fs/btrfs/inode.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/kernel.h>
   7 #include <linux/bio.h>
   8 #include <linux/buffer_head.h>
   9 #include <linux/file.h>
  10 #include <linux/fs.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/highmem.h>
  13 #include <linux/time.h>
  14 #include <linux/init.h>
  15 #include <linux/string.h>
  16 #include <linux/backing-dev.h>
  17 #include <linux/writeback.h>
  18 #include <linux/compat.h>
  19 #include <linux/xattr.h>
  20 #include <linux/posix_acl.h>
  21 #include <linux/falloc.h>
  22 #include <linux/slab.h>
  23 #include <linux/ratelimit.h>
  24 #include <linux/btrfs.h>
  25 #include <linux/blkdev.h>
  26 #include <linux/posix_acl_xattr.h>
  27 #include <linux/uio.h>
  28 #include <linux/magic.h>
  29 #include <linux/iversion.h>
  30 #include <asm/unaligned.h>
  31 #include "ctree.h"
  32 #include "disk-io.h"
  33 #include "transaction.h"
  34 #include "btrfs_inode.h"
  35 #include "print-tree.h"
  36 #include "ordered-data.h"
  37 #include "xattr.h"
  38 #include "tree-log.h"
  39 #include "volumes.h"
  40 #include "compression.h"
  41 #include "locking.h"
  42 #include "free-space-cache.h"
  43 #include "inode-map.h"
  44 #include "backref.h"
  45 #include "props.h"
  46 #include "qgroup.h"
  47 #include "dedupe.h"
  48
  49 struct btrfs_iget_args {
  50         struct btrfs_key *location;
  51         struct btrfs_root *root;
  52 };
  53
  54 struct btrfs_dio_data {
  55         u64 reserve;
  56         u64 unsubmitted_oe_range_start;
  57         u64 unsubmitted_oe_range_end;
  58         int overwrite;
  59 };
  60
  61 static const struct inode_operations btrfs_dir_inode_operations;
  62 static const struct inode_operations btrfs_symlink_inode_operations;
  63 static const struct inode_operations btrfs_dir_ro_inode_operations;
  64 static const struct inode_operations btrfs_special_inode_operations;
  65 static const struct inode_operations btrfs_file_inode_operations;
  66 static const struct address_space_operations btrfs_aops;
  67 static const struct address_space_operations btrfs_symlink_aops;
  68 static const struct file_operations btrfs_dir_file_operations;
  69 static const struct extent_io_ops btrfs_extent_io_ops;
  70
  71 static struct kmem_cache *btrfs_inode_cachep;
  72 struct kmem_cache *btrfs_trans_handle_cachep;
  73 struct kmem_cache *btrfs_path_cachep;
  74 struct kmem_cache *btrfs_free_space_cachep;
  75
  76 #define S_SHIFT 12
  77 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  78         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  79         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  80         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  81         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  82         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  83         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
  84         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
  85 };
  86
  87 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
  88 static int btrfs_truncate(struct inode *inode, bool skip_writeback);
  89 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
  90 static noinline int cow_file_range(struct inode *inode,
  91                                    struct page *locked_page,
  92                                    u64 start, u64 end, u64 delalloc_end,
  93                                    int *page_started, unsigned long *nr_written,
  94                                    int unlock, struct btrfs_dedupe_hash *hash);
  95 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
  96                                        u64 orig_start, u64 block_start,
  97                                        u64 block_len, u64 orig_block_len,
  98                                        u64 ram_bytes, int compress_type,
  99                                        int type);
 100
 101 static void __endio_write_update_ordered(struct inode *inode,
 102                                          const u64 offset, const u64 bytes,
 103                                          const bool uptodate);
 104
 105 /*
 106  * Cleanup all submitted ordered extents in specified range to handle errors
 107  * from the fill_dellaloc() callback.
 108  *
 109  * NOTE: caller must ensure that when an error happens, it can not call
 110  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 111  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 112  * to be released, which we want to happen only when finishing the ordered
 113  * extent (btrfs_finish_ordered_io()). Also note that the caller of the
 114  * fill_delalloc() callback already does proper cleanup for the first page of
 115  * the range, that is, it invokes the callback writepage_end_io_hook() for the
 116  * range of the first page.
 117  */
 118 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 119                                                  const u64 offset,
 120                                                  const u64 bytes)
 121 {
 122         unsigned long index = offset >> PAGE_SHIFT;
 123         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
 124         struct page *page;
 125
 126         while (index <= end_index) {
 127                 page = find_get_page(inode->i_mapping, index);
 128                 index++;
 129                 if (!page)
 130                         continue;
 131                 ClearPagePrivate2(page);
 132                 put_page(page);
 133         }
 134         return __endio_write_update_ordered(inode, offset + PAGE_SIZE,
 135                                             bytes - PAGE_SIZE, false);
 136 }
 137
 138 static int btrfs_dirty_inode(struct inode *inode);
 139
 140 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 141 void btrfs_test_inode_set_ops(struct inode *inode)
 142 {
 143         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 144 }
 145 #endif
 146
 147 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 148                                      struct inode *inode,  struct inode *dir,
 149                                      const struct qstr *qstr)
 150 {
 151         int err;
 152
 153         err = btrfs_init_acl(trans, inode, dir);
 154         if (!err)
 155                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 156         return err;
 157 }
 158
 159 /*
 160  * this does all the hard work for inserting an inline extent into
 161  * the btree.  The caller should have done a btrfs_drop_extents so that
 162  * no overlapping inline items exist in the btree
 163  */
 164 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 165                                 struct btrfs_path *path, int extent_inserted,
 166                                 struct btrfs_root *root, struct inode *inode,
 167                                 u64 start, size_t size, size_t compressed_size,
 168                                 int compress_type,
 169                                 struct page **compressed_pages)
 170 {
 171         struct extent_buffer *leaf;
 172         struct page *page = NULL;
 173         char *kaddr;
 174         unsigned long ptr;
 175         struct btrfs_file_extent_item *ei;
 176         int ret;
 177         size_t cur_size = size;
 178         unsigned long offset;
 179
 180         if (compressed_size && compressed_pages)
 181                 cur_size = compressed_size;
 182
 183         inode_add_bytes(inode, size);
 184
 185         if (!extent_inserted) {
 186                 struct btrfs_key key;
 187                 size_t datasize;
 188
 189                 key.objectid = btrfs_ino(BTRFS_I(inode));
 190                 key.offset = start;
 191                 key.type = BTRFS_EXTENT_DATA_KEY;
 192
 193                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 194                 path->leave_spinning = 1;
 195                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 196                                               datasize);
 197                 if (ret)
 198                         goto fail;
 199         }
 200         leaf = path->nodes[0];
 201         ei = btrfs_item_ptr(leaf, path->slots[0],
 202                             struct btrfs_file_extent_item);
 203         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 204         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 205         btrfs_set_file_extent_encryption(leaf, ei, 0);
 206         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 207         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 208         ptr = btrfs_file_extent_inline_start(ei);
 209
 210         if (compress_type != BTRFS_COMPRESS_NONE) {
 211                 struct page *cpage;
 212                 int i = 0;
 213                 while (compressed_size > 0) {
 214                         cpage = compressed_pages[i];
 215                         cur_size = min_t(unsigned long, compressed_size,
 216                                        PAGE_SIZE);
 217
 218                         kaddr = kmap_atomic(cpage);
 219                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 220                         kunmap_atomic(kaddr);
 221
 222                         i++;
 223                         ptr += cur_size;
 224                         compressed_size -= cur_size;
 225                 }
 226                 btrfs_set_file_extent_compression(leaf, ei,
 227                                                   compress_type);
 228         } else {
 229                 page = find_get_page(inode->i_mapping,
 230                                      start >> PAGE_SHIFT);
 231                 btrfs_set_file_extent_compression(leaf, ei, 0);
 232                 kaddr = kmap_atomic(page);
 233                 offset = start & (PAGE_SIZE - 1);
 234                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 235                 kunmap_atomic(kaddr);
 236                 put_page(page);
 237         }
 238         btrfs_mark_buffer_dirty(leaf);
 239         btrfs_release_path(path);
 240
 241         /*
 242          * we're an inline extent, so nobody can
 243          * extend the file past i_size without locking
 244          * a page we already have locked.
 245          *
 246          * We must do any isize and inode updates
 247          * before we unlock the pages.  Otherwise we
 248          * could end up racing with unlink.
 249          */
 250         BTRFS_I(inode)->disk_i_size = inode->i_size;
 251         ret = btrfs_update_inode(trans, root, inode);
 252
 253 fail:
 254         return ret;
 255 }
 256
 257
 258 /*
 259  * conditionally insert an inline extent into the file.  This
 260  * does the checks required to make sure the data is small enough
 261  * to fit as an inline extent.
 262  */
 263 static noinline int cow_file_range_inline(struct inode *inode, u64 start,
 264                                           u64 end, size_t compressed_size,
 265                                           int compress_type,
 266                                           struct page **compressed_pages)
 267 {
 268         struct btrfs_root *root = BTRFS_I(inode)->root;
 269         struct btrfs_fs_info *fs_info = root->fs_info;
 270         struct btrfs_trans_handle *trans;
 271         u64 isize = i_size_read(inode);
 272         u64 actual_end = min(end + 1, isize);
 273         u64 inline_len = actual_end - start;
 274         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 275         u64 data_len = inline_len;
 276         int ret;
 277         struct btrfs_path *path;
 278         int extent_inserted = 0;
 279         u32 extent_item_size;
 280
 281         if (compressed_size)
 282                 data_len = compressed_size;
 283
 284         if (start > 0 ||
 285             actual_end > fs_info->sectorsize ||
 286             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 287             (!compressed_size &&
 288             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 289             end + 1 < isize ||
 290             data_len > fs_info->max_inline) {
 291                 return 1;
 292         }
 293
 294         path = btrfs_alloc_path();
 295         if (!path)
 296                 return -ENOMEM;
 297
 298         trans = btrfs_join_transaction(root);
 299         if (IS_ERR(trans)) {
 300                 btrfs_free_path(path);
 301                 return PTR_ERR(trans);
 302         }
 303         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
 304
 305         if (compressed_size && compressed_pages)
 306                 extent_item_size = btrfs_file_extent_calc_inline_size(
 307                    compressed_size);
 308         else
 309                 extent_item_size = btrfs_file_extent_calc_inline_size(
 310                     inline_len);
 311
 312         ret = __btrfs_drop_extents(trans, root, inode, path,
 313                                    start, aligned_end, NULL,
 314                                    1, 1, extent_item_size, &extent_inserted);
 315         if (ret) {
 316                 btrfs_abort_transaction(trans, ret);
 317                 goto out;
 318         }
 319
 320         if (isize > actual_end)
 321                 inline_len = min_t(u64, isize, actual_end);
 322         ret = insert_inline_extent(trans, path, extent_inserted,
 323                                    root, inode, start,
 324                                    inline_len, compressed_size,
 325                                    compress_type, compressed_pages);
 326         if (ret && ret != -ENOSPC) {
 327                 btrfs_abort_transaction(trans, ret);
 328                 goto out;
 329         } else if (ret == -ENOSPC) {
 330                 ret = 1;
 331                 goto out;
 332         }
 333
 334         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 335         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 336 out:
 337         /*
 338          * Don't forget to free the reserved space, as for inlined extent
 339          * it won't count as data extent, free them directly here.
 340          * And at reserve time, it's always aligned to page size, so
 341          * just free one page here.
 342          */
 343         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
 344         btrfs_free_path(path);
 345         btrfs_end_transaction(trans);
 346         return ret;
 347 }
 348
 349 struct async_extent {
 350         u64 start;
 351         u64 ram_size;
 352         u64 compressed_size;
 353         struct page **pages;
 354         unsigned long nr_pages;
 355         int compress_type;
 356         struct list_head list;
 357 };
 358
 359 struct async_cow {
 360         struct inode *inode;
 361         struct btrfs_root *root;
 362         struct page *locked_page;
 363         u64 start;
 364         u64 end;
 365         unsigned int write_flags;
 366         struct list_head extents;
 367         struct btrfs_work work;
 368 };
 369
 370 static noinline int add_async_extent(struct async_cow *cow,
 371                                      u64 start, u64 ram_size,
 372                                      u64 compressed_size,
 373                                      struct page **pages,
 374                                      unsigned long nr_pages,
 375                                      int compress_type)
 376 {
 377         struct async_extent *async_extent;
 378
 379         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 380         BUG_ON(!async_extent); /* -ENOMEM */
 381         async_extent->start = start;
 382         async_extent->ram_size = ram_size;
 383         async_extent->compressed_size = compressed_size;
 384         async_extent->pages = pages;
 385         async_extent->nr_pages = nr_pages;
 386         async_extent->compress_type = compress_type;
 387         list_add_tail(&async_extent->list, &cow->extents);
 388         return 0;
 389 }
 390
 391 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
 392 {
 393         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 394
 395         /* force compress */
 396         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 397                 return 1;
 398         /* defrag ioctl */
 399         if (BTRFS_I(inode)->defrag_compress)
 400                 return 1;
 401         /* bad compression ratios */
 402         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 403                 return 0;
 404         if (btrfs_test_opt(fs_info, COMPRESS) ||
 405             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 406             BTRFS_I(inode)->prop_compress)
 407                 return btrfs_compress_heuristic(inode, start, end);
 408         return 0;
 409 }
 410
 411 static inline void inode_should_defrag(struct btrfs_inode *inode,
 412                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 413 {
 414         /* If this is a small write inside eof, kick off a defrag */
 415         if (num_bytes < small_write &&
 416             (start > 0 || end + 1 < inode->disk_i_size))
 417                 btrfs_add_inode_defrag(NULL, inode);
 418 }
 419
 420 /*
 421  * we create compressed extents in two phases.  The first
 422  * phase compresses a range of pages that have already been
 423  * locked (both pages and state bits are locked).
 424  *
 425  * This is done inside an ordered work queue, and the compression
 426  * is spread across many cpus.  The actual IO submission is step
 427  * two, and the ordered work queue takes care of making sure that
 428  * happens in the same order things were put onto the queue by
 429  * writepages and friends.
 430  *
 431  * If this code finds it can't get good compression, it puts an
 432  * entry onto the work queue to write the uncompressed bytes.  This
 433  * makes sure that both compressed inodes and uncompressed inodes
 434  * are written in the same order that the flusher thread sent them
 435  * down.
 436  */
 437 static noinline void compress_file_range(struct inode *inode,
 438                                         struct page *locked_page,
 439                                         u64 start, u64 end,
 440                                         struct async_cow *async_cow,
 441                                         int *num_added)
 442 {
 443         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 444         u64 blocksize = fs_info->sectorsize;
 445         u64 actual_end;
 446         u64 isize = i_size_read(inode);
 447         int ret = 0;
 448         struct page **pages = NULL;
 449         unsigned long nr_pages;
 450         unsigned long total_compressed = 0;
 451         unsigned long total_in = 0;
 452         int i;
 453         int will_compress;
 454         int compress_type = fs_info->compress_type;
 455         int redirty = 0;
 456
 457         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 458                         SZ_16K);
 459
 460         actual_end = min_t(u64, isize, end + 1);
 461 again:
 462         will_compress = 0;
 463         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 464         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 465         nr_pages = min_t(unsigned long, nr_pages,
 466                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 467
 468         /*
 469          * we don't want to send crud past the end of i_size through
 470          * compression, that's just a waste of CPU time.  So, if the
 471          * end of the file is before the start of our current
 472          * requested range of bytes, we bail out to the uncompressed
 473          * cleanup code that can deal with all of this.
 474          *
 475          * It isn't really the fastest way to fix things, but this is a
 476          * very uncommon corner.
 477          */
 478         if (actual_end <= start)
 479                 goto cleanup_and_bail_uncompressed;
 480
 481         total_compressed = actual_end - start;
 482
 483         /*
 484          * skip compression for a small file range(<=blocksize) that
 485          * isn't an inline extent, since it doesn't save disk space at all.
 486          */
 487         if (total_compressed <= blocksize &&
 488            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 489                 goto cleanup_and_bail_uncompressed;
 490
 491         total_compressed = min_t(unsigned long, total_compressed,
 492                         BTRFS_MAX_UNCOMPRESSED);
 493         total_in = 0;
 494         ret = 0;
 495
 496         /*
 497          * we do compression for mount -o compress and when the
 498          * inode has not been flagged as nocompress.  This flag can
 499          * change at any time if we discover bad compression ratios.
 500          */
 501         if (inode_need_compress(inode, start, end)) {
 502                 WARN_ON(pages);
 503                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 504                 if (!pages) {
 505                         /* just bail out to the uncompressed code */
 506                         goto cont;
 507                 }
 508
 509                 if (BTRFS_I(inode)->defrag_compress)
 510                         compress_type = BTRFS_I(inode)->defrag_compress;
 511                 else if (BTRFS_I(inode)->prop_compress)
 512                         compress_type = BTRFS_I(inode)->prop_compress;
 513
 514                 /*
 515                  * we need to call clear_page_dirty_for_io on each
 516                  * page in the range.  Otherwise applications with the file
 517                  * mmap'd can wander in and change the page contents while
 518                  * we are compressing them.
 519                  *
 520                  * If the compression fails for any reason, we set the pages
 521                  * dirty again later on.
 522                  *
 523                  * Note that the remaining part is redirtied, the start pointer
 524                  * has moved, the end is the original one.
 525                  */
 526                 if (!redirty) {
 527                         extent_range_clear_dirty_for_io(inode, start, end);
 528                         redirty = 1;
 529                 }
 530
 531                 /* Compression level is applied here and only here */
 532                 ret = btrfs_compress_pages(
 533                         compress_type | (fs_info->compress_level << 4),
 534                                            inode->i_mapping, start,
 535                                            pages,
 536                                            &nr_pages,
 537                                            &total_in,
 538                                            &total_compressed);
 539
 540                 if (!ret) {
 541                         unsigned long offset = total_compressed &
 542                                 (PAGE_SIZE - 1);
 543                         struct page *page = pages[nr_pages - 1];
 544                         char *kaddr;
 545
 546                         /* zero the tail end of the last page, we might be
 547                          * sending it down to disk
 548                          */
 549                         if (offset) {
 550                                 kaddr = kmap_atomic(page);
 551                                 memset(kaddr + offset, 0,
 552                                        PAGE_SIZE - offset);
 553                                 kunmap_atomic(kaddr);
 554                         }
 555                         will_compress = 1;
 556                 }
 557         }
 558 cont:
 559         if (start == 0) {
 560                 /* lets try to make an inline extent */
 561                 if (ret || total_in < actual_end) {
 562                         /* we didn't compress the entire range, try
 563                          * to make an uncompressed inline extent.
 564                          */
 565                         ret = cow_file_range_inline(inode, start, end, 0,
 566                                                     BTRFS_COMPRESS_NONE, NULL);
 567                 } else {
 568                         /* try making a compressed inline extent */
 569                         ret = cow_file_range_inline(inode, start, end,
 570                                                     total_compressed,
 571                                                     compress_type, pages);
 572                 }
 573                 if (ret <= 0) {
 574                         unsigned long clear_flags = EXTENT_DELALLOC |
 575                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 576                                 EXTENT_DO_ACCOUNTING;
 577                         unsigned long page_error_op;
 578
 579                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 580
 581                         /*
 582                          * inline extent creation worked or returned error,
 583                          * we don't need to create any more async work items.
 584                          * Unlock and free up our temp pages.
 585                          *
 586                          * We use DO_ACCOUNTING here because we need the
 587                          * delalloc_release_metadata to be done _after_ we drop
 588                          * our outstanding extent for clearing delalloc for this
 589                          * range.
 590                          */
 591                         extent_clear_unlock_delalloc(inode, start, end, end,
 592                                                      NULL, clear_flags,
 593                                                      PAGE_UNLOCK |
 594                                                      PAGE_CLEAR_DIRTY |
 595                                                      PAGE_SET_WRITEBACK |
 596                                                      page_error_op |
 597                                                      PAGE_END_WRITEBACK);
 598                         goto free_pages_out;
 599                 }
 600         }
 601
 602         if (will_compress) {
 603                 /*
 604                  * we aren't doing an inline extent round the compressed size
 605                  * up to a block size boundary so the allocator does sane
 606                  * things
 607                  */
 608                 total_compressed = ALIGN(total_compressed, blocksize);
 609
 610                 /*
 611                  * one last check to make sure the compression is really a
 612                  * win, compare the page count read with the blocks on disk,
 613                  * compression must free at least one sector size
 614                  */
 615                 total_in = ALIGN(total_in, PAGE_SIZE);
 616                 if (total_compressed + blocksize <= total_in) {
 617                         *num_added += 1;
 618
 619                         /*
 620                          * The async work queues will take care of doing actual
 621                          * allocation on disk for these compressed pages, and
 622                          * will submit them to the elevator.
 623                          */
 624                         add_async_extent(async_cow, start, total_in,
 625                                         total_compressed, pages, nr_pages,
 626                                         compress_type);
 627
 628                         if (start + total_in < end) {
 629                                 start += total_in;
 630                                 pages = NULL;
 631                                 cond_resched();
 632                                 goto again;
 633                         }
 634                         return;
 635                 }
 636         }
 637         if (pages) {
 638                 /*
 639                  * the compression code ran but failed to make things smaller,
 640                  * free any pages it allocated and our page pointer array
 641                  */
 642                 for (i = 0; i < nr_pages; i++) {
 643                         WARN_ON(pages[i]->mapping);
 644                         put_page(pages[i]);
 645                 }
 646                 kfree(pages);
 647                 pages = NULL;
 648                 total_compressed = 0;
 649                 nr_pages = 0;
 650
 651                 /* flag the file so we don't compress in the future */
 652                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 653                     !(BTRFS_I(inode)->prop_compress)) {
 654                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 655                 }
 656         }
 657 cleanup_and_bail_uncompressed:
 658         /*
 659          * No compression, but we still need to write the pages in the file
 660          * we've been given so far.  redirty the locked page if it corresponds
 661          * to our extent and set things up for the async work queue to run
 662          * cow_file_range to do the normal delalloc dance.
 663          */
 664         if (page_offset(locked_page) >= start &&
 665             page_offset(locked_page) <= end)
 666                 __set_page_dirty_nobuffers(locked_page);
 667                 /* unlocked later on in the async handlers */
 668
 669         if (redirty)
 670                 extent_range_redirty_for_io(inode, start, end);
 671         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
 672                          BTRFS_COMPRESS_NONE);
 673         *num_added += 1;
 674
 675         return;
 676
 677 free_pages_out:
 678         for (i = 0; i < nr_pages; i++) {
 679                 WARN_ON(pages[i]->mapping);
 680                 put_page(pages[i]);
 681         }
 682         kfree(pages);
 683 }
 684
 685 static void free_async_extent_pages(struct async_extent *async_extent)
 686 {
 687         int i;
 688
 689         if (!async_extent->pages)
 690                 return;
 691
 692         for (i = 0; i < async_extent->nr_pages; i++) {
 693                 WARN_ON(async_extent->pages[i]->mapping);
 694                 put_page(async_extent->pages[i]);
 695         }
 696         kfree(async_extent->pages);
 697         async_extent->nr_pages = 0;
 698         async_extent->pages = NULL;
 699 }
 700
 701 /*
 702  * phase two of compressed writeback.  This is the ordered portion
 703  * of the code, which only gets called in the order the work was
 704  * queued.  We walk all the async extents created by compress_file_range
 705  * and send them down to the disk.
 706  */
 707 static noinline void submit_compressed_extents(struct inode *inode,
 708                                               struct async_cow *async_cow)
 709 {
 710         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 711         struct async_extent *async_extent;
 712         u64 alloc_hint = 0;
 713         struct btrfs_key ins;
 714         struct extent_map *em;
 715         struct btrfs_root *root = BTRFS_I(inode)->root;
 716         struct extent_io_tree *io_tree;
 717         int ret = 0;
 718
 719 again:
 720         while (!list_empty(&async_cow->extents)) {
 721                 async_extent = list_entry(async_cow->extents.next,
 722                                           struct async_extent, list);
 723                 list_del(&async_extent->list);
 724
 725                 io_tree = &BTRFS_I(inode)->io_tree;
 726
 727 retry:
 728                 /* did the compression code fall back to uncompressed IO? */
 729                 if (!async_extent->pages) {
 730                         int page_started = 0;
 731                         unsigned long nr_written = 0;
 732
 733                         lock_extent(io_tree, async_extent->start,
 734                                          async_extent->start +
 735                                          async_extent->ram_size - 1);
 736
 737                         /* allocate blocks */
 738                         ret = cow_file_range(inode, async_cow->locked_page,
 739                                              async_extent->start,
 740                                              async_extent->start +
 741                                              async_extent->ram_size - 1,
 742                                              async_extent->start +
 743                                              async_extent->ram_size - 1,
 744                                              &page_started, &nr_written, 0,
 745                                              NULL);
 746
 747                         /* JDM XXX */
 748
 749                         /*
 750                          * if page_started, cow_file_range inserted an
 751                          * inline extent and took care of all the unlocking
 752                          * and IO for us.  Otherwise, we need to submit
 753                          * all those pages down to the drive.
 754                          */
 755                         if (!page_started && !ret)
 756                                 extent_write_locked_range(inode,
 757                                                   async_extent->start,
 758                                                   async_extent->start +
 759                                                   async_extent->ram_size - 1,
 760                                                   WB_SYNC_ALL);
 761                         else if (ret)
 762                                 unlock_page(async_cow->locked_page);
 763                         kfree(async_extent);
 764                         cond_resched();
 765                         continue;
 766                 }
 767
 768                 lock_extent(io_tree, async_extent->start,
 769                             async_extent->start + async_extent->ram_size - 1);
 770
 771                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 772                                            async_extent->compressed_size,
 773                                            async_extent->compressed_size,
 774                                            0, alloc_hint, &ins, 1, 1);
 775                 if (ret) {
 776                         free_async_extent_pages(async_extent);
 777
 778                         if (ret == -ENOSPC) {
 779                                 unlock_extent(io_tree, async_extent->start,
 780                                               async_extent->start +
 781                                               async_extent->ram_size - 1);
 782
 783                                 /*
 784                                  * we need to redirty the pages if we decide to
 785                                  * fallback to uncompressed IO, otherwise we
 786                                  * will not submit these pages down to lower
 787                                  * layers.
 788                                  */
 789                                 extent_range_redirty_for_io(inode,
 790                                                 async_extent->start,
 791                                                 async_extent->start +
 792                                                 async_extent->ram_size - 1);
 793
 794                                 goto retry;
 795                         }
 796                         goto out_free;
 797                 }
 798                 /*
 799                  * here we're doing allocation and writeback of the
 800                  * compressed pages
 801                  */
 802                 em = create_io_em(inode, async_extent->start,
 803                                   async_extent->ram_size, /* len */
 804                                   async_extent->start, /* orig_start */
 805                                   ins.objectid, /* block_start */
 806                                   ins.offset, /* block_len */
 807                                   ins.offset, /* orig_block_len */
 808                                   async_extent->ram_size, /* ram_bytes */
 809                                   async_extent->compress_type,
 810                                   BTRFS_ORDERED_COMPRESSED);
 811                 if (IS_ERR(em))
 812                         /* ret value is not necessary due to void function */
 813                         goto out_free_reserve;
 814                 free_extent_map(em);
 815
 816                 ret = btrfs_add_ordered_extent_compress(inode,
 817                                                 async_extent->start,
 818                                                 ins.objectid,
 819                                                 async_extent->ram_size,
 820                                                 ins.offset,
 821                                                 BTRFS_ORDERED_COMPRESSED,
 822                                                 async_extent->compress_type);
 823                 if (ret) {
 824                         btrfs_drop_extent_cache(BTRFS_I(inode),
 825                                                 async_extent->start,
 826                                                 async_extent->start +
 827                                                 async_extent->ram_size - 1, 0);
 828                         goto out_free_reserve;
 829                 }
 830                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 831
 832                 /*
 833                  * clear dirty, set writeback and unlock the pages.
 834                  */
 835                 extent_clear_unlock_delalloc(inode, async_extent->start,
 836                                 async_extent->start +
 837                                 async_extent->ram_size - 1,
 838                                 async_extent->start +
 839                                 async_extent->ram_size - 1,
 840                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 841                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 842                                 PAGE_SET_WRITEBACK);
 843                 if (btrfs_submit_compressed_write(inode,
 844                                     async_extent->start,
 845                                     async_extent->ram_size,
 846                                     ins.objectid,
 847                                     ins.offset, async_extent->pages,
 848                                     async_extent->nr_pages,
 849                                     async_cow->write_flags)) {
 850                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 851                         struct page *p = async_extent->pages[0];
 852                         const u64 start = async_extent->start;
 853                         const u64 end = start + async_extent->ram_size - 1;
 854
 855                         p->mapping = inode->i_mapping;
 856                         tree->ops->writepage_end_io_hook(p, start, end,
 857                                                          NULL, 0);
 858                         p->mapping = NULL;
 859                         extent_clear_unlock_delalloc(inode, start, end, end,
 860                                                      NULL, 0,
 861                                                      PAGE_END_WRITEBACK |
 862                                                      PAGE_SET_ERROR);
 863                         free_async_extent_pages(async_extent);
 864                 }
 865                 alloc_hint = ins.objectid + ins.offset;
 866                 kfree(async_extent);
 867                 cond_resched();
 868         }
 869         return;
 870 out_free_reserve:
 871         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 872         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 873 out_free:
 874         extent_clear_unlock_delalloc(inode, async_extent->start,
 875                                      async_extent->start +
 876                                      async_extent->ram_size - 1,
 877                                      async_extent->start +
 878                                      async_extent->ram_size - 1,
 879                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 880                                      EXTENT_DELALLOC_NEW |
 881                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 882                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 883                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 884                                      PAGE_SET_ERROR);
 885         free_async_extent_pages(async_extent);
 886         kfree(async_extent);
 887         goto again;
 888 }
 889
 890 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 891                                       u64 num_bytes)
 892 {
 893         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 894         struct extent_map *em;
 895         u64 alloc_hint = 0;
 896
 897         read_lock(&em_tree->lock);
 898         em = search_extent_mapping(em_tree, start, num_bytes);
 899         if (em) {
 900                 /*
 901                  * if block start isn't an actual block number then find the
 902                  * first block in this inode and use that as a hint.  If that
 903                  * block is also bogus then just don't worry about it.
 904                  */
 905                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 906                         free_extent_map(em);
 907                         em = search_extent_mapping(em_tree, 0, 0);
 908                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 909                                 alloc_hint = em->block_start;
 910                         if (em)
 911                                 free_extent_map(em);
 912                 } else {
 913                         alloc_hint = em->block_start;
 914                         free_extent_map(em);
 915                 }
 916         }
 917         read_unlock(&em_tree->lock);
 918
 919         return alloc_hint;
 920 }
 921
 922 /*
 923  * when extent_io.c finds a delayed allocation range in the file,
 924  * the call backs end up in this code.  The basic idea is to
 925  * allocate extents on disk for the range, and create ordered data structs
 926  * in ram to track those extents.
 927  *
 928  * locked_page is the page that writepage had locked already.  We use
 929  * it to make sure we don't do extra locks or unlocks.
 930  *
 931  * *page_started is set to one if we unlock locked_page and do everything
 932  * required to start IO on it.  It may be clean and already done with
 933  * IO when we return.
 934  */
 935 static noinline int cow_file_range(struct inode *inode,
 936                                    struct page *locked_page,
 937                                    u64 start, u64 end, u64 delalloc_end,
 938                                    int *page_started, unsigned long *nr_written,
 939                                    int unlock, struct btrfs_dedupe_hash *hash)
 940 {
 941         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 942         struct btrfs_root *root = BTRFS_I(inode)->root;
 943         u64 alloc_hint = 0;
 944         u64 num_bytes;
 945         unsigned long ram_size;
 946         u64 cur_alloc_size = 0;
 947         u64 blocksize = fs_info->sectorsize;
 948         struct btrfs_key ins;
 949         struct extent_map *em;
 950         unsigned clear_bits;
 951         unsigned long page_ops;
 952         bool extent_reserved = false;
 953         int ret = 0;
 954
 955         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 956                 WARN_ON_ONCE(1);
 957                 ret = -EINVAL;
 958                 goto out_unlock;
 959         }
 960
 961         num_bytes = ALIGN(end - start + 1, blocksize);
 962         num_bytes = max(blocksize,  num_bytes);
 963         ASSERT(num_bytes <= btrfs_super_total_bytes(fs_info->super_copy));
 964
 965         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
 966
 967         if (start == 0) {
 968                 /* lets try to make an inline extent */
 969                 ret = cow_file_range_inline(inode, start, end, 0,
 970                                             BTRFS_COMPRESS_NONE, NULL);
 971                 if (ret == 0) {
 972                         /*
 973                          * We use DO_ACCOUNTING here because we need the
 974                          * delalloc_release_metadata to be run _after_ we drop
 975                          * our outstanding extent for clearing delalloc for this
 976                          * range.
 977                          */
 978                         extent_clear_unlock_delalloc(inode, start, end,
 979                                      delalloc_end, NULL,
 980                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 981                                      EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 982                                      EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
 983                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 984                                      PAGE_END_WRITEBACK);
 985                         *nr_written = *nr_written +
 986                              (end - start + PAGE_SIZE) / PAGE_SIZE;
 987                         *page_started = 1;
 988                         goto out;
 989                 } else if (ret < 0) {
 990                         goto out_unlock;
 991                 }
 992         }
 993
 994         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 995         btrfs_drop_extent_cache(BTRFS_I(inode), start,
 996                         start + num_bytes - 1, 0);
 997
 998         while (num_bytes > 0) {
 999                 cur_alloc_size = num_bytes;
1000                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1001                                            fs_info->sectorsize, 0, alloc_hint,
1002                                            &ins, 1, 1);
1003                 if (ret < 0)
1004                         goto out_unlock;
1005                 cur_alloc_size = ins.offset;
1006                 extent_reserved = true;
1007
1008                 ram_size = ins.offset;
1009                 em = create_io_em(inode, start, ins.offset, /* len */
1010                                   start, /* orig_start */
1011                                   ins.objectid, /* block_start */
1012                                   ins.offset, /* block_len */
1013                                   ins.offset, /* orig_block_len */
1014                                   ram_size, /* ram_bytes */
1015                                   BTRFS_COMPRESS_NONE, /* compress_type */
1016                                   BTRFS_ORDERED_REGULAR /* type */);
1017                 if (IS_ERR(em)) {
1018                         ret = PTR_ERR(em);
1019                         goto out_reserve;
1020                 }
1021                 free_extent_map(em);
1022
1023                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1024                                                ram_size, cur_alloc_size, 0);
1025                 if (ret)
1026                         goto out_drop_extent_cache;
1027
1028                 if (root->root_key.objectid ==
1029                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1030                         ret = btrfs_reloc_clone_csums(inode, start,
1031                                                       cur_alloc_size);
1032                         /*
1033                          * Only drop cache here, and process as normal.
1034                          *
1035                          * We must not allow extent_clear_unlock_delalloc()
1036                          * at out_unlock label to free meta of this ordered
1037                          * extent, as its meta should be freed by
1038                          * btrfs_finish_ordered_io().
1039                          *
1040                          * So we must continue until @start is increased to
1041                          * skip current ordered extent.
1042                          */
1043                         if (ret)
1044                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1045                                                 start + ram_size - 1, 0);
1046                 }
1047
1048                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1049
1050                 /* we're not doing compressed IO, don't unlock the first
1051                  * page (which the caller expects to stay locked), don't
1052                  * clear any dirty bits and don't set any writeback bits
1053                  *
1054                  * Do set the Private2 bit so we know this page was properly
1055                  * setup for writepage
1056                  */
1057                 page_ops = unlock ? PAGE_UNLOCK : 0;
1058                 page_ops |= PAGE_SET_PRIVATE2;
1059
1060                 extent_clear_unlock_delalloc(inode, start,
1061                                              start + ram_size - 1,
1062                                              delalloc_end, locked_page,
1063                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1064                                              page_ops);
1065                 if (num_bytes < cur_alloc_size)
1066                         num_bytes = 0;
1067                 else
1068                         num_bytes -= cur_alloc_size;
1069                 alloc_hint = ins.objectid + ins.offset;
1070                 start += cur_alloc_size;
1071                 extent_reserved = false;
1072
1073                 /*
1074                  * btrfs_reloc_clone_csums() error, since start is increased
1075                  * extent_clear_unlock_delalloc() at out_unlock label won't
1076                  * free metadata of current ordered extent, we're OK to exit.
1077                  */
1078                 if (ret)
1079                         goto out_unlock;
1080         }
1081 out:
1082         return ret;
1083
1084 out_drop_extent_cache:
1085         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1086 out_reserve:
1087         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1088         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1089 out_unlock:
1090         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1091                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1092         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1093                 PAGE_END_WRITEBACK;
1094         /*
1095          * If we reserved an extent for our delalloc range (or a subrange) and
1096          * failed to create the respective ordered extent, then it means that
1097          * when we reserved the extent we decremented the extent's size from
1098          * the data space_info's bytes_may_use counter and incremented the
1099          * space_info's bytes_reserved counter by the same amount. We must make
1100          * sure extent_clear_unlock_delalloc() does not try to decrement again
1101          * the data space_info's bytes_may_use counter, therefore we do not pass
1102          * it the flag EXTENT_CLEAR_DATA_RESV.
1103          */
1104         if (extent_reserved) {
1105                 extent_clear_unlock_delalloc(inode, start,
1106                                              start + cur_alloc_size,
1107                                              start + cur_alloc_size,
1108                                              locked_page,
1109                                              clear_bits,
1110                                              page_ops);
1111                 start += cur_alloc_size;
1112                 if (start >= end)
1113                         goto out;
1114         }
1115         extent_clear_unlock_delalloc(inode, start, end, delalloc_end,
1116                                      locked_page,
1117                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1118                                      page_ops);
1119         goto out;
1120 }
1121
1122 /*
1123  * work queue call back to started compression on a file and pages
1124  */
1125 static noinline void async_cow_start(struct btrfs_work *work)
1126 {
1127         struct async_cow *async_cow;
1128         int num_added = 0;
1129         async_cow = container_of(work, struct async_cow, work);
1130
1131         compress_file_range(async_cow->inode, async_cow->locked_page,
1132                             async_cow->start, async_cow->end, async_cow,
1133                             &num_added);
1134         if (num_added == 0) {
1135                 btrfs_add_delayed_iput(async_cow->inode);
1136                 async_cow->inode = NULL;
1137         }
1138 }
1139
1140 /*
1141  * work queue call back to submit previously compressed pages
1142  */
1143 static noinline void async_cow_submit(struct btrfs_work *work)
1144 {
1145         struct btrfs_fs_info *fs_info;
1146         struct async_cow *async_cow;
1147         struct btrfs_root *root;
1148         unsigned long nr_pages;
1149
1150         async_cow = container_of(work, struct async_cow, work);
1151
1152         root = async_cow->root;
1153         fs_info = root->fs_info;
1154         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1155                 PAGE_SHIFT;
1156
1157         /* atomic_sub_return implies a barrier */
1158         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1159             5 * SZ_1M)
1160                 cond_wake_up_nomb(&fs_info->async_submit_wait);
1161
1162         if (async_cow->inode)
1163                 submit_compressed_extents(async_cow->inode, async_cow);
1164 }
1165
1166 static noinline void async_cow_free(struct btrfs_work *work)
1167 {
1168         struct async_cow *async_cow;
1169         async_cow = container_of(work, struct async_cow, work);
1170         if (async_cow->inode)
1171                 btrfs_add_delayed_iput(async_cow->inode);
1172         kfree(async_cow);
1173 }
1174
1175 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1176                                 u64 start, u64 end, int *page_started,
1177                                 unsigned long *nr_written,
1178                                 unsigned int write_flags)
1179 {
1180         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1181         struct async_cow *async_cow;
1182         struct btrfs_root *root = BTRFS_I(inode)->root;
1183         unsigned long nr_pages;
1184         u64 cur_end;
1185
1186         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1187                          1, 0, NULL);
1188         while (start < end) {
1189                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1190                 BUG_ON(!async_cow); /* -ENOMEM */
1191                 async_cow->inode = igrab(inode);
1192                 async_cow->root = root;
1193                 async_cow->locked_page = locked_page;
1194                 async_cow->start = start;
1195                 async_cow->write_flags = write_flags;
1196
1197                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1198                     !btrfs_test_opt(fs_info, FORCE_COMPRESS))
1199                         cur_end = end;
1200                 else
1201                         cur_end = min(end, start + SZ_512K - 1);
1202
1203                 async_cow->end = cur_end;
1204                 INIT_LIST_HEAD(&async_cow->extents);
1205
1206                 btrfs_init_work(&async_cow->work,
1207                                 btrfs_delalloc_helper,
1208                                 async_cow_start, async_cow_submit,
1209                                 async_cow_free);
1210
1211                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1212                         PAGE_SHIFT;
1213                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1214
1215                 btrfs_queue_work(fs_info->delalloc_workers, &async_cow->work);
1216
1217                 *nr_written += nr_pages;
1218                 start = cur_end + 1;
1219         }
1220         *page_started = 1;
1221         return 0;
1222 }
1223
1224 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1225                                         u64 bytenr, u64 num_bytes)
1226 {
1227         int ret;
1228         struct btrfs_ordered_sum *sums;
1229         LIST_HEAD(list);
1230
1231         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1232                                        bytenr + num_bytes - 1, &list, 0);
1233         if (ret == 0 && list_empty(&list))
1234                 return 0;
1235
1236         while (!list_empty(&list)) {
1237                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1238                 list_del(&sums->list);
1239                 kfree(sums);
1240         }
1241         if (ret < 0)
1242                 return ret;
1243         return 1;
1244 }
1245
1246 /*
1247  * when nowcow writeback call back.  This checks for snapshots or COW copies
1248  * of the extents that exist in the file, and COWs the file as required.
1249  *
1250  * If no cow copies or snapshots exist, we write directly to the existing
1251  * blocks on disk
1252  */
1253 static noinline int run_delalloc_nocow(struct inode *inode,
1254                                        struct page *locked_page,
1255                               u64 start, u64 end, int *page_started, int force,
1256                               unsigned long *nr_written)
1257 {
1258         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1259         struct btrfs_root *root = BTRFS_I(inode)->root;
1260         struct extent_buffer *leaf;
1261         struct btrfs_path *path;
1262         struct btrfs_file_extent_item *fi;
1263         struct btrfs_key found_key;
1264         struct extent_map *em;
1265         u64 cow_start;
1266         u64 cur_offset;
1267         u64 extent_end;
1268         u64 extent_offset;
1269         u64 disk_bytenr;
1270         u64 num_bytes;
1271         u64 disk_num_bytes;
1272         u64 ram_bytes;
1273         int extent_type;
1274         int ret, err;
1275         int type;
1276         int nocow;
1277         int check_prev = 1;
1278         bool nolock;
1279         u64 ino = btrfs_ino(BTRFS_I(inode));
1280
1281         path = btrfs_alloc_path();
1282         if (!path) {
1283                 extent_clear_unlock_delalloc(inode, start, end, end,
1284                                              locked_page,
1285                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1286                                              EXTENT_DO_ACCOUNTING |
1287                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1288                                              PAGE_CLEAR_DIRTY |
1289                                              PAGE_SET_WRITEBACK |
1290                                              PAGE_END_WRITEBACK);
1291                 return -ENOMEM;
1292         }
1293
1294         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
1295
1296         cow_start = (u64)-1;
1297         cur_offset = start;
1298         while (1) {
1299                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1300                                                cur_offset, 0);
1301                 if (ret < 0)
1302                         goto error;
1303                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1304                         leaf = path->nodes[0];
1305                         btrfs_item_key_to_cpu(leaf, &found_key,
1306                                               path->slots[0] - 1);
1307                         if (found_key.objectid == ino &&
1308                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1309                                 path->slots[0]--;
1310                 }
1311                 check_prev = 0;
1312 next_slot:
1313                 leaf = path->nodes[0];
1314                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1315                         ret = btrfs_next_leaf(root, path);
1316                         if (ret < 0) {
1317                                 if (cow_start != (u64)-1)
1318                                         cur_offset = cow_start;
1319                                 goto error;
1320                         }
1321                         if (ret > 0)
1322                                 break;
1323                         leaf = path->nodes[0];
1324                 }
1325
1326                 nocow = 0;
1327                 disk_bytenr = 0;
1328                 num_bytes = 0;
1329                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1330
1331                 if (found_key.objectid > ino)
1332                         break;
1333                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1334                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1335                         path->slots[0]++;
1336                         goto next_slot;
1337                 }
1338                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1339                     found_key.offset > end)
1340                         break;
1341
1342                 if (found_key.offset > cur_offset) {
1343                         extent_end = found_key.offset;
1344                         extent_type = 0;
1345                         goto out_check;
1346                 }
1347
1348                 fi = btrfs_item_ptr(leaf, path->slots[0],
1349                                     struct btrfs_file_extent_item);
1350                 extent_type = btrfs_file_extent_type(leaf, fi);
1351
1352                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1353                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1354                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1355                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1356                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1357                         extent_end = found_key.offset +
1358                                 btrfs_file_extent_num_bytes(leaf, fi);
1359                         disk_num_bytes =
1360                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1361                         if (extent_end <= start) {
1362                                 path->slots[0]++;
1363                                 goto next_slot;
1364                         }
1365                         if (disk_bytenr == 0)
1366                                 goto out_check;
1367                         if (btrfs_file_extent_compression(leaf, fi) ||
1368                             btrfs_file_extent_encryption(leaf, fi) ||
1369                             btrfs_file_extent_other_encoding(leaf, fi))
1370                                 goto out_check;
1371                         /*
1372                          * Do the same check as in btrfs_cross_ref_exist but
1373                          * without the unnecessary search.
1374                          */
1375                         if (btrfs_file_extent_generation(leaf, fi) <=
1376                             btrfs_root_last_snapshot(&root->root_item))
1377                                 goto out_check;
1378                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1379                                 goto out_check;
1380                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1381                                 goto out_check;
1382                         ret = btrfs_cross_ref_exist(root, ino,
1383                                                     found_key.offset -
1384                                                     extent_offset, disk_bytenr);
1385                         if (ret) {
1386                                 /*
1387                                  * ret could be -EIO if the above fails to read
1388                                  * metadata.
1389                                  */
1390                                 if (ret < 0) {
1391                                         if (cow_start != (u64)-1)
1392                                                 cur_offset = cow_start;
1393                                         goto error;
1394                                 }
1395
1396                                 WARN_ON_ONCE(nolock);
1397                                 goto out_check;
1398                         }
1399                         disk_bytenr += extent_offset;
1400                         disk_bytenr += cur_offset - found_key.offset;
1401                         num_bytes = min(end + 1, extent_end) - cur_offset;
1402                         /*
1403                          * if there are pending snapshots for this root,
1404                          * we fall into common COW way.
1405                          */
1406                         if (!nolock) {
1407                                 err = btrfs_start_write_no_snapshotting(root);
1408                                 if (!err)
1409                                         goto out_check;
1410                         }
1411                         /*
1412                          * force cow if csum exists in the range.
1413                          * this ensure that csum for a given extent are
1414                          * either valid or do not exist.
1415                          */
1416                         ret = csum_exist_in_range(fs_info, disk_bytenr,
1417                                                   num_bytes);
1418                         if (ret) {
1419                                 if (!nolock)
1420                                         btrfs_end_write_no_snapshotting(root);
1421
1422                                 /*
1423                                  * ret could be -EIO if the above fails to read
1424                                  * metadata.
1425                                  */
1426                                 if (ret < 0) {
1427                                         if (cow_start != (u64)-1)
1428                                                 cur_offset = cow_start;
1429                                         goto error;
1430                                 }
1431                                 WARN_ON_ONCE(nolock);
1432                                 goto out_check;
1433                         }
1434                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr)) {
1435                                 if (!nolock)
1436                                         btrfs_end_write_no_snapshotting(root);
1437                                 goto out_check;
1438                         }
1439                         nocow = 1;
1440                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1441                         extent_end = found_key.offset +
1442                                 btrfs_file_extent_ram_bytes(leaf, fi);
1443                         extent_end = ALIGN(extent_end,
1444                                            fs_info->sectorsize);
1445                 } else {
1446                         BUG_ON(1);
1447                 }
1448 out_check:
1449                 if (extent_end <= start) {
1450                         path->slots[0]++;
1451                         if (!nolock && nocow)
1452                                 btrfs_end_write_no_snapshotting(root);
1453                         if (nocow)
1454                                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1455                         goto next_slot;
1456                 }
1457                 if (!nocow) {
1458                         if (cow_start == (u64)-1)
1459                                 cow_start = cur_offset;
1460                         cur_offset = extent_end;
1461                         if (cur_offset > end)
1462                                 break;
1463                         path->slots[0]++;
1464                         goto next_slot;
1465                 }
1466
1467                 btrfs_release_path(path);
1468                 if (cow_start != (u64)-1) {
1469                         ret = cow_file_range(inode, locked_page,
1470                                              cow_start, found_key.offset - 1,
1471                                              end, page_started, nr_written, 1,
1472                                              NULL);
1473                         if (ret) {
1474                                 if (!nolock && nocow)
1475                                         btrfs_end_write_no_snapshotting(root);
1476                                 if (nocow)
1477                                         btrfs_dec_nocow_writers(fs_info,
1478                                                                 disk_bytenr);
1479                                 goto error;
1480                         }
1481                         cow_start = (u64)-1;
1482                 }
1483
1484                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1485                         u64 orig_start = found_key.offset - extent_offset;
1486
1487                         em = create_io_em(inode, cur_offset, num_bytes,
1488                                           orig_start,
1489                                           disk_bytenr, /* block_start */
1490                                           num_bytes, /* block_len */
1491                                           disk_num_bytes, /* orig_block_len */
1492                                           ram_bytes, BTRFS_COMPRESS_NONE,
1493                                           BTRFS_ORDERED_PREALLOC);
1494                         if (IS_ERR(em)) {
1495                                 if (!nolock && nocow)
1496                                         btrfs_end_write_no_snapshotting(root);
1497                                 if (nocow)
1498                                         btrfs_dec_nocow_writers(fs_info,
1499                                                                 disk_bytenr);
1500                                 ret = PTR_ERR(em);
1501                                 goto error;
1502                         }
1503                         free_extent_map(em);
1504                 }
1505
1506                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1507                         type = BTRFS_ORDERED_PREALLOC;
1508                 } else {
1509                         type = BTRFS_ORDERED_NOCOW;
1510                 }
1511
1512                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1513                                                num_bytes, num_bytes, type);
1514                 if (nocow)
1515                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1516                 BUG_ON(ret); /* -ENOMEM */
1517
1518                 if (root->root_key.objectid ==
1519                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1520                         /*
1521                          * Error handled later, as we must prevent
1522                          * extent_clear_unlock_delalloc() in error handler
1523                          * from freeing metadata of created ordered extent.
1524                          */
1525                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1526                                                       num_bytes);
1527
1528                 extent_clear_unlock_delalloc(inode, cur_offset,
1529                                              cur_offset + num_bytes - 1, end,
1530                                              locked_page, EXTENT_LOCKED |
1531                                              EXTENT_DELALLOC |
1532                                              EXTENT_CLEAR_DATA_RESV,
1533                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1534
1535                 if (!nolock && nocow)
1536                         btrfs_end_write_no_snapshotting(root);
1537                 cur_offset = extent_end;
1538
1539                 /*
1540                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1541                  * handler, as metadata for created ordered extent will only
1542                  * be freed by btrfs_finish_ordered_io().
1543                  */
1544                 if (ret)
1545                         goto error;
1546                 if (cur_offset > end)
1547                         break;
1548         }
1549         btrfs_release_path(path);
1550
1551         if (cur_offset <= end && cow_start == (u64)-1) {
1552                 cow_start = cur_offset;
1553                 cur_offset = end;
1554         }
1555
1556         if (cow_start != (u64)-1) {
1557                 ret = cow_file_range(inode, locked_page, cow_start, end, end,
1558                                      page_started, nr_written, 1, NULL);
1559                 if (ret)
1560                         goto error;
1561         }
1562
1563 error:
1564         if (ret && cur_offset < end)
1565                 extent_clear_unlock_delalloc(inode, cur_offset, end, end,
1566                                              locked_page, EXTENT_LOCKED |
1567                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1568                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1569                                              PAGE_CLEAR_DIRTY |
1570                                              PAGE_SET_WRITEBACK |
1571                                              PAGE_END_WRITEBACK);
1572         btrfs_free_path(path);
1573         return ret;
1574 }
1575
1576 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1577 {
1578
1579         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1580             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1581                 return 0;
1582
1583         /*
1584          * @defrag_bytes is a hint value, no spinlock held here,
1585          * if is not zero, it means the file is defragging.
1586          * Force cow if given extent needs to be defragged.
1587          */
1588         if (BTRFS_I(inode)->defrag_bytes &&
1589             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1590                            EXTENT_DEFRAG, 0, NULL))
1591                 return 1;
1592
1593         return 0;
1594 }
1595
1596 /*
1597  * extent_io.c call back to do delayed allocation processing
1598  */
1599 static int run_delalloc_range(void *private_data, struct page *locked_page,
1600                               u64 start, u64 end, int *page_started,
1601                               unsigned long *nr_written,
1602                               struct writeback_control *wbc)
1603 {
1604         struct inode *inode = private_data;
1605         int ret;
1606         int force_cow = need_force_cow(inode, start, end);
1607         unsigned int write_flags = wbc_to_write_flags(wbc);
1608
1609         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1610                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1611                                          page_started, 1, nr_written);
1612         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1613                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1614                                          page_started, 0, nr_written);
1615         } else if (!inode_need_compress(inode, start, end)) {
1616                 ret = cow_file_range(inode, locked_page, start, end, end,
1617                                       page_started, nr_written, 1, NULL);
1618         } else {
1619                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1620                         &BTRFS_I(inode)->runtime_flags);
1621                 ret = cow_file_range_async(inode, locked_page, start, end,
1622                                            page_started, nr_written,
1623                                            write_flags);
1624         }
1625         if (ret)
1626                 btrfs_cleanup_ordered_extents(inode, start, end - start + 1);
1627         return ret;
1628 }
1629
1630 static void btrfs_split_extent_hook(void *private_data,
1631                                     struct extent_state *orig, u64 split)
1632 {
1633         struct inode *inode = private_data;
1634         u64 size;
1635
1636         /* not delalloc, ignore it */
1637         if (!(orig->state & EXTENT_DELALLOC))
1638                 return;
1639
1640         size = orig->end - orig->start + 1;
1641         if (size > BTRFS_MAX_EXTENT_SIZE) {
1642                 u32 num_extents;
1643                 u64 new_size;
1644
1645                 /*
1646                  * See the explanation in btrfs_merge_extent_hook, the same
1647                  * applies here, just in reverse.
1648                  */
1649                 new_size = orig->end - split + 1;
1650                 num_extents = count_max_extents(new_size);
1651                 new_size = split - orig->start;
1652                 num_extents += count_max_extents(new_size);
1653                 if (count_max_extents(size) >= num_extents)
1654                         return;
1655         }
1656
1657         spin_lock(&BTRFS_I(inode)->lock);
1658         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
1659         spin_unlock(&BTRFS_I(inode)->lock);
1660 }
1661
1662 /*
1663  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1664  * extents so we can keep track of new extents that are just merged onto old
1665  * extents, such as when we are doing sequential writes, so we can properly
1666  * account for the metadata space we'll need.
1667  */
1668 static void btrfs_merge_extent_hook(void *private_data,
1669                                     struct extent_state *new,
1670                                     struct extent_state *other)
1671 {
1672         struct inode *inode = private_data;
1673         u64 new_size, old_size;
1674         u32 num_extents;
1675
1676         /* not delalloc, ignore it */
1677         if (!(other->state & EXTENT_DELALLOC))
1678                 return;
1679
1680         if (new->start > other->start)
1681                 new_size = new->end - other->start + 1;
1682         else
1683                 new_size = other->end - new->start + 1;
1684
1685         /* we're not bigger than the max, unreserve the space and go */
1686         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1687                 spin_lock(&BTRFS_I(inode)->lock);
1688                 btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1689                 spin_unlock(&BTRFS_I(inode)->lock);
1690                 return;
1691         }
1692
1693         /*
1694          * We have to add up either side to figure out how many extents were
1695          * accounted for before we merged into one big extent.  If the number of
1696          * extents we accounted for is <= the amount we need for the new range
1697          * then we can return, otherwise drop.  Think of it like this
1698          *
1699          * [ 4k][MAX_SIZE]
1700          *
1701          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1702          * need 2 outstanding extents, on one side we have 1 and the other side
1703          * we have 1 so they are == and we can return.  But in this case
1704          *
1705          * [MAX_SIZE+4k][MAX_SIZE+4k]
1706          *
1707          * Each range on their own accounts for 2 extents, but merged together
1708          * they are only 3 extents worth of accounting, so we need to drop in
1709          * this case.
1710          */
1711         old_size = other->end - other->start + 1;
1712         num_extents = count_max_extents(old_size);
1713         old_size = new->end - new->start + 1;
1714         num_extents += count_max_extents(old_size);
1715         if (count_max_extents(new_size) >= num_extents)
1716                 return;
1717
1718         spin_lock(&BTRFS_I(inode)->lock);
1719         btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1720         spin_unlock(&BTRFS_I(inode)->lock);
1721 }
1722
1723 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1724                                       struct inode *inode)
1725 {
1726         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1727
1728         spin_lock(&root->delalloc_lock);
1729         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1730                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1731                               &root->delalloc_inodes);
1732                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1733                         &BTRFS_I(inode)->runtime_flags);
1734                 root->nr_delalloc_inodes++;
1735                 if (root->nr_delalloc_inodes == 1) {
1736                         spin_lock(&fs_info->delalloc_root_lock);
1737                         BUG_ON(!list_empty(&root->delalloc_root));
1738                         list_add_tail(&root->delalloc_root,
1739                                       &fs_info->delalloc_roots);
1740                         spin_unlock(&fs_info->delalloc_root_lock);
1741                 }
1742         }
1743         spin_unlock(&root->delalloc_lock);
1744 }
1745
1746
1747 void __btrfs_del_delalloc_inode(struct btrfs_root *root,
1748                                 struct btrfs_inode *inode)
1749 {
1750         struct btrfs_fs_info *fs_info = root->fs_info;
1751
1752         if (!list_empty(&inode->delalloc_inodes)) {
1753                 list_del_init(&inode->delalloc_inodes);
1754                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1755                           &inode->runtime_flags);
1756                 root->nr_delalloc_inodes--;
1757                 if (!root->nr_delalloc_inodes) {
1758                         ASSERT(list_empty(&root->delalloc_inodes));
1759                         spin_lock(&fs_info->delalloc_root_lock);
1760                         BUG_ON(list_empty(&root->delalloc_root));
1761                         list_del_init(&root->delalloc_root);
1762                         spin_unlock(&fs_info->delalloc_root_lock);
1763                 }
1764         }
1765 }
1766
1767 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1768                                      struct btrfs_inode *inode)
1769 {
1770         spin_lock(&root->delalloc_lock);
1771         __btrfs_del_delalloc_inode(root, inode);
1772         spin_unlock(&root->delalloc_lock);
1773 }
1774
1775 /*
1776  * extent_io.c set_bit_hook, used to track delayed allocation
1777  * bytes in this file, and to maintain the list of inodes that
1778  * have pending delalloc work to be done.
1779  */
1780 static void btrfs_set_bit_hook(void *private_data,
1781                                struct extent_state *state, unsigned *bits)
1782 {
1783         struct inode *inode = private_data;
1784
1785         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1786
1787         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1788                 WARN_ON(1);
1789         /*
1790          * set_bit and clear bit hooks normally require _irqsave/restore
1791          * but in this case, we are only testing for the DELALLOC
1792          * bit, which is only set or cleared with irqs on
1793          */
1794         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1795                 struct btrfs_root *root = BTRFS_I(inode)->root;
1796                 u64 len = state->end + 1 - state->start;
1797                 u32 num_extents = count_max_extents(len);
1798                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1799
1800                 spin_lock(&BTRFS_I(inode)->lock);
1801                 btrfs_mod_outstanding_extents(BTRFS_I(inode), num_extents);
1802                 spin_unlock(&BTRFS_I(inode)->lock);
1803
1804                 /* For sanity tests */
1805                 if (btrfs_is_testing(fs_info))
1806                         return;
1807
1808                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
1809                                          fs_info->delalloc_batch);
1810                 spin_lock(&BTRFS_I(inode)->lock);
1811                 BTRFS_I(inode)->delalloc_bytes += len;
1812                 if (*bits & EXTENT_DEFRAG)
1813                         BTRFS_I(inode)->defrag_bytes += len;
1814                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1815                                          &BTRFS_I(inode)->runtime_flags))
1816                         btrfs_add_delalloc_inodes(root, inode);
1817                 spin_unlock(&BTRFS_I(inode)->lock);
1818         }
1819
1820         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1821             (*bits & EXTENT_DELALLOC_NEW)) {
1822                 spin_lock(&BTRFS_I(inode)->lock);
1823                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1824                         state->start;
1825                 spin_unlock(&BTRFS_I(inode)->lock);
1826         }
1827 }
1828
1829 /*
1830  * extent_io.c clear_bit_hook, see set_bit_hook for why
1831  */
1832 static void btrfs_clear_bit_hook(void *private_data,
1833                                  struct extent_state *state,
1834                                  unsigned *bits)
1835 {
1836         struct btrfs_inode *inode = BTRFS_I((struct inode *)private_data);
1837         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1838         u64 len = state->end + 1 - state->start;
1839         u32 num_extents = count_max_extents(len);
1840
1841         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
1842                 spin_lock(&inode->lock);
1843                 inode->defrag_bytes -= len;
1844                 spin_unlock(&inode->lock);
1845         }
1846
1847         /*
1848          * set_bit and clear bit hooks normally require _irqsave/restore
1849          * but in this case, we are only testing for the DELALLOC
1850          * bit, which is only set or cleared with irqs on
1851          */
1852         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1853                 struct btrfs_root *root = inode->root;
1854                 bool do_list = !btrfs_is_free_space_inode(inode);
1855
1856                 spin_lock(&inode->lock);
1857                 btrfs_mod_outstanding_extents(inode, -num_extents);
1858                 spin_unlock(&inode->lock);
1859
1860                 /*
1861                  * We don't reserve metadata space for space cache inodes so we
1862                  * don't need to call dellalloc_release_metadata if there is an
1863                  * error.
1864                  */
1865                 if (*bits & EXTENT_CLEAR_META_RESV &&
1866                     root != fs_info->tree_root)
1867                         btrfs_delalloc_release_metadata(inode, len, false);
1868
1869                 /* For sanity tests. */
1870                 if (btrfs_is_testing(fs_info))
1871                         return;
1872
1873                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
1874                     do_list && !(state->state & EXTENT_NORESERVE) &&
1875                     (*bits & EXTENT_CLEAR_DATA_RESV))
1876                         btrfs_free_reserved_data_space_noquota(
1877                                         &inode->vfs_inode,
1878                                         state->start, len);
1879
1880                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
1881                                          fs_info->delalloc_batch);
1882                 spin_lock(&inode->lock);
1883                 inode->delalloc_bytes -= len;
1884                 if (do_list && inode->delalloc_bytes == 0 &&
1885                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1886                                         &inode->runtime_flags))
1887                         btrfs_del_delalloc_inode(root, inode);
1888                 spin_unlock(&inode->lock);
1889         }
1890
1891         if ((state->state & EXTENT_DELALLOC_NEW) &&
1892             (*bits & EXTENT_DELALLOC_NEW)) {
1893                 spin_lock(&inode->lock);
1894                 ASSERT(inode->new_delalloc_bytes >= len);
1895                 inode->new_delalloc_bytes -= len;
1896                 spin_unlock(&inode->lock);
1897         }
1898 }
1899
1900 /*
1901  * Merge bio hook, this must check the chunk tree to make sure we don't create
1902  * bios that span stripes or chunks
1903  *
1904  * return 1 if page cannot be merged to bio
1905  * return 0 if page can be merged to bio
1906  * return error otherwise
1907  */
1908 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1909                          size_t size, struct bio *bio,
1910                          unsigned long bio_flags)
1911 {
1912         struct inode *inode = page->mapping->host;
1913         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1914         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1915         u64 length = 0;
1916         u64 map_length;
1917         int ret;
1918
1919         if (bio_flags & EXTENT_BIO_COMPRESSED)
1920                 return 0;
1921
1922         length = bio->bi_iter.bi_size;
1923         map_length = length;
1924         ret = btrfs_map_block(fs_info, btrfs_op(bio), logical, &map_length,
1925                               NULL, 0);
1926         if (ret < 0)
1927                 return ret;
1928         if (map_length < length + size)
1929                 return 1;
1930         return 0;
1931 }
1932
1933 /*
1934  * in order to insert checksums into the metadata in large chunks,
1935  * we wait until bio submission time.   All the pages in the bio are
1936  * checksummed and sums are attached onto the ordered extent record.
1937  *
1938  * At IO completion time the cums attached on the ordered extent record
1939  * are inserted into the btree
1940  */
1941 static blk_status_t btrfs_submit_bio_start(void *private_data, struct bio *bio,
1942                                     u64 bio_offset)
1943 {
1944         struct inode *inode = private_data;
1945         blk_status_t ret = 0;
1946
1947         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1948         BUG_ON(ret); /* -ENOMEM */
1949         return 0;
1950 }
1951
1952 /*
1953  * in order to insert checksums into the metadata in large chunks,
1954  * we wait until bio submission time.   All the pages in the bio are
1955  * checksummed and sums are attached onto the ordered extent record.
1956  *
1957  * At IO completion time the cums attached on the ordered extent record
1958  * are inserted into the btree
1959  */
1960 blk_status_t btrfs_submit_bio_done(void *private_data, struct bio *bio,
1961                           int mirror_num)
1962 {
1963         struct inode *inode = private_data;
1964         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1965         blk_status_t ret;
1966
1967         ret = btrfs_map_bio(fs_info, bio, mirror_num, 1);
1968         if (ret) {
1969                 bio->bi_status = ret;
1970                 bio_endio(bio);
1971         }
1972         return ret;
1973 }
1974
1975 /*
1976  * extent_io.c submission hook. This does the right thing for csum calculation
1977  * on write, or reading the csums from the tree before a read.
1978  *
1979  * Rules about async/sync submit,
1980  * a) read:                             sync submit
1981  *
1982  * b) write without checksum:           sync submit
1983  *
1984  * c) write with checksum:
1985  *    c-1) if bio is issued by fsync:   sync submit
1986  *         (sync_writers != 0)
1987  *
1988  *    c-2) if root is reloc root:       sync submit
1989  *         (only in case of buffered IO)
1990  *
1991  *    c-3) otherwise:                   async submit
1992  */
1993 static blk_status_t btrfs_submit_bio_hook(void *private_data, struct bio *bio,
1994                                  int mirror_num, unsigned long bio_flags,
1995                                  u64 bio_offset)
1996 {
1997         struct inode *inode = private_data;
1998         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1999         struct btrfs_root *root = BTRFS_I(inode)->root;
2000         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
2001         blk_status_t ret = 0;
2002         int skip_sum;
2003         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
2004
2005         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
2006
2007         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
2008                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
2009
2010         if (bio_op(bio) != REQ_OP_WRITE) {
2011                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
2012                 if (ret)
2013                         goto out;
2014
2015                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
2016                         ret = btrfs_submit_compressed_read(inode, bio,
2017                                                            mirror_num,
2018                                                            bio_flags);
2019                         goto out;
2020                 } else if (!skip_sum) {
2021                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
2022                         if (ret)
2023                                 goto out;
2024                 }
2025                 goto mapit;
2026         } else if (async && !skip_sum) {
2027                 /* csum items have already been cloned */
2028                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2029                         goto mapit;
2030                 /* we're doing a write, do the async checksumming */
2031                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
2032                                           bio_offset, inode,
2033                                           btrfs_submit_bio_start);
2034                 goto out;
2035         } else if (!skip_sum) {
2036                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2037                 if (ret)
2038                         goto out;
2039         }
2040
2041 mapit:
2042         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
2043
2044 out:
2045         if (ret) {
2046                 bio->bi_status = ret;
2047                 bio_endio(bio);
2048         }
2049         return ret;
2050 }
2051
2052 /*
2053  * given a list of ordered sums record them in the inode.  This happens
2054  * at IO completion time based on sums calculated at bio submission time.
2055  */
2056 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2057                              struct inode *inode, struct list_head *list)
2058 {
2059         struct btrfs_ordered_sum *sum;
2060         int ret;
2061
2062         list_for_each_entry(sum, list, list) {
2063                 trans->adding_csums = true;
2064                 ret = btrfs_csum_file_blocks(trans,
2065                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2066                 trans->adding_csums = false;
2067                 if (ret)
2068                         return ret;
2069         }
2070         return 0;
2071 }
2072
2073 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2074                               unsigned int extra_bits,
2075                               struct extent_state **cached_state, int dedupe)
2076 {
2077         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
2078         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2079                                    extra_bits, cached_state);
2080 }
2081
2082 /* see btrfs_writepage_start_hook for details on why this is required */
2083 struct btrfs_writepage_fixup {
2084         struct page *page;
2085         struct btrfs_work work;
2086 };
2087
2088 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2089 {
2090         struct btrfs_writepage_fixup *fixup;
2091         struct btrfs_ordered_extent *ordered;
2092         struct extent_state *cached_state = NULL;
2093         struct extent_changeset *data_reserved = NULL;
2094         struct page *page;
2095         struct inode *inode;
2096         u64 page_start;
2097         u64 page_end;
2098         int ret;
2099
2100         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2101         page = fixup->page;
2102 again:
2103         lock_page(page);
2104         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2105                 ClearPageChecked(page);
2106                 goto out_page;
2107         }
2108
2109         inode = page->mapping->host;
2110         page_start = page_offset(page);
2111         page_end = page_offset(page) + PAGE_SIZE - 1;
2112
2113         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2114                          &cached_state);
2115
2116         /* already ordered? We're done */
2117         if (PagePrivate2(page))
2118                 goto out;
2119
2120         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2121                                         PAGE_SIZE);
2122         if (ordered) {
2123                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2124                                      page_end, &cached_state);
2125                 unlock_page(page);
2126                 btrfs_start_ordered_extent(inode, ordered, 1);
2127                 btrfs_put_ordered_extent(ordered);
2128                 goto again;
2129         }
2130
2131         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2132                                            PAGE_SIZE);
2133         if (ret) {
2134                 mapping_set_error(page->mapping, ret);
2135                 end_extent_writepage(page, ret, page_start, page_end);
2136                 ClearPageChecked(page);
2137                 goto out;
2138          }
2139
2140         ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
2141                                         &cached_state, 0);
2142         if (ret) {
2143                 mapping_set_error(page->mapping, ret);
2144                 end_extent_writepage(page, ret, page_start, page_end);
2145                 ClearPageChecked(page);
2146                 goto out;
2147         }
2148
2149         ClearPageChecked(page);
2150         set_page_dirty(page);
2151         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE, false);
2152 out:
2153         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2154                              &cached_state);
2155 out_page:
2156         unlock_page(page);
2157         put_page(page);
2158         kfree(fixup);
2159         extent_changeset_free(data_reserved);
2160 }
2161
2162 /*
2163  * There are a few paths in the higher layers of the kernel that directly
2164  * set the page dirty bit without asking the filesystem if it is a
2165  * good idea.  This causes problems because we want to make sure COW
2166  * properly happens and the data=ordered rules are followed.
2167  *
2168  * In our case any range that doesn't have the ORDERED bit set
2169  * hasn't been properly setup for IO.  We kick off an async process
2170  * to fix it up.  The async helper will wait for ordered extents, set
2171  * the delalloc bit and make it safe to write the page.
2172  */
2173 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2174 {
2175         struct inode *inode = page->mapping->host;
2176         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2177         struct btrfs_writepage_fixup *fixup;
2178
2179         /* this page is properly in the ordered list */
2180         if (TestClearPagePrivate2(page))
2181                 return 0;
2182
2183         if (PageChecked(page))
2184                 return -EAGAIN;
2185
2186         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2187         if (!fixup)
2188                 return -EAGAIN;
2189
2190         SetPageChecked(page);
2191         get_page(page);
2192         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2193                         btrfs_writepage_fixup_worker, NULL, NULL);
2194         fixup->page = page;
2195         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2196         return -EBUSY;
2197 }
2198
2199 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2200                                        struct inode *inode, u64 file_pos,
2201                                        u64 disk_bytenr, u64 disk_num_bytes,
2202                                        u64 num_bytes, u64 ram_bytes,
2203                                        u8 compression, u8 encryption,
2204                                        u16 other_encoding, int extent_type)
2205 {
2206         struct btrfs_root *root = BTRFS_I(inode)->root;
2207         struct btrfs_file_extent_item *fi;
2208         struct btrfs_path *path;
2209         struct extent_buffer *leaf;
2210         struct btrfs_key ins;
2211         u64 qg_released;
2212         int extent_inserted = 0;
2213         int ret;
2214
2215         path = btrfs_alloc_path();
2216         if (!path)
2217                 return -ENOMEM;
2218
2219         /*
2220          * we may be replacing one extent in the tree with another.
2221          * The new extent is pinned in the extent map, and we don't want
2222          * to drop it from the cache until it is completely in the btree.
2223          *
2224          * So, tell btrfs_drop_extents to leave this extent in the cache.
2225          * the caller is expected to unpin it and allow it to be merged
2226          * with the others.
2227          */
2228         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2229                                    file_pos + num_bytes, NULL, 0,
2230                                    1, sizeof(*fi), &extent_inserted);
2231         if (ret)
2232                 goto out;
2233
2234         if (!extent_inserted) {
2235                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2236                 ins.offset = file_pos;
2237                 ins.type = BTRFS_EXTENT_DATA_KEY;
2238
2239                 path->leave_spinning = 1;
2240                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2241                                               sizeof(*fi));
2242                 if (ret)
2243                         goto out;
2244         }
2245         leaf = path->nodes[0];
2246         fi = btrfs_item_ptr(leaf, path->slots[0],
2247                             struct btrfs_file_extent_item);
2248         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2249         btrfs_set_file_extent_type(leaf, fi, extent_type);
2250         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2251         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2252         btrfs_set_file_extent_offset(leaf, fi, 0);
2253         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2254         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2255         btrfs_set_file_extent_compression(leaf, fi, compression);
2256         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2257         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2258
2259         btrfs_mark_buffer_dirty(leaf);
2260         btrfs_release_path(path);
2261
2262         inode_add_bytes(inode, num_bytes);
2263
2264         ins.objectid = disk_bytenr;
2265         ins.offset = disk_num_bytes;
2266         ins.type = BTRFS_EXTENT_ITEM_KEY;
2267
2268         /*
2269          * Release the reserved range from inode dirty range map, as it is
2270          * already moved into delayed_ref_head
2271          */
2272         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2273         if (ret < 0)
2274                 goto out;
2275         qg_released = ret;
2276         ret = btrfs_alloc_reserved_file_extent(trans, root,
2277                                                btrfs_ino(BTRFS_I(inode)),
2278                                                file_pos, qg_released, &ins);
2279 out:
2280         btrfs_free_path(path);
2281
2282         return ret;
2283 }
2284
2285 /* snapshot-aware defrag */
2286 struct sa_defrag_extent_backref {
2287         struct rb_node node;
2288         struct old_sa_defrag_extent *old;
2289         u64 root_id;
2290         u64 inum;
2291         u64 file_pos;
2292         u64 extent_offset;
2293         u64 num_bytes;
2294         u64 generation;
2295 };
2296
2297 struct old_sa_defrag_extent {
2298         struct list_head list;
2299         struct new_sa_defrag_extent *new;
2300
2301         u64 extent_offset;
2302         u64 bytenr;
2303         u64 offset;
2304         u64 len;
2305         int count;
2306 };
2307
2308 struct new_sa_defrag_extent {
2309         struct rb_root root;
2310         struct list_head head;
2311         struct btrfs_path *path;
2312         struct inode *inode;
2313         u64 file_pos;
2314         u64 len;
2315         u64 bytenr;
2316         u64 disk_len;
2317         u8 compress_type;
2318 };
2319
2320 static int backref_comp(struct sa_defrag_extent_backref *b1,
2321                         struct sa_defrag_extent_backref *b2)
2322 {
2323         if (b1->root_id < b2->root_id)
2324                 return -1;
2325         else if (b1->root_id > b2->root_id)
2326                 return 1;
2327
2328         if (b1->inum < b2->inum)
2329                 return -1;
2330         else if (b1->inum > b2->inum)
2331                 return 1;
2332
2333         if (b1->file_pos < b2->file_pos)
2334                 return -1;
2335         else if (b1->file_pos > b2->file_pos)
2336                 return 1;
2337
2338         /*
2339          * [------------------------------] ===> (a range of space)
2340          *     |<--->|   |<---->| =============> (fs/file tree A)
2341          * |<---------------------------->| ===> (fs/file tree B)
2342          *
2343          * A range of space can refer to two file extents in one tree while
2344          * refer to only one file extent in another tree.
2345          *
2346          * So we may process a disk offset more than one time(two extents in A)
2347          * and locate at the same extent(one extent in B), then insert two same
2348          * backrefs(both refer to the extent in B).
2349          */
2350         return 0;
2351 }
2352
2353 static void backref_insert(struct rb_root *root,
2354                            struct sa_defrag_extent_backref *backref)
2355 {
2356         struct rb_node **p = &root->rb_node;
2357         struct rb_node *parent = NULL;
2358         struct sa_defrag_extent_backref *entry;
2359         int ret;
2360
2361         while (*p) {
2362                 parent = *p;
2363                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2364
2365                 ret = backref_comp(backref, entry);
2366                 if (ret < 0)
2367                         p = &(*p)->rb_left;
2368                 else
2369                         p = &(*p)->rb_right;
2370         }
2371
2372         rb_link_node(&backref->node, parent, p);
2373         rb_insert_color(&backref->node, root);
2374 }
2375
2376 /*
2377  * Note the backref might has changed, and in this case we just return 0.
2378  */
2379 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2380                                        void *ctx)
2381 {
2382         struct btrfs_file_extent_item *extent;
2383         struct old_sa_defrag_extent *old = ctx;
2384         struct new_sa_defrag_extent *new = old->new;
2385         struct btrfs_path *path = new->path;
2386         struct btrfs_key key;
2387         struct btrfs_root *root;
2388         struct sa_defrag_extent_backref *backref;
2389         struct extent_buffer *leaf;
2390         struct inode *inode = new->inode;
2391         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2392         int slot;
2393         int ret;
2394         u64 extent_offset;
2395         u64 num_bytes;
2396
2397         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2398             inum == btrfs_ino(BTRFS_I(inode)))
2399                 return 0;
2400
2401         key.objectid = root_id;
2402         key.type = BTRFS_ROOT_ITEM_KEY;
2403         key.offset = (u64)-1;
2404
2405         root = btrfs_read_fs_root_no_name(fs_info, &key);
2406         if (IS_ERR(root)) {
2407                 if (PTR_ERR(root) == -ENOENT)
2408                         return 0;
2409                 WARN_ON(1);
2410                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2411                          inum, offset, root_id);
2412                 return PTR_ERR(root);
2413         }
2414
2415         key.objectid = inum;
2416         key.type = BTRFS_EXTENT_DATA_KEY;
2417         if (offset > (u64)-1 << 32)
2418                 key.offset = 0;
2419         else
2420                 key.offset = offset;
2421
2422         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2423         if (WARN_ON(ret < 0))
2424                 return ret;
2425         ret = 0;
2426
2427         while (1) {
2428                 cond_resched();
2429
2430                 leaf = path->nodes[0];
2431                 slot = path->slots[0];
2432
2433                 if (slot >= btrfs_header_nritems(leaf)) {
2434                         ret = btrfs_next_leaf(root, path);
2435                         if (ret < 0) {
2436                                 goto out;
2437                         } else if (ret > 0) {
2438                                 ret = 0;
2439                                 goto out;
2440                         }
2441                         continue;
2442                 }
2443
2444                 path->slots[0]++;
2445
2446                 btrfs_item_key_to_cpu(leaf, &key, slot);
2447
2448                 if (key.objectid > inum)
2449                         goto out;
2450
2451                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2452                         continue;
2453
2454                 extent = btrfs_item_ptr(leaf, slot,
2455                                         struct btrfs_file_extent_item);
2456
2457                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2458                         continue;
2459
2460                 /*
2461                  * 'offset' refers to the exact key.offset,
2462                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2463                  * (key.offset - extent_offset).
2464                  */
2465                 if (key.offset != offset)
2466                         continue;
2467
2468                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2469                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2470
2471                 if (extent_offset >= old->extent_offset + old->offset +
2472                     old->len || extent_offset + num_bytes <=
2473                     old->extent_offset + old->offset)
2474                         continue;
2475                 break;
2476         }
2477
2478         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2479         if (!backref) {
2480                 ret = -ENOENT;
2481                 goto out;
2482         }
2483
2484         backref->root_id = root_id;
2485         backref->inum = inum;
2486         backref->file_pos = offset;
2487         backref->num_bytes = num_bytes;
2488         backref->extent_offset = extent_offset;
2489         backref->generation = btrfs_file_extent_generation(leaf, extent);
2490         backref->old = old;
2491         backref_insert(&new->root, backref);
2492         old->count++;
2493 out:
2494         btrfs_release_path(path);
2495         WARN_ON(ret);
2496         return ret;
2497 }
2498
2499 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2500                                    struct new_sa_defrag_extent *new)
2501 {
2502         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2503         struct old_sa_defrag_extent *old, *tmp;
2504         int ret;
2505
2506         new->path = path;
2507
2508         list_for_each_entry_safe(old, tmp, &new->head, list) {
2509                 ret = iterate_inodes_from_logical(old->bytenr +
2510                                                   old->extent_offset, fs_info,
2511                                                   path, record_one_backref,
2512                                                   old, false);
2513                 if (ret < 0 && ret != -ENOENT)
2514                         return false;
2515
2516                 /* no backref to be processed for this extent */
2517                 if (!old->count) {
2518                         list_del(&old->list);
2519                         kfree(old);
2520                 }
2521         }
2522
2523         if (list_empty(&new->head))
2524                 return false;
2525
2526         return true;
2527 }
2528
2529 static int relink_is_mergable(struct extent_buffer *leaf,
2530                               struct btrfs_file_extent_item *fi,
2531                               struct new_sa_defrag_extent *new)
2532 {
2533         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2534                 return 0;
2535
2536         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2537                 return 0;
2538
2539         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2540                 return 0;
2541
2542         if (btrfs_file_extent_encryption(leaf, fi) ||
2543             btrfs_file_extent_other_encoding(leaf, fi))
2544                 return 0;
2545
2546         return 1;
2547 }
2548
2549 /*
2550  * Note the backref might has changed, and in this case we just return 0.
2551  */
2552 static noinline int relink_extent_backref(struct btrfs_path *path,
2553                                  struct sa_defrag_extent_backref *prev,
2554                                  struct sa_defrag_extent_backref *backref)
2555 {
2556         struct btrfs_file_extent_item *extent;
2557         struct btrfs_file_extent_item *item;
2558         struct btrfs_ordered_extent *ordered;
2559         struct btrfs_trans_handle *trans;
2560         struct btrfs_root *root;
2561         struct btrfs_key key;
2562         struct extent_buffer *leaf;
2563         struct old_sa_defrag_extent *old = backref->old;
2564         struct new_sa_defrag_extent *new = old->new;
2565         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2566         struct inode *inode;
2567         struct extent_state *cached = NULL;
2568         int ret = 0;
2569         u64 start;
2570         u64 len;
2571         u64 lock_start;
2572         u64 lock_end;
2573         bool merge = false;
2574         int index;
2575
2576         if (prev && prev->root_id == backref->root_id &&
2577             prev->inum == backref->inum &&
2578             prev->file_pos + prev->num_bytes == backref->file_pos)
2579                 merge = true;
2580
2581         /* step 1: get root */
2582         key.objectid = backref->root_id;
2583         key.type = BTRFS_ROOT_ITEM_KEY;
2584         key.offset = (u64)-1;
2585
2586         index = srcu_read_lock(&fs_info->subvol_srcu);
2587
2588         root = btrfs_read_fs_root_no_name(fs_info, &key);
2589         if (IS_ERR(root)) {
2590                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2591                 if (PTR_ERR(root) == -ENOENT)
2592                         return 0;
2593                 return PTR_ERR(root);
2594         }
2595
2596         if (btrfs_root_readonly(root)) {
2597                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2598                 return 0;
2599         }
2600
2601         /* step 2: get inode */
2602         key.objectid = backref->inum;
2603         key.type = BTRFS_INODE_ITEM_KEY;
2604         key.offset = 0;
2605
2606         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2607         if (IS_ERR(inode)) {
2608                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2609                 return 0;
2610         }
2611
2612         srcu_read_unlock(&fs_info->subvol_srcu, index);
2613
2614         /* step 3: relink backref */
2615         lock_start = backref->file_pos;
2616         lock_end = backref->file_pos + backref->num_bytes - 1;
2617         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2618                          &cached);
2619
2620         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2621         if (ordered) {
2622                 btrfs_put_ordered_extent(ordered);
2623                 goto out_unlock;
2624         }
2625
2626         trans = btrfs_join_transaction(root);
2627         if (IS_ERR(trans)) {
2628                 ret = PTR_ERR(trans);
2629                 goto out_unlock;
2630         }
2631
2632         key.objectid = backref->inum;
2633         key.type = BTRFS_EXTENT_DATA_KEY;
2634         key.offset = backref->file_pos;
2635
2636         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2637         if (ret < 0) {
2638                 goto out_free_path;
2639         } else if (ret > 0) {
2640                 ret = 0;
2641                 goto out_free_path;
2642         }
2643
2644         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2645                                 struct btrfs_file_extent_item);
2646
2647         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2648             backref->generation)
2649                 goto out_free_path;
2650
2651         btrfs_release_path(path);
2652
2653         start = backref->file_pos;
2654         if (backref->extent_offset < old->extent_offset + old->offset)
2655                 start += old->extent_offset + old->offset -
2656                          backref->extent_offset;
2657
2658         len = min(backref->extent_offset + backref->num_bytes,
2659                   old->extent_offset + old->offset + old->len);
2660         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2661
2662         ret = btrfs_drop_extents(trans, root, inode, start,
2663                                  start + len, 1);
2664         if (ret)
2665                 goto out_free_path;
2666 again:
2667         key.objectid = btrfs_ino(BTRFS_I(inode));
2668         key.type = BTRFS_EXTENT_DATA_KEY;
2669         key.offset = start;
2670
2671         path->leave_spinning = 1;
2672         if (merge) {
2673                 struct btrfs_file_extent_item *fi;
2674                 u64 extent_len;
2675                 struct btrfs_key found_key;
2676
2677                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2678                 if (ret < 0)
2679                         goto out_free_path;
2680
2681                 path->slots[0]--;
2682                 leaf = path->nodes[0];
2683                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2684
2685                 fi = btrfs_item_ptr(leaf, path->slots[0],
2686                                     struct btrfs_file_extent_item);
2687                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2688
2689                 if (extent_len + found_key.offset == start &&
2690                     relink_is_mergable(leaf, fi, new)) {
2691                         btrfs_set_file_extent_num_bytes(leaf, fi,
2692                                                         extent_len + len);
2693                         btrfs_mark_buffer_dirty(leaf);
2694                         inode_add_bytes(inode, len);
2695
2696                         ret = 1;
2697                         goto out_free_path;
2698                 } else {
2699                         merge = false;
2700                         btrfs_release_path(path);
2701                         goto again;
2702                 }
2703         }
2704
2705         ret = btrfs_insert_empty_item(trans, root, path, &key,
2706                                         sizeof(*extent));
2707         if (ret) {
2708                 btrfs_abort_transaction(trans, ret);
2709                 goto out_free_path;
2710         }
2711
2712         leaf = path->nodes[0];
2713         item = btrfs_item_ptr(leaf, path->slots[0],
2714                                 struct btrfs_file_extent_item);
2715         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2716         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2717         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2718         btrfs_set_file_extent_num_bytes(leaf, item, len);
2719         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2720         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2721         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2722         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2723         btrfs_set_file_extent_encryption(leaf, item, 0);
2724         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2725
2726         btrfs_mark_buffer_dirty(leaf);
2727         inode_add_bytes(inode, len);
2728         btrfs_release_path(path);
2729
2730         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2731                         new->disk_len, 0,
2732                         backref->root_id, backref->inum,
2733                         new->file_pos); /* start - extent_offset */
2734         if (ret) {
2735                 btrfs_abort_transaction(trans, ret);
2736                 goto out_free_path;
2737         }
2738
2739         ret = 1;
2740 out_free_path:
2741         btrfs_release_path(path);
2742         path->leave_spinning = 0;
2743         btrfs_end_transaction(trans);
2744 out_unlock:
2745         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2746                              &cached);
2747         iput(inode);
2748         return ret;
2749 }
2750
2751 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2752 {
2753         struct old_sa_defrag_extent *old, *tmp;
2754
2755         if (!new)
2756                 return;
2757
2758         list_for_each_entry_safe(old, tmp, &new->head, list) {
2759                 kfree(old);
2760         }
2761         kfree(new);
2762 }
2763
2764 static void relink_file_extents(struct new_sa_defrag_extent *new)
2765 {
2766         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2767         struct btrfs_path *path;
2768         struct sa_defrag_extent_backref *backref;
2769         struct sa_defrag_extent_backref *prev = NULL;
2770         struct inode *inode;
2771         struct rb_node *node;
2772         int ret;
2773
2774         inode = new->inode;
2775
2776         path = btrfs_alloc_path();
2777         if (!path)
2778                 return;
2779
2780         if (!record_extent_backrefs(path, new)) {
2781                 btrfs_free_path(path);
2782                 goto out;
2783         }
2784         btrfs_release_path(path);
2785
2786         while (1) {
2787                 node = rb_first(&new->root);
2788                 if (!node)
2789                         break;
2790                 rb_erase(node, &new->root);
2791
2792                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2793
2794                 ret = relink_extent_backref(path, prev, backref);
2795                 WARN_ON(ret < 0);
2796
2797                 kfree(prev);
2798
2799                 if (ret == 1)
2800                         prev = backref;
2801                 else
2802                         prev = NULL;
2803                 cond_resched();
2804         }
2805         kfree(prev);
2806
2807         btrfs_free_path(path);
2808 out:
2809         free_sa_defrag_extent(new);
2810
2811         atomic_dec(&fs_info->defrag_running);
2812         wake_up(&fs_info->transaction_wait);
2813 }
2814
2815 static struct new_sa_defrag_extent *
2816 record_old_file_extents(struct inode *inode,
2817                         struct btrfs_ordered_extent *ordered)
2818 {
2819         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2820         struct btrfs_root *root = BTRFS_I(inode)->root;
2821         struct btrfs_path *path;
2822         struct btrfs_key key;
2823         struct old_sa_defrag_extent *old;
2824         struct new_sa_defrag_extent *new;
2825         int ret;
2826
2827         new = kmalloc(sizeof(*new), GFP_NOFS);
2828         if (!new)
2829                 return NULL;
2830
2831         new->inode = inode;
2832         new->file_pos = ordered->file_offset;
2833         new->len = ordered->len;
2834         new->bytenr = ordered->start;
2835         new->disk_len = ordered->disk_len;
2836         new->compress_type = ordered->compress_type;
2837         new->root = RB_ROOT;
2838         INIT_LIST_HEAD(&new->head);
2839
2840         path = btrfs_alloc_path();
2841         if (!path)
2842                 goto out_kfree;
2843
2844         key.objectid = btrfs_ino(BTRFS_I(inode));
2845         key.type = BTRFS_EXTENT_DATA_KEY;
2846         key.offset = new->file_pos;
2847
2848         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2849         if (ret < 0)
2850                 goto out_free_path;
2851         if (ret > 0 && path->slots[0] > 0)
2852                 path->slots[0]--;
2853
2854         /* find out all the old extents for the file range */
2855         while (1) {
2856                 struct btrfs_file_extent_item *extent;
2857                 struct extent_buffer *l;
2858                 int slot;
2859                 u64 num_bytes;
2860                 u64 offset;
2861                 u64 end;
2862                 u64 disk_bytenr;
2863                 u64 extent_offset;
2864
2865                 l = path->nodes[0];
2866                 slot = path->slots[0];
2867
2868                 if (slot >= btrfs_header_nritems(l)) {
2869                         ret = btrfs_next_leaf(root, path);
2870                         if (ret < 0)
2871                                 goto out_free_path;
2872                         else if (ret > 0)
2873                                 break;
2874                         continue;
2875                 }
2876
2877                 btrfs_item_key_to_cpu(l, &key, slot);
2878
2879                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2880                         break;
2881                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2882                         break;
2883                 if (key.offset >= new->file_pos + new->len)
2884                         break;
2885
2886                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2887
2888                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2889                 if (key.offset + num_bytes < new->file_pos)
2890                         goto next;
2891
2892                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2893                 if (!disk_bytenr)
2894                         goto next;
2895
2896                 extent_offset = btrfs_file_extent_offset(l, extent);
2897
2898                 old = kmalloc(sizeof(*old), GFP_NOFS);
2899                 if (!old)
2900                         goto out_free_path;
2901
2902                 offset = max(new->file_pos, key.offset);
2903                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2904
2905                 old->bytenr = disk_bytenr;
2906                 old->extent_offset = extent_offset;
2907                 old->offset = offset - key.offset;
2908                 old->len = end - offset;
2909                 old->new = new;
2910                 old->count = 0;
2911                 list_add_tail(&old->list, &new->head);
2912 next:
2913                 path->slots[0]++;
2914                 cond_resched();
2915         }
2916
2917         btrfs_free_path(path);
2918         atomic_inc(&fs_info->defrag_running);
2919
2920         return new;
2921
2922 out_free_path:
2923         btrfs_free_path(path);
2924 out_kfree:
2925         free_sa_defrag_extent(new);
2926         return NULL;
2927 }
2928
2929 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2930                                          u64 start, u64 len)
2931 {
2932         struct btrfs_block_group_cache *cache;
2933
2934         cache = btrfs_lookup_block_group(fs_info, start);
2935         ASSERT(cache);
2936
2937         spin_lock(&cache->lock);
2938         cache->delalloc_bytes -= len;
2939         spin_unlock(&cache->lock);
2940
2941         btrfs_put_block_group(cache);
2942 }
2943
2944 /* as ordered data IO finishes, this gets called so we can finish
2945  * an ordered extent if the range of bytes in the file it covers are
2946  * fully written.
2947  */
2948 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2949 {
2950         struct inode *inode = ordered_extent->inode;
2951         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2952         struct btrfs_root *root = BTRFS_I(inode)->root;
2953         struct btrfs_trans_handle *trans = NULL;
2954         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2955         struct extent_state *cached_state = NULL;
2956         struct new_sa_defrag_extent *new = NULL;
2957         int compress_type = 0;
2958         int ret = 0;
2959         u64 logical_len = ordered_extent->len;
2960         bool nolock;
2961         bool truncated = false;
2962         bool range_locked = false;
2963         bool clear_new_delalloc_bytes = false;
2964
2965         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2966             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
2967             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
2968                 clear_new_delalloc_bytes = true;
2969
2970         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
2971
2972         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2973                 ret = -EIO;
2974                 goto out;
2975         }
2976
2977         btrfs_free_io_failure_record(BTRFS_I(inode),
2978                         ordered_extent->file_offset,
2979                         ordered_extent->file_offset +
2980                         ordered_extent->len - 1);
2981
2982         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2983                 truncated = true;
2984                 logical_len = ordered_extent->truncated_len;
2985                 /* Truncated the entire extent, don't bother adding */
2986                 if (!logical_len)
2987                         goto out;
2988         }
2989
2990         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2991                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2992
2993                 /*
2994                  * For mwrite(mmap + memset to write) case, we still reserve
2995                  * space for NOCOW range.
2996                  * As NOCOW won't cause a new delayed ref, just free the space
2997                  */
2998                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
2999                                        ordered_extent->len);
3000                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3001                 if (nolock)
3002                         trans = btrfs_join_transaction_nolock(root);
3003                 else
3004                         trans = btrfs_join_transaction(root);
3005                 if (IS_ERR(trans)) {
3006                         ret = PTR_ERR(trans);
3007                         trans = NULL;
3008                         goto out;
3009                 }
3010                 trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3011                 ret = btrfs_update_inode_fallback(trans, root, inode);
3012                 if (ret) /* -ENOMEM or corruption */
3013                         btrfs_abort_transaction(trans, ret);
3014                 goto out;
3015         }
3016
3017         range_locked = true;
3018         lock_extent_bits(io_tree, ordered_extent->file_offset,
3019                          ordered_extent->file_offset + ordered_extent->len - 1,
3020                          &cached_state);
3021
3022         ret = test_range_bit(io_tree, ordered_extent->file_offset,
3023                         ordered_extent->file_offset + ordered_extent->len - 1,
3024                         EXTENT_DEFRAG, 0, cached_state);
3025         if (ret) {
3026                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
3027                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
3028                         /* the inode is shared */
3029                         new = record_old_file_extents(inode, ordered_extent);
3030
3031                 clear_extent_bit(io_tree, ordered_extent->file_offset,
3032                         ordered_extent->file_offset + ordered_extent->len - 1,
3033                         EXTENT_DEFRAG, 0, 0, &cached_state);
3034         }
3035
3036         if (nolock)
3037                 trans = btrfs_join_transaction_nolock(root);
3038         else
3039                 trans = btrfs_join_transaction(root);
3040         if (IS_ERR(trans)) {
3041                 ret = PTR_ERR(trans);
3042                 trans = NULL;
3043                 goto out;
3044         }
3045
3046         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3047
3048         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
3049                 compress_type = ordered_extent->compress_type;
3050         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
3051                 BUG_ON(compress_type);
3052                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3053                                        ordered_extent->len);
3054                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
3055                                                 ordered_extent->file_offset,
3056                                                 ordered_extent->file_offset +
3057                                                 logical_len);
3058         } else {
3059                 BUG_ON(root == fs_info->tree_root);
3060                 ret = insert_reserved_file_extent(trans, inode,
3061                                                 ordered_extent->file_offset,
3062                                                 ordered_extent->start,
3063                                                 ordered_extent->disk_len,
3064                                                 logical_len, logical_len,
3065                                                 compress_type, 0, 0,
3066                                                 BTRFS_FILE_EXTENT_REG);
3067                 if (!ret)
3068                         btrfs_release_delalloc_bytes(fs_info,
3069                                                      ordered_extent->start,
3070                                                      ordered_extent->disk_len);
3071         }
3072         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
3073                            ordered_extent->file_offset, ordered_extent->len,
3074                            trans->transid);
3075         if (ret < 0) {
3076                 btrfs_abort_transaction(trans, ret);
3077                 goto out;
3078         }
3079
3080         ret = add_pending_csums(trans, inode, &ordered_extent->list);
3081         if (ret) {
3082                 btrfs_abort_transaction(trans, ret);
3083                 goto out;
3084         }
3085
3086         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3087         ret = btrfs_update_inode_fallback(trans, root, inode);
3088         if (ret) { /* -ENOMEM or corruption */
3089                 btrfs_abort_transaction(trans, ret);
3090                 goto out;
3091         }
3092         ret = 0;
3093 out:
3094         if (range_locked || clear_new_delalloc_bytes) {
3095                 unsigned int clear_bits = 0;
3096
3097                 if (range_locked)
3098                         clear_bits |= EXTENT_LOCKED;
3099                 if (clear_new_delalloc_bytes)
3100                         clear_bits |= EXTENT_DELALLOC_NEW;
3101                 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3102                                  ordered_extent->file_offset,
3103                                  ordered_extent->file_offset +
3104                                  ordered_extent->len - 1,
3105                                  clear_bits,
3106                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
3107                                  0, &cached_state);
3108         }
3109
3110         if (trans)
3111                 btrfs_end_transaction(trans);
3112
3113         if (ret || truncated) {
3114                 u64 start, end;
3115
3116                 if (truncated)
3117                         start = ordered_extent->file_offset + logical_len;
3118                 else
3119                         start = ordered_extent->file_offset;
3120                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3121                 clear_extent_uptodate(io_tree, start, end, NULL);
3122
3123                 /* Drop the cache for the part of the extent we didn't write. */
3124                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3125
3126                 /*
3127                  * If the ordered extent had an IOERR or something else went
3128                  * wrong we need to return the space for this ordered extent
3129                  * back to the allocator.  We only free the extent in the
3130                  * truncated case if we didn't write out the extent at all.
3131                  */
3132                 if ((ret || !logical_len) &&
3133                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3134                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3135                         btrfs_free_reserved_extent(fs_info,
3136                                                    ordered_extent->start,
3137                                                    ordered_extent->disk_len, 1);
3138         }
3139
3140
3141         /*
3142          * This needs to be done to make sure anybody waiting knows we are done
3143          * updating everything for this ordered extent.
3144          */
3145         btrfs_remove_ordered_extent(inode, ordered_extent);
3146
3147         /* for snapshot-aware defrag */
3148         if (new) {
3149                 if (ret) {
3150                         free_sa_defrag_extent(new);
3151                         atomic_dec(&fs_info->defrag_running);
3152                 } else {
3153                         relink_file_extents(new);
3154                 }
3155         }
3156
3157         /* once for us */
3158         btrfs_put_ordered_extent(ordered_extent);
3159         /* once for the tree */
3160         btrfs_put_ordered_extent(ordered_extent);
3161
3162         /* Try to release some metadata so we don't get an OOM but don't wait */
3163         btrfs_btree_balance_dirty_nodelay(fs_info);
3164
3165         return ret;
3166 }
3167
3168 static void finish_ordered_fn(struct btrfs_work *work)
3169 {
3170         struct btrfs_ordered_extent *ordered_extent;
3171         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3172         btrfs_finish_ordered_io(ordered_extent);
3173 }
3174
3175 static void btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3176                                 struct extent_state *state, int uptodate)
3177 {
3178         struct inode *inode = page->mapping->host;
3179         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3180         struct btrfs_ordered_extent *ordered_extent = NULL;
3181         struct btrfs_workqueue *wq;
3182         btrfs_work_func_t func;
3183
3184         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3185
3186         ClearPagePrivate2(page);
3187         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3188                                             end - start + 1, uptodate))
3189                 return;
3190
3191         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
3192                 wq = fs_info->endio_freespace_worker;
3193                 func = btrfs_freespace_write_helper;
3194         } else {
3195                 wq = fs_info->endio_write_workers;
3196                 func = btrfs_endio_write_helper;
3197         }
3198
3199         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3200                         NULL);
3201         btrfs_queue_work(wq, &ordered_extent->work);
3202 }
3203
3204 static int __readpage_endio_check(struct inode *inode,
3205                                   struct btrfs_io_bio *io_bio,
3206                                   int icsum, struct page *page,
3207                                   int pgoff, u64 start, size_t len)
3208 {
3209         char *kaddr;
3210         u32 csum_expected;
3211         u32 csum = ~(u32)0;
3212
3213         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3214
3215         kaddr = kmap_atomic(page);
3216         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3217         btrfs_csum_final(csum, (u8 *)&csum);
3218         if (csum != csum_expected)
3219                 goto zeroit;
3220
3221         kunmap_atomic(kaddr);
3222         return 0;
3223 zeroit:
3224         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
3225                                     io_bio->mirror_num);
3226         memset(kaddr + pgoff, 1, len);
3227         flush_dcache_page(page);
3228         kunmap_atomic(kaddr);
3229         return -EIO;
3230 }
3231
3232 /*
3233  * when reads are done, we need to check csums to verify the data is correct
3234  * if there's a match, we allow the bio to finish.  If not, the code in
3235  * extent_io.c will try to find good copies for us.
3236  */
3237 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3238                                       u64 phy_offset, struct page *page,
3239                                       u64 start, u64 end, int mirror)
3240 {
3241         size_t offset = start - page_offset(page);
3242         struct inode *inode = page->mapping->host;
3243         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3244         struct btrfs_root *root = BTRFS_I(inode)->root;
3245
3246         if (PageChecked(page)) {
3247                 ClearPageChecked(page);
3248                 return 0;
3249         }
3250
3251         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3252                 return 0;
3253
3254         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3255             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3256                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3257                 return 0;
3258         }
3259
3260         phy_offset >>= inode->i_sb->s_blocksize_bits;
3261         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3262                                       start, (size_t)(end - start + 1));
3263 }
3264
3265 /*
3266  * btrfs_add_delayed_iput - perform a delayed iput on @inode
3267  *
3268  * @inode: The inode we want to perform iput on
3269  *
3270  * This function uses the generic vfs_inode::i_count to track whether we should
3271  * just decrement it (in case it's > 1) or if this is the last iput then link
3272  * the inode to the delayed iput machinery. Delayed iputs are processed at
3273  * transaction commit time/superblock commit/cleaner kthread.
3274  */
3275 void btrfs_add_delayed_iput(struct inode *inode)
3276 {
3277         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3278         struct btrfs_inode *binode = BTRFS_I(inode);
3279
3280         if (atomic_add_unless(&inode->i_count, -1, 1))
3281                 return;
3282
3283         spin_lock(&fs_info->delayed_iput_lock);
3284         ASSERT(list_empty(&binode->delayed_iput));
3285         list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3286         spin_unlock(&fs_info->delayed_iput_lock);
3287 }
3288
3289 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3290 {
3291
3292         spin_lock(&fs_info->delayed_iput_lock);
3293         while (!list_empty(&fs_info->delayed_iputs)) {
3294                 struct btrfs_inode *inode;
3295
3296                 inode = list_first_entry(&fs_info->delayed_iputs,
3297                                 struct btrfs_inode, delayed_iput);
3298                 list_del_init(&inode->delayed_iput);
3299                 spin_unlock(&fs_info->delayed_iput_lock);
3300                 iput(&inode->vfs_inode);
3301                 spin_lock(&fs_info->delayed_iput_lock);
3302         }
3303         spin_unlock(&fs_info->delayed_iput_lock);
3304 }
3305
3306 /*
3307  * This creates an orphan entry for the given inode in case something goes wrong
3308  * in the middle of an unlink.
3309  */
3310 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
3311                      struct btrfs_inode *inode)
3312 {
3313         int ret;
3314
3315         ret = btrfs_insert_orphan_item(trans, inode->root, btrfs_ino(inode));
3316         if (ret && ret != -EEXIST) {
3317                 btrfs_abort_transaction(trans, ret);
3318                 return ret;
3319         }
3320
3321         return 0;
3322 }
3323
3324 /*
3325  * We have done the delete so we can go ahead and remove the orphan item for
3326  * this particular inode.
3327  */
3328 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3329                             struct btrfs_inode *inode)
3330 {
3331         return btrfs_del_orphan_item(trans, inode->root, btrfs_ino(inode));
3332 }
3333
3334 /*
3335  * this cleans up any orphans that may be left on the list from the last use
3336  * of this root.
3337  */
3338 int btrfs_orphan_cleanup(struct btrfs_root *root)
3339 {
3340         struct btrfs_fs_info *fs_info = root->fs_info;
3341         struct btrfs_path *path;
3342         struct extent_buffer *leaf;
3343         struct btrfs_key key, found_key;
3344         struct btrfs_trans_handle *trans;
3345         struct inode *inode;
3346         u64 last_objectid = 0;
3347         int ret = 0, nr_unlink = 0;
3348
3349         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3350                 return 0;
3351
3352         path = btrfs_alloc_path();
3353         if (!path) {
3354                 ret = -ENOMEM;
3355                 goto out;
3356         }
3357         path->reada = READA_BACK;
3358
3359         key.objectid = BTRFS_ORPHAN_OBJECTID;
3360         key.type = BTRFS_ORPHAN_ITEM_KEY;
3361         key.offset = (u64)-1;
3362
3363         while (1) {
3364                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3365                 if (ret < 0)
3366                         goto out;
3367
3368                 /*
3369                  * if ret == 0 means we found what we were searching for, which
3370                  * is weird, but possible, so only screw with path if we didn't
3371                  * find the key and see if we have stuff that matches
3372                  */
3373                 if (ret > 0) {
3374                         ret = 0;
3375                         if (path->slots[0] == 0)
3376                                 break;
3377                         path->slots[0]--;
3378                 }
3379
3380                 /* pull out the item */
3381                 leaf = path->nodes[0];
3382                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3383
3384                 /* make sure the item matches what we want */
3385                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3386                         break;
3387                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3388                         break;
3389
3390                 /* release the path since we're done with it */
3391                 btrfs_release_path(path);
3392
3393                 /*
3394                  * this is where we are basically btrfs_lookup, without the
3395                  * crossing root thing.  we store the inode number in the
3396                  * offset of the orphan item.
3397                  */
3398
3399                 if (found_key.offset == last_objectid) {
3400                         btrfs_err(fs_info,
3401                                   "Error removing orphan entry, stopping orphan cleanup");
3402                         ret = -EINVAL;
3403                         goto out;
3404                 }
3405
3406                 last_objectid = found_key.offset;
3407
3408                 found_key.objectid = found_key.offset;
3409                 found_key.type = BTRFS_INODE_ITEM_KEY;
3410                 found_key.offset = 0;
3411                 inode = btrfs_iget(fs_info->sb, &found_key, root, NULL);
3412                 ret = PTR_ERR_OR_ZERO(inode);
3413                 if (ret && ret != -ENOENT)
3414                         goto out;
3415
3416                 if (ret == -ENOENT && root == fs_info->tree_root) {
3417                         struct btrfs_root *dead_root;
3418                         struct btrfs_fs_info *fs_info = root->fs_info;
3419                         int is_dead_root = 0;
3420
3421                         /*
3422                          * this is an orphan in the tree root. Currently these
3423                          * could come from 2 sources:
3424                          *  a) a snapshot deletion in progress
3425                          *  b) a free space cache inode
3426                          * We need to distinguish those two, as the snapshot
3427                          * orphan must not get deleted.
3428                          * find_dead_roots already ran before us, so if this
3429                          * is a snapshot deletion, we should find the root
3430                          * in the dead_roots list
3431                          */
3432                         spin_lock(&fs_info->trans_lock);
3433                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3434                                             root_list) {
3435                                 if (dead_root->root_key.objectid ==
3436                                     found_key.objectid) {
3437                                         is_dead_root = 1;
3438                                         break;
3439                                 }
3440                         }
3441                         spin_unlock(&fs_info->trans_lock);
3442                         if (is_dead_root) {
3443                                 /* prevent this orphan from being found again */
3444                                 key.offset = found_key.objectid - 1;
3445                                 continue;
3446                         }
3447
3448                 }
3449
3450                 /*
3451                  * If we have an inode with links, there are a couple of
3452                  * possibilities. Old kernels (before v3.12) used to create an
3453                  * orphan item for truncate indicating that there were possibly
3454                  * extent items past i_size that needed to be deleted. In v3.12,
3455                  * truncate was changed to update i_size in sync with the extent
3456                  * items, but the (useless) orphan item was still created. Since
3457                  * v4.18, we don't create the orphan item for truncate at all.
3458                  *
3459                  * So, this item could mean that we need to do a truncate, but
3460                  * only if this filesystem was last used on a pre-v3.12 kernel
3461                  * and was not cleanly unmounted. The odds of that are quite
3462                  * slim, and it's a pain to do the truncate now, so just delete
3463                  * the orphan item.
3464                  *
3465                  * It's also possible that this orphan item was supposed to be
3466                  * deleted but wasn't. The inode number may have been reused,
3467                  * but either way, we can delete the orphan item.
3468                  */
3469                 if (ret == -ENOENT || inode->i_nlink) {
3470                         if (!ret)
3471                                 iput(inode);
3472                         trans = btrfs_start_transaction(root, 1);
3473                         if (IS_ERR(trans)) {
3474                                 ret = PTR_ERR(trans);
3475                                 goto out;
3476                         }
3477                         btrfs_debug(fs_info, "auto deleting %Lu",
3478                                     found_key.objectid);
3479                         ret = btrfs_del_orphan_item(trans, root,
3480                                                     found_key.objectid);
3481                         btrfs_end_transaction(trans);
3482                         if (ret)
3483                                 goto out;
3484                         continue;
3485                 }
3486
3487                 nr_unlink++;
3488
3489                 /* this will do delete_inode and everything for us */
3490                 iput(inode);
3491                 if (ret)
3492                         goto out;
3493         }
3494         /* release the path since we're done with it */
3495         btrfs_release_path(path);
3496
3497         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3498
3499         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3500                 trans = btrfs_join_transaction(root);
3501                 if (!IS_ERR(trans))
3502                         btrfs_end_transaction(trans);
3503         }
3504
3505         if (nr_unlink)
3506                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3507
3508 out:
3509         if (ret)
3510                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3511         btrfs_free_path(path);
3512         return ret;
3513 }
3514
3515 /*
3516  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3517  * don't find any xattrs, we know there can't be any acls.
3518  *
3519  * slot is the slot the inode is in, objectid is the objectid of the inode
3520  */
3521 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3522                                           int slot, u64 objectid,
3523                                           int *first_xattr_slot)
3524 {
3525         u32 nritems = btrfs_header_nritems(leaf);
3526         struct btrfs_key found_key;
3527         static u64 xattr_access = 0;
3528         static u64 xattr_default = 0;
3529         int scanned = 0;
3530
3531         if (!xattr_access) {
3532                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3533                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3534                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3535                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3536         }
3537
3538         slot++;
3539         *first_xattr_slot = -1;
3540         while (slot < nritems) {
3541                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3542
3543                 /* we found a different objectid, there must not be acls */
3544                 if (found_key.objectid != objectid)
3545                         return 0;
3546
3547                 /* we found an xattr, assume we've got an acl */
3548                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3549                         if (*first_xattr_slot == -1)
3550                                 *first_xattr_slot = slot;
3551                         if (found_key.offset == xattr_access ||
3552                             found_key.offset == xattr_default)
3553                                 return 1;
3554                 }
3555
3556                 /*
3557                  * we found a key greater than an xattr key, there can't
3558                  * be any acls later on
3559                  */
3560                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3561                         return 0;
3562
3563                 slot++;
3564                 scanned++;
3565
3566                 /*
3567                  * it goes inode, inode backrefs, xattrs, extents,
3568                  * so if there are a ton of hard links to an inode there can
3569                  * be a lot of backrefs.  Don't waste time searching too hard,
3570                  * this is just an optimization
3571                  */
3572                 if (scanned >= 8)
3573                         break;
3574         }
3575         /* we hit the end of the leaf before we found an xattr or
3576          * something larger than an xattr.  We have to assume the inode
3577          * has acls
3578          */
3579         if (*first_xattr_slot == -1)
3580                 *first_xattr_slot = slot;
3581         return 1;
3582 }
3583
3584 /*
3585  * read an inode from the btree into the in-memory inode
3586  */
3587 static int btrfs_read_locked_inode(struct inode *inode)
3588 {
3589         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3590         struct btrfs_path *path;
3591         struct extent_buffer *leaf;
3592         struct btrfs_inode_item *inode_item;
3593         struct btrfs_root *root = BTRFS_I(inode)->root;
3594         struct btrfs_key location;
3595         unsigned long ptr;
3596         int maybe_acls;
3597         u32 rdev;
3598         int ret;
3599         bool filled = false;
3600         int first_xattr_slot;
3601
3602         ret = btrfs_fill_inode(inode, &rdev);
3603         if (!ret)
3604                 filled = true;
3605
3606         path = btrfs_alloc_path();
3607         if (!path)
3608                 return -ENOMEM;
3609
3610         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3611
3612         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3613         if (ret) {
3614                 btrfs_free_path(path);
3615                 return ret;
3616         }
3617
3618         leaf = path->nodes[0];
3619
3620         if (filled)
3621                 goto cache_index;
3622
3623         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3624                                     struct btrfs_inode_item);
3625         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3626         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3627         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3628         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3629         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3630
3631         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3632         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3633
3634         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3635         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3636
3637         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3638         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3639
3640         BTRFS_I(inode)->i_otime.tv_sec =
3641                 btrfs_timespec_sec(leaf, &inode_item->otime);
3642         BTRFS_I(inode)->i_otime.tv_nsec =
3643                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3644
3645         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3646         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3647         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3648
3649         inode_set_iversion_queried(inode,
3650                                    btrfs_inode_sequence(leaf, inode_item));
3651         inode->i_generation = BTRFS_I(inode)->generation;
3652         inode->i_rdev = 0;
3653         rdev = btrfs_inode_rdev(leaf, inode_item);
3654
3655         BTRFS_I(inode)->index_cnt = (u64)-1;
3656         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3657
3658 cache_index:
3659         /*
3660          * If we were modified in the current generation and evicted from memory
3661          * and then re-read we need to do a full sync since we don't have any
3662          * idea about which extents were modified before we were evicted from
3663          * cache.
3664          *
3665          * This is required for both inode re-read from disk and delayed inode
3666          * in delayed_nodes_tree.
3667          */
3668         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3669                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3670                         &BTRFS_I(inode)->runtime_flags);
3671
3672         /*
3673          * We don't persist the id of the transaction where an unlink operation
3674          * against the inode was last made. So here we assume the inode might
3675          * have been evicted, and therefore the exact value of last_unlink_trans
3676          * lost, and set it to last_trans to avoid metadata inconsistencies
3677          * between the inode and its parent if the inode is fsync'ed and the log
3678          * replayed. For example, in the scenario:
3679          *
3680          * touch mydir/foo
3681          * ln mydir/foo mydir/bar
3682          * sync
3683          * unlink mydir/bar
3684          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3685          * xfs_io -c fsync mydir/foo
3686          * <power failure>
3687          * mount fs, triggers fsync log replay
3688          *
3689          * We must make sure that when we fsync our inode foo we also log its
3690          * parent inode, otherwise after log replay the parent still has the
3691          * dentry with the "bar" name but our inode foo has a link count of 1
3692          * and doesn't have an inode ref with the name "bar" anymore.
3693          *
3694          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3695          * but it guarantees correctness at the expense of occasional full
3696          * transaction commits on fsync if our inode is a directory, or if our
3697          * inode is not a directory, logging its parent unnecessarily.
3698          */
3699         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3700
3701         path->slots[0]++;
3702         if (inode->i_nlink != 1 ||
3703             path->slots[0] >= btrfs_header_nritems(leaf))
3704                 goto cache_acl;
3705
3706         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3707         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3708                 goto cache_acl;
3709
3710         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3711         if (location.type == BTRFS_INODE_REF_KEY) {
3712                 struct btrfs_inode_ref *ref;
3713
3714                 ref = (struct btrfs_inode_ref *)ptr;
3715                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3716         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3717                 struct btrfs_inode_extref *extref;
3718
3719                 extref = (struct btrfs_inode_extref *)ptr;
3720                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3721                                                                      extref);
3722         }
3723 cache_acl:
3724         /*
3725          * try to precache a NULL acl entry for files that don't have
3726          * any xattrs or acls
3727          */
3728         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3729                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3730         if (first_xattr_slot != -1) {
3731                 path->slots[0] = first_xattr_slot;
3732                 ret = btrfs_load_inode_props(inode, path);
3733                 if (ret)
3734                         btrfs_err(fs_info,
3735                                   "error loading props for ino %llu (root %llu): %d",
3736                                   btrfs_ino(BTRFS_I(inode)),
3737                                   root->root_key.objectid, ret);
3738         }
3739         btrfs_free_path(path);
3740
3741         if (!maybe_acls)
3742                 cache_no_acl(inode);
3743
3744         switch (inode->i_mode & S_IFMT) {
3745         case S_IFREG:
3746                 inode->i_mapping->a_ops = &btrfs_aops;
3747                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3748                 inode->i_fop = &btrfs_file_operations;
3749                 inode->i_op = &btrfs_file_inode_operations;
3750                 break;
3751         case S_IFDIR:
3752                 inode->i_fop = &btrfs_dir_file_operations;
3753                 inode->i_op = &btrfs_dir_inode_operations;
3754                 break;
3755         case S_IFLNK:
3756                 inode->i_op = &btrfs_symlink_inode_operations;
3757                 inode_nohighmem(inode);
3758                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3759                 break;
3760         default:
3761                 inode->i_op = &btrfs_special_inode_operations;
3762                 init_special_inode(inode, inode->i_mode, rdev);
3763                 break;
3764         }
3765
3766         btrfs_sync_inode_flags_to_i_flags(inode);
3767         return 0;
3768 }
3769
3770 /*
3771  * given a leaf and an inode, copy the inode fields into the leaf
3772  */
3773 static void fill_inode_item(struct btrfs_trans_handle *trans,
3774                             struct extent_buffer *leaf,
3775                             struct btrfs_inode_item *item,
3776                             struct inode *inode)
3777 {
3778         struct btrfs_map_token token;
3779
3780         btrfs_init_map_token(&token);
3781
3782         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3783         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3784         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3785                                    &token);
3786         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3787         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3788
3789         btrfs_set_token_timespec_sec(leaf, &item->atime,
3790                                      inode->i_atime.tv_sec, &token);
3791         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3792                                       inode->i_atime.tv_nsec, &token);
3793
3794         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3795                                      inode->i_mtime.tv_sec, &token);
3796         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3797                                       inode->i_mtime.tv_nsec, &token);
3798
3799         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3800                                      inode->i_ctime.tv_sec, &token);
3801         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3802                                       inode->i_ctime.tv_nsec, &token);
3803
3804         btrfs_set_token_timespec_sec(leaf, &item->otime,
3805                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3806         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3807                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3808
3809         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3810                                      &token);
3811         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3812                                          &token);
3813         btrfs_set_token_inode_sequence(leaf, item, inode_peek_iversion(inode),
3814                                        &token);
3815         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3816         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3817         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3818         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3819 }
3820
3821 /*
3822  * copy everything in the in-memory inode into the btree.
3823  */
3824 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3825                                 struct btrfs_root *root, struct inode *inode)
3826 {
3827         struct btrfs_inode_item *inode_item;
3828         struct btrfs_path *path;
3829         struct extent_buffer *leaf;
3830         int ret;
3831
3832         path = btrfs_alloc_path();
3833         if (!path)
3834                 return -ENOMEM;
3835
3836         path->leave_spinning = 1;
3837         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3838                                  1);
3839         if (ret) {
3840                 if (ret > 0)
3841                         ret = -ENOENT;
3842                 goto failed;
3843         }
3844
3845         leaf = path->nodes[0];
3846         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3847                                     struct btrfs_inode_item);
3848
3849         fill_inode_item(trans, leaf, inode_item, inode);
3850         btrfs_mark_buffer_dirty(leaf);
3851         btrfs_set_inode_last_trans(trans, inode);
3852         ret = 0;
3853 failed:
3854         btrfs_free_path(path);
3855         return ret;
3856 }
3857
3858 /*
3859  * copy everything in the in-memory inode into the btree.
3860  */
3861 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3862                                 struct btrfs_root *root, struct inode *inode)
3863 {
3864         struct btrfs_fs_info *fs_info = root->fs_info;
3865         int ret;
3866
3867         /*
3868          * If the inode is a free space inode, we can deadlock during commit
3869          * if we put it into the delayed code.
3870          *
3871          * The data relocation inode should also be directly updated
3872          * without delay
3873          */
3874         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
3875             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3876             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
3877                 btrfs_update_root_times(trans, root);
3878
3879                 ret = btrfs_delayed_update_inode(trans, root, inode);
3880                 if (!ret)
3881                         btrfs_set_inode_last_trans(trans, inode);
3882                 return ret;
3883         }
3884
3885         return btrfs_update_inode_item(trans, root, inode);
3886 }
3887
3888 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3889                                          struct btrfs_root *root,
3890                                          struct inode *inode)
3891 {
3892         int ret;
3893
3894         ret = btrfs_update_inode(trans, root, inode);
3895         if (ret == -ENOSPC)
3896                 return btrfs_update_inode_item(trans, root, inode);
3897         return ret;
3898 }
3899
3900 /*
3901  * unlink helper that gets used here in inode.c and in the tree logging
3902  * recovery code.  It remove a link in a directory with a given name, and
3903  * also drops the back refs in the inode to the directory
3904  */
3905 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3906                                 struct btrfs_root *root,
3907                                 struct btrfs_inode *dir,
3908                                 struct btrfs_inode *inode,
3909                                 const char *name, int name_len)
3910 {
3911         struct btrfs_fs_info *fs_info = root->fs_info;
3912         struct btrfs_path *path;
3913         int ret = 0;
3914         struct extent_buffer *leaf;
3915         struct btrfs_dir_item *di;
3916         struct btrfs_key key;
3917         u64 index;
3918         u64 ino = btrfs_ino(inode);
3919         u64 dir_ino = btrfs_ino(dir);
3920
3921         path = btrfs_alloc_path();
3922         if (!path) {
3923                 ret = -ENOMEM;
3924                 goto out;
3925         }
3926
3927         path->leave_spinning = 1;
3928         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3929                                     name, name_len, -1);
3930         if (IS_ERR(di)) {
3931                 ret = PTR_ERR(di);
3932                 goto err;
3933         }
3934         if (!di) {
3935                 ret = -ENOENT;
3936                 goto err;
3937         }
3938         leaf = path->nodes[0];
3939         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3940         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3941         if (ret)
3942                 goto err;
3943         btrfs_release_path(path);
3944
3945         /*
3946          * If we don't have dir index, we have to get it by looking up
3947          * the inode ref, since we get the inode ref, remove it directly,
3948          * it is unnecessary to do delayed deletion.
3949          *
3950          * But if we have dir index, needn't search inode ref to get it.
3951          * Since the inode ref is close to the inode item, it is better
3952          * that we delay to delete it, and just do this deletion when
3953          * we update the inode item.
3954          */
3955         if (inode->dir_index) {
3956                 ret = btrfs_delayed_delete_inode_ref(inode);
3957                 if (!ret) {
3958                         index = inode->dir_index;
3959                         goto skip_backref;
3960                 }
3961         }
3962
3963         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
3964                                   dir_ino, &index);
3965         if (ret) {
3966                 btrfs_info(fs_info,
3967                         "failed to delete reference to %.*s, inode %llu parent %llu",
3968                         name_len, name, ino, dir_ino);
3969                 btrfs_abort_transaction(trans, ret);
3970                 goto err;
3971         }
3972 skip_backref:
3973         ret = btrfs_delete_delayed_dir_index(trans, dir, index);
3974         if (ret) {
3975                 btrfs_abort_transaction(trans, ret);
3976                 goto err;
3977         }
3978
3979         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
3980                         dir_ino);
3981         if (ret != 0 && ret != -ENOENT) {
3982                 btrfs_abort_transaction(trans, ret);
3983                 goto err;
3984         }
3985
3986         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
3987                         index);
3988         if (ret == -ENOENT)
3989                 ret = 0;
3990         else if (ret)
3991                 btrfs_abort_transaction(trans, ret);
3992 err:
3993         btrfs_free_path(path);
3994         if (ret)
3995                 goto out;
3996
3997         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
3998         inode_inc_iversion(&inode->vfs_inode);
3999         inode_inc_iversion(&dir->vfs_inode);
4000         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
4001                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
4002         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
4003 out:
4004         return ret;
4005 }
4006
4007 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4008                        struct btrfs_root *root,
4009                        struct btrfs_inode *dir, struct btrfs_inode *inode,
4010                        const char *name, int name_len)
4011 {
4012         int ret;
4013         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4014         if (!ret) {
4015                 drop_nlink(&inode->vfs_inode);
4016                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
4017         }
4018         return ret;
4019 }
4020
4021 /*
4022  * helper to start transaction for unlink and rmdir.
4023  *
4024  * unlink and rmdir are special in btrfs, they do not always free space, so
4025  * if we cannot make our reservations the normal way try and see if there is
4026  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4027  * allow the unlink to occur.
4028  */
4029 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4030 {
4031         struct btrfs_root *root = BTRFS_I(dir)->root;
4032
4033         /*
4034          * 1 for the possible orphan item
4035          * 1 for the dir item
4036          * 1 for the dir index
4037          * 1 for the inode ref
4038          * 1 for the inode
4039          */
4040         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4041 }
4042
4043 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4044 {
4045         struct btrfs_root *root = BTRFS_I(dir)->root;
4046         struct btrfs_trans_handle *trans;
4047         struct inode *inode = d_inode(dentry);
4048         int ret;
4049
4050         trans = __unlink_start_trans(dir);
4051         if (IS_ERR(trans))
4052                 return PTR_ERR(trans);
4053
4054         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
4055                         0);
4056
4057         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4058                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4059                         dentry->d_name.len);
4060         if (ret)
4061                 goto out;
4062
4063         if (inode->i_nlink == 0) {
4064                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
4065                 if (ret)
4066                         goto out;
4067         }
4068
4069 out:
4070         btrfs_end_transaction(trans);
4071         btrfs_btree_balance_dirty(root->fs_info);
4072         return ret;
4073 }
4074
4075 static int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4076                                struct inode *dir, u64 objectid,
4077                                const char *name, int name_len)
4078 {
4079         struct btrfs_root *root = BTRFS_I(dir)->root;
4080         struct btrfs_path *path;
4081         struct extent_buffer *leaf;
4082         struct btrfs_dir_item *di;
4083         struct btrfs_key key;
4084         u64 index;
4085         int ret;
4086         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
4087
4088         path = btrfs_alloc_path();
4089         if (!path)
4090                 return -ENOMEM;
4091
4092         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4093                                    name, name_len, -1);
4094         if (IS_ERR_OR_NULL(di)) {
4095                 if (!di)
4096                         ret = -ENOENT;
4097                 else
4098                         ret = PTR_ERR(di);
4099                 goto out;
4100         }
4101
4102         leaf = path->nodes[0];
4103         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4104         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4105         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4106         if (ret) {
4107                 btrfs_abort_transaction(trans, ret);
4108                 goto out;
4109         }
4110         btrfs_release_path(path);
4111
4112         ret = btrfs_del_root_ref(trans, objectid, root->root_key.objectid,
4113                                  dir_ino, &index, name, name_len);
4114         if (ret < 0) {
4115                 if (ret != -ENOENT) {
4116                         btrfs_abort_transaction(trans, ret);
4117                         goto out;
4118                 }
4119                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4120                                                  name, name_len);
4121                 if (IS_ERR_OR_NULL(di)) {
4122                         if (!di)
4123                                 ret = -ENOENT;
4124                         else
4125                                 ret = PTR_ERR(di);
4126                         btrfs_abort_transaction(trans, ret);
4127                         goto out;
4128                 }
4129
4130                 leaf = path->nodes[0];
4131                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4132                 index = key.offset;
4133         }
4134         btrfs_release_path(path);
4135
4136         ret = btrfs_delete_delayed_dir_index(trans, BTRFS_I(dir), index);
4137         if (ret) {
4138                 btrfs_abort_transaction(trans, ret);
4139                 goto out;
4140         }
4141
4142         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
4143         inode_inc_iversion(dir);
4144         dir->i_mtime = dir->i_ctime = current_time(dir);
4145         ret = btrfs_update_inode_fallback(trans, root, dir);
4146         if (ret)
4147                 btrfs_abort_transaction(trans, ret);
4148 out:
4149         btrfs_free_path(path);
4150         return ret;
4151 }
4152
4153 /*
4154  * Helper to check if the subvolume references other subvolumes or if it's
4155  * default.
4156  */
4157 static noinline int may_destroy_subvol(struct btrfs_root *root)
4158 {
4159         struct btrfs_fs_info *fs_info = root->fs_info;
4160         struct btrfs_path *path;
4161         struct btrfs_dir_item *di;
4162         struct btrfs_key key;
4163         u64 dir_id;
4164         int ret;
4165
4166         path = btrfs_alloc_path();
4167         if (!path)
4168                 return -ENOMEM;
4169
4170         /* Make sure this root isn't set as the default subvol */
4171         dir_id = btrfs_super_root_dir(fs_info->super_copy);
4172         di = btrfs_lookup_dir_item(NULL, fs_info->tree_root, path,
4173                                    dir_id, "default", 7, 0);
4174         if (di && !IS_ERR(di)) {
4175                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
4176                 if (key.objectid == root->root_key.objectid) {
4177                         ret = -EPERM;
4178                         btrfs_err(fs_info,
4179                                   "deleting default subvolume %llu is not allowed",
4180                                   key.objectid);
4181                         goto out;
4182                 }
4183                 btrfs_release_path(path);
4184         }
4185
4186         key.objectid = root->root_key.objectid;
4187         key.type = BTRFS_ROOT_REF_KEY;
4188         key.offset = (u64)-1;
4189
4190         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
4191         if (ret < 0)
4192                 goto out;
4193         BUG_ON(ret == 0);
4194
4195         ret = 0;
4196         if (path->slots[0] > 0) {
4197                 path->slots[0]--;
4198                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
4199                 if (key.objectid == root->root_key.objectid &&
4200                     key.type == BTRFS_ROOT_REF_KEY)
4201                         ret = -ENOTEMPTY;
4202         }
4203 out:
4204         btrfs_free_path(path);
4205         return ret;
4206 }
4207
4208 /* Delete all dentries for inodes belonging to the root */
4209 static void btrfs_prune_dentries(struct btrfs_root *root)
4210 {
4211         struct btrfs_fs_info *fs_info = root->fs_info;
4212         struct rb_node *node;
4213         struct rb_node *prev;
4214         struct btrfs_inode *entry;
4215         struct inode *inode;
4216         u64 objectid = 0;
4217
4218         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
4219                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
4220
4221         spin_lock(&root->inode_lock);
4222 again:
4223         node = root->inode_tree.rb_node;
4224         prev = NULL;
4225         while (node) {
4226                 prev = node;
4227                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4228
4229                 if (objectid < btrfs_ino(entry))
4230                         node = node->rb_left;
4231                 else if (objectid > btrfs_ino(entry))
4232                         node = node->rb_right;
4233                 else
4234                         break;
4235         }
4236         if (!node) {
4237                 while (prev) {
4238                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
4239                         if (objectid <= btrfs_ino(entry)) {
4240                                 node = prev;
4241                                 break;
4242                         }
4243                         prev = rb_next(prev);
4244                 }
4245         }
4246         while (node) {
4247                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4248                 objectid = btrfs_ino(entry) + 1;
4249                 inode = igrab(&entry->vfs_inode);
4250                 if (inode) {
4251                         spin_unlock(&root->inode_lock);
4252                         if (atomic_read(&inode->i_count) > 1)
4253                                 d_prune_aliases(inode);
4254                         /*
4255                          * btrfs_drop_inode will have it removed from the inode
4256                          * cache when its usage count hits zero.
4257                          */
4258                         iput(inode);
4259                         cond_resched();
4260                         spin_lock(&root->inode_lock);
4261                         goto again;
4262                 }
4263
4264                 if (cond_resched_lock(&root->inode_lock))
4265                         goto again;
4266
4267                 node = rb_next(node);
4268         }
4269         spin_unlock(&root->inode_lock);
4270 }
4271
4272 int btrfs_delete_subvolume(struct inode *dir, struct dentry *dentry)
4273 {
4274         struct btrfs_fs_info *fs_info = btrfs_sb(dentry->d_sb);
4275         struct btrfs_root *root = BTRFS_I(dir)->root;
4276         struct inode *inode = d_inode(dentry);
4277         struct btrfs_root *dest = BTRFS_I(inode)->root;
4278         struct btrfs_trans_handle *trans;
4279         struct btrfs_block_rsv block_rsv;
4280         u64 root_flags;
4281         int ret;
4282         int err;
4283
4284         /*
4285          * Don't allow to delete a subvolume with send in progress. This is
4286          * inside the inode lock so the error handling that has to drop the bit
4287          * again is not run concurrently.
4288          */
4289         spin_lock(&dest->root_item_lock);
4290         root_flags = btrfs_root_flags(&dest->root_item);
4291         if (dest->send_in_progress == 0) {
4292                 btrfs_set_root_flags(&dest->root_item,
4293                                 root_flags | BTRFS_ROOT_SUBVOL_DEAD);
4294                 spin_unlock(&dest->root_item_lock);
4295         } else {
4296                 spin_unlock(&dest->root_item_lock);
4297                 btrfs_warn(fs_info,
4298                            "attempt to delete subvolume %llu during send",
4299                            dest->root_key.objectid);
4300                 return -EPERM;
4301         }
4302
4303         down_write(&fs_info->subvol_sem);
4304
4305         err = may_destroy_subvol(dest);
4306         if (err)
4307                 goto out_up_write;
4308
4309         btrfs_init_block_rsv(&block_rsv, BTRFS_BLOCK_RSV_TEMP);
4310         /*
4311          * One for dir inode,
4312          * two for dir entries,
4313          * two for root ref/backref.
4314          */
4315         err = btrfs_subvolume_reserve_metadata(root, &block_rsv, 5, true);
4316         if (err)
4317                 goto out_up_write;
4318
4319         trans = btrfs_start_transaction(root, 0);
4320         if (IS_ERR(trans)) {
4321                 err = PTR_ERR(trans);
4322                 goto out_release;
4323         }
4324         trans->block_rsv = &block_rsv;
4325         trans->bytes_reserved = block_rsv.size;
4326
4327         btrfs_record_snapshot_destroy(trans, BTRFS_I(dir));
4328
4329         ret = btrfs_unlink_subvol(trans, dir, dest->root_key.objectid,
4330                                   dentry->d_name.name, dentry->d_name.len);
4331         if (ret) {
4332                 err = ret;
4333                 btrfs_abort_transaction(trans, ret);
4334                 goto out_end_trans;
4335         }
4336
4337         btrfs_record_root_in_trans(trans, dest);
4338
4339         memset(&dest->root_item.drop_progress, 0,
4340                 sizeof(dest->root_item.drop_progress));
4341         dest->root_item.drop_level = 0;
4342         btrfs_set_root_refs(&dest->root_item, 0);
4343
4344         if (!test_and_set_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &dest->state)) {
4345                 ret = btrfs_insert_orphan_item(trans,
4346                                         fs_info->tree_root,
4347                                         dest->root_key.objectid);
4348                 if (ret) {
4349                         btrfs_abort_transaction(trans, ret);
4350                         err = ret;
4351                         goto out_end_trans;
4352                 }
4353         }
4354
4355         ret = btrfs_uuid_tree_remove(trans, dest->root_item.uuid,
4356                                   BTRFS_UUID_KEY_SUBVOL,
4357                                   dest->root_key.objectid);
4358         if (ret && ret != -ENOENT) {
4359                 btrfs_abort_transaction(trans, ret);
4360                 err = ret;
4361                 goto out_end_trans;
4362         }
4363         if (!btrfs_is_empty_uuid(dest->root_item.received_uuid)) {
4364                 ret = btrfs_uuid_tree_remove(trans,
4365                                           dest->root_item.received_uuid,
4366                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
4367                                           dest->root_key.objectid);
4368                 if (ret && ret != -ENOENT) {
4369                         btrfs_abort_transaction(trans, ret);
4370                         err = ret;
4371                         goto out_end_trans;
4372                 }
4373         }
4374
4375 out_end_trans:
4376         trans->block_rsv = NULL;
4377         trans->bytes_reserved = 0;
4378         ret = btrfs_end_transaction(trans);
4379         if (ret && !err)
4380                 err = ret;
4381         inode->i_flags |= S_DEAD;
4382 out_release:
4383         btrfs_subvolume_release_metadata(fs_info, &block_rsv);
4384 out_up_write:
4385         up_write(&fs_info->subvol_sem);
4386         if (err) {
4387                 spin_lock(&dest->root_item_lock);
4388                 root_flags = btrfs_root_flags(&dest->root_item);
4389                 btrfs_set_root_flags(&dest->root_item,
4390                                 root_flags & ~BTRFS_ROOT_SUBVOL_DEAD);
4391                 spin_unlock(&dest->root_item_lock);
4392         } else {
4393                 d_invalidate(dentry);
4394                 btrfs_prune_dentries(dest);
4395                 ASSERT(dest->send_in_progress == 0);
4396
4397                 /* the last ref */
4398                 if (dest->ino_cache_inode) {
4399                         iput(dest->ino_cache_inode);
4400                         dest->ino_cache_inode = NULL;
4401                 }
4402         }
4403
4404         return err;
4405 }
4406
4407 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4408 {
4409         struct inode *inode = d_inode(dentry);
4410         int err = 0;
4411         struct btrfs_root *root = BTRFS_I(dir)->root;
4412         struct btrfs_trans_handle *trans;
4413         u64 last_unlink_trans;
4414
4415         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4416                 return -ENOTEMPTY;
4417         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4418                 return btrfs_delete_subvolume(dir, dentry);
4419
4420         trans = __unlink_start_trans(dir);
4421         if (IS_ERR(trans))
4422                 return PTR_ERR(trans);
4423
4424         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4425                 err = btrfs_unlink_subvol(trans, dir,
4426                                           BTRFS_I(inode)->location.objectid,
4427                                           dentry->d_name.name,
4428                                           dentry->d_name.len);
4429                 goto out;
4430         }
4431
4432         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4433         if (err)
4434                 goto out;
4435
4436         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4437
4438         /* now the directory is empty */
4439         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4440                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4441                         dentry->d_name.len);
4442         if (!err) {
4443                 btrfs_i_size_write(BTRFS_I(inode), 0);
4444                 /*
4445                  * Propagate the last_unlink_trans value of the deleted dir to
4446                  * its parent directory. This is to prevent an unrecoverable
4447                  * log tree in the case we do something like this:
4448                  * 1) create dir foo
4449                  * 2) create snapshot under dir foo
4450                  * 3) delete the snapshot
4451                  * 4) rmdir foo
4452                  * 5) mkdir foo
4453                  * 6) fsync foo or some file inside foo
4454                  */
4455                 if (last_unlink_trans >= trans->transid)
4456                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4457         }
4458 out:
4459         btrfs_end_transaction(trans);
4460         btrfs_btree_balance_dirty(root->fs_info);
4461
4462         return err;
4463 }
4464
4465 static int truncate_space_check(struct btrfs_trans_handle *trans,
4466                                 struct btrfs_root *root,
4467                                 u64 bytes_deleted)
4468 {
4469         struct btrfs_fs_info *fs_info = root->fs_info;
4470         int ret;
4471
4472         /*
4473          * This is only used to apply pressure to the enospc system, we don't
4474          * intend to use this reservation at all.
4475          */
4476         bytes_deleted = btrfs_csum_bytes_to_leaves(fs_info, bytes_deleted);
4477         bytes_deleted *= fs_info->nodesize;
4478         ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
4479                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4480         if (!ret) {
4481                 trace_btrfs_space_reservation(fs_info, "transaction",
4482                                               trans->transid,
4483                                               bytes_deleted, 1);
4484                 trans->bytes_reserved += bytes_deleted;
4485         }
4486         return ret;
4487
4488 }
4489
4490 /*
4491  * Return this if we need to call truncate_block for the last bit of the
4492  * truncate.
4493  */
4494 #define NEED_TRUNCATE_BLOCK 1
4495
4496 /*
4497  * this can truncate away extent items, csum items and directory items.
4498  * It starts at a high offset and removes keys until it can't find
4499  * any higher than new_size
4500  *
4501  * csum items that cross the new i_size are truncated to the new size
4502  * as well.
4503  *
4504  * min_type is the minimum key type to truncate down to.  If set to 0, this
4505  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4506  */
4507 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4508                                struct btrfs_root *root,
4509                                struct inode *inode,
4510                                u64 new_size, u32 min_type)
4511 {
4512         struct btrfs_fs_info *fs_info = root->fs_info;
4513         struct btrfs_path *path;
4514         struct extent_buffer *leaf;
4515         struct btrfs_file_extent_item *fi;
4516         struct btrfs_key key;
4517         struct btrfs_key found_key;
4518         u64 extent_start = 0;
4519         u64 extent_num_bytes = 0;
4520         u64 extent_offset = 0;
4521         u64 item_end = 0;
4522         u64 last_size = new_size;
4523         u32 found_type = (u8)-1;
4524         int found_extent;
4525         int del_item;
4526         int pending_del_nr = 0;
4527         int pending_del_slot = 0;
4528         int extent_type = -1;
4529         int ret;
4530         u64 ino = btrfs_ino(BTRFS_I(inode));
4531         u64 bytes_deleted = 0;
4532         bool be_nice = false;
4533         bool should_throttle = false;
4534         bool should_end = false;
4535
4536         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4537
4538         /*
4539          * for non-free space inodes and ref cows, we want to back off from
4540          * time to time
4541          */
4542         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4543             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4544                 be_nice = true;
4545
4546         path = btrfs_alloc_path();
4547         if (!path)
4548                 return -ENOMEM;
4549         path->reada = READA_BACK;
4550
4551         /*
4552          * We want to drop from the next block forward in case this new size is
4553          * not block aligned since we will be keeping the last block of the
4554          * extent just the way it is.
4555          */
4556         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4557             root == fs_info->tree_root)
4558                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4559                                         fs_info->sectorsize),
4560                                         (u64)-1, 0);
4561
4562         /*
4563          * This function is also used to drop the items in the log tree before
4564          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4565          * it is used to drop the loged items. So we shouldn't kill the delayed
4566          * items.
4567          */
4568         if (min_type == 0 && root == BTRFS_I(inode)->root)
4569                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4570
4571         key.objectid = ino;
4572         key.offset = (u64)-1;
4573         key.type = (u8)-1;
4574
4575 search_again:
4576         /*
4577          * with a 16K leaf size and 128MB extents, you can actually queue
4578          * up a huge file in a single leaf.  Most of the time that
4579          * bytes_deleted is > 0, it will be huge by the time we get here
4580          */
4581         if (be_nice && bytes_deleted > SZ_32M &&
4582             btrfs_should_end_transaction(trans)) {
4583                 ret = -EAGAIN;
4584                 goto out;
4585         }
4586
4587         path->leave_spinning = 1;
4588         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4589         if (ret < 0)
4590                 goto out;
4591
4592         if (ret > 0) {
4593                 ret = 0;
4594                 /* there are no items in the tree for us to truncate, we're
4595                  * done
4596                  */
4597                 if (path->slots[0] == 0)
4598                         goto out;
4599                 path->slots[0]--;
4600         }
4601
4602         while (1) {
4603                 fi = NULL;
4604                 leaf = path->nodes[0];
4605                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4606                 found_type = found_key.type;
4607
4608                 if (found_key.objectid != ino)
4609                         break;
4610
4611                 if (found_type < min_type)
4612                         break;
4613
4614                 item_end = found_key.offset;
4615                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4616                         fi = btrfs_item_ptr(leaf, path->slots[0],
4617                                             struct btrfs_file_extent_item);
4618                         extent_type = btrfs_file_extent_type(leaf, fi);
4619                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4620                                 item_end +=
4621                                     btrfs_file_extent_num_bytes(leaf, fi);
4622
4623                                 trace_btrfs_truncate_show_fi_regular(
4624                                         BTRFS_I(inode), leaf, fi,
4625                                         found_key.offset);
4626                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4627                                 item_end += btrfs_file_extent_ram_bytes(leaf,
4628                                                                         fi);
4629
4630                                 trace_btrfs_truncate_show_fi_inline(
4631                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4632                                         found_key.offset);
4633                         }
4634                         item_end--;
4635                 }
4636                 if (found_type > min_type) {
4637                         del_item = 1;
4638                 } else {
4639                         if (item_end < new_size)
4640                                 break;
4641                         if (found_key.offset >= new_size)
4642                                 del_item = 1;
4643                         else
4644                                 del_item = 0;
4645                 }
4646                 found_extent = 0;
4647                 /* FIXME, shrink the extent if the ref count is only 1 */
4648                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4649                         goto delete;
4650
4651                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4652                         u64 num_dec;
4653                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4654                         if (!del_item) {
4655                                 u64 orig_num_bytes =
4656                                         btrfs_file_extent_num_bytes(leaf, fi);
4657                                 extent_num_bytes = ALIGN(new_size -
4658                                                 found_key.offset,
4659                                                 fs_info->sectorsize);
4660                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4661                                                          extent_num_bytes);
4662                                 num_dec = (orig_num_bytes -
4663                                            extent_num_bytes);
4664                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4665                                              &root->state) &&
4666                                     extent_start != 0)
4667                                         inode_sub_bytes(inode, num_dec);
4668                                 btrfs_mark_buffer_dirty(leaf);
4669                         } else {
4670                                 extent_num_bytes =
4671                                         btrfs_file_extent_disk_num_bytes(leaf,
4672                                                                          fi);
4673                                 extent_offset = found_key.offset -
4674                                         btrfs_file_extent_offset(leaf, fi);
4675
4676                                 /* FIXME blocksize != 4096 */
4677                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4678                                 if (extent_start != 0) {
4679                                         found_extent = 1;
4680                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4681                                                      &root->state))
4682                                                 inode_sub_bytes(inode, num_dec);
4683                                 }
4684                         }
4685                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4686                         /*
4687                          * we can't truncate inline items that have had
4688                          * special encodings
4689                          */
4690                         if (!del_item &&
4691                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4692                             btrfs_file_extent_other_encoding(leaf, fi) == 0 &&
4693                             btrfs_file_extent_compression(leaf, fi) == 0) {
4694                                 u32 size = (u32)(new_size - found_key.offset);
4695
4696                                 btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4697                                 size = btrfs_file_extent_calc_inline_size(size);
4698                                 btrfs_truncate_item(root->fs_info, path, size, 1);
4699                         } else if (!del_item) {
4700                                 /*
4701                                  * We have to bail so the last_size is set to
4702                                  * just before this extent.
4703                                  */
4704                                 ret = NEED_TRUNCATE_BLOCK;
4705                                 break;
4706                         }
4707
4708                         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4709                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4710                 }
4711 delete:
4712                 if (del_item)
4713                         last_size = found_key.offset;
4714                 else
4715                         last_size = new_size;
4716                 if (del_item) {
4717                         if (!pending_del_nr) {
4718                                 /* no pending yet, add ourselves */
4719                                 pending_del_slot = path->slots[0];
4720                                 pending_del_nr = 1;
4721                         } else if (pending_del_nr &&
4722                                    path->slots[0] + 1 == pending_del_slot) {
4723                                 /* hop on the pending chunk */
4724                                 pending_del_nr++;
4725                                 pending_del_slot = path->slots[0];
4726                         } else {
4727                                 BUG();
4728                         }
4729                 } else {
4730                         break;
4731                 }
4732                 should_throttle = false;
4733
4734                 if (found_extent &&
4735                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4736                      root == fs_info->tree_root)) {
4737                         btrfs_set_path_blocking(path);
4738                         bytes_deleted += extent_num_bytes;
4739                         ret = btrfs_free_extent(trans, root, extent_start,
4740                                                 extent_num_bytes, 0,
4741                                                 btrfs_header_owner(leaf),
4742                                                 ino, extent_offset);
4743                         if (ret) {
4744                                 btrfs_abort_transaction(trans, ret);
4745                                 break;
4746                         }
4747                         if (btrfs_should_throttle_delayed_refs(trans, fs_info))
4748                                 btrfs_async_run_delayed_refs(fs_info,
4749                                         trans->delayed_ref_updates * 2,
4750                                         trans->transid, 0);
4751                         if (be_nice) {
4752                                 if (truncate_space_check(trans, root,
4753                                                          extent_num_bytes)) {
4754                                         should_end = true;
4755                                 }
4756                                 if (btrfs_should_throttle_delayed_refs(trans,
4757                                                                        fs_info))
4758                                         should_throttle = true;
4759                         }
4760                 }
4761
4762                 if (found_type == BTRFS_INODE_ITEM_KEY)
4763                         break;
4764
4765                 if (path->slots[0] == 0 ||
4766                     path->slots[0] != pending_del_slot ||
4767                     should_throttle || should_end) {
4768                         if (pending_del_nr) {
4769                                 ret = btrfs_del_items(trans, root, path,
4770                                                 pending_del_slot,
4771                                                 pending_del_nr);
4772                                 if (ret) {
4773                                         btrfs_abort_transaction(trans, ret);
4774                                         break;
4775                                 }
4776                                 pending_del_nr = 0;
4777                         }
4778                         btrfs_release_path(path);
4779                         if (should_throttle) {
4780                                 unsigned long updates = trans->delayed_ref_updates;
4781                                 if (updates) {
4782                                         trans->delayed_ref_updates = 0;
4783                                         ret = btrfs_run_delayed_refs(trans,
4784                                                                    updates * 2);
4785                                         if (ret)
4786                                                 break;
4787                                 }
4788                         }
4789                         /*
4790                          * if we failed to refill our space rsv, bail out
4791                          * and let the transaction restart
4792                          */
4793                         if (should_end) {
4794                                 ret = -EAGAIN;
4795                                 break;
4796                         }
4797                         goto search_again;
4798                 } else {
4799                         path->slots[0]--;
4800                 }
4801         }
4802 out:
4803         if (ret >= 0 && pending_del_nr) {
4804                 int err;
4805
4806                 err = btrfs_del_items(trans, root, path, pending_del_slot,
4807                                       pending_del_nr);
4808                 if (err) {
4809                         btrfs_abort_transaction(trans, err);
4810                         ret = err;
4811                 }
4812         }
4813         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4814                 ASSERT(last_size >= new_size);
4815                 if (!ret && last_size > new_size)
4816                         last_size = new_size;
4817                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4818         }
4819
4820         btrfs_free_path(path);
4821
4822         if (be_nice && bytes_deleted > SZ_32M && (ret >= 0 || ret == -EAGAIN)) {
4823                 unsigned long updates = trans->delayed_ref_updates;
4824                 int err;
4825
4826                 if (updates) {
4827                         trans->delayed_ref_updates = 0;
4828                         err = btrfs_run_delayed_refs(trans, updates * 2);
4829                         if (err)
4830                                 ret = err;
4831                 }
4832         }
4833         return ret;
4834 }
4835
4836 /*
4837  * btrfs_truncate_block - read, zero a chunk and write a block
4838  * @inode - inode that we're zeroing
4839  * @from - the offset to start zeroing
4840  * @len - the length to zero, 0 to zero the entire range respective to the
4841  *      offset
4842  * @front - zero up to the offset instead of from the offset on
4843  *
4844  * This will find the block for the "from" offset and cow the block and zero the
4845  * part we want to zero.  This is used with truncate and hole punching.
4846  */
4847 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4848                         int front)
4849 {
4850         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4851         struct address_space *mapping = inode->i_mapping;
4852         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4853         struct btrfs_ordered_extent *ordered;
4854         struct extent_state *cached_state = NULL;
4855         struct extent_changeset *data_reserved = NULL;
4856         char *kaddr;
4857         u32 blocksize = fs_info->sectorsize;
4858         pgoff_t index = from >> PAGE_SHIFT;
4859         unsigned offset = from & (blocksize - 1);
4860         struct page *page;
4861         gfp_t mask = btrfs_alloc_write_mask(mapping);
4862         int ret = 0;
4863         u64 block_start;
4864         u64 block_end;
4865
4866         if (IS_ALIGNED(offset, blocksize) &&
4867             (!len || IS_ALIGNED(len, blocksize)))
4868                 goto out;
4869
4870         block_start = round_down(from, blocksize);
4871         block_end = block_start + blocksize - 1;
4872
4873         ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
4874                                            block_start, blocksize);
4875         if (ret)
4876                 goto out;
4877
4878 again:
4879         page = find_or_create_page(mapping, index, mask);
4880         if (!page) {
4881                 btrfs_delalloc_release_space(inode, data_reserved,
4882                                              block_start, blocksize, true);
4883                 btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize, true);
4884                 ret = -ENOMEM;
4885                 goto out;
4886         }
4887
4888         if (!PageUptodate(page)) {
4889                 ret = btrfs_readpage(NULL, page);
4890                 lock_page(page);
4891                 if (page->mapping != mapping) {
4892                         unlock_page(page);
4893                         put_page(page);
4894                         goto again;
4895                 }
4896                 if (!PageUptodate(page)) {
4897                         ret = -EIO;
4898                         goto out_unlock;
4899                 }
4900         }
4901         wait_on_page_writeback(page);
4902
4903         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4904         set_page_extent_mapped(page);
4905
4906         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4907         if (ordered) {
4908                 unlock_extent_cached(io_tree, block_start, block_end,
4909                                      &cached_state);
4910                 unlock_page(page);
4911                 put_page(page);
4912                 btrfs_start_ordered_extent(inode, ordered, 1);
4913                 btrfs_put_ordered_extent(ordered);
4914                 goto again;
4915         }
4916
4917         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4918                           EXTENT_DIRTY | EXTENT_DELALLOC |
4919                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4920                           0, 0, &cached_state);
4921
4922         ret = btrfs_set_extent_delalloc(inode, block_start, block_end, 0,
4923                                         &cached_state, 0);
4924         if (ret) {
4925                 unlock_extent_cached(io_tree, block_start, block_end,
4926                                      &cached_state);
4927                 goto out_unlock;
4928         }
4929
4930         if (offset != blocksize) {
4931                 if (!len)
4932                         len = blocksize - offset;
4933                 kaddr = kmap(page);
4934                 if (front)
4935                         memset(kaddr + (block_start - page_offset(page)),
4936                                 0, offset);
4937                 else
4938                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4939                                 0, len);
4940                 flush_dcache_page(page);
4941                 kunmap(page);
4942         }
4943         ClearPageChecked(page);
4944         set_page_dirty(page);
4945         unlock_extent_cached(io_tree, block_start, block_end, &cached_state);
4946
4947 out_unlock:
4948         if (ret)
4949                 btrfs_delalloc_release_space(inode, data_reserved, block_start,
4950                                              blocksize, true);
4951         btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize, (ret != 0));
4952         unlock_page(page);
4953         put_page(page);
4954 out:
4955         extent_changeset_free(data_reserved);
4956         return ret;
4957 }
4958
4959 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4960                              u64 offset, u64 len)
4961 {
4962         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4963         struct btrfs_trans_handle *trans;
4964         int ret;
4965
4966         /*
4967          * Still need to make sure the inode looks like it's been updated so
4968          * that any holes get logged if we fsync.
4969          */
4970         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
4971                 BTRFS_I(inode)->last_trans = fs_info->generation;
4972                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4973                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4974                 return 0;
4975         }
4976
4977         /*
4978          * 1 - for the one we're dropping
4979          * 1 - for the one we're adding
4980          * 1 - for updating the inode.
4981          */
4982         trans = btrfs_start_transaction(root, 3);
4983         if (IS_ERR(trans))
4984                 return PTR_ERR(trans);
4985
4986         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4987         if (ret) {
4988                 btrfs_abort_transaction(trans, ret);
4989                 btrfs_end_transaction(trans);
4990                 return ret;
4991         }
4992
4993         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
4994                         offset, 0, 0, len, 0, len, 0, 0, 0);
4995         if (ret)
4996                 btrfs_abort_transaction(trans, ret);
4997         else
4998                 btrfs_update_inode(trans, root, inode);
4999         btrfs_end_transaction(trans);
5000         return ret;
5001 }
5002
5003 /*
5004  * This function puts in dummy file extents for the area we're creating a hole
5005  * for.  So if we are truncating this file to a larger size we need to insert
5006  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
5007  * the range between oldsize and size
5008  */
5009 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
5010 {
5011         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5012         struct btrfs_root *root = BTRFS_I(inode)->root;
5013         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5014         struct extent_map *em = NULL;
5015         struct extent_state *cached_state = NULL;
5016         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
5017         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
5018         u64 block_end = ALIGN(size, fs_info->sectorsize);
5019         u64 last_byte;
5020         u64 cur_offset;
5021         u64 hole_size;
5022         int err = 0;
5023
5024         /*
5025          * If our size started in the middle of a block we need to zero out the
5026          * rest of the block before we expand the i_size, otherwise we could
5027          * expose stale data.
5028          */
5029         err = btrfs_truncate_block(inode, oldsize, 0, 0);
5030         if (err)
5031                 return err;
5032
5033         if (size <= hole_start)
5034                 return 0;
5035
5036         while (1) {
5037                 struct btrfs_ordered_extent *ordered;
5038
5039                 lock_extent_bits(io_tree, hole_start, block_end - 1,
5040                                  &cached_state);
5041                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), hole_start,
5042                                                      block_end - hole_start);
5043                 if (!ordered)
5044                         break;
5045                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
5046                                      &cached_state);
5047                 btrfs_start_ordered_extent(inode, ordered, 1);
5048                 btrfs_put_ordered_extent(ordered);
5049         }
5050
5051         cur_offset = hole_start;
5052         while (1) {
5053                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
5054                                 block_end - cur_offset, 0);
5055                 if (IS_ERR(em)) {
5056                         err = PTR_ERR(em);
5057                         em = NULL;
5058                         break;
5059                 }
5060                 last_byte = min(extent_map_end(em), block_end);
5061                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
5062                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
5063                         struct extent_map *hole_em;
5064                         hole_size = last_byte - cur_offset;
5065
5066                         err = maybe_insert_hole(root, inode, cur_offset,
5067                                                 hole_size);
5068                         if (err)
5069                                 break;
5070                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
5071                                                 cur_offset + hole_size - 1, 0);
5072                         hole_em = alloc_extent_map();
5073                         if (!hole_em) {
5074                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
5075                                         &BTRFS_I(inode)->runtime_flags);
5076                                 goto next;
5077                         }
5078                         hole_em->start = cur_offset;
5079                         hole_em->len = hole_size;
5080                         hole_em->orig_start = cur_offset;
5081
5082                         hole_em->block_start = EXTENT_MAP_HOLE;
5083                         hole_em->block_len = 0;
5084                         hole_em->orig_block_len = 0;
5085                         hole_em->ram_bytes = hole_size;
5086                         hole_em->bdev = fs_info->fs_devices->latest_bdev;
5087                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
5088                         hole_em->generation = fs_info->generation;
5089
5090                         while (1) {
5091                                 write_lock(&em_tree->lock);
5092                                 err = add_extent_mapping(em_tree, hole_em, 1);
5093                                 write_unlock(&em_tree->lock);
5094                                 if (err != -EEXIST)
5095                                         break;
5096                                 btrfs_drop_extent_cache(BTRFS_I(inode),
5097                                                         cur_offset,
5098                                                         cur_offset +
5099                                                         hole_size - 1, 0);
5100                         }
5101                         free_extent_map(hole_em);
5102                 }
5103 next:
5104                 free_extent_map(em);
5105                 em = NULL;
5106                 cur_offset = last_byte;
5107                 if (cur_offset >= block_end)
5108                         break;
5109         }
5110         free_extent_map(em);
5111         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state);
5112         return err;
5113 }
5114
5115 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
5116 {
5117         struct btrfs_root *root = BTRFS_I(inode)->root;
5118         struct btrfs_trans_handle *trans;
5119         loff_t oldsize = i_size_read(inode);
5120         loff_t newsize = attr->ia_size;
5121         int mask = attr->ia_valid;
5122         int ret;
5123
5124         /*
5125          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
5126          * special case where we need to update the times despite not having
5127          * these flags set.  For all other operations the VFS set these flags
5128          * explicitly if it wants a timestamp update.
5129          */
5130         if (newsize != oldsize) {
5131                 inode_inc_iversion(inode);
5132                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
5133                         inode->i_ctime = inode->i_mtime =
5134                                 current_time(inode);
5135         }
5136
5137         if (newsize > oldsize) {
5138                 /*
5139                  * Don't do an expanding truncate while snapshotting is ongoing.
5140                  * This is to ensure the snapshot captures a fully consistent
5141                  * state of this file - if the snapshot captures this expanding
5142                  * truncation, it must capture all writes that happened before
5143                  * this truncation.
5144                  */
5145                 btrfs_wait_for_snapshot_creation(root);
5146                 ret = btrfs_cont_expand(inode, oldsize, newsize);
5147                 if (ret) {
5148                         btrfs_end_write_no_snapshotting(root);
5149                         return ret;
5150                 }
5151
5152                 trans = btrfs_start_transaction(root, 1);
5153                 if (IS_ERR(trans)) {
5154                         btrfs_end_write_no_snapshotting(root);
5155                         return PTR_ERR(trans);
5156                 }
5157
5158                 i_size_write(inode, newsize);
5159                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
5160                 pagecache_isize_extended(inode, oldsize, newsize);
5161                 ret = btrfs_update_inode(trans, root, inode);
5162                 btrfs_end_write_no_snapshotting(root);
5163                 btrfs_end_transaction(trans);
5164         } else {
5165
5166                 /*
5167                  * We're truncating a file that used to have good data down to
5168                  * zero. Make sure it gets into the ordered flush list so that
5169                  * any new writes get down to disk quickly.
5170                  */
5171                 if (newsize == 0)
5172                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5173                                 &BTRFS_I(inode)->runtime_flags);
5174
5175                 truncate_setsize(inode, newsize);
5176
5177                 /* Disable nonlocked read DIO to avoid the end less truncate */
5178                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
5179                 inode_dio_wait(inode);
5180                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
5181
5182                 ret = btrfs_truncate(inode, newsize == oldsize);
5183                 if (ret && inode->i_nlink) {
5184                         int err;
5185
5186                         /*
5187                          * Truncate failed, so fix up the in-memory size. We
5188                          * adjusted disk_i_size down as we removed extents, so
5189                          * wait for disk_i_size to be stable and then update the
5190                          * in-memory size to match.
5191                          */
5192                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
5193                         if (err)
5194                                 return err;
5195                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5196                 }
5197         }
5198
5199         return ret;
5200 }
5201
5202 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5203 {
5204         struct inode *inode = d_inode(dentry);
5205         struct btrfs_root *root = BTRFS_I(inode)->root;
5206         int err;
5207
5208         if (btrfs_root_readonly(root))
5209                 return -EROFS;
5210
5211         err = setattr_prepare(dentry, attr);
5212         if (err)
5213                 return err;
5214
5215         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5216                 err = btrfs_setsize(inode, attr);
5217                 if (err)
5218                         return err;
5219         }
5220
5221         if (attr->ia_valid) {
5222                 setattr_copy(inode, attr);
5223                 inode_inc_iversion(inode);
5224                 err = btrfs_dirty_inode(inode);
5225
5226                 if (!err && attr->ia_valid & ATTR_MODE)
5227                         err = posix_acl_chmod(inode, inode->i_mode);
5228         }
5229
5230         return err;
5231 }
5232
5233 /*
5234  * While truncating the inode pages during eviction, we get the VFS calling
5235  * btrfs_invalidatepage() against each page of the inode. This is slow because
5236  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5237  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5238  * extent_state structures over and over, wasting lots of time.
5239  *
5240  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5241  * those expensive operations on a per page basis and do only the ordered io
5242  * finishing, while we release here the extent_map and extent_state structures,
5243  * without the excessive merging and splitting.
5244  */
5245 static void evict_inode_truncate_pages(struct inode *inode)
5246 {
5247         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5248         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5249         struct rb_node *node;
5250
5251         ASSERT(inode->i_state & I_FREEING);
5252         truncate_inode_pages_final(&inode->i_data);
5253
5254         write_lock(&map_tree->lock);
5255         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5256                 struct extent_map *em;
5257
5258                 node = rb_first(&map_tree->map);
5259                 em = rb_entry(node, struct extent_map, rb_node);
5260                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5261                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5262                 remove_extent_mapping(map_tree, em);
5263                 free_extent_map(em);
5264                 if (need_resched()) {
5265                         write_unlock(&map_tree->lock);
5266                         cond_resched();
5267                         write_lock(&map_tree->lock);
5268                 }
5269         }
5270         write_unlock(&map_tree->lock);
5271
5272         /*
5273          * Keep looping until we have no more ranges in the io tree.
5274          * We can have ongoing bios started by readpages (called from readahead)
5275          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5276          * still in progress (unlocked the pages in the bio but did not yet
5277          * unlocked the ranges in the io tree). Therefore this means some
5278          * ranges can still be locked and eviction started because before
5279          * submitting those bios, which are executed by a separate task (work
5280          * queue kthread), inode references (inode->i_count) were not taken
5281          * (which would be dropped in the end io callback of each bio).
5282          * Therefore here we effectively end up waiting for those bios and
5283          * anyone else holding locked ranges without having bumped the inode's
5284          * reference count - if we don't do it, when they access the inode's
5285          * io_tree to unlock a range it may be too late, leading to an
5286          * use-after-free issue.
5287          */
5288         spin_lock(&io_tree->lock);
5289         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5290                 struct extent_state *state;
5291                 struct extent_state *cached_state = NULL;
5292                 u64 start;
5293                 u64 end;
5294
5295                 node = rb_first(&io_tree->state);
5296                 state = rb_entry(node, struct extent_state, rb_node);
5297                 start = state->start;
5298                 end = state->end;
5299                 spin_unlock(&io_tree->lock);
5300
5301                 lock_extent_bits(io_tree, start, end, &cached_state);
5302
5303                 /*
5304                  * If still has DELALLOC flag, the extent didn't reach disk,
5305                  * and its reserved space won't be freed by delayed_ref.
5306                  * So we need to free its reserved space here.
5307                  * (Refer to comment in btrfs_invalidatepage, case 2)
5308                  *
5309                  * Note, end is the bytenr of last byte, so we need + 1 here.
5310                  */
5311                 if (state->state & EXTENT_DELALLOC)
5312                         btrfs_qgroup_free_data(inode, NULL, start, end - start + 1);
5313
5314                 clear_extent_bit(io_tree, start, end,
5315                                  EXTENT_LOCKED | EXTENT_DIRTY |
5316                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5317                                  EXTENT_DEFRAG, 1, 1, &cached_state);
5318
5319                 cond_resched();
5320                 spin_lock(&io_tree->lock);
5321         }
5322         spin_unlock(&io_tree->lock);
5323 }
5324
5325 static struct btrfs_trans_handle *evict_refill_and_join(struct btrfs_root *root,
5326                                                         struct btrfs_block_rsv *rsv,
5327                                                         u64 min_size)
5328 {
5329         struct btrfs_fs_info *fs_info = root->fs_info;
5330         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5331         int failures = 0;
5332
5333         for (;;) {
5334                 struct btrfs_trans_handle *trans;
5335                 int ret;
5336
5337                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5338                                              BTRFS_RESERVE_FLUSH_LIMIT);
5339
5340                 if (ret && ++failures > 2) {
5341                         btrfs_warn(fs_info,
5342                                    "could not allocate space for a delete; will truncate on mount");
5343                         return ERR_PTR(-ENOSPC);
5344                 }
5345
5346                 trans = btrfs_join_transaction(root);
5347                 if (IS_ERR(trans) || !ret)
5348                         return trans;
5349
5350                 /*
5351                  * Try to steal from the global reserve if there is space for
5352                  * it.
5353                  */
5354                 if (!btrfs_check_space_for_delayed_refs(trans, fs_info) &&
5355                     !btrfs_block_rsv_migrate(global_rsv, rsv, min_size, 0))
5356                         return trans;
5357
5358                 /* If not, commit and try again. */
5359                 ret = btrfs_commit_transaction(trans);
5360                 if (ret)
5361                         return ERR_PTR(ret);
5362         }
5363 }
5364
5365 void btrfs_evict_inode(struct inode *inode)
5366 {
5367         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5368         struct btrfs_trans_handle *trans;
5369         struct btrfs_root *root = BTRFS_I(inode)->root;
5370         struct btrfs_block_rsv *rsv;
5371         u64 min_size;
5372         int ret;
5373
5374         trace_btrfs_inode_evict(inode);
5375
5376         if (!root) {
5377                 clear_inode(inode);
5378                 return;
5379         }
5380
5381         min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
5382
5383         evict_inode_truncate_pages(inode);
5384
5385         if (inode->i_nlink &&
5386             ((btrfs_root_refs(&root->root_item) != 0 &&
5387               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5388              btrfs_is_free_space_inode(BTRFS_I(inode))))
5389                 goto no_delete;
5390
5391         if (is_bad_inode(inode))
5392                 goto no_delete;
5393         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5394         if (!special_file(inode->i_mode))
5395                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5396
5397         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5398
5399         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
5400                 goto no_delete;
5401
5402         if (inode->i_nlink > 0) {
5403                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5404                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5405                 goto no_delete;
5406         }
5407
5408         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5409         if (ret)
5410                 goto no_delete;
5411
5412         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5413         if (!rsv)
5414                 goto no_delete;
5415         rsv->size = min_size;
5416         rsv->failfast = 1;
5417
5418         btrfs_i_size_write(BTRFS_I(inode), 0);
5419
5420         while (1) {
5421                 trans = evict_refill_and_join(root, rsv, min_size);
5422                 if (IS_ERR(trans))
5423                         goto free_rsv;
5424
5425                 trans->block_rsv = rsv;
5426
5427                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5428                 trans->block_rsv = &fs_info->trans_block_rsv;
5429                 btrfs_end_transaction(trans);
5430                 btrfs_btree_balance_dirty(fs_info);
5431                 if (ret && ret != -ENOSPC && ret != -EAGAIN)
5432                         goto free_rsv;
5433                 else if (!ret)
5434                         break;
5435         }
5436
5437         /*
5438          * Errors here aren't a big deal, it just means we leave orphan items in
5439          * the tree. They will be cleaned up on the next mount. If the inode
5440          * number gets reused, cleanup deletes the orphan item without doing
5441          * anything, and unlink reuses the existing orphan item.
5442          *
5443          * If it turns out that we are dropping too many of these, we might want
5444          * to add a mechanism for retrying these after a commit.
5445          */
5446         trans = evict_refill_and_join(root, rsv, min_size);
5447         if (!IS_ERR(trans)) {
5448                 trans->block_rsv = rsv;
5449                 btrfs_orphan_del(trans, BTRFS_I(inode));
5450                 trans->block_rsv = &fs_info->trans_block_rsv;
5451                 btrfs_end_transaction(trans);
5452         }
5453
5454         if (!(root == fs_info->tree_root ||
5455               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5456                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5457
5458 free_rsv:
5459         btrfs_free_block_rsv(fs_info, rsv);
5460 no_delete:
5461         /*
5462          * If we didn't successfully delete, the orphan item will still be in
5463          * the tree and we'll retry on the next mount. Again, we might also want
5464          * to retry these periodically in the future.
5465          */
5466         btrfs_remove_delayed_node(BTRFS_I(inode));
5467         clear_inode(inode);
5468 }
5469
5470 /*
5471  * this returns the key found in the dir entry in the location pointer.
5472  * If no dir entries were found, returns -ENOENT.
5473  * If found a corrupted location in dir entry, returns -EUCLEAN.
5474  */
5475 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5476                                struct btrfs_key *location)
5477 {
5478         const char *name = dentry->d_name.name;
5479         int namelen = dentry->d_name.len;
5480         struct btrfs_dir_item *di;
5481         struct btrfs_path *path;
5482         struct btrfs_root *root = BTRFS_I(dir)->root;
5483         int ret = 0;
5484
5485         path = btrfs_alloc_path();
5486         if (!path)
5487                 return -ENOMEM;
5488
5489         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5490                         name, namelen, 0);
5491         if (!di) {
5492                 ret = -ENOENT;
5493                 goto out;
5494         }
5495         if (IS_ERR(di)) {
5496                 ret = PTR_ERR(di);
5497                 goto out;
5498         }
5499
5500         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5501         if (location->type != BTRFS_INODE_ITEM_KEY &&
5502             location->type != BTRFS_ROOT_ITEM_KEY) {
5503                 ret = -EUCLEAN;
5504                 btrfs_warn(root->fs_info,
5505 "%s gets something invalid in DIR_ITEM (name %s, directory ino %llu, location(%llu %u %llu))",
5506                            __func__, name, btrfs_ino(BTRFS_I(dir)),
5507                            location->objectid, location->type, location->offset);
5508         }
5509 out:
5510         btrfs_free_path(path);
5511         return ret;
5512 }
5513
5514 /*
5515  * when we hit a tree root in a directory, the btrfs part of the inode
5516  * needs to be changed to reflect the root directory of the tree root.  This
5517  * is kind of like crossing a mount point.
5518  */
5519 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5520                                     struct inode *dir,
5521                                     struct dentry *dentry,
5522                                     struct btrfs_key *location,
5523                                     struct btrfs_root **sub_root)
5524 {
5525         struct btrfs_path *path;
5526         struct btrfs_root *new_root;
5527         struct btrfs_root_ref *ref;
5528         struct extent_buffer *leaf;
5529         struct btrfs_key key;
5530         int ret;
5531         int err = 0;
5532
5533         path = btrfs_alloc_path();
5534         if (!path) {
5535                 err = -ENOMEM;
5536                 goto out;
5537         }
5538
5539         err = -ENOENT;
5540         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5541         key.type = BTRFS_ROOT_REF_KEY;
5542         key.offset = location->objectid;
5543
5544         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5545         if (ret) {
5546                 if (ret < 0)
5547                         err = ret;
5548                 goto out;
5549         }
5550
5551         leaf = path->nodes[0];
5552         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5553         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5554             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5555                 goto out;
5556
5557         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5558                                    (unsigned long)(ref + 1),
5559                                    dentry->d_name.len);
5560         if (ret)
5561                 goto out;
5562
5563         btrfs_release_path(path);
5564
5565         new_root = btrfs_read_fs_root_no_name(fs_info, location);
5566         if (IS_ERR(new_root)) {
5567                 err = PTR_ERR(new_root);
5568                 goto out;
5569         }
5570
5571         *sub_root = new_root;
5572         location->objectid = btrfs_root_dirid(&new_root->root_item);
5573         location->type = BTRFS_INODE_ITEM_KEY;
5574         location->offset = 0;
5575         err = 0;
5576 out:
5577         btrfs_free_path(path);
5578         return err;
5579 }
5580
5581 static void inode_tree_add(struct inode *inode)
5582 {
5583         struct btrfs_root *root = BTRFS_I(inode)->root;
5584         struct btrfs_inode *entry;
5585         struct rb_node **p;
5586         struct rb_node *parent;
5587         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5588         u64 ino = btrfs_ino(BTRFS_I(inode));
5589
5590         if (inode_unhashed(inode))
5591                 return;
5592         parent = NULL;
5593         spin_lock(&root->inode_lock);
5594         p = &root->inode_tree.rb_node;
5595         while (*p) {
5596                 parent = *p;
5597                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5598
5599                 if (ino < btrfs_ino(entry))
5600                         p = &parent->rb_left;
5601                 else if (ino > btrfs_ino(entry))
5602                         p = &parent->rb_right;
5603                 else {
5604                         WARN_ON(!(entry->vfs_inode.i_state &
5605                                   (I_WILL_FREE | I_FREEING)));
5606                         rb_replace_node(parent, new, &root->inode_tree);
5607                         RB_CLEAR_NODE(parent);
5608                         spin_unlock(&root->inode_lock);
5609                         return;
5610                 }
5611         }
5612         rb_link_node(new, parent, p);
5613         rb_insert_color(new, &root->inode_tree);
5614         spin_unlock(&root->inode_lock);
5615 }
5616
5617 static void inode_tree_del(struct inode *inode)
5618 {
5619         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5620         struct btrfs_root *root = BTRFS_I(inode)->root;
5621         int empty = 0;
5622
5623         spin_lock(&root->inode_lock);
5624         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5625                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5626                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5627                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5628         }
5629         spin_unlock(&root->inode_lock);
5630
5631         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5632                 synchronize_srcu(&fs_info->subvol_srcu);
5633                 spin_lock(&root->inode_lock);
5634                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5635                 spin_unlock(&root->inode_lock);
5636                 if (empty)
5637                         btrfs_add_dead_root(root);
5638         }
5639 }
5640
5641
5642 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5643 {
5644         struct btrfs_iget_args *args = p;
5645         inode->i_ino = args->location->objectid;
5646         memcpy(&BTRFS_I(inode)->location, args->location,
5647                sizeof(*args->location));
5648         BTRFS_I(inode)->root = args->root;
5649         return 0;
5650 }
5651
5652 static int btrfs_find_actor(struct inode *inode, void *opaque)
5653 {
5654         struct btrfs_iget_args *args = opaque;
5655         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5656                 args->root == BTRFS_I(inode)->root;
5657 }
5658
5659 static struct inode *btrfs_iget_locked(struct super_block *s,
5660                                        struct btrfs_key *location,
5661                                        struct btrfs_root *root)
5662 {
5663         struct inode *inode;
5664         struct btrfs_iget_args args;
5665         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5666
5667         args.location = location;
5668         args.root = root;
5669
5670         inode = iget5_locked(s, hashval, btrfs_find_actor,
5671                              btrfs_init_locked_inode,
5672                              (void *)&args);
5673         return inode;
5674 }
5675
5676 /* Get an inode object given its location and corresponding root.
5677  * Returns in *is_new if the inode was read from disk
5678  */
5679 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5680                          struct btrfs_root *root, int *new)
5681 {
5682         struct inode *inode;
5683
5684         inode = btrfs_iget_locked(s, location, root);
5685         if (!inode)
5686                 return ERR_PTR(-ENOMEM);
5687
5688         if (inode->i_state & I_NEW) {
5689                 int ret;
5690
5691                 ret = btrfs_read_locked_inode(inode);
5692                 if (!ret) {
5693                         inode_tree_add(inode);
5694                         unlock_new_inode(inode);
5695                         if (new)
5696                                 *new = 1;
5697                 } else {
5698                         iget_failed(inode);
5699                         /*
5700                          * ret > 0 can come from btrfs_search_slot called by
5701                          * btrfs_read_locked_inode, this means the inode item
5702                          * was not found.
5703                          */
5704                         if (ret > 0)
5705                                 ret = -ENOENT;
5706                         inode = ERR_PTR(ret);
5707                 }
5708         }
5709
5710         return inode;
5711 }
5712
5713 static struct inode *new_simple_dir(struct super_block *s,
5714                                     struct btrfs_key *key,
5715                                     struct btrfs_root *root)
5716 {
5717         struct inode *inode = new_inode(s);
5718
5719         if (!inode)
5720                 return ERR_PTR(-ENOMEM);
5721
5722         BTRFS_I(inode)->root = root;
5723         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5724         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5725
5726         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5727         inode->i_op = &btrfs_dir_ro_inode_operations;
5728         inode->i_opflags &= ~IOP_XATTR;
5729         inode->i_fop = &simple_dir_operations;
5730         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5731         inode->i_mtime = current_time(inode);
5732         inode->i_atime = inode->i_mtime;
5733         inode->i_ctime = inode->i_mtime;
5734         BTRFS_I(inode)->i_otime = inode->i_mtime;
5735
5736         return inode;
5737 }
5738
5739 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5740 {
5741         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5742         struct inode *inode;
5743         struct btrfs_root *root = BTRFS_I(dir)->root;
5744         struct btrfs_root *sub_root = root;
5745         struct btrfs_key location;
5746         int index;
5747         int ret = 0;
5748
5749         if (dentry->d_name.len > BTRFS_NAME_LEN)
5750                 return ERR_PTR(-ENAMETOOLONG);
5751
5752         ret = btrfs_inode_by_name(dir, dentry, &location);
5753         if (ret < 0)
5754                 return ERR_PTR(ret);
5755
5756         if (location.type == BTRFS_INODE_ITEM_KEY) {
5757                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5758                 return inode;
5759         }
5760
5761         index = srcu_read_lock(&fs_info->subvol_srcu);
5762         ret = fixup_tree_root_location(fs_info, dir, dentry,
5763                                        &location, &sub_root);
5764         if (ret < 0) {
5765                 if (ret != -ENOENT)
5766                         inode = ERR_PTR(ret);
5767                 else
5768                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5769         } else {
5770                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5771         }
5772         srcu_read_unlock(&fs_info->subvol_srcu, index);
5773
5774         if (!IS_ERR(inode) && root != sub_root) {
5775                 down_read(&fs_info->cleanup_work_sem);
5776                 if (!sb_rdonly(inode->i_sb))
5777                         ret = btrfs_orphan_cleanup(sub_root);
5778                 up_read(&fs_info->cleanup_work_sem);
5779                 if (ret) {
5780                         iput(inode);
5781                         inode = ERR_PTR(ret);
5782                 }
5783         }
5784
5785         return inode;
5786 }
5787
5788 static int btrfs_dentry_delete(const struct dentry *dentry)
5789 {
5790         struct btrfs_root *root;
5791         struct inode *inode = d_inode(dentry);
5792
5793         if (!inode && !IS_ROOT(dentry))
5794                 inode = d_inode(dentry->d_parent);
5795
5796         if (inode) {
5797                 root = BTRFS_I(inode)->root;
5798                 if (btrfs_root_refs(&root->root_item) == 0)
5799                         return 1;
5800
5801                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5802                         return 1;
5803         }
5804         return 0;
5805 }
5806
5807 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5808                                    unsigned int flags)
5809 {
5810         struct inode *inode = btrfs_lookup_dentry(dir, dentry);
5811
5812         if (inode == ERR_PTR(-ENOENT))
5813                 inode = NULL;
5814         return d_splice_alias(inode, dentry);
5815 }
5816
5817 unsigned char btrfs_filetype_table[] = {
5818         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5819 };
5820
5821 /*
5822  * All this infrastructure exists because dir_emit can fault, and we are holding
5823  * the tree lock when doing readdir.  For now just allocate a buffer and copy
5824  * our information into that, and then dir_emit from the buffer.  This is
5825  * similar to what NFS does, only we don't keep the buffer around in pagecache
5826  * because I'm afraid I'll mess that up.  Long term we need to make filldir do
5827  * copy_to_user_inatomic so we don't have to worry about page faulting under the
5828  * tree lock.
5829  */
5830 static int btrfs_opendir(struct inode *inode, struct file *file)
5831 {
5832         struct btrfs_file_private *private;
5833
5834         private = kzalloc(sizeof(struct btrfs_file_private), GFP_KERNEL);
5835         if (!private)
5836                 return -ENOMEM;
5837         private->filldir_buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
5838         if (!private->filldir_buf) {
5839                 kfree(private);
5840                 return -ENOMEM;
5841         }
5842         file->private_data = private;
5843         return 0;
5844 }
5845
5846 struct dir_entry {
5847         u64 ino;
5848         u64 offset;
5849         unsigned type;
5850         int name_len;
5851 };
5852
5853 static int btrfs_filldir(void *addr, int entries, struct dir_context *ctx)
5854 {
5855         while (entries--) {
5856                 struct dir_entry *entry = addr;
5857                 char *name = (char *)(entry + 1);
5858
5859                 ctx->pos = get_unaligned(&entry->offset);
5860                 if (!dir_emit(ctx, name, get_unaligned(&entry->name_len),
5861                                          get_unaligned(&entry->ino),
5862                                          get_unaligned(&entry->type)))
5863                         return 1;
5864                 addr += sizeof(struct dir_entry) +
5865                         get_unaligned(&entry->name_len);
5866                 ctx->pos++;
5867         }
5868         return 0;
5869 }
5870
5871 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5872 {
5873         struct inode *inode = file_inode(file);
5874         struct btrfs_root *root = BTRFS_I(inode)->root;
5875         struct btrfs_file_private *private = file->private_data;
5876         struct btrfs_dir_item *di;
5877         struct btrfs_key key;
5878         struct btrfs_key found_key;
5879         struct btrfs_path *path;
5880         void *addr;
5881         struct list_head ins_list;
5882         struct list_head del_list;
5883         int ret;
5884         struct extent_buffer *leaf;
5885         int slot;
5886         char *name_ptr;
5887         int name_len;
5888         int entries = 0;
5889         int total_len = 0;
5890         bool put = false;
5891         struct btrfs_key location;
5892
5893         if (!dir_emit_dots(file, ctx))
5894                 return 0;
5895
5896         path = btrfs_alloc_path();
5897         if (!path)
5898                 return -ENOMEM;
5899
5900         addr = private->filldir_buf;
5901         path->reada = READA_FORWARD;
5902
5903         INIT_LIST_HEAD(&ins_list);
5904         INIT_LIST_HEAD(&del_list);
5905         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
5906
5907 again:
5908         key.type = BTRFS_DIR_INDEX_KEY;
5909         key.offset = ctx->pos;
5910         key.objectid = btrfs_ino(BTRFS_I(inode));
5911
5912         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5913         if (ret < 0)
5914                 goto err;
5915
5916         while (1) {
5917                 struct dir_entry *entry;
5918
5919                 leaf = path->nodes[0];
5920                 slot = path->slots[0];
5921                 if (slot >= btrfs_header_nritems(leaf)) {
5922                         ret = btrfs_next_leaf(root, path);
5923                         if (ret < 0)
5924                                 goto err;
5925                         else if (ret > 0)
5926                                 break;
5927                         continue;
5928                 }
5929
5930                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5931
5932                 if (found_key.objectid != key.objectid)
5933                         break;
5934                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
5935                         break;
5936                 if (found_key.offset < ctx->pos)
5937                         goto next;
5938                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
5939                         goto next;
5940                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5941                 name_len = btrfs_dir_name_len(leaf, di);
5942                 if ((total_len + sizeof(struct dir_entry) + name_len) >=
5943                     PAGE_SIZE) {
5944                         btrfs_release_path(path);
5945                         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
5946                         if (ret)
5947                                 goto nopos;
5948                         addr = private->filldir_buf;
5949                         entries = 0;
5950                         total_len = 0;
5951                         goto again;
5952                 }
5953
5954                 entry = addr;
5955                 put_unaligned(name_len, &entry->name_len);
5956                 name_ptr = (char *)(entry + 1);
5957                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
5958                                    name_len);
5959                 put_unaligned(btrfs_filetype_table[btrfs_dir_type(leaf, di)],
5960                                 &entry->type);
5961                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
5962                 put_unaligned(location.objectid, &entry->ino);
5963                 put_unaligned(found_key.offset, &entry->offset);
5964                 entries++;
5965                 addr += sizeof(struct dir_entry) + name_len;
5966                 total_len += sizeof(struct dir_entry) + name_len;
5967 next:
5968                 path->slots[0]++;
5969         }
5970         btrfs_release_path(path);
5971
5972         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
5973         if (ret)
5974                 goto nopos;
5975
5976         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
5977         if (ret)
5978                 goto nopos;
5979
5980         /*
5981          * Stop new entries from being returned after we return the last
5982          * entry.
5983          *
5984          * New directory entries are assigned a strictly increasing
5985          * offset.  This means that new entries created during readdir
5986          * are *guaranteed* to be seen in the future by that readdir.
5987          * This has broken buggy programs which operate on names as
5988          * they're returned by readdir.  Until we re-use freed offsets
5989          * we have this hack to stop new entries from being returned
5990          * under the assumption that they'll never reach this huge
5991          * offset.
5992          *
5993          * This is being careful not to overflow 32bit loff_t unless the
5994          * last entry requires it because doing so has broken 32bit apps
5995          * in the past.
5996          */
5997         if (ctx->pos >= INT_MAX)
5998                 ctx->pos = LLONG_MAX;
5999         else
6000                 ctx->pos = INT_MAX;
6001 nopos:
6002         ret = 0;
6003 err:
6004         if (put)
6005                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
6006         btrfs_free_path(path);
6007         return ret;
6008 }
6009
6010 /*
6011  * This is somewhat expensive, updating the tree every time the
6012  * inode changes.  But, it is most likely to find the inode in cache.
6013  * FIXME, needs more benchmarking...there are no reasons other than performance
6014  * to keep or drop this code.
6015  */
6016 static int btrfs_dirty_inode(struct inode *inode)
6017 {
6018         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6019         struct btrfs_root *root = BTRFS_I(inode)->root;
6020         struct btrfs_trans_handle *trans;
6021         int ret;
6022
6023         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6024                 return 0;
6025
6026         trans = btrfs_join_transaction(root);
6027         if (IS_ERR(trans))
6028                 return PTR_ERR(trans);
6029
6030         ret = btrfs_update_inode(trans, root, inode);
6031         if (ret && ret == -ENOSPC) {
6032                 /* whoops, lets try again with the full transaction */
6033                 btrfs_end_transaction(trans);
6034                 trans = btrfs_start_transaction(root, 1);
6035                 if (IS_ERR(trans))
6036                         return PTR_ERR(trans);
6037
6038                 ret = btrfs_update_inode(trans, root, inode);
6039         }
6040         btrfs_end_transaction(trans);
6041         if (BTRFS_I(inode)->delayed_node)
6042                 btrfs_balance_delayed_items(fs_info);
6043
6044         return ret;
6045 }
6046
6047 /*
6048  * This is a copy of file_update_time.  We need this so we can return error on
6049  * ENOSPC for updating the inode in the case of file write and mmap writes.
6050  */
6051 static int btrfs_update_time(struct inode *inode, struct timespec64 *now,
6052                              int flags)
6053 {
6054         struct btrfs_root *root = BTRFS_I(inode)->root;
6055         bool dirty = flags & ~S_VERSION;
6056
6057         if (btrfs_root_readonly(root))
6058                 return -EROFS;
6059
6060         if (flags & S_VERSION)
6061                 dirty |= inode_maybe_inc_iversion(inode, dirty);
6062         if (flags & S_CTIME)
6063                 inode->i_ctime = *now;
6064         if (flags & S_MTIME)
6065                 inode->i_mtime = *now;
6066         if (flags & S_ATIME)
6067                 inode->i_atime = *now;
6068         return dirty ? btrfs_dirty_inode(inode) : 0;
6069 }
6070
6071 /*
6072  * find the highest existing sequence number in a directory
6073  * and then set the in-memory index_cnt variable to reflect
6074  * free sequence numbers
6075  */
6076 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
6077 {
6078         struct btrfs_root *root = inode->root;
6079         struct btrfs_key key, found_key;
6080         struct btrfs_path *path;
6081         struct extent_buffer *leaf;
6082         int ret;
6083
6084         key.objectid = btrfs_ino(inode);
6085         key.type = BTRFS_DIR_INDEX_KEY;
6086         key.offset = (u64)-1;
6087
6088         path = btrfs_alloc_path();
6089         if (!path)
6090                 return -ENOMEM;
6091
6092         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6093         if (ret < 0)
6094                 goto out;
6095         /* FIXME: we should be able to handle this */
6096         if (ret == 0)
6097                 goto out;
6098         ret = 0;
6099
6100         /*
6101          * MAGIC NUMBER EXPLANATION:
6102          * since we search a directory based on f_pos we have to start at 2
6103          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6104          * else has to start at 2
6105          */
6106         if (path->slots[0] == 0) {
6107                 inode->index_cnt = 2;
6108                 goto out;
6109         }
6110
6111         path->slots[0]--;
6112
6113         leaf = path->nodes[0];
6114         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6115
6116         if (found_key.objectid != btrfs_ino(inode) ||
6117             found_key.type != BTRFS_DIR_INDEX_KEY) {
6118                 inode->index_cnt = 2;
6119                 goto out;
6120         }
6121
6122         inode->index_cnt = found_key.offset + 1;
6123 out:
6124         btrfs_free_path(path);
6125         return ret;
6126 }
6127
6128 /*
6129  * helper to find a free sequence number in a given directory.  This current
6130  * code is very simple, later versions will do smarter things in the btree
6131  */
6132 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
6133 {
6134         int ret = 0;
6135
6136         if (dir->index_cnt == (u64)-1) {
6137                 ret = btrfs_inode_delayed_dir_index_count(dir);
6138                 if (ret) {
6139                         ret = btrfs_set_inode_index_count(dir);
6140                         if (ret)
6141                                 return ret;
6142                 }
6143         }
6144
6145         *index = dir->index_cnt;
6146         dir->index_cnt++;
6147
6148         return ret;
6149 }
6150
6151 static int btrfs_insert_inode_locked(struct inode *inode)
6152 {
6153         struct btrfs_iget_args args;
6154         args.location = &BTRFS_I(inode)->location;
6155         args.root = BTRFS_I(inode)->root;
6156
6157         return insert_inode_locked4(inode,
6158                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6159                    btrfs_find_actor, &args);
6160 }
6161
6162 /*
6163  * Inherit flags from the parent inode.
6164  *
6165  * Currently only the compression flags and the cow flags are inherited.
6166  */
6167 static void btrfs_inherit_iflags(struct inode *inode, struct inode *dir)
6168 {
6169         unsigned int flags;
6170
6171         if (!dir)
6172                 return;
6173
6174         flags = BTRFS_I(dir)->flags;
6175
6176         if (flags & BTRFS_INODE_NOCOMPRESS) {
6177                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_COMPRESS;
6178                 BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
6179         } else if (flags & BTRFS_INODE_COMPRESS) {
6180                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NOCOMPRESS;
6181                 BTRFS_I(inode)->flags |= BTRFS_INODE_COMPRESS;
6182         }
6183
6184         if (flags & BTRFS_INODE_NODATACOW) {
6185                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
6186                 if (S_ISREG(inode->i_mode))
6187                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6188         }
6189
6190         btrfs_sync_inode_flags_to_i_flags(inode);
6191 }
6192
6193 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6194                                      struct btrfs_root *root,
6195                                      struct inode *dir,
6196                                      const char *name, int name_len,
6197                                      u64 ref_objectid, u64 objectid,
6198                                      umode_t mode, u64 *index)
6199 {
6200         struct btrfs_fs_info *fs_info = root->fs_info;
6201         struct inode *inode;
6202         struct btrfs_inode_item *inode_item;
6203         struct btrfs_key *location;
6204         struct btrfs_path *path;
6205         struct btrfs_inode_ref *ref;
6206         struct btrfs_key key[2];
6207         u32 sizes[2];
6208         int nitems = name ? 2 : 1;
6209         unsigned long ptr;
6210         int ret;
6211
6212         path = btrfs_alloc_path();
6213         if (!path)
6214                 return ERR_PTR(-ENOMEM);
6215
6216         inode = new_inode(fs_info->sb);
6217         if (!inode) {
6218                 btrfs_free_path(path);
6219                 return ERR_PTR(-ENOMEM);
6220         }
6221
6222         /*
6223          * O_TMPFILE, set link count to 0, so that after this point,
6224          * we fill in an inode item with the correct link count.
6225          */
6226         if (!name)
6227                 set_nlink(inode, 0);
6228
6229         /*
6230          * we have to initialize this early, so we can reclaim the inode
6231          * number if we fail afterwards in this function.
6232          */
6233         inode->i_ino = objectid;
6234
6235         if (dir && name) {
6236                 trace_btrfs_inode_request(dir);
6237
6238                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
6239                 if (ret) {
6240                         btrfs_free_path(path);
6241                         iput(inode);
6242                         return ERR_PTR(ret);
6243                 }
6244         } else if (dir) {
6245                 *index = 0;
6246         }
6247         /*
6248          * index_cnt is ignored for everything but a dir,
6249          * btrfs_set_inode_index_count has an explanation for the magic
6250          * number
6251          */
6252         BTRFS_I(inode)->index_cnt = 2;
6253         BTRFS_I(inode)->dir_index = *index;
6254         BTRFS_I(inode)->root = root;
6255         BTRFS_I(inode)->generation = trans->transid;
6256         inode->i_generation = BTRFS_I(inode)->generation;
6257
6258         /*
6259          * We could have gotten an inode number from somebody who was fsynced
6260          * and then removed in this same transaction, so let's just set full
6261          * sync since it will be a full sync anyway and this will blow away the
6262          * old info in the log.
6263          */
6264         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6265
6266         key[0].objectid = objectid;
6267         key[0].type = BTRFS_INODE_ITEM_KEY;
6268         key[0].offset = 0;
6269
6270         sizes[0] = sizeof(struct btrfs_inode_item);
6271
6272         if (name) {
6273                 /*
6274                  * Start new inodes with an inode_ref. This is slightly more
6275                  * efficient for small numbers of hard links since they will
6276                  * be packed into one item. Extended refs will kick in if we
6277                  * add more hard links than can fit in the ref item.
6278                  */
6279                 key[1].objectid = objectid;
6280                 key[1].type = BTRFS_INODE_REF_KEY;
6281                 key[1].offset = ref_objectid;
6282
6283                 sizes[1] = name_len + sizeof(*ref);
6284         }
6285
6286         location = &BTRFS_I(inode)->location;
6287         location->objectid = objectid;
6288         location->offset = 0;
6289         location->type = BTRFS_INODE_ITEM_KEY;
6290
6291         ret = btrfs_insert_inode_locked(inode);
6292         if (ret < 0) {
6293                 iput(inode);
6294                 goto fail;
6295         }
6296
6297         path->leave_spinning = 1;
6298         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6299         if (ret != 0)
6300                 goto fail_unlock;
6301
6302         inode_init_owner(inode, dir, mode);
6303         inode_set_bytes(inode, 0);
6304
6305         inode->i_mtime = current_time(inode);
6306         inode->i_atime = inode->i_mtime;
6307         inode->i_ctime = inode->i_mtime;
6308         BTRFS_I(inode)->i_otime = inode->i_mtime;
6309
6310         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6311                                   struct btrfs_inode_item);
6312         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6313                              sizeof(*inode_item));
6314         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6315
6316         if (name) {
6317                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6318                                      struct btrfs_inode_ref);
6319                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6320                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6321                 ptr = (unsigned long)(ref + 1);
6322                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6323         }
6324
6325         btrfs_mark_buffer_dirty(path->nodes[0]);
6326         btrfs_free_path(path);
6327
6328         btrfs_inherit_iflags(inode, dir);
6329
6330         if (S_ISREG(mode)) {
6331                 if (btrfs_test_opt(fs_info, NODATASUM))
6332                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6333                 if (btrfs_test_opt(fs_info, NODATACOW))
6334                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6335                                 BTRFS_INODE_NODATASUM;
6336         }
6337
6338         inode_tree_add(inode);
6339
6340         trace_btrfs_inode_new(inode);
6341         btrfs_set_inode_last_trans(trans, inode);
6342
6343         btrfs_update_root_times(trans, root);
6344
6345         ret = btrfs_inode_inherit_props(trans, inode, dir);
6346         if (ret)
6347                 btrfs_err(fs_info,
6348                           "error inheriting props for ino %llu (root %llu): %d",
6349                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6350
6351         return inode;
6352
6353 fail_unlock:
6354         discard_new_inode(inode);
6355 fail:
6356         if (dir && name)
6357                 BTRFS_I(dir)->index_cnt--;
6358         btrfs_free_path(path);
6359         return ERR_PTR(ret);
6360 }
6361
6362 static inline u8 btrfs_inode_type(struct inode *inode)
6363 {
6364         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6365 }
6366
6367 /*
6368  * utility function to add 'inode' into 'parent_inode' with
6369  * a give name and a given sequence number.
6370  * if 'add_backref' is true, also insert a backref from the
6371  * inode to the parent directory.
6372  */
6373 int btrfs_add_link(struct btrfs_trans_handle *trans,
6374                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6375                    const char *name, int name_len, int add_backref, u64 index)
6376 {
6377         int ret = 0;
6378         struct btrfs_key key;
6379         struct btrfs_root *root = parent_inode->root;
6380         u64 ino = btrfs_ino(inode);
6381         u64 parent_ino = btrfs_ino(parent_inode);
6382
6383         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6384                 memcpy(&key, &inode->root->root_key, sizeof(key));
6385         } else {
6386                 key.objectid = ino;
6387                 key.type = BTRFS_INODE_ITEM_KEY;
6388                 key.offset = 0;
6389         }
6390
6391         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6392                 ret = btrfs_add_root_ref(trans, key.objectid,
6393                                          root->root_key.objectid, parent_ino,
6394                                          index, name, name_len);
6395         } else if (add_backref) {
6396                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6397                                              parent_ino, index);
6398         }
6399
6400         /* Nothing to clean up yet */
6401         if (ret)
6402                 return ret;
6403
6404         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6405                                     parent_inode, &key,
6406                                     btrfs_inode_type(&inode->vfs_inode), index);
6407         if (ret == -EEXIST || ret == -EOVERFLOW)
6408                 goto fail_dir_item;
6409         else if (ret) {
6410                 btrfs_abort_transaction(trans, ret);
6411                 return ret;
6412         }
6413
6414         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6415                            name_len * 2);
6416         inode_inc_iversion(&parent_inode->vfs_inode);
6417         parent_inode->vfs_inode.i_mtime = parent_inode->vfs_inode.i_ctime =
6418                 current_time(&parent_inode->vfs_inode);
6419         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6420         if (ret)
6421                 btrfs_abort_transaction(trans, ret);
6422         return ret;
6423
6424 fail_dir_item:
6425         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6426                 u64 local_index;
6427                 int err;
6428                 err = btrfs_del_root_ref(trans, key.objectid,
6429                                          root->root_key.objectid, parent_ino,
6430                                          &local_index, name, name_len);
6431
6432         } else if (add_backref) {
6433                 u64 local_index;
6434                 int err;
6435
6436                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6437                                           ino, parent_ino, &local_index);
6438         }
6439         return ret;
6440 }
6441
6442 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6443                             struct btrfs_inode *dir, struct dentry *dentry,
6444                             struct btrfs_inode *inode, int backref, u64 index)
6445 {
6446         int err = btrfs_add_link(trans, dir, inode,
6447                                  dentry->d_name.name, dentry->d_name.len,
6448                                  backref, index);
6449         if (err > 0)
6450                 err = -EEXIST;
6451         return err;
6452 }
6453
6454 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6455                         umode_t mode, dev_t rdev)
6456 {
6457         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6458         struct btrfs_trans_handle *trans;
6459         struct btrfs_root *root = BTRFS_I(dir)->root;
6460         struct inode *inode = NULL;
6461         int err;
6462         u64 objectid;
6463         u64 index = 0;
6464
6465         /*
6466          * 2 for inode item and ref
6467          * 2 for dir items
6468          * 1 for xattr if selinux is on
6469          */
6470         trans = btrfs_start_transaction(root, 5);
6471         if (IS_ERR(trans))
6472                 return PTR_ERR(trans);
6473
6474         err = btrfs_find_free_ino(root, &objectid);
6475         if (err)
6476                 goto out_unlock;
6477
6478         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6479                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6480                         mode, &index);
6481         if (IS_ERR(inode)) {
6482                 err = PTR_ERR(inode);
6483                 inode = NULL;
6484                 goto out_unlock;
6485         }
6486
6487         /*
6488         * If the active LSM wants to access the inode during
6489         * d_instantiate it needs these. Smack checks to see
6490         * if the filesystem supports xattrs by looking at the
6491         * ops vector.
6492         */
6493         inode->i_op = &btrfs_special_inode_operations;
6494         init_special_inode(inode, inode->i_mode, rdev);
6495
6496         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6497         if (err)
6498                 goto out_unlock;
6499
6500         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6501                         0, index);
6502         if (err)
6503                 goto out_unlock;
6504
6505         btrfs_update_inode(trans, root, inode);
6506         d_instantiate_new(dentry, inode);
6507
6508 out_unlock:
6509         btrfs_end_transaction(trans);
6510         btrfs_btree_balance_dirty(fs_info);
6511         if (err && inode) {
6512                 inode_dec_link_count(inode);
6513                 discard_new_inode(inode);
6514         }
6515         return err;
6516 }
6517
6518 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6519                         umode_t mode, bool excl)
6520 {
6521         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6522         struct btrfs_trans_handle *trans;
6523         struct btrfs_root *root = BTRFS_I(dir)->root;
6524         struct inode *inode = NULL;
6525         int err;
6526         u64 objectid;
6527         u64 index = 0;
6528
6529         /*
6530          * 2 for inode item and ref
6531          * 2 for dir items
6532          * 1 for xattr if selinux is on
6533          */
6534         trans = btrfs_start_transaction(root, 5);
6535         if (IS_ERR(trans))
6536                 return PTR_ERR(trans);
6537
6538         err = btrfs_find_free_ino(root, &objectid);
6539         if (err)
6540                 goto out_unlock;
6541
6542         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6543                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6544                         mode, &index);
6545         if (IS_ERR(inode)) {
6546                 err = PTR_ERR(inode);
6547                 inode = NULL;
6548                 goto out_unlock;
6549         }
6550         /*
6551         * If the active LSM wants to access the inode during
6552         * d_instantiate it needs these. Smack checks to see
6553         * if the filesystem supports xattrs by looking at the
6554         * ops vector.
6555         */
6556         inode->i_fop = &btrfs_file_operations;
6557         inode->i_op = &btrfs_file_inode_operations;
6558         inode->i_mapping->a_ops = &btrfs_aops;
6559
6560         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6561         if (err)
6562                 goto out_unlock;
6563
6564         err = btrfs_update_inode(trans, root, inode);
6565         if (err)
6566                 goto out_unlock;
6567
6568         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6569                         0, index);
6570         if (err)
6571                 goto out_unlock;
6572
6573         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6574         d_instantiate_new(dentry, inode);
6575
6576 out_unlock:
6577         btrfs_end_transaction(trans);
6578         if (err && inode) {
6579                 inode_dec_link_count(inode);
6580                 discard_new_inode(inode);
6581         }
6582         btrfs_btree_balance_dirty(fs_info);
6583         return err;
6584 }
6585
6586 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6587                       struct dentry *dentry)
6588 {
6589         struct btrfs_trans_handle *trans = NULL;
6590         struct btrfs_root *root = BTRFS_I(dir)->root;
6591         struct inode *inode = d_inode(old_dentry);
6592         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6593         u64 index;
6594         int err;
6595         int drop_inode = 0;
6596
6597         /* do not allow sys_link's with other subvols of the same device */
6598         if (root->objectid != BTRFS_I(inode)->root->objectid)
6599                 return -EXDEV;
6600
6601         if (inode->i_nlink >= BTRFS_LINK_MAX)
6602                 return -EMLINK;
6603
6604         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6605         if (err)
6606                 goto fail;
6607
6608         /*
6609          * 2 items for inode and inode ref
6610          * 2 items for dir items
6611          * 1 item for parent inode
6612          * 1 item for orphan item deletion if O_TMPFILE
6613          */
6614         trans = btrfs_start_transaction(root, inode->i_nlink ? 5 : 6);
6615         if (IS_ERR(trans)) {
6616                 err = PTR_ERR(trans);
6617                 trans = NULL;
6618                 goto fail;
6619         }
6620
6621         /* There are several dir indexes for this inode, clear the cache. */
6622         BTRFS_I(inode)->dir_index = 0ULL;
6623         inc_nlink(inode);
6624         inode_inc_iversion(inode);
6625         inode->i_ctime = current_time(inode);
6626         ihold(inode);
6627         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6628
6629         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6630                         1, index);
6631
6632         if (err) {
6633                 drop_inode = 1;
6634         } else {
6635                 struct dentry *parent = dentry->d_parent;
6636                 err = btrfs_update_inode(trans, root, inode);
6637                 if (err)
6638                         goto fail;
6639                 if (inode->i_nlink == 1) {
6640                         /*
6641                          * If new hard link count is 1, it's a file created
6642                          * with open(2) O_TMPFILE flag.
6643                          */
6644                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6645                         if (err)
6646                                 goto fail;
6647                 }
6648                 d_instantiate(dentry, inode);
6649                 btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent);
6650         }
6651
6652 fail:
6653         if (trans)
6654                 btrfs_end_transaction(trans);
6655         if (drop_inode) {
6656                 inode_dec_link_count(inode);
6657                 iput(inode);
6658         }
6659         btrfs_btree_balance_dirty(fs_info);
6660         return err;
6661 }
6662
6663 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6664 {
6665         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6666         struct inode *inode = NULL;
6667         struct btrfs_trans_handle *trans;
6668         struct btrfs_root *root = BTRFS_I(dir)->root;
6669         int err = 0;
6670         int drop_on_err = 0;
6671         u64 objectid = 0;
6672         u64 index = 0;
6673
6674         /*
6675          * 2 items for inode and ref
6676          * 2 items for dir items
6677          * 1 for xattr if selinux is on
6678          */
6679         trans = btrfs_start_transaction(root, 5);
6680         if (IS_ERR(trans))
6681                 return PTR_ERR(trans);
6682
6683         err = btrfs_find_free_ino(root, &objectid);
6684         if (err)
6685                 goto out_fail;
6686
6687         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6688                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6689                         S_IFDIR | mode, &index);
6690         if (IS_ERR(inode)) {
6691                 err = PTR_ERR(inode);
6692                 inode = NULL;
6693                 goto out_fail;
6694         }
6695
6696         drop_on_err = 1;
6697         /* these must be set before we unlock the inode */
6698         inode->i_op = &btrfs_dir_inode_operations;
6699         inode->i_fop = &btrfs_dir_file_operations;
6700
6701         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6702         if (err)
6703                 goto out_fail;
6704
6705         btrfs_i_size_write(BTRFS_I(inode), 0);
6706         err = btrfs_update_inode(trans, root, inode);
6707         if (err)
6708                 goto out_fail;
6709
6710         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6711                         dentry->d_name.name,
6712                         dentry->d_name.len, 0, index);
6713         if (err)
6714                 goto out_fail;
6715
6716         d_instantiate_new(dentry, inode);
6717         drop_on_err = 0;
6718
6719 out_fail:
6720         btrfs_end_transaction(trans);
6721         if (err && inode) {
6722                 inode_dec_link_count(inode);
6723                 discard_new_inode(inode);
6724         }
6725         btrfs_btree_balance_dirty(fs_info);
6726         return err;
6727 }
6728
6729 static noinline int uncompress_inline(struct btrfs_path *path,
6730                                       struct page *page,
6731                                       size_t pg_offset, u64 extent_offset,
6732                                       struct btrfs_file_extent_item *item)
6733 {
6734         int ret;
6735         struct extent_buffer *leaf = path->nodes[0];
6736         char *tmp;
6737         size_t max_size;
6738         unsigned long inline_size;
6739         unsigned long ptr;
6740         int compress_type;
6741
6742         WARN_ON(pg_offset != 0);
6743         compress_type = btrfs_file_extent_compression(leaf, item);
6744         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6745         inline_size = btrfs_file_extent_inline_item_len(leaf,
6746                                         btrfs_item_nr(path->slots[0]));
6747         tmp = kmalloc(inline_size, GFP_NOFS);
6748         if (!tmp)
6749                 return -ENOMEM;
6750         ptr = btrfs_file_extent_inline_start(item);
6751
6752         read_extent_buffer(leaf, tmp, ptr, inline_size);
6753
6754         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6755         ret = btrfs_decompress(compress_type, tmp, page,
6756                                extent_offset, inline_size, max_size);
6757
6758         /*
6759          * decompression code contains a memset to fill in any space between the end
6760          * of the uncompressed data and the end of max_size in case the decompressed
6761          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6762          * the end of an inline extent and the beginning of the next block, so we
6763          * cover that region here.
6764          */
6765
6766         if (max_size + pg_offset < PAGE_SIZE) {
6767                 char *map = kmap(page);
6768                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6769                 kunmap(page);
6770         }
6771         kfree(tmp);
6772         return ret;
6773 }
6774
6775 /*
6776  * a bit scary, this does extent mapping from logical file offset to the disk.
6777  * the ugly parts come from merging extents from the disk with the in-ram
6778  * representation.  This gets more complex because of the data=ordered code,
6779  * where the in-ram extents might be locked pending data=ordered completion.
6780  *
6781  * This also copies inline extents directly into the page.
6782  */
6783 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6784                 struct page *page,
6785             size_t pg_offset, u64 start, u64 len,
6786                 int create)
6787 {
6788         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6789         int ret;
6790         int err = 0;
6791         u64 extent_start = 0;
6792         u64 extent_end = 0;
6793         u64 objectid = btrfs_ino(inode);
6794         u32 found_type;
6795         struct btrfs_path *path = NULL;
6796         struct btrfs_root *root = inode->root;
6797         struct btrfs_file_extent_item *item;
6798         struct extent_buffer *leaf;
6799         struct btrfs_key found_key;
6800         struct extent_map *em = NULL;
6801         struct extent_map_tree *em_tree = &inode->extent_tree;
6802         struct extent_io_tree *io_tree = &inode->io_tree;
6803         const bool new_inline = !page || create;
6804
6805         read_lock(&em_tree->lock);
6806         em = lookup_extent_mapping(em_tree, start, len);
6807         if (em)
6808                 em->bdev = fs_info->fs_devices->latest_bdev;
6809         read_unlock(&em_tree->lock);
6810
6811         if (em) {
6812                 if (em->start > start || em->start + em->len <= start)
6813                         free_extent_map(em);
6814                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6815                         free_extent_map(em);
6816                 else
6817                         goto out;
6818         }
6819         em = alloc_extent_map();
6820         if (!em) {
6821                 err = -ENOMEM;
6822                 goto out;
6823         }
6824         em->bdev = fs_info->fs_devices->latest_bdev;
6825         em->start = EXTENT_MAP_HOLE;
6826         em->orig_start = EXTENT_MAP_HOLE;
6827         em->len = (u64)-1;
6828         em->block_len = (u64)-1;
6829
6830         if (!path) {
6831                 path = btrfs_alloc_path();
6832                 if (!path) {
6833                         err = -ENOMEM;
6834                         goto out;
6835                 }
6836                 /*
6837                  * Chances are we'll be called again, so go ahead and do
6838                  * readahead
6839                  */
6840                 path->reada = READA_FORWARD;
6841         }
6842
6843         ret = btrfs_lookup_file_extent(NULL, root, path, objectid, start, 0);
6844         if (ret < 0) {
6845                 err = ret;
6846                 goto out;
6847         }
6848
6849         if (ret != 0) {
6850                 if (path->slots[0] == 0)
6851                         goto not_found;
6852                 path->slots[0]--;
6853         }
6854
6855         leaf = path->nodes[0];
6856         item = btrfs_item_ptr(leaf, path->slots[0],
6857                               struct btrfs_file_extent_item);
6858         /* are we inside the extent that was found? */
6859         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6860         found_type = found_key.type;
6861         if (found_key.objectid != objectid ||
6862             found_type != BTRFS_EXTENT_DATA_KEY) {
6863                 /*
6864                  * If we backup past the first extent we want to move forward
6865                  * and see if there is an extent in front of us, otherwise we'll
6866                  * say there is a hole for our whole search range which can
6867                  * cause problems.
6868                  */
6869                 extent_end = start;
6870                 goto next;
6871         }
6872
6873         found_type = btrfs_file_extent_type(leaf, item);
6874         extent_start = found_key.offset;
6875         if (found_type == BTRFS_FILE_EXTENT_REG ||
6876             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6877                 extent_end = extent_start +
6878                        btrfs_file_extent_num_bytes(leaf, item);
6879
6880                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
6881                                                        extent_start);
6882         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6883                 size_t size;
6884
6885                 size = btrfs_file_extent_ram_bytes(leaf, item);
6886                 extent_end = ALIGN(extent_start + size,
6887                                    fs_info->sectorsize);
6888
6889                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
6890                                                       path->slots[0],
6891                                                       extent_start);
6892         }
6893 next:
6894         if (start >= extent_end) {
6895                 path->slots[0]++;
6896                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6897                         ret = btrfs_next_leaf(root, path);
6898                         if (ret < 0) {
6899                                 err = ret;
6900                                 goto out;
6901                         }
6902                         if (ret > 0)
6903                                 goto not_found;
6904                         leaf = path->nodes[0];
6905                 }
6906                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6907                 if (found_key.objectid != objectid ||
6908                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6909                         goto not_found;
6910                 if (start + len <= found_key.offset)
6911                         goto not_found;
6912                 if (start > found_key.offset)
6913                         goto next;
6914                 em->start = start;
6915                 em->orig_start = start;
6916                 em->len = found_key.offset - start;
6917                 goto not_found_em;
6918         }
6919
6920         btrfs_extent_item_to_extent_map(inode, path, item,
6921                         new_inline, em);
6922
6923         if (found_type == BTRFS_FILE_EXTENT_REG ||
6924             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6925                 goto insert;
6926         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6927                 unsigned long ptr;
6928                 char *map;
6929                 size_t size;
6930                 size_t extent_offset;
6931                 size_t copy_size;
6932
6933                 if (new_inline)
6934                         goto out;
6935
6936                 size = btrfs_file_extent_ram_bytes(leaf, item);
6937                 extent_offset = page_offset(page) + pg_offset - extent_start;
6938                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
6939                                   size - extent_offset);
6940                 em->start = extent_start + extent_offset;
6941                 em->len = ALIGN(copy_size, fs_info->sectorsize);
6942                 em->orig_block_len = em->len;
6943                 em->orig_start = em->start;
6944                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6945                 if (!PageUptodate(page)) {
6946                         if (btrfs_file_extent_compression(leaf, item) !=
6947                             BTRFS_COMPRESS_NONE) {
6948                                 ret = uncompress_inline(path, page, pg_offset,
6949                                                         extent_offset, item);
6950                                 if (ret) {
6951                                         err = ret;
6952                                         goto out;
6953                                 }
6954                         } else {
6955                                 map = kmap(page);
6956                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6957                                                    copy_size);
6958                                 if (pg_offset + copy_size < PAGE_SIZE) {
6959                                         memset(map + pg_offset + copy_size, 0,
6960                                                PAGE_SIZE - pg_offset -
6961                                                copy_size);
6962                                 }
6963                                 kunmap(page);
6964                         }
6965                         flush_dcache_page(page);
6966                 }
6967                 set_extent_uptodate(io_tree, em->start,
6968                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6969                 goto insert;
6970         }
6971 not_found:
6972         em->start = start;
6973         em->orig_start = start;
6974         em->len = len;
6975 not_found_em:
6976         em->block_start = EXTENT_MAP_HOLE;
6977 insert:
6978         btrfs_release_path(path);
6979         if (em->start > start || extent_map_end(em) <= start) {
6980                 btrfs_err(fs_info,
6981                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
6982                           em->start, em->len, start, len);
6983                 err = -EIO;
6984                 goto out;
6985         }
6986
6987         err = 0;
6988         write_lock(&em_tree->lock);
6989         err = btrfs_add_extent_mapping(fs_info, em_tree, &em, start, len);
6990         write_unlock(&em_tree->lock);
6991 out:
6992
6993         trace_btrfs_get_extent(root, inode, em);
6994
6995         btrfs_free_path(path);
6996         if (err) {
6997                 free_extent_map(em);
6998                 return ERR_PTR(err);
6999         }
7000         BUG_ON(!em); /* Error is always set */
7001         return em;
7002 }
7003
7004 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
7005                 struct page *page,
7006                 size_t pg_offset, u64 start, u64 len,
7007                 int create)
7008 {
7009         struct extent_map *em;
7010         struct extent_map *hole_em = NULL;
7011         u64 range_start = start;
7012         u64 end;
7013         u64 found;
7014         u64 found_end;
7015         int err = 0;
7016
7017         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7018         if (IS_ERR(em))
7019                 return em;
7020         /*
7021          * If our em maps to:
7022          * - a hole or
7023          * - a pre-alloc extent,
7024          * there might actually be delalloc bytes behind it.
7025          */
7026         if (em->block_start != EXTENT_MAP_HOLE &&
7027             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7028                 return em;
7029         else
7030                 hole_em = em;
7031
7032         /* check to see if we've wrapped (len == -1 or similar) */
7033         end = start + len;
7034         if (end < start)
7035                 end = (u64)-1;
7036         else
7037                 end -= 1;
7038
7039         em = NULL;
7040
7041         /* ok, we didn't find anything, lets look for delalloc */
7042         found = count_range_bits(&inode->io_tree, &range_start,
7043                                  end, len, EXTENT_DELALLOC, 1);
7044         found_end = range_start + found;
7045         if (found_end < range_start)
7046                 found_end = (u64)-1;
7047
7048         /*
7049          * we didn't find anything useful, return
7050          * the original results from get_extent()
7051          */
7052         if (range_start > end || found_end <= start) {
7053                 em = hole_em;
7054                 hole_em = NULL;
7055                 goto out;
7056         }
7057
7058         /* adjust the range_start to make sure it doesn't
7059          * go backwards from the start they passed in
7060          */
7061         range_start = max(start, range_start);
7062         found = found_end - range_start;
7063
7064         if (found > 0) {
7065                 u64 hole_start = start;
7066                 u64 hole_len = len;
7067
7068                 em = alloc_extent_map();
7069                 if (!em) {
7070                         err = -ENOMEM;
7071                         goto out;
7072                 }
7073                 /*
7074                  * when btrfs_get_extent can't find anything it
7075                  * returns one huge hole
7076                  *
7077                  * make sure what it found really fits our range, and
7078                  * adjust to make sure it is based on the start from
7079                  * the caller
7080                  */
7081                 if (hole_em) {
7082                         u64 calc_end = extent_map_end(hole_em);
7083
7084                         if (calc_end <= start || (hole_em->start > end)) {
7085                                 free_extent_map(hole_em);
7086                                 hole_em = NULL;
7087                         } else {
7088                                 hole_start = max(hole_em->start, start);
7089                                 hole_len = calc_end - hole_start;
7090                         }
7091                 }
7092                 em->bdev = NULL;
7093                 if (hole_em && range_start > hole_start) {
7094                         /* our hole starts before our delalloc, so we
7095                          * have to return just the parts of the hole
7096                          * that go until  the delalloc starts
7097                          */
7098                         em->len = min(hole_len,
7099                                       range_start - hole_start);
7100                         em->start = hole_start;
7101                         em->orig_start = hole_start;
7102                         /*
7103                          * don't adjust block start at all,
7104                          * it is fixed at EXTENT_MAP_HOLE
7105                          */
7106                         em->block_start = hole_em->block_start;
7107                         em->block_len = hole_len;
7108                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7109                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7110                 } else {
7111                         em->start = range_start;
7112                         em->len = found;
7113                         em->orig_start = range_start;
7114                         em->block_start = EXTENT_MAP_DELALLOC;
7115                         em->block_len = found;
7116                 }
7117         } else {
7118                 return hole_em;
7119         }
7120 out:
7121
7122         free_extent_map(hole_em);
7123         if (err) {
7124                 free_extent_map(em);
7125                 return ERR_PTR(err);
7126         }
7127         return em;
7128 }
7129
7130 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7131                                                   const u64 start,
7132                                                   const u64 len,
7133                                                   const u64 orig_start,
7134                                                   const u64 block_start,
7135                                                   const u64 block_len,
7136                                                   const u64 orig_block_len,
7137                                                   const u64 ram_bytes,
7138                                                   const int type)
7139 {
7140         struct extent_map *em = NULL;
7141         int ret;
7142
7143         if (type != BTRFS_ORDERED_NOCOW) {
7144                 em = create_io_em(inode, start, len, orig_start,
7145                                   block_start, block_len, orig_block_len,
7146                                   ram_bytes,
7147                                   BTRFS_COMPRESS_NONE, /* compress_type */
7148                                   type);
7149                 if (IS_ERR(em))
7150                         goto out;
7151         }
7152         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7153                                            len, block_len, type);
7154         if (ret) {
7155                 if (em) {
7156                         free_extent_map(em);
7157                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
7158                                                 start + len - 1, 0);
7159                 }
7160                 em = ERR_PTR(ret);
7161         }
7162  out:
7163
7164         return em;
7165 }
7166
7167 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7168                                                   u64 start, u64 len)
7169 {
7170         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7171         struct btrfs_root *root = BTRFS_I(inode)->root;
7172         struct extent_map *em;
7173         struct btrfs_key ins;
7174         u64 alloc_hint;
7175         int ret;
7176
7177         alloc_hint = get_extent_allocation_hint(inode, start, len);
7178         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
7179                                    0, alloc_hint, &ins, 1, 1);
7180         if (ret)
7181                 return ERR_PTR(ret);
7182
7183         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7184                                      ins.objectid, ins.offset, ins.offset,
7185                                      ins.offset, BTRFS_ORDERED_REGULAR);
7186         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
7187         if (IS_ERR(em))
7188                 btrfs_free_reserved_extent(fs_info, ins.objectid,
7189                                            ins.offset, 1);
7190
7191         return em;
7192 }
7193
7194 /*
7195  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7196  * block must be cow'd
7197  */
7198 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7199                               u64 *orig_start, u64 *orig_block_len,
7200                               u64 *ram_bytes)
7201 {
7202         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7203         struct btrfs_path *path;
7204         int ret;
7205         struct extent_buffer *leaf;
7206         struct btrfs_root *root = BTRFS_I(inode)->root;
7207         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7208         struct btrfs_file_extent_item *fi;
7209         struct btrfs_key key;
7210         u64 disk_bytenr;
7211         u64 backref_offset;
7212         u64 extent_end;
7213         u64 num_bytes;
7214         int slot;
7215         int found_type;
7216         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7217
7218         path = btrfs_alloc_path();
7219         if (!path)
7220                 return -ENOMEM;
7221
7222         ret = btrfs_lookup_file_extent(NULL, root, path,
7223                         btrfs_ino(BTRFS_I(inode)), offset, 0);
7224         if (ret < 0)
7225                 goto out;
7226
7227         slot = path->slots[0];
7228         if (ret == 1) {
7229                 if (slot == 0) {
7230                         /* can't find the item, must cow */
7231                         ret = 0;
7232                         goto out;
7233                 }
7234                 slot--;
7235         }
7236         ret = 0;
7237         leaf = path->nodes[0];
7238         btrfs_item_key_to_cpu(leaf, &key, slot);
7239         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
7240             key.type != BTRFS_EXTENT_DATA_KEY) {
7241                 /* not our file or wrong item type, must cow */
7242                 goto out;
7243         }
7244
7245         if (key.offset > offset) {
7246                 /* Wrong offset, must cow */
7247                 goto out;
7248         }
7249
7250         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7251         found_type = btrfs_file_extent_type(leaf, fi);
7252         if (found_type != BTRFS_FILE_EXTENT_REG &&
7253             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7254                 /* not a regular extent, must cow */
7255                 goto out;
7256         }
7257
7258         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7259                 goto out;
7260
7261         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7262         if (extent_end <= offset)
7263                 goto out;
7264
7265         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7266         if (disk_bytenr == 0)
7267                 goto out;
7268
7269         if (btrfs_file_extent_compression(leaf, fi) ||
7270             btrfs_file_extent_encryption(leaf, fi) ||
7271             btrfs_file_extent_other_encoding(leaf, fi))
7272                 goto out;
7273
7274         /*
7275          * Do the same check as in btrfs_cross_ref_exist but without the
7276          * unnecessary search.
7277          */
7278         if (btrfs_file_extent_generation(leaf, fi) <=
7279             btrfs_root_last_snapshot(&root->root_item))
7280                 goto out;
7281
7282         backref_offset = btrfs_file_extent_offset(leaf, fi);
7283
7284         if (orig_start) {
7285                 *orig_start = key.offset - backref_offset;
7286                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7287                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7288         }
7289
7290         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7291                 goto out;
7292
7293         num_bytes = min(offset + *len, extent_end) - offset;
7294         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7295                 u64 range_end;
7296
7297                 range_end = round_up(offset + num_bytes,
7298                                      root->fs_info->sectorsize) - 1;
7299                 ret = test_range_bit(io_tree, offset, range_end,
7300                                      EXTENT_DELALLOC, 0, NULL);
7301                 if (ret) {
7302                         ret = -EAGAIN;
7303                         goto out;
7304                 }
7305         }
7306
7307         btrfs_release_path(path);
7308
7309         /*
7310          * look for other files referencing this extent, if we
7311          * find any we must cow
7312          */
7313
7314         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7315                                     key.offset - backref_offset, disk_bytenr);
7316         if (ret) {
7317                 ret = 0;
7318                 goto out;
7319         }
7320
7321         /*
7322          * adjust disk_bytenr and num_bytes to cover just the bytes
7323          * in this extent we are about to write.  If there
7324          * are any csums in that range we have to cow in order
7325          * to keep the csums correct
7326          */
7327         disk_bytenr += backref_offset;
7328         disk_bytenr += offset - key.offset;
7329         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7330                 goto out;
7331         /*
7332          * all of the above have passed, it is safe to overwrite this extent
7333          * without cow
7334          */
7335         *len = num_bytes;
7336         ret = 1;
7337 out:
7338         btrfs_free_path(path);
7339         return ret;
7340 }
7341
7342 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7343                               struct extent_state **cached_state, int writing)
7344 {
7345         struct btrfs_ordered_extent *ordered;
7346         int ret = 0;
7347
7348         while (1) {
7349                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7350                                  cached_state);
7351                 /*
7352                  * We're concerned with the entire range that we're going to be
7353                  * doing DIO to, so we need to make sure there's no ordered
7354                  * extents in this range.
7355                  */
7356                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7357                                                      lockend - lockstart + 1);
7358
7359                 /*
7360                  * We need to make sure there are no buffered pages in this
7361                  * range either, we could have raced between the invalidate in
7362                  * generic_file_direct_write and locking the extent.  The
7363                  * invalidate needs to happen so that reads after a write do not
7364                  * get stale data.
7365                  */
7366                 if (!ordered &&
7367                     (!writing || !filemap_range_has_page(inode->i_mapping,
7368                                                          lockstart, lockend)))
7369                         break;
7370
7371                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7372                                      cached_state);
7373
7374                 if (ordered) {
7375                         /*
7376                          * If we are doing a DIO read and the ordered extent we
7377                          * found is for a buffered write, we can not wait for it
7378                          * to complete and retry, because if we do so we can
7379                          * deadlock with concurrent buffered writes on page
7380                          * locks. This happens only if our DIO read covers more
7381                          * than one extent map, if at this point has already
7382                          * created an ordered extent for a previous extent map
7383                          * and locked its range in the inode's io tree, and a
7384                          * concurrent write against that previous extent map's
7385                          * range and this range started (we unlock the ranges
7386                          * in the io tree only when the bios complete and
7387                          * buffered writes always lock pages before attempting
7388                          * to lock range in the io tree).
7389                          */
7390                         if (writing ||
7391                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7392                                 btrfs_start_ordered_extent(inode, ordered, 1);
7393                         else
7394                                 ret = -ENOTBLK;
7395                         btrfs_put_ordered_extent(ordered);
7396                 } else {
7397                         /*
7398                          * We could trigger writeback for this range (and wait
7399                          * for it to complete) and then invalidate the pages for
7400                          * this range (through invalidate_inode_pages2_range()),
7401                          * but that can lead us to a deadlock with a concurrent
7402                          * call to readpages() (a buffered read or a defrag call
7403                          * triggered a readahead) on a page lock due to an
7404                          * ordered dio extent we created before but did not have
7405                          * yet a corresponding bio submitted (whence it can not
7406                          * complete), which makes readpages() wait for that
7407                          * ordered extent to complete while holding a lock on
7408                          * that page.
7409                          */
7410                         ret = -ENOTBLK;
7411                 }
7412
7413                 if (ret)
7414                         break;
7415
7416                 cond_resched();
7417         }
7418
7419         return ret;
7420 }
7421
7422 /* The callers of this must take lock_extent() */
7423 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7424                                        u64 orig_start, u64 block_start,
7425                                        u64 block_len, u64 orig_block_len,
7426                                        u64 ram_bytes, int compress_type,
7427                                        int type)
7428 {
7429         struct extent_map_tree *em_tree;
7430         struct extent_map *em;
7431         struct btrfs_root *root = BTRFS_I(inode)->root;
7432         int ret;
7433
7434         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7435                type == BTRFS_ORDERED_COMPRESSED ||
7436                type == BTRFS_ORDERED_NOCOW ||
7437                type == BTRFS_ORDERED_REGULAR);
7438
7439         em_tree = &BTRFS_I(inode)->extent_tree;
7440         em = alloc_extent_map();
7441         if (!em)
7442                 return ERR_PTR(-ENOMEM);
7443
7444         em->start = start;
7445         em->orig_start = orig_start;
7446         em->len = len;
7447         em->block_len = block_len;
7448         em->block_start = block_start;
7449         em->bdev = root->fs_info->fs_devices->latest_bdev;
7450         em->orig_block_len = orig_block_len;
7451         em->ram_bytes = ram_bytes;
7452         em->generation = -1;
7453         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7454         if (type == BTRFS_ORDERED_PREALLOC) {
7455                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7456         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7457                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7458                 em->compress_type = compress_type;
7459         }
7460
7461         do {
7462                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7463                                 em->start + em->len - 1, 0);
7464                 write_lock(&em_tree->lock);
7465                 ret = add_extent_mapping(em_tree, em, 1);
7466                 write_unlock(&em_tree->lock);
7467                 /*
7468                  * The caller has taken lock_extent(), who could race with us
7469                  * to add em?
7470                  */
7471         } while (ret == -EEXIST);
7472
7473         if (ret) {
7474                 free_extent_map(em);
7475                 return ERR_PTR(ret);
7476         }
7477
7478         /* em got 2 refs now, callers needs to do free_extent_map once. */
7479         return em;
7480 }
7481
7482
7483 static int btrfs_get_blocks_direct_read(struct extent_map *em,
7484                                         struct buffer_head *bh_result,
7485                                         struct inode *inode,
7486                                         u64 start, u64 len)
7487 {
7488         if (em->block_start == EXTENT_MAP_HOLE ||
7489                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7490                 return -ENOENT;
7491
7492         len = min(len, em->len - (start - em->start));
7493
7494         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7495                 inode->i_blkbits;
7496         bh_result->b_size = len;
7497         bh_result->b_bdev = em->bdev;
7498         set_buffer_mapped(bh_result);
7499
7500         return 0;
7501 }
7502
7503 static int btrfs_get_blocks_direct_write(struct extent_map **map,
7504                                          struct buffer_head *bh_result,
7505                                          struct inode *inode,
7506                                          struct btrfs_dio_data *dio_data,
7507                                          u64 start, u64 len)
7508 {
7509         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7510         struct extent_map *em = *map;
7511         int ret = 0;
7512
7513         /*
7514          * We don't allocate a new extent in the following cases
7515          *
7516          * 1) The inode is marked as NODATACOW. In this case we'll just use the
7517          * existing extent.
7518          * 2) The extent is marked as PREALLOC. We're good to go here and can
7519          * just use the extent.
7520          *
7521          */
7522         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7523             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7524              em->block_start != EXTENT_MAP_HOLE)) {
7525                 int type;
7526                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7527
7528                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7529                         type = BTRFS_ORDERED_PREALLOC;
7530                 else
7531                         type = BTRFS_ORDERED_NOCOW;
7532                 len = min(len, em->len - (start - em->start));
7533                 block_start = em->block_start + (start - em->start);
7534
7535                 if (can_nocow_extent(inode, start, &len, &orig_start,
7536                                      &orig_block_len, &ram_bytes) == 1 &&
7537                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7538                         struct extent_map *em2;
7539
7540                         em2 = btrfs_create_dio_extent(inode, start, len,
7541                                                       orig_start, block_start,
7542                                                       len, orig_block_len,
7543                                                       ram_bytes, type);
7544                         btrfs_dec_nocow_writers(fs_info, block_start);
7545                         if (type == BTRFS_ORDERED_PREALLOC) {
7546                                 free_extent_map(em);
7547                                 *map = em = em2;
7548                         }
7549
7550                         if (em2 && IS_ERR(em2)) {
7551                                 ret = PTR_ERR(em2);
7552                                 goto out;
7553                         }
7554                         /*
7555                          * For inode marked NODATACOW or extent marked PREALLOC,
7556                          * use the existing or preallocated extent, so does not
7557                          * need to adjust btrfs_space_info's bytes_may_use.
7558                          */
7559                         btrfs_free_reserved_data_space_noquota(inode, start,
7560                                                                len);
7561                         goto skip_cow;
7562                 }
7563         }
7564
7565         /* this will cow the extent */
7566         len = bh_result->b_size;
7567         free_extent_map(em);
7568         *map = em = btrfs_new_extent_direct(inode, start, len);
7569         if (IS_ERR(em)) {
7570                 ret = PTR_ERR(em);
7571                 goto out;
7572         }
7573
7574         len = min(len, em->len - (start - em->start));
7575
7576 skip_cow:
7577         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7578                 inode->i_blkbits;
7579         bh_result->b_size = len;
7580         bh_result->b_bdev = em->bdev;
7581         set_buffer_mapped(bh_result);
7582
7583         if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7584                 set_buffer_new(bh_result);
7585
7586         /*
7587          * Need to update the i_size under the extent lock so buffered
7588          * readers will get the updated i_size when we unlock.
7589          */
7590         if (!dio_data->overwrite && start + len > i_size_read(inode))
7591                 i_size_write(inode, start + len);
7592
7593         WARN_ON(dio_data->reserve < len);
7594         dio_data->reserve -= len;
7595         dio_data->unsubmitted_oe_range_end = start + len;
7596         current->journal_info = dio_data;
7597 out:
7598         return ret;
7599 }
7600
7601 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7602                                    struct buffer_head *bh_result, int create)
7603 {
7604         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7605         struct extent_map *em;
7606         struct extent_state *cached_state = NULL;
7607         struct btrfs_dio_data *dio_data = NULL;
7608         u64 start = iblock << inode->i_blkbits;
7609         u64 lockstart, lockend;
7610         u64 len = bh_result->b_size;
7611         int unlock_bits = EXTENT_LOCKED;
7612         int ret = 0;
7613
7614         if (create)
7615                 unlock_bits |= EXTENT_DIRTY;
7616         else
7617                 len = min_t(u64, len, fs_info->sectorsize);
7618
7619         lockstart = start;
7620         lockend = start + len - 1;
7621
7622         if (current->journal_info) {
7623                 /*
7624                  * Need to pull our outstanding extents and set journal_info to NULL so
7625                  * that anything that needs to check if there's a transaction doesn't get
7626                  * confused.
7627                  */
7628                 dio_data = current->journal_info;
7629                 current->journal_info = NULL;
7630         }
7631
7632         /*
7633          * If this errors out it's because we couldn't invalidate pagecache for
7634          * this range and we need to fallback to buffered.
7635          */
7636         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7637                                create)) {
7638                 ret = -ENOTBLK;
7639                 goto err;
7640         }
7641
7642         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
7643         if (IS_ERR(em)) {
7644                 ret = PTR_ERR(em);
7645                 goto unlock_err;
7646         }
7647
7648         /*
7649          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7650          * io.  INLINE is special, and we could probably kludge it in here, but
7651          * it's still buffered so for safety lets just fall back to the generic
7652          * buffered path.
7653          *
7654          * For COMPRESSED we _have_ to read the entire extent in so we can
7655          * decompress it, so there will be buffering required no matter what we
7656          * do, so go ahead and fallback to buffered.
7657          *
7658          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7659          * to buffered IO.  Don't blame me, this is the price we pay for using
7660          * the generic code.
7661          */
7662         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7663             em->block_start == EXTENT_MAP_INLINE) {
7664                 free_extent_map(em);
7665                 ret = -ENOTBLK;
7666                 goto unlock_err;
7667         }
7668
7669         if (create) {
7670                 ret = btrfs_get_blocks_direct_write(&em, bh_result, inode,
7671                                                     dio_data, start, len);
7672                 if (ret < 0)
7673                         goto unlock_err;
7674
7675                 /* clear and unlock the entire range */
7676                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7677                                  unlock_bits, 1, 0, &cached_state);
7678         } else {
7679                 ret = btrfs_get_blocks_direct_read(em, bh_result, inode,
7680                                                    start, len);
7681                 /* Can be negative only if we read from a hole */
7682                 if (ret < 0) {
7683                         ret = 0;
7684                         free_extent_map(em);
7685                         goto unlock_err;
7686                 }
7687                 /*
7688                  * We need to unlock only the end area that we aren't using.
7689                  * The rest is going to be unlocked by the endio routine.
7690                  */
7691                 lockstart = start + bh_result->b_size;
7692                 if (lockstart < lockend) {
7693                         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7694                                          lockend, unlock_bits, 1, 0,
7695                                          &cached_state);
7696                 } else {
7697                         free_extent_state(cached_state);
7698                 }
7699         }
7700
7701         free_extent_map(em);
7702
7703         return 0;
7704
7705 unlock_err:
7706         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7707                          unlock_bits, 1, 0, &cached_state);
7708 err:
7709         if (dio_data)
7710                 current->journal_info = dio_data;
7711         return ret;
7712 }
7713
7714 static inline blk_status_t submit_dio_repair_bio(struct inode *inode,
7715                                                  struct bio *bio,
7716                                                  int mirror_num)
7717 {
7718         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7719         blk_status_t ret;
7720
7721         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7722
7723         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
7724         if (ret)
7725                 return ret;
7726
7727         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
7728
7729         return ret;
7730 }
7731
7732 static int btrfs_check_dio_repairable(struct inode *inode,
7733                                       struct bio *failed_bio,
7734                                       struct io_failure_record *failrec,
7735                                       int failed_mirror)
7736 {
7737         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7738         int num_copies;
7739
7740         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
7741         if (num_copies == 1) {
7742                 /*
7743                  * we only have a single copy of the data, so don't bother with
7744                  * all the retry and error correction code that follows. no
7745                  * matter what the error is, it is very likely to persist.
7746                  */
7747                 btrfs_debug(fs_info,
7748                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
7749                         num_copies, failrec->this_mirror, failed_mirror);
7750                 return 0;
7751         }
7752
7753         failrec->failed_mirror = failed_mirror;
7754         failrec->this_mirror++;
7755         if (failrec->this_mirror == failed_mirror)
7756                 failrec->this_mirror++;
7757
7758         if (failrec->this_mirror > num_copies) {
7759                 btrfs_debug(fs_info,
7760                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
7761                         num_copies, failrec->this_mirror, failed_mirror);
7762                 return 0;
7763         }
7764
7765         return 1;
7766 }
7767
7768 static blk_status_t dio_read_error(struct inode *inode, struct bio *failed_bio,
7769                                    struct page *page, unsigned int pgoff,
7770                                    u64 start, u64 end, int failed_mirror,
7771                                    bio_end_io_t *repair_endio, void *repair_arg)
7772 {
7773         struct io_failure_record *failrec;
7774         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7775         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
7776         struct bio *bio;
7777         int isector;
7778         unsigned int read_mode = 0;
7779         int segs;
7780         int ret;
7781         blk_status_t status;
7782         struct bio_vec bvec;
7783
7784         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
7785
7786         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7787         if (ret)
7788                 return errno_to_blk_status(ret);
7789
7790         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7791                                          failed_mirror);
7792         if (!ret) {
7793                 free_io_failure(failure_tree, io_tree, failrec);
7794                 return BLK_STS_IOERR;
7795         }
7796
7797         segs = bio_segments(failed_bio);
7798         bio_get_first_bvec(failed_bio, &bvec);
7799         if (segs > 1 ||
7800             (bvec.bv_len > btrfs_inode_sectorsize(inode)))
7801                 read_mode |= REQ_FAILFAST_DEV;
7802
7803         isector = start - btrfs_io_bio(failed_bio)->logical;
7804         isector >>= inode->i_sb->s_blocksize_bits;
7805         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7806                                 pgoff, isector, repair_endio, repair_arg);
7807         bio->bi_opf = REQ_OP_READ | read_mode;
7808
7809         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7810                     "repair DIO read error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d",
7811                     read_mode, failrec->this_mirror, failrec->in_validation);
7812
7813         status = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
7814         if (status) {
7815                 free_io_failure(failure_tree, io_tree, failrec);
7816                 bio_put(bio);
7817         }
7818
7819         return status;
7820 }
7821
7822 struct btrfs_retry_complete {
7823         struct completion done;
7824         struct inode *inode;
7825         u64 start;
7826         int uptodate;
7827 };
7828
7829 static void btrfs_retry_endio_nocsum(struct bio *bio)
7830 {
7831         struct btrfs_retry_complete *done = bio->bi_private;
7832         struct inode *inode = done->inode;
7833         struct bio_vec *bvec;
7834         struct extent_io_tree *io_tree, *failure_tree;
7835         int i;
7836
7837         if (bio->bi_status)
7838                 goto end;
7839
7840         ASSERT(bio->bi_vcnt == 1);
7841         io_tree = &BTRFS_I(inode)->io_tree;
7842         failure_tree = &BTRFS_I(inode)->io_failure_tree;
7843         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(inode));
7844
7845         done->uptodate = 1;
7846         ASSERT(!bio_flagged(bio, BIO_CLONED));
7847         bio_for_each_segment_all(bvec, bio, i)
7848                 clean_io_failure(BTRFS_I(inode)->root->fs_info, failure_tree,
7849                                  io_tree, done->start, bvec->bv_page,
7850                                  btrfs_ino(BTRFS_I(inode)), 0);
7851 end:
7852         complete(&done->done);
7853         bio_put(bio);
7854 }
7855
7856 static blk_status_t __btrfs_correct_data_nocsum(struct inode *inode,
7857                                                 struct btrfs_io_bio *io_bio)
7858 {
7859         struct btrfs_fs_info *fs_info;
7860         struct bio_vec bvec;
7861         struct bvec_iter iter;
7862         struct btrfs_retry_complete done;
7863         u64 start;
7864         unsigned int pgoff;
7865         u32 sectorsize;
7866         int nr_sectors;
7867         blk_status_t ret;
7868         blk_status_t err = BLK_STS_OK;
7869
7870         fs_info = BTRFS_I(inode)->root->fs_info;
7871         sectorsize = fs_info->sectorsize;
7872
7873         start = io_bio->logical;
7874         done.inode = inode;
7875         io_bio->bio.bi_iter = io_bio->iter;
7876
7877         bio_for_each_segment(bvec, &io_bio->bio, iter) {
7878                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
7879                 pgoff = bvec.bv_offset;
7880
7881 next_block_or_try_again:
7882                 done.uptodate = 0;
7883                 done.start = start;
7884                 init_completion(&done.done);
7885
7886                 ret = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
7887                                 pgoff, start, start + sectorsize - 1,
7888                                 io_bio->mirror_num,
7889                                 btrfs_retry_endio_nocsum, &done);
7890                 if (ret) {
7891                         err = ret;
7892                         goto next;
7893                 }
7894
7895                 wait_for_completion_io(&done.done);
7896
7897                 if (!done.uptodate) {
7898                         /* We might have another mirror, so try again */
7899                         goto next_block_or_try_again;
7900                 }
7901
7902 next:
7903                 start += sectorsize;
7904
7905                 nr_sectors--;
7906                 if (nr_sectors) {
7907                         pgoff += sectorsize;
7908                         ASSERT(pgoff < PAGE_SIZE);
7909                         goto next_block_or_try_again;
7910                 }
7911         }
7912
7913         return err;
7914 }
7915
7916 static void btrfs_retry_endio(struct bio *bio)
7917 {
7918         struct btrfs_retry_complete *done = bio->bi_private;
7919         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7920         struct extent_io_tree *io_tree, *failure_tree;
7921         struct inode *inode = done->inode;
7922         struct bio_vec *bvec;
7923         int uptodate;
7924         int ret;
7925         int i;
7926
7927         if (bio->bi_status)
7928                 goto end;
7929
7930         uptodate = 1;
7931
7932         ASSERT(bio->bi_vcnt == 1);
7933         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(done->inode));
7934
7935         io_tree = &BTRFS_I(inode)->io_tree;
7936         failure_tree = &BTRFS_I(inode)->io_failure_tree;
7937
7938         ASSERT(!bio_flagged(bio, BIO_CLONED));
7939         bio_for_each_segment_all(bvec, bio, i) {
7940                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
7941                                              bvec->bv_offset, done->start,
7942                                              bvec->bv_len);
7943                 if (!ret)
7944                         clean_io_failure(BTRFS_I(inode)->root->fs_info,
7945                                          failure_tree, io_tree, done->start,
7946                                          bvec->bv_page,
7947                                          btrfs_ino(BTRFS_I(inode)),
7948                                          bvec->bv_offset);
7949                 else
7950                         uptodate = 0;
7951         }
7952
7953         done->uptodate = uptodate;
7954 end:
7955         complete(&done->done);
7956         bio_put(bio);
7957 }
7958
7959 static blk_status_t __btrfs_subio_endio_read(struct inode *inode,
7960                 struct btrfs_io_bio *io_bio, blk_status_t err)
7961 {
7962         struct btrfs_fs_info *fs_info;
7963         struct bio_vec bvec;
7964         struct bvec_iter iter;
7965         struct btrfs_retry_complete done;
7966         u64 start;
7967         u64 offset = 0;
7968         u32 sectorsize;
7969         int nr_sectors;
7970         unsigned int pgoff;
7971         int csum_pos;
7972         bool uptodate = (err == 0);
7973         int ret;
7974         blk_status_t status;
7975
7976         fs_info = BTRFS_I(inode)->root->fs_info;
7977         sectorsize = fs_info->sectorsize;
7978
7979         err = BLK_STS_OK;
7980         start = io_bio->logical;
7981         done.inode = inode;
7982         io_bio->bio.bi_iter = io_bio->iter;
7983
7984         bio_for_each_segment(bvec, &io_bio->bio, iter) {
7985                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
7986
7987                 pgoff = bvec.bv_offset;
7988 next_block:
7989                 if (uptodate) {
7990                         csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
7991                         ret = __readpage_endio_check(inode, io_bio, csum_pos,
7992                                         bvec.bv_page, pgoff, start, sectorsize);
7993                         if (likely(!ret))
7994                                 goto next;
7995                 }
7996 try_again:
7997                 done.uptodate = 0;
7998                 done.start = start;
7999                 init_completion(&done.done);
8000
8001                 status = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8002                                         pgoff, start, start + sectorsize - 1,
8003                                         io_bio->mirror_num, btrfs_retry_endio,
8004                                         &done);
8005                 if (status) {
8006                         err = status;
8007                         goto next;
8008                 }
8009
8010                 wait_for_completion_io(&done.done);
8011
8012                 if (!done.uptodate) {
8013                         /* We might have another mirror, so try again */
8014                         goto try_again;
8015                 }
8016 next:
8017                 offset += sectorsize;
8018                 start += sectorsize;
8019
8020                 ASSERT(nr_sectors);
8021
8022                 nr_sectors--;
8023                 if (nr_sectors) {
8024                         pgoff += sectorsize;
8025                         ASSERT(pgoff < PAGE_SIZE);
8026                         goto next_block;
8027                 }
8028         }
8029
8030         return err;
8031 }
8032
8033 static blk_status_t btrfs_subio_endio_read(struct inode *inode,
8034                 struct btrfs_io_bio *io_bio, blk_status_t err)
8035 {
8036         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8037
8038         if (skip_csum) {
8039                 if (unlikely(err))
8040                         return __btrfs_correct_data_nocsum(inode, io_bio);
8041                 else
8042                         return BLK_STS_OK;
8043         } else {
8044                 return __btrfs_subio_endio_read(inode, io_bio, err);
8045         }
8046 }
8047
8048 static void btrfs_endio_direct_read(struct bio *bio)
8049 {
8050         struct btrfs_dio_private *dip = bio->bi_private;
8051         struct inode *inode = dip->inode;
8052         struct bio *dio_bio;
8053         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8054         blk_status_t err = bio->bi_status;
8055
8056         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8057                 err = btrfs_subio_endio_read(inode, io_bio, err);
8058
8059         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8060                       dip->logical_offset + dip->bytes - 1);
8061         dio_bio = dip->dio_bio;
8062
8063         kfree(dip);
8064
8065         dio_bio->bi_status = err;
8066         dio_end_io(dio_bio);
8067
8068         if (io_bio->end_io)
8069                 io_bio->end_io(io_bio, blk_status_to_errno(err));
8070         bio_put(bio);
8071 }
8072
8073 static void __endio_write_update_ordered(struct inode *inode,
8074                                          const u64 offset, const u64 bytes,
8075                                          const bool uptodate)
8076 {
8077         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8078         struct btrfs_ordered_extent *ordered = NULL;
8079         struct btrfs_workqueue *wq;
8080         btrfs_work_func_t func;
8081         u64 ordered_offset = offset;
8082         u64 ordered_bytes = bytes;
8083         u64 last_offset;
8084
8085         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
8086                 wq = fs_info->endio_freespace_worker;
8087                 func = btrfs_freespace_write_helper;
8088         } else {
8089                 wq = fs_info->endio_write_workers;
8090                 func = btrfs_endio_write_helper;
8091         }
8092
8093         while (ordered_offset < offset + bytes) {
8094                 last_offset = ordered_offset;
8095                 if (btrfs_dec_test_first_ordered_pending(inode, &ordered,
8096                                                            &ordered_offset,
8097                                                            ordered_bytes,
8098                                                            uptodate)) {
8099                         btrfs_init_work(&ordered->work, func,
8100                                         finish_ordered_fn,
8101                                         NULL, NULL);
8102                         btrfs_queue_work(wq, &ordered->work);
8103                 }
8104                 /*
8105                  * If btrfs_dec_test_ordered_pending does not find any ordered
8106                  * extent in the range, we can exit.
8107                  */
8108                 if (ordered_offset == last_offset)
8109                         return;
8110                 /*
8111                  * Our bio might span multiple ordered extents. In this case
8112                  * we keep goin until we have accounted the whole dio.
8113                  */
8114                 if (ordered_offset < offset + bytes) {
8115                         ordered_bytes = offset + bytes - ordered_offset;
8116                         ordered = NULL;
8117                 }
8118         }
8119 }
8120
8121 static void btrfs_endio_direct_write(struct bio *bio)
8122 {
8123         struct btrfs_dio_private *dip = bio->bi_private;
8124         struct bio *dio_bio = dip->dio_bio;
8125
8126         __endio_write_update_ordered(dip->inode, dip->logical_offset,
8127                                      dip->bytes, !bio->bi_status);
8128
8129         kfree(dip);
8130
8131         dio_bio->bi_status = bio->bi_status;
8132         dio_end_io(dio_bio);
8133         bio_put(bio);
8134 }
8135
8136 static blk_status_t btrfs_submit_bio_start_direct_io(void *private_data,
8137                                     struct bio *bio, u64 offset)
8138 {
8139         struct inode *inode = private_data;
8140         blk_status_t ret;
8141         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
8142         BUG_ON(ret); /* -ENOMEM */
8143         return 0;
8144 }
8145
8146 static void btrfs_end_dio_bio(struct bio *bio)
8147 {
8148         struct btrfs_dio_private *dip = bio->bi_private;
8149         blk_status_t err = bio->bi_status;
8150
8151         if (err)
8152                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8153                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8154                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
8155                            bio->bi_opf,
8156                            (unsigned long long)bio->bi_iter.bi_sector,
8157                            bio->bi_iter.bi_size, err);
8158
8159         if (dip->subio_endio)
8160                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8161
8162         if (err) {
8163                 /*
8164                  * We want to perceive the errors flag being set before
8165                  * decrementing the reference count. We don't need a barrier
8166                  * since atomic operations with a return value are fully
8167                  * ordered as per atomic_t.txt
8168                  */
8169                 dip->errors = 1;
8170         }
8171
8172         /* if there are more bios still pending for this dio, just exit */
8173         if (!atomic_dec_and_test(&dip->pending_bios))
8174                 goto out;
8175
8176         if (dip->errors) {
8177                 bio_io_error(dip->orig_bio);
8178         } else {
8179                 dip->dio_bio->bi_status = BLK_STS_OK;
8180                 bio_endio(dip->orig_bio);
8181         }
8182 out:
8183         bio_put(bio);
8184 }
8185
8186 static inline blk_status_t btrfs_lookup_and_bind_dio_csum(struct inode *inode,
8187                                                  struct btrfs_dio_private *dip,
8188                                                  struct bio *bio,
8189                                                  u64 file_offset)
8190 {
8191         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8192         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8193         blk_status_t ret;
8194
8195         /*
8196          * We load all the csum data we need when we submit
8197          * the first bio to reduce the csum tree search and
8198          * contention.
8199          */
8200         if (dip->logical_offset == file_offset) {
8201                 ret = btrfs_lookup_bio_sums_dio(inode, dip->orig_bio,
8202                                                 file_offset);
8203                 if (ret)
8204                         return ret;
8205         }
8206
8207         if (bio == dip->orig_bio)
8208                 return 0;
8209
8210         file_offset -= dip->logical_offset;
8211         file_offset >>= inode->i_sb->s_blocksize_bits;
8212         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8213
8214         return 0;
8215 }
8216
8217 static inline blk_status_t btrfs_submit_dio_bio(struct bio *bio,
8218                 struct inode *inode, u64 file_offset, int async_submit)
8219 {
8220         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8221         struct btrfs_dio_private *dip = bio->bi_private;
8222         bool write = bio_op(bio) == REQ_OP_WRITE;
8223         blk_status_t ret;
8224
8225         /* Check btrfs_submit_bio_hook() for rules about async submit. */
8226         if (async_submit)
8227                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8228
8229         if (!write) {
8230                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
8231                 if (ret)
8232                         goto err;
8233         }
8234
8235         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
8236                 goto map;
8237
8238         if (write && async_submit) {
8239                 ret = btrfs_wq_submit_bio(fs_info, bio, 0, 0,
8240                                           file_offset, inode,
8241                                           btrfs_submit_bio_start_direct_io);
8242                 goto err;
8243         } else if (write) {
8244                 /*
8245                  * If we aren't doing async submit, calculate the csum of the
8246                  * bio now.
8247                  */
8248                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
8249                 if (ret)
8250                         goto err;
8251         } else {
8252                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
8253                                                      file_offset);
8254                 if (ret)
8255                         goto err;
8256         }
8257 map:
8258         ret = btrfs_map_bio(fs_info, bio, 0, 0);
8259 err:
8260         return ret;
8261 }
8262
8263 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip)
8264 {
8265         struct inode *inode = dip->inode;
8266         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8267         struct bio *bio;
8268         struct bio *orig_bio = dip->orig_bio;
8269         u64 start_sector = orig_bio->bi_iter.bi_sector;
8270         u64 file_offset = dip->logical_offset;
8271         u64 map_length;
8272         int async_submit = 0;
8273         u64 submit_len;
8274         int clone_offset = 0;
8275         int clone_len;
8276         int ret;
8277         blk_status_t status;
8278
8279         map_length = orig_bio->bi_iter.bi_size;
8280         submit_len = map_length;
8281         ret = btrfs_map_block(fs_info, btrfs_op(orig_bio), start_sector << 9,
8282                               &map_length, NULL, 0);
8283         if (ret)
8284                 return -EIO;
8285
8286         if (map_length >= submit_len) {
8287                 bio = orig_bio;
8288                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8289                 goto submit;
8290         }
8291
8292         /* async crcs make it difficult to collect full stripe writes. */
8293         if (btrfs_data_alloc_profile(fs_info) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8294                 async_submit = 0;
8295         else
8296                 async_submit = 1;
8297
8298         /* bio split */
8299         ASSERT(map_length <= INT_MAX);
8300         atomic_inc(&dip->pending_bios);
8301         do {
8302                 clone_len = min_t(int, submit_len, map_length);
8303
8304                 /*
8305                  * This will never fail as it's passing GPF_NOFS and
8306                  * the allocation is backed by btrfs_bioset.
8307                  */
8308                 bio = btrfs_bio_clone_partial(orig_bio, clone_offset,
8309                                               clone_len);
8310                 bio->bi_private = dip;
8311                 bio->bi_end_io = btrfs_end_dio_bio;
8312                 btrfs_io_bio(bio)->logical = file_offset;
8313
8314                 ASSERT(submit_len >= clone_len);
8315                 submit_len -= clone_len;
8316                 if (submit_len == 0)
8317                         break;
8318
8319                 /*
8320                  * Increase the count before we submit the bio so we know
8321                  * the end IO handler won't happen before we increase the
8322                  * count. Otherwise, the dip might get freed before we're
8323                  * done setting it up.
8324                  */
8325                 atomic_inc(&dip->pending_bios);
8326
8327                 status = btrfs_submit_dio_bio(bio, inode, file_offset,
8328                                                 async_submit);
8329                 if (status) {
8330                         bio_put(bio);
8331                         atomic_dec(&dip->pending_bios);
8332                         goto out_err;
8333                 }
8334
8335                 clone_offset += clone_len;
8336                 start_sector += clone_len >> 9;
8337                 file_offset += clone_len;
8338
8339                 map_length = submit_len;
8340                 ret = btrfs_map_block(fs_info, btrfs_op(orig_bio),
8341                                       start_sector << 9, &map_length, NULL, 0);
8342                 if (ret)
8343                         goto out_err;
8344         } while (submit_len > 0);
8345
8346 submit:
8347         status = btrfs_submit_dio_bio(bio, inode, file_offset, async_submit);
8348         if (!status)
8349                 return 0;
8350
8351         bio_put(bio);
8352 out_err:
8353         dip->errors = 1;
8354         /*
8355          * Before atomic variable goto zero, we must  make sure dip->errors is
8356          * perceived to be set. This ordering is ensured by the fact that an
8357          * atomic operations with a return value are fully ordered as per
8358          * atomic_t.txt
8359          */
8360         if (atomic_dec_and_test(&dip->pending_bios))
8361                 bio_io_error(dip->orig_bio);
8362
8363         /* bio_end_io() will handle error, so we needn't return it */
8364         return 0;
8365 }
8366
8367 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8368                                 loff_t file_offset)
8369 {
8370         struct btrfs_dio_private *dip = NULL;
8371         struct bio *bio = NULL;
8372         struct btrfs_io_bio *io_bio;
8373         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8374         int ret = 0;
8375
8376         bio = btrfs_bio_clone(dio_bio);
8377
8378         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8379         if (!dip) {
8380                 ret = -ENOMEM;
8381                 goto free_ordered;
8382         }
8383
8384         dip->private = dio_bio->bi_private;
8385         dip->inode = inode;
8386         dip->logical_offset = file_offset;
8387         dip->bytes = dio_bio->bi_iter.bi_size;
8388         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8389         bio->bi_private = dip;
8390         dip->orig_bio = bio;
8391         dip->dio_bio = dio_bio;
8392         atomic_set(&dip->pending_bios, 0);
8393         io_bio = btrfs_io_bio(bio);
8394         io_bio->logical = file_offset;
8395
8396         if (write) {
8397                 bio->bi_end_io = btrfs_endio_direct_write;
8398         } else {
8399                 bio->bi_end_io = btrfs_endio_direct_read;
8400                 dip->subio_endio = btrfs_subio_endio_read;
8401         }
8402
8403         /*
8404          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8405          * even if we fail to submit a bio, because in such case we do the
8406          * corresponding error handling below and it must not be done a second
8407          * time by btrfs_direct_IO().
8408          */
8409         if (write) {
8410                 struct btrfs_dio_data *dio_data = current->journal_info;
8411
8412                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8413                         dip->bytes;
8414                 dio_data->unsubmitted_oe_range_start =
8415                         dio_data->unsubmitted_oe_range_end;
8416         }
8417
8418         ret = btrfs_submit_direct_hook(dip);
8419         if (!ret)
8420                 return;
8421
8422         if (io_bio->end_io)
8423                 io_bio->end_io(io_bio, ret);
8424
8425 free_ordered:
8426         /*
8427          * If we arrived here it means either we failed to submit the dip
8428          * or we either failed to clone the dio_bio or failed to allocate the
8429          * dip. If we cloned the dio_bio and allocated the dip, we can just
8430          * call bio_endio against our io_bio so that we get proper resource
8431          * cleanup if we fail to submit the dip, otherwise, we must do the
8432          * same as btrfs_endio_direct_[write|read] because we can't call these
8433          * callbacks - they require an allocated dip and a clone of dio_bio.
8434          */
8435         if (bio && dip) {
8436                 bio_io_error(bio);
8437                 /*
8438                  * The end io callbacks free our dip, do the final put on bio
8439                  * and all the cleanup and final put for dio_bio (through
8440                  * dio_end_io()).
8441                  */
8442                 dip = NULL;
8443                 bio = NULL;
8444         } else {
8445                 if (write)
8446                         __endio_write_update_ordered(inode,
8447                                                 file_offset,
8448                                                 dio_bio->bi_iter.bi_size,
8449                                                 false);
8450                 else
8451                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8452                               file_offset + dio_bio->bi_iter.bi_size - 1);
8453
8454                 dio_bio->bi_status = BLK_STS_IOERR;
8455                 /*
8456                  * Releases and cleans up our dio_bio, no need to bio_put()
8457                  * nor bio_endio()/bio_io_error() against dio_bio.
8458                  */
8459                 dio_end_io(dio_bio);
8460         }
8461         if (bio)
8462                 bio_put(bio);
8463         kfree(dip);
8464 }
8465
8466 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8467                                const struct iov_iter *iter, loff_t offset)
8468 {
8469         int seg;
8470         int i;
8471         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8472         ssize_t retval = -EINVAL;
8473
8474         if (offset & blocksize_mask)
8475                 goto out;
8476
8477         if (iov_iter_alignment(iter) & blocksize_mask)
8478                 goto out;
8479
8480         /* If this is a write we don't need to check anymore */
8481         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8482                 return 0;
8483         /*
8484          * Check to make sure we don't have duplicate iov_base's in this
8485          * iovec, if so return EINVAL, otherwise we'll get csum errors
8486          * when reading back.
8487          */
8488         for (seg = 0; seg < iter->nr_segs; seg++) {
8489                 for (i = seg + 1; i < iter->nr_segs; i++) {
8490                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8491                                 goto out;
8492                 }
8493         }
8494         retval = 0;
8495 out:
8496         return retval;
8497 }
8498
8499 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8500 {
8501         struct file *file = iocb->ki_filp;
8502         struct inode *inode = file->f_mapping->host;
8503         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8504         struct btrfs_dio_data dio_data = { 0 };
8505         struct extent_changeset *data_reserved = NULL;
8506         loff_t offset = iocb->ki_pos;
8507         size_t count = 0;
8508         int flags = 0;
8509         bool wakeup = true;
8510         bool relock = false;
8511         ssize_t ret;
8512
8513         if (check_direct_IO(fs_info, iter, offset))
8514                 return 0;
8515
8516         inode_dio_begin(inode);
8517
8518         /*
8519          * The generic stuff only does filemap_write_and_wait_range, which
8520          * isn't enough if we've written compressed pages to this area, so
8521          * we need to flush the dirty pages again to make absolutely sure
8522          * that any outstanding dirty pages are on disk.
8523          */
8524         count = iov_iter_count(iter);
8525         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8526                      &BTRFS_I(inode)->runtime_flags))
8527                 filemap_fdatawrite_range(inode->i_mapping, offset,
8528                                          offset + count - 1);
8529
8530         if (iov_iter_rw(iter) == WRITE) {
8531                 /*
8532                  * If the write DIO is beyond the EOF, we need update
8533                  * the isize, but it is protected by i_mutex. So we can
8534                  * not unlock the i_mutex at this case.
8535                  */
8536                 if (offset + count <= inode->i_size) {
8537                         dio_data.overwrite = 1;
8538                         inode_unlock(inode);
8539                         relock = true;
8540                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
8541                         ret = -EAGAIN;
8542                         goto out;
8543                 }
8544                 ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
8545                                                    offset, count);
8546                 if (ret)
8547                         goto out;
8548
8549                 /*
8550                  * We need to know how many extents we reserved so that we can
8551                  * do the accounting properly if we go over the number we
8552                  * originally calculated.  Abuse current->journal_info for this.
8553                  */
8554                 dio_data.reserve = round_up(count,
8555                                             fs_info->sectorsize);
8556                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8557                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8558                 current->journal_info = &dio_data;
8559                 down_read(&BTRFS_I(inode)->dio_sem);
8560         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8561                                      &BTRFS_I(inode)->runtime_flags)) {
8562                 inode_dio_end(inode);
8563                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8564                 wakeup = false;
8565         }
8566
8567         ret = __blockdev_direct_IO(iocb, inode,
8568                                    fs_info->fs_devices->latest_bdev,
8569                                    iter, btrfs_get_blocks_direct, NULL,
8570                                    btrfs_submit_direct, flags);
8571         if (iov_iter_rw(iter) == WRITE) {
8572                 up_read(&BTRFS_I(inode)->dio_sem);
8573                 current->journal_info = NULL;
8574                 if (ret < 0 && ret != -EIOCBQUEUED) {
8575                         if (dio_data.reserve)
8576                                 btrfs_delalloc_release_space(inode, data_reserved,
8577                                         offset, dio_data.reserve, true);
8578                         /*
8579                          * On error we might have left some ordered extents
8580                          * without submitting corresponding bios for them, so
8581                          * cleanup them up to avoid other tasks getting them
8582                          * and waiting for them to complete forever.
8583                          */
8584                         if (dio_data.unsubmitted_oe_range_start <
8585                             dio_data.unsubmitted_oe_range_end)
8586                                 __endio_write_update_ordered(inode,
8587                                         dio_data.unsubmitted_oe_range_start,
8588                                         dio_data.unsubmitted_oe_range_end -
8589                                         dio_data.unsubmitted_oe_range_start,
8590                                         false);
8591                 } else if (ret >= 0 && (size_t)ret < count)
8592                         btrfs_delalloc_release_space(inode, data_reserved,
8593                                         offset, count - (size_t)ret, true);
8594                 btrfs_delalloc_release_extents(BTRFS_I(inode), count, false);
8595         }
8596 out:
8597         if (wakeup)
8598                 inode_dio_end(inode);
8599         if (relock)
8600                 inode_lock(inode);
8601
8602         extent_changeset_free(data_reserved);
8603         return ret;
8604 }
8605
8606 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8607
8608 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8609                 __u64 start, __u64 len)
8610 {
8611         int     ret;
8612
8613         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8614         if (ret)
8615                 return ret;
8616
8617         return extent_fiemap(inode, fieinfo, start, len);
8618 }
8619
8620 int btrfs_readpage(struct file *file, struct page *page)
8621 {
8622         struct extent_io_tree *tree;
8623         tree = &BTRFS_I(page->mapping->host)->io_tree;
8624         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8625 }
8626
8627 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8628 {
8629         struct inode *inode = page->mapping->host;
8630         int ret;
8631
8632         if (current->flags & PF_MEMALLOC) {
8633                 redirty_page_for_writepage(wbc, page);
8634                 unlock_page(page);
8635                 return 0;
8636         }
8637
8638         /*
8639          * If we are under memory pressure we will call this directly from the
8640          * VM, we need to make sure we have the inode referenced for the ordered
8641          * extent.  If not just return like we didn't do anything.
8642          */
8643         if (!igrab(inode)) {
8644                 redirty_page_for_writepage(wbc, page);
8645                 return AOP_WRITEPAGE_ACTIVATE;
8646         }
8647         ret = extent_write_full_page(page, wbc);
8648         btrfs_add_delayed_iput(inode);
8649         return ret;
8650 }
8651
8652 static int btrfs_writepages(struct address_space *mapping,
8653                             struct writeback_control *wbc)
8654 {
8655         return extent_writepages(mapping, wbc);
8656 }
8657
8658 static int
8659 btrfs_readpages(struct file *file, struct address_space *mapping,
8660                 struct list_head *pages, unsigned nr_pages)
8661 {
8662         return extent_readpages(mapping, pages, nr_pages);
8663 }
8664
8665 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8666 {
8667         int ret = try_release_extent_mapping(page, gfp_flags);
8668         if (ret == 1) {
8669                 ClearPagePrivate(page);
8670                 set_page_private(page, 0);
8671                 put_page(page);
8672         }
8673         return ret;
8674 }
8675
8676 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8677 {
8678         if (PageWriteback(page) || PageDirty(page))
8679                 return 0;
8680         return __btrfs_releasepage(page, gfp_flags);
8681 }
8682
8683 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8684                                  unsigned int length)
8685 {
8686         struct inode *inode = page->mapping->host;
8687         struct extent_io_tree *tree;
8688         struct btrfs_ordered_extent *ordered;
8689         struct extent_state *cached_state = NULL;
8690         u64 page_start = page_offset(page);
8691         u64 page_end = page_start + PAGE_SIZE - 1;
8692         u64 start;
8693         u64 end;
8694         int inode_evicting = inode->i_state & I_FREEING;
8695
8696         /*
8697          * we have the page locked, so new writeback can't start,
8698          * and the dirty bit won't be cleared while we are here.
8699          *
8700          * Wait for IO on this page so that we can safely clear
8701          * the PagePrivate2 bit and do ordered accounting
8702          */
8703         wait_on_page_writeback(page);
8704
8705         tree = &BTRFS_I(inode)->io_tree;
8706         if (offset) {
8707                 btrfs_releasepage(page, GFP_NOFS);
8708                 return;
8709         }
8710
8711         if (!inode_evicting)
8712                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8713 again:
8714         start = page_start;
8715         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
8716                                         page_end - start + 1);
8717         if (ordered) {
8718                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8719                 /*
8720                  * IO on this page will never be started, so we need
8721                  * to account for any ordered extents now
8722                  */
8723                 if (!inode_evicting)
8724                         clear_extent_bit(tree, start, end,
8725                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8726                                          EXTENT_DELALLOC_NEW |
8727                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8728                                          EXTENT_DEFRAG, 1, 0, &cached_state);
8729                 /*
8730                  * whoever cleared the private bit is responsible
8731                  * for the finish_ordered_io
8732                  */
8733                 if (TestClearPagePrivate2(page)) {
8734                         struct btrfs_ordered_inode_tree *tree;
8735                         u64 new_len;
8736
8737                         tree = &BTRFS_I(inode)->ordered_tree;
8738
8739                         spin_lock_irq(&tree->lock);
8740                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8741                         new_len = start - ordered->file_offset;
8742                         if (new_len < ordered->truncated_len)
8743                                 ordered->truncated_len = new_len;
8744                         spin_unlock_irq(&tree->lock);
8745
8746                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8747                                                            start,
8748                                                            end - start + 1, 1))
8749                                 btrfs_finish_ordered_io(ordered);
8750                 }
8751                 btrfs_put_ordered_extent(ordered);
8752                 if (!inode_evicting) {
8753                         cached_state = NULL;
8754                         lock_extent_bits(tree, start, end,
8755                                          &cached_state);
8756                 }
8757
8758                 start = end + 1;
8759                 if (start < page_end)
8760                         goto again;
8761         }
8762
8763         /*
8764          * Qgroup reserved space handler
8765          * Page here will be either
8766          * 1) Already written to disk
8767          *    In this case, its reserved space is released from data rsv map
8768          *    and will be freed by delayed_ref handler finally.
8769          *    So even we call qgroup_free_data(), it won't decrease reserved
8770          *    space.
8771          * 2) Not written to disk
8772          *    This means the reserved space should be freed here. However,
8773          *    if a truncate invalidates the page (by clearing PageDirty)
8774          *    and the page is accounted for while allocating extent
8775          *    in btrfs_check_data_free_space() we let delayed_ref to
8776          *    free the entire extent.
8777          */
8778         if (PageDirty(page))
8779                 btrfs_qgroup_free_data(inode, NULL, page_start, PAGE_SIZE);
8780         if (!inode_evicting) {
8781                 clear_extent_bit(tree, page_start, page_end,
8782                                  EXTENT_LOCKED | EXTENT_DIRTY |
8783                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8784                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
8785                                  &cached_state);
8786
8787                 __btrfs_releasepage(page, GFP_NOFS);
8788         }
8789
8790         ClearPageChecked(page);
8791         if (PagePrivate(page)) {
8792                 ClearPagePrivate(page);
8793                 set_page_private(page, 0);
8794                 put_page(page);
8795         }
8796 }
8797
8798 /*
8799  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8800  * called from a page fault handler when a page is first dirtied. Hence we must
8801  * be careful to check for EOF conditions here. We set the page up correctly
8802  * for a written page which means we get ENOSPC checking when writing into
8803  * holes and correct delalloc and unwritten extent mapping on filesystems that
8804  * support these features.
8805  *
8806  * We are not allowed to take the i_mutex here so we have to play games to
8807  * protect against truncate races as the page could now be beyond EOF.  Because
8808  * truncate_setsize() writes the inode size before removing pages, once we have
8809  * the page lock we can determine safely if the page is beyond EOF. If it is not
8810  * beyond EOF, then the page is guaranteed safe against truncation until we
8811  * unlock the page.
8812  */
8813 vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf)
8814 {
8815         struct page *page = vmf->page;
8816         struct inode *inode = file_inode(vmf->vma->vm_file);
8817         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8818         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8819         struct btrfs_ordered_extent *ordered;
8820         struct extent_state *cached_state = NULL;
8821         struct extent_changeset *data_reserved = NULL;
8822         char *kaddr;
8823         unsigned long zero_start;
8824         loff_t size;
8825         vm_fault_t ret;
8826         int ret2;
8827         int reserved = 0;
8828         u64 reserved_space;
8829         u64 page_start;
8830         u64 page_end;
8831         u64 end;
8832
8833         reserved_space = PAGE_SIZE;
8834
8835         sb_start_pagefault(inode->i_sb);
8836         page_start = page_offset(page);
8837         page_end = page_start + PAGE_SIZE - 1;
8838         end = page_end;
8839
8840         /*
8841          * Reserving delalloc space after obtaining the page lock can lead to
8842          * deadlock. For example, if a dirty page is locked by this function
8843          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8844          * dirty page write out, then the btrfs_writepage() function could
8845          * end up waiting indefinitely to get a lock on the page currently
8846          * being processed by btrfs_page_mkwrite() function.
8847          */
8848         ret2 = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
8849                                            reserved_space);
8850         if (!ret2) {
8851                 ret2 = file_update_time(vmf->vma->vm_file);
8852                 reserved = 1;
8853         }
8854         if (ret2) {
8855                 ret = vmf_error(ret2);
8856                 if (reserved)
8857                         goto out;
8858                 goto out_noreserve;
8859         }
8860
8861         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8862 again:
8863         lock_page(page);
8864         size = i_size_read(inode);
8865
8866         if ((page->mapping != inode->i_mapping) ||
8867             (page_start >= size)) {
8868                 /* page got truncated out from underneath us */
8869                 goto out_unlock;
8870         }
8871         wait_on_page_writeback(page);
8872
8873         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8874         set_page_extent_mapped(page);
8875
8876         /*
8877          * we can't set the delalloc bits if there are pending ordered
8878          * extents.  Drop our locks and wait for them to finish
8879          */
8880         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
8881                         PAGE_SIZE);
8882         if (ordered) {
8883                 unlock_extent_cached(io_tree, page_start, page_end,
8884                                      &cached_state);
8885                 unlock_page(page);
8886                 btrfs_start_ordered_extent(inode, ordered, 1);
8887                 btrfs_put_ordered_extent(ordered);
8888                 goto again;
8889         }
8890
8891         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
8892                 reserved_space = round_up(size - page_start,
8893                                           fs_info->sectorsize);
8894                 if (reserved_space < PAGE_SIZE) {
8895                         end = page_start + reserved_space - 1;
8896                         btrfs_delalloc_release_space(inode, data_reserved,
8897                                         page_start, PAGE_SIZE - reserved_space,
8898                                         true);
8899                 }
8900         }
8901
8902         /*
8903          * page_mkwrite gets called when the page is firstly dirtied after it's
8904          * faulted in, but write(2) could also dirty a page and set delalloc
8905          * bits, thus in this case for space account reason, we still need to
8906          * clear any delalloc bits within this page range since we have to
8907          * reserve data&meta space before lock_page() (see above comments).
8908          */
8909         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
8910                           EXTENT_DIRTY | EXTENT_DELALLOC |
8911                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
8912                           0, 0, &cached_state);
8913
8914         ret2 = btrfs_set_extent_delalloc(inode, page_start, end, 0,
8915                                         &cached_state, 0);
8916         if (ret2) {
8917                 unlock_extent_cached(io_tree, page_start, page_end,
8918                                      &cached_state);
8919                 ret = VM_FAULT_SIGBUS;
8920                 goto out_unlock;
8921         }
8922         ret2 = 0;
8923
8924         /* page is wholly or partially inside EOF */
8925         if (page_start + PAGE_SIZE > size)
8926                 zero_start = size & ~PAGE_MASK;
8927         else
8928                 zero_start = PAGE_SIZE;
8929
8930         if (zero_start != PAGE_SIZE) {
8931                 kaddr = kmap(page);
8932                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
8933                 flush_dcache_page(page);
8934                 kunmap(page);
8935         }
8936         ClearPageChecked(page);
8937         set_page_dirty(page);
8938         SetPageUptodate(page);
8939
8940         BTRFS_I(inode)->last_trans = fs_info->generation;
8941         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
8942         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
8943
8944         unlock_extent_cached(io_tree, page_start, page_end, &cached_state);
8945
8946         if (!ret2) {
8947                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE, true);
8948                 sb_end_pagefault(inode->i_sb);
8949                 extent_changeset_free(data_reserved);
8950                 return VM_FAULT_LOCKED;
8951         }
8952
8953 out_unlock:
8954         unlock_page(page);
8955 out:
8956         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE, (ret != 0));
8957         btrfs_delalloc_release_space(inode, data_reserved, page_start,
8958                                      reserved_space, (ret != 0));
8959 out_noreserve:
8960         sb_end_pagefault(inode->i_sb);
8961         extent_changeset_free(data_reserved);
8962         return ret;
8963 }
8964
8965 static int btrfs_truncate(struct inode *inode, bool skip_writeback)
8966 {
8967         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8968         struct btrfs_root *root = BTRFS_I(inode)->root;
8969         struct btrfs_block_rsv *rsv;
8970         int ret;
8971         struct btrfs_trans_handle *trans;
8972         u64 mask = fs_info->sectorsize - 1;
8973         u64 min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
8974
8975         if (!skip_writeback) {
8976                 ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
8977                                                (u64)-1);
8978                 if (ret)
8979                         return ret;
8980         }
8981
8982         /*
8983          * Yes ladies and gentlemen, this is indeed ugly.  We have a couple of
8984          * things going on here:
8985          *
8986          * 1) We need to reserve space to update our inode.
8987          *
8988          * 2) We need to have something to cache all the space that is going to
8989          * be free'd up by the truncate operation, but also have some slack
8990          * space reserved in case it uses space during the truncate (thank you
8991          * very much snapshotting).
8992          *
8993          * And we need these to be separate.  The fact is we can use a lot of
8994          * space doing the truncate, and we have no earthly idea how much space
8995          * we will use, so we need the truncate reservation to be separate so it
8996          * doesn't end up using space reserved for updating the inode.  We also
8997          * need to be able to stop the transaction and start a new one, which
8998          * means we need to be able to update the inode several times, and we
8999          * have no idea of knowing how many times that will be, so we can't just
9000          * reserve 1 item for the entirety of the operation, so that has to be
9001          * done separately as well.
9002          *
9003          * So that leaves us with
9004          *
9005          * 1) rsv - for the truncate reservation, which we will steal from the
9006          * transaction reservation.
9007          * 2) fs_info->trans_block_rsv - this will have 1 items worth left for
9008          * updating the inode.
9009          */
9010         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
9011         if (!rsv)
9012                 return -ENOMEM;
9013         rsv->size = min_size;
9014         rsv->failfast = 1;
9015
9016         /*
9017          * 1 for the truncate slack space
9018          * 1 for updating the inode.
9019          */
9020         trans = btrfs_start_transaction(root, 2);
9021         if (IS_ERR(trans)) {
9022                 ret = PTR_ERR(trans);
9023                 goto out;
9024         }
9025
9026         /* Migrate the slack space for the truncate to our reserve */
9027         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
9028                                       min_size, 0);
9029         BUG_ON(ret);
9030
9031         /*
9032          * So if we truncate and then write and fsync we normally would just
9033          * write the extents that changed, which is a problem if we need to
9034          * first truncate that entire inode.  So set this flag so we write out
9035          * all of the extents in the inode to the sync log so we're completely
9036          * safe.
9037          */
9038         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9039         trans->block_rsv = rsv;
9040
9041         while (1) {
9042                 ret = btrfs_truncate_inode_items(trans, root, inode,
9043                                                  inode->i_size,
9044                                                  BTRFS_EXTENT_DATA_KEY);
9045                 trans->block_rsv = &fs_info->trans_block_rsv;
9046                 if (ret != -ENOSPC && ret != -EAGAIN)
9047                         break;
9048
9049                 ret = btrfs_update_inode(trans, root, inode);
9050                 if (ret)
9051                         break;
9052
9053                 btrfs_end_transaction(trans);
9054                 btrfs_btree_balance_dirty(fs_info);
9055
9056                 trans = btrfs_start_transaction(root, 2);
9057                 if (IS_ERR(trans)) {
9058                         ret = PTR_ERR(trans);
9059                         trans = NULL;
9060                         break;
9061                 }
9062
9063                 btrfs_block_rsv_release(fs_info, rsv, -1);
9064                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
9065                                               rsv, min_size, 0);
9066                 BUG_ON(ret);    /* shouldn't happen */
9067                 trans->block_rsv = rsv;
9068         }
9069
9070         /*
9071          * We can't call btrfs_truncate_block inside a trans handle as we could
9072          * deadlock with freeze, if we got NEED_TRUNCATE_BLOCK then we know
9073          * we've truncated everything except the last little bit, and can do
9074          * btrfs_truncate_block and then update the disk_i_size.
9075          */
9076         if (ret == NEED_TRUNCATE_BLOCK) {
9077                 btrfs_end_transaction(trans);
9078                 btrfs_btree_balance_dirty(fs_info);
9079
9080                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
9081                 if (ret)
9082                         goto out;
9083                 trans = btrfs_start_transaction(root, 1);
9084                 if (IS_ERR(trans)) {
9085                         ret = PTR_ERR(trans);
9086                         goto out;
9087                 }
9088                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
9089         }
9090
9091         if (trans) {
9092                 int ret2;
9093
9094                 trans->block_rsv = &fs_info->trans_block_rsv;
9095                 ret2 = btrfs_update_inode(trans, root, inode);
9096                 if (ret2 && !ret)
9097                         ret = ret2;
9098
9099                 ret2 = btrfs_end_transaction(trans);
9100                 if (ret2 && !ret)
9101                         ret = ret2;
9102                 btrfs_btree_balance_dirty(fs_info);
9103         }
9104 out:
9105         btrfs_free_block_rsv(fs_info, rsv);
9106
9107         return ret;
9108 }
9109
9110 /*
9111  * create a new subvolume directory/inode (helper for the ioctl).
9112  */
9113 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9114                              struct btrfs_root *new_root,
9115                              struct btrfs_root *parent_root,
9116                              u64 new_dirid)
9117 {
9118         struct inode *inode;
9119         int err;
9120         u64 index = 0;
9121
9122         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9123                                 new_dirid, new_dirid,
9124                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9125                                 &index);
9126         if (IS_ERR(inode))
9127                 return PTR_ERR(inode);
9128         inode->i_op = &btrfs_dir_inode_operations;
9129         inode->i_fop = &btrfs_dir_file_operations;
9130
9131         set_nlink(inode, 1);
9132         btrfs_i_size_write(BTRFS_I(inode), 0);
9133         unlock_new_inode(inode);
9134
9135         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9136         if (err)
9137                 btrfs_err(new_root->fs_info,
9138                           "error inheriting subvolume %llu properties: %d",
9139                           new_root->root_key.objectid, err);
9140
9141         err = btrfs_update_inode(trans, new_root, inode);
9142
9143         iput(inode);
9144         return err;
9145 }
9146
9147 struct inode *btrfs_alloc_inode(struct super_block *sb)
9148 {
9149         struct btrfs_fs_info *fs_info = btrfs_sb(sb);
9150         struct btrfs_inode *ei;
9151         struct inode *inode;
9152
9153         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_KERNEL);
9154         if (!ei)
9155                 return NULL;
9156
9157         ei->root = NULL;
9158         ei->generation = 0;
9159         ei->last_trans = 0;
9160         ei->last_sub_trans = 0;
9161         ei->logged_trans = 0;
9162         ei->delalloc_bytes = 0;
9163         ei->new_delalloc_bytes = 0;
9164         ei->defrag_bytes = 0;
9165         ei->disk_i_size = 0;
9166         ei->flags = 0;
9167         ei->csum_bytes = 0;
9168         ei->index_cnt = (u64)-1;
9169         ei->dir_index = 0;
9170         ei->last_unlink_trans = 0;
9171         ei->last_log_commit = 0;
9172
9173         spin_lock_init(&ei->lock);
9174         ei->outstanding_extents = 0;
9175         if (sb->s_magic != BTRFS_TEST_MAGIC)
9176                 btrfs_init_metadata_block_rsv(fs_info, &ei->block_rsv,
9177                                               BTRFS_BLOCK_RSV_DELALLOC);
9178         ei->runtime_flags = 0;
9179         ei->prop_compress = BTRFS_COMPRESS_NONE;
9180         ei->defrag_compress = BTRFS_COMPRESS_NONE;
9181
9182         ei->delayed_node = NULL;
9183
9184         ei->i_otime.tv_sec = 0;
9185         ei->i_otime.tv_nsec = 0;
9186
9187         inode = &ei->vfs_inode;
9188         extent_map_tree_init(&ei->extent_tree);
9189         extent_io_tree_init(&ei->io_tree, inode);
9190         extent_io_tree_init(&ei->io_failure_tree, inode);
9191         ei->io_tree.track_uptodate = 1;
9192         ei->io_failure_tree.track_uptodate = 1;
9193         atomic_set(&ei->sync_writers, 0);
9194         mutex_init(&ei->log_mutex);
9195         mutex_init(&ei->delalloc_mutex);
9196         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9197         INIT_LIST_HEAD(&ei->delalloc_inodes);
9198         INIT_LIST_HEAD(&ei->delayed_iput);
9199         RB_CLEAR_NODE(&ei->rb_node);
9200         init_rwsem(&ei->dio_sem);
9201
9202         return inode;
9203 }
9204
9205 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9206 void btrfs_test_destroy_inode(struct inode *inode)
9207 {
9208         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9209         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9210 }
9211 #endif
9212
9213 static void btrfs_i_callback(struct rcu_head *head)
9214 {
9215         struct inode *inode = container_of(head, struct inode, i_rcu);
9216         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9217 }
9218
9219 void btrfs_destroy_inode(struct inode *inode)
9220 {
9221         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9222         struct btrfs_ordered_extent *ordered;
9223         struct btrfs_root *root = BTRFS_I(inode)->root;
9224
9225         WARN_ON(!hlist_empty(&inode->i_dentry));
9226         WARN_ON(inode->i_data.nrpages);
9227         WARN_ON(BTRFS_I(inode)->block_rsv.reserved);
9228         WARN_ON(BTRFS_I(inode)->block_rsv.size);
9229         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9230         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9231         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
9232         WARN_ON(BTRFS_I(inode)->csum_bytes);
9233         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9234
9235         /*
9236          * This can happen where we create an inode, but somebody else also
9237          * created the same inode and we need to destroy the one we already
9238          * created.
9239          */
9240         if (!root)
9241                 goto free;
9242
9243         while (1) {
9244                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9245                 if (!ordered)
9246                         break;
9247                 else {
9248                         btrfs_err(fs_info,
9249                                   "found ordered extent %llu %llu on inode cleanup",
9250                                   ordered->file_offset, ordered->len);
9251                         btrfs_remove_ordered_extent(inode, ordered);
9252                         btrfs_put_ordered_extent(ordered);
9253                         btrfs_put_ordered_extent(ordered);
9254                 }
9255         }
9256         btrfs_qgroup_check_reserved_leak(inode);
9257         inode_tree_del(inode);
9258         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9259 free:
9260         call_rcu(&inode->i_rcu, btrfs_i_callback);
9261 }
9262
9263 int btrfs_drop_inode(struct inode *inode)
9264 {
9265         struct btrfs_root *root = BTRFS_I(inode)->root;
9266
9267         if (root == NULL)
9268                 return 1;
9269
9270         /* the snap/subvol tree is on deleting */
9271         if (btrfs_root_refs(&root->root_item) == 0)
9272                 return 1;
9273         else
9274                 return generic_drop_inode(inode);
9275 }
9276
9277 static void init_once(void *foo)
9278 {
9279         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9280
9281         inode_init_once(&ei->vfs_inode);
9282 }
9283
9284 void __cold btrfs_destroy_cachep(void)
9285 {
9286         /*
9287          * Make sure all delayed rcu free inodes are flushed before we
9288          * destroy cache.
9289          */
9290         rcu_barrier();
9291         kmem_cache_destroy(btrfs_inode_cachep);
9292         kmem_cache_destroy(btrfs_trans_handle_cachep);
9293         kmem_cache_destroy(btrfs_path_cachep);
9294         kmem_cache_destroy(btrfs_free_space_cachep);
9295 }
9296
9297 int __init btrfs_init_cachep(void)
9298 {
9299         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9300                         sizeof(struct btrfs_inode), 0,
9301                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9302                         init_once);
9303         if (!btrfs_inode_cachep)
9304                 goto fail;
9305
9306         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9307                         sizeof(struct btrfs_trans_handle), 0,
9308                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9309         if (!btrfs_trans_handle_cachep)
9310                 goto fail;
9311
9312         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9313                         sizeof(struct btrfs_path), 0,
9314                         SLAB_MEM_SPREAD, NULL);
9315         if (!btrfs_path_cachep)
9316                 goto fail;
9317
9318         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9319                         sizeof(struct btrfs_free_space), 0,
9320                         SLAB_MEM_SPREAD, NULL);
9321         if (!btrfs_free_space_cachep)
9322                 goto fail;
9323
9324         return 0;
9325 fail:
9326         btrfs_destroy_cachep();
9327         return -ENOMEM;
9328 }
9329
9330 static int btrfs_getattr(const struct path *path, struct kstat *stat,
9331                          u32 request_mask, unsigned int flags)
9332 {
9333         u64 delalloc_bytes;
9334         struct inode *inode = d_inode(path->dentry);
9335         u32 blocksize = inode->i_sb->s_blocksize;
9336         u32 bi_flags = BTRFS_I(inode)->flags;
9337
9338         stat->result_mask |= STATX_BTIME;
9339         stat->btime.tv_sec = BTRFS_I(inode)->i_otime.tv_sec;
9340         stat->btime.tv_nsec = BTRFS_I(inode)->i_otime.tv_nsec;
9341         if (bi_flags & BTRFS_INODE_APPEND)
9342                 stat->attributes |= STATX_ATTR_APPEND;
9343         if (bi_flags & BTRFS_INODE_COMPRESS)
9344                 stat->attributes |= STATX_ATTR_COMPRESSED;
9345         if (bi_flags & BTRFS_INODE_IMMUTABLE)
9346                 stat->attributes |= STATX_ATTR_IMMUTABLE;
9347         if (bi_flags & BTRFS_INODE_NODUMP)
9348                 stat->attributes |= STATX_ATTR_NODUMP;
9349
9350         stat->attributes_mask |= (STATX_ATTR_APPEND |
9351                                   STATX_ATTR_COMPRESSED |
9352                                   STATX_ATTR_IMMUTABLE |
9353                                   STATX_ATTR_NODUMP);
9354
9355         generic_fillattr(inode, stat);
9356         stat->dev = BTRFS_I(inode)->root->anon_dev;
9357
9358         spin_lock(&BTRFS_I(inode)->lock);
9359         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
9360         spin_unlock(&BTRFS_I(inode)->lock);
9361         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9362                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9363         return 0;
9364 }
9365
9366 static int btrfs_rename_exchange(struct inode *old_dir,
9367                               struct dentry *old_dentry,
9368                               struct inode *new_dir,
9369                               struct dentry *new_dentry)
9370 {
9371         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9372         struct btrfs_trans_handle *trans;
9373         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9374         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9375         struct inode *new_inode = new_dentry->d_inode;
9376         struct inode *old_inode = old_dentry->d_inode;
9377         struct timespec64 ctime = current_time(old_inode);
9378         struct dentry *parent;
9379         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9380         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
9381         u64 old_idx = 0;
9382         u64 new_idx = 0;
9383         u64 root_objectid;
9384         int ret;
9385         int ret2;
9386         bool root_log_pinned = false;
9387         bool dest_log_pinned = false;
9388
9389         /* we only allow rename subvolume link between subvolumes */
9390         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9391                 return -EXDEV;
9392
9393         /* close the race window with snapshot create/destroy ioctl */
9394         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9395                 down_read(&fs_info->subvol_sem);
9396         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9397                 down_read(&fs_info->subvol_sem);
9398
9399         /*
9400          * We want to reserve the absolute worst case amount of items.  So if
9401          * both inodes are subvols and we need to unlink them then that would
9402          * require 4 item modifications, but if they are both normal inodes it
9403          * would require 5 item modifications, so we'll assume their normal
9404          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9405          * should cover the worst case number of items we'll modify.
9406          */
9407         trans = btrfs_start_transaction(root, 12);
9408         if (IS_ERR(trans)) {
9409                 ret = PTR_ERR(trans);
9410                 goto out_notrans;
9411         }
9412
9413         /*
9414          * We need to find a free sequence number both in the source and
9415          * in the destination directory for the exchange.
9416          */
9417         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
9418         if (ret)
9419                 goto out_fail;
9420         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
9421         if (ret)
9422                 goto out_fail;
9423
9424         BTRFS_I(old_inode)->dir_index = 0ULL;
9425         BTRFS_I(new_inode)->dir_index = 0ULL;
9426
9427         /* Reference for the source. */
9428         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9429                 /* force full log commit if subvolume involved. */
9430                 btrfs_set_log_full_commit(fs_info, trans);
9431         } else {
9432                 btrfs_pin_log_trans(root);
9433                 root_log_pinned = true;
9434                 ret = btrfs_insert_inode_ref(trans, dest,
9435                                              new_dentry->d_name.name,
9436                                              new_dentry->d_name.len,
9437                                              old_ino,
9438                                              btrfs_ino(BTRFS_I(new_dir)),
9439                                              old_idx);
9440                 if (ret)
9441                         goto out_fail;
9442         }
9443
9444         /* And now for the dest. */
9445         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9446                 /* force full log commit if subvolume involved. */
9447                 btrfs_set_log_full_commit(fs_info, trans);
9448         } else {
9449                 btrfs_pin_log_trans(dest);
9450                 dest_log_pinned = true;
9451                 ret = btrfs_insert_inode_ref(trans, root,
9452                                              old_dentry->d_name.name,
9453                                              old_dentry->d_name.len,
9454                                              new_ino,
9455                                              btrfs_ino(BTRFS_I(old_dir)),
9456                                              new_idx);
9457                 if (ret)
9458                         goto out_fail;
9459         }
9460
9461         /* Update inode version and ctime/mtime. */
9462         inode_inc_iversion(old_dir);
9463         inode_inc_iversion(new_dir);
9464         inode_inc_iversion(old_inode);
9465         inode_inc_iversion(new_inode);
9466         old_dir->i_ctime = old_dir->i_mtime = ctime;
9467         new_dir->i_ctime = new_dir->i_mtime = ctime;
9468         old_inode->i_ctime = ctime;
9469         new_inode->i_ctime = ctime;
9470
9471         if (old_dentry->d_parent != new_dentry->d_parent) {
9472                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9473                                 BTRFS_I(old_inode), 1);
9474                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
9475                                 BTRFS_I(new_inode), 1);
9476         }
9477
9478         /* src is a subvolume */
9479         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9480                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9481                 ret = btrfs_unlink_subvol(trans, old_dir, root_objectid,
9482                                           old_dentry->d_name.name,
9483                                           old_dentry->d_name.len);
9484         } else { /* src is an inode */
9485                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9486                                            BTRFS_I(old_dentry->d_inode),
9487                                            old_dentry->d_name.name,
9488                                            old_dentry->d_name.len);
9489                 if (!ret)
9490                         ret = btrfs_update_inode(trans, root, old_inode);
9491         }
9492         if (ret) {
9493                 btrfs_abort_transaction(trans, ret);
9494                 goto out_fail;
9495         }
9496
9497         /* dest is a subvolume */
9498         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9499                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9500                 ret = btrfs_unlink_subvol(trans, new_dir, root_objectid,
9501                                           new_dentry->d_name.name,
9502                                           new_dentry->d_name.len);
9503         } else { /* dest is an inode */
9504                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9505                                            BTRFS_I(new_dentry->d_inode),
9506                                            new_dentry->d_name.name,
9507                                            new_dentry->d_name.len);
9508                 if (!ret)
9509                         ret = btrfs_update_inode(trans, dest, new_inode);
9510         }
9511         if (ret) {
9512                 btrfs_abort_transaction(trans, ret);
9513                 goto out_fail;
9514         }
9515
9516         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9517                              new_dentry->d_name.name,
9518                              new_dentry->d_name.len, 0, old_idx);
9519         if (ret) {
9520                 btrfs_abort_transaction(trans, ret);
9521                 goto out_fail;
9522         }
9523
9524         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
9525                              old_dentry->d_name.name,
9526                              old_dentry->d_name.len, 0, new_idx);
9527         if (ret) {
9528                 btrfs_abort_transaction(trans, ret);
9529                 goto out_fail;
9530         }
9531
9532         if (old_inode->i_nlink == 1)
9533                 BTRFS_I(old_inode)->dir_index = old_idx;
9534         if (new_inode->i_nlink == 1)
9535                 BTRFS_I(new_inode)->dir_index = new_idx;
9536
9537         if (root_log_pinned) {
9538                 parent = new_dentry->d_parent;
9539                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
9540                                 parent);
9541                 btrfs_end_log_trans(root);
9542                 root_log_pinned = false;
9543         }
9544         if (dest_log_pinned) {
9545                 parent = old_dentry->d_parent;
9546                 btrfs_log_new_name(trans, BTRFS_I(new_inode), BTRFS_I(new_dir),
9547                                 parent);
9548                 btrfs_end_log_trans(dest);
9549                 dest_log_pinned = false;
9550         }
9551 out_fail:
9552         /*
9553          * If we have pinned a log and an error happened, we unpin tasks
9554          * trying to sync the log and force them to fallback to a transaction
9555          * commit if the log currently contains any of the inodes involved in
9556          * this rename operation (to ensure we do not persist a log with an
9557          * inconsistent state for any of these inodes or leading to any
9558          * inconsistencies when replayed). If the transaction was aborted, the
9559          * abortion reason is propagated to userspace when attempting to commit
9560          * the transaction. If the log does not contain any of these inodes, we
9561          * allow the tasks to sync it.
9562          */
9563         if (ret && (root_log_pinned || dest_log_pinned)) {
9564                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9565                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9566                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9567                     (new_inode &&
9568                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9569                         btrfs_set_log_full_commit(fs_info, trans);
9570
9571                 if (root_log_pinned) {
9572                         btrfs_end_log_trans(root);
9573                         root_log_pinned = false;
9574                 }
9575                 if (dest_log_pinned) {
9576                         btrfs_end_log_trans(dest);
9577                         dest_log_pinned = false;
9578                 }
9579         }
9580         ret2 = btrfs_end_transaction(trans);
9581         ret = ret ? ret : ret2;
9582 out_notrans:
9583         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9584                 up_read(&fs_info->subvol_sem);
9585         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9586                 up_read(&fs_info->subvol_sem);
9587
9588         return ret;
9589 }
9590
9591 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9592                                      struct btrfs_root *root,
9593                                      struct inode *dir,
9594                                      struct dentry *dentry)
9595 {
9596         int ret;
9597         struct inode *inode;
9598         u64 objectid;
9599         u64 index;
9600
9601         ret = btrfs_find_free_ino(root, &objectid);
9602         if (ret)
9603                 return ret;
9604
9605         inode = btrfs_new_inode(trans, root, dir,
9606                                 dentry->d_name.name,
9607                                 dentry->d_name.len,
9608                                 btrfs_ino(BTRFS_I(dir)),
9609                                 objectid,
9610                                 S_IFCHR | WHITEOUT_MODE,
9611                                 &index);
9612
9613         if (IS_ERR(inode)) {
9614                 ret = PTR_ERR(inode);
9615                 return ret;
9616         }
9617
9618         inode->i_op = &btrfs_special_inode_operations;
9619         init_special_inode(inode, inode->i_mode,
9620                 WHITEOUT_DEV);
9621
9622         ret = btrfs_init_inode_security(trans, inode, dir,
9623                                 &dentry->d_name);
9624         if (ret)
9625                 goto out;
9626
9627         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9628                                 BTRFS_I(inode), 0, index);
9629         if (ret)
9630                 goto out;
9631
9632         ret = btrfs_update_inode(trans, root, inode);
9633 out:
9634         unlock_new_inode(inode);
9635         if (ret)
9636                 inode_dec_link_count(inode);
9637         iput(inode);
9638
9639         return ret;
9640 }
9641
9642 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9643                            struct inode *new_dir, struct dentry *new_dentry,
9644                            unsigned int flags)
9645 {
9646         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9647         struct btrfs_trans_handle *trans;
9648         unsigned int trans_num_items;
9649         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9650         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9651         struct inode *new_inode = d_inode(new_dentry);
9652         struct inode *old_inode = d_inode(old_dentry);
9653         u64 index = 0;
9654         u64 root_objectid;
9655         int ret;
9656         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9657         bool log_pinned = false;
9658
9659         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9660                 return -EPERM;
9661
9662         /* we only allow rename subvolume link between subvolumes */
9663         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9664                 return -EXDEV;
9665
9666         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9667             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
9668                 return -ENOTEMPTY;
9669
9670         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9671             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9672                 return -ENOTEMPTY;
9673
9674
9675         /* check for collisions, even if the  name isn't there */
9676         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9677                              new_dentry->d_name.name,
9678                              new_dentry->d_name.len);
9679
9680         if (ret) {
9681                 if (ret == -EEXIST) {
9682                         /* we shouldn't get
9683                          * eexist without a new_inode */
9684                         if (WARN_ON(!new_inode)) {
9685                                 return ret;
9686                         }
9687                 } else {
9688                         /* maybe -EOVERFLOW */
9689                         return ret;
9690                 }
9691         }
9692         ret = 0;
9693
9694         /*
9695          * we're using rename to replace one file with another.  Start IO on it
9696          * now so  we don't add too much work to the end of the transaction
9697          */
9698         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9699                 filemap_flush(old_inode->i_mapping);
9700
9701         /* close the racy window with snapshot create/destroy ioctl */
9702         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9703                 down_read(&fs_info->subvol_sem);
9704         /*
9705          * We want to reserve the absolute worst case amount of items.  So if
9706          * both inodes are subvols and we need to unlink them then that would
9707          * require 4 item modifications, but if they are both normal inodes it
9708          * would require 5 item modifications, so we'll assume they are normal
9709          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9710          * should cover the worst case number of items we'll modify.
9711          * If our rename has the whiteout flag, we need more 5 units for the
9712          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9713          * when selinux is enabled).
9714          */
9715         trans_num_items = 11;
9716         if (flags & RENAME_WHITEOUT)
9717                 trans_num_items += 5;
9718         trans = btrfs_start_transaction(root, trans_num_items);
9719         if (IS_ERR(trans)) {
9720                 ret = PTR_ERR(trans);
9721                 goto out_notrans;
9722         }
9723
9724         if (dest != root)
9725                 btrfs_record_root_in_trans(trans, dest);
9726
9727         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
9728         if (ret)
9729                 goto out_fail;
9730
9731         BTRFS_I(old_inode)->dir_index = 0ULL;
9732         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9733                 /* force full log commit if subvolume involved. */
9734                 btrfs_set_log_full_commit(fs_info, trans);
9735         } else {
9736                 btrfs_pin_log_trans(root);
9737                 log_pinned = true;
9738                 ret = btrfs_insert_inode_ref(trans, dest,
9739                                              new_dentry->d_name.name,
9740                                              new_dentry->d_name.len,
9741                                              old_ino,
9742                                              btrfs_ino(BTRFS_I(new_dir)), index);
9743                 if (ret)
9744                         goto out_fail;
9745         }
9746
9747         inode_inc_iversion(old_dir);
9748         inode_inc_iversion(new_dir);
9749         inode_inc_iversion(old_inode);
9750         old_dir->i_ctime = old_dir->i_mtime =
9751         new_dir->i_ctime = new_dir->i_mtime =
9752         old_inode->i_ctime = current_time(old_dir);
9753
9754         if (old_dentry->d_parent != new_dentry->d_parent)
9755                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9756                                 BTRFS_I(old_inode), 1);
9757
9758         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9759                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9760                 ret = btrfs_unlink_subvol(trans, old_dir, root_objectid,
9761                                         old_dentry->d_name.name,
9762                                         old_dentry->d_name.len);
9763         } else {
9764                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9765                                         BTRFS_I(d_inode(old_dentry)),
9766                                         old_dentry->d_name.name,
9767                                         old_dentry->d_name.len);
9768                 if (!ret)
9769                         ret = btrfs_update_inode(trans, root, old_inode);
9770         }
9771         if (ret) {
9772                 btrfs_abort_transaction(trans, ret);
9773                 goto out_fail;
9774         }
9775
9776         if (new_inode) {
9777                 inode_inc_iversion(new_inode);
9778                 new_inode->i_ctime = current_time(new_inode);
9779                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
9780                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9781                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9782                         ret = btrfs_unlink_subvol(trans, new_dir, root_objectid,
9783                                                 new_dentry->d_name.name,
9784                                                 new_dentry->d_name.len);
9785                         BUG_ON(new_inode->i_nlink == 0);
9786                 } else {
9787                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9788                                                  BTRFS_I(d_inode(new_dentry)),
9789                                                  new_dentry->d_name.name,
9790                                                  new_dentry->d_name.len);
9791                 }
9792                 if (!ret && new_inode->i_nlink == 0)
9793                         ret = btrfs_orphan_add(trans,
9794                                         BTRFS_I(d_inode(new_dentry)));
9795                 if (ret) {
9796                         btrfs_abort_transaction(trans, ret);
9797                         goto out_fail;
9798                 }
9799         }
9800
9801         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9802                              new_dentry->d_name.name,
9803                              new_dentry->d_name.len, 0, index);
9804         if (ret) {
9805                 btrfs_abort_transaction(trans, ret);
9806                 goto out_fail;
9807         }
9808
9809         if (old_inode->i_nlink == 1)
9810                 BTRFS_I(old_inode)->dir_index = index;
9811
9812         if (log_pinned) {
9813                 struct dentry *parent = new_dentry->d_parent;
9814
9815                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
9816                                 parent);
9817                 btrfs_end_log_trans(root);
9818                 log_pinned = false;
9819         }
9820
9821         if (flags & RENAME_WHITEOUT) {
9822                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
9823                                                 old_dentry);
9824
9825                 if (ret) {
9826                         btrfs_abort_transaction(trans, ret);
9827                         goto out_fail;
9828                 }
9829         }
9830 out_fail:
9831         /*
9832          * If we have pinned the log and an error happened, we unpin tasks
9833          * trying to sync the log and force them to fallback to a transaction
9834          * commit if the log currently contains any of the inodes involved in
9835          * this rename operation (to ensure we do not persist a log with an
9836          * inconsistent state for any of these inodes or leading to any
9837          * inconsistencies when replayed). If the transaction was aborted, the
9838          * abortion reason is propagated to userspace when attempting to commit
9839          * the transaction. If the log does not contain any of these inodes, we
9840          * allow the tasks to sync it.
9841          */
9842         if (ret && log_pinned) {
9843                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9844                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9845                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9846                     (new_inode &&
9847                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9848                         btrfs_set_log_full_commit(fs_info, trans);
9849
9850                 btrfs_end_log_trans(root);
9851                 log_pinned = false;
9852         }
9853         btrfs_end_transaction(trans);
9854 out_notrans:
9855         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9856                 up_read(&fs_info->subvol_sem);
9857
9858         return ret;
9859 }
9860
9861 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9862                          struct inode *new_dir, struct dentry *new_dentry,
9863                          unsigned int flags)
9864 {
9865         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
9866                 return -EINVAL;
9867
9868         if (flags & RENAME_EXCHANGE)
9869                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
9870                                           new_dentry);
9871
9872         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
9873 }
9874
9875 struct btrfs_delalloc_work {
9876         struct inode *inode;
9877         struct completion completion;
9878         struct list_head list;
9879         struct btrfs_work work;
9880 };
9881
9882 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9883 {
9884         struct btrfs_delalloc_work *delalloc_work;
9885         struct inode *inode;
9886
9887         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9888                                      work);
9889         inode = delalloc_work->inode;
9890         filemap_flush(inode->i_mapping);
9891         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9892                                 &BTRFS_I(inode)->runtime_flags))
9893                 filemap_flush(inode->i_mapping);
9894
9895         iput(inode);
9896         complete(&delalloc_work->completion);
9897 }
9898
9899 static struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode)
9900 {
9901         struct btrfs_delalloc_work *work;
9902
9903         work = kmalloc(sizeof(*work), GFP_NOFS);
9904         if (!work)
9905                 return NULL;
9906
9907         init_completion(&work->completion);
9908         INIT_LIST_HEAD(&work->list);
9909         work->inode = inode;
9910         WARN_ON_ONCE(!inode);
9911         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
9912                         btrfs_run_delalloc_work, NULL, NULL);
9913
9914         return work;
9915 }
9916
9917 /*
9918  * some fairly slow code that needs optimization. This walks the list
9919  * of all the inodes with pending delalloc and forces them to disk.
9920  */
9921 static int start_delalloc_inodes(struct btrfs_root *root, int nr)
9922 {
9923         struct btrfs_inode *binode;
9924         struct inode *inode;
9925         struct btrfs_delalloc_work *work, *next;
9926         struct list_head works;
9927         struct list_head splice;
9928         int ret = 0;
9929
9930         INIT_LIST_HEAD(&works);
9931         INIT_LIST_HEAD(&splice);
9932
9933         mutex_lock(&root->delalloc_mutex);
9934         spin_lock(&root->delalloc_lock);
9935         list_splice_init(&root->delalloc_inodes, &splice);
9936         while (!list_empty(&splice)) {
9937                 binode = list_entry(splice.next, struct btrfs_inode,
9938                                     delalloc_inodes);
9939
9940                 list_move_tail(&binode->delalloc_inodes,
9941                                &root->delalloc_inodes);
9942                 inode = igrab(&binode->vfs_inode);
9943                 if (!inode) {
9944                         cond_resched_lock(&root->delalloc_lock);
9945                         continue;
9946                 }
9947                 spin_unlock(&root->delalloc_lock);
9948
9949                 work = btrfs_alloc_delalloc_work(inode);
9950                 if (!work) {
9951                         iput(inode);
9952                         ret = -ENOMEM;
9953                         goto out;
9954                 }
9955                 list_add_tail(&work->list, &works);
9956                 btrfs_queue_work(root->fs_info->flush_workers,
9957                                  &work->work);
9958                 ret++;
9959                 if (nr != -1 && ret >= nr)
9960                         goto out;
9961                 cond_resched();
9962                 spin_lock(&root->delalloc_lock);
9963         }
9964         spin_unlock(&root->delalloc_lock);
9965
9966 out:
9967         list_for_each_entry_safe(work, next, &works, list) {
9968                 list_del_init(&work->list);
9969                 wait_for_completion(&work->completion);
9970                 kfree(work);
9971         }
9972
9973         if (!list_empty(&splice)) {
9974                 spin_lock(&root->delalloc_lock);
9975                 list_splice_tail(&splice, &root->delalloc_inodes);
9976                 spin_unlock(&root->delalloc_lock);
9977         }
9978         mutex_unlock(&root->delalloc_mutex);
9979         return ret;
9980 }
9981
9982 int btrfs_start_delalloc_inodes(struct btrfs_root *root)
9983 {
9984         struct btrfs_fs_info *fs_info = root->fs_info;
9985         int ret;
9986
9987         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
9988                 return -EROFS;
9989
9990         ret = start_delalloc_inodes(root, -1);
9991         if (ret > 0)
9992                 ret = 0;
9993         return ret;
9994 }
9995
9996 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int nr)
9997 {
9998         struct btrfs_root *root;
9999         struct list_head splice;
10000         int ret;
10001
10002         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10003                 return -EROFS;
10004
10005         INIT_LIST_HEAD(&splice);
10006
10007         mutex_lock(&fs_info->delalloc_root_mutex);
10008         spin_lock(&fs_info->delalloc_root_lock);
10009         list_splice_init(&fs_info->delalloc_roots, &splice);
10010         while (!list_empty(&splice) && nr) {
10011                 root = list_first_entry(&splice, struct btrfs_root,
10012                                         delalloc_root);
10013                 root = btrfs_grab_fs_root(root);
10014                 BUG_ON(!root);
10015                 list_move_tail(&root->delalloc_root,
10016                                &fs_info->delalloc_roots);
10017                 spin_unlock(&fs_info->delalloc_root_lock);
10018
10019                 ret = start_delalloc_inodes(root, nr);
10020                 btrfs_put_fs_root(root);
10021                 if (ret < 0)
10022                         goto out;
10023
10024                 if (nr != -1) {
10025                         nr -= ret;
10026                         WARN_ON(nr < 0);
10027                 }
10028                 spin_lock(&fs_info->delalloc_root_lock);
10029         }
10030         spin_unlock(&fs_info->delalloc_root_lock);
10031
10032         ret = 0;
10033 out:
10034         if (!list_empty(&splice)) {
10035                 spin_lock(&fs_info->delalloc_root_lock);
10036                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10037                 spin_unlock(&fs_info->delalloc_root_lock);
10038         }
10039         mutex_unlock(&fs_info->delalloc_root_mutex);
10040         return ret;
10041 }
10042
10043 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10044                          const char *symname)
10045 {
10046         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10047         struct btrfs_trans_handle *trans;
10048         struct btrfs_root *root = BTRFS_I(dir)->root;
10049         struct btrfs_path *path;
10050         struct btrfs_key key;
10051         struct inode *inode = NULL;
10052         int err;
10053         u64 objectid;
10054         u64 index = 0;
10055         int name_len;
10056         int datasize;
10057         unsigned long ptr;
10058         struct btrfs_file_extent_item *ei;
10059         struct extent_buffer *leaf;
10060
10061         name_len = strlen(symname);
10062         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
10063                 return -ENAMETOOLONG;
10064
10065         /*
10066          * 2 items for inode item and ref
10067          * 2 items for dir items
10068          * 1 item for updating parent inode item
10069          * 1 item for the inline extent item
10070          * 1 item for xattr if selinux is on
10071          */
10072         trans = btrfs_start_transaction(root, 7);
10073         if (IS_ERR(trans))
10074                 return PTR_ERR(trans);
10075
10076         err = btrfs_find_free_ino(root, &objectid);
10077         if (err)
10078                 goto out_unlock;
10079
10080         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10081                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
10082                                 objectid, S_IFLNK|S_IRWXUGO, &index);
10083         if (IS_ERR(inode)) {
10084                 err = PTR_ERR(inode);
10085                 inode = NULL;
10086                 goto out_unlock;
10087         }
10088
10089         /*
10090         * If the active LSM wants to access the inode during
10091         * d_instantiate it needs these. Smack checks to see
10092         * if the filesystem supports xattrs by looking at the
10093         * ops vector.
10094         */
10095         inode->i_fop = &btrfs_file_operations;
10096         inode->i_op = &btrfs_file_inode_operations;
10097         inode->i_mapping->a_ops = &btrfs_aops;
10098         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10099
10100         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10101         if (err)
10102                 goto out_unlock;
10103
10104         path = btrfs_alloc_path();
10105         if (!path) {
10106                 err = -ENOMEM;
10107                 goto out_unlock;
10108         }
10109         key.objectid = btrfs_ino(BTRFS_I(inode));
10110         key.offset = 0;
10111         key.type = BTRFS_EXTENT_DATA_KEY;
10112         datasize = btrfs_file_extent_calc_inline_size(name_len);
10113         err = btrfs_insert_empty_item(trans, root, path, &key,
10114                                       datasize);
10115         if (err) {
10116                 btrfs_free_path(path);
10117                 goto out_unlock;
10118         }
10119         leaf = path->nodes[0];
10120         ei = btrfs_item_ptr(leaf, path->slots[0],
10121                             struct btrfs_file_extent_item);
10122         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10123         btrfs_set_file_extent_type(leaf, ei,
10124                                    BTRFS_FILE_EXTENT_INLINE);
10125         btrfs_set_file_extent_encryption(leaf, ei, 0);
10126         btrfs_set_file_extent_compression(leaf, ei, 0);
10127         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10128         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10129
10130         ptr = btrfs_file_extent_inline_start(ei);
10131         write_extent_buffer(leaf, symname, ptr, name_len);
10132         btrfs_mark_buffer_dirty(leaf);
10133         btrfs_free_path(path);
10134
10135         inode->i_op = &btrfs_symlink_inode_operations;
10136         inode_nohighmem(inode);
10137         inode->i_mapping->a_ops = &btrfs_symlink_aops;
10138         inode_set_bytes(inode, name_len);
10139         btrfs_i_size_write(BTRFS_I(inode), name_len);
10140         err = btrfs_update_inode(trans, root, inode);
10141         /*
10142          * Last step, add directory indexes for our symlink inode. This is the
10143          * last step to avoid extra cleanup of these indexes if an error happens
10144          * elsewhere above.
10145          */
10146         if (!err)
10147                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
10148                                 BTRFS_I(inode), 0, index);
10149         if (err)
10150                 goto out_unlock;
10151
10152         d_instantiate_new(dentry, inode);
10153
10154 out_unlock:
10155         btrfs_end_transaction(trans);
10156         if (err && inode) {
10157                 inode_dec_link_count(inode);
10158                 discard_new_inode(inode);
10159         }
10160         btrfs_btree_balance_dirty(fs_info);
10161         return err;
10162 }
10163
10164 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10165                                        u64 start, u64 num_bytes, u64 min_size,
10166                                        loff_t actual_len, u64 *alloc_hint,
10167                                        struct btrfs_trans_handle *trans)
10168 {
10169         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
10170         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10171         struct extent_map *em;
10172         struct btrfs_root *root = BTRFS_I(inode)->root;
10173         struct btrfs_key ins;
10174         u64 cur_offset = start;
10175         u64 i_size;
10176         u64 cur_bytes;
10177         u64 last_alloc = (u64)-1;
10178         int ret = 0;
10179         bool own_trans = true;
10180         u64 end = start + num_bytes - 1;
10181
10182         if (trans)
10183                 own_trans = false;
10184         while (num_bytes > 0) {
10185                 if (own_trans) {
10186                         trans = btrfs_start_transaction(root, 3);
10187                         if (IS_ERR(trans)) {
10188                                 ret = PTR_ERR(trans);
10189                                 break;
10190                         }
10191                 }
10192
10193                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10194                 cur_bytes = max(cur_bytes, min_size);
10195                 /*
10196                  * If we are severely fragmented we could end up with really
10197                  * small allocations, so if the allocator is returning small
10198                  * chunks lets make its job easier by only searching for those
10199                  * sized chunks.
10200                  */
10201                 cur_bytes = min(cur_bytes, last_alloc);
10202                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10203                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10204                 if (ret) {
10205                         if (own_trans)
10206                                 btrfs_end_transaction(trans);
10207                         break;
10208                 }
10209                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
10210
10211                 last_alloc = ins.offset;
10212                 ret = insert_reserved_file_extent(trans, inode,
10213                                                   cur_offset, ins.objectid,
10214                                                   ins.offset, ins.offset,
10215                                                   ins.offset, 0, 0, 0,
10216                                                   BTRFS_FILE_EXTENT_PREALLOC);
10217                 if (ret) {
10218                         btrfs_free_reserved_extent(fs_info, ins.objectid,
10219                                                    ins.offset, 0);
10220                         btrfs_abort_transaction(trans, ret);
10221                         if (own_trans)
10222                                 btrfs_end_transaction(trans);
10223                         break;
10224                 }
10225
10226                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10227                                         cur_offset + ins.offset -1, 0);
10228
10229                 em = alloc_extent_map();
10230                 if (!em) {
10231                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10232                                 &BTRFS_I(inode)->runtime_flags);
10233                         goto next;
10234                 }
10235
10236                 em->start = cur_offset;
10237                 em->orig_start = cur_offset;
10238                 em->len = ins.offset;
10239                 em->block_start = ins.objectid;
10240                 em->block_len = ins.offset;
10241                 em->orig_block_len = ins.offset;
10242                 em->ram_bytes = ins.offset;
10243                 em->bdev = fs_info->fs_devices->latest_bdev;
10244                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10245                 em->generation = trans->transid;
10246
10247                 while (1) {
10248                         write_lock(&em_tree->lock);
10249                         ret = add_extent_mapping(em_tree, em, 1);
10250                         write_unlock(&em_tree->lock);
10251                         if (ret != -EEXIST)
10252                                 break;
10253                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10254                                                 cur_offset + ins.offset - 1,
10255                                                 0);
10256                 }
10257                 free_extent_map(em);
10258 next:
10259                 num_bytes -= ins.offset;
10260                 cur_offset += ins.offset;
10261                 *alloc_hint = ins.objectid + ins.offset;
10262
10263                 inode_inc_iversion(inode);
10264                 inode->i_ctime = current_time(inode);
10265                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10266                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10267                     (actual_len > inode->i_size) &&
10268                     (cur_offset > inode->i_size)) {
10269                         if (cur_offset > actual_len)
10270                                 i_size = actual_len;
10271                         else
10272                                 i_size = cur_offset;
10273                         i_size_write(inode, i_size);
10274                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10275                 }
10276
10277                 ret = btrfs_update_inode(trans, root, inode);
10278
10279                 if (ret) {
10280                         btrfs_abort_transaction(trans, ret);
10281                         if (own_trans)
10282                                 btrfs_end_transaction(trans);
10283                         break;
10284                 }
10285
10286                 if (own_trans)
10287                         btrfs_end_transaction(trans);
10288         }
10289         if (cur_offset < end)
10290                 btrfs_free_reserved_data_space(inode, NULL, cur_offset,
10291                         end - cur_offset + 1);
10292         return ret;
10293 }
10294
10295 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10296                               u64 start, u64 num_bytes, u64 min_size,
10297                               loff_t actual_len, u64 *alloc_hint)
10298 {
10299         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10300                                            min_size, actual_len, alloc_hint,
10301                                            NULL);
10302 }
10303
10304 int btrfs_prealloc_file_range_trans(struct inode *inode,
10305                                     struct btrfs_trans_handle *trans, int mode,
10306                                     u64 start, u64 num_bytes, u64 min_size,
10307                                     loff_t actual_len, u64 *alloc_hint)
10308 {
10309         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10310                                            min_size, actual_len, alloc_hint, trans);
10311 }
10312
10313 static int btrfs_set_page_dirty(struct page *page)
10314 {
10315         return __set_page_dirty_nobuffers(page);
10316 }
10317
10318 static int btrfs_permission(struct inode *inode, int mask)
10319 {
10320         struct btrfs_root *root = BTRFS_I(inode)->root;
10321         umode_t mode = inode->i_mode;
10322
10323         if (mask & MAY_WRITE &&
10324             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10325                 if (btrfs_root_readonly(root))
10326                         return -EROFS;
10327                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10328                         return -EACCES;
10329         }
10330         return generic_permission(inode, mask);
10331 }
10332
10333 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10334 {
10335         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10336         struct btrfs_trans_handle *trans;
10337         struct btrfs_root *root = BTRFS_I(dir)->root;
10338         struct inode *inode = NULL;
10339         u64 objectid;
10340         u64 index;
10341         int ret = 0;
10342
10343         /*
10344          * 5 units required for adding orphan entry
10345          */
10346         trans = btrfs_start_transaction(root, 5);
10347         if (IS_ERR(trans))
10348                 return PTR_ERR(trans);
10349
10350         ret = btrfs_find_free_ino(root, &objectid);
10351         if (ret)
10352                 goto out;
10353
10354         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10355                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
10356         if (IS_ERR(inode)) {
10357                 ret = PTR_ERR(inode);
10358                 inode = NULL;
10359                 goto out;
10360         }
10361
10362         inode->i_fop = &btrfs_file_operations;
10363         inode->i_op = &btrfs_file_inode_operations;
10364
10365         inode->i_mapping->a_ops = &btrfs_aops;
10366         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10367
10368         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10369         if (ret)
10370                 goto out;
10371
10372         ret = btrfs_update_inode(trans, root, inode);
10373         if (ret)
10374                 goto out;
10375         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10376         if (ret)
10377                 goto out;
10378
10379         /*
10380          * We set number of links to 0 in btrfs_new_inode(), and here we set
10381          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10382          * through:
10383          *
10384          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10385          */
10386         set_nlink(inode, 1);
10387         d_tmpfile(dentry, inode);
10388         unlock_new_inode(inode);
10389         mark_inode_dirty(inode);
10390 out:
10391         btrfs_end_transaction(trans);
10392         if (ret && inode)
10393                 discard_new_inode(inode);
10394         btrfs_btree_balance_dirty(fs_info);
10395         return ret;
10396 }
10397
10398 __attribute__((const))
10399 static int btrfs_readpage_io_failed_hook(struct page *page, int failed_mirror)
10400 {
10401         return -EAGAIN;
10402 }
10403
10404 static void btrfs_check_extent_io_range(void *private_data, const char *caller,
10405                                         u64 start, u64 end)
10406 {
10407         struct inode *inode = private_data;
10408         u64 isize;
10409
10410         isize = i_size_read(inode);
10411         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
10412                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
10413                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
10414                         caller, btrfs_ino(BTRFS_I(inode)), isize, start, end);
10415         }
10416 }
10417
10418 void btrfs_set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
10419 {
10420         struct inode *inode = tree->private_data;
10421         unsigned long index = start >> PAGE_SHIFT;
10422         unsigned long end_index = end >> PAGE_SHIFT;
10423         struct page *page;
10424
10425         while (index <= end_index) {
10426                 page = find_get_page(inode->i_mapping, index);
10427                 ASSERT(page); /* Pages should be in the extent_io_tree */
10428                 set_page_writeback(page);
10429                 put_page(page);
10430                 index++;
10431         }
10432 }
10433
10434 static const struct inode_operations btrfs_dir_inode_operations = {
10435         .getattr        = btrfs_getattr,
10436         .lookup         = btrfs_lookup,
10437         .create         = btrfs_create,
10438         .unlink         = btrfs_unlink,
10439         .link           = btrfs_link,
10440         .mkdir          = btrfs_mkdir,
10441         .rmdir          = btrfs_rmdir,
10442         .rename         = btrfs_rename2,
10443         .symlink        = btrfs_symlink,
10444         .setattr        = btrfs_setattr,
10445         .mknod          = btrfs_mknod,
10446         .listxattr      = btrfs_listxattr,
10447         .permission     = btrfs_permission,
10448         .get_acl        = btrfs_get_acl,
10449         .set_acl        = btrfs_set_acl,
10450         .update_time    = btrfs_update_time,
10451         .tmpfile        = btrfs_tmpfile,
10452 };
10453 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10454         .lookup         = btrfs_lookup,
10455         .permission     = btrfs_permission,
10456         .update_time    = btrfs_update_time,
10457 };
10458
10459 static const struct file_operations btrfs_dir_file_operations = {
10460         .llseek         = generic_file_llseek,
10461         .read           = generic_read_dir,
10462         .iterate_shared = btrfs_real_readdir,
10463         .open           = btrfs_opendir,
10464         .unlocked_ioctl = btrfs_ioctl,
10465 #ifdef CONFIG_COMPAT
10466         .compat_ioctl   = btrfs_compat_ioctl,
10467 #endif
10468         .release        = btrfs_release_file,
10469         .fsync          = btrfs_sync_file,
10470 };
10471
10472 static const struct extent_io_ops btrfs_extent_io_ops = {
10473         /* mandatory callbacks */
10474         .submit_bio_hook = btrfs_submit_bio_hook,
10475         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10476         .readpage_io_failed_hook = btrfs_readpage_io_failed_hook,
10477
10478         /* optional callbacks */
10479         .fill_delalloc = run_delalloc_range,
10480         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10481         .writepage_start_hook = btrfs_writepage_start_hook,
10482         .set_bit_hook = btrfs_set_bit_hook,
10483         .clear_bit_hook = btrfs_clear_bit_hook,
10484         .merge_extent_hook = btrfs_merge_extent_hook,
10485         .split_extent_hook = btrfs_split_extent_hook,
10486         .check_extent_io_range = btrfs_check_extent_io_range,
10487 };
10488
10489 /*
10490  * btrfs doesn't support the bmap operation because swapfiles
10491  * use bmap to make a mapping of extents in the file.  They assume
10492  * these extents won't change over the life of the file and they
10493  * use the bmap result to do IO directly to the drive.
10494  *
10495  * the btrfs bmap call would return logical addresses that aren't
10496  * suitable for IO and they also will change frequently as COW
10497  * operations happen.  So, swapfile + btrfs == corruption.
10498  *
10499  * For now we're avoiding this by dropping bmap.
10500  */
10501 static const struct address_space_operations btrfs_aops = {
10502         .readpage       = btrfs_readpage,
10503         .writepage      = btrfs_writepage,
10504         .writepages     = btrfs_writepages,
10505         .readpages      = btrfs_readpages,
10506         .direct_IO      = btrfs_direct_IO,
10507         .invalidatepage = btrfs_invalidatepage,
10508         .releasepage    = btrfs_releasepage,
10509         .set_page_dirty = btrfs_set_page_dirty,
10510         .error_remove_page = generic_error_remove_page,
10511 };
10512
10513 static const struct address_space_operations btrfs_symlink_aops = {
10514         .readpage       = btrfs_readpage,
10515         .writepage      = btrfs_writepage,
10516         .invalidatepage = btrfs_invalidatepage,
10517         .releasepage    = btrfs_releasepage,
10518 };
10519
10520 static const struct inode_operations btrfs_file_inode_operations = {
10521         .getattr        = btrfs_getattr,
10522         .setattr        = btrfs_setattr,
10523         .listxattr      = btrfs_listxattr,
10524         .permission     = btrfs_permission,
10525         .fiemap         = btrfs_fiemap,
10526         .get_acl        = btrfs_get_acl,
10527         .set_acl        = btrfs_set_acl,
10528         .update_time    = btrfs_update_time,
10529 };
10530 static const struct inode_operations btrfs_special_inode_operations = {
10531         .getattr        = btrfs_getattr,
10532         .setattr        = btrfs_setattr,
10533         .permission     = btrfs_permission,
10534         .listxattr      = btrfs_listxattr,
10535         .get_acl        = btrfs_get_acl,
10536         .set_acl        = btrfs_set_acl,
10537         .update_time    = btrfs_update_time,
10538 };
10539 static const struct inode_operations btrfs_symlink_inode_operations = {
10540         .get_link       = page_get_link,
10541         .getattr        = btrfs_getattr,
10542         .setattr        = btrfs_setattr,
10543         .permission     = btrfs_permission,
10544         .listxattr      = btrfs_listxattr,
10545         .update_time    = btrfs_update_time,
10546 };
10547
10548 const struct dentry_operations btrfs_dentry_operations = {
10549         .d_delete       = btrfs_dentry_delete,
10550 };