;
5843                         else if (ret > 0)
5844                                 break;
5845                         continue;
5846                 }
5847
5848                 item = btrfs_item_nr(slot);
5849                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5850
5851                 if (found_key.objectid != key.objectid)
5852                         break;
5853                 if (found_key.type != key_type)
5854                         break;
5855                 if (found_key.offset < ctx->pos)
5856                         goto next;
5857                 if (key_type == BTRFS_DIR_INDEX_KEY &&
5858                     btrfs_should_delete_dir_index(&del_list,
5859                                                   found_key.offset))
5860                         goto next;
5861
5862                 ctx->pos = found_key.offset;
5863                 is_curr = 1;
5864
5865                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5866                 di_cur = 0;
5867                 di_total = btrfs_item_size(leaf, item);
5868
5869                 while (di_cur < di_total) {
5870                         struct btrfs_key location;
5871
5872                         if (verify_dir_item(root, leaf, di))
5873                                 break;
5874
5875                         name_len = btrfs_dir_name_len(leaf, di);
5876                         if (name_len <= sizeof(tmp_name)) {
5877                                 name_ptr = tmp_name;
5878                         } else {
5879                                 name_ptr = kmalloc(name_len, GFP_KERNEL);
5880                                 if (!name_ptr) {
5881                                         ret = -ENOMEM;
5882                                         goto err;
5883                                 }
5884                         }
5885                         read_extent_buffer(leaf, name_ptr,
5886                                            (unsigned long)(di + 1), name_len);
5887
5888                         d_type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5889                         btrfs_dir_item_key_to_cpu(leaf, di, &location);
5890
5891
5892                         /* is this a reference to our own snapshot? If so
5893                          * skip it.
5894                          *
5895                          * In contrast to old kernels, we insert the snapshot's
5896                          * dir item and dir index after it has been created, so
5897                          * we won't find a reference to our own snapshot. We
5898                          * still keep the following code for backward
5899                          * compatibility.
5900                          */
5901                         if (location.type == BTRFS_ROOT_ITEM_KEY &&
5902                             location.objectid == root->root_key.objectid) {
5903                                 over = 0;
5904                                 goto skip;
5905                         }
5906                         over = !dir_emit(ctx, name_ptr, name_len,
5907                                        location.objectid, d_type);
5908
5909 skip:
5910                         if (name_ptr != tmp_name)
5911                                 kfree(name_ptr);
5912
5913                         if (over)
5914                                 goto nopos;
5915                         emitted = true;
5916                         di_len = btrfs_dir_name_len(leaf, di) +
5917                                  btrfs_dir_data_len(leaf, di) + sizeof(*di);
5918                         di_cur += di_len;
5919                         di = (struct btrfs_dir_item *)((char *)di + di_len);
5920                 }
5921 next:
5922                 path->slots[0]++;
5923         }
5924
5925         if (key_type == BTRFS_DIR_INDEX_KEY) {
5926                 if (is_curr)
5927                         ctx->pos++;
5928                 ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list, &emitted);
5929                 if (ret)
5930                         goto nopos;
5931         }
5932
5933         /*
5934          * If we haven't emitted any dir entry, we must not touch ctx->pos as
5935          * it was was set to the termination value in previous call. We assume
5936          * that "." and ".." were emitted if we reach this point and set the
5937          * termination value as well for an empty directory.
5938          */
5939         if (ctx->pos > 2 && !emitted)
5940                 goto nopos;
5941
5942         /* Reached end of directory/root. Bump pos past the last item. */
5943         ctx->pos++;
5944
5945         /*
5946          * Stop new entries from being returned after we return the last
5947          * entry.
5948          *
5949          * New directory entries are assigned a strictly increasing
5950          * offset.  This means that new entries created during readdir
5951          * are *guaranteed* to be seen in the future by that readdir.
5952          * This has broken buggy programs which operate on names as
5953          * they're returned by readdir.  Until we re-use freed offsets
5954          * we have this hack to stop new entries from being returned
5955          * under the assumption that they'll never reach this huge
5956          * offset.
5957          *
5958          * This is being careful not to overflow 32bit loff_t unless the
5959          * last entry requires it because doing so has broken 32bit apps
5960          * in the past.
5961          */
5962         if (key_type == BTRFS_DIR_INDEX_KEY) {
5963                 if (ctx->pos >= INT_MAX)
5964                         ctx->pos = LLONG_MAX;
5965                 else
5966                         ctx->pos = INT_MAX;
5967         }
5968 nopos:
5969         ret = 0;
5970 err:
5971         if (put)
5972                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
5973         btrfs_free_path(path);
5974         return ret;
5975 }
5976
5977 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
5978 {
5979         struct btrfs_root *root = BTRFS_I(inode)->root;
5980         struct btrfs_trans_handle *trans;
5981         int ret = 0;
5982         bool nolock = false;
5983
5984         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5985                 return 0;
5986
5987         if (btrfs_fs_closing(root->fs_info) && btrfs_is_free_space_inode(inode))
5988                 nolock = true;
5989
5990         if (wbc->sync_mode == WB_SYNC_ALL) {
5991                 if (nolock)
5992                         trans = btrfs_join_transaction_nolock(root);
5993                 else
5994                         trans = btrfs_join_transaction(root);
5995                 if (IS_ERR(trans))
5996                         return PTR_ERR(trans);
5997                 ret = btrfs_commit_transaction(trans, root);
5998         }
5999         return ret;
6000 }
6001
6002 /*
6003  * This is somewhat expensive, updating the tree every time the
6004  * inode changes.  But, it is most likely to find the inode in cache.
6005  * FIXME, needs more benchmarking...there are no reasons other than performance
6006  * to keep or drop this code.
6007  */
6008 static int btrfs_dirty_inode(struct inode *inode)
6009 {
6010         struct btrfs_root *root = BTRFS_I(inode)->root;
6011         struct btrfs_trans_handle *trans;
6012         int ret;
6013
6014         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6015                 return 0;
6016
6017         trans = btrfs_join_transaction(root);
6018         if (IS_ERR(trans))
6019                 return PTR_ERR(trans);
6020
6021         ret = btrfs_update_inode(trans, root, inode);
6022         if (ret && ret == -ENOSPC) {
6023                 /* whoops, lets try again with the full transaction */
6024                 btrfs_end_transaction(trans, root);
6025                 trans = btrfs_start_transaction(root, 1);
6026                 if (IS_ERR(trans))
6027                         return PTR_ERR(trans);
6028
6029                 ret = btrfs_update_inode(trans, root, inode);
6030         }
6031         btrfs_end_transaction(trans, root);
6032         if (BTRFS_I(inode)->delayed_node)
6033                 btrfs_balance_delayed_items(root);
6034
6035         return ret;
6036 }
6037
6038 /*
6039  * This is a copy of file_update_time.  We need this so we can return error on
6040  * ENOSPC for updating the inode in the case of file write and mmap writes.
6041  */
6042 static int btrfs_update_time(struct inode *inode, struct timespec *now,
6043                              int flags)
6044 {
6045         struct btrfs_root *root = BTRFS_I(inode)->root;
6046
6047         if (btrfs_root_readonly(root))
6048                 return -EROFS;
6049
6050         if (flags & S_VERSION)
6051                 inode_inc_iversion(inode);
6052         if (flags & S_CTIME)
6053                 inode->i_ctime = *now;
6054         if (flags & S_MTIME)
6055                 inode->i_mtime = *now;
6056         if (flags & S_ATIME)
6057                 inode->i_atime = *now;
6058         return btrfs_dirty_inode(inode);
6059 }
6060
6061 /*
6062  * find the highest existing sequence number in a directory
6063  * and then set the in-memory index_cnt variable to reflect
6064  * free sequence numbers
6065  */
6066 static int btrfs_set_inode_index_count(struct inode *inode)
6067 {
6068         struct btrfs_root *root = BTRFS_I(inode)->root;
6069         struct btrfs_key key, found_key;
6070         struct btrfs_path *path;
6071         struct extent_buffer *leaf;
6072         int ret;
6073
6074         key.objectid = btrfs_ino(inode);
6075         key.type = BTRFS_DIR_INDEX_KEY;
6076         key.offset = (u64)-1;
6077
6078         path = btrfs_alloc_path();
6079         if (!path)
6080                 return -ENOMEM;
6081
6082         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6083         if (ret < 0)
6084                 goto out;
6085         /* FIXME: we should be able to handle this */
6086         if (ret == 0)
6087                 goto out;
6088         ret = 0;
6089
6090         /*
6091          * MAGIC NUMBER EXPLANATION:
6092          * since we search a directory based on f_pos we have to start at 2
6093          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6094          * else has to start at 2
6095          */
6096         if (path->slots[0] == 0) {
6097                 BTRFS_I(inode)->index_cnt = 2;
6098                 goto out;
6099         }
6100
6101         path->slots[0]--;
6102
6103         leaf = path->nodes[0];
6104         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6105
6106         if (found_key.objectid != btrfs_ino(inode) ||
6107             found_key.type != BTRFS_DIR_INDEX_KEY) {
6108                 BTRFS_I(inode)->index_cnt = 2;
6109                 goto out;
6110         }
6111
6112         BTRFS_I(inode)->index_cnt = found_key.offset + 1;
6113 out:
6114         btrfs_free_path(path);
6115         return ret;
6116 }
6117
6118 /*
6119  * helper to find a free sequence number in a given directory.  This current
6120  * code is very simple, later versions will do smarter things in the btree
6121  */
6122 int btrfs_set_inode_index(struct inode *dir, u64 *index)
6123 {
6124         int ret = 0;
6125
6126         if (BTRFS_I(dir)->index_cnt == (u64)-1) {
6127                 ret = btrfs_inode_delayed_dir_index_count(dir);
6128                 if (ret) {
6129                         ret = btrfs_set_inode_index_count(dir);
6130                         if (ret)
6131                                 return ret;
6132                 }
6133         }
6134
6135         *index = BTRFS_I(dir)->index_cnt;
6136         BTRFS_I(dir)->index_cnt++;
6137
6138         return ret;
6139 }
6140
6141 static int btrfs_insert_inode_locked(struct inode *inode)
6142 {
6143         struct btrfs_iget_args args;
6144         args.location = &BTRFS_I(inode)->location;
6145         args.root = BTRFS_I(inode)->root;
6146
6147         return insert_inode_locked4(inode,
6148                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6149                    btrfs_find_actor, &args);
6150 }
6151
6152 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6153                                      struct btrfs_root *root,
6154                                      struct inode *dir,
6155                                      const char *name, int name_len,
6156                                      u64 ref_objectid, u64 objectid,
6157                                      umode_t mode, u64 *index)
6158 {
6159         struct inode *inode;
6160         struct btrfs_inode_item *inode_item;
6161         struct btrfs_key *location;
6162         struct btrfs_path *path;
6163         struct btrfs_inode_ref *ref;
6164         struct btrfs_key key[2];
6165         u32 sizes[2];
6166         int nitems = name ? 2 : 1;
6167         unsigned long ptr;
6168         int ret;
6169
6170         path = btrfs_alloc_path();
6171         if (!path)
6172                 return ERR_PTR(-ENOMEM);
6173
6174         inode = new_inode(root->fs_info->sb);
6175         if (!inode) {
6176                 btrfs_free_path(path);
6177                 return ERR_PTR(-ENOMEM);
6178         }
6179
6180         /*
6181          * O_TMPFILE, set link count to 0, so that after this point,
6182          * we fill in an inode item with the correct link count.
6183          */
6184         if (!name)
6185                 set_nlink(inode, 0);
6186
6187         /*
6188          * we have to initialize this early, so we can reclaim the inode
6189          * number if we fail afterwards in this function.
6190          */
6191         inode->i_ino = objectid;
6192
6193         if (dir && name) {
6194                 trace_btrfs_inode_request(dir);
6195
6196                 ret = btrfs_set_inode_index(dir, index);
6197                 if (ret) {
6198                         btrfs_free_path(path);
6199                         iput(inode);
6200                         return ERR_PTR(ret);
6201                 }
6202         } else if (dir) {
6203                 *index = 0;
6204         }
6205         /*
6206          * index_cnt is ignored for everything but a dir,
6207          * btrfs_get_inode_index_count has an explanation for the magic
6208          * number
6209          */
6210         BTRFS_I(inode)->index_cnt = 2;
6211         BTRFS_I(inode)->dir_index = *index;
6212         BTRFS_I(inode)->root = root;
6213         BTRFS_I(inode)->generation = trans->transid;
6214         inode->i_generation = BTRFS_I(inode)->generation;
6215
6216         /*
6217          * We could have gotten an inode number from somebody who was fsynced
6218          * and then removed in this same transaction, so let's just set full
6219          * sync since it will be a full sync anyway and this will blow away the
6220          * old info in the log.
6221          */
6222         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6223
6224         key[0].objectid = objectid;
6225         key[0].type = BTRFS_INODE_ITEM_KEY;
6226         key[0].offset = 0;
6227
6228         sizes[0] = sizeof(struct btrfs_inode_item);
6229
6230         if (name) {
6231                 /*
6232                  * Start new inodes with an inode_ref. This is slightly more
6233                  * efficient for small numbers of hard links since they will
6234                  * be packed into one item. Extended refs will kick in if we
6235                  * add more hard links than can fit in the ref item.
6236                  */
6237                 key[1].objectid = objectid;
6238                 key[1].type = BTRFS_INODE_REF_KEY;
6239                 key[1].offset = ref_objectid;
6240
6241                 sizes[1] = name_len + sizeof(*ref);
6242         }
6243
6244         location = &BTRFS_I(inode)->location;
6245         location->objectid = objectid;
6246         location->offset = 0;
6247         location->type = BTRFS_INODE_ITEM_KEY;
6248
6249         ret = btrfs_insert_inode_locked(inode);
6250         if (ret < 0)
6251                 goto fail;
6252
6253         path->leave_spinning = 1;
6254         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6255         if (ret != 0)
6256                 goto fail_unlock;
6257
6258         inode_init_owner(inode, dir, mode);
6259         inode_set_bytes(inode, 0);
6260
6261         inode->i_mtime = current_fs_time(inode->i_sb);
6262         inode->i_atime = inode->i_mtime;
6263         inode->i_ctime = inode->i_mtime;
6264         BTRFS_I(inode)->i_otime = inode->i_mtime;
6265
6266         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6267                                   struct btrfs_inode_item);
6268         memset_extent_buffer(path->nodes[0], 0, (unsigned long)inode_item,
6269                              sizeof(*inode_item));
6270         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6271
6272         if (name) {
6273                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6274                                      struct btrfs_inode_ref);
6275                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6276                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6277                 ptr = (unsigned long)(ref + 1);
6278                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6279         }
6280
6281         btrfs_mark_buffer_dirty(path->nodes[0]);
6282         btrfs_free_path(path);
6283
6284         btrfs_inherit_iflags(inode, dir);
6285
6286         if (S_ISREG(mode)) {
6287                 if (btrfs_test_opt(root->fs_info, NODATASUM))
6288                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6289                 if (btrfs_test_opt(root->fs_info, NODATACOW))
6290                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6291                                 BTRFS_INODE_NODATASUM;
6292         }
6293
6294         inode_tree_add(inode);
6295
6296         trace_btrfs_inode_new(inode);
6297         btrfs_set_inode_last_trans(trans, inode);
6298
6299         btrfs_update_root_times(trans, root);
6300
6301         ret = btrfs_inode_inherit_props(trans, inode, dir);
6302         if (ret)
6303                 btrfs_err(root->fs_info,
6304                           "error inheriting props for ino %llu (root %llu): %d",
6305                           btrfs_ino(inode), root->root_key.objectid, ret);
6306
6307         return inode;
6308
6309 fail_unlock:
6310         unlock_new_inode(inode);
6311 fail:
6312         if (dir && name)
6313                 BTRFS_I(dir)->index_cnt--;
6314         btrfs_free_path(path);
6315         iput(inode);
6316         return ERR_PTR(ret);
6317 }
6318
6319 static inline u8 btrfs_inode_type(struct inode *inode)
6320 {
6321         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6322 }
6323
6324 /*
6325  * utility function to add 'inode' into 'parent_inode' with
6326  * a give name and a given sequence number.
6327  * if 'add_backref' is true, also insert a backref from the
6328  * inode to the parent directory.
6329  */
6330 int btrfs_add_link(struct btrfs_trans_handle *trans,
6331                    struct inode *parent_inode, struct inode *inode,
6332                    const char *name, int name_len, int add_backref, u64 index)
6333 {
6334         int ret = 0;
6335         struct btrfs_key key;
6336         struct btrfs_root *root = BTRFS_I(parent_inode)->root;
6337         u64 ino = btrfs_ino(inode);
6338         u64 parent_ino = btrfs_ino(parent_inode);
6339
6340         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6341                 memcpy(&key, &BTRFS_I(inode)->root->root_key, sizeof(key));
6342         } else {
6343                 key.objectid = ino;
6344                 key.type = BTRFS_INODE_ITEM_KEY;
6345                 key.offset = 0;
6346         }
6347
6348         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6349                 ret = btrfs_add_root_ref(trans, root->fs_info->tree_root,
6350                                          key.objectid, root->root_key.objectid,
6351                                          parent_ino, index, name, name_len);
6352         } else if (add_backref) {
6353                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6354                                              parent_ino, index);
6355         }
6356
6357         /* Nothing to clean up yet */
6358         if (ret)
6359                 return ret;
6360
6361         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6362                                     parent_inode, &key,
6363                                     btrfs_inode_type(inode), index);
6364         if (ret == -EEXIST || ret == -EOVERFLOW)
6365                 goto fail_dir_item;
6366         else if (ret) {
6367                 btrfs_abort_transaction(trans, ret);
6368                 return ret;
6369         }
6370
6371         btrfs_i_size_write(parent_inode, parent_inode->i_size +
6372                            name_len * 2);
6373         inode_inc_iversion(parent_inode);
6374         parent_inode->i_mtime = parent_inode->i_ctime =
6375                 current_fs_time(parent_inode->i_sb);
6376         ret = btrfs_update_inode(trans, root, parent_inode);
6377         if (ret)
6378                 btrfs_abort_transaction(trans, ret);
6379         return ret;
6380
6381 fail_dir_item:
6382         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6383                 u64 local_index;
6384                 int err;
6385                 err = btrfs_del_root_ref(trans, root->fs_info->tree_root,
6386                                  key.objectid, root->root_key.objectid,
6387                                  parent_ino, &local_index, name, name_len);
6388
6389         } else if (add_backref) {
6390                 u64 local_index;
6391                 int err;
6392
6393                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6394                                           ino, parent_ino, &local_index);
6395         }
6396         return ret;
6397 }
6398
6399 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6400                             struct inode *dir, struct dentry *dentry,
6401                             struct inode *inode, int backref, u64 index)
6402 {
6403         int err = btrfs_add_link(trans, dir, inode,
6404                                  dentry->d_name.name, dentry->d_name.len,
6405                                  backref, index);
6406         if (err > 0)
6407                 err = -EEXIST;
6408         return err;
6409 }
6410
6411 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6412                         umode_t mode, dev_t rdev)
6413 {
6414         struct btrfs_trans_handle *trans;
6415         struct btrfs_root *root = BTRFS_I(dir)->root;
6416         struct inode *inode = NULL;
6417         int err;
6418         int drop_inode = 0;
6419         u64 objectid;
6420         u64 index = 0;
6421
6422         /*
6423          * 2 for inode item and ref
6424          * 2 for dir items
6425          * 1 for xattr if selinux is on
6426          */
6427         trans = btrfs_start_transaction(root, 5);
6428         if (IS_ERR(trans))
6429                 return PTR_ERR(trans);
6430
6431         err = btrfs_find_free_ino(root, &objectid);
6432         if (err)
6433                 goto out_unlock;
6434
6435         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6436                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6437                                 mode, &index);
6438         if (IS_ERR(inode)) {
6439                 err = PTR_ERR(inode);
6440                 goto out_unlock;
6441         }
6442
6443         /*
6444         * If the active LSM wants to access the inode during
6445         * d_instantiate it needs these. Smack checks to see
6446         * if the filesystem supports xattrs by looking at the
6447         * ops vector.
6448         */
6449         inode->i_op = &btrfs_special_inode_operations;
6450         init_special_inode(inode, inode->i_mode, rdev);
6451
6452         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6453         if (err)
6454                 goto out_unlock_inode;
6455
6456         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6457         if (err) {
6458                 goto out_unlock_inode;
6459         } else {
6460                 btrfs_update_inode(trans, root, inode);
6461                 unlock_new_inode(inode);
6462                 d_instantiate(dentry, inode);
6463         }
6464
6465 out_unlock:
6466         btrfs_end_transaction(trans, root);
6467         btrfs_balance_delayed_items(root);
6468         btrfs_btree_balance_dirty(root);
6469         if (drop_inode) {
6470                 inode_dec_link_count(inode);
6471                 iput(inode);
6472         }
6473         return err;
6474
6475 out_unlock_inode:
6476         drop_inode = 1;
6477         unlock_new_inode(inode);
6478         goto out_unlock;
6479
6480 }
6481
6482 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6483                         umode_t mode, bool excl)
6484 {
6485         struct btrfs_trans_handle *trans;
6486         struct btrfs_root *root = BTRFS_I(dir)->root;
6487         struct inode *inode = NULL;
6488         int drop_inode_on_err = 0;
6489         int err;
6490         u64 objectid;
6491         u64 index = 0;
6492
6493         /*
6494          * 2 for inode item and ref
6495          * 2 for dir items
6496          * 1 for xattr if selinux is on
6497          */
6498         trans = btrfs_start_transaction(root, 5);
6499         if (IS_ERR(trans))
6500                 return PTR_ERR(trans);
6501
6502         err = btrfs_find_free_ino(root, &objectid);
6503         if (err)
6504                 goto out_unlock;
6505
6506         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6507                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6508                                 mode, &index);
6509         if (IS_ERR(inode)) {
6510                 err = PTR_ERR(inode);
6511                 goto out_unlock;
6512         }
6513         drop_inode_on_err = 1;
6514         /*
6515         * If the active LSM wants to access the inode during
6516         * d_instantiate it needs these. Smack checks to see
6517         * if the filesystem supports xattrs by looking at the
6518         * ops vector.
6519         */
6520         inode->i_fop = &btrfs_file_operations;
6521         inode->i_op = &btrfs_file_inode_operations;
6522         inode->i_mapping->a_ops = &btrfs_aops;
6523
6524         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6525         if (err)
6526                 goto out_unlock_inode;
6527
6528         err = btrfs_update_inode(trans, root, inode);
6529         if (err)
6530                 goto out_unlock_inode;
6531
6532         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6533         if (err)
6534                 goto out_unlock_inode;
6535
6536         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6537         unlock_new_inode(inode);
6538         d_instantiate(dentry, inode);
6539
6540 out_unlock:
6541         btrfs_end_transaction(trans, root);
6542         if (err && drop_inode_on_err) {
6543                 inode_dec_link_count(inode);
6544                 iput(inode);
6545         }
6546         btrfs_balance_delayed_items(root);
6547         btrfs_btree_balance_dirty(root);
6548         return err;
6549
6550 out_unlock_inode:
6551         unlock_new_inode(inode);
6552         goto out_unlock;
6553
6554 }
6555
6556 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6557                       struct dentry *dentry)
6558 {
6559         struct btrfs_trans_handle *trans = NULL;
6560         struct btrfs_root *root = BTRFS_I(dir)->root;
6561         struct inode *inode = d_inode(old_dentry);
6562         u64 index;
6563         int err;
6564         int drop_inode = 0;
6565
6566         /* do not allow sys_link's with other subvols of the same device */
6567         if (root->objectid != BTRFS_I(inode)->root->objectid)
6568                 return -EXDEV;
6569
6570         if (inode->i_nlink >= BTRFS_LINK_MAX)
6571                 return -EMLINK;
6572
6573         err = btrfs_set_inode_index(dir, &index);
6574         if (err)
6575                 goto fail;
6576
6577         /*
6578          * 2 items for inode and inode ref
6579          * 2 items for dir items
6580          * 1 item for parent inode
6581          */
6582         trans = btrfs_start_transaction(root, 5);
6583         if (IS_ERR(trans)) {
6584                 err = PTR_ERR(trans);
6585                 trans = NULL;
6586                 goto fail;
6587         }
6588
6589         /* There are several dir indexes for this inode, clear the cache. */
6590         BTRFS_I(inode)->dir_index = 0ULL;
6591         inc_nlink(inode);
6592         inode_inc_iversion(inode);
6593         inode->i_ctime = current_fs_time(inode->i_sb);
6594         ihold(inode);
6595         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6596
6597         err = btrfs_add_nondir(trans, dir, dentry, inode, 1, index);
6598
6599         if (err) {
6600                 drop_inode = 1;
6601         } else {
6602                 struct dentry *parent = dentry->d_parent;
6603                 err = btrfs_update_inode(trans, root, inode);
6604                 if (err)
6605                         goto fail;
6606                 if (inode->i_nlink == 1) {
6607                         /*
6608                          * If new hard link count is 1, it's a file created
6609                          * with open(2) O_TMPFILE flag.
6610                          */
6611                         err = btrfs_orphan_del(trans, inode);
6612                         if (err)
6613                                 goto fail;
6614                 }
6615                 d_instantiate(dentry, inode);
6616                 btrfs_log_new_name(trans, inode, NULL, parent);
6617         }
6618
6619         btrfs_balance_delayed_items(root);
6620 fail:
6621         if (trans)
6622                 btrfs_end_transaction(trans, root);
6623         if (drop_inode) {
6624                 inode_dec_link_count(inode);
6625                 iput(inode);
6626         }
6627         btrfs_btree_balance_dirty(root);
6628         return err;
6629 }
6630
6631 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6632 {
6633         struct inode *inode = NULL;
6634         struct btrfs_trans_handle *trans;
6635         struct btrfs_root *root = BTRFS_I(dir)->root;
6636         int err = 0;
6637         int drop_on_err = 0;
6638         u64 objectid = 0;
6639         u64 index = 0;
6640
6641         /*
6642          * 2 items for inode and ref
6643          * 2 items for dir items
6644          * 1 for xattr if selinux is on
6645          */
6646         trans = btrfs_start_transaction(root, 5);
6647         if (IS_ERR(trans))
6648                 return PTR_ERR(trans);
6649
6650         err = btrfs_find_free_ino(root, &objectid);
6651         if (err)
6652                 goto out_fail;
6653
6654         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6655                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6656                                 S_IFDIR | mode, &index);
6657         if (IS_ERR(inode)) {
6658                 err = PTR_ERR(inode);
6659                 goto out_fail;
6660         }
6661
6662         drop_on_err = 1;
6663         /* these must be set before we unlock the inode */
6664         inode->i_op = &btrfs_dir_inode_operations;
6665         inode->i_fop = &btrfs_dir_file_operations;
6666
6667         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6668         if (err)
6669                 goto out_fail_inode;
6670
6671         btrfs_i_size_write(inode, 0);
6672         err = btrfs_update_inode(trans, root, inode);
6673         if (err)
6674                 goto out_fail_inode;
6675
6676         err = btrfs_add_link(trans, dir, inode, dentry->d_name.name,
6677                              dentry->d_name.len, 0, index);
6678         if (err)
6679                 goto out_fail_inode;
6680
6681         d_instantiate(dentry, inode);
6682         /*
6683          * mkdir is special.  We're unlocking after we call d_instantiate
6684          * to avoid a race with nfsd calling d_instantiate.
6685          */
6686         unlock_new_inode(inode);
6687         drop_on_err = 0;
6688
6689 out_fail:
6690         btrfs_end_transaction(trans, root);
6691         if (drop_on_err) {
6692                 inode_dec_link_count(inode);
6693                 iput(inode);
6694         }
6695         btrfs_balance_delayed_items(root);
6696         btrfs_btree_balance_dirty(root);
6697         return err;
6698
6699 out_fail_inode:
6700         unlock_new_inode(inode);
6701         goto out_fail;
6702 }
6703
6704 /* Find next extent map of a given extent map, caller needs to ensure locks */
6705 static struct extent_map *next_extent_map(struct extent_map *em)
6706 {
6707         struct rb_node *next;
6708
6709         next = rb_next(&em->rb_node);
6710         if (!next)
6711                 return NULL;
6712         return container_of(next, struct extent_map, rb_node);
6713 }
6714
6715 static struct extent_map *prev_extent_map(struct extent_map *em)
6716 {
6717         struct rb_node *prev;
6718
6719         prev = rb_prev(&em->rb_node);
6720         if (!prev)
6721                 return NULL;
6722         return container_of(prev, struct extent_map, rb_node);
6723 }
6724
6725 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6726  * the existing extent is the nearest extent to map_start,
6727  * and an extent that you want to insert, deal with overlap and insert
6728  * the best fitted new extent into the tree.
6729  */
6730 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6731                                 struct extent_map *existing,
6732                                 struct extent_map *em,
6733                                 u64 map_start)
6734 {
6735         struct extent_map *prev;
6736         struct extent_map *next;
6737         u64 start;
6738         u64 end;
6739         u64 start_diff;
6740
6741         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6742
6743         if (existing->start > map_start) {
6744                 next = existing;
6745                 prev = prev_extent_map(next);
6746         } else {
6747                 prev = existing;
6748                 next = next_extent_map(prev);
6749         }
6750
6751         start = prev ? extent_map_end(prev) : em->start;
6752         start = max_t(u64, start, em->start);
6753         end = next ? next->start : extent_map_end(em);
6754         end = min_t(u64, end, extent_map_end(em));
6755         start_diff = start - em->start;
6756         em->start = start;
6757         em->len = end - start;
6758         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6759             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6760                 em->block_start += start_diff;
6761                 em->block_len -= start_diff;
6762         }
6763         return add_extent_mapping(em_tree, em, 0);
6764 }
6765
6766 static noinline int uncompress_inline(struct btrfs_path *path,
6767                                       struct page *page,
6768                                       size_t pg_offset, u64 extent_offset,
6769                                       struct btrfs_file_extent_item *item)
6770 {
6771         int ret;
6772         struct extent_buffer *leaf = path->nodes[0];
6773         char *tmp;
6774         size_t max_size;
6775         unsigned long inline_size;
6776         unsigned long ptr;
6777         int compress_type;
6778
6779         WARN_ON(pg_offset != 0);
6780         compress_type = btrfs_file_extent_compression(leaf, item);
6781         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6782         inline_size = btrfs_file_extent_inline_item_len(leaf,
6783                                         btrfs_item_nr(path->slots[0]));
6784         tmp = kmalloc(inline_size, GFP_NOFS);
6785         if (!tmp)
6786                 return -ENOMEM;
6787         ptr = btrfs_file_extent_inline_start(item);
6788
6789         read_extent_buffer(leaf, tmp, ptr, inline_size);
6790
6791         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6792         ret = btrfs_decompress(compress_type, tmp, page,
6793                                extent_offset, inline_size, max_size);
6794         kfree(tmp);
6795         return ret;
6796 }
6797
6798 /*
6799  * a bit scary, this does extent mapping from logical file offset to the disk.
6800  * the ugly parts come from merging extents from the disk with the in-ram
6801  * representation.  This gets more complex because of the data=ordered code,
6802  * where the in-ram extents might be locked pending data=ordered completion.
6803  *
6804  * This also copies inline extents directly into the page.
6805  */
6806
6807 struct extent_map *btrfs_get_extent(struct inode *inode, struct page *page,
6808                                     size_t pg_offset, u64 start, u64 len,
6809                                     int create)
6810 {
6811         int ret;
6812         int err = 0;
6813         u64 extent_start = 0;
6814         u64 extent_end = 0;
6815         u64 objectid = btrfs_ino(inode);
6816         u32 found_type;
6817         struct btrfs_path *path = NULL;
6818         struct btrfs_root *root = BTRFS_I(inode)->root;
6819         struct btrfs_file_extent_item *item;
6820         struct extent_buffer *leaf;
6821         struct btrfs_key found_key;
6822         struct extent_map *em = NULL;
6823         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
6824         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
6825         struct btrfs_trans_handle *trans = NULL;
6826         const bool new_inline = !page || create;
6827
6828 again:
6829         read_lock(&em_tree->lock);
6830         em = lookup_extent_mapping(em_tree, start, len);
6831         if (em)
6832                 em->bdev = root->fs_info->fs_devices->latest_bdev;
6833         read_unlock(&em_tree->lock);
6834
6835         if (em) {
6836                 if (em->start > start || em->start + em->len <= start)
6837                         free_extent_map(em);
6838                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6839                         free_extent_map(em);
6840                 else
6841                         goto out;
6842         }
6843         em = alloc_extent_map();
6844         if (!em) {
6845                 err = -ENOMEM;
6846                 goto out;
6847         }
6848         em->bdev = root->fs_info->fs_devices->latest_bdev;
6849         em->start = EXTENT_MAP_HOLE;
6850         em->orig_start = EXTENT_MAP_HOLE;
6851         em->len = (u64)-1;
6852         em->block_len = (u64)-1;
6853
6854         if (!path) {
6855                 path = btrfs_alloc_path();
6856                 if (!path) {
6857                         err = -ENOMEM;
6858                         goto out;
6859                 }
6860                 /*
6861                  * Chances are we'll be called again, so go ahead and do
6862                  * readahead
6863                  */
6864                 path->reada = READA_FORWARD;
6865         }
6866
6867         ret = btrfs_lookup_file_extent(trans, root, path,
6868                                        objectid, start, trans != NULL);
6869         if (ret < 0) {
6870                 err = ret;
6871                 goto out;
6872         }
6873
6874         if (ret != 0) {
6875                 if (path->slots[0] == 0)
6876                         goto not_found;
6877                 path->slots[0]--;
6878         }
6879
6880         leaf = path->nodes[0];
6881         item = btrfs_item_ptr(leaf, path->slots[0],
6882                               struct btrfs_file_extent_item);
6883         /* are we inside the extent that was found? */
6884         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6885         found_type = found_key.type;
6886         if (found_key.objectid != objectid ||
6887             found_type != BTRFS_EXTENT_DATA_KEY) {
6888                 /*
6889                  * If we backup past the first extent we want to move forward
6890                  * and see if there is an extent in front of us, otherwise we'll
6891                  * say there is a hole for our whole search range which can
6892                  * cause problems.
6893                  */
6894                 extent_end = start;
6895                 goto next;
6896         }
6897
6898         found_type = btrfs_file_extent_type(leaf, item);
6899         extent_start = found_key.offset;
6900         if (found_type == BTRFS_FILE_EXTENT_REG ||
6901             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6902                 extent_end = extent_start +
6903                        btrfs_file_extent_num_bytes(leaf, item);
6904         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6905                 size_t size;
6906                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6907                 extent_end = ALIGN(extent_start + size, root->sectorsize);
6908         }
6909 next:
6910         if (start >= extent_end) {
6911                 path->slots[0]++;
6912                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6913                         ret = btrfs_next_leaf(root, path);
6914                         if (ret < 0) {
6915                                 err = ret;
6916                                 goto out;
6917                         }
6918                         if (ret > 0)
6919                                 goto not_found;
6920                         leaf = path->nodes[0];
6921                 }
6922                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6923                 if (found_key.objectid != objectid ||
6924                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6925                         goto not_found;
6926                 if (start + len <= found_key.offset)
6927                         goto not_found;
6928                 if (start > found_key.offset)
6929                         goto next;
6930                 em->start = start;
6931                 em->orig_start = start;
6932                 em->len = found_key.offset - start;
6933                 goto not_found_em;
6934         }
6935
6936         btrfs_extent_item_to_extent_map(inode, path, item, new_inline, em);
6937
6938         if (found_type == BTRFS_FILE_EXTENT_REG ||
6939             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6940                 goto insert;
6941         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6942                 unsigned long ptr;
6943                 char *map;
6944                 size_t size;
6945                 size_t extent_offset;
6946                 size_t copy_size;
6947
6948                 if (new_inline)
6949                         goto out;
6950
6951                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6952                 extent_offset = page_offset(page) + pg_offset - extent_start;
6953                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
6954                                   size - extent_offset);
6955                 em->start = extent_start + extent_offset;
6956                 em->len = ALIGN(copy_size, root->sectorsize);
6957                 em->orig_block_len = em->len;
6958                 em->orig_start = em->start;
6959                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6960                 if (create == 0 && !PageUptodate(page)) {
6961                         if (btrfs_file_extent_compression(leaf, item) !=
6962                             BTRFS_COMPRESS_NONE) {
6963                                 ret = uncompress_inline(path, page, pg_offset,
6964                                                         extent_offset, item);
6965                                 if (ret) {
6966                                         err = ret;
6967                                         goto out;
6968                                 }
6969                         } else {
6970                                 map = kmap(page);
6971                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6972                                                    copy_size);
6973                                 if (pg_offset + copy_size < PAGE_SIZE) {
6974                                         memset(map + pg_offset + copy_size, 0,
6975                                                PAGE_SIZE - pg_offset -
6976                                                copy_size);
6977                                 }
6978                                 kunmap(page);
6979                         }
6980                         flush_dcache_page(page);
6981                 } else if (create && PageUptodate(page)) {
6982                         BUG();
6983                         if (!trans) {
6984                                 kunmap(page);
6985                                 free_extent_map(em);
6986                                 em = NULL;
6987
6988                                 btrfs_release_path(path);
6989                                 trans = btrfs_join_transaction(root);
6990
6991                                 if (IS_ERR(trans))
6992                                         return ERR_CAST(trans);
6993                                 goto again;
6994                         }
6995                         map = kmap(page);
6996                         write_extent_buffer(leaf, map + pg_offset, ptr,
6997                                             copy_size);
6998                         kunmap(page);
6999                         btrfs_mark_buffer_dirty(leaf);
7000                 }
7001                 set_extent_uptodate(io_tree, em->start,
7002                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
7003                 goto insert;
7004         }
7005 not_found:
7006         em->start = start;
7007         em->orig_start = start;
7008         em->len = len;
7009 not_found_em:
7010         em->block_start = EXTENT_MAP_HOLE;
7011         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
7012 insert:
7013         btrfs_release_path(path);
7014         if (em->start > start || extent_map_end(em) <= start) {
7015                 btrfs_err(root->fs_info, "bad extent! em: [%llu %llu] passed [%llu %llu]",
7016                         em->start, em->len, start, len);
7017                 err = -EIO;
7018                 goto out;
7019         }
7020
7021         err = 0;
7022         write_lock(&em_tree->lock);
7023         ret = add_extent_mapping(em_tree, em, 0);
7024         /* it is possible that someone inserted the extent into the tree
7025          * while we had the lock dropped.  It is also possible that
7026          * an overlapping map exists in the tree
7027          */
7028         if (ret == -EEXIST) {
7029                 struct extent_map *existing;
7030
7031                 ret = 0;
7032
7033                 existing = search_extent_mapping(em_tree, start, len);
7034                 /*
7035                  * existing will always be non-NULL, since there must be
7036                  * extent causing the -EEXIST.
7037                  */
7038                 if (existing->start == em->start &&
7039                     extent_map_end(existing) == extent_map_end(em) &&
7040                     em->block_start == existing->block_start) {
7041                         /*
7042                          * these two extents are the same, it happens
7043                          * with inlines especially
7044                          */
7045                         free_extent_map(em);
7046                         em = existing;
7047                         err = 0;
7048
7049                 } else if (start >= extent_map_end(existing) ||
7050                     start <= existing->start) {
7051                         /*
7052                          * The existing extent map is the one nearest to
7053                          * the [start, start + len) range which overlaps
7054                          */
7055                         err = merge_extent_mapping(em_tree, existing,
7056                                                    em, start);
7057                         free_extent_map(existing);
7058                         if (err) {
7059                                 free_extent_map(em);
7060                                 em = NULL;
7061                         }
7062                 } else {
7063                         free_extent_map(em);
7064                         em = existing;
7065                         err = 0;
7066                 }
7067         }
7068         write_unlock(&em_tree->lock);
7069 out:
7070
7071         trace_btrfs_get_extent(root, em);
7072
7073         btrfs_free_path(path);
7074         if (trans) {
7075                 ret = btrfs_end_transaction(trans, root);
7076                 if (!err)
7077                         err = ret;
7078         }
7079         if (err) {
7080                 free_extent_map(em);
7081                 return ERR_PTR(err);
7082         }
7083         BUG_ON(!em); /* Error is always set */
7084         return em;
7085 }
7086
7087 struct extent_map *btrfs_get_extent_fiemap(struct inode *inode, struct page *page,
7088                                            size_t pg_offset, u64 start, u64 len,
7089                                            int create)
7090 {
7091         struct extent_map *em;
7092         struct extent_map *hole_em = NULL;
7093         u64 range_start = start;
7094         u64 end;
7095         u64 found;
7096         u64 found_end;
7097         int err = 0;
7098
7099         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7100         if (IS_ERR(em))
7101                 return em;
7102         if (em) {
7103                 /*
7104                  * if our em maps to
7105                  * -  a hole or
7106                  * -  a pre-alloc extent,
7107                  * there might actually be delalloc bytes behind it.
7108                  */
7109                 if (em->block_start != EXTENT_MAP_HOLE &&
7110                     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7111                         return em;
7112                 else
7113                         hole_em = em;
7114         }
7115
7116         /* check to see if we've wrapped (len == -1 or similar) */
7117         end = start + len;
7118         if (end < start)
7119                 end = (u64)-1;
7120         else
7121                 end -= 1;
7122
7123         em = NULL;
7124
7125         /* ok, we didn't find anything, lets look for delalloc */
7126         found = count_range_bits(&BTRFS_I(inode)->io_tree, &range_start,
7127                                  end, len, EXTENT_DELALLOC, 1);
7128         found_end = range_start + found;
7129         if (found_end < range_start)
7130                 found_end = (u64)-1;
7131
7132         /*
7133          * we didn't find anything useful, return
7134          * the original results from get_extent()
7135          */
7136         if (range_start > end || found_end <= start) {
7137                 em = hole_em;
7138                 hole_em = NULL;
7139                 goto out;
7140         }
7141
7142         /* adjust the range_start to make sure it doesn't
7143          * go backwards from the start they passed in
7144          */
7145         range_start = max(start, range_start);
7146         found = found_end - range_start;
7147
7148         if (found > 0) {
7149                 u64 hole_start = start;
7150                 u64 hole_len = len;
7151
7152                 em = alloc_extent_map();
7153                 if (!em) {
7154                         err = -ENOMEM;
7155                         goto out;
7156                 }
7157                 /*
7158                  * when btrfs_get_extent can't find anything it
7159                  * returns one huge hole
7160                  *
7161                  * make sure what it found really fits our range, and
7162                  * adjust to make sure it is based on the start from
7163                  * the caller
7164                  */
7165                 if (hole_em) {
7166                         u64 calc_end = extent_map_end(hole_em);
7167
7168                         if (calc_end <= start || (hole_em->start > end)) {
7169                                 free_extent_map(hole_em);
7170                                 hole_em = NULL;
7171                         } else {
7172                                 hole_start = max(hole_em->start, start);
7173                                 hole_len = calc_end - hole_start;
7174                         }
7175                 }
7176                 em->bdev = NULL;
7177                 if (hole_em && range_start > hole_start) {
7178                         /* our hole starts before our delalloc, so we
7179                          * have to return just the parts of the hole
7180                          * that go until  the delalloc starts
7181                          */
7182                         em->len = min(hole_len,
7183                                       range_start - hole_start);
7184                         em->start = hole_start;
7185                         em->orig_start = hole_start;
7186                         /*
7187                          * don't adjust block start at all,
7188                          * it is fixed at EXTENT_MAP_HOLE
7189                          */
7190                         em->block_start = hole_em->block_start;
7191                         em->block_len = hole_len;
7192                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7193                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7194                 } else {
7195                         em->start = range_start;
7196                         em->len = found;
7197                         em->orig_start = range_start;
7198                         em->block_start = EXTENT_MAP_DELALLOC;
7199                         em->block_len = found;
7200                 }
7201         } else if (hole_em) {
7202                 return hole_em;
7203         }
7204 out:
7205
7206         free_extent_map(hole_em);
7207         if (err) {
7208                 free_extent_map(em);
7209                 return ERR_PTR(err);
7210         }
7211         return em;
7212 }
7213
7214 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7215                                                   const u64 start,
7216                                                   const u64 len,
7217                                                   const u64 orig_start,
7218                                                   const u64 block_start,
7219                                                   const u64 block_len,
7220                                                   const u64 orig_block_len,
7221                                                   const u64 ram_bytes,
7222                                                   const int type)
7223 {
7224         struct extent_map *em = NULL;
7225         int ret;
7226
7227         down_read(&BTRFS_I(inode)->dio_sem);
7228         if (type != BTRFS_ORDERED_NOCOW) {
7229                 em = create_pinned_em(inode, start, len, orig_start,
7230                                       block_start, block_len, orig_block_len,
7231                                       ram_bytes, type);
7232                 if (IS_ERR(em))
7233                         goto out;
7234         }
7235         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7236                                            len, block_len, type);
7237         if (ret) {
7238                 if (em) {
7239                         free_extent_map(em);
7240                         btrfs_drop_extent_cache(inode, start,
7241                                                 start + len - 1, 0);
7242                 }
7243                 em = ERR_PTR(ret);
7244         }
7245  out:
7246         up_read(&BTRFS_I(inode)->dio_sem);
7247
7248         return em;
7249 }
7250
7251 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7252                                                   u64 start, u64 len)
7253 {
7254         struct btrfs_root *root = BTRFS_I(inode)->root;
7255         struct extent_map *em;
7256         struct btrfs_key ins;
7257         u64 alloc_hint;
7258         int ret;
7259
7260         alloc_hint = get_extent_allocation_hint(inode, start, len);
7261         ret = btrfs_reserve_extent(root, len, len, root->sectorsize, 0,
7262                                    alloc_hint, &ins, 1, 1);
7263         if (ret)
7264                 return ERR_PTR(ret);
7265
7266         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7267                                      ins.objectid, ins.offset, ins.offset,
7268                                      ins.offset, 0);
7269         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
7270         if (IS_ERR(em))
7271                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7272
7273         return em;
7274 }
7275
7276 /*
7277  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7278  * block must be cow'd
7279  */
7280 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7281                               u64 *orig_start, u64 *orig_block_len,
7282                               u64 *ram_bytes)
7283 {
7284         struct btrfs_trans_handle *trans;
7285         struct btrfs_path *path;
7286         int ret;
7287         struct extent_buffer *leaf;
7288         struct btrfs_root *root = BTRFS_I(inode)->root;
7289         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7290         struct btrfs_file_extent_item *fi;
7291         struct btrfs_key key;
7292         u64 disk_bytenr;
7293         u64 backref_offset;
7294         u64 extent_end;
7295         u64 num_bytes;
7296         int slot;
7297         int found_type;
7298         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7299
7300         path = btrfs_alloc_path();
7301         if (!path)
7302                 return -ENOMEM;
7303
7304         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode),
7305                                        offset, 0);
7306         if (ret < 0)
7307                 goto out;
7308
7309         slot = path->slots[0];
7310         if (ret == 1) {
7311                 if (slot == 0) {
7312                         /* can't find the item, must cow */
7313                         ret = 0;
7314                         goto out;
7315                 }
7316                 slot--;
7317         }
7318         ret = 0;
7319         leaf = path->nodes[0];
7320         btrfs_item_key_to_cpu(leaf, &key, slot);
7321         if (key.objectid != btrfs_ino(inode) ||
7322             key.type != BTRFS_EXTENT_DATA_KEY) {
7323                 /* not our file or wrong item type, must cow */
7324                 goto out;
7325         }
7326
7327         if (key.offset > offset) {
7328                 /* Wrong offset, must cow */
7329                 goto out;
7330         }
7331
7332         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7333         found_type = btrfs_file_extent_type(leaf, fi);
7334         if (found_type != BTRFS_FILE_EXTENT_REG &&
7335             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7336                 /* not a regular extent, must cow */
7337                 goto out;
7338         }
7339
7340         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7341                 goto out;
7342
7343         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7344         if (extent_end <= offset)
7345                 goto out;
7346
7347         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7348         if (disk_bytenr == 0)
7349                 goto out;
7350
7351         if (btrfs_file_extent_compression(leaf, fi) ||
7352             btrfs_file_extent_encryption(leaf, fi) ||
7353             btrfs_file_extent_other_encoding(leaf, fi))
7354                 goto out;
7355
7356         backref_offset = btrfs_file_extent_offset(leaf, fi);
7357
7358         if (orig_start) {
7359                 *orig_start = key.offset - backref_offset;
7360                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7361                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7362         }
7363
7364         if (btrfs_extent_readonly(root, disk_bytenr))
7365                 goto out;
7366
7367         num_bytes = min(offset + *len, extent_end) - offset;
7368         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7369                 u64 range_end;
7370
7371                 range_end = round_up(offset + num_bytes, root->sectorsize) - 1;
7372                 ret = test_range_bit(io_tree, offset, range_end,
7373                                      EXTENT_DELALLOC, 0, NULL);
7374                 if (ret) {
7375                         ret = -EAGAIN;
7376                         goto out;
7377                 }
7378         }
7379
7380         btrfs_release_path(path);
7381
7382         /*
7383          * look for other files referencing this extent, if we
7384          * find any we must cow
7385          */
7386         trans = btrfs_join_transaction(root);
7387         if (IS_ERR(trans)) {
7388                 ret = 0;
7389                 goto out;
7390         }
7391
7392         ret = btrfs_cross_ref_exist(trans, root, btrfs_ino(inode),
7393                                     key.offset - backref_offset, disk_bytenr);
7394         btrfs_end_transaction(trans, root);
7395         if (ret) {
7396                 ret = 0;
7397                 goto out;
7398         }
7399
7400         /*
7401          * adjust disk_bytenr and num_bytes to cover just the bytes
7402          * in this extent we are about to write.  If there
7403          * are any csums in that range we have to cow in order
7404          * to keep the csums correct
7405          */
7406         disk_bytenr += backref_offset;
7407         disk_bytenr += offset - key.offset;
7408         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
7409                                 goto out;
7410         /*
7411          * all of the above have passed, it is safe to overwrite this extent
7412          * without cow
7413          */
7414         *len = num_bytes;
7415         ret = 1;
7416 out:
7417         btrfs_free_path(path);
7418         return ret;
7419 }
7420
7421 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7422 {
7423         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7424         int found = false;
7425         void **pagep = NULL;
7426         struct page *page = NULL;
7427         int start_idx;
7428         int end_idx;
7429
7430         start_idx = start >> PAGE_SHIFT;
7431
7432         /*
7433          * end is the last byte in the last page.  end == start is legal
7434          */
7435         end_idx = end >> PAGE_SHIFT;
7436
7437         rcu_read_lock();
7438
7439         /* Most of the code in this while loop is lifted from
7440          * find_get_page.  It's been modified to begin searching from a
7441          * page and return just the first page found in that range.  If the
7442          * found idx is less than or equal to the end idx then we know that
7443          * a page exists.  If no pages are found or if those pages are
7444          * outside of the range then we're fine (yay!) */
7445         while (page == NULL &&
7446                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7447                 page = radix_tree_deref_slot(pagep);
7448                 if (unlikely(!page))
7449                         break;
7450
7451                 if (radix_tree_exception(page)) {
7452                         if (radix_tree_deref_retry(page)) {
7453                                 page = NULL;
7454                                 continue;
7455                         }
7456                         /*
7457                          * Otherwise, shmem/tmpfs must be storing a swap entry
7458                          * here as an exceptional entry: so return it without
7459                          * attempting to raise page count.
7460                          */
7461                         page = NULL;
7462                         break; /* TODO: Is this relevant for this use case? */
7463                 }
7464
7465                 if (!page_cache_get_speculative(page)) {
7466                         page = NULL;
7467                         continue;
7468                 }
7469
7470                 /*
7471                  * Has the page moved?
7472                  * This is part of the lockless pagecache protocol. See
7473                  * include/linux/pagemap.h for details.
7474                  */
7475                 if (unlikely(page != *pagep)) {
7476                         put_page(page);
7477                         page = NULL;
7478                 }
7479         }
7480
7481         if (page) {
7482                 if (page->index <= end_idx)
7483                         found = true;
7484                 put_page(page);
7485         }
7486
7487         rcu_read_unlock();
7488         return found;
7489 }
7490
7491 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7492                               struct extent_state **cached_state, int writing)
7493 {
7494         struct btrfs_ordered_extent *ordered;
7495         int ret = 0;
7496
7497         while (1) {
7498                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7499                                  cached_state);
7500                 /*
7501                  * We're concerned with the entire range that we're going to be
7502                  * doing DIO to, so we need to make sure there's no ordered
7503                  * extents in this range.
7504                  */
7505                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
7506                                                      lockend - lockstart + 1);
7507
7508                 /*
7509                  * We need to make sure there are no buffered pages in this
7510                  * range either, we could have raced between the invalidate in
7511                  * generic_file_direct_write and locking the extent.  The
7512                  * invalidate needs to happen so that reads after a write do not
7513                  * get stale data.
7514                  */
7515                 if (!ordered &&
7516                     (!writing ||
7517                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7518                         break;
7519
7520                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7521                                      cached_state, GFP_NOFS);
7522
7523                 if (ordered) {
7524                         /*
7525                          * If we are doing a DIO read and the ordered extent we
7526                          * found is for a buffered write, we can not wait for it
7527                          * to complete and retry, because if we do so we can
7528                          * deadlock with concurrent buffered writes on page
7529                          * locks. This happens only if our DIO read covers more
7530                          * than one extent map, if at this point has already
7531                          * created an ordered extent for a previous extent map
7532                          * and locked its range in the inode's io tree, and a
7533                          * concurrent write against that previous extent map's
7534                          * range and this range started (we unlock the ranges
7535                          * in the io tree only when the bios complete and
7536                          * buffered writes always lock pages before attempting
7537                          * to lock range in the io tree).
7538                          */
7539                         if (writing ||
7540                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7541                                 btrfs_start_ordered_extent(inode, ordered, 1);
7542                         else
7543                                 ret = -ENOTBLK;
7544                         btrfs_put_ordered_extent(ordered);
7545                 } else {
7546                         /*
7547                          * We could trigger writeback for this range (and wait
7548                          * for it to complete) and then invalidate the pages for
7549                          * this range (through invalidate_inode_pages2_range()),
7550                          * but that can lead us to a deadlock with a concurrent
7551                          * call to readpages() (a buffered read or a defrag call
7552                          * triggered a readahead) on a page lock due to an
7553                          * ordered dio extent we created before but did not have
7554                          * yet a corresponding bio submitted (whence it can not
7555                          * complete), which makes readpages() wait for that
7556                          * ordered extent to complete while holding a lock on
7557                          * that page.
7558                          */
7559                         ret = -ENOTBLK;
7560                 }
7561
7562                 if (ret)
7563                         break;
7564
7565                 cond_resched();
7566         }
7567
7568         return ret;
7569 }
7570
7571 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
7572                                            u64 len, u64 orig_start,
7573                                            u64 block_start, u64 block_len,
7574                                            u64 orig_block_len, u64 ram_bytes,
7575                                            int type)
7576 {
7577         struct extent_map_tree *em_tree;
7578         struct extent_map *em;
7579         struct btrfs_root *root = BTRFS_I(inode)->root;
7580         int ret;
7581
7582         em_tree = &BTRFS_I(inode)->extent_tree;
7583         em = alloc_extent_map();
7584         if (!em)
7585                 return ERR_PTR(-ENOMEM);
7586
7587         em->start = start;
7588         em->orig_start = orig_start;
7589         em->mod_start = start;
7590         em->mod_len = len;
7591         em->len = len;
7592         em->block_len = block_len;
7593         em->block_start = block_start;
7594         em->bdev = root->fs_info->fs_devices->latest_bdev;
7595         em->orig_block_len = orig_block_len;
7596         em->ram_bytes = ram_bytes;
7597         em->generation = -1;
7598         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7599         if (type == BTRFS_ORDERED_PREALLOC)
7600                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7601
7602         do {
7603                 btrfs_drop_extent_cache(inode, em->start,
7604                                 em->start + em->len - 1, 0);
7605                 write_lock(&em_tree->lock);
7606                 ret = add_extent_mapping(em_tree, em, 1);
7607                 write_unlock(&em_tree->lock);
7608         } while (ret == -EEXIST);
7609
7610         if (ret) {
7611                 free_extent_map(em);
7612                 return ERR_PTR(ret);
7613         }
7614
7615         return em;
7616 }
7617
7618 static void adjust_dio_outstanding_extents(struct inode *inode,
7619                                            struct btrfs_dio_data *dio_data,
7620                                            const u64 len)
7621 {
7622         unsigned num_extents;
7623
7624         num_extents = (unsigned) div64_u64(len + BTRFS_MAX_EXTENT_SIZE - 1,
7625                                            BTRFS_MAX_EXTENT_SIZE);
7626         /*
7627          * If we have an outstanding_extents count still set then we're
7628          * within our reservation, otherwise we need to adjust our inode
7629          * counter appropriately.
7630          */
7631         if (dio_data->outstanding_extents) {
7632                 dio_data->outstanding_extents -= num_extents;
7633         } else {
7634                 spin_lock(&BTRFS_I(inode)->lock);
7635                 BTRFS_I(inode)->outstanding_extents += num_extents;
7636                 spin_unlock(&BTRFS_I(inode)->lock);
7637         }
7638 }
7639
7640 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7641                                    struct buffer_head *bh_result, int create)
7642 {
7643         struct extent_map *em;
7644         struct btrfs_root *root = BTRFS_I(inode)->root;
7645         struct extent_state *cached_state = NULL;
7646         struct btrfs_dio_data *dio_data = NULL;
7647         u64 start = iblock << inode->i_blkbits;
7648         u64 lockstart, lockend;
7649         u64 len = bh_result->b_size;
7650         int unlock_bits = EXTENT_LOCKED;
7651         int ret = 0;
7652
7653         if (create)
7654                 unlock_bits |= EXTENT_DIRTY;
7655         else
7656                 len = min_t(u64, len, root->sectorsize);
7657
7658         lockstart = start;
7659         lockend = start + len - 1;
7660
7661         if (current->journal_info) {
7662                 /*
7663                  * Need to pull our outstanding extents and set journal_info to NULL so
7664                  * that anything that needs to check if there's a transaction doesn't get
7665                  * confused.
7666                  */
7667                 dio_data = current->journal_info;
7668                 current->journal_info = NULL;
7669         }
7670
7671         /*
7672          * If this errors out it's because we couldn't invalidate pagecache for
7673          * this range and we need to fallback to buffered.
7674          */
7675         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7676                                create)) {
7677                 ret = -ENOTBLK;
7678                 goto err;
7679         }
7680
7681         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7682         if (IS_ERR(em)) {
7683                 ret = PTR_ERR(em);
7684                 goto unlock_err;
7685         }
7686
7687         /*
7688          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7689          * io.  INLINE is special, and we could probably kludge it in here, but
7690          * it's still buffered so for safety lets just fall back to the generic
7691          * buffered path.
7692          *
7693          * For COMPRESSED we _have_ to read the entire extent in so we can
7694          * decompress it, so there will be buffering required no matter what we
7695          * do, so go ahead and fallback to buffered.
7696          *
7697          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7698          * to buffered IO.  Don't blame me, this is the price we pay for using
7699          * the generic code.
7700          */
7701         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7702             em->block_start == EXTENT_MAP_INLINE) {
7703                 free_extent_map(em);
7704                 ret = -ENOTBLK;
7705                 goto unlock_err;
7706         }
7707
7708         /* Just a good old fashioned hole, return */
7709         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7710                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7711                 free_extent_map(em);
7712                 goto unlock_err;
7713         }
7714
7715         /*
7716          * We don't allocate a new extent in the following cases
7717          *
7718          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7719          * existing extent.
7720          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7721          * just use the extent.
7722          *
7723          */
7724         if (!create) {
7725                 len = min(len, em->len - (start - em->start));
7726                 lockstart = start + len;
7727                 goto unlock;
7728         }
7729
7730         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7731             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7732              em->block_start != EXTENT_MAP_HOLE)) {
7733                 int type;
7734                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7735
7736                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7737                         type = BTRFS_ORDERED_PREALLOC;
7738                 else
7739                         type = BTRFS_ORDERED_NOCOW;
7740                 len = min(len, em->len - (start - em->start));
7741                 block_start = em->block_start + (start - em->start);
7742
7743                 if (can_nocow_extent(inode, start, &len, &orig_start,
7744                                      &orig_block_len, &ram_bytes) == 1 &&
7745                     btrfs_inc_nocow_writers(root->fs_info, block_start)) {
7746                         struct extent_map *em2;
7747
7748                         em2 = btrfs_create_dio_extent(inode, start, len,
7749                                                       orig_start, block_start,
7750                                                       len, orig_block_len,
7751                                                       ram_bytes, type);
7752                         btrfs_dec_nocow_writers(root->fs_info, block_start);
7753                         if (type == BTRFS_ORDERED_PREALLOC) {
7754                                 free_extent_map(em);
7755                                 em = em2;
7756                         }
7757                         if (em2 && IS_ERR(em2)) {
7758                                 ret = PTR_ERR(em2);
7759                                 goto unlock_err;
7760                         }
7761                         /*
7762                          * For inode marked NODATACOW or extent marked PREALLOC,
7763                          * use the existing or preallocated extent, so does not
7764                          * need to adjust btrfs_space_info's bytes_may_use.
7765                          */
7766                         btrfs_free_reserved_data_space_noquota(inode,
7767                                         start, len);
7768                         goto unlock;
7769                 }
7770         }
7771
7772         /*
7773          * this will cow the extent, reset the len in case we changed
7774          * it above
7775          */
7776         len = bh_result->b_size;
7777         free_extent_map(em);
7778         em = btrfs_new_extent_direct(inode, start, len);
7779         if (IS_ERR(em)) {
7780                 ret = PTR_ERR(em);
7781                 goto unlock_err;
7782         }
7783         len = min(len, em->len - (start - em->start));
7784 unlock:
7785         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7786                 inode->i_blkbits;
7787         bh_result->b_size = len;
7788         bh_result->b_bdev = em->bdev;
7789         set_buffer_mapped(bh_result);
7790         if (create) {
7791                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7792                         set_buffer_new(bh_result);
7793
7794                 /*
7795                  * Need to update the i_size under the extent lock so buffered
7796                  * readers will get the updated i_size when we unlock.
7797                  */
7798                 if (start + len > i_size_read(inode))
7799                         i_size_write(inode, start + len);
7800
7801                 adjust_dio_outstanding_extents(inode, dio_data, len);
7802                 WARN_ON(dio_data->reserve < len);
7803                 dio_data->reserve -= len;
7804                 dio_data->unsubmitted_oe_range_end = start + len;
7805                 current->journal_info = dio_data;
7806         }
7807
7808         /*
7809          * In the case of write we need to clear and unlock the entire range,
7810          * in the case of read we need to unlock only the end area that we
7811          * aren't using if there is any left over space.
7812          */
7813         if (lockstart < lockend) {
7814                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7815                                  lockend, unlock_bits, 1, 0,
7816                                  &cached_state, GFP_NOFS);
7817         } else {
7818                 free_extent_state(cached_state);
7819         }
7820
7821         free_extent_map(em);
7822
7823         return 0;
7824
7825 unlock_err:
7826         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7827                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
7828 err:
7829         if (dio_data)
7830                 current->journal_info = dio_data;
7831         /*
7832          * Compensate the delalloc release we do in btrfs_direct_IO() when we
7833          * write less data then expected, so that we don't underflow our inode's
7834          * outstanding extents counter.
7835          */
7836         if (create && dio_data)
7837                 adjust_dio_outstanding_extents(inode, dio_data, len);
7838
7839         return ret;
7840 }
7841
7842 static inline int submit_dio_repair_bio(struct inode *inode, struct bio *bio,
7843                                         int mirror_num)
7844 {
7845         struct btrfs_root *root = BTRFS_I(inode)->root;
7846         int ret;
7847
7848         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7849
7850         bio_get(bio);
7851
7852         ret = btrfs_bio_wq_end_io(root->fs_info, bio,
7853                                   BTRFS_WQ_ENDIO_DIO_REPAIR);
7854         if (ret)
7855                 goto err;
7856
7857         ret = btrfs_map_bio(root, bio, mirror_num, 0);
7858 err:
7859         bio_put(bio);
7860         return ret;
7861 }
7862
7863 static int btrfs_check_dio_repairable(struct inode *inode,
7864                                       struct bio *failed_bio,
7865                                       struct io_failure_record *failrec,
7866                                       int failed_mirror)
7867 {
7868         int num_copies;
7869
7870         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
7871                                       failrec->logical, failrec->len);
7872         if (num_copies == 1) {
7873                 /*
7874                  * we only have a single copy of the data, so don't bother with
7875                  * all the retry and error correction code that follows. no
7876                  * matter what the error is, it is very likely to persist.
7877                  */
7878                 pr_debug("Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
7879                          num_copies, failrec->this_mirror, failed_mirror);
7880                 return 0;
7881         }
7882
7883         failrec->failed_mirror = failed_mirror;
7884         failrec->this_mirror++;
7885         if (failrec->this_mirror == failed_mirror)
7886                 failrec->this_mirror++;
7887
7888         if (failrec->this_mirror > num_copies) {
7889                 pr_debug("Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
7890                          num_copies, failrec->this_mirror, failed_mirror);
7891                 return 0;
7892         }
7893
7894         return 1;
7895 }
7896
7897 static int dio_read_error(struct inode *inode, struct bio *failed_bio,
7898                         struct page *page, unsigned int pgoff,
7899                         u64 start, u64 end, int failed_mirror,
7900                         bio_end_io_t *repair_endio, void *repair_arg)
7901 {
7902         struct io_failure_record *failrec;
7903         struct bio *bio;
7904         int isector;
7905         int read_mode;
7906         int ret;
7907
7908         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
7909
7910         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7911         if (ret)
7912                 return ret;
7913
7914         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7915                                          failed_mirror);
7916         if (!ret) {
7917                 free_io_failure(inode, failrec);
7918                 return -EIO;
7919         }
7920
7921         if ((failed_bio->bi_vcnt > 1)
7922                 || (failed_bio->bi_io_vec->bv_len
7923                         > BTRFS_I(inode)->root->sectorsize))
7924                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
7925         else
7926                 read_mode = READ_SYNC;
7927
7928         isector = start - btrfs_io_bio(failed_bio)->logical;
7929         isector >>= inode->i_sb->s_blocksize_bits;
7930         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7931                                 pgoff, isector, repair_endio, repair_arg);
7932         if (!bio) {
7933                 free_io_failure(inode, failrec);
7934                 return -EIO;
7935         }
7936         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
7937
7938         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7939                     "Repair DIO Read Error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d\n",
7940                     read_mode, failrec->this_mirror, failrec->in_validation);
7941
7942         ret = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
7943         if (ret) {
7944                 free_io_failure(inode, failrec);
7945                 bio_put(bio);
7946         }
7947
7948         return ret;
7949 }
7950
7951 struct btrfs_retry_complete {
7952         struct completion done;
7953         struct inode *inode;
7954         u64 start;
7955         int uptodate;
7956 };
7957
7958 static void btrfs_retry_endio_nocsum(struct bio *bio)
7959 {
7960         struct btrfs_retry_complete *done = bio->bi_private;
7961         struct inode *inode;
7962         struct bio_vec *bvec;
7963         int i;
7964
7965         if (bio->bi_error)
7966                 goto end;
7967
7968         ASSERT(bio->bi_vcnt == 1);
7969         inode = bio->bi_io_vec->bv_page->mapping->host;
7970         ASSERT(bio->bi_io_vec->bv_len == BTRFS_I(inode)->root->sectorsize);
7971
7972         done->uptodate = 1;
7973         bio_for_each_segment_all(bvec, bio, i)
7974                 clean_io_failure(done->inode, done->start, bvec->bv_page, 0);
7975 end:
7976         complete(&done->done);
7977         bio_put(bio);
7978 }
7979
7980 static int __btrfs_correct_data_nocsum(struct inode *inode,
7981                                        struct btrfs_io_bio *io_bio)
7982 {
7983         struct btrfs_fs_info *fs_info;
7984         struct bio_vec *bvec;
7985         struct btrfs_retry_complete done;
7986         u64 start;
7987         unsigned int pgoff;
7988         u32 sectorsize;
7989         int nr_sectors;
7990         int i;
7991         int ret;
7992
7993         fs_info = BTRFS_I(inode)->root->fs_info;
7994         sectorsize = BTRFS_I(inode)->root->sectorsize;
7995
7996         start = io_bio->logical;
7997         done.inode = inode;
7998
7999         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
8000                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
8001                 pgoff = bvec->bv_offset;
8002
8003 next_block_or_try_again:
8004                 done.uptodate = 0;
8005                 done.start = start;
8006                 init_completion(&done.done);
8007
8008                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
8009                                 pgoff, start, start + sectorsize - 1,
8010                                 io_bio->mirror_num,
8011                                 btrfs_retry_endio_nocsum, &done);
8012                 if (ret)
8013                         return ret;
8014
8015                 wait_for_completion(&done.done);
8016
8017                 if (!done.uptodate) {
8018                         /* We might have another mirror, so try again */
8019                         goto next_block_or_try_again;
8020                 }
8021
8022                 start += sectorsize;
8023
8024                 if (nr_sectors--) {
8025                         pgoff += sectorsize;
8026                         goto next_block_or_try_again;
8027                 }
8028         }
8029
8030         return 0;
8031 }
8032
8033 static void btrfs_retry_endio(struct bio *bio)
8034 {
8035         struct btrfs_retry_complete *done = bio->bi_private;
8036         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8037         struct inode *inode;
8038         struct bio_vec *bvec;
8039         u64 start;
8040         int uptodate;
8041         int ret;
8042         int i;
8043
8044         if (bio->bi_error)
8045                 goto end;
8046
8047         uptodate = 1;
8048
8049         start = done->start;
8050
8051         ASSERT(bio->bi_vcnt == 1);
8052         inode = bio->bi_io_vec->bv_page->mapping->host;
8053         ASSERT(bio->bi_io_vec->bv_len == BTRFS_I(inode)->root->sectorsize);
8054
8055         bio_for_each_segment_all(bvec, bio, i) {
8056                 ret = __readpage_endio_check(done->inode, io_bio, i,
8057                                         bvec->bv_page, bvec->bv_offset,
8058                                         done->start, bvec->bv_len);
8059                 if (!ret)
8060                         clean_io_failure(done->inode, done->start,
8061                                         bvec->bv_page, bvec->bv_offset);
8062                 else
8063                         uptodate = 0;
8064         }
8065
8066         done->uptodate = uptodate;
8067 end:
8068         complete(&done->done);
8069         bio_put(bio);
8070 }
8071
8072 static int __btrfs_subio_endio_read(struct inode *inode,
8073                                     struct btrfs_io_bio *io_bio, int err)
8074 {
8075         struct btrfs_fs_info *fs_info;
8076         struct bio_vec *bvec;
8077         struct btrfs_retry_complete done;
8078         u64 start;
8079         u64 offset = 0;
8080         u32 sectorsize;
8081         int nr_sectors;
8082         unsigned int pgoff;
8083         int csum_pos;
8084         int i;
8085         int ret;
8086
8087         fs_info = BTRFS_I(inode)->root->fs_info;
8088         sectorsize = BTRFS_I(inode)->root->sectorsize;
8089
8090         err = 0;
8091         start = io_bio->logical;
8092         done.inode = inode;
8093
8094         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
8095                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
8096
8097                 pgoff = bvec->bv_offset;
8098 next_block:
8099                 csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8100                 ret = __readpage_endio_check(inode, io_bio, csum_pos,
8101                                         bvec->bv_page, pgoff, start,
8102                                         sectorsize);
8103                 if (likely(!ret))
8104                         goto next;
8105 try_again:
8106                 done.uptodate = 0;
8107                 done.start = start;
8108                 init_completion(&done.done);
8109
8110                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
8111                                 pgoff, start, start + sectorsize - 1,
8112                                 io_bio->mirror_num,
8113                                 btrfs_retry_endio, &done);
8114                 if (ret) {
8115                         err = ret;
8116                         goto next;
8117                 }
8118
8119                 wait_for_completion(&done.done);
8120
8121                 if (!done.uptodate) {
8122                         /* We might have another mirror, so try again */
8123                         goto try_again;
8124                 }
8125 next:
8126                 offset += sectorsize;
8127                 start += sectorsize;
8128
8129                 ASSERT(nr_sectors);
8130
8131                 if (--nr_sectors) {
8132                         pgoff += sectorsize;
8133                         goto next_block;
8134                 }
8135         }
8136
8137         return err;
8138 }
8139
8140 static int btrfs_subio_endio_read(struct inode *inode,
8141                                   struct btrfs_io_bio *io_bio, int err)
8142 {
8143         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8144
8145         if (skip_csum) {
8146                 if (unlikely(err))
8147                         return __btrfs_correct_data_nocsum(inode, io_bio);
8148                 else
8149                         return 0;
8150         } else {
8151                 return __btrfs_subio_endio_read(inode, io_bio, err);
8152         }
8153 }
8154
8155 static void btrfs_endio_direct_read(struct bio *bio)
8156 {
8157         struct btrfs_dio_private *dip = bio->bi_private;
8158         struct inode *inode = dip->inode;
8159         struct bio *dio_bio;
8160         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8161         int err = bio->bi_error;
8162
8163         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8164                 err = btrfs_subio_endio_read(inode, io_bio, err);
8165
8166         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8167                       dip->logical_offset + dip->bytes - 1);
8168         dio_bio = dip->dio_bio;
8169
8170         kfree(dip);
8171
8172         dio_bio->bi_error = bio->bi_error;
8173         dio_end_io(dio_bio, bio->bi_error);
8174
8175         if (io_bio->end_io)
8176                 io_bio->end_io(io_bio, err);
8177         bio_put(bio);
8178 }
8179
8180 static void btrfs_endio_direct_write_update_ordered(struct inode *inode,
8181                                                     const u64 offset,
8182                                                     const u64 bytes,
8183                                                     const int uptodate)
8184 {
8185         struct btrfs_root *root = BTRFS_I(inode)->root;
8186         struct btrfs_ordered_extent *ordered = NULL;
8187         u64 ordered_offset = offset;
8188         u64 ordered_bytes = bytes;
8189         int ret;
8190
8191 again:
8192         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8193                                                    &ordered_offset,
8194                                                    ordered_bytes,
8195                                                    uptodate);
8196         if (!ret)
8197                 goto out_test;
8198
8199         btrfs_init_work(&ordered->work, btrfs_endio_write_helper,
8200                         finish_ordered_fn, NULL, NULL);
8201         btrfs_queue_work(root->fs_info->endio_write_workers,
8202                          &ordered->work);
8203 out_test:
8204         /*
8205          * our bio might span multiple ordered extents.  If we haven't
8206          * completed the accounting for the whole dio, go back and try again
8207          */
8208         if (ordered_offset < offset + bytes) {
8209                 ordered_bytes = offset + bytes - ordered_offset;
8210                 ordered = NULL;
8211                 goto again;
8212         }
8213 }
8214
8215 static void btrfs_endio_direct_write(struct bio *bio)
8216 {
8217         struct btrfs_dio_private *dip = bio->bi_private;
8218         struct bio *dio_bio = dip->dio_bio;
8219
8220         btrfs_endio_direct_write_update_ordered(dip->inode,
8221                                                 dip->logical_offset,
8222                                                 dip->bytes,
8223                                                 !bio->bi_error);
8224
8225         kfree(dip);
8226
8227         dio_bio->bi_error = bio->bi_error;
8228         dio_end_io(dio_bio, bio->bi_error);
8229         bio_put(bio);
8230 }
8231
8232 static int __btrfs_submit_bio_start_direct_io(struct inode *inode,
8233                                     struct bio *bio, int mirror_num,
8234                                     unsigned long bio_flags, u64 offset)
8235 {
8236         int ret;
8237         struct btrfs_root *root = BTRFS_I(inode)->root;
8238         ret = btrfs_csum_one_bio(root, inode, bio, offset, 1);
8239         BUG_ON(ret); /* -ENOMEM */
8240         return 0;
8241 }
8242
8243 static void btrfs_end_dio_bio(struct bio *bio)
8244 {
8245         struct btrfs_dio_private *dip = bio->bi_private;
8246         int err = bio->bi_error;
8247
8248         if (err)
8249                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8250                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8251                            btrfs_ino(dip->inode), bio_op(bio), bio->bi_opf,
8252                            (unsigned long long)bio->bi_iter.bi_sector,
8253                            bio->bi_iter.bi_size, err);
8254
8255         if (dip->subio_endio)
8256                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8257
8258         if (err) {
8259                 dip->errors = 1;
8260
8261                 /*
8262                  * before atomic variable goto zero, we must make sure
8263                  * dip->errors is perceived to be set.
8264                  */
8265                 smp_mb__before_atomic();
8266         }
8267
8268         /* if there are more bios still pending for this dio, just exit */
8269         if (!atomic_dec_and_test(&dip->pending_bios))
8270                 goto out;
8271
8272         if (dip->errors) {
8273                 bio_io_error(dip->orig_bio);
8274         } else {
8275                 dip->dio_bio->bi_error = 0;
8276                 bio_endio(dip->orig_bio);
8277         }
8278 out:
8279         bio_put(bio);
8280 }
8281
8282 static struct bio *btrfs_dio_bio_alloc(struct block_device *bdev,
8283                                        u64 first_sector, gfp_t gfp_flags)
8284 {
8285         struct bio *bio;
8286         bio = btrfs_bio_alloc(bdev, first_sector, BIO_MAX_PAGES, gfp_flags);
8287         if (bio)
8288                 bio_associate_current(bio);
8289         return bio;
8290 }
8291
8292 static inline int btrfs_lookup_and_bind_dio_csum(struct btrfs_root *root,
8293                                                  struct inode *inode,
8294                                                  struct btrfs_dio_private *dip,
8295                                                  struct bio *bio,
8296                                                  u64 file_offset)
8297 {
8298         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8299         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8300         int ret;
8301
8302         /*
8303          * We load all the csum data we need when we submit
8304          * the first bio to reduce the csum tree search and
8305          * contention.
8306          */
8307         if (dip->logical_offset == file_offset) {
8308                 ret = btrfs_lookup_bio_sums_dio(root, inode, dip->orig_bio,
8309                                                 file_offset);
8310                 if (ret)
8311                         return ret;
8312         }
8313
8314         if (bio == dip->orig_bio)
8315                 return 0;
8316
8317         file_offset -= dip->logical_offset;
8318         file_offset >>= inode->i_sb->s_blocksize_bits;
8319         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8320
8321         return 0;
8322 }
8323
8324 static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
8325                                          u64 file_offset, int skip_sum,
8326                                          int async_submit)
8327 {
8328         struct btrfs_dio_private *dip = bio->bi_private;
8329         bool write = bio_op(bio) == REQ_OP_WRITE;
8330         struct btrfs_root *root = BTRFS_I(inode)->root;
8331         int ret;
8332
8333         if (async_submit)
8334                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8335
8336         bio_get(bio);
8337
8338         if (!write) {
8339                 ret = btrfs_bio_wq_end_io(root->fs_info, bio,
8340                                 BTRFS_WQ_ENDIO_DATA);
8341                 if (ret)
8342                         goto err;
8343         }
8344
8345         if (skip_sum)
8346                 goto map;
8347
8348         if (write && async_submit) {
8349                 ret = btrfs_wq_submit_bio(root->fs_info,
8350                                    inode, bio, 0, 0, file_offset,
8351                                    __btrfs_submit_bio_start_direct_io,
8352                                    __btrfs_submit_bio_done);
8353                 goto err;
8354         } else if (write) {
8355                 /*
8356                  * If we aren't doing async submit, calculate the csum of the
8357                  * bio now.
8358                  */
8359                 ret = btrfs_csum_one_bio(root, inode, bio, file_offset, 1);
8360                 if (ret)
8361                         goto err;
8362         } else {
8363                 ret = btrfs_lookup_and_bind_dio_csum(root, inode, dip, bio,
8364                                                      file_offset);
8365                 if (ret)
8366                         goto err;
8367         }
8368 map:
8369         ret = btrfs_map_bio(root, bio, 0, async_submit);
8370 err:
8371         bio_put(bio);
8372         return ret;
8373 }
8374
8375 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip,
8376                                     int skip_sum)
8377 {
8378         struct inode *inode = dip->inode;
8379         struct btrfs_root *root = BTRFS_I(inode)->root;
8380         struct bio *bio;
8381         struct bio *orig_bio = dip->orig_bio;
8382         struct bio_vec *bvec = orig_bio->bi_io_vec;
8383         u64 start_sector = orig_bio->bi_iter.bi_sector;
8384         u64 file_offset = dip->logical_offset;
8385         u64 submit_len = 0;
8386         u64 map_length;
8387         u32 blocksize = root->sectorsize;
8388         int async_submit = 0;
8389         int nr_sectors;
8390         int ret;
8391         int i;
8392
8393         map_length = orig_bio->bi_iter.bi_size;
8394         ret = btrfs_map_block(root->fs_info, bio_op(orig_bio),
8395                               start_sector << 9, &map_length, NULL, 0);
8396         if (ret)
8397                 return -EIO;
8398
8399         if (map_length >= orig_bio->bi_iter.bi_size) {
8400                 bio = orig_bio;
8401                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8402                 goto submit;
8403         }
8404
8405         /* async crcs make it difficult to collect full stripe writes. */
8406         if (btrfs_get_alloc_profile(root, 1) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8407                 async_submit = 0;
8408         else
8409                 async_submit = 1;
8410
8411         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev, start_sector, GFP_NOFS);
8412         if (!bio)
8413                 return -ENOMEM;
8414
8415         bio_set_op_attrs(bio, bio_op(orig_bio), bio_flags(orig_bio));
8416         bio->bi_private = dip;
8417         bio->bi_end_io = btrfs_end_dio_bio;
8418         btrfs_io_bio(bio)->logical = file_offset;
8419         atomic_inc(&dip->pending_bios);
8420
8421         while (bvec <= (orig_bio->bi_io_vec + orig_bio->bi_vcnt - 1)) {
8422                 nr_sectors = BTRFS_BYTES_TO_BLKS(root->fs_info, bvec->bv_len);
8423                 i = 0;
8424 next_block:
8425                 if (unlikely(map_length < submit_len + blocksize ||
8426                     bio_add_page(bio, bvec->bv_page, blocksize,
8427                             bvec->bv_offset + (i * blocksize)) < blocksize)) {
8428                         /*
8429                          * inc the count before we submit the bio so
8430                          * we know the end IO handler won't happen before
8431                          * we inc the count. Otherwise, the dip might get freed
8432                          * before we're done setting it up
8433                          */
8434                         atomic_inc(&dip->pending_bios);
8435                         ret = __btrfs_submit_dio_bio(bio, inode,
8436                                                      file_offset, skip_sum,
8437                                                      async_submit);
8438                         if (ret) {
8439                                 bio_put(bio);
8440                                 atomic_dec(&dip->pending_bios);
8441                                 goto out_err;
8442                         }
8443
8444                         start_sector += submit_len >> 9;
8445                         file_offset += submit_len;
8446
8447                         submit_len = 0;
8448
8449                         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev,
8450                                                   start_sector, GFP_NOFS);
8451                         if (!bio)
8452                                 goto out_err;
8453                         bio_set_op_attrs(bio, bio_op(orig_bio),
8454                                          bio_flags(orig_bio));
8455                         bio->bi_private = dip;
8456                         bio->bi_end_io = btrfs_end_dio_bio;
8457                         btrfs_io_bio(bio)->logical = file_offset;
8458
8459                         map_length = orig_bio->bi_iter.bi_size;
8460                         ret = btrfs_map_block(root->fs_info, bio_op(orig_bio),
8461                                               start_sector << 9,
8462                                               &map_length, NULL, 0);
8463                         if (ret) {
8464                                 bio_put(bio);
8465                                 goto out_err;
8466                         }
8467
8468                         goto next_block;
8469                 } else {
8470                         submit_len += blocksize;
8471                         if (--nr_sectors) {
8472                                 i++;
8473                                 goto next_block;
8474                         }
8475                         bvec++;
8476                 }
8477         }
8478
8479 submit:
8480         ret = __btrfs_submit_dio_bio(bio, inode, file_offset, skip_sum,
8481                                      async_submit);
8482         if (!ret)
8483                 return 0;
8484
8485         bio_put(bio);
8486 out_err:
8487         dip->errors = 1;
8488         /*
8489          * before atomic variable goto zero, we must
8490          * make sure dip->errors is perceived to be set.
8491          */
8492         smp_mb__before_atomic();
8493         if (atomic_dec_and_test(&dip->pending_bios))
8494                 bio_io_error(dip->orig_bio);
8495
8496         /* bio_end_io() will handle error, so we needn't return it */
8497         return 0;
8498 }
8499
8500 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8501                                 loff_t file_offset)
8502 {
8503         struct btrfs_dio_private *dip = NULL;
8504         struct bio *io_bio = NULL;
8505         struct btrfs_io_bio *btrfs_bio;
8506         int skip_sum;
8507         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8508         int ret = 0;
8509
8510         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8511
8512         io_bio = btrfs_bio_clone(dio_bio, GFP_NOFS);
8513         if (!io_bio) {
8514                 ret = -ENOMEM;
8515                 goto free_ordered;
8516         }
8517
8518         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8519         if (!dip) {
8520                 ret = -ENOMEM;
8521                 goto free_ordered;
8522         }
8523
8524         dip->private = dio_bio->bi_private;
8525         dip->inode = inode;
8526         dip->logical_offset = file_offset;
8527         dip->bytes = dio_bio->bi_iter.bi_size;
8528         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8529         io_bio->bi_private = dip;
8530         dip->orig_bio = io_bio;
8531         dip->dio_bio = dio_bio;
8532         atomic_set(&dip->pending_bios, 0);
8533         btrfs_bio = btrfs_io_bio(io_bio);
8534         btrfs_bio->logical = file_offset;
8535
8536         if (write) {
8537                 io_bio->bi_end_io = btrfs_endio_direct_write;
8538         } else {
8539                 io_bio->bi_end_io = btrfs_endio_direct_read;
8540                 dip->subio_endio = btrfs_subio_endio_read;
8541         }
8542
8543         /*
8544          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8545          * even if we fail to submit a bio, because in such case we do the
8546          * corresponding error handling below and it must not be done a second
8547          * time by btrfs_direct_IO().
8548          */
8549         if (write) {
8550                 struct btrfs_dio_data *dio_data = current->journal_info;
8551
8552                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8553                         dip->bytes;
8554                 dio_data->unsubmitted_oe_range_start =
8555                         dio_data->unsubmitted_oe_range_end;
8556         }
8557
8558         ret = btrfs_submit_direct_hook(dip, skip_sum);
8559         if (!ret)
8560                 return;
8561
8562         if (btrfs_bio->end_io)
8563                 btrfs_bio->end_io(btrfs_bio, ret);
8564
8565 free_ordered:
8566         /*
8567          * If we arrived here it means either we failed to submit the dip
8568          * or we either failed to clone the dio_bio or failed to allocate the
8569          * dip. If we cloned the dio_bio and allocated the dip, we can just
8570          * call bio_endio against our io_bio so that we get proper resource
8571          * cleanup if we fail to submit the dip, otherwise, we must do the
8572          * same as btrfs_endio_direct_[write|read] because we can't call these
8573          * callbacks - they require an allocated dip and a clone of dio_bio.
8574          */
8575         if (io_bio && dip) {
8576                 io_bio->bi_error = -EIO;
8577                 bio_endio(io_bio);
8578                 /*
8579                  * The end io callbacks free our dip, do the final put on io_bio
8580                  * and all the cleanup and final put for dio_bio (through
8581                  * dio_end_io()).
8582                  */
8583                 dip = NULL;
8584                 io_bio = NULL;
8585         } else {
8586                 if (write)
8587                         btrfs_endio_direct_write_update_ordered(inode,
8588                                                 file_offset,
8589                                                 dio_bio->bi_iter.bi_size,
8590                                                 0);
8591                 else
8592                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8593                               file_offset + dio_bio->bi_iter.bi_size - 1);
8594
8595                 dio_bio->bi_error = -EIO;
8596                 /*
8597                  * Releases and cleans up our dio_bio, no need to bio_put()
8598                  * nor bio_endio()/bio_io_error() against dio_bio.
8599                  */
8600                 dio_end_io(dio_bio, ret);
8601         }
8602         if (io_bio)
8603                 bio_put(io_bio);
8604         kfree(dip);
8605 }
8606
8607 static ssize_t check_direct_IO(struct btrfs_root *root, struct kiocb *iocb,
8608                         const struct iov_iter *iter, loff_t offset)
8609 {
8610         int seg;
8611         int i;
8612         unsigned blocksize_mask = root->sectorsize - 1;
8613         ssize_t retval = -EINVAL;
8614
8615         if (offset & blocksize_mask)
8616                 goto out;
8617
8618         if (iov_iter_alignment(iter) & blocksize_mask)
8619                 goto out;
8620
8621         /* If this is a write we don't need to check anymore */
8622         if (iov_iter_rw(iter) == WRITE)
8623                 return 0;
8624         /*
8625          * Check to make sure we don't have duplicate iov_base's in this
8626          * iovec, if so return EINVAL, otherwise we'll get csum errors
8627          * when reading back.
8628          */
8629         for (seg = 0; seg < iter->nr_segs; seg++) {
8630                 for (i = seg + 1; i < iter->nr_segs; i++) {
8631                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8632                                 goto out;
8633                 }
8634         }
8635         retval = 0;
8636 out:
8637         return retval;
8638 }
8639
8640 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8641 {
8642         struct file *file = iocb->ki_filp;
8643         struct inode *inode = file->f_mapping->host;
8644         struct btrfs_root *root = BTRFS_I(inode)->root;
8645         struct btrfs_dio_data dio_data = { 0 };
8646         loff_t offset = iocb->ki_pos;
8647         size_t count = 0;
8648         int flags = 0;
8649         bool wakeup = true;
8650         bool relock = false;
8651         ssize_t ret;
8652
8653         if (check_direct_IO(BTRFS_I(inode)->root, iocb, iter, offset))
8654                 return 0;
8655
8656         inode_dio_begin(inode);
8657         smp_mb__after_atomic();
8658
8659         /*
8660          * The generic stuff only does filemap_write_and_wait_range, which
8661          * isn't enough if we've written compressed pages to this area, so
8662          * we need to flush the dirty pages again to make absolutely sure
8663          * that any outstanding dirty pages are on disk.
8664          */
8665         count = iov_iter_count(iter);
8666         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8667                      &BTRFS_I(inode)->runtime_flags))
8668                 filemap_fdatawrite_range(inode->i_mapping, offset,
8669                                          offset + count - 1);
8670
8671         if (iov_iter_rw(iter) == WRITE) {
8672                 /*
8673                  * If the write DIO is beyond the EOF, we need update
8674                  * the isize, but it is protected by i_mutex. So we can
8675                  * not unlock the i_mutex at this case.
8676                  */
8677                 if (offset + count <= inode->i_size) {
8678                         inode_unlock(inode);
8679                         relock = true;
8680                 }
8681                 ret = btrfs_delalloc_reserve_space(inode, offset, count);
8682                 if (ret)
8683                         goto out;
8684                 dio_data.outstanding_extents = div64_u64(count +
8685                                                 BTRFS_MAX_EXTENT_SIZE - 1,
8686                                                 BTRFS_MAX_EXTENT_SIZE);
8687
8688                 /*
8689                  * We need to know how many extents we reserved so that we can
8690                  * do the accounting properly if we go over the number we
8691                  * originally calculated.  Abuse current->journal_info for this.
8692                  */
8693                 dio_data.reserve = round_up(count, root->sectorsize);
8694                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8695                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8696                 current->journal_info = &dio_data;
8697         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8698                                      &BTRFS_I(inode)->runtime_flags)) {
8699                 inode_dio_end(inode);
8700                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8701                 wakeup = false;
8702         }
8703
8704         ret = __blockdev_direct_IO(iocb, inode,
8705                                    BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev,
8706                                    iter, btrfs_get_blocks_direct, NULL,
8707                                    btrfs_submit_direct, flags);
8708         if (iov_iter_rw(iter) == WRITE) {
8709                 current->journal_info = NULL;
8710                 if (ret < 0 && ret != -EIOCBQUEUED) {
8711                         if (dio_data.reserve)
8712                                 btrfs_delalloc_release_space(inode, offset,
8713                                                              dio_data.reserve);
8714                         /*
8715                          * On error we might have left some ordered extents
8716                          * without submitting corresponding bios for them, so
8717                          * cleanup them up to avoid other tasks getting them
8718                          * and waiting for them to complete forever.
8719                          */
8720                         if (dio_data.unsubmitted_oe_range_start <
8721                             dio_data.unsubmitted_oe_range_end)
8722                                 btrfs_endio_direct_write_update_ordered(inode,
8723                                         dio_data.unsubmitted_oe_range_start,
8724                                         dio_data.unsubmitted_oe_range_end -
8725                                         dio_data.unsubmitted_oe_range_start,
8726                                         0);
8727                 } else if (ret >= 0 && (size_t)ret < count)
8728                         btrfs_delalloc_release_space(inode, offset,
8729                                                      count - (size_t)ret);
8730         }
8731 out:
8732         if (wakeup)
8733                 inode_dio_end(inode);
8734         if (relock)
8735                 inode_lock(inode);
8736
8737         return ret;
8738 }
8739
8740 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8741
8742 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8743                 __u64 start, __u64 len)
8744 {
8745         int     ret;
8746
8747         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8748         if (ret)
8749                 return ret;
8750
8751         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8752 }
8753
8754 int btrfs_readpage(struct file *file, struct page *page)
8755 {
8756         struct extent_io_tree *tree;
8757         tree = &BTRFS_I(page->mapping->host)->io_tree;
8758         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8759 }
8760
8761 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8762 {
8763         struct extent_io_tree *tree;
8764         struct inode *inode = page->mapping->host;
8765         int ret;
8766
8767         if (current->flags & PF_MEMALLOC) {
8768                 redirty_page_for_writepage(wbc, page);
8769                 unlock_page(page);
8770                 return 0;
8771         }
8772
8773         /*
8774          * If we are under memory pressure we will call this directly from the
8775          * VM, we need to make sure we have the inode referenced for the ordered
8776          * extent.  If not just return like we didn't do anything.
8777          */
8778         if (!igrab(inode)) {
8779                 redirty_page_for_writepage(wbc, page);
8780                 return AOP_WRITEPAGE_ACTIVATE;
8781         }
8782         tree = &BTRFS_I(page->mapping->host)->io_tree;
8783         ret = extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8784         btrfs_add_delayed_iput(inode);
8785         return ret;
8786 }
8787
8788 static int btrfs_writepages(struct address_space *mapping,
8789                             struct writeback_control *wbc)
8790 {
8791         struct extent_io_tree *tree;
8792
8793         tree = &BTRFS_I(mapping->host)->io_tree;
8794         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8795 }
8796
8797 static int
8798 btrfs_readpages(struct file *file, struct address_space *mapping,
8799                 struct list_head *pages, unsigned nr_pages)
8800 {
8801         struct extent_io_tree *tree;
8802         tree = &BTRFS_I(mapping->host)->io_tree;
8803         return extent_readpages(tree, mapping, pages, nr_pages,
8804                                 btrfs_get_extent);
8805 }
8806 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8807 {
8808         struct extent_io_tree *tree;
8809         struct extent_map_tree *map;
8810         int ret;
8811
8812         tree = &BTRFS_I(page->mapping->host)->io_tree;
8813         map = &BTRFS_I(page->mapping->host)->extent_tree;
8814         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8815         if (ret == 1) {
8816                 ClearPagePrivate(page);
8817                 set_page_private(page, 0);
8818                 put_page(page);
8819         }
8820         return ret;
8821 }
8822
8823 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8824 {
8825         if (PageWriteback(page) || PageDirty(page))
8826                 return 0;
8827         return __btrfs_releasepage(page, gfp_flags & GFP_NOFS);
8828 }
8829
8830 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8831                                  unsigned int length)
8832 {
8833         struct inode *inode = page->mapping->host;
8834         struct extent_io_tree *tree;
8835         struct btrfs_ordered_extent *ordered;
8836         struct extent_state *cached_state = NULL;
8837         u64 page_start = page_offset(page);
8838         u64 page_end = page_start + PAGE_SIZE - 1;
8839         u64 start;
8840         u64 end;
8841         int inode_evicting = inode->i_state & I_FREEING;
8842
8843         /*
8844          * we have the page locked, so new writeback can't start,
8845          * and the dirty bit won't be cleared while we are here.
8846          *
8847          * Wait for IO on this page so that we can safely clear
8848          * the PagePrivate2 bit and do ordered accounting
8849          */
8850         wait_on_page_writeback(page);
8851
8852         tree = &BTRFS_I(inode)->io_tree;
8853         if (offset) {
8854                 btrfs_releasepage(page, GFP_NOFS);
8855                 return;
8856         }
8857
8858         if (!inode_evicting)
8859                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8860 again:
8861         start = page_start;
8862         ordered = btrfs_lookup_ordered_range(inode, start,
8863                                         page_end - start + 1);
8864         if (ordered) {
8865                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8866                 /*
8867                  * IO on this page will never be started, so we need
8868                  * to account for any ordered extents now
8869                  */
8870                 if (!inode_evicting)
8871                         clear_extent_bit(tree, start, end,
8872                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8873                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8874                                          EXTENT_DEFRAG, 1, 0, &cached_state,
8875                                          GFP_NOFS);
8876                 /*
8877                  * whoever cleared the private bit is responsible
8878                  * for the finish_ordered_io
8879                  */
8880                 if (TestClearPagePrivate2(page)) {
8881                         struct btrfs_ordered_inode_tree *tree;
8882                         u64 new_len;
8883
8884                         tree = &BTRFS_I(inode)->ordered_tree;
8885
8886                         spin_lock_irq(&tree->lock);
8887                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8888                         new_len = start - ordered->file_offset;
8889                         if (new_len < ordered->truncated_len)
8890                                 ordered->truncated_len = new_len;
8891                         spin_unlock_irq(&tree->lock);
8892
8893                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8894                                                            start,
8895                                                            end - start + 1, 1))
8896                                 btrfs_finish_ordered_io(ordered);
8897                 }
8898                 btrfs_put_ordered_extent(ordered);
8899                 if (!inode_evicting) {
8900                         cached_state = NULL;
8901                         lock_extent_bits(tree, start, end,
8902                                          &cached_state);
8903                 }
8904
8905                 start = end + 1;
8906                 if (start < page_end)
8907                         goto again;
8908         }
8909
8910         /*
8911          * Qgroup reserved space handler
8912          * Page here will be either
8913          * 1) Already written to disk
8914          *    In this case, its reserved space is released from data rsv map
8915          *    and will be freed by delayed_ref handler finally.
8916          *    So even we call qgroup_free_data(), it won't decrease reserved
8917          *    space.
8918          * 2) Not written to disk
8919          *    This means the reserved space should be freed here.
8920          */
8921         btrfs_qgroup_free_data(inode, page_start, PAGE_SIZE);
8922         if (!inode_evicting) {
8923                 clear_extent_bit(tree, page_start, page_end,
8924                                  EXTENT_LOCKED | EXTENT_DIRTY |
8925                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
8926                                  EXTENT_DEFRAG, 1, 1,
8927                                  &cached_state, GFP_NOFS);
8928
8929                 __btrfs_releasepage(page, GFP_NOFS);
8930         }
8931
8932         ClearPageChecked(page);
8933         if (PagePrivate(page)) {
8934                 ClearPagePrivate(page);
8935                 set_page_private(page, 0);
8936                 put_page(page);
8937         }
8938 }
8939
8940 /*
8941  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8942  * called from a page fault handler when a page is first dirtied. Hence we must
8943  * be careful to check for EOF conditions here. We set the page up correctly
8944  * for a written page which means we get ENOSPC checking when writing into
8945  * holes and correct delalloc and unwritten extent mapping on filesystems that
8946  * support these features.
8947  *
8948  * We are not allowed to take the i_mutex here so we have to play games to
8949  * protect against truncate races as the page could now be beyond EOF.  Because
8950  * vmtruncate() writes the inode size before removing pages, once we have the
8951  * page lock we can determine safely if the page is beyond EOF. If it is not
8952  * beyond EOF, then the page is guaranteed safe against truncation until we
8953  * unlock the page.
8954  */
8955 int btrfs_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
8956 {
8957         struct page *page = vmf->page;
8958         struct inode *inode = file_inode(vma->vm_file);
8959         struct btrfs_root *root = BTRFS_I(inode)->root;
8960         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8961         struct btrfs_ordered_extent *ordered;
8962         struct extent_state *cached_state = NULL;
8963         char *kaddr;
8964         unsigned long zero_start;
8965         loff_t size;
8966         int ret;
8967         int reserved = 0;
8968         u64 reserved_space;
8969         u64 page_start;
8970         u64 page_end;
8971         u64 end;
8972
8973         reserved_space = PAGE_SIZE;
8974
8975         sb_start_pagefault(inode->i_sb);
8976         page_start = page_offset(page);
8977         page_end = page_start + PAGE_SIZE - 1;
8978         end = page_end;
8979
8980         /*
8981          * Reserving delalloc space after obtaining the page lock can lead to
8982          * deadlock. For example, if a dirty page is locked by this function
8983          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8984          * dirty page write out, then the btrfs_writepage() function could
8985          * end up waiting indefinitely to get a lock on the page currently
8986          * being processed by btrfs_page_mkwrite() function.
8987          */
8988         ret = btrfs_delalloc_reserve_space(inode, page_start,
8989                                            reserved_space);
8990         if (!ret) {
8991                 ret = file_update_time(vma->vm_file);
8992                 reserved = 1;
8993         }
8994         if (ret) {
8995                 if (ret == -ENOMEM)
8996                         ret = VM_FAULT_OOM;
8997                 else /* -ENOSPC, -EIO, etc */
8998                         ret = VM_FAULT_SIGBUS;
8999                 if (reserved)
9000                         goto out;
9001                 goto out_noreserve;
9002         }
9003
9004         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
9005 again:
9006         lock_page(page);
9007         size = i_size_read(inode);
9008
9009         if ((page->mapping != inode->i_mapping) ||
9010             (page_start >= size)) {
9011                 /* page got truncated out from underneath us */
9012                 goto out_unlock;
9013         }
9014         wait_on_page_writeback(page);
9015
9016         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
9017         set_page_extent_mapped(page);
9018
9019         /*
9020          * we can't set the delalloc bits if there are pending ordered
9021          * extents.  Drop our locks and wait for them to finish
9022          */
9023         ordered = btrfs_lookup_ordered_range(inode, page_start, page_end);
9024         if (ordered) {
9025                 unlock_extent_cached(io_tree, page_start, page_end,
9026                                      &cached_state, GFP_NOFS);
9027                 unlock_page(page);
9028                 btrfs_start_ordered_extent(inode, ordered, 1);
9029                 btrfs_put_ordered_extent(ordered);
9030                 goto again;
9031         }
9032
9033         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
9034                 reserved_space = round_up(size - page_start, root->sectorsize);
9035                 if (reserved_space < PAGE_SIZE) {
9036                         end = page_start + reserved_space - 1;
9037                         spin_lock(&BTRFS_I(inode)->lock);
9038                         BTRFS_I(inode)->outstanding_extents++;
9039                         spin_unlock(&BTRFS_I(inode)->lock);
9040                         btrfs_delalloc_release_space(inode, page_start,
9041                                                 PAGE_SIZE - reserved_space);
9042                 }
9043         }
9044
9045         /*
9046          * XXX - page_mkwrite gets called every time the page is dirtied, even
9047          * if it was already dirty, so for space accounting reasons we need to
9048          * clear any delalloc bits for the range we are fixing to save.  There
9049          * is probably a better way to do this, but for now keep consistent with
9050          * prepare_pages in the normal write path.
9051          */
9052         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9053                           EXTENT_DIRTY | EXTENT_DELALLOC |
9054                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
9055                           0, 0, &cached_state, GFP_NOFS);
9056
9057         ret = btrfs_set_extent_delalloc(inode, page_start, end,
9058                                         &cached_state);
9059         if (ret) {
9060                 unlock_extent_cached(io_tree, page_start, page_end,
9061                                      &cached_state, GFP_NOFS);
9062                 ret = VM_FAULT_SIGBUS;
9063                 goto out_unlock;
9064         }
9065         ret = 0;
9066
9067         /* page is wholly or partially inside EOF */
9068         if (page_start + PAGE_SIZE > size)
9069                 zero_start = size & ~PAGE_MASK;
9070         else
9071                 zero_start = PAGE_SIZE;
9072
9073         if (zero_start != PAGE_SIZE) {
9074                 kaddr = kmap(page);
9075                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9076                 flush_dcache_page(page);
9077                 kunmap(page);
9078         }
9079         ClearPageChecked(page);
9080         set_page_dirty(page);
9081         SetPageUptodate(page);
9082
9083         BTRFS_I(inode)->last_trans = root->fs_info->generation;
9084         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9085         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9086
9087         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
9088
9089 out_unlock:
9090         if (!ret) {
9091                 sb_end_pagefault(inode->i_sb);
9092                 return VM_FAULT_LOCKED;
9093         }
9094         unlock_page(page);
9095 out:
9096         btrfs_delalloc_release_space(inode, page_start, reserved_space);
9097 out_noreserve:
9098         sb_end_pagefault(inode->i_sb);
9099         return ret;
9100 }
9101
9102 static int btrfs_truncate(struct inode *inode)
9103 {
9104         struct btrfs_root *root = BTRFS_I(inode)->root;
9105         struct btrfs_block_rsv *rsv;
9106         int ret = 0;
9107         int err = 0;
9108         struct btrfs_trans_handle *trans;
9109         u64 mask = root->sectorsize - 1;
9110         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
9111
9112         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9113                                        (u64)-1);
9114         if (ret)
9115                 return ret;
9116
9117         /*
9118          * Yes ladies and gentlemen, this is indeed ugly.  The fact is we have
9119          * 3 things going on here
9120          *
9121          * 1) We need to reserve space for our orphan item and the space to
9122          * delete our orphan item.  Lord knows we don't want to have a dangling
9123          * orphan item because we didn't reserve space to remove it.
9124          *
9125          * 2) We need to reserve space to update our inode.
9126          *
9127          * 3) We need to have something to cache all the space that is going to
9128          * be free'd up by the truncate operation, but also have some slack
9129          * space reserved in case it uses space during the truncate (thank you
9130          * very much snapshotting).
9131          *
9132          * And we need these to all be separate.  The fact is we can use a lot of
9133          * space doing the truncate, and we have no earthly idea how much space
9134          * we will use, so we need the truncate reservation to be separate so it
9135          * doesn't end up using space reserved for updating the inode or
9136          * removing the orphan item.  We also need to be able to stop the
9137          * transaction and start a new one, which means we need to be able to
9138          * update the inode several times, and we have no idea of knowing how
9139          * many times that will be, so we can't just reserve 1 item for the
9140          * entirety of the operation, so that has to be done separately as well.
9141          * Then there is the orphan item, which does indeed need to be held on
9142          * to for the whole operation, and we need nobody to touch this reserved
9143          * space except the orphan code.
9144          *
9145          * So that leaves us with
9146          *
9147          * 1) root->orphan_block_rsv - for the orphan deletion.
9148          * 2) rsv - for the truncate reservation, which we will steal from the
9149          * transaction reservation.
9150          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
9151          * updating the inode.
9152          */
9153         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
9154         if (!rsv)
9155                 return -ENOMEM;
9156         rsv->size = min_size;
9157         rsv->failfast = 1;
9158
9159         /*
9160          * 1 for the truncate slack space
9161          * 1 for updating the inode.
9162          */
9163         trans = btrfs_start_transaction(root, 2);
9164         if (IS_ERR(trans)) {
9165                 err = PTR_ERR(trans);
9166                 goto out;
9167         }
9168
9169         /* Migrate the slack space for the truncate to our reserve */
9170         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
9171                                       min_size, 0);
9172         BUG_ON(ret);
9173
9174         /*
9175          * So if we truncate and then write and fsync we normally would just
9176          * write the extents that changed, which is a problem if we need to
9177          * first truncate that entire inode.  So set this flag so we write out
9178          * all of the extents in the inode to the sync log so we're completely
9179          * safe.
9180          */
9181         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9182         trans->block_rsv = rsv;
9183
9184         while (1) {
9185                 ret = btrfs_truncate_inode_items(trans, root, inode,
9186                                                  inode->i_size,
9187                                                  BTRFS_EXTENT_DATA_KEY);
9188                 if (ret != -ENOSPC && ret != -EAGAIN) {
9189                         err = ret;
9190                         break;
9191                 }
9192
9193                 trans->block_rsv = &root->fs_info->trans_block_rsv;
9194                 ret = btrfs_update_inode(trans, root, inode);
9195                 if (ret) {
9196                         err = ret;
9197                         break;
9198                 }
9199
9200                 btrfs_end_transaction(trans, root);
9201                 btrfs_btree_balance_dirty(root);
9202
9203                 trans = btrfs_start_transaction(root, 2);
9204                 if (IS_ERR(trans)) {
9205                         ret = err = PTR_ERR(trans);
9206                         trans = NULL;
9207                         break;
9208                 }
9209
9210                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
9211                                               rsv, min_size, 0);
9212                 BUG_ON(ret);    /* shouldn't happen */
9213                 trans->block_rsv = rsv;
9214         }
9215
9216         if (ret == 0 && inode->i_nlink > 0) {
9217                 trans->block_rsv = root->orphan_block_rsv;
9218                 ret = btrfs_orphan_del(trans, inode);
9219                 if (ret)
9220                         err = ret;
9221         }
9222
9223         if (trans) {
9224                 trans->block_rsv = &root->fs_info->trans_block_rsv;
9225                 ret = btrfs_update_inode(trans, root, inode);
9226                 if (ret && !err)
9227                         err = ret;
9228
9229                 ret = btrfs_end_transaction(trans, root);
9230                 btrfs_btree_balance_dirty(root);
9231         }
9232 out:
9233         btrfs_free_block_rsv(root, rsv);
9234
9235         if (ret && !err)
9236                 err = ret;
9237
9238         return err;
9239 }
9240
9241 /*
9242  * create a new subvolume directory/inode (helper for the ioctl).
9243  */
9244 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9245                              struct btrfs_root *new_root,
9246                              struct btrfs_root *parent_root,
9247                              u64 new_dirid)
9248 {
9249         struct inode *inode;
9250         int err;
9251         u64 index = 0;
9252
9253         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9254                                 new_dirid, new_dirid,
9255                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9256                                 &index);
9257         if (IS_ERR(inode))
9258                 return PTR_ERR(inode);
9259         inode->i_op = &btrfs_dir_inode_operations;
9260         inode->i_fop = &btrfs_dir_file_operations;
9261
9262         set_nlink(inode, 1);
9263         btrfs_i_size_write(inode, 0);
9264         unlock_new_inode(inode);
9265
9266         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9267         if (err)
9268                 btrfs_err(new_root->fs_info,
9269                           "error inheriting subvolume %llu properties: %d",
9270                           new_root->root_key.objectid, err);
9271
9272         err = btrfs_update_inode(trans, new_root, inode);
9273
9274         iput(inode);
9275         return err;
9276 }
9277
9278 struct inode *btrfs_alloc_inode(struct super_block *sb)
9279 {
9280         struct btrfs_inode *ei;
9281         struct inode *inode;
9282
9283         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9284         if (!ei)
9285                 return NULL;
9286
9287         ei->root = NULL;
9288         ei->generation = 0;
9289         ei->last_trans = 0;
9290         ei->last_sub_trans = 0;
9291         ei->logged_trans = 0;
9292         ei->delalloc_bytes = 0;
9293         ei->defrag_bytes = 0;
9294         ei->disk_i_size = 0;
9295         ei->flags = 0;
9296         ei->csum_bytes = 0;
9297         ei->index_cnt = (u64)-1;
9298         ei->dir_index = 0;
9299         ei->last_unlink_trans = 0;
9300         ei->last_log_commit = 0;
9301         ei->delayed_iput_count = 0;
9302
9303         spin_lock_init(&ei->lock);
9304         ei->outstanding_extents = 0;
9305         ei->reserved_extents = 0;
9306
9307         ei->runtime_flags = 0;
9308         ei->force_compress = BTRFS_COMPRESS_NONE;
9309
9310         ei->delayed_node = NULL;
9311
9312         ei->i_otime.tv_sec = 0;
9313         ei->i_otime.tv_nsec = 0;
9314
9315         inode = &ei->vfs_inode;
9316         extent_map_tree_init(&ei->extent_tree);
9317         extent_io_tree_init(&ei->io_tree, &inode->i_data);
9318         extent_io_tree_init(&ei->io_failure_tree, &inode->i_data);
9319         ei->io_tree.track_uptodate = 1;
9320         ei->io_failure_tree.track_uptodate = 1;
9321         atomic_set(&ei->sync_writers, 0);
9322         mutex_init(&ei->log_mutex);
9323         mutex_init(&ei->delalloc_mutex);
9324         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9325         INIT_LIST_HEAD(&ei->delalloc_inodes);
9326         INIT_LIST_HEAD(&ei->delayed_iput);
9327         RB_CLEAR_NODE(&ei->rb_node);
9328         init_rwsem(&ei->dio_sem);
9329
9330         return inode;
9331 }
9332
9333 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9334 void btrfs_test_destroy_inode(struct inode *inode)
9335 {
9336         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9337         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9338 }
9339 #endif
9340
9341 static void btrfs_i_callback(struct rcu_head *head)
9342 {
9343         struct inode *inode = container_of(head, struct inode, i_rcu);
9344         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9345 }
9346
9347 void btrfs_destroy_inode(struct inode *inode)
9348 {
9349         struct btrfs_ordered_extent *ordered;
9350         struct btrfs_root *root = BTRFS_I(inode)->root;
9351
9352         WARN_ON(!hlist_empty(&inode->i_dentry));
9353         WARN_ON(inode->i_data.nrpages);
9354         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9355         WARN_ON(BTRFS_I(inode)->reserved_extents);
9356         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9357         WARN_ON(BTRFS_I(inode)->csum_bytes);
9358         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9359
9360         /*
9361          * This can happen where we create an inode, but somebody else also
9362          * created the same inode and we need to destroy the one we already
9363          * created.
9364          */
9365         if (!root)
9366                 goto free;
9367
9368         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9369                      &BTRFS_I(inode)->runtime_flags)) {
9370                 btrfs_info(root->fs_info, "inode %llu still on the orphan list",
9371                         btrfs_ino(inode));
9372                 atomic_dec(&root->orphan_inodes);
9373         }
9374
9375         while (1) {
9376                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9377                 if (!ordered)
9378                         break;
9379                 else {
9380                         btrfs_err(root->fs_info, "found ordered extent %llu %llu on inode cleanup",
9381                                 ordered->file_offset, ordered->len);
9382                         btrfs_remove_ordered_extent(inode, ordered);
9383                         btrfs_put_ordered_extent(ordered);
9384                         btrfs_put_ordered_extent(ordered);
9385                 }
9386         }
9387         btrfs_qgroup_check_reserved_leak(inode);
9388         inode_tree_del(inode);
9389         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9390 free:
9391         call_rcu(&inode->i_rcu, btrfs_i_callback);
9392 }
9393
9394 int btrfs_drop_inode(struct inode *inode)
9395 {
9396         struct btrfs_root *root = BTRFS_I(inode)->root;
9397
9398         if (root == NULL)
9399                 return 1;
9400
9401         /* the snap/subvol tree is on deleting */
9402         if (btrfs_root_refs(&root->root_item) == 0)
9403                 return 1;
9404         else
9405                 return generic_drop_inode(inode);
9406 }
9407
9408 static void init_once(void *foo)
9409 {
9410         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9411
9412         inode_init_once(&ei->vfs_inode);
9413 }
9414
9415 void btrfs_destroy_cachep(void)
9416 {
9417         /*
9418          * Make sure all delayed rcu free inodes are flushed before we
9419          * destroy cache.
9420          */
9421         rcu_barrier();
9422         kmem_cache_destroy(btrfs_inode_cachep);
9423         kmem_cache_destroy(btrfs_trans_handle_cachep);
9424         kmem_cache_destroy(btrfs_transaction_cachep);
9425         kmem_cache_destroy(btrfs_path_cachep);
9426         kmem_cache_destroy(btrfs_free_space_cachep);
9427 }
9428
9429 int btrfs_init_cachep(void)
9430 {
9431         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9432                         sizeof(struct btrfs_inode), 0,
9433                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9434                         init_once);
9435         if (!btrfs_inode_cachep)
9436                 goto fail;
9437
9438         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9439                         sizeof(struct btrfs_trans_handle), 0,
9440                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9441         if (!btrfs_trans_handle_cachep)
9442                 goto fail;
9443
9444         btrfs_transaction_cachep = kmem_cache_create("btrfs_transaction",
9445                         sizeof(struct btrfs_transaction), 0,
9446                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9447         if (!btrfs_transaction_cachep)
9448                 goto fail;
9449
9450         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9451                         sizeof(struct btrfs_path), 0,
9452                         SLAB_MEM_SPREAD, NULL);
9453         if (!btrfs_path_cachep)
9454                 goto fail;
9455
9456         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9457                         sizeof(struct btrfs_free_space), 0,
9458                         SLAB_MEM_SPREAD, NULL);
9459         if (!btrfs_free_space_cachep)
9460                 goto fail;
9461
9462         return 0;
9463 fail:
9464         btrfs_destroy_cachep();
9465         return -ENOMEM;
9466 }
9467
9468 static int btrfs_getattr(struct vfsmount *mnt,
9469                          struct dentry *dentry, struct kstat *stat)
9470 {
9471         u64 delalloc_bytes;
9472         struct inode *inode = d_inode(dentry);
9473         u32 blocksize = inode->i_sb->s_blocksize;
9474
9475         generic_fillattr(inode, stat);
9476         stat->dev = BTRFS_I(inode)->root->anon_dev;
9477
9478         spin_lock(&BTRFS_I(inode)->lock);
9479         delalloc_bytes = BTRFS_I(inode)->delalloc_bytes;
9480         spin_unlock(&BTRFS_I(inode)->lock);
9481         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9482                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9483         return 0;
9484 }
9485
9486 static int btrfs_rename_exchange(struct inode *old_dir,
9487                               struct dentry *old_dentry,
9488                               struct inode *new_dir,
9489                               struct dentry *new_dentry)
9490 {
9491         struct btrfs_trans_handle *trans;
9492         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9493         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9494         struct inode *new_inode = new_dentry->d_inode;
9495         struct inode *old_inode = old_dentry->d_inode;
9496         struct timespec ctime = CURRENT_TIME;
9497         struct dentry *parent;
9498         u64 old_ino = btrfs_ino(old_inode);
9499         u64 new_ino = btrfs_ino(new_inode);
9500         u64 old_idx = 0;
9501         u64 new_idx = 0;
9502         u64 root_objectid;
9503         int ret;
9504         bool root_log_pinned = false;
9505         bool dest_log_pinned = false;
9506
9507         /* we only allow rename subvolume link between subvolumes */
9508         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9509                 return -EXDEV;
9510
9511         /* close the race window with snapshot create/destroy ioctl */
9512         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9513                 down_read(&root->fs_info->subvol_sem);
9514         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9515                 down_read(&dest->fs_info->subvol_sem);
9516
9517         /*
9518          * We want to reserve the absolute worst case amount of items.  So if
9519          * both inodes are subvols and we need to unlink them then that would
9520          * require 4 item modifications, but if they are both normal inodes it
9521          * would require 5 item modifications, so we'll assume their normal
9522          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9523          * should cover the worst case number of items we'll modify.
9524          */
9525         trans = btrfs_start_transaction(root, 12);
9526         if (IS_ERR(trans)) {
9527                 ret = PTR_ERR(trans);
9528                 goto out_notrans;
9529         }
9530
9531         /*
9532          * We need to find a free sequence number both in the source and
9533          * in the destination directory for the exchange.
9534          */
9535         ret = btrfs_set_inode_index(new_dir, &old_idx);
9536         if (ret)
9537                 goto out_fail;
9538         ret = btrfs_set_inode_index(old_dir, &new_idx);
9539         if (ret)
9540                 goto out_fail;
9541
9542         BTRFS_I(old_inode)->dir_index = 0ULL;
9543         BTRFS_I(new_inode)->dir_index = 0ULL;
9544
9545         /* Reference for the source. */
9546         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9547                 /* force full log commit if subvolume involved. */
9548                 btrfs_set_log_full_commit(root->fs_info, trans);
9549         } else {
9550                 btrfs_pin_log_trans(root);
9551                 root_log_pinned = true;
9552                 ret = btrfs_insert_inode_ref(trans, dest,
9553                                              new_dentry->d_name.name,
9554                                              new_dentry->d_name.len,
9555                                              old_ino,
9556                                              btrfs_ino(new_dir), old_idx);
9557                 if (ret)
9558                         goto out_fail;
9559         }
9560
9561         /* And now for the dest. */
9562         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9563                 /* force full log commit if subvolume involved. */
9564                 btrfs_set_log_full_commit(dest->fs_info, trans);
9565         } else {
9566                 btrfs_pin_log_trans(dest);
9567                 dest_log_pinned = true;
9568                 ret = btrfs_insert_inode_ref(trans, root,
9569                                              old_dentry->d_name.name,
9570                                              old_dentry->d_name.len,
9571                                              new_ino,
9572                                              btrfs_ino(old_dir), new_idx);
9573                 if (ret)
9574                         goto out_fail;
9575         }
9576
9577         /* Update inode version and ctime/mtime. */
9578         inode_inc_iversion(old_dir);
9579         inode_inc_iversion(new_dir);
9580         inode_inc_iversion(old_inode);
9581         inode_inc_iversion(new_inode);
9582         old_dir->i_ctime = old_dir->i_mtime = ctime;
9583         new_dir->i_ctime = new_dir->i_mtime = ctime;
9584         old_inode->i_ctime = ctime;
9585         new_inode->i_ctime = ctime;
9586
9587         if (old_dentry->d_parent != new_dentry->d_parent) {
9588                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9589                 btrfs_record_unlink_dir(trans, new_dir, new_inode, 1);
9590         }
9591
9592         /* src is a subvolume */
9593         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9594                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9595                 ret = btrfs_unlink_subvol(trans, root, old_dir,
9596                                           root_objectid,
9597                                           old_dentry->d_name.name,
9598                                           old_dentry->d_name.len);
9599         } else { /* src is an inode */
9600                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9601                                            old_dentry->d_inode,
9602                                            old_dentry->d_name.name,
9603                                            old_dentry->d_name.len);
9604                 if (!ret)
9605                         ret = btrfs_update_inode(trans, root, old_inode);
9606         }
9607         if (ret) {
9608                 btrfs_abort_transaction(trans, ret);
9609                 goto out_fail;
9610         }
9611
9612         /* dest is a subvolume */
9613         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9614                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9615                 ret = btrfs_unlink_subvol(trans, dest, new_dir,
9616                                           root_objectid,
9617                                           new_dentry->d_name.name,
9618                                           new_dentry->d_name.len);
9619         } else { /* dest is an inode */
9620                 ret = __btrfs_unlink_inode(trans, dest, new_dir,
9621                                            new_dentry->d_inode,
9622                                            new_dentry->d_name.name,
9623                                            new_dentry->d_name.len);
9624                 if (!ret)
9625                         ret = btrfs_update_inode(trans, dest, new_inode);
9626         }
9627         if (ret) {
9628                 btrfs_abort_transaction(trans, ret);
9629                 goto out_fail;
9630         }
9631
9632         ret = btrfs_add_link(trans, new_dir, old_inode,
9633                              new_dentry->d_name.name,
9634                              new_dentry->d_name.len, 0, old_idx);
9635         if (ret) {
9636                 btrfs_abort_transaction(trans, ret);
9637                 goto out_fail;
9638         }
9639
9640         ret = btrfs_add_link(trans, old_dir, new_inode,
9641                              old_dentry->d_name.name,
9642                              old_dentry->d_name.len, 0, new_idx);
9643         if (ret) {
9644                 btrfs_abort_transaction(trans, ret);
9645                 goto out_fail;
9646         }
9647
9648         if (old_inode->i_nlink == 1)
9649                 BTRFS_I(old_inode)->dir_index = old_idx;
9650         if (new_inode->i_nlink == 1)
9651                 BTRFS_I(new_inode)->dir_index = new_idx;
9652
9653         if (root_log_pinned) {
9654                 parent = new_dentry->d_parent;
9655                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9656                 btrfs_end_log_trans(root);
9657                 root_log_pinned = false;
9658         }
9659         if (dest_log_pinned) {
9660                 parent = old_dentry->d_parent;
9661                 btrfs_log_new_name(trans, new_inode, new_dir, parent);
9662                 btrfs_end_log_trans(dest);
9663                 dest_log_pinned = false;
9664         }
9665 out_fail:
9666         /*
9667          * If we have pinned a log and an error happened, we unpin tasks
9668          * trying to sync the log and force them to fallback to a transaction
9669          * commit if the log currently contains any of the inodes involved in
9670          * this rename operation (to ensure we do not persist a log with an
9671          * inconsistent state for any of these inodes or leading to any
9672          * inconsistencies when replayed). If the transaction was aborted, the
9673          * abortion reason is propagated to userspace when attempting to commit
9674          * the transaction. If the log does not contain any of these inodes, we
9675          * allow the tasks to sync it.
9676          */
9677         if (ret && (root_log_pinned || dest_log_pinned)) {
9678                 if (btrfs_inode_in_log(old_dir, root->fs_info->generation) ||
9679                     btrfs_inode_in_log(new_dir, root->fs_info->generation) ||
9680                     btrfs_inode_in_log(old_inode, root->fs_info->generation) ||
9681                     (new_inode &&
9682                      btrfs_inode_in_log(new_inode, root->fs_info->generation)))
9683                     btrfs_set_log_full_commit(root->fs_info, trans);
9684
9685                 if (root_log_pinned) {
9686                         btrfs_end_log_trans(root);
9687                         root_log_pinned = false;
9688                 }
9689                 if (dest_log_pinned) {
9690                         btrfs_end_log_trans(dest);
9691                         dest_log_pinned = false;
9692                 }
9693         }
9694         ret = btrfs_end_transaction(trans, root);
9695 out_notrans:
9696         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9697                 up_read(&dest->fs_info->subvol_sem);
9698         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9699                 up_read(&root->fs_info->subvol_sem);
9700
9701         return ret;
9702 }
9703
9704 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9705                                      struct btrfs_root *root,
9706                                      struct inode *dir,
9707                                      struct dentry *dentry)
9708 {
9709         int ret;
9710         struct inode *inode;
9711         u64 objectid;
9712         u64 index;
9713
9714         ret = btrfs_find_free_ino(root, &objectid);
9715         if (ret)
9716                 return ret;
9717
9718         inode = btrfs_new_inode(trans, root, dir,
9719                                 dentry->d_name.name,
9720                                 dentry->d_name.len,
9721                                 btrfs_ino(dir),
9722                                 objectid,
9723                                 S_IFCHR | WHITEOUT_MODE,
9724                                 &index);
9725
9726         if (IS_ERR(inode)) {
9727                 ret = PTR_ERR(inode);
9728                 return ret;
9729         }
9730
9731         inode->i_op = &btrfs_special_inode_operations;
9732         init_special_inode(inode, inode->i_mode,
9733                 WHITEOUT_DEV);
9734
9735         ret = btrfs_init_inode_security(trans, inode, dir,
9736                                 &dentry->d_name);
9737         if (ret)
9738                 goto out;
9739
9740         ret = btrfs_add_nondir(trans, dir, dentry,
9741                                 inode, 0, index);
9742         if (ret)
9743                 goto out;
9744
9745         ret = btrfs_update_inode(trans, root, inode);
9746 out:
9747         unlock_new_inode(inode);
9748         if (ret)
9749                 inode_dec_link_count(inode);
9750         iput(inode);
9751
9752         return ret;
9753 }
9754
9755 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9756                            struct inode *new_dir, struct dentry *new_dentry,
9757                            unsigned int flags)
9758 {
9759         struct btrfs_trans_handle *trans;
9760         unsigned int trans_num_items;
9761         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9762         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9763         struct inode *new_inode = d_inode(new_dentry);
9764         struct inode *old_inode = d_inode(old_dentry);
9765         u64 index = 0;
9766         u64 root_objectid;
9767         int ret;
9768         u64 old_ino = btrfs_ino(old_inode);
9769         bool log_pinned = false;
9770
9771         if (btrfs_ino(new_dir) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9772                 return -EPERM;
9773
9774         /* we only allow rename subvolume link between subvolumes */
9775         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9776                 return -EXDEV;
9777
9778         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9779             (new_inode && btrfs_ino(new_inode) == BTRFS_FIRST_FREE_OBJECTID))
9780                 return -ENOTEMPTY;
9781
9782         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9783             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9784                 return -ENOTEMPTY;
9785
9786
9787         /* check for collisions, even if the  name isn't there */
9788         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9789                              new_dentry->d_name.name,
9790                              new_dentry->d_name.len);
9791
9792         if (ret) {
9793                 if (ret == -EEXIST) {
9794                         /* we shouldn't get
9795                          * eexist without a new_inode */
9796                         if (WARN_ON(!new_inode)) {
9797                                 return ret;
9798                         }
9799                 } else {
9800                         /* maybe -EOVERFLOW */
9801                         return ret;
9802                 }
9803         }
9804         ret = 0;
9805
9806         /*
9807          * we're using rename to replace one file with another.  Start IO on it
9808          * now so  we don't add too much work to the end of the transaction
9809          */
9810         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9811                 filemap_flush(old_inode->i_mapping);
9812
9813         /* close the racy window with snapshot create/destroy ioctl */
9814         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9815                 down_read(&root->fs_info->subvol_sem);
9816         /*
9817          * We want to reserve the absolute worst case amount of items.  So if
9818          * both inodes are subvols and we need to unlink them then that would
9819          * require 4 item modifications, but if they are both normal inodes it
9820          * would require 5 item modifications, so we'll assume they are normal
9821          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9822          * should cover the worst case number of items we'll modify.
9823          * If our rename has the whiteout flag, we need more 5 units for the
9824          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9825          * when selinux is enabled).
9826          */
9827         trans_num_items = 11;
9828         if (flags & RENAME_WHITEOUT)
9829                 trans_num_items += 5;
9830         trans = btrfs_start_transaction(root, trans_num_items);
9831         if (IS_ERR(trans)) {
9832                 ret = PTR_ERR(trans);
9833                 goto out_notrans;
9834         }
9835
9836         if (dest != root)
9837                 btrfs_record_root_in_trans(trans, dest);
9838
9839         ret = btrfs_set_inode_index(new_dir, &index);
9840         if (ret)
9841                 goto out_fail;
9842
9843         BTRFS_I(old_inode)->dir_index = 0ULL;
9844         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9845                 /* force full log commit if subvolume involved. */
9846                 btrfs_set_log_full_commit(root->fs_info, trans);
9847         } else {
9848                 btrfs_pin_log_trans(root);
9849                 log_pinned = true;
9850                 ret = btrfs_insert_inode_ref(trans, dest,
9851                                              new_dentry->d_name.name,
9852                                              new_dentry->d_name.len,
9853                                              old_ino,
9854                                              btrfs_ino(new_dir), index);
9855                 if (ret)
9856                         goto out_fail;
9857         }
9858
9859         inode_inc_iversion(old_dir);
9860         inode_inc_iversion(new_dir);
9861         inode_inc_iversion(old_inode);
9862         old_dir->i_ctime = old_dir->i_mtime =
9863         new_dir->i_ctime = new_dir->i_mtime =
9864         old_inode->i_ctime = current_fs_time(old_dir->i_sb);
9865
9866         if (old_dentry->d_parent != new_dentry->d_parent)
9867                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9868
9869         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9870                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9871                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9872                                         old_dentry->d_name.name,
9873                                         old_dentry->d_name.len);
9874         } else {
9875                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9876                                         d_inode(old_dentry),
9877                                         old_dentry->d_name.name,
9878                                         old_dentry->d_name.len);
9879                 if (!ret)
9880                         ret = btrfs_update_inode(trans, root, old_inode);
9881         }
9882         if (ret) {
9883                 btrfs_abort_transaction(trans, ret);
9884                 goto out_fail;
9885         }
9886
9887         if (new_inode) {
9888                 inode_inc_iversion(new_inode);
9889                 new_inode->i_ctime = current_fs_time(new_inode->i_sb);
9890                 if (unlikely(btrfs_ino(new_inode) ==
9891                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9892                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9893                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9894                                                 root_objectid,
9895                                                 new_dentry->d_name.name,
9896                                                 new_dentry->d_name.len);
9897                         BUG_ON(new_inode->i_nlink == 0);
9898                 } else {
9899                         ret = btrfs_unlink_inode(trans, dest, new_dir,
9900                                                  d_inode(new_dentry),
9901                                                  new_dentry->d_name.name,
9902                                                  new_dentry->d_name.len);
9903                 }
9904                 if (!ret && new_inode->i_nlink == 0)
9905                         ret = btrfs_orphan_add(trans, d_inode(new_dentry));
9906                 if (ret) {
9907                         btrfs_abort_transaction(trans, ret);
9908                         goto out_fail;
9909                 }
9910         }
9911
9912         ret = btrfs_add_link(trans, new_dir, old_inode,
9913                              new_dentry->d_name.name,
9914                              new_dentry->d_name.len, 0, index);
9915         if (ret) {
9916                 btrfs_abort_transaction(trans, ret);
9917                 goto out_fail;
9918         }
9919
9920         if (old_inode->i_nlink == 1)
9921                 BTRFS_I(old_inode)->dir_index = index;
9922
9923         if (log_pinned) {
9924                 struct dentry *parent = new_dentry->d_parent;
9925
9926                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9927                 btrfs_end_log_trans(root);
9928                 log_pinned = false;
9929         }
9930
9931         if (flags & RENAME_WHITEOUT) {
9932                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
9933                                                 old_dentry);
9934
9935                 if (ret) {
9936                         btrfs_abort_transaction(trans, ret);
9937                         goto out_fail;
9938                 }
9939         }
9940 out_fail:
9941         /*
9942          * If we have pinned the log and an error happened, we unpin tasks
9943          * trying to sync the log and force them to fallback to a transaction
9944          * commit if the log currently contains any of the inodes involved in
9945          * this rename operation (to ensure we do not persist a log with an
9946          * inconsistent state for any of these inodes or leading to any
9947          * inconsistencies when replayed). If the transaction was aborted, the
9948          * abortion reason is propagated to userspace when attempting to commit
9949          * the transaction. If the log does not contain any of these inodes, we
9950          * allow the tasks to sync it.
9951          */
9952         if (ret && log_pinned) {
9953                 if (btrfs_inode_in_log(old_dir, root->fs_info->generation) ||
9954                     btrfs_inode_in_log(new_dir, root->fs_info->generation) ||
9955                     btrfs_inode_in_log(old_inode, root->fs_info->generation) ||
9956                     (new_inode &&
9957                      btrfs_inode_in_log(new_inode, root->fs_info->generation)))
9958                     btrfs_set_log_full_commit(root->fs_info, trans);
9959
9960                 btrfs_end_log_trans(root);
9961                 log_pinned = false;
9962         }
9963         btrfs_end_transaction(trans, root);
9964 out_notrans:
9965         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9966                 up_read(&root->fs_info->subvol_sem);
9967
9968         return ret;
9969 }
9970
9971 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9972                          struct inode *new_dir, struct dentry *new_dentry,
9973                          unsigned int flags)
9974 {
9975         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
9976                 return -EINVAL;
9977
9978         if (flags & RENAME_EXCHANGE)
9979                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
9980                                           new_dentry);
9981
9982         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
9983 }
9984
9985 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9986 {
9987         struct btrfs_delalloc_work *delalloc_work;
9988         struct inode *inode;
9989
9990         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9991                                      work);
9992         inode = delalloc_work->inode;
9993         filemap_flush(inode->i_mapping);
9994         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9995                                 &BTRFS_I(inode)->runtime_flags))
9996                 filemap_flush(inode->i_mapping);
9997
9998         if (delalloc_work->delay_iput)
9999                 btrfs_add_delayed_iput(inode);
10000         else
10001                 iput(inode);
10002         complete(&delalloc_work->completion);
10003 }
10004
10005 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
10006                                                     int delay_iput)
10007 {
10008         struct btrfs_delalloc_work *work;
10009
10010         work = kmalloc(sizeof(*work), GFP_NOFS);
10011         if (!work)
10012                 return NULL;
10013
10014         init_completion(&work->completion);
10015         INIT_LIST_HEAD(&work->list);
10016         work->inode = inode;
10017         work->delay_iput = delay_iput;
10018         WARN_ON_ONCE(!inode);
10019         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
10020                         btrfs_run_delalloc_work, NULL, NULL);
10021
10022         return work;
10023 }
10024
10025 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
10026 {
10027         wait_for_completion(&work->completion);
10028         kfree(work);
10029 }
10030
10031 /*
10032  * some fairly slow code that needs optimization. This walks the list
10033  * of all the inodes with pending delalloc and forces them to disk.
10034  */
10035 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
10036                                    int nr)
10037 {
10038         struct btrfs_inode *binode;
10039         struct inode *inode;
10040         struct btrfs_delalloc_work *work, *next;
10041         struct list_head works;
10042         struct list_head splice;
10043         int ret = 0;
10044
10045         INIT_LIST_HEAD(&works);
10046         INIT_LIST_HEAD(&splice);
10047
10048         mutex_lock(&root->delalloc_mutex);
10049         spin_lock(&root->delalloc_lock);
10050         list_splice_init(&root->delalloc_inodes, &splice);
10051         while (!list_empty(&splice)) {
10052                 binode = list_entry(splice.next, struct btrfs_inode,
10053                                     delalloc_inodes);
10054
10055                 list_move_tail(&binode->delalloc_inodes,
10056                                &root->delalloc_inodes);
10057                 inode = igrab(&binode->vfs_inode);
10058                 if (!inode) {
10059                         cond_resched_lock(&root->delalloc_lock);
10060                         continue;
10061                 }
10062                 spin_unlock(&root->delalloc_lock);
10063
10064                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
10065                 if (!work) {
10066                         if (delay_iput)
10067                                 btrfs_add_delayed_iput(inode);
10068                         else
10069                                 iput(inode);
10070                         ret = -ENOMEM;
10071                         goto out;
10072                 }
10073                 list_add_tail(&work->list, &works);
10074                 btrfs_queue_work(root->fs_info->flush_workers,
10075                                  &work->work);
10076                 ret++;
10077                 if (nr != -1 && ret >= nr)
10078                         goto out;
10079                 cond_resched();
10080                 spin_lock(&root->delalloc_lock);
10081         }
10082         spin_unlock(&root->delalloc_lock);
10083
10084 out:
10085         list_for_each_entry_safe(work, next, &works, list) {
10086                 list_del_init(&work->list);
10087                 btrfs_wait_and_free_delalloc_work(work);
10088         }
10089
10090         if (!list_empty_careful(&splice)) {
10091                 spin_lock(&root->delalloc_lock);
10092                 list_splice_tail(&splice, &root->delalloc_inodes);
10093                 spin_unlock(&root->delalloc_lock);
10094         }
10095         mutex_unlock(&root->delalloc_mutex);
10096         return ret;
10097 }
10098
10099 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
10100 {
10101         int ret;
10102
10103         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
10104                 return -EROFS;
10105
10106         ret = __start_delalloc_inodes(root, delay_iput, -1);
10107         if (ret > 0)
10108                 ret = 0;
10109         /*
10110          * the filemap_flush will queue IO into the worker threads, but
10111          * we have to make sure the IO is actually started and that
10112          * ordered extents get created before we return
10113          */
10114         atomic_inc(&root->fs_info->async_submit_draining);
10115         while (atomic_read(&root->fs_info->nr_async_submits) ||
10116               atomic_read(&root->fs_info->async_delalloc_pages)) {
10117                 wait_event(root->fs_info->async_submit_wait,
10118                    (atomic_read(&root->fs_info->nr_async_submits) == 0 &&
10119                     atomic_read(&root->fs_info->async_delalloc_pages) == 0));
10120         }
10121         atomic_dec(&root->fs_info->async_submit_draining);
10122         return ret;
10123 }
10124
10125 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
10126                                int nr)
10127 {
10128         struct btrfs_root *root;
10129         struct list_head splice;
10130         int ret;
10131
10132         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10133                 return -EROFS;
10134
10135         INIT_LIST_HEAD(&splice);
10136
10137         mutex_lock(&fs_info->delalloc_root_mutex);
10138         spin_lock(&fs_info->delalloc_root_lock);
10139         list_splice_init(&fs_info->delalloc_roots, &splice);
10140         while (!list_empty(&splice) && nr) {
10141                 root = list_first_entry(&splice, struct btrfs_root,
10142                                         delalloc_root);
10143                 root = btrfs_grab_fs_root(root);
10144                 BUG_ON(!root);
10145                 list_move_tail(&root->delalloc_root,
10146                                &fs_info->delalloc_roots);
10147                 spin_unlock(&fs_info->delalloc_root_lock);
10148
10149                 ret = __start_delalloc_inodes(root, delay_iput, nr);
10150                 btrfs_put_fs_root(root);
10151                 if (ret < 0)
10152                         goto out;
10153
10154                 if (nr != -1) {
10155                         nr -= ret;
10156                         WARN_ON(nr < 0);
10157                 }
10158                 spin_lock(&fs_info->delalloc_root_lock);
10159         }
10160         spin_unlock(&fs_info->delalloc_root_lock);
10161
10162         ret = 0;
10163         atomic_inc(&fs_info->async_submit_draining);
10164         while (atomic_read(&fs_info->nr_async_submits) ||
10165               atomic_read(&fs_info->async_delalloc_pages)) {
10166                 wait_event(fs_info->async_submit_wait,
10167                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
10168                     atomic_read(&fs_info->async_delalloc_pages) == 0));
10169         }
10170         atomic_dec(&fs_info->async_submit_draining);
10171 out:
10172         if (!list_empty_careful(&splice)) {
10173                 spin_lock(&fs_info->delalloc_root_lock);
10174                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10175                 spin_unlock(&fs_info->delalloc_root_lock);
10176         }
10177         mutex_unlock(&fs_info->delalloc_root_mutex);
10178         return ret;
10179 }
10180
10181 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10182                          const char *symname)
10183 {
10184         struct btrfs_trans_handle *trans;
10185         struct btrfs_root *root = BTRFS_I(dir)->root;
10186         struct btrfs_path *path;
10187         struct btrfs_key key;
10188         struct inode *inode = NULL;
10189         int err;
10190         int drop_inode = 0;
10191         u64 objectid;
10192         u64 index = 0;
10193         int name_len;
10194         int datasize;
10195         unsigned long ptr;
10196         struct btrfs_file_extent_item *ei;
10197         struct extent_buffer *leaf;
10198
10199         name_len = strlen(symname);
10200         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(root))
10201                 return -ENAMETOOLONG;
10202
10203         /*
10204          * 2 items for inode item and ref
10205          * 2 items for dir items
10206          * 1 item for updating parent inode item
10207          * 1 item for the inline extent item
10208          * 1 item for xattr if selinux is on
10209          */
10210         trans = btrfs_start_transaction(root, 7);
10211         if (IS_ERR(trans))
10212                 return PTR_ERR(trans);
10213
10214         err = btrfs_find_free_ino(root, &objectid);
10215         if (err)
10216                 goto out_unlock;
10217
10218         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10219                                 dentry->d_name.len, btrfs_ino(dir), objectid,
10220                                 S_IFLNK|S_IRWXUGO, &index);
10221         if (IS_ERR(inode)) {
10222                 err = PTR_ERR(inode);
10223                 goto out_unlock;
10224         }
10225
10226         /*
10227         * If the active LSM wants to access the inode during
10228         * d_instantiate it needs these. Smack checks to see
10229         * if the filesystem supports xattrs by looking at the
10230         * ops vector.
10231         */
10232         inode->i_fop = &btrfs_file_operations;
10233         inode->i_op = &btrfs_file_inode_operations;
10234         inode->i_mapping->a_ops = &btrfs_aops;
10235         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10236
10237         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10238         if (err)
10239                 goto out_unlock_inode;
10240
10241         path = btrfs_alloc_path();
10242         if (!path) {
10243                 err = -ENOMEM;
10244                 goto out_unlock_inode;
10245         }
10246         key.objectid = btrfs_ino(inode);
10247         key.offset = 0;
10248         key.type = BTRFS_EXTENT_DATA_KEY;
10249         datasize = btrfs_file_extent_calc_inline_size(name_len);
10250         err = btrfs_insert_empty_item(trans, root, path, &key,
10251                                       datasize);
10252         if (err) {
10253                 btrfs_free_path(path);
10254                 goto out_unlock_inode;
10255         }
10256         leaf = path->nodes[0];
10257         ei = btrfs_item_ptr(leaf, path->slots[0],
10258                             struct btrfs_file_extent_item);
10259         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10260         btrfs_set_file_extent_type(leaf, ei,
10261                                    BTRFS_FILE_EXTENT_INLINE);
10262         btrfs_set_file_extent_encryption(leaf, ei, 0);
10263         btrfs_set_file_extent_compression(leaf, ei, 0);
10264         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10265         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10266
10267         ptr = btrfs_file_extent_inline_start(ei);
10268         write_extent_buffer(leaf, symname, ptr, name_len);
10269         btrfs_mark_buffer_dirty(leaf);
10270         btrfs_free_path(path);
10271
10272         inode->i_op = &btrfs_symlink_inode_operations;
10273         inode_nohighmem(inode);
10274         inode->i_mapping->a_ops = &btrfs_symlink_aops;
10275         inode_set_bytes(inode, name_len);
10276         btrfs_i_size_write(inode, name_len);
10277         err = btrfs_update_inode(trans, root, inode);
10278         /*
10279          * Last step, add directory indexes for our symlink inode. This is the
10280          * last step to avoid extra cleanup of these indexes if an error happens
10281          * elsewhere above.
10282          */
10283         if (!err)
10284                 err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
10285         if (err) {
10286                 drop_inode = 1;
10287                 goto out_unlock_inode;
10288         }
10289
10290         unlock_new_inode(inode);
10291         d_instantiate(dentry, inode);
10292
10293 out_unlock:
10294         btrfs_end_transaction(trans, root);
10295         if (drop_inode) {
10296                 inode_dec_link_count(inode);
10297                 iput(inode);
10298         }
10299         btrfs_btree_balance_dirty(root);
10300         return err;
10301
10302 out_unlock_inode:
10303         drop_inode = 1;
10304         unlock_new_inode(inode);
10305         goto out_unlock;
10306 }
10307
10308 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10309                                        u64 start, u64 num_bytes, u64 min_size,
10310                                        loff_t actual_len, u64 *alloc_hint,
10311                                        struct btrfs_trans_handle *trans)
10312 {
10313         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10314         struct extent_map *em;
10315         struct btrfs_root *root = BTRFS_I(inode)->root;
10316         struct btrfs_key ins;
10317         u64 cur_offset = start;
10318         u64 i_size;
10319         u64 cur_bytes;
10320         u64 last_alloc = (u64)-1;
10321         int ret = 0;
10322         bool own_trans = true;
10323         u64 end = start + num_bytes - 1;
10324
10325         if (trans)
10326                 own_trans = false;
10327         while (num_bytes > 0) {
10328                 if (own_trans) {
10329                         trans = btrfs_start_transaction(root, 3);
10330                         if (IS_ERR(trans)) {
10331                                 ret = PTR_ERR(trans);
10332                                 break;
10333                         }
10334                 }
10335
10336                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10337                 cur_bytes = max(cur_bytes, min_size);
10338                 /*
10339                  * If we are severely fragmented we could end up with really
10340                  * small allocations, so if the allocator is returning small
10341                  * chunks lets make its job easier by only searching for those
10342                  * sized chunks.
10343                  */
10344                 cur_bytes = min(cur_bytes, last_alloc);
10345                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10346                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10347                 if (ret) {
10348                         if (own_trans)
10349                                 btrfs_end_transaction(trans, root);
10350                         break;
10351                 }
10352                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
10353
10354                 last_alloc = ins.offset;
10355                 ret = insert_reserved_file_extent(trans, inode,
10356                                                   cur_offset, ins.objectid,
10357                                                   ins.offset, ins.offset,
10358                                                   ins.offset, 0, 0, 0,
10359                                                   BTRFS_FILE_EXTENT_PREALLOC);
10360                 if (ret) {
10361                         btrfs_free_reserved_extent(root, ins.objectid,
10362                                                    ins.offset, 0);
10363                         btrfs_abort_transaction(trans, ret);
10364                         if (own_trans)
10365                                 btrfs_end_transaction(trans, root);
10366                         break;
10367                 }
10368
10369                 btrfs_drop_extent_cache(inode, cur_offset,
10370                                         cur_offset + ins.offset -1, 0);
10371
10372                 em = alloc_extent_map();
10373                 if (!em) {
10374                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10375                                 &BTRFS_I(inode)->runtime_flags);
10376                         goto next;
10377                 }
10378
10379                 em->start = cur_offset;
10380                 em->orig_start = cur_offset;
10381                 em->len = ins.offset;
10382                 em->block_start = ins.objectid;
10383                 em->block_len = ins.offset;
10384                 em->orig_block_len = ins.offset;
10385                 em->ram_bytes = ins.offset;
10386                 em->bdev = root->fs_info->fs_devices->latest_bdev;
10387                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10388                 em->generation = trans->transid;
10389
10390                 while (1) {
10391                         write_lock(&em_tree->lock);
10392                         ret = add_extent_mapping(em_tree, em, 1);
10393                         write_unlock(&em_tree->lock);
10394                         if (ret != -EEXIST)
10395                                 break;
10396                         btrfs_drop_extent_cache(inode, cur_offset,
10397                                                 cur_offset + ins.offset - 1,
10398                                                 0);
10399                 }
10400                 free_extent_map(em);
10401 next:
10402                 num_bytes -= ins.offset;
10403                 cur_offset += ins.offset;
10404                 *alloc_hint = ins.objectid + ins.offset;
10405
10406                 inode_inc_iversion(inode);
10407                 inode->i_ctime = current_fs_time(inode->i_sb);
10408                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10409                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10410                     (actual_len > inode->i_size) &&
10411                     (cur_offset > inode->i_size)) {
10412                         if (cur_offset > actual_len)
10413                                 i_size = actual_len;
10414                         else
10415                                 i_size = cur_offset;
10416                         i_size_write(inode, i_size);
10417                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10418                 }
10419
10420                 ret = btrfs_update_inode(trans, root, inode);
10421
10422                 if (ret) {
10423                         btrfs_abort_transaction(trans, ret);
10424                         if (own_trans)
10425                                 btrfs_end_transaction(trans, root);
10426                         break;
10427                 }
10428
10429                 if (own_trans)
10430                         btrfs_end_transaction(trans, root);
10431         }
10432         if (cur_offset < end)
10433                 btrfs_free_reserved_data_space(inode, cur_offset,
10434                         end - cur_offset + 1);
10435         return ret;
10436 }
10437
10438 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10439                               u64 start, u64 num_bytes, u64 min_size,
10440                               loff_t actual_len, u64 *alloc_hint)
10441 {
10442         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10443                                            min_size, actual_len, alloc_hint,
10444                                            NULL);
10445 }
10446
10447 int btrfs_prealloc_file_range_trans(struct inode *inode,
10448                                     struct btrfs_trans_handle *trans, int mode,
10449                                     u64 start, u64 num_bytes, u64 min_size,
10450                                     loff_t actual_len, u64 *alloc_hint)
10451 {
10452         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10453                                            min_size, actual_len, alloc_hint, trans);
10454 }
10455
10456 static int btrfs_set_page_dirty(struct page *page)
10457 {
10458         return __set_page_dirty_nobuffers(page);
10459 }
10460
10461 static int btrfs_permission(struct inode *inode, int mask)
10462 {
10463         struct btrfs_root *root = BTRFS_I(inode)->root;
10464         umode_t mode = inode->i_mode;
10465
10466         if (mask & MAY_WRITE &&
10467             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10468                 if (btrfs_root_readonly(root))
10469                         return -EROFS;
10470                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10471                         return -EACCES;
10472         }
10473         return generic_permission(inode, mask);
10474 }
10475
10476 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10477 {
10478         struct btrfs_trans_handle *trans;
10479         struct btrfs_root *root = BTRFS_I(dir)->root;
10480         struct inode *inode = NULL;
10481         u64 objectid;
10482         u64 index;
10483         int ret = 0;
10484
10485         /*
10486          * 5 units required for adding orphan entry
10487          */
10488         trans = btrfs_start_transaction(root, 5);
10489         if (IS_ERR(trans))
10490                 return PTR_ERR(trans);
10491
10492         ret = btrfs_find_free_ino(root, &objectid);
10493         if (ret)
10494                 goto out;
10495
10496         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10497                                 btrfs_ino(dir), objectid, mode, &index);
10498         if (IS_ERR(inode)) {
10499                 ret = PTR_ERR(inode);
10500                 inode = NULL;
10501                 goto out;
10502         }
10503
10504         inode->i_fop = &btrfs_file_operations;
10505         inode->i_op = &btrfs_file_inode_operations;
10506
10507         inode->i_mapping->a_ops = &btrfs_aops;
10508         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10509
10510         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10511         if (ret)
10512                 goto out_inode;
10513
10514         ret = btrfs_update_inode(trans, root, inode);
10515         if (ret)
10516                 goto out_inode;
10517         ret = btrfs_orphan_add(trans, inode);
10518         if (ret)
10519                 goto out_inode;
10520
10521         /*
10522          * We set number of links to 0 in btrfs_new_inode(), and here we set
10523          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10524          * through:
10525          *
10526          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10527          */
10528         set_nlink(inode, 1);
10529         unlock_new_inode(inode);
10530         d_tmpfile(dentry, inode);
10531         mark_inode_dirty(inode);
10532
10533 out:
10534         btrfs_end_transaction(trans, root);
10535         if (ret)
10536                 iput(inode);
10537         btrfs_balance_delayed_items(root);
10538         btrfs_btree_balance_dirty(root);
10539         return ret;
10540
10541 out_inode:
10542         unlock_new_inode(inode);
10543         goto out;
10544
10545 }
10546
10547 /* Inspired by filemap_check_errors() */
10548 int btrfs_inode_check_errors(struct inode *inode)
10549 {
10550         int ret = 0;
10551
10552         if (test_bit(AS_ENOSPC, &inode->i_mapping->flags) &&
10553             test_and_clear_bit(AS_ENOSPC, &inode->i_mapping->flags))
10554                 ret = -ENOSPC;
10555         if (test_bit(AS_EIO, &inode->i_mapping->flags) &&
10556             test_and_clear_bit(AS_EIO, &inode->i_mapping->flags))
10557                 ret = -EIO;
10558
10559         return ret;
10560 }
10561
10562 static const struct inode_operations btrfs_dir_inode_operations = {
10563         .getattr        = btrfs_getattr,
10564         .lookup         = btrfs_lookup,
10565         .create         = btrfs_create,
10566         .unlink         = btrfs_unlink,
10567         .link           = btrfs_link,
10568         .mkdir          = btrfs_mkdir,
10569         .rmdir          = btrfs_rmdir,
10570         .rename2        = btrfs_rename2,
10571         .symlink        = btrfs_symlink,
10572         .setattr        = btrfs_setattr,
10573         .mknod          = btrfs_mknod,
10574         .setxattr       = generic_setxattr,
10575         .getxattr       = generic_getxattr,
10576         .listxattr      = btrfs_listxattr,
10577         .removexattr    = generic_removexattr,
10578         .permission     = btrfs_permission,
10579         .get_acl        = btrfs_get_acl,
10580         .set_acl        = btrfs_set_acl,
10581         .update_time    = btrfs_update_time,
10582         .tmpfile        = btrfs_tmpfile,
10583 };
10584 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10585         .lookup         = btrfs_lookup,
10586         .permission     = btrfs_permission,
10587         .get_acl        = btrfs_get_acl,
10588         .set_acl        = btrfs_set_acl,
10589         .update_time    = btrfs_update_time,
10590 };
10591
10592 static const struct file_operations btrfs_dir_file_operations = {
10593         .llseek         = generic_file_llseek,
10594         .read           = generic_read_dir,
10595         .iterate_shared = btrfs_real_readdir,
10596         .unlocked_ioctl = btrfs_ioctl,
10597 #ifdef CONFIG_COMPAT
10598         .compat_ioctl   = btrfs_compat_ioctl,
10599 #endif
10600         .release        = btrfs_release_file,
10601         .fsync          = btrfs_sync_file,
10602 };
10603
10604 static const struct extent_io_ops btrfs_extent_io_ops = {
10605         .fill_delalloc = run_delalloc_range,
10606         .submit_bio_hook = btrfs_submit_bio_hook,
10607         .merge_bio_hook = btrfs_merge_bio_hook,
10608         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10609         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10610         .writepage_start_hook = btrfs_writepage_start_hook,
10611         .set_bit_hook = btrfs_set_bit_hook,
10612         .clear_bit_hook = btrfs_clear_bit_hook,
10613         .merge_extent_hook = btrfs_merge_extent_hook,
10614         .split_extent_hook = btrfs_split_extent_hook,
10615 };
10616
10617 /*
10618  * btrfs doesn't support the bmap operation because swapfiles
10619  * use bmap to make a mapping of extents in the file.  They assume
10620  * these extents won't change over the life of the file and they
10621  * use the bmap result to do IO directly to the drive.
10622  *
10623  * the btrfs bmap call would return logical addresses that aren't
10624  * suitable for IO and they also will change frequently as COW
10625  * operations happen.  So, swapfile + btrfs == corruption.
10626  *
10627  * For now we're avoiding this by dropping bmap.
10628  */
10629 static const struct address_space_operations btrfs_aops = {
10630         .readpage       = btrfs_readpage,
10631         .writepage      = btrfs_writepage,
10632         .writepages     = btrfs_writepages,
10633         .readpages      = btrfs_readpages,
10634         .direct_IO      = btrfs_direct_IO,
10635         .invalidatepage = btrfs_invalidatepage,
10636         .releasepage    = btrfs_releasepage,
10637         .set_page_dirty = btrfs_set_page_dirty,
10638         .error_remove_page = generic_error_remove_page,
10639 };
10640
10641 static const struct address_space_operations btrfs_symlink_aops = {
10642         .readpage       = btrfs_readpage,
10643         .writepage      = btrfs_writepage,
10644         .invalidatepage = btrfs_invalidatepage,
10645         .releasepage    = btrfs_releasepage,
10646 };
10647
10648 static const struct inode_operations btrfs_file_inode_operations = {
10649         .getattr        = btrfs_getattr,
10650         .setattr        = btrfs_setattr,
10651         .setxattr       = generic_setxattr,
10652         .getxattr       = generic_getxattr,
10653         .listxattr      = btrfs_listxattr,
10654         .removexattr    = generic_removexattr,
10655         .permission     = btrfs_permission,
10656         .fiemap         = btrfs_fiemap,
10657         .get_acl        = btrfs_get_acl,
10658         .set_acl        = btrfs_set_acl,
10659         .update_time    = btrfs_update_time,
10660 };
10661 static const struct inode_operations btrfs_special_inode_operations = {
10662         .getattr        = btrfs_getattr,
10663         .setattr        = btrfs_setattr,
10664         .permission     = btrfs_permission,
10665         .setxattr       = generic_setxattr,
10666         .getxattr       = generic_getxattr,
10667         .listxattr      = btrfs_listxattr,
10668         .removexattr    = generic_removexattr,
10669         .get_acl        = btrfs_get_acl,
10670         .set_acl        = btrfs_set_acl,
10671         .update_time    = btrfs_update_time,
10672 };
10673 static const struct inode_operations btrfs_symlink_inode_operations = {
10674         .readlink       = generic_readlink,
10675         .get_link       = page_get_link,
10676         .getattr        = btrfs_getattr,
10677         .setattr        = btrfs_setattr,
10678         .permission     = btrfs_permission,
10679         .setxattr       = generic_setxattr,
10680         .getxattr       = generic_getxattr,
10681         .listxattr      = btrfs_listxattr,
10682         .removexattr    = generic_removexattr,
10683         .update_time    = btrfs_update_time,
10684 };
10685
10686 const struct dentry_operations btrfs_dentry_operations = {
10687         .d_delete       = btrfs_dentry_delete,
10688         .d_release      = btrfs_dentry_release,
10689 };