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