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