fs/btrfs/inode.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/kernel.h>
  20 #include <linux/bio.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/time.h>
  27 #include <linux/init.h>
  28 #include <linux/string.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/mpage.h>
  31 #include <linux/swap.h>
  32 #include <linux/writeback.h>
  33 #include <linux/compat.h>
  34 #include <linux/bit_spinlock.h>
  35 #include <linux/xattr.h>
  36 #include <linux/posix_acl.h>
  37 #include <linux/falloc.h>
  38 #include <linux/slab.h>
  39 #include <linux/ratelimit.h>
  40 #include <linux/mount.h>
  41 #include <linux/btrfs.h>
  42 #include <linux/blkdev.h>
  43 #include <linux/posix_acl_xattr.h>
  44 #include <linux/uio.h>
  45 #include <linux/magic.h>
  46 #include <linux/iversion.h>
  47 #include "ctree.h"
  48 #include "disk-io.h"
  49 #include "transaction.h"
  50 #include "btrfs_inode.h"
  51 #include "print-tree.h"
  52 #include "ordered-data.h"
  53 #include "xattr.h"
  54 #include "tree-log.h"
  55 #include "volumes.h"
  56 #include "compression.h"
  57 #include "locking.h"
  58 #include "free-space-cache.h"
  59 #include "inode-map.h"
  60 #include "backref.h"
  61 #include "hash.h"
  62 #include "props.h"
  63 #include "qgroup.h"
  64 #include "dedupe.h"
  65
  66 struct btrfs_iget_args {
  67         struct btrfs_key *location;
  68         struct btrfs_root *root;
  69 };
  70
  71 struct btrfs_dio_data {
  72         u64 reserve;
  73         u64 unsubmitted_oe_range_start;
  74         u64 unsubmitted_oe_range_end;
  75         int overwrite;
  76 };
  77
  78 static const struct inode_operations btrfs_dir_inode_operations;
  79 static const struct inode_operations btrfs_symlink_inode_operations;
  80 static const struct inode_operations btrfs_dir_ro_inode_operations;
  81 static const struct inode_operations btrfs_special_inode_operations;
  82 static const struct inode_operations btrfs_file_inode_operations;
  83 static const struct address_space_operations btrfs_aops;
  84 static const struct address_space_operations btrfs_symlink_aops;
  85 static const struct file_operations btrfs_dir_file_operations;
  86 static const struct extent_io_ops btrfs_extent_io_ops;
  87
  88 static struct kmem_cache *btrfs_inode_cachep;
  89 struct kmem_cache *btrfs_trans_handle_cachep;
  90 struct kmem_cache *btrfs_path_cachep;
  91 struct kmem_cache *btrfs_free_space_cachep;
  92
  93 #define S_SHIFT 12
  94 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  95         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  96         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  97         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  98         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  99         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
 100         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 101         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 102 };
 103
 104 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
 105 static int btrfs_truncate(struct inode *inode);
 106 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
 107 static noinline int cow_file_range(struct inode *inode,
 108                                    struct page *locked_page,
 109                                    u64 start, u64 end, u64 delalloc_end,
 110                                    int *page_started, unsigned long *nr_written,
 111                                    int unlock, struct btrfs_dedupe_hash *hash);
 112 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
 113                                        u64 orig_start, u64 block_start,
 114                                        u64 block_len, u64 orig_block_len,
 115                                        u64 ram_bytes, int compress_type,
 116                                        int type);
 117
 118 static void __endio_write_update_ordered(struct inode *inode,
 119                                          const u64 offset, const u64 bytes,
 120                                          const bool uptodate);
 121
 122 /*
 123  * Cleanup all submitted ordered extents in specified range to handle errors
 124  * from the fill_dellaloc() callback.
 125  *
 126  * NOTE: caller must ensure that when an error happens, it can not call
 127  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 128  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 129  * to be released, which we want to happen only when finishing the ordered
 130  * extent (btrfs_finish_ordered_io()). Also note that the caller of the
 131  * fill_delalloc() callback already does proper cleanup for the first page of
 132  * the range, that is, it invokes the callback writepage_end_io_hook() for the
 133  * range of the first page.
 134  */
 135 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 136                                                  const u64 offset,
 137                                                  const u64 bytes)
 138 {
 139         unsigned long index = offset >> PAGE_SHIFT;
 140         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
 141         struct page *page;
 142
 143         while (index <= end_index) {
 144                 page = find_get_page(inode->i_mapping, index);
 145                 index++;
 146                 if (!page)
 147                         continue;
 148                 ClearPagePrivate2(page);
 149                 put_page(page);
 150         }
 151         return __endio_write_update_ordered(inode, offset + PAGE_SIZE,
 152                                             bytes - PAGE_SIZE, false);
 153 }
 154
 155 static int btrfs_dirty_inode(struct inode *inode);
 156
 157 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 158 void btrfs_test_inode_set_ops(struct inode *inode)
 159 {
 160         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 161 }
 162 #endif
 163
 164 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 165                                      struct inode *inode,  struct inode *dir,
 166                                      const struct qstr *qstr)
 167 {
 168         int err;
 169
 170         err = btrfs_init_acl(trans, inode, dir);
 171         if (!err)
 172                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 173         return err;
 174 }
 175
 176 /*
 177  * this does all the hard work for inserting an inline extent into
 178  * the btree.  The caller should have done a btrfs_drop_extents so that
 179  * no overlapping inline items exist in the btree
 180  */
 181 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 182                                 struct btrfs_path *path, int extent_inserted,
 183                                 struct btrfs_root *root, struct inode *inode,
 184                                 u64 start, size_t size, size_t compressed_size,
 185                                 int compress_type,
 186                                 struct page **compressed_pages)
 187 {
 188         struct extent_buffer *leaf;
 189         struct page *page = NULL;
 190         char *kaddr;
 191         unsigned long ptr;
 192         struct btrfs_file_extent_item *ei;
 193         int ret;
 194         size_t cur_size = size;
 195         unsigned long offset;
 196
 197         if (compressed_size && compressed_pages)
 198                 cur_size = compressed_size;
 199
 200         inode_add_bytes(inode, size);
 201
 202         if (!extent_inserted) {
 203                 struct btrfs_key key;
 204                 size_t datasize;
 205
 206                 key.objectid = btrfs_ino(BTRFS_I(inode));
 207                 key.offset = start;
 208                 key.type = BTRFS_EXTENT_DATA_KEY;
 209
 210                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 211                 path->leave_spinning = 1;
 212                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 213                                               datasize);
 214                 if (ret)
 215                         goto fail;
 216         }
 217         leaf = path->nodes[0];
 218         ei = btrfs_item_ptr(leaf, path->slots[0],
 219                             struct btrfs_file_extent_item);
 220         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 221         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 222         btrfs_set_file_extent_encryption(leaf, ei, 0);
 223         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 224         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 225         ptr = btrfs_file_extent_inline_start(ei);
 226
 227         if (compress_type != BTRFS_COMPRESS_NONE) {
 228                 struct page *cpage;
 229                 int i = 0;
 230                 while (compressed_size > 0) {
 231                         cpage = compressed_pages[i];
 232                         cur_size = min_t(unsigned long, compressed_size,
 233                                        PAGE_SIZE);
 234
 235                         kaddr = kmap_atomic(cpage);
 236                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 237                         kunmap_atomic(kaddr);
 238
 239                         i++;
 240                         ptr += cur_size;
 241                         compressed_size -= cur_size;
 242                 }
 243                 btrfs_set_file_extent_compression(leaf, ei,
 244                                                   compress_type);
 245         } else {
 246                 page = find_get_page(inode->i_mapping,
 247                                      start >> PAGE_SHIFT);
 248                 btrfs_set_file_extent_compression(leaf, ei, 0);
 249                 kaddr = kmap_atomic(page);
 250                 offset = start & (PAGE_SIZE - 1);
 251                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 252                 kunmap_atomic(kaddr);
 253                 put_page(page);
 254         }
 255         btrfs_mark_buffer_dirty(leaf);
 256         btrfs_release_path(path);
 257
 258         /*
 259          * we're an inline extent, so nobody can
 260          * extend the file past i_size without locking
 261          * a page we already have locked.
 262          *
 263          * We must do any isize and inode updates
 264          * before we unlock the pages.  Otherwise we
 265          * could end up racing with unlink.
 266          */
 267         BTRFS_I(inode)->disk_i_size = inode->i_size;
 268         ret = btrfs_update_inode(trans, root, inode);
 269
 270 fail:
 271         return ret;
 272 }
 273
 274
 275 /*
 276  * conditionally insert an inline extent into the file.  This
 277  * does the checks required to make sure the data is small enough
 278  * to fit as an inline extent.
 279  */
 280 static noinline int cow_file_range_inline(struct btrfs_root *root,
 281                                           struct inode *inode, u64 start,
 282                                           u64 end, size_t compressed_size,
 283                                           int compress_type,
 284                                           struct page **compressed_pages)
 285 {
 286         struct btrfs_fs_info *fs_info = root->fs_info;
 287         struct btrfs_trans_handle *trans;
 288         u64 isize = i_size_read(inode);
 289         u64 actual_end = min(end + 1, isize);
 290         u64 inline_len = actual_end - start;
 291         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 292         u64 data_len = inline_len;
 293         int ret;
 294         struct btrfs_path *path;
 295         int extent_inserted = 0;
 296         u32 extent_item_size;
 297
 298         if (compressed_size)
 299                 data_len = compressed_size;
 300
 301         if (start > 0 ||
 302             actual_end > fs_info->sectorsize ||
 303             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 304             (!compressed_size &&
 305             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 306             end + 1 < isize ||
 307             data_len > fs_info->max_inline) {
 308                 return 1;
 309         }
 310
 311         path = btrfs_alloc_path();
 312         if (!path)
 313                 return -ENOMEM;
 314
 315         trans = btrfs_join_transaction(root);
 316         if (IS_ERR(trans)) {
 317                 btrfs_free_path(path);
 318                 return PTR_ERR(trans);
 319         }
 320         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
 321
 322         if (compressed_size && compressed_pages)
 323                 extent_item_size = btrfs_file_extent_calc_inline_size(
 324                    compressed_size);
 325         else
 326                 extent_item_size = btrfs_file_extent_calc_inline_size(
 327                     inline_len);
 328
 329         ret = __btrfs_drop_extents(trans, root, inode, path,
 330                                    start, aligned_end, NULL,
 331                                    1, 1, extent_item_size, &extent_inserted);
 332         if (ret) {
 333                 btrfs_abort_transaction(trans, ret);
 334                 goto out;
 335         }
 336
 337         if (isize > actual_end)
 338                 inline_len = min_t(u64, isize, actual_end);
 339         ret = insert_inline_extent(trans, path, extent_inserted,
 340                                    root, inode, start,
 341                                    inline_len, compressed_size,
 342                                    compress_type, compressed_pages);
 343         if (ret && ret != -ENOSPC) {
 344                 btrfs_abort_transaction(trans, ret);
 345                 goto out;
 346         } else if (ret == -ENOSPC) {
 347                 ret = 1;
 348                 goto out;
 349         }
 350
 351         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 352         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 353 out:
 354         /*
 355          * Don't forget to free the reserved space, as for inlined extent
 356          * it won't count as data extent, free them directly here.
 357          * And at reserve time, it's always aligned to page size, so
 358          * just free one page here.
 359          */
 360         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
 361         btrfs_free_path(path);
 362         btrfs_end_transaction(trans);
 363         return ret;
 364 }
 365
 366 struct async_extent {
 367         u64 start;
 368         u64 ram_size;
 369         u64 compressed_size;
 370         struct page **pages;
 371         unsigned long nr_pages;
 372         int compress_type;
 373         struct list_head list;
 374 };
 375
 376 struct async_cow {
 377         struct inode *inode;
 378         struct btrfs_root *root;
 379         struct page *locked_page;
 380         u64 start;
 381         u64 end;
 382         unsigned int write_flags;
 383         struct list_head extents;
 384         struct btrfs_work work;
 385 };
 386
 387 static noinline int add_async_extent(struct async_cow *cow,
 388                                      u64 start, u64 ram_size,
 389                                      u64 compressed_size,
 390                                      struct page **pages,
 391                                      unsigned long nr_pages,
 392                                      int compress_type)
 393 {
 394         struct async_extent *async_extent;
 395
 396         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 397         BUG_ON(!async_extent); /* -ENOMEM */
 398         async_extent->start = start;
 399         async_extent->ram_size = ram_size;
 400         async_extent->compressed_size = compressed_size;
 401         async_extent->pages = pages;
 402         async_extent->nr_pages = nr_pages;
 403         async_extent->compress_type = compress_type;
 404         list_add_tail(&async_extent->list, &cow->extents);
 405         return 0;
 406 }
 407
 408 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
 409 {
 410         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 411
 412         /* force compress */
 413         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 414                 return 1;
 415         /* defrag ioctl */
 416         if (BTRFS_I(inode)->defrag_compress)
 417                 return 1;
 418         /* bad compression ratios */
 419         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 420                 return 0;
 421         if (btrfs_test_opt(fs_info, COMPRESS) ||
 422             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 423             BTRFS_I(inode)->prop_compress)
 424                 return btrfs_compress_heuristic(inode, start, end);
 425         return 0;
 426 }
 427
 428 static inline void inode_should_defrag(struct btrfs_inode *inode,
 429                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 430 {
 431         /* If this is a small write inside eof, kick off a defrag */
 432         if (num_bytes < small_write &&
 433             (start > 0 || end + 1 < inode->disk_i_size))
 434                 btrfs_add_inode_defrag(NULL, inode);
 435 }
 436
 437 /*
 438  * we create compressed extents in two phases.  The first
 439  * phase compresses a range of pages that have already been
 440  * locked (both pages and state bits are locked).
 441  *
 442  * This is done inside an ordered work queue, and the compression
 443  * is spread across many cpus.  The actual IO submission is step
 444  * two, and the ordered work queue takes care of making sure that
 445  * happens in the same order things were put onto the queue by
 446  * writepages and friends.
 447  *
 448  * If this code finds it can't get good compression, it puts an
 449  * entry onto the work queue to write the uncompressed bytes.  This
 450  * makes sure that both compressed inodes and uncompressed inodes
 451  * are written in the same order that the flusher thread sent them
 452  * down.
 453  */
 454 static noinline void compress_file_range(struct inode *inode,
 455                                         struct page *locked_page,
 456                                         u64 start, u64 end,
 457                                         struct async_cow *async_cow,
 458                                         int *num_added)
 459 {
 460         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 461         struct btrfs_root *root = BTRFS_I(inode)->root;
 462         u64 blocksize = fs_info->sectorsize;
 463         u64 actual_end;
 464         u64 isize = i_size_read(inode);
 465         int ret = 0;
 466         struct page **pages = NULL;
 467         unsigned long nr_pages;
 468         unsigned long total_compressed = 0;
 469         unsigned long total_in = 0;
 470         int i;
 471         int will_compress;
 472         int compress_type = fs_info->compress_type;
 473         int redirty = 0;
 474
 475         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 476                         SZ_16K);
 477
 478         actual_end = min_t(u64, isize, end + 1);
 479 again:
 480         will_compress = 0;
 481         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 482         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 483         nr_pages = min_t(unsigned long, nr_pages,
 484                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 485
 486         /*
 487          * we don't want to send crud past the end of i_size through
 488          * compression, that's just a waste of CPU time.  So, if the
 489          * end of the file is before the start of our current
 490          * requested range of bytes, we bail out to the uncompressed
 491          * cleanup code that can deal with all of this.
 492          *
 493          * It isn't really the fastest way to fix things, but this is a
 494          * very uncommon corner.
 495          */
 496         if (actual_end <= start)
 497                 goto cleanup_and_bail_uncompressed;
 498
 499         total_compressed = actual_end - start;
 500
 501         /*
 502          * skip compression for a small file range(<=blocksize) that
 503          * isn't an inline extent, since it doesn't save disk space at all.
 504          */
 505         if (total_compressed <= blocksize &&
 506            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 507                 goto cleanup_and_bail_uncompressed;
 508
 509         total_compressed = min_t(unsigned long, total_compressed,
 510                         BTRFS_MAX_UNCOMPRESSED);
 511         total_in = 0;
 512         ret = 0;
 513
 514         /*
 515          * we do compression for mount -o compress and when the
 516          * inode has not been flagged as nocompress.  This flag can
 517          * change at any time if we discover bad compression ratios.
 518          */
 519         if (inode_need_compress(inode, start, end)) {
 520                 WARN_ON(pages);
 521                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 522                 if (!pages) {
 523                         /* just bail out to the uncompressed code */
 524                         goto cont;
 525                 }
 526
 527                 if (BTRFS_I(inode)->defrag_compress)
 528                         compress_type = BTRFS_I(inode)->defrag_compress;
 529                 else if (BTRFS_I(inode)->prop_compress)
 530                         compress_type = BTRFS_I(inode)->prop_compress;
 531
 532                 /*
 533                  * we need to call clear_page_dirty_for_io on each
 534                  * page in the range.  Otherwise applications with the file
 535                  * mmap'd can wander in and change the page contents while
 536                  * we are compressing them.
 537                  *
 538                  * If the compression fails for any reason, we set the pages
 539                  * dirty again later on.
 540                  */
 541                 extent_range_clear_dirty_for_io(inode, start, end);
 542                 redirty = 1;
 543
 544                 /* Compression level is applied here and only here */
 545                 ret = btrfs_compress_pages(
 546                         compress_type | (fs_info->compress_level << 4),
 547                                            inode->i_mapping, start,
 548                                            pages,
 549                                            &nr_pages,
 550                                            &total_in,
 551                                            &total_compressed);
 552
 553                 if (!ret) {
 554                         unsigned long offset = total_compressed &
 555                                 (PAGE_SIZE - 1);
 556                         struct page *page = pages[nr_pages - 1];
 557                         char *kaddr;
 558
 559                         /* zero the tail end of the last page, we might be
 560                          * sending it down to disk
 561                          */
 562                         if (offset) {
 563                                 kaddr = kmap_atomic(page);
 564                                 memset(kaddr + offset, 0,
 565                                        PAGE_SIZE - offset);
 566                                 kunmap_atomic(kaddr);
 567                         }
 568                         will_compress = 1;
 569                 }
 570         }
 571 cont:
 572         if (start == 0) {
 573                 /* lets try to make an inline extent */
 574                 if (ret || total_in < actual_end) {
 575                         /* we didn't compress the entire range, try
 576                          * to make an uncompressed inline extent.
 577                          */
 578                         ret = cow_file_range_inline(root, inode, start, end,
 579                                             0, BTRFS_COMPRESS_NONE, NULL);
 580                 } else {
 581                         /* try making a compressed inline extent */
 582                         ret = cow_file_range_inline(root, inode, start, end,
 583                                                     total_compressed,
 584                                                     compress_type, pages);
 585                 }
 586                 if (ret <= 0) {
 587                         unsigned long clear_flags = EXTENT_DELALLOC |
 588                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 589                                 EXTENT_DO_ACCOUNTING;
 590                         unsigned long page_error_op;
 591
 592                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 593
 594                         /*
 595                          * inline extent creation worked or returned error,
 596                          * we don't need to create any more async work items.
 597                          * Unlock and free up our temp pages.
 598                          *
 599                          * We use DO_ACCOUNTING here because we need the
 600                          * delalloc_release_metadata to be done _after_ we drop
 601                          * our outstanding extent for clearing delalloc for this
 602                          * range.
 603                          */
 604                         extent_clear_unlock_delalloc(inode, start, end, end,
 605                                                      NULL, clear_flags,
 606                                                      PAGE_UNLOCK |
 607                                                      PAGE_CLEAR_DIRTY |
 608                                                      PAGE_SET_WRITEBACK |
 609                                                      page_error_op |
 610                                                      PAGE_END_WRITEBACK);
 611                         goto free_pages_out;
 612                 }
 613         }
 614
 615         if (will_compress) {
 616                 /*
 617                  * we aren't doing an inline extent round the compressed size
 618                  * up to a block size boundary so the allocator does sane
 619                  * things
 620                  */
 621                 total_compressed = ALIGN(total_compressed, blocksize);
 622
 623                 /*
 624                  * one last check to make sure the compression is really a
 625                  * win, compare the page count read with the blocks on disk,
 626                  * compression must free at least one sector size
 627                  */
 628                 total_in = ALIGN(total_in, PAGE_SIZE);
 629                 if (total_compressed + blocksize <= total_in) {
 630                         *num_added += 1;
 631
 632                         /*
 633                          * The async work queues will take care of doing actual
 634                          * allocation on disk for these compressed pages, and
 635                          * will submit them to the elevator.
 636                          */
 637                         add_async_extent(async_cow, start, total_in,
 638                                         total_compressed, pages, nr_pages,
 639                                         compress_type);
 640
 641                         if (start + total_in < end) {
 642                                 start += total_in;
 643                                 pages = NULL;
 644                                 cond_resched();
 645                                 goto again;
 646                         }
 647                         return;
 648                 }
 649         }
 650         if (pages) {
 651                 /*
 652                  * the compression code ran but failed to make things smaller,
 653                  * free any pages it allocated and our page pointer array
 654                  */
 655                 for (i = 0; i < nr_pages; i++) {
 656                         WARN_ON(pages[i]->mapping);
 657                         put_page(pages[i]);
 658                 }
 659                 kfree(pages);
 660                 pages = NULL;
 661                 total_compressed = 0;
 662                 nr_pages = 0;
 663
 664                 /* flag the file so we don't compress in the future */
 665                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 666                     !(BTRFS_I(inode)->prop_compress)) {
 667                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 668                 }
 669         }
 670 cleanup_and_bail_uncompressed:
 671         /*
 672          * No compression, but we still need to write the pages in the file
 673          * we've been given so far.  redirty the locked page if it corresponds
 674          * to our extent and set things up for the async work queue to run
 675          * cow_file_range to do the normal delalloc dance.
 676          */
 677         if (page_offset(locked_page) >= start &&
 678             page_offset(locked_page) <= end)
 679                 __set_page_dirty_nobuffers(locked_page);
 680                 /* unlocked later on in the async handlers */
 681
 682         if (redirty)
 683                 extent_range_redirty_for_io(inode, start, end);
 684         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
 685                          BTRFS_COMPRESS_NONE);
 686         *num_added += 1;
 687
 688         return;
 689
 690 free_pages_out:
 691         for (i = 0; i < nr_pages; i++) {
 692                 WARN_ON(pages[i]->mapping);
 693                 put_page(pages[i]);
 694         }
 695         kfree(pages);
 696 }
 697
 698 static void free_async_extent_pages(struct async_extent *async_extent)
 699 {
 700         int i;
 701
 702         if (!async_extent->pages)
 703                 return;
 704
 705         for (i = 0; i < async_extent->nr_pages; i++) {
 706                 WARN_ON(async_extent->pages[i]->mapping);
 707                 put_page(async_extent->pages[i]);
 708         }
 709         kfree(async_extent->pages);
 710         async_extent->nr_pages = 0;
 711         async_extent->pages = NULL;
 712 }
 713
 714 /*
 715  * phase two of compressed writeback.  This is the ordered portion
 716  * of the code, which only gets called in the order the work was
 717  * queued.  We walk all the async extents created by compress_file_range
 718  * and send them down to the disk.
 719  */
 720 static noinline void submit_compressed_extents(struct inode *inode,
 721                                               struct async_cow *async_cow)
 722 {
 723         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 724         struct async_extent *async_extent;
 725         u64 alloc_hint = 0;
 726         struct btrfs_key ins;
 727         struct extent_map *em;
 728         struct btrfs_root *root = BTRFS_I(inode)->root;
 729         struct extent_io_tree *io_tree;
 730         int ret = 0;
 731
 732 again:
 733         while (!list_empty(&async_cow->extents)) {
 734                 async_extent = list_entry(async_cow->extents.next,
 735                                           struct async_extent, list);
 736                 list_del(&async_extent->list);
 737
 738                 io_tree = &BTRFS_I(inode)->io_tree;
 739
 740 retry:
 741                 /* did the compression code fall back to uncompressed IO? */
 742                 if (!async_extent->pages) {
 743                         int page_started = 0;
 744                         unsigned long nr_written = 0;
 745
 746                         lock_extent(io_tree, async_extent->start,
 747                                          async_extent->start +
 748                                          async_extent->ram_size - 1);
 749
 750                         /* allocate blocks */
 751                         ret = cow_file_range(inode, async_cow->locked_page,
 752                                              async_extent->start,
 753                                              async_extent->start +
 754                                              async_extent->ram_size - 1,
 755                                              async_extent->start +
 756                                              async_extent->ram_size - 1,
 757                                              &page_started, &nr_written, 0,
 758                                              NULL);
 759
 760                         /* JDM XXX */
 761
 762                         /*
 763                          * if page_started, cow_file_range inserted an
 764                          * inline extent and took care of all the unlocking
 765                          * and IO for us.  Otherwise, we need to submit
 766                          * all those pages down to the drive.
 767                          */
 768                         if (!page_started && !ret)
 769                                 extent_write_locked_range(io_tree,
 770                                                   inode, async_extent->start,
 771                                                   async_extent->start +
 772                                                   async_extent->ram_size - 1,
 773                                                   btrfs_get_extent,
 774                                                   WB_SYNC_ALL);
 775                         else if (ret)
 776                                 unlock_page(async_cow->locked_page);
 777                         kfree(async_extent);
 778                         cond_resched();
 779                         continue;
 780                 }
 781
 782                 lock_extent(io_tree, async_extent->start,
 783                             async_extent->start + async_extent->ram_size - 1);
 784
 785                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 786                                            async_extent->compressed_size,
 787                                            async_extent->compressed_size,
 788                                            0, alloc_hint, &ins, 1, 1);
 789                 if (ret) {
 790                         free_async_extent_pages(async_extent);
 791
 792                         if (ret == -ENOSPC) {
 793                                 unlock_extent(io_tree, async_extent->start,
 794                                               async_extent->start +
 795                                               async_extent->ram_size - 1);
 796
 797                                 /*
 798                                  * we need to redirty the pages if we decide to
 799                                  * fallback to uncompressed IO, otherwise we
 800                                  * will not submit these pages down to lower
 801                                  * layers.
 802                                  */
 803                                 extent_range_redirty_for_io(inode,
 804                                                 async_extent->start,
 805                                                 async_extent->start +
 806                                                 async_extent->ram_size - 1);
 807
 808                                 goto retry;
 809                         }
 810                         goto out_free;
 811                 }
 812                 /*
 813                  * here we're doing allocation and writeback of the
 814                  * compressed pages
 815                  */
 816                 em = create_io_em(inode, async_extent->start,
 817                                   async_extent->ram_size, /* len */
 818                                   async_extent->start, /* orig_start */
 819                                   ins.objectid, /* block_start */
 820                                   ins.offset, /* block_len */
 821                                   ins.offset, /* orig_block_len */
 822                                   async_extent->ram_size, /* ram_bytes */
 823                                   async_extent->compress_type,
 824                                   BTRFS_ORDERED_COMPRESSED);
 825                 if (IS_ERR(em))
 826                         /* ret value is not necessary due to void function */
 827                         goto out_free_reserve;
 828                 free_extent_map(em);
 829
 830                 ret = btrfs_add_ordered_extent_compress(inode,
 831                                                 async_extent->start,
 832                                                 ins.objectid,
 833                                                 async_extent->ram_size,
 834                                                 ins.offset,
 835                                                 BTRFS_ORDERED_COMPRESSED,
 836                                                 async_extent->compress_type);
 837                 if (ret) {
 838                         btrfs_drop_extent_cache(BTRFS_I(inode),
 839                                                 async_extent->start,
 840                                                 async_extent->start +
 841                                                 async_extent->ram_size - 1, 0);
 842                         goto out_free_reserve;
 843                 }
 844                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 845
 846                 /*
 847                  * clear dirty, set writeback and unlock the pages.
 848                  */
 849                 extent_clear_unlock_delalloc(inode, async_extent->start,
 850                                 async_extent->start +
 851                                 async_extent->ram_size - 1,
 852                                 async_extent->start +
 853                                 async_extent->ram_size - 1,
 854                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 855                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 856                                 PAGE_SET_WRITEBACK);
 857                 if (btrfs_submit_compressed_write(inode,
 858                                     async_extent->start,
 859                                     async_extent->ram_size,
 860                                     ins.objectid,
 861                                     ins.offset, async_extent->pages,
 862                                     async_extent->nr_pages,
 863                                     async_cow->write_flags)) {
 864                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 865                         struct page *p = async_extent->pages[0];
 866                         const u64 start = async_extent->start;
 867                         const u64 end = start + async_extent->ram_size - 1;
 868
 869                         p->mapping = inode->i_mapping;
 870                         tree->ops->writepage_end_io_hook(p, start, end,
 871                                                          NULL, 0);
 872                         p->mapping = NULL;
 873                         extent_clear_unlock_delalloc(inode, start, end, end,
 874                                                      NULL, 0,
 875                                                      PAGE_END_WRITEBACK |
 876                                                      PAGE_SET_ERROR);
 877                         free_async_extent_pages(async_extent);
 878                 }
 879                 alloc_hint = ins.objectid + ins.offset;
 880                 kfree(async_extent);
 881                 cond_resched();
 882         }
 883         return;
 884 out_free_reserve:
 885         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 886         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 887 out_free:
 888         extent_clear_unlock_delalloc(inode, async_extent->start,
 889                                      async_extent->start +
 890                                      async_extent->ram_size - 1,
 891                                      async_extent->start +
 892                                      async_extent->ram_size - 1,
 893                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 894                                      EXTENT_DELALLOC_NEW |
 895                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 896                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 897                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 898                                      PAGE_SET_ERROR);
 899         free_async_extent_pages(async_extent);
 900         kfree(async_extent);
 901         goto again;
 902 }
 903
 904 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 905                                       u64 num_bytes)
 906 {
 907         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 908         struct extent_map *em;
 909         u64 alloc_hint = 0;
 910
 911         read_lock(&em_tree->lock);
 912         em = search_extent_mapping(em_tree, start, num_bytes);
 913         if (em) {
 914                 /*
 915                  * if block start isn't an actual block number then find the
 916                  * first block in this inode and use that as a hint.  If that
 917                  * block is also bogus then just don't worry about it.
 918                  */
 919                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 920                         free_extent_map(em);
 921                         em = search_extent_mapping(em_tree, 0, 0);
 922                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 923                                 alloc_hint = em->block_start;
 924                         if (em)
 925                                 free_extent_map(em);
 926                 } else {
 927                         alloc_hint = em->block_start;
 928                         free_extent_map(em);
 929                 }
 930         }
 931         read_unlock(&em_tree->lock);
 932
 933         return alloc_hint;
 934 }
 935
 936 /*
 937  * when extent_io.c finds a delayed allocation range in the file,
 938  * the call backs end up in this code.  The basic idea is to
 939  * allocate extents on disk for the range, and create ordered data structs
 940  * in ram to track those extents.
 941  *
 942  * locked_page is the page that writepage had locked already.  We use
 943  * it to make sure we don't do extra locks or unlocks.
 944  *
 945  * *page_started is set to one if we unlock locked_page and do everything
 946  * required to start IO on it.  It may be clean and already done with
 947  * IO when we return.
 948  */
 949 static noinline int cow_file_range(struct inode *inode,
 950                                    struct page *locked_page,
 951                                    u64 start, u64 end, u64 delalloc_end,
 952                                    int *page_started, unsigned long *nr_written,
 953                                    int unlock, struct btrfs_dedupe_hash *hash)
 954 {
 955         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 956         struct btrfs_root *root = BTRFS_I(inode)->root;
 957         u64 alloc_hint = 0;
 958         u64 num_bytes;
 959         unsigned long ram_size;
 960         u64 disk_num_bytes;
 961         u64 cur_alloc_size = 0;
 962         u64 blocksize = fs_info->sectorsize;
 963         struct btrfs_key ins;
 964         struct extent_map *em;
 965         unsigned clear_bits;
 966         unsigned long page_ops;
 967         bool extent_reserved = false;
 968         int ret = 0;
 969
 970         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 971                 WARN_ON_ONCE(1);
 972                 ret = -EINVAL;
 973                 goto out_unlock;
 974         }
 975
 976         num_bytes = ALIGN(end - start + 1, blocksize);
 977         num_bytes = max(blocksize,  num_bytes);
 978         disk_num_bytes = num_bytes;
 979
 980         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
 981
 982         if (start == 0) {
 983                 /* lets try to make an inline extent */
 984                 ret = cow_file_range_inline(root, inode, start, end, 0,
 985                                         BTRFS_COMPRESS_NONE, NULL);
 986                 if (ret == 0) {
 987                         /*
 988                          * We use DO_ACCOUNTING here because we need the
 989                          * delalloc_release_metadata to be run _after_ we drop
 990                          * our outstanding extent for clearing delalloc for this
 991                          * range.
 992                          */
 993                         extent_clear_unlock_delalloc(inode, start, end,
 994                                      delalloc_end, NULL,
 995                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 996                                      EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 997                                      EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
 998                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 999                                      PAGE_END_WRITEBACK);
1000                         *nr_written = *nr_written +
1001                              (end - start + PAGE_SIZE) / PAGE_SIZE;
1002                         *page_started = 1;
1003                         goto out;
1004                 } else if (ret < 0) {
1005                         goto out_unlock;
1006                 }
1007         }
1008
1009         BUG_ON(disk_num_bytes >
1010                btrfs_super_total_bytes(fs_info->super_copy));
1011
1012         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
1013         btrfs_drop_extent_cache(BTRFS_I(inode), start,
1014                         start + num_bytes - 1, 0);
1015
1016         while (disk_num_bytes > 0) {
1017                 cur_alloc_size = disk_num_bytes;
1018                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1019                                            fs_info->sectorsize, 0, alloc_hint,
1020                                            &ins, 1, 1);
1021                 if (ret < 0)
1022                         goto out_unlock;
1023                 cur_alloc_size = ins.offset;
1024                 extent_reserved = true;
1025
1026                 ram_size = ins.offset;
1027                 em = create_io_em(inode, start, ins.offset, /* len */
1028                                   start, /* orig_start */
1029                                   ins.objectid, /* block_start */
1030                                   ins.offset, /* block_len */
1031                                   ins.offset, /* orig_block_len */
1032                                   ram_size, /* ram_bytes */
1033                                   BTRFS_COMPRESS_NONE, /* compress_type */
1034                                   BTRFS_ORDERED_REGULAR /* type */);
1035                 if (IS_ERR(em))
1036                         goto out_reserve;
1037                 free_extent_map(em);
1038
1039                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1040                                                ram_size, cur_alloc_size, 0);
1041                 if (ret)
1042                         goto out_drop_extent_cache;
1043
1044                 if (root->root_key.objectid ==
1045                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1046                         ret = btrfs_reloc_clone_csums(inode, start,
1047                                                       cur_alloc_size);
1048                         /*
1049                          * Only drop cache here, and process as normal.
1050                          *
1051                          * We must not allow extent_clear_unlock_delalloc()
1052                          * at out_unlock label to free meta of this ordered
1053                          * extent, as its meta should be freed by
1054                          * btrfs_finish_ordered_io().
1055                          *
1056                          * So we must continue until @start is increased to
1057                          * skip current ordered extent.
1058                          */
1059                         if (ret)
1060                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1061                                                 start + ram_size - 1, 0);
1062                 }
1063
1064                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1065
1066                 /* we're not doing compressed IO, don't unlock the first
1067                  * page (which the caller expects to stay locked), don't
1068                  * clear any dirty bits and don't set any writeback bits
1069                  *
1070                  * Do set the Private2 bit so we know this page was properly
1071                  * setup for writepage
1072                  */
1073                 page_ops = unlock ? PAGE_UNLOCK : 0;
1074                 page_ops |= PAGE_SET_PRIVATE2;
1075
1076                 extent_clear_unlock_delalloc(inode, start,
1077                                              start + ram_size - 1,
1078                                              delalloc_end, locked_page,
1079                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1080                                              page_ops);
1081                 if (disk_num_bytes < cur_alloc_size)
1082                         disk_num_bytes = 0;
1083                 else
1084                         disk_num_bytes -= cur_alloc_size;
1085                 num_bytes -= cur_alloc_size;
1086                 alloc_hint = ins.objectid + ins.offset;
1087                 start += cur_alloc_size;
1088                 extent_reserved = false;
1089
1090                 /*
1091                  * btrfs_reloc_clone_csums() error, since start is increased
1092                  * extent_clear_unlock_delalloc() at out_unlock label won't
1093                  * free metadata of current ordered extent, we're OK to exit.
1094                  */
1095                 if (ret)
1096                         goto out_unlock;
1097         }
1098 out:
1099         return ret;
1100
1101 out_drop_extent_cache:
1102         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1103 out_reserve:
1104         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1105         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1106 out_unlock:
1107         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1108                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1109         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1110                 PAGE_END_WRITEBACK;
1111         /*
1112          * If we reserved an extent for our delalloc range (or a subrange) and
1113          * failed to create the respective ordered extent, then it means that
1114          * when we reserved the extent we decremented the extent's size from
1115          * the data space_info's bytes_may_use counter and incremented the
1116          * space_info's bytes_reserved counter by the same amount. We must make
1117          * sure extent_clear_unlock_delalloc() does not try to decrement again
1118          * the data space_info's bytes_may_use counter, therefore we do not pass
1119          * it the flag EXTENT_CLEAR_DATA_RESV.
1120          */
1121         if (extent_reserved) {
1122                 extent_clear_unlock_delalloc(inode, start,
1123                                              start + cur_alloc_size,
1124                                              start + cur_alloc_size,
1125                                              locked_page,
1126                                              clear_bits,
1127                                              page_ops);
1128                 start += cur_alloc_size;
1129                 if (start >= end)
1130                         goto out;
1131         }
1132         extent_clear_unlock_delalloc(inode, start, end, delalloc_end,
1133                                      locked_page,
1134                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1135                                      page_ops);
1136         goto out;
1137 }
1138
1139 /*
1140  * work queue call back to started compression on a file and pages
1141  */
1142 static noinline void async_cow_start(struct btrfs_work *work)
1143 {
1144         struct async_cow *async_cow;
1145         int num_added = 0;
1146         async_cow = container_of(work, struct async_cow, work);
1147
1148         compress_file_range(async_cow->inode, async_cow->locked_page,
1149                             async_cow->start, async_cow->end, async_cow,
1150                             &num_added);
1151         if (num_added == 0) {
1152                 btrfs_add_delayed_iput(async_cow->inode);
1153                 async_cow->inode = NULL;
1154         }
1155 }
1156
1157 /*
1158  * work queue call back to submit previously compressed pages
1159  */
1160 static noinline void async_cow_submit(struct btrfs_work *work)
1161 {
1162         struct btrfs_fs_info *fs_info;
1163         struct async_cow *async_cow;
1164         struct btrfs_root *root;
1165         unsigned long nr_pages;
1166
1167         async_cow = container_of(work, struct async_cow, work);
1168
1169         root = async_cow->root;
1170         fs_info = root->fs_info;
1171         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1172                 PAGE_SHIFT;
1173
1174         /*
1175          * atomic_sub_return implies a barrier for waitqueue_active
1176          */
1177         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1178             5 * SZ_1M &&
1179             waitqueue_active(&fs_info->async_submit_wait))
1180                 wake_up(&fs_info->async_submit_wait);
1181
1182         if (async_cow->inode)
1183                 submit_compressed_extents(async_cow->inode, async_cow);
1184 }
1185
1186 static noinline void async_cow_free(struct btrfs_work *work)
1187 {
1188         struct async_cow *async_cow;
1189         async_cow = container_of(work, struct async_cow, work);
1190         if (async_cow->inode)
1191                 btrfs_add_delayed_iput(async_cow->inode);
1192         kfree(async_cow);
1193 }
1194
1195 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1196                                 u64 start, u64 end, int *page_started,
1197                                 unsigned long *nr_written,
1198                                 unsigned int write_flags)
1199 {
1200         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1201         struct async_cow *async_cow;
1202         struct btrfs_root *root = BTRFS_I(inode)->root;
1203         unsigned long nr_pages;
1204         u64 cur_end;
1205
1206         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1207                          1, 0, NULL, GFP_NOFS);
1208         while (start < end) {
1209                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1210                 BUG_ON(!async_cow); /* -ENOMEM */
1211                 async_cow->inode = igrab(inode);
1212                 async_cow->root = root;
1213                 async_cow->locked_page = locked_page;
1214                 async_cow->start = start;
1215                 async_cow->write_flags = write_flags;
1216
1217                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1218                     !btrfs_test_opt(fs_info, FORCE_COMPRESS))
1219                         cur_end = end;
1220                 else
1221                         cur_end = min(end, start + SZ_512K - 1);
1222
1223                 async_cow->end = cur_end;
1224                 INIT_LIST_HEAD(&async_cow->extents);
1225
1226                 btrfs_init_work(&async_cow->work,
1227                                 btrfs_delalloc_helper,
1228                                 async_cow_start, async_cow_submit,
1229                                 async_cow_free);
1230
1231                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1232                         PAGE_SHIFT;
1233                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1234
1235                 btrfs_queue_work(fs_info->delalloc_workers, &async_cow->work);
1236
1237                 *nr_written += nr_pages;
1238                 start = cur_end + 1;
1239         }
1240         *page_started = 1;
1241         return 0;
1242 }
1243
1244 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1245                                         u64 bytenr, u64 num_bytes)
1246 {
1247         int ret;
1248         struct btrfs_ordered_sum *sums;
1249         LIST_HEAD(list);
1250
1251         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1252                                        bytenr + num_bytes - 1, &list, 0);
1253         if (ret == 0 && list_empty(&list))
1254                 return 0;
1255
1256         while (!list_empty(&list)) {
1257                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1258                 list_del(&sums->list);
1259                 kfree(sums);
1260         }
1261         return 1;
1262 }
1263
1264 /*
1265  * when nowcow writeback call back.  This checks for snapshots or COW copies
1266  * of the extents that exist in the file, and COWs the file as required.
1267  *
1268  * If no cow copies or snapshots exist, we write directly to the existing
1269  * blocks on disk
1270  */
1271 static noinline int run_delalloc_nocow(struct inode *inode,
1272                                        struct page *locked_page,
1273                               u64 start, u64 end, int *page_started, int force,
1274                               unsigned long *nr_written)
1275 {
1276         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1277         struct btrfs_root *root = BTRFS_I(inode)->root;
1278         struct extent_buffer *leaf;
1279         struct btrfs_path *path;
1280         struct btrfs_file_extent_item *fi;
1281         struct btrfs_key found_key;
1282         struct extent_map *em;
1283         u64 cow_start;
1284         u64 cur_offset;
1285         u64 extent_end;
1286         u64 extent_offset;
1287         u64 disk_bytenr;
1288         u64 num_bytes;
1289         u64 disk_num_bytes;
1290         u64 ram_bytes;
1291         int extent_type;
1292         int ret, err;
1293         int type;
1294         int nocow;
1295         int check_prev = 1;
1296         bool nolock;
1297         u64 ino = btrfs_ino(BTRFS_I(inode));
1298
1299         path = btrfs_alloc_path();
1300         if (!path) {
1301                 extent_clear_unlock_delalloc(inode, start, end, end,
1302                                              locked_page,
1303                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1304                                              EXTENT_DO_ACCOUNTING |
1305                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1306                                              PAGE_CLEAR_DIRTY |
1307                                              PAGE_SET_WRITEBACK |
1308                                              PAGE_END_WRITEBACK);
1309                 return -ENOMEM;
1310         }
1311
1312         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
1313
1314         cow_start = (u64)-1;
1315         cur_offset = start;
1316         while (1) {
1317                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1318                                                cur_offset, 0);
1319                 if (ret < 0)
1320                         goto error;
1321                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1322                         leaf = path->nodes[0];
1323                         btrfs_item_key_to_cpu(leaf, &found_key,
1324                                               path->slots[0] - 1);
1325                         if (found_key.objectid == ino &&
1326                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1327                                 path->slots[0]--;
1328                 }
1329                 check_prev = 0;
1330 next_slot:
1331                 leaf = path->nodes[0];
1332                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1333                         ret = btrfs_next_leaf(root, path);
1334                         if (ret < 0)
1335                                 goto error;
1336                         if (ret > 0)
1337                                 break;
1338                         leaf = path->nodes[0];
1339                 }
1340
1341                 nocow = 0;
1342                 disk_bytenr = 0;
1343                 num_bytes = 0;
1344                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1345
1346                 if (found_key.objectid > ino)
1347                         break;
1348                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1349                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1350                         path->slots[0]++;
1351                         goto next_slot;
1352                 }
1353                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1354                     found_key.offset > end)
1355                         break;
1356
1357                 if (found_key.offset > cur_offset) {
1358                         extent_end = found_key.offset;
1359                         extent_type = 0;
1360                         goto out_check;
1361                 }
1362
1363                 fi = btrfs_item_ptr(leaf, path->slots[0],
1364                                     struct btrfs_file_extent_item);
1365                 extent_type = btrfs_file_extent_type(leaf, fi);
1366
1367                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1368                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1369                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1370                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1371                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1372                         extent_end = found_key.offset +
1373                                 btrfs_file_extent_num_bytes(leaf, fi);
1374                         disk_num_bytes =
1375                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1376                         if (extent_end <= start) {
1377                                 path->slots[0]++;
1378                                 goto next_slot;
1379                         }
1380                         if (disk_bytenr == 0)
1381                                 goto out_check;
1382                         if (btrfs_file_extent_compression(leaf, fi) ||
1383                             btrfs_file_extent_encryption(leaf, fi) ||
1384                             btrfs_file_extent_other_encoding(leaf, fi))
1385                                 goto out_check;
1386                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1387                                 goto out_check;
1388                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1389                                 goto out_check;
1390                         if (btrfs_cross_ref_exist(root, ino,
1391                                                   found_key.offset -
1392                                                   extent_offset, disk_bytenr))
1393                                 goto out_check;
1394                         disk_bytenr += extent_offset;
1395                         disk_bytenr += cur_offset - found_key.offset;
1396                         num_bytes = min(end + 1, extent_end) - cur_offset;
1397                         /*
1398                          * if there are pending snapshots for this root,
1399                          * we fall into common COW way.
1400                          */
1401                         if (!nolock) {
1402                                 err = btrfs_start_write_no_snapshotting(root);
1403                                 if (!err)
1404                                         goto out_check;
1405                         }
1406                         /*
1407                          * force cow if csum exists in the range.
1408                          * this ensure that csum for a given extent are
1409                          * either valid or do not exist.
1410                          */
1411                         if (csum_exist_in_range(fs_info, disk_bytenr,
1412                                                 num_bytes)) {
1413                                 if (!nolock)
1414                                         btrfs_end_write_no_snapshotting(root);
1415                                 goto out_check;
1416                         }
1417                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr)) {
1418                                 if (!nolock)
1419                                         btrfs_end_write_no_snapshotting(root);
1420                                 goto out_check;
1421                         }
1422                         nocow = 1;
1423                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1424                         extent_end = found_key.offset +
1425                                 btrfs_file_extent_inline_len(leaf,
1426                                                      path->slots[0], fi);
1427                         extent_end = ALIGN(extent_end,
1428                                            fs_info->sectorsize);
1429                 } else {
1430                         BUG_ON(1);
1431                 }
1432 out_check:
1433                 if (extent_end <= start) {
1434                         path->slots[0]++;
1435                         if (!nolock && nocow)
1436                                 btrfs_end_write_no_snapshotting(root);
1437                         if (nocow)
1438                                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1439                         goto next_slot;
1440                 }
1441                 if (!nocow) {
1442                         if (cow_start == (u64)-1)
1443                                 cow_start = cur_offset;
1444                         cur_offset = extent_end;
1445                         if (cur_offset > end)
1446                                 break;
1447                         path->slots[0]++;
1448                         goto next_slot;
1449                 }
1450
1451                 btrfs_release_path(path);
1452                 if (cow_start != (u64)-1) {
1453                         ret = cow_file_range(inode, locked_page,
1454                                              cow_start, found_key.offset - 1,
1455                                              end, page_started, nr_written, 1,
1456                                              NULL);
1457                         if (ret) {
1458                                 if (!nolock && nocow)
1459                                         btrfs_end_write_no_snapshotting(root);
1460                                 if (nocow)
1461                                         btrfs_dec_nocow_writers(fs_info,
1462                                                                 disk_bytenr);
1463                                 goto error;
1464                         }
1465                         cow_start = (u64)-1;
1466                 }
1467
1468                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1469                         u64 orig_start = found_key.offset - extent_offset;
1470
1471                         em = create_io_em(inode, cur_offset, num_bytes,
1472                                           orig_start,
1473                                           disk_bytenr, /* block_start */
1474                                           num_bytes, /* block_len */
1475                                           disk_num_bytes, /* orig_block_len */
1476                                           ram_bytes, BTRFS_COMPRESS_NONE,
1477                                           BTRFS_ORDERED_PREALLOC);
1478                         if (IS_ERR(em)) {
1479                                 if (!nolock && nocow)
1480                                         btrfs_end_write_no_snapshotting(root);
1481                                 if (nocow)
1482                                         btrfs_dec_nocow_writers(fs_info,
1483                                                                 disk_bytenr);
1484                                 ret = PTR_ERR(em);
1485                                 goto error;
1486                         }
1487                         free_extent_map(em);
1488                 }
1489
1490                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1491                         type = BTRFS_ORDERED_PREALLOC;
1492                 } else {
1493                         type = BTRFS_ORDERED_NOCOW;
1494                 }
1495
1496                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1497                                                num_bytes, num_bytes, type);
1498                 if (nocow)
1499                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1500                 BUG_ON(ret); /* -ENOMEM */
1501
1502                 if (root->root_key.objectid ==
1503                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1504                         /*
1505                          * Error handled later, as we must prevent
1506                          * extent_clear_unlock_delalloc() in error handler
1507                          * from freeing metadata of created ordered extent.
1508                          */
1509                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1510                                                       num_bytes);
1511
1512                 extent_clear_unlock_delalloc(inode, cur_offset,
1513                                              cur_offset + num_bytes - 1, end,
1514                                              locked_page, EXTENT_LOCKED |
1515                                              EXTENT_DELALLOC |
1516                                              EXTENT_CLEAR_DATA_RESV,
1517                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1518
1519                 if (!nolock && nocow)
1520                         btrfs_end_write_no_snapshotting(root);
1521                 cur_offset = extent_end;
1522
1523                 /*
1524                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1525                  * handler, as metadata for created ordered extent will only
1526                  * be freed by btrfs_finish_ordered_io().
1527                  */
1528                 if (ret)
1529                         goto error;
1530                 if (cur_offset > end)
1531                         break;
1532         }
1533         btrfs_release_path(path);
1534
1535         if (cur_offset <= end && cow_start == (u64)-1) {
1536                 cow_start = cur_offset;
1537                 cur_offset = end;
1538         }
1539
1540         if (cow_start != (u64)-1) {
1541                 ret = cow_file_range(inode, locked_page, cow_start, end, end,
1542                                      page_started, nr_written, 1, NULL);
1543                 if (ret)
1544                         goto error;
1545         }
1546
1547 error:
1548         if (ret && cur_offset < end)
1549                 extent_clear_unlock_delalloc(inode, cur_offset, end, end,
1550                                              locked_page, EXTENT_LOCKED |
1551                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1552                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1553                                              PAGE_CLEAR_DIRTY |
1554                                              PAGE_SET_WRITEBACK |
1555                                              PAGE_END_WRITEBACK);
1556         btrfs_free_path(path);
1557         return ret;
1558 }
1559
1560 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1561 {
1562
1563         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1564             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1565                 return 0;
1566
1567         /*
1568          * @defrag_bytes is a hint value, no spinlock held here,
1569          * if is not zero, it means the file is defragging.
1570          * Force cow if given extent needs to be defragged.
1571          */
1572         if (BTRFS_I(inode)->defrag_bytes &&
1573             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1574                            EXTENT_DEFRAG, 0, NULL))
1575                 return 1;
1576
1577         return 0;
1578 }
1579
1580 /*
1581  * extent_io.c call back to do delayed allocation processing
1582  */
1583 static int run_delalloc_range(void *private_data, struct page *locked_page,
1584                               u64 start, u64 end, int *page_started,
1585                               unsigned long *nr_written,
1586                               struct writeback_control *wbc)
1587 {
1588         struct inode *inode = private_data;
1589         int ret;
1590         int force_cow = need_force_cow(inode, start, end);
1591         unsigned int write_flags = wbc_to_write_flags(wbc);
1592
1593         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1594                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1595                                          page_started, 1, nr_written);
1596         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1597                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1598                                          page_started, 0, nr_written);
1599         } else if (!inode_need_compress(inode, start, end)) {
1600                 ret = cow_file_range(inode, locked_page, start, end, end,
1601                                       page_started, nr_written, 1, NULL);
1602         } else {
1603                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1604                         &BTRFS_I(inode)->runtime_flags);
1605                 ret = cow_file_range_async(inode, locked_page, start, end,
1606                                            page_started, nr_written,
1607                                            write_flags);
1608         }
1609         if (ret)
1610                 btrfs_cleanup_ordered_extents(inode, start, end - start + 1);
1611         return ret;
1612 }
1613
1614 static void btrfs_split_extent_hook(void *private_data,
1615                                     struct extent_state *orig, u64 split)
1616 {
1617         struct inode *inode = private_data;
1618         u64 size;
1619
1620         /* not delalloc, ignore it */
1621         if (!(orig->state & EXTENT_DELALLOC))
1622                 return;
1623
1624         size = orig->end - orig->start + 1;
1625         if (size > BTRFS_MAX_EXTENT_SIZE) {
1626                 u32 num_extents;
1627                 u64 new_size;
1628
1629                 /*
1630                  * See the explanation in btrfs_merge_extent_hook, the same
1631                  * applies here, just in reverse.
1632                  */
1633                 new_size = orig->end - split + 1;
1634                 num_extents = count_max_extents(new_size);
1635                 new_size = split - orig->start;
1636                 num_extents += count_max_extents(new_size);
1637                 if (count_max_extents(size) >= num_extents)
1638                         return;
1639         }
1640
1641         spin_lock(&BTRFS_I(inode)->lock);
1642         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
1643         spin_unlock(&BTRFS_I(inode)->lock);
1644 }
1645
1646 /*
1647  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1648  * extents so we can keep track of new extents that are just merged onto old
1649  * extents, such as when we are doing sequential writes, so we can properly
1650  * account for the metadata space we'll need.
1651  */
1652 static void btrfs_merge_extent_hook(void *private_data,
1653                                     struct extent_state *new,
1654                                     struct extent_state *other)
1655 {
1656         struct inode *inode = private_data;
1657         u64 new_size, old_size;
1658         u32 num_extents;
1659
1660         /* not delalloc, ignore it */
1661         if (!(other->state & EXTENT_DELALLOC))
1662                 return;
1663
1664         if (new->start > other->start)
1665                 new_size = new->end - other->start + 1;
1666         else
1667                 new_size = other->end - new->start + 1;
1668
1669         /* we're not bigger than the max, unreserve the space and go */
1670         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1671                 spin_lock(&BTRFS_I(inode)->lock);
1672                 btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1673                 spin_unlock(&BTRFS_I(inode)->lock);
1674                 return;
1675         }
1676
1677         /*
1678          * We have to add up either side to figure out how many extents were
1679          * accounted for before we merged into one big extent.  If the number of
1680          * extents we accounted for is <= the amount we need for the new range
1681          * then we can return, otherwise drop.  Think of it like this
1682          *
1683          * [ 4k][MAX_SIZE]
1684          *
1685          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1686          * need 2 outstanding extents, on one side we have 1 and the other side
1687          * we have 1 so they are == and we can return.  But in this case
1688          *
1689          * [MAX_SIZE+4k][MAX_SIZE+4k]
1690          *
1691          * Each range on their own accounts for 2 extents, but merged together
1692          * they are only 3 extents worth of accounting, so we need to drop in
1693          * this case.
1694          */
1695         old_size = other->end - other->start + 1;
1696         num_extents = count_max_extents(old_size);
1697         old_size = new->end - new->start + 1;
1698         num_extents += count_max_extents(old_size);
1699         if (count_max_extents(new_size) >= num_extents)
1700                 return;
1701
1702         spin_lock(&BTRFS_I(inode)->lock);
1703         btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1704         spin_unlock(&BTRFS_I(inode)->lock);
1705 }
1706
1707 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1708                                       struct inode *inode)
1709 {
1710         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1711
1712         spin_lock(&root->delalloc_lock);
1713         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1714                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1715                               &root->delalloc_inodes);
1716                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1717                         &BTRFS_I(inode)->runtime_flags);
1718                 root->nr_delalloc_inodes++;
1719                 if (root->nr_delalloc_inodes == 1) {
1720                         spin_lock(&fs_info->delalloc_root_lock);
1721                         BUG_ON(!list_empty(&root->delalloc_root));
1722                         list_add_tail(&root->delalloc_root,
1723                                       &fs_info->delalloc_roots);
1724                         spin_unlock(&fs_info->delalloc_root_lock);
1725                 }
1726         }
1727         spin_unlock(&root->delalloc_lock);
1728 }
1729
1730 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1731                                      struct btrfs_inode *inode)
1732 {
1733         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1734
1735         spin_lock(&root->delalloc_lock);
1736         if (!list_empty(&inode->delalloc_inodes)) {
1737                 list_del_init(&inode->delalloc_inodes);
1738                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1739                           &inode->runtime_flags);
1740                 root->nr_delalloc_inodes--;
1741                 if (!root->nr_delalloc_inodes) {
1742                         spin_lock(&fs_info->delalloc_root_lock);
1743                         BUG_ON(list_empty(&root->delalloc_root));
1744                         list_del_init(&root->delalloc_root);
1745                         spin_unlock(&fs_info->delalloc_root_lock);
1746                 }
1747         }
1748         spin_unlock(&root->delalloc_lock);
1749 }
1750
1751 /*
1752  * extent_io.c set_bit_hook, used to track delayed allocation
1753  * bytes in this file, and to maintain the list of inodes that
1754  * have pending delalloc work to be done.
1755  */
1756 static void btrfs_set_bit_hook(void *private_data,
1757                                struct extent_state *state, unsigned *bits)
1758 {
1759         struct inode *inode = private_data;
1760
1761         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1762
1763         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1764                 WARN_ON(1);
1765         /*
1766          * set_bit and clear bit hooks normally require _irqsave/restore
1767          * but in this case, we are only testing for the DELALLOC
1768          * bit, which is only set or cleared with irqs on
1769          */
1770         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1771                 struct btrfs_root *root = BTRFS_I(inode)->root;
1772                 u64 len = state->end + 1 - state->start;
1773                 u32 num_extents = count_max_extents(len);
1774                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1775
1776                 spin_lock(&BTRFS_I(inode)->lock);
1777                 btrfs_mod_outstanding_extents(BTRFS_I(inode), num_extents);
1778                 spin_unlock(&BTRFS_I(inode)->lock);
1779
1780                 /* For sanity tests */
1781                 if (btrfs_is_testing(fs_info))
1782                         return;
1783
1784                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
1785                                          fs_info->delalloc_batch);
1786                 spin_lock(&BTRFS_I(inode)->lock);
1787                 BTRFS_I(inode)->delalloc_bytes += len;
1788                 if (*bits & EXTENT_DEFRAG)
1789                         BTRFS_I(inode)->defrag_bytes += len;
1790                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1791                                          &BTRFS_I(inode)->runtime_flags))
1792                         btrfs_add_delalloc_inodes(root, inode);
1793                 spin_unlock(&BTRFS_I(inode)->lock);
1794         }
1795
1796         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1797             (*bits & EXTENT_DELALLOC_NEW)) {
1798                 spin_lock(&BTRFS_I(inode)->lock);
1799                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1800                         state->start;
1801                 spin_unlock(&BTRFS_I(inode)->lock);
1802         }
1803 }
1804
1805 /*
1806  * extent_io.c clear_bit_hook, see set_bit_hook for why
1807  */
1808 static void btrfs_clear_bit_hook(void *private_data,
1809                                  struct extent_state *state,
1810                                  unsigned *bits)
1811 {
1812         struct btrfs_inode *inode = BTRFS_I((struct inode *)private_data);
1813         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1814         u64 len = state->end + 1 - state->start;
1815         u32 num_extents = count_max_extents(len);
1816
1817         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
1818                 spin_lock(&inode->lock);
1819                 inode->defrag_bytes -= len;
1820                 spin_unlock(&inode->lock);
1821         }
1822
1823         /*
1824          * set_bit and clear bit hooks normally require _irqsave/restore
1825          * but in this case, we are only testing for the DELALLOC
1826          * bit, which is only set or cleared with irqs on
1827          */
1828         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1829                 struct btrfs_root *root = inode->root;
1830                 bool do_list = !btrfs_is_free_space_inode(inode);
1831
1832                 spin_lock(&inode->lock);
1833                 btrfs_mod_outstanding_extents(inode, -num_extents);
1834                 spin_unlock(&inode->lock);
1835
1836                 /*
1837                  * We don't reserve metadata space for space cache inodes so we
1838                  * don't need to call dellalloc_release_metadata if there is an
1839                  * error.
1840                  */
1841                 if (*bits & EXTENT_CLEAR_META_RESV &&
1842                     root != fs_info->tree_root)
1843                         btrfs_delalloc_release_metadata(inode, len);
1844
1845                 /* For sanity tests. */
1846                 if (btrfs_is_testing(fs_info))
1847                         return;
1848
1849                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
1850                     do_list && !(state->state & EXTENT_NORESERVE) &&
1851                     (*bits & EXTENT_CLEAR_DATA_RESV))
1852                         btrfs_free_reserved_data_space_noquota(
1853                                         &inode->vfs_inode,
1854                                         state->start, len);
1855
1856                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
1857                                          fs_info->delalloc_batch);
1858                 spin_lock(&inode->lock);
1859                 inode->delalloc_bytes -= len;
1860                 if (do_list && inode->delalloc_bytes == 0 &&
1861                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1862                                         &inode->runtime_flags))
1863                         btrfs_del_delalloc_inode(root, inode);
1864                 spin_unlock(&inode->lock);
1865         }
1866
1867         if ((state->state & EXTENT_DELALLOC_NEW) &&
1868             (*bits & EXTENT_DELALLOC_NEW)) {
1869                 spin_lock(&inode->lock);
1870                 ASSERT(inode->new_delalloc_bytes >= len);
1871                 inode->new_delalloc_bytes -= len;
1872                 spin_unlock(&inode->lock);
1873         }
1874 }
1875
1876 /*
1877  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1878  * we don't create bios that span stripes or chunks
1879  *
1880  * return 1 if page cannot be merged to bio
1881  * return 0 if page can be merged to bio
1882  * return error otherwise
1883  */
1884 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1885                          size_t size, struct bio *bio,
1886                          unsigned long bio_flags)
1887 {
1888         struct inode *inode = page->mapping->host;
1889         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1890         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1891         u64 length = 0;
1892         u64 map_length;
1893         int ret;
1894
1895         if (bio_flags & EXTENT_BIO_COMPRESSED)
1896                 return 0;
1897
1898         length = bio->bi_iter.bi_size;
1899         map_length = length;
1900         ret = btrfs_map_block(fs_info, btrfs_op(bio), logical, &map_length,
1901                               NULL, 0);
1902         if (ret < 0)
1903                 return ret;
1904         if (map_length < length + size)
1905                 return 1;
1906         return 0;
1907 }
1908
1909 /*
1910  * in order to insert checksums into the metadata in large chunks,
1911  * we wait until bio submission time.   All the pages in the bio are
1912  * checksummed and sums are attached onto the ordered extent record.
1913  *
1914  * At IO completion time the cums attached on the ordered extent record
1915  * are inserted into the btree
1916  */
1917 static blk_status_t __btrfs_submit_bio_start(void *private_data, struct bio *bio,
1918                                     int mirror_num, unsigned long bio_flags,
1919                                     u64 bio_offset)
1920 {
1921         struct inode *inode = private_data;
1922         blk_status_t ret = 0;
1923
1924         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1925         BUG_ON(ret); /* -ENOMEM */
1926         return 0;
1927 }
1928
1929 /*
1930  * in order to insert checksums into the metadata in large chunks,
1931  * we wait until bio submission time.   All the pages in the bio are
1932  * checksummed and sums are attached onto the ordered extent record.
1933  *
1934  * At IO completion time the cums attached on the ordered extent record
1935  * are inserted into the btree
1936  */
1937 static blk_status_t __btrfs_submit_bio_done(void *private_data, struct bio *bio,
1938                           int mirror_num, unsigned long bio_flags,
1939                           u64 bio_offset)
1940 {
1941         struct inode *inode = private_data;
1942         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1943         blk_status_t ret;
1944
1945         ret = btrfs_map_bio(fs_info, bio, mirror_num, 1);
1946         if (ret) {
1947                 bio->bi_status = ret;
1948                 bio_endio(bio);
1949         }
1950         return ret;
1951 }
1952
1953 /*
1954  * extent_io.c submission hook. This does the right thing for csum calculation
1955  * on write, or reading the csums from the tree before a read
1956  */
1957 static blk_status_t btrfs_submit_bio_hook(void *private_data, struct bio *bio,
1958                                  int mirror_num, unsigned long bio_flags,
1959                                  u64 bio_offset)
1960 {
1961         struct inode *inode = private_data;
1962         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1963         struct btrfs_root *root = BTRFS_I(inode)->root;
1964         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1965         blk_status_t ret = 0;
1966         int skip_sum;
1967         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1968
1969         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1970
1971         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
1972                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1973
1974         if (bio_op(bio) != REQ_OP_WRITE) {
1975                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
1976                 if (ret)
1977                         goto out;
1978
1979                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1980                         ret = btrfs_submit_compressed_read(inode, bio,
1981                                                            mirror_num,
1982                                                            bio_flags);
1983                         goto out;
1984                 } else if (!skip_sum) {
1985                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
1986                         if (ret)
1987                                 goto out;
1988                 }
1989                 goto mapit;
1990         } else if (async && !skip_sum) {
1991                 /* csum items have already been cloned */
1992                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1993                         goto mapit;
1994                 /* we're doing a write, do the async checksumming */
1995                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
1996                                           bio_offset, inode,
1997                                           __btrfs_submit_bio_start,
1998                                           __btrfs_submit_bio_done);
1999                 goto out;
2000         } else if (!skip_sum) {
2001                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2002                 if (ret)
2003                         goto out;
2004         }
2005
2006 mapit:
2007         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
2008
2009 out:
2010         if (ret) {
2011                 bio->bi_status = ret;
2012                 bio_endio(bio);
2013         }
2014         return ret;
2015 }
2016
2017 /*
2018  * given a list of ordered sums record them in the inode.  This happens
2019  * at IO completion time based on sums calculated at bio submission time.
2020  */
2021 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2022                              struct inode *inode, struct list_head *list)
2023 {
2024         struct btrfs_ordered_sum *sum;
2025
2026         list_for_each_entry(sum, list, list) {
2027                 trans->adding_csums = 1;
2028                 btrfs_csum_file_blocks(trans,
2029                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2030                 trans->adding_csums = 0;
2031         }
2032         return 0;
2033 }
2034
2035 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2036                               unsigned int extra_bits,
2037                               struct extent_state **cached_state, int dedupe)
2038 {
2039         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
2040         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2041                                    extra_bits, cached_state);
2042 }
2043
2044 /* see btrfs_writepage_start_hook for details on why this is required */
2045 struct btrfs_writepage_fixup {
2046         struct page *page;
2047         struct btrfs_work work;
2048 };
2049
2050 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2051 {
2052         struct btrfs_writepage_fixup *fixup;
2053         struct btrfs_ordered_extent *ordered;
2054         struct extent_state *cached_state = NULL;
2055         struct extent_changeset *data_reserved = NULL;
2056         struct page *page;
2057         struct inode *inode;
2058         u64 page_start;
2059         u64 page_end;
2060         int ret;
2061
2062         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2063         page = fixup->page;
2064 again:
2065         lock_page(page);
2066         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2067                 ClearPageChecked(page);
2068                 goto out_page;
2069         }
2070
2071         inode = page->mapping->host;
2072         page_start = page_offset(page);
2073         page_end = page_offset(page) + PAGE_SIZE - 1;
2074
2075         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2076                          &cached_state);
2077
2078         /* already ordered? We're done */
2079         if (PagePrivate2(page))
2080                 goto out;
2081
2082         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2083                                         PAGE_SIZE);
2084         if (ordered) {
2085                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2086                                      page_end, &cached_state, GFP_NOFS);
2087                 unlock_page(page);
2088                 btrfs_start_ordered_extent(inode, ordered, 1);
2089                 btrfs_put_ordered_extent(ordered);
2090                 goto again;
2091         }
2092
2093         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2094                                            PAGE_SIZE);
2095         if (ret) {
2096                 mapping_set_error(page->mapping, ret);
2097                 end_extent_writepage(page, ret, page_start, page_end);
2098                 ClearPageChecked(page);
2099                 goto out;
2100          }
2101
2102         btrfs_set_extent_delalloc(inode, page_start, page_end, 0, &cached_state,
2103                                   0);
2104         ClearPageChecked(page);
2105         set_page_dirty(page);
2106         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
2107 out:
2108         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2109                              &cached_state, GFP_NOFS);
2110 out_page:
2111         unlock_page(page);
2112         put_page(page);
2113         kfree(fixup);
2114         extent_changeset_free(data_reserved);
2115 }
2116
2117 /*
2118  * There are a few paths in the higher layers of the kernel that directly
2119  * set the page dirty bit without asking the filesystem if it is a
2120  * good idea.  This causes problems because we want to make sure COW
2121  * properly happens and the data=ordered rules are followed.
2122  *
2123  * In our case any range that doesn't have the ORDERED bit set
2124  * hasn't been properly setup for IO.  We kick off an async process
2125  * to fix it up.  The async helper will wait for ordered extents, set
2126  * the delalloc bit and make it safe to write the page.
2127  */
2128 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2129 {
2130         struct inode *inode = page->mapping->host;
2131         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2132         struct btrfs_writepage_fixup *fixup;
2133
2134         /* this page is properly in the ordered list */
2135         if (TestClearPagePrivate2(page))
2136                 return 0;
2137
2138         if (PageChecked(page))
2139                 return -EAGAIN;
2140
2141         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2142         if (!fixup)
2143                 return -EAGAIN;
2144
2145         SetPageChecked(page);
2146         get_page(page);
2147         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2148                         btrfs_writepage_fixup_worker, NULL, NULL);
2149         fixup->page = page;
2150         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2151         return -EBUSY;
2152 }
2153
2154 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2155                                        struct inode *inode, u64 file_pos,
2156                                        u64 disk_bytenr, u64 disk_num_bytes,
2157                                        u64 num_bytes, u64 ram_bytes,
2158                                        u8 compression, u8 encryption,
2159                                        u16 other_encoding, int extent_type)
2160 {
2161         struct btrfs_root *root = BTRFS_I(inode)->root;
2162         struct btrfs_file_extent_item *fi;
2163         struct btrfs_path *path;
2164         struct extent_buffer *leaf;
2165         struct btrfs_key ins;
2166         u64 qg_released;
2167         int extent_inserted = 0;
2168         int ret;
2169
2170         path = btrfs_alloc_path();
2171         if (!path)
2172                 return -ENOMEM;
2173
2174         /*
2175          * we may be replacing one extent in the tree with another.
2176          * The new extent is pinned in the extent map, and we don't want
2177          * to drop it from the cache until it is completely in the btree.
2178          *
2179          * So, tell btrfs_drop_extents to leave this extent in the cache.
2180          * the caller is expected to unpin it and allow it to be merged
2181          * with the others.
2182          */
2183         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2184                                    file_pos + num_bytes, NULL, 0,
2185                                    1, sizeof(*fi), &extent_inserted);
2186         if (ret)
2187                 goto out;
2188
2189         if (!extent_inserted) {
2190                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2191                 ins.offset = file_pos;
2192                 ins.type = BTRFS_EXTENT_DATA_KEY;
2193
2194                 path->leave_spinning = 1;
2195                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2196                                               sizeof(*fi));
2197                 if (ret)
2198                         goto out;
2199         }
2200         leaf = path->nodes[0];
2201         fi = btrfs_item_ptr(leaf, path->slots[0],
2202                             struct btrfs_file_extent_item);
2203         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2204         btrfs_set_file_extent_type(leaf, fi, extent_type);
2205         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2206         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2207         btrfs_set_file_extent_offset(leaf, fi, 0);
2208         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2209         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2210         btrfs_set_file_extent_compression(leaf, fi, compression);
2211         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2212         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2213
2214         btrfs_mark_buffer_dirty(leaf);
2215         btrfs_release_path(path);
2216
2217         inode_add_bytes(inode, num_bytes);
2218
2219         ins.objectid = disk_bytenr;
2220         ins.offset = disk_num_bytes;
2221         ins.type = BTRFS_EXTENT_ITEM_KEY;
2222
2223         /*
2224          * Release the reserved range from inode dirty range map, as it is
2225          * already moved into delayed_ref_head
2226          */
2227         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2228         if (ret < 0)
2229                 goto out;
2230         qg_released = ret;
2231         ret = btrfs_alloc_reserved_file_extent(trans, root,
2232                                                btrfs_ino(BTRFS_I(inode)),
2233                                                file_pos, qg_released, &ins);
2234 out:
2235         btrfs_free_path(path);
2236
2237         return ret;
2238 }
2239
2240 /* snapshot-aware defrag */
2241 struct sa_defrag_extent_backref {
2242         struct rb_node node;
2243         struct old_sa_defrag_extent *old;
2244         u64 root_id;
2245         u64 inum;
2246         u64 file_pos;
2247         u64 extent_offset;
2248         u64 num_bytes;
2249         u64 generation;
2250 };
2251
2252 struct old_sa_defrag_extent {
2253         struct list_head list;
2254         struct new_sa_defrag_extent *new;
2255
2256         u64 extent_offset;
2257         u64 bytenr;
2258         u64 offset;
2259         u64 len;
2260         int count;
2261 };
2262
2263 struct new_sa_defrag_extent {
2264         struct rb_root root;
2265         struct list_head head;
2266         struct btrfs_path *path;
2267         struct inode *inode;
2268         u64 file_pos;
2269         u64 len;
2270         u64 bytenr;
2271         u64 disk_len;
2272         u8 compress_type;
2273 };
2274
2275 static int backref_comp(struct sa_defrag_extent_backref *b1,
2276                         struct sa_defrag_extent_backref *b2)
2277 {
2278         if (b1->root_id < b2->root_id)
2279                 return -1;
2280         else if (b1->root_id > b2->root_id)
2281                 return 1;
2282
2283         if (b1->inum < b2->inum)
2284                 return -1;
2285         else if (b1->inum > b2->inum)
2286                 return 1;
2287
2288         if (b1->file_pos < b2->file_pos)
2289                 return -1;
2290         else if (b1->file_pos > b2->file_pos)
2291                 return 1;
2292
2293         /*
2294          * [------------------------------] ===> (a range of space)
2295          *     |<--->|   |<---->| =============> (fs/file tree A)
2296          * |<---------------------------->| ===> (fs/file tree B)
2297          *
2298          * A range of space can refer to two file extents in one tree while
2299          * refer to only one file extent in another tree.
2300          *
2301          * So we may process a disk offset more than one time(two extents in A)
2302          * and locate at the same extent(one extent in B), then insert two same
2303          * backrefs(both refer to the extent in B).
2304          */
2305         return 0;
2306 }
2307
2308 static void backref_insert(struct rb_root *root,
2309                            struct sa_defrag_extent_backref *backref)
2310 {
2311         struct rb_node **p = &root->rb_node;
2312         struct rb_node *parent = NULL;
2313         struct sa_defrag_extent_backref *entry;
2314         int ret;
2315
2316         while (*p) {
2317                 parent = *p;
2318                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2319
2320                 ret = backref_comp(backref, entry);
2321                 if (ret < 0)
2322                         p = &(*p)->rb_left;
2323                 else
2324                         p = &(*p)->rb_right;
2325         }
2326
2327         rb_link_node(&backref->node, parent, p);
2328         rb_insert_color(&backref->node, root);
2329 }
2330
2331 /*
2332  * Note the backref might has changed, and in this case we just return 0.
2333  */
2334 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2335                                        void *ctx)
2336 {
2337         struct btrfs_file_extent_item *extent;
2338         struct old_sa_defrag_extent *old = ctx;
2339         struct new_sa_defrag_extent *new = old->new;
2340         struct btrfs_path *path = new->path;
2341         struct btrfs_key key;
2342         struct btrfs_root *root;
2343         struct sa_defrag_extent_backref *backref;
2344         struct extent_buffer *leaf;
2345         struct inode *inode = new->inode;
2346         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2347         int slot;
2348         int ret;
2349         u64 extent_offset;
2350         u64 num_bytes;
2351
2352         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2353             inum == btrfs_ino(BTRFS_I(inode)))
2354                 return 0;
2355
2356         key.objectid = root_id;
2357         key.type = BTRFS_ROOT_ITEM_KEY;
2358         key.offset = (u64)-1;
2359
2360         root = btrfs_read_fs_root_no_name(fs_info, &key);
2361         if (IS_ERR(root)) {
2362                 if (PTR_ERR(root) == -ENOENT)
2363                         return 0;
2364                 WARN_ON(1);
2365                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2366                          inum, offset, root_id);
2367                 return PTR_ERR(root);
2368         }
2369
2370         key.objectid = inum;
2371         key.type = BTRFS_EXTENT_DATA_KEY;
2372         if (offset > (u64)-1 << 32)
2373                 key.offset = 0;
2374         else
2375                 key.offset = offset;
2376
2377         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2378         if (WARN_ON(ret < 0))
2379                 return ret;
2380         ret = 0;
2381
2382         while (1) {
2383                 cond_resched();
2384
2385                 leaf = path->nodes[0];
2386                 slot = path->slots[0];
2387
2388                 if (slot >= btrfs_header_nritems(leaf)) {
2389                         ret = btrfs_next_leaf(root, path);
2390                         if (ret < 0) {
2391                                 goto out;
2392                         } else if (ret > 0) {
2393                                 ret = 0;
2394                                 goto out;
2395                         }
2396                         continue;
2397                 }
2398
2399                 path->slots[0]++;
2400
2401                 btrfs_item_key_to_cpu(leaf, &key, slot);
2402
2403                 if (key.objectid > inum)
2404                         goto out;
2405
2406                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2407                         continue;
2408
2409                 extent = btrfs_item_ptr(leaf, slot,
2410                                         struct btrfs_file_extent_item);
2411
2412                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2413                         continue;
2414
2415                 /*
2416                  * 'offset' refers to the exact key.offset,
2417                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2418                  * (key.offset - extent_offset).
2419                  */
2420                 if (key.offset != offset)
2421                         continue;
2422
2423                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2424                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2425
2426                 if (extent_offset >= old->extent_offset + old->offset +
2427                     old->len || extent_offset + num_bytes <=
2428                     old->extent_offset + old->offset)
2429                         continue;
2430                 break;
2431         }
2432
2433         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2434         if (!backref) {
2435                 ret = -ENOENT;
2436                 goto out;
2437         }
2438
2439         backref->root_id = root_id;
2440         backref->inum = inum;
2441         backref->file_pos = offset;
2442         backref->num_bytes = num_bytes;
2443         backref->extent_offset = extent_offset;
2444         backref->generation = btrfs_file_extent_generation(leaf, extent);
2445         backref->old = old;
2446         backref_insert(&new->root, backref);
2447         old->count++;
2448 out:
2449         btrfs_release_path(path);
2450         WARN_ON(ret);
2451         return ret;
2452 }
2453
2454 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2455                                    struct new_sa_defrag_extent *new)
2456 {
2457         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2458         struct old_sa_defrag_extent *old, *tmp;
2459         int ret;
2460
2461         new->path = path;
2462
2463         list_for_each_entry_safe(old, tmp, &new->head, list) {
2464                 ret = iterate_inodes_from_logical(old->bytenr +
2465                                                   old->extent_offset, fs_info,
2466                                                   path, record_one_backref,
2467                                                   old, false);
2468                 if (ret < 0 && ret != -ENOENT)
2469                         return false;
2470
2471                 /* no backref to be processed for this extent */
2472                 if (!old->count) {
2473                         list_del(&old->list);
2474                         kfree(old);
2475                 }
2476         }
2477
2478         if (list_empty(&new->head))
2479                 return false;
2480
2481         return true;
2482 }
2483
2484 static int relink_is_mergable(struct extent_buffer *leaf,
2485                               struct btrfs_file_extent_item *fi,
2486                               struct new_sa_defrag_extent *new)
2487 {
2488         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2489                 return 0;
2490
2491         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2492                 return 0;
2493
2494         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2495                 return 0;
2496
2497         if (btrfs_file_extent_encryption(leaf, fi) ||
2498             btrfs_file_extent_other_encoding(leaf, fi))
2499                 return 0;
2500
2501         return 1;
2502 }
2503
2504 /*
2505  * Note the backref might has changed, and in this case we just return 0.
2506  */
2507 static noinline int relink_extent_backref(struct btrfs_path *path,
2508                                  struct sa_defrag_extent_backref *prev,
2509                                  struct sa_defrag_extent_backref *backref)
2510 {
2511         struct btrfs_file_extent_item *extent;
2512         struct btrfs_file_extent_item *item;
2513         struct btrfs_ordered_extent *ordered;
2514         struct btrfs_trans_handle *trans;
2515         struct btrfs_root *root;
2516         struct btrfs_key key;
2517         struct extent_buffer *leaf;
2518         struct old_sa_defrag_extent *old = backref->old;
2519         struct new_sa_defrag_extent *new = old->new;
2520         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2521         struct inode *inode;
2522         struct extent_state *cached = NULL;
2523         int ret = 0;
2524         u64 start;
2525         u64 len;
2526         u64 lock_start;
2527         u64 lock_end;
2528         bool merge = false;
2529         int index;
2530
2531         if (prev && prev->root_id == backref->root_id &&
2532             prev->inum == backref->inum &&
2533             prev->file_pos + prev->num_bytes == backref->file_pos)
2534                 merge = true;
2535
2536         /* step 1: get root */
2537         key.objectid = backref->root_id;
2538         key.type = BTRFS_ROOT_ITEM_KEY;
2539         key.offset = (u64)-1;
2540
2541         index = srcu_read_lock(&fs_info->subvol_srcu);
2542
2543         root = btrfs_read_fs_root_no_name(fs_info, &key);
2544         if (IS_ERR(root)) {
2545                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2546                 if (PTR_ERR(root) == -ENOENT)
2547                         return 0;
2548                 return PTR_ERR(root);
2549         }
2550
2551         if (btrfs_root_readonly(root)) {
2552                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2553                 return 0;
2554         }
2555
2556         /* step 2: get inode */
2557         key.objectid = backref->inum;
2558         key.type = BTRFS_INODE_ITEM_KEY;
2559         key.offset = 0;
2560
2561         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2562         if (IS_ERR(inode)) {
2563                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2564                 return 0;
2565         }
2566
2567         srcu_read_unlock(&fs_info->subvol_srcu, index);
2568
2569         /* step 3: relink backref */
2570         lock_start = backref->file_pos;
2571         lock_end = backref->file_pos + backref->num_bytes - 1;
2572         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2573                          &cached);
2574
2575         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2576         if (ordered) {
2577                 btrfs_put_ordered_extent(ordered);
2578                 goto out_unlock;
2579         }
2580
2581         trans = btrfs_join_transaction(root);
2582         if (IS_ERR(trans)) {
2583                 ret = PTR_ERR(trans);
2584                 goto out_unlock;
2585         }
2586
2587         key.objectid = backref->inum;
2588         key.type = BTRFS_EXTENT_DATA_KEY;
2589         key.offset = backref->file_pos;
2590
2591         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2592         if (ret < 0) {
2593                 goto out_free_path;
2594         } else if (ret > 0) {
2595                 ret = 0;
2596                 goto out_free_path;
2597         }
2598
2599         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2600                                 struct btrfs_file_extent_item);
2601
2602         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2603             backref->generation)
2604                 goto out_free_path;
2605
2606         btrfs_release_path(path);
2607
2608         start = backref->file_pos;
2609         if (backref->extent_offset < old->extent_offset + old->offset)
2610                 start += old->extent_offset + old->offset -
2611                          backref->extent_offset;
2612
2613         len = min(backref->extent_offset + backref->num_bytes,
2614                   old->extent_offset + old->offset + old->len);
2615         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2616
2617         ret = btrfs_drop_extents(trans, root, inode, start,
2618                                  start + len, 1);
2619         if (ret)
2620                 goto out_free_path;
2621 again:
2622         key.objectid = btrfs_ino(BTRFS_I(inode));
2623         key.type = BTRFS_EXTENT_DATA_KEY;
2624         key.offset = start;
2625
2626         path->leave_spinning = 1;
2627         if (merge) {
2628                 struct btrfs_file_extent_item *fi;
2629                 u64 extent_len;
2630                 struct btrfs_key found_key;
2631
2632                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2633                 if (ret < 0)
2634                         goto out_free_path;
2635
2636                 path->slots[0]--;
2637                 leaf = path->nodes[0];
2638                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2639
2640                 fi = btrfs_item_ptr(leaf, path->slots[0],
2641                                     struct btrfs_file_extent_item);
2642                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2643
2644                 if (extent_len + found_key.offset == start &&
2645                     relink_is_mergable(leaf, fi, new)) {
2646                         btrfs_set_file_extent_num_bytes(leaf, fi,
2647                                                         extent_len + len);
2648                         btrfs_mark_buffer_dirty(leaf);
2649                         inode_add_bytes(inode, len);
2650
2651                         ret = 1;
2652                         goto out_free_path;
2653                 } else {
2654                         merge = false;
2655                         btrfs_release_path(path);
2656                         goto again;
2657                 }
2658         }
2659
2660         ret = btrfs_insert_empty_item(trans, root, path, &key,
2661                                         sizeof(*extent));
2662         if (ret) {
2663                 btrfs_abort_transaction(trans, ret);
2664                 goto out_free_path;
2665         }
2666
2667         leaf = path->nodes[0];
2668         item = btrfs_item_ptr(leaf, path->slots[0],
2669                                 struct btrfs_file_extent_item);
2670         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2671         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2672         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2673         btrfs_set_file_extent_num_bytes(leaf, item, len);
2674         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2675         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2676         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2677         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2678         btrfs_set_file_extent_encryption(leaf, item, 0);
2679         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2680
2681         btrfs_mark_buffer_dirty(leaf);
2682         inode_add_bytes(inode, len);
2683         btrfs_release_path(path);
2684
2685         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2686                         new->disk_len, 0,
2687                         backref->root_id, backref->inum,
2688                         new->file_pos); /* start - extent_offset */
2689         if (ret) {
2690                 btrfs_abort_transaction(trans, ret);
2691                 goto out_free_path;
2692         }
2693
2694         ret = 1;
2695 out_free_path:
2696         btrfs_release_path(path);
2697         path->leave_spinning = 0;
2698         btrfs_end_transaction(trans);
2699 out_unlock:
2700         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2701                              &cached, GFP_NOFS);
2702         iput(inode);
2703         return ret;
2704 }
2705
2706 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2707 {
2708         struct old_sa_defrag_extent *old, *tmp;
2709
2710         if (!new)
2711                 return;
2712
2713         list_for_each_entry_safe(old, tmp, &new->head, list) {
2714                 kfree(old);
2715         }
2716         kfree(new);
2717 }
2718
2719 static void relink_file_extents(struct new_sa_defrag_extent *new)
2720 {
2721         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2722         struct btrfs_path *path;
2723         struct sa_defrag_extent_backref *backref;
2724         struct sa_defrag_extent_backref *prev = NULL;
2725         struct inode *inode;
2726         struct btrfs_root *root;
2727         struct rb_node *node;
2728         int ret;
2729
2730         inode = new->inode;
2731         root = BTRFS_I(inode)->root;
2732
2733         path = btrfs_alloc_path();
2734         if (!path)
2735                 return;
2736
2737         if (!record_extent_backrefs(path, new)) {
2738                 btrfs_free_path(path);
2739                 goto out;
2740         }
2741         btrfs_release_path(path);
2742
2743         while (1) {
2744                 node = rb_first(&new->root);
2745                 if (!node)
2746                         break;
2747                 rb_erase(node, &new->root);
2748
2749                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2750
2751                 ret = relink_extent_backref(path, prev, backref);
2752                 WARN_ON(ret < 0);
2753
2754                 kfree(prev);
2755
2756                 if (ret == 1)
2757                         prev = backref;
2758                 else
2759                         prev = NULL;
2760                 cond_resched();
2761         }
2762         kfree(prev);
2763
2764         btrfs_free_path(path);
2765 out:
2766         free_sa_defrag_extent(new);
2767
2768         atomic_dec(&fs_info->defrag_running);
2769         wake_up(&fs_info->transaction_wait);
2770 }
2771
2772 static struct new_sa_defrag_extent *
2773 record_old_file_extents(struct inode *inode,
2774                         struct btrfs_ordered_extent *ordered)
2775 {
2776         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2777         struct btrfs_root *root = BTRFS_I(inode)->root;
2778         struct btrfs_path *path;
2779         struct btrfs_key key;
2780         struct old_sa_defrag_extent *old;
2781         struct new_sa_defrag_extent *new;
2782         int ret;
2783
2784         new = kmalloc(sizeof(*new), GFP_NOFS);
2785         if (!new)
2786                 return NULL;
2787
2788         new->inode = inode;
2789         new->file_pos = ordered->file_offset;
2790         new->len = ordered->len;
2791         new->bytenr = ordered->start;
2792         new->disk_len = ordered->disk_len;
2793         new->compress_type = ordered->compress_type;
2794         new->root = RB_ROOT;
2795         INIT_LIST_HEAD(&new->head);
2796
2797         path = btrfs_alloc_path();
2798         if (!path)
2799                 goto out_kfree;
2800
2801         key.objectid = btrfs_ino(BTRFS_I(inode));
2802         key.type = BTRFS_EXTENT_DATA_KEY;
2803         key.offset = new->file_pos;
2804
2805         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2806         if (ret < 0)
2807                 goto out_free_path;
2808         if (ret > 0 && path->slots[0] > 0)
2809                 path->slots[0]--;
2810
2811         /* find out all the old extents for the file range */
2812         while (1) {
2813                 struct btrfs_file_extent_item *extent;
2814                 struct extent_buffer *l;
2815                 int slot;
2816                 u64 num_bytes;
2817                 u64 offset;
2818                 u64 end;
2819                 u64 disk_bytenr;
2820                 u64 extent_offset;
2821
2822                 l = path->nodes[0];
2823                 slot = path->slots[0];
2824
2825                 if (slot >= btrfs_header_nritems(l)) {
2826                         ret = btrfs_next_leaf(root, path);
2827                         if (ret < 0)
2828                                 goto out_free_path;
2829                         else if (ret > 0)
2830                                 break;
2831                         continue;
2832                 }
2833
2834                 btrfs_item_key_to_cpu(l, &key, slot);
2835
2836                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2837                         break;
2838                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2839                         break;
2840                 if (key.offset >= new->file_pos + new->len)
2841                         break;
2842
2843                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2844
2845                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2846                 if (key.offset + num_bytes < new->file_pos)
2847                         goto next;
2848
2849                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2850                 if (!disk_bytenr)
2851                         goto next;
2852
2853                 extent_offset = btrfs_file_extent_offset(l, extent);
2854
2855                 old = kmalloc(sizeof(*old), GFP_NOFS);
2856                 if (!old)
2857                         goto out_free_path;
2858
2859                 offset = max(new->file_pos, key.offset);
2860                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2861
2862                 old->bytenr = disk_bytenr;
2863                 old->extent_offset = extent_offset;
2864                 old->offset = offset - key.offset;
2865                 old->len = end - offset;
2866                 old->new = new;
2867                 old->count = 0;
2868                 list_add_tail(&old->list, &new->head);
2869 next:
2870                 path->slots[0]++;
2871                 cond_resched();
2872         }
2873
2874         btrfs_free_path(path);
2875         atomic_inc(&fs_info->defrag_running);
2876
2877         return new;
2878
2879 out_free_path:
2880         btrfs_free_path(path);
2881 out_kfree:
2882         free_sa_defrag_extent(new);
2883         return NULL;
2884 }
2885
2886 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2887                                          u64 start, u64 len)
2888 {
2889         struct btrfs_block_group_cache *cache;
2890
2891         cache = btrfs_lookup_block_group(fs_info, start);
2892         ASSERT(cache);
2893
2894         spin_lock(&cache->lock);
2895         cache->delalloc_bytes -= len;
2896         spin_unlock(&cache->lock);
2897
2898         btrfs_put_block_group(cache);
2899 }
2900
2901 /* as ordered data IO finishes, this gets called so we can finish
2902  * an ordered extent if the range of bytes in the file it covers are
2903  * fully written.
2904  */
2905 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2906 {
2907         struct inode *inode = ordered_extent->inode;
2908         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2909         struct btrfs_root *root = BTRFS_I(inode)->root;
2910         struct btrfs_trans_handle *trans = NULL;
2911         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2912         struct extent_state *cached_state = NULL;
2913         struct new_sa_defrag_extent *new = NULL;
2914         int compress_type = 0;
2915         int ret = 0;
2916         u64 logical_len = ordered_extent->len;
2917         bool nolock;
2918         bool truncated = false;
2919         bool range_locked = false;
2920         bool clear_new_delalloc_bytes = false;
2921
2922         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2923             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
2924             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
2925                 clear_new_delalloc_bytes = true;
2926
2927         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
2928
2929         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2930                 ret = -EIO;
2931                 goto out;
2932         }
2933
2934         btrfs_free_io_failure_record(BTRFS_I(inode),
2935                         ordered_extent->file_offset,
2936                         ordered_extent->file_offset +
2937                         ordered_extent->len - 1);
2938
2939         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2940                 truncated = true;
2941                 logical_len = ordered_extent->truncated_len;
2942                 /* Truncated the entire extent, don't bother adding */
2943                 if (!logical_len)
2944                         goto out;
2945         }
2946
2947         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2948                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2949
2950                 /*
2951                  * For mwrite(mmap + memset to write) case, we still reserve
2952                  * space for NOCOW range.
2953                  * As NOCOW won't cause a new delayed ref, just free the space
2954                  */
2955                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
2956                                        ordered_extent->len);
2957                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2958                 if (nolock)
2959                         trans = btrfs_join_transaction_nolock(root);
2960                 else
2961                         trans = btrfs_join_transaction(root);
2962                 if (IS_ERR(trans)) {
2963                         ret = PTR_ERR(trans);
2964                         trans = NULL;
2965                         goto out;
2966                 }
2967                 trans->block_rsv = &BTRFS_I(inode)->block_rsv;
2968                 ret = btrfs_update_inode_fallback(trans, root, inode);
2969                 if (ret) /* -ENOMEM or corruption */
2970                         btrfs_abort_transaction(trans, ret);
2971                 goto out;
2972         }
2973
2974         range_locked = true;
2975         lock_extent_bits(io_tree, ordered_extent->file_offset,
2976                          ordered_extent->file_offset + ordered_extent->len - 1,
2977                          &cached_state);
2978
2979         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2980                         ordered_extent->file_offset + ordered_extent->len - 1,
2981                         EXTENT_DEFRAG, 0, cached_state);
2982         if (ret) {
2983                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2984                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
2985                         /* the inode is shared */
2986                         new = record_old_file_extents(inode, ordered_extent);
2987
2988                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2989                         ordered_extent->file_offset + ordered_extent->len - 1,
2990                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2991         }
2992
2993         if (nolock)
2994                 trans = btrfs_join_transaction_nolock(root);
2995         else
2996                 trans = btrfs_join_transaction(root);
2997         if (IS_ERR(trans)) {
2998                 ret = PTR_ERR(trans);
2999                 trans = NULL;
3000                 goto out;
3001         }
3002
3003         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3004
3005         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
3006                 compress_type = ordered_extent->compress_type;
3007         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
3008                 BUG_ON(compress_type);
3009                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3010                                        ordered_extent->len);
3011                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
3012                                                 ordered_extent->file_offset,
3013                                                 ordered_extent->file_offset +
3014                                                 logical_len);
3015         } else {
3016                 BUG_ON(root == fs_info->tree_root);
3017                 ret = insert_reserved_file_extent(trans, inode,
3018                                                 ordered_extent->file_offset,
3019                                                 ordered_extent->start,
3020                                                 ordered_extent->disk_len,
3021                                                 logical_len, logical_len,
3022                                                 compress_type, 0, 0,
3023                                                 BTRFS_FILE_EXTENT_REG);
3024                 if (!ret)
3025                         btrfs_release_delalloc_bytes(fs_info,
3026                                                      ordered_extent->start,
3027                                                      ordered_extent->disk_len);
3028         }
3029         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
3030                            ordered_extent->file_offset, ordered_extent->len,
3031                            trans->transid);
3032         if (ret < 0) {
3033                 btrfs_abort_transaction(trans, ret);
3034                 goto out;
3035         }
3036
3037         add_pending_csums(trans, inode, &ordered_extent->list);
3038
3039         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3040         ret = btrfs_update_inode_fallback(trans, root, inode);
3041         if (ret) { /* -ENOMEM or corruption */
3042                 btrfs_abort_transaction(trans, ret);
3043                 goto out;
3044         }
3045         ret = 0;
3046 out:
3047         if (range_locked || clear_new_delalloc_bytes) {
3048                 unsigned int clear_bits = 0;
3049
3050                 if (range_locked)
3051                         clear_bits |= EXTENT_LOCKED;
3052                 if (clear_new_delalloc_bytes)
3053                         clear_bits |= EXTENT_DELALLOC_NEW;
3054                 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3055                                  ordered_extent->file_offset,
3056                                  ordered_extent->file_offset +
3057                                  ordered_extent->len - 1,
3058                                  clear_bits,
3059                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
3060                                  0, &cached_state, GFP_NOFS);
3061         }
3062
3063         if (trans)
3064                 btrfs_end_transaction(trans);
3065
3066         if (ret || truncated) {
3067                 u64 start, end;
3068
3069                 if (truncated)
3070                         start = ordered_extent->file_offset + logical_len;
3071                 else
3072                         start = ordered_extent->file_offset;
3073                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3074                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
3075
3076                 /* Drop the cache for the part of the extent we didn't write. */
3077                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3078
3079                 /*
3080                  * If the ordered extent had an IOERR or something else went
3081                  * wrong we need to return the space for this ordered extent
3082                  * back to the allocator.  We only free the extent in the
3083                  * truncated case if we didn't write out the extent at all.
3084                  */
3085                 if ((ret || !logical_len) &&
3086                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3087                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3088                         btrfs_free_reserved_extent(fs_info,
3089                                                    ordered_extent->start,
3090                                                    ordered_extent->disk_len, 1);
3091         }
3092
3093
3094         /*
3095          * This needs to be done to make sure anybody waiting knows we are done
3096          * updating everything for this ordered extent.
3097          */
3098         btrfs_remove_ordered_extent(inode, ordered_extent);
3099
3100         /* for snapshot-aware defrag */
3101         if (new) {
3102                 if (ret) {
3103                         free_sa_defrag_extent(new);
3104                         atomic_dec(&fs_info->defrag_running);
3105                 } else {
3106                         relink_file_extents(new);
3107                 }
3108         }
3109
3110         /* once for us */
3111         btrfs_put_ordered_extent(ordered_extent);
3112         /* once for the tree */
3113         btrfs_put_ordered_extent(ordered_extent);
3114
3115         return ret;
3116 }
3117
3118 static void finish_ordered_fn(struct btrfs_work *work)
3119 {
3120         struct btrfs_ordered_extent *ordered_extent;
3121         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3122         btrfs_finish_ordered_io(ordered_extent);
3123 }
3124
3125 static void btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3126                                 struct extent_state *state, int uptodate)
3127 {
3128         struct inode *inode = page->mapping->host;
3129         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3130         struct btrfs_ordered_extent *ordered_extent = NULL;
3131         struct btrfs_workqueue *wq;
3132         btrfs_work_func_t func;
3133
3134         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3135
3136         ClearPagePrivate2(page);
3137         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3138                                             end - start + 1, uptodate))
3139                 return;
3140
3141         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
3142                 wq = fs_info->endio_freespace_worker;
3143                 func = btrfs_freespace_write_helper;
3144         } else {
3145                 wq = fs_info->endio_write_workers;
3146                 func = btrfs_endio_write_helper;
3147         }
3148
3149         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3150                         NULL);
3151         btrfs_queue_work(wq, &ordered_extent->work);
3152 }
3153
3154 static int __readpage_endio_check(struct inode *inode,
3155                                   struct btrfs_io_bio *io_bio,
3156                                   int icsum, struct page *page,
3157                                   int pgoff, u64 start, size_t len)
3158 {
3159         char *kaddr;
3160         u32 csum_expected;
3161         u32 csum = ~(u32)0;
3162
3163         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3164
3165         kaddr = kmap_atomic(page);
3166         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3167         btrfs_csum_final(csum, (u8 *)&csum);
3168         if (csum != csum_expected)
3169                 goto zeroit;
3170
3171         kunmap_atomic(kaddr);
3172         return 0;
3173 zeroit:
3174         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
3175                                     io_bio->mirror_num);
3176         memset(kaddr + pgoff, 1, len);
3177         flush_dcache_page(page);
3178         kunmap_atomic(kaddr);
3179         return -EIO;
3180 }
3181
3182 /*
3183  * when reads are done, we need to check csums to verify the data is correct
3184  * if there's a match, we allow the bio to finish.  If not, the code in
3185  * extent_io.c will try to find good copies for us.
3186  */
3187 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3188                                       u64 phy_offset, struct page *page,
3189                                       u64 start, u64 end, int mirror)
3190 {
3191         size_t offset = start - page_offset(page);
3192         struct inode *inode = page->mapping->host;
3193         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3194         struct btrfs_root *root = BTRFS_I(inode)->root;
3195
3196         if (PageChecked(page)) {
3197                 ClearPageChecked(page);
3198                 return 0;
3199         }
3200
3201         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3202                 return 0;
3203
3204         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3205             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3206                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3207                 return 0;
3208         }
3209
3210         phy_offset >>= inode->i_sb->s_blocksize_bits;
3211         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3212                                       start, (size_t)(end - start + 1));
3213 }
3214
3215 void btrfs_add_delayed_iput(struct inode *inode)
3216 {
3217         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3218         struct btrfs_inode *binode = BTRFS_I(inode);
3219
3220         if (atomic_add_unless(&inode->i_count, -1, 1))
3221                 return;
3222
3223         spin_lock(&fs_info->delayed_iput_lock);
3224         if (binode->delayed_iput_count == 0) {
3225                 ASSERT(list_empty(&binode->delayed_iput));
3226                 list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3227         } else {
3228                 binode->delayed_iput_count++;
3229         }
3230         spin_unlock(&fs_info->delayed_iput_lock);
3231 }
3232
3233 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3234 {
3235
3236         spin_lock(&fs_info->delayed_iput_lock);
3237         while (!list_empty(&fs_info->delayed_iputs)) {
3238                 struct btrfs_inode *inode;
3239
3240                 inode = list_first_entry(&fs_info->delayed_iputs,
3241                                 struct btrfs_inode, delayed_iput);
3242                 if (inode->delayed_iput_count) {
3243                         inode->delayed_iput_count--;
3244                         list_move_tail(&inode->delayed_iput,
3245                                         &fs_info->delayed_iputs);
3246                 } else {
3247                         list_del_init(&inode->delayed_iput);
3248                 }
3249                 spin_unlock(&fs_info->delayed_iput_lock);
3250                 iput(&inode->vfs_inode);
3251                 spin_lock(&fs_info->delayed_iput_lock);
3252         }
3253         spin_unlock(&fs_info->delayed_iput_lock);
3254 }
3255
3256 /*
3257  * This is called in transaction commit time. If there are no orphan
3258  * files in the subvolume, it removes orphan item and frees block_rsv
3259  * structure.
3260  */
3261 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3262                               struct btrfs_root *root)
3263 {
3264         struct btrfs_fs_info *fs_info = root->fs_info;
3265         struct btrfs_block_rsv *block_rsv;
3266         int ret;
3267
3268         if (atomic_read(&root->orphan_inodes) ||
3269             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3270                 return;
3271
3272         spin_lock(&root->orphan_lock);
3273         if (atomic_read(&root->orphan_inodes)) {
3274                 spin_unlock(&root->orphan_lock);
3275                 return;
3276         }
3277
3278         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3279                 spin_unlock(&root->orphan_lock);
3280                 return;
3281         }
3282
3283         block_rsv = root->orphan_block_rsv;
3284         root->orphan_block_rsv = NULL;
3285         spin_unlock(&root->orphan_lock);
3286
3287         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3288             btrfs_root_refs(&root->root_item) > 0) {
3289                 ret = btrfs_del_orphan_item(trans, fs_info->tree_root,
3290                                             root->root_key.objectid);
3291                 if (ret)
3292                         btrfs_abort_transaction(trans, ret);
3293                 else
3294                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3295                                   &root->state);
3296         }
3297
3298         if (block_rsv) {
3299                 WARN_ON(block_rsv->size > 0);
3300                 btrfs_free_block_rsv(fs_info, block_rsv);
3301         }
3302 }
3303
3304 /*
3305  * This creates an orphan entry for the given inode in case something goes
3306  * wrong in the middle of an unlink/truncate.
3307  *
3308  * NOTE: caller of this function should reserve 5 units of metadata for
3309  *       this function.
3310  */
3311 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
3312                 struct btrfs_inode *inode)
3313 {
3314         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
3315         struct btrfs_root *root = inode->root;
3316         struct btrfs_block_rsv *block_rsv = NULL;
3317         int reserve = 0;
3318         int insert = 0;
3319         int ret;
3320
3321         if (!root->orphan_block_rsv) {
3322                 block_rsv = btrfs_alloc_block_rsv(fs_info,
3323                                                   BTRFS_BLOCK_RSV_TEMP);
3324                 if (!block_rsv)
3325                         return -ENOMEM;
3326         }
3327
3328         spin_lock(&root->orphan_lock);
3329         if (!root->orphan_block_rsv) {
3330                 root->orphan_block_rsv = block_rsv;
3331         } else if (block_rsv) {
3332                 btrfs_free_block_rsv(fs_info, block_rsv);
3333                 block_rsv = NULL;
3334         }
3335
3336         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3337                               &inode->runtime_flags)) {
3338 #if 0
3339                 /*
3340                  * For proper ENOSPC handling, we should do orphan
3341                  * cleanup when mounting. But this introduces backward
3342                  * compatibility issue.
3343                  */
3344                 if (!xchg(&root->orphan_item_inserted, 1))
3345                         insert = 2;
3346                 else
3347                         insert = 1;
3348 #endif
3349                 insert = 1;
3350                 atomic_inc(&root->orphan_inodes);
3351         }
3352
3353         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3354                               &inode->runtime_flags))
3355                 reserve = 1;
3356         spin_unlock(&root->orphan_lock);
3357
3358         /* grab metadata reservation from transaction handle */
3359         if (reserve) {
3360                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3361                 ASSERT(!ret);
3362                 if (ret) {
3363                         atomic_dec(&root->orphan_inodes);
3364                         clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3365                                   &inode->runtime_flags);
3366                         if (insert)
3367                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3368                                           &inode->runtime_flags);
3369                         return ret;
3370                 }
3371         }
3372
3373         /* insert an orphan item to track this unlinked/truncated file */
3374         if (insert >= 1) {
3375                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3376                 if (ret) {
3377                         atomic_dec(&root->orphan_inodes);
3378                         if (reserve) {
3379                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3380                                           &inode->runtime_flags);
3381                                 btrfs_orphan_release_metadata(inode);
3382                         }
3383                         if (ret != -EEXIST) {
3384                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3385                                           &inode->runtime_flags);
3386                                 btrfs_abort_transaction(trans, ret);
3387                                 return ret;
3388                         }
3389                 }
3390                 ret = 0;
3391         }
3392
3393         /* insert an orphan item to track subvolume contains orphan files */
3394         if (insert >= 2) {
3395                 ret = btrfs_insert_orphan_item(trans, fs_info->tree_root,
3396                                                root->root_key.objectid);
3397                 if (ret && ret != -EEXIST) {
3398                         btrfs_abort_transaction(trans, ret);
3399                         return ret;
3400                 }
3401         }
3402         return 0;
3403 }
3404
3405 /*
3406  * We have done the truncate/delete so we can go ahead and remove the orphan
3407  * item for this particular inode.
3408  */
3409 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3410                             struct btrfs_inode *inode)
3411 {
3412         struct btrfs_root *root = inode->root;
3413         int delete_item = 0;
3414         int release_rsv = 0;
3415         int ret = 0;
3416
3417         spin_lock(&root->orphan_lock);
3418         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3419                                &inode->runtime_flags))
3420                 delete_item = 1;
3421
3422         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3423                                &inode->runtime_flags))
3424                 release_rsv = 1;
3425         spin_unlock(&root->orphan_lock);
3426
3427         if (delete_item) {
3428                 atomic_dec(&root->orphan_inodes);
3429                 if (trans)
3430                         ret = btrfs_del_orphan_item(trans, root,
3431                                                     btrfs_ino(inode));
3432         }
3433
3434         if (release_rsv)
3435                 btrfs_orphan_release_metadata(inode);
3436
3437         return ret;
3438 }
3439
3440 /*
3441  * this cleans up any orphans that may be left on the list from the last use
3442  * of this root.
3443  */
3444 int btrfs_orphan_cleanup(struct btrfs_root *root)
3445 {
3446         struct btrfs_fs_info *fs_info = root->fs_info;
3447         struct btrfs_path *path;
3448         struct extent_buffer *leaf;
3449         struct btrfs_key key, found_key;
3450         struct btrfs_trans_handle *trans;
3451         struct inode *inode;
3452         u64 last_objectid = 0;
3453         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3454
3455         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3456                 return 0;
3457
3458         path = btrfs_alloc_path();
3459         if (!path) {
3460                 ret = -ENOMEM;
3461                 goto out;
3462         }
3463         path->reada = READA_BACK;
3464
3465         key.objectid = BTRFS_ORPHAN_OBJECTID;
3466         key.type = BTRFS_ORPHAN_ITEM_KEY;
3467         key.offset = (u64)-1;
3468
3469         while (1) {
3470                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3471                 if (ret < 0)
3472                         goto out;
3473
3474                 /*
3475                  * if ret == 0 means we found what we were searching for, which
3476                  * is weird, but possible, so only screw with path if we didn't
3477                  * find the key and see if we have stuff that matches
3478                  */
3479                 if (ret > 0) {
3480                         ret = 0;
3481                         if (path->slots[0] == 0)
3482                                 break;
3483                         path->slots[0]--;
3484                 }
3485
3486                 /* pull out the item */
3487                 leaf = path->nodes[0];
3488                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3489
3490                 /* make sure the item matches what we want */
3491                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3492                         break;
3493                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3494                         break;
3495
3496                 /* release the path since we're done with it */
3497                 btrfs_release_path(path);
3498
3499                 /*
3500                  * this is where we are basically btrfs_lookup, without the
3501                  * crossing root thing.  we store the inode number in the
3502                  * offset of the orphan item.
3503                  */
3504
3505                 if (found_key.offset == last_objectid) {
3506                         btrfs_err(fs_info,
3507                                   "Error removing orphan entry, stopping orphan cleanup");
3508                         ret = -EINVAL;
3509                         goto out;
3510                 }
3511
3512                 last_objectid = found_key.offset;
3513
3514                 found_key.objectid = found_key.offset;
3515                 found_key.type = BTRFS_INODE_ITEM_KEY;
3516                 found_key.offset = 0;
3517                 inode = btrfs_iget(fs_info->sb, &found_key, root, NULL);
3518                 ret = PTR_ERR_OR_ZERO(inode);
3519                 if (ret && ret != -ENOENT)
3520                         goto out;
3521
3522                 if (ret == -ENOENT && root == fs_info->tree_root) {
3523                         struct btrfs_root *dead_root;
3524                         struct btrfs_fs_info *fs_info = root->fs_info;
3525                         int is_dead_root = 0;
3526
3527                         /*
3528                          * this is an orphan in the tree root. Currently these
3529                          * could come from 2 sources:
3530                          *  a) a snapshot deletion in progress
3531                          *  b) a free space cache inode
3532                          * We need to distinguish those two, as the snapshot
3533                          * orphan must not get deleted.
3534                          * find_dead_roots already ran before us, so if this
3535                          * is a snapshot deletion, we should find the root
3536                          * in the dead_roots list
3537                          */
3538                         spin_lock(&fs_info->trans_lock);
3539                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3540                                             root_list) {
3541                                 if (dead_root->root_key.objectid ==
3542                                     found_key.objectid) {
3543                                         is_dead_root = 1;
3544                                         break;
3545                                 }
3546                         }
3547                         spin_unlock(&fs_info->trans_lock);
3548                         if (is_dead_root) {
3549                                 /* prevent this orphan from being found again */
3550                                 key.offset = found_key.objectid - 1;
3551                                 continue;
3552                         }
3553                 }
3554                 /*
3555                  * Inode is already gone but the orphan item is still there,
3556                  * kill the orphan item.
3557                  */
3558                 if (ret == -ENOENT) {
3559                         trans = btrfs_start_transaction(root, 1);
3560                         if (IS_ERR(trans)) {
3561                                 ret = PTR_ERR(trans);
3562                                 goto out;
3563                         }
3564                         btrfs_debug(fs_info, "auto deleting %Lu",
3565                                     found_key.objectid);
3566                         ret = btrfs_del_orphan_item(trans, root,
3567                                                     found_key.objectid);
3568                         btrfs_end_transaction(trans);
3569                         if (ret)
3570                                 goto out;
3571                         continue;
3572                 }
3573
3574                 /*
3575                  * add this inode to the orphan list so btrfs_orphan_del does
3576                  * the proper thing when we hit it
3577                  */
3578                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3579                         &BTRFS_I(inode)->runtime_flags);
3580                 atomic_inc(&root->orphan_inodes);
3581
3582                 /* if we have links, this was a truncate, lets do that */
3583                 if (inode->i_nlink) {
3584                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3585                                 iput(inode);
3586                                 continue;
3587                         }
3588                         nr_truncate++;
3589
3590                         /* 1 for the orphan item deletion. */
3591                         trans = btrfs_start_transaction(root, 1);
3592                         if (IS_ERR(trans)) {
3593                                 iput(inode);
3594                                 ret = PTR_ERR(trans);
3595                                 goto out;
3596                         }
3597                         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
3598                         btrfs_end_transaction(trans);
3599                         if (ret) {
3600                                 iput(inode);
3601                                 goto out;
3602                         }
3603
3604                         ret = btrfs_truncate(inode);
3605                         if (ret)
3606                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
3607                 } else {
3608                         nr_unlink++;
3609                 }
3610
3611                 /* this will do delete_inode and everything for us */
3612                 iput(inode);
3613                 if (ret)
3614                         goto out;
3615         }
3616         /* release the path since we're done with it */
3617         btrfs_release_path(path);
3618
3619         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3620
3621         if (root->orphan_block_rsv)
3622                 btrfs_block_rsv_release(fs_info, root->orphan_block_rsv,
3623                                         (u64)-1);
3624
3625         if (root->orphan_block_rsv ||
3626             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3627                 trans = btrfs_join_transaction(root);
3628                 if (!IS_ERR(trans))
3629                         btrfs_end_transaction(trans);
3630         }
3631
3632         if (nr_unlink)
3633                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3634         if (nr_truncate)
3635                 btrfs_debug(fs_info, "truncated %d orphans", nr_truncate);
3636
3637 out:
3638         if (ret)
3639                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3640         btrfs_free_path(path);
3641         return ret;
3642 }
3643
3644 /*
3645  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3646  * don't find any xattrs, we know there can't be any acls.
3647  *
3648  * slot is the slot the inode is in, objectid is the objectid of the inode
3649  */
3650 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3651                                           int slot, u64 objectid,
3652                                           int *first_xattr_slot)
3653 {
3654         u32 nritems = btrfs_header_nritems(leaf);
3655         struct btrfs_key found_key;
3656         static u64 xattr_access = 0;
3657         static u64 xattr_default = 0;
3658         int scanned = 0;
3659
3660         if (!xattr_access) {
3661                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3662                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3663                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3664                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3665         }
3666
3667         slot++;
3668         *first_xattr_slot = -1;
3669         while (slot < nritems) {
3670                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3671
3672                 /* we found a different objectid, there must not be acls */
3673                 if (found_key.objectid != objectid)
3674                         return 0;
3675
3676                 /* we found an xattr, assume we've got an acl */
3677                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3678                         if (*first_xattr_slot == -1)
3679                                 *first_xattr_slot = slot;
3680                         if (found_key.offset == xattr_access ||
3681                             found_key.offset == xattr_default)
3682                                 return 1;
3683                 }
3684
3685                 /*
3686                  * we found a key greater than an xattr key, there can't
3687                  * be any acls later on
3688                  */
3689                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3690                         return 0;
3691
3692                 slot++;
3693                 scanned++;
3694
3695                 /*
3696                  * it goes inode, inode backrefs, xattrs, extents,
3697                  * so if there are a ton of hard links to an inode there can
3698                  * be a lot of backrefs.  Don't waste time searching too hard,
3699                  * this is just an optimization
3700                  */
3701                 if (scanned >= 8)
3702                         break;
3703         }
3704         /* we hit the end of the leaf before we found an xattr or
3705          * something larger than an xattr.  We have to assume the inode
3706          * has acls
3707          */
3708         if (*first_xattr_slot == -1)
3709                 *first_xattr_slot = slot;
3710         return 1;
3711 }
3712
3713 /*
3714  * read an inode from the btree into the in-memory inode
3715  */
3716 static int btrfs_read_locked_inode(struct inode *inode)
3717 {
3718         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3719         struct btrfs_path *path;
3720         struct extent_buffer *leaf;
3721         struct btrfs_inode_item *inode_item;
3722         struct btrfs_root *root = BTRFS_I(inode)->root;
3723         struct btrfs_key location;
3724         unsigned long ptr;
3725         int maybe_acls;
3726         u32 rdev;
3727         int ret;
3728         bool filled = false;
3729         int first_xattr_slot;
3730
3731         ret = btrfs_fill_inode(inode, &rdev);
3732         if (!ret)
3733                 filled = true;
3734
3735         path = btrfs_alloc_path();
3736         if (!path) {
3737                 ret = -ENOMEM;
3738                 goto make_bad;
3739         }
3740
3741         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3742
3743         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3744         if (ret) {
3745                 if (ret > 0)
3746                         ret = -ENOENT;
3747                 goto make_bad;
3748         }
3749
3750         leaf = path->nodes[0];
3751
3752         if (filled)
3753                 goto cache_index;
3754
3755         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3756                                     struct btrfs_inode_item);
3757         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3758         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3759         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3760         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3761         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3762
3763         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3764         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3765
3766         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3767         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3768
3769         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3770         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3771
3772         BTRFS_I(inode)->i_otime.tv_sec =
3773                 btrfs_timespec_sec(leaf, &inode_item->otime);
3774         BTRFS_I(inode)->i_otime.tv_nsec =
3775                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3776
3777         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3778         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3779         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3780
3781         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3782         inode->i_generation = BTRFS_I(inode)->generation;
3783         inode->i_rdev = 0;
3784         rdev = btrfs_inode_rdev(leaf, inode_item);
3785
3786         BTRFS_I(inode)->index_cnt = (u64)-1;
3787         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3788
3789 cache_index:
3790         /*
3791          * If we were modified in the current generation and evicted from memory
3792          * and then re-read we need to do a full sync since we don't have any
3793          * idea about which extents were modified before we were evicted from
3794          * cache.
3795          *
3796          * This is required for both inode re-read from disk and delayed inode
3797          * in delayed_nodes_tree.
3798          */
3799         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3800                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3801                         &BTRFS_I(inode)->runtime_flags);
3802
3803         /*
3804          * We don't persist the id of the transaction where an unlink operation
3805          * against the inode was last made. So here we assume the inode might
3806          * have been evicted, and therefore the exact value of last_unlink_trans
3807          * lost, and set it to last_trans to avoid metadata inconsistencies
3808          * between the inode and its parent if the inode is fsync'ed and the log
3809          * replayed. For example, in the scenario:
3810          *
3811          * touch mydir/foo
3812          * ln mydir/foo mydir/bar
3813          * sync
3814          * unlink mydir/bar
3815          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3816          * xfs_io -c fsync mydir/foo
3817          * <power failure>
3818          * mount fs, triggers fsync log replay
3819          *
3820          * We must make sure that when we fsync our inode foo we also log its
3821          * parent inode, otherwise after log replay the parent still has the
3822          * dentry with the "bar" name but our inode foo has a link count of 1
3823          * and doesn't have an inode ref with the name "bar" anymore.
3824          *
3825          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3826          * but it guarantees correctness at the expense of occasional full
3827          * transaction commits on fsync if our inode is a directory, or if our
3828          * inode is not a directory, logging its parent unnecessarily.
3829          */
3830         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3831
3832         path->slots[0]++;
3833         if (inode->i_nlink != 1 ||
3834             path->slots[0] >= btrfs_header_nritems(leaf))
3835                 goto cache_acl;
3836
3837         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3838         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3839                 goto cache_acl;
3840
3841         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3842         if (location.type == BTRFS_INODE_REF_KEY) {
3843                 struct btrfs_inode_ref *ref;
3844
3845                 ref = (struct btrfs_inode_ref *)ptr;
3846                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3847         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3848                 struct btrfs_inode_extref *extref;
3849
3850                 extref = (struct btrfs_inode_extref *)ptr;
3851                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3852                                                                      extref);
3853         }
3854 cache_acl:
3855         /*
3856          * try to precache a NULL acl entry for files that don't have
3857          * any xattrs or acls
3858          */
3859         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3860                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3861         if (first_xattr_slot != -1) {
3862                 path->slots[0] = first_xattr_slot;
3863                 ret = btrfs_load_inode_props(inode, path);
3864                 if (ret)
3865                         btrfs_err(fs_info,
3866                                   "error loading props for ino %llu (root %llu): %d",
3867                                   btrfs_ino(BTRFS_I(inode)),
3868                                   root->root_key.objectid, ret);
3869         }
3870         btrfs_free_path(path);
3871
3872         if (!maybe_acls)
3873                 cache_no_acl(inode);
3874
3875         switch (inode->i_mode & S_IFMT) {
3876         case S_IFREG:
3877                 inode->i_mapping->a_ops = &btrfs_aops;
3878                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3879                 inode->i_fop = &btrfs_file_operations;
3880                 inode->i_op = &btrfs_file_inode_operations;
3881                 break;
3882         case S_IFDIR:
3883                 inode->i_fop = &btrfs_dir_file_operations;
3884                 inode->i_op = &btrfs_dir_inode_operations;
3885                 break;
3886         case S_IFLNK:
3887                 inode->i_op = &btrfs_symlink_inode_operations;
3888                 inode_nohighmem(inode);
3889                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3890                 break;
3891         default:
3892                 inode->i_op = &btrfs_special_inode_operations;
3893                 init_special_inode(inode, inode->i_mode, rdev);
3894                 break;
3895         }
3896
3897         btrfs_update_iflags(inode);
3898         return 0;
3899
3900 make_bad:
3901         btrfs_free_path(path);
3902         make_bad_inode(inode);
3903         return ret;
3904 }
3905
3906 /*
3907  * given a leaf and an inode, copy the inode fields into the leaf
3908  */
3909 static void fill_inode_item(struct btrfs_trans_handle *trans,
3910                             struct extent_buffer *leaf,
3911                             struct btrfs_inode_item *item,
3912                             struct inode *inode)
3913 {
3914         struct btrfs_map_token token;
3915
3916         btrfs_init_map_token(&token);
3917
3918         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3919         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3920         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3921                                    &token);
3922         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3923         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3924
3925         btrfs_set_token_timespec_sec(leaf, &item->atime,
3926                                      inode->i_atime.tv_sec, &token);
3927         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3928                                       inode->i_atime.tv_nsec, &token);
3929
3930         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3931                                      inode->i_mtime.tv_sec, &token);
3932         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3933                                       inode->i_mtime.tv_nsec, &token);
3934
3935         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3936                                      inode->i_ctime.tv_sec, &token);
3937         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3938                                       inode->i_ctime.tv_nsec, &token);
3939
3940         btrfs_set_token_timespec_sec(leaf, &item->otime,
3941                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3942         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3943                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3944
3945         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3946                                      &token);
3947         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3948                                          &token);
3949         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3950         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3951         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3952         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3953         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3954 }
3955
3956 /*
3957  * copy everything in the in-memory inode into the btree.
3958  */
3959 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3960                                 struct btrfs_root *root, struct inode *inode)
3961 {
3962         struct btrfs_inode_item *inode_item;
3963         struct btrfs_path *path;
3964         struct extent_buffer *leaf;
3965         int ret;
3966
3967         path = btrfs_alloc_path();
3968         if (!path)
3969                 return -ENOMEM;
3970
3971         path->leave_spinning = 1;
3972         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3973                                  1);
3974         if (ret) {
3975                 if (ret > 0)
3976                         ret = -ENOENT;
3977                 goto failed;
3978         }
3979
3980         leaf = path->nodes[0];
3981         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3982                                     struct btrfs_inode_item);
3983
3984         fill_inode_item(trans, leaf, inode_item, inode);
3985         btrfs_mark_buffer_dirty(leaf);
3986         btrfs_set_inode_last_trans(trans, inode);
3987         ret = 0;
3988 failed:
3989         btrfs_free_path(path);
3990         return ret;
3991 }
3992
3993 /*
3994  * copy everything in the in-memory inode into the btree.
3995  */
3996 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3997                                 struct btrfs_root *root, struct inode *inode)
3998 {
3999         struct btrfs_fs_info *fs_info = root->fs_info;
4000         int ret;
4001
4002         /*
4003          * If the inode is a free space inode, we can deadlock during commit
4004          * if we put it into the delayed code.
4005          *
4006          * The data relocation inode should also be directly updated
4007          * without delay
4008          */
4009         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
4010             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
4011             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
4012                 btrfs_update_root_times(trans, root);
4013
4014                 ret = btrfs_delayed_update_inode(trans, root, inode);
4015                 if (!ret)
4016                         btrfs_set_inode_last_trans(trans, inode);
4017                 return ret;
4018         }
4019
4020         return btrfs_update_inode_item(trans, root, inode);
4021 }
4022
4023 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
4024                                          struct btrfs_root *root,
4025                                          struct inode *inode)
4026 {
4027         int ret;
4028
4029         ret = btrfs_update_inode(trans, root, inode);
4030         if (ret == -ENOSPC)
4031                 return btrfs_update_inode_item(trans, root, inode);
4032         return ret;
4033 }
4034
4035 /*
4036  * unlink helper that gets used here in inode.c and in the tree logging
4037  * recovery code.  It remove a link in a directory with a given name, and
4038  * also drops the back refs in the inode to the directory
4039  */
4040 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4041                                 struct btrfs_root *root,
4042                                 struct btrfs_inode *dir,
4043                                 struct btrfs_inode *inode,
4044                                 const char *name, int name_len)
4045 {
4046         struct btrfs_fs_info *fs_info = root->fs_info;
4047         struct btrfs_path *path;
4048         int ret = 0;
4049         struct extent_buffer *leaf;
4050         struct btrfs_dir_item *di;
4051         struct btrfs_key key;
4052         u64 index;
4053         u64 ino = btrfs_ino(inode);
4054         u64 dir_ino = btrfs_ino(dir);
4055
4056         path = btrfs_alloc_path();
4057         if (!path) {
4058                 ret = -ENOMEM;
4059                 goto out;
4060         }
4061
4062         path->leave_spinning = 1;
4063         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4064                                     name, name_len, -1);
4065         if (IS_ERR(di)) {
4066                 ret = PTR_ERR(di);
4067                 goto err;
4068         }
4069         if (!di) {
4070                 ret = -ENOENT;
4071                 goto err;
4072         }
4073         leaf = path->nodes[0];
4074         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4075         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4076         if (ret)
4077                 goto err;
4078         btrfs_release_path(path);
4079
4080         /*
4081          * If we don't have dir index, we have to get it by looking up
4082          * the inode ref, since we get the inode ref, remove it directly,
4083          * it is unnecessary to do delayed deletion.
4084          *
4085          * But if we have dir index, needn't search inode ref to get it.
4086          * Since the inode ref is close to the inode item, it is better
4087          * that we delay to delete it, and just do this deletion when
4088          * we update the inode item.
4089          */
4090         if (inode->dir_index) {
4091                 ret = btrfs_delayed_delete_inode_ref(inode);
4092                 if (!ret) {
4093                         index = inode->dir_index;
4094                         goto skip_backref;
4095                 }
4096         }
4097
4098         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4099                                   dir_ino, &index);
4100         if (ret) {
4101                 btrfs_info(fs_info,
4102                         "failed to delete reference to %.*s, inode %llu parent %llu",
4103                         name_len, name, ino, dir_ino);
4104                 btrfs_abort_transaction(trans, ret);
4105                 goto err;
4106         }
4107 skip_backref:
4108         ret = btrfs_delete_delayed_dir_index(trans, fs_info, dir, index);
4109         if (ret) {
4110                 btrfs_abort_transaction(trans, ret);
4111                 goto err;
4112         }
4113
4114         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
4115                         dir_ino);
4116         if (ret != 0 && ret != -ENOENT) {
4117                 btrfs_abort_transaction(trans, ret);
4118                 goto err;
4119         }
4120
4121         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
4122                         index);
4123         if (ret == -ENOENT)
4124                 ret = 0;
4125         else if (ret)
4126                 btrfs_abort_transaction(trans, ret);
4127 err:
4128         btrfs_free_path(path);
4129         if (ret)
4130                 goto out;
4131
4132         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
4133         inode_inc_iversion(&inode->vfs_inode);
4134         inode_inc_iversion(&dir->vfs_inode);
4135         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
4136                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
4137         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
4138 out:
4139         return ret;
4140 }
4141
4142 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4143                        struct btrfs_root *root,
4144                        struct btrfs_inode *dir, struct btrfs_inode *inode,
4145                        const char *name, int name_len)
4146 {
4147         int ret;
4148         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4149         if (!ret) {
4150                 drop_nlink(&inode->vfs_inode);
4151                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
4152         }
4153         return ret;
4154 }
4155
4156 /*
4157  * helper to start transaction for unlink and rmdir.
4158  *
4159  * unlink and rmdir are special in btrfs, they do not always free space, so
4160  * if we cannot make our reservations the normal way try and see if there is
4161  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4162  * allow the unlink to occur.
4163  */
4164 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4165 {
4166         struct btrfs_root *root = BTRFS_I(dir)->root;
4167
4168         /*
4169          * 1 for the possible orphan item
4170          * 1 for the dir item
4171          * 1 for the dir index
4172          * 1 for the inode ref
4173          * 1 for the inode
4174          */
4175         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4176 }
4177
4178 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4179 {
4180         struct btrfs_root *root = BTRFS_I(dir)->root;
4181         struct btrfs_trans_handle *trans;
4182         struct inode *inode = d_inode(dentry);
4183         int ret;
4184
4185         trans = __unlink_start_trans(dir);
4186         if (IS_ERR(trans))
4187                 return PTR_ERR(trans);
4188
4189         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
4190                         0);
4191
4192         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4193                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4194                         dentry->d_name.len);
4195         if (ret)
4196                 goto out;
4197
4198         if (inode->i_nlink == 0) {
4199                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
4200                 if (ret)
4201                         goto out;
4202         }
4203
4204 out:
4205         btrfs_end_transaction(trans);
4206         btrfs_btree_balance_dirty(root->fs_info);
4207         return ret;
4208 }
4209
4210 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4211                         struct btrfs_root *root,
4212                         struct inode *dir, u64 objectid,
4213                         const char *name, int name_len)
4214 {
4215         struct btrfs_fs_info *fs_info = root->fs_info;
4216         struct btrfs_path *path;
4217         struct extent_buffer *leaf;
4218         struct btrfs_dir_item *di;
4219         struct btrfs_key key;
4220         u64 index;
4221         int ret;
4222         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
4223
4224         path = btrfs_alloc_path();
4225         if (!path)
4226                 return -ENOMEM;
4227
4228         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4229                                    name, name_len, -1);
4230         if (IS_ERR_OR_NULL(di)) {
4231                 if (!di)
4232                         ret = -ENOENT;
4233                 else
4234                         ret = PTR_ERR(di);
4235                 goto out;
4236         }
4237
4238         leaf = path->nodes[0];
4239         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4240         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4241         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4242         if (ret) {
4243                 btrfs_abort_transaction(trans, ret);
4244                 goto out;
4245         }
4246         btrfs_release_path(path);
4247
4248         ret = btrfs_del_root_ref(trans, fs_info, objectid,
4249                                  root->root_key.objectid, dir_ino,
4250                                  &index, name, name_len);
4251         if (ret < 0) {
4252                 if (ret != -ENOENT) {
4253                         btrfs_abort_transaction(trans, ret);
4254                         goto out;
4255                 }
4256                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4257                                                  name, name_len);
4258                 if (IS_ERR_OR_NULL(di)) {
4259                         if (!di)
4260                                 ret = -ENOENT;
4261                         else
4262                                 ret = PTR_ERR(di);
4263                         btrfs_abort_transaction(trans, ret);
4264                         goto out;
4265                 }
4266
4267                 leaf = path->nodes[0];
4268                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4269                 btrfs_release_path(path);
4270                 index = key.offset;
4271         }
4272         btrfs_release_path(path);
4273
4274         ret = btrfs_delete_delayed_dir_index(trans, fs_info, BTRFS_I(dir), index);
4275         if (ret) {
4276                 btrfs_abort_transaction(trans, ret);
4277                 goto out;
4278         }
4279
4280         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
4281         inode_inc_iversion(dir);
4282         dir->i_mtime = dir->i_ctime = current_time(dir);
4283         ret = btrfs_update_inode_fallback(trans, root, dir);
4284         if (ret)
4285                 btrfs_abort_transaction(trans, ret);
4286 out:
4287         btrfs_free_path(path);
4288         return ret;
4289 }
4290
4291 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4292 {
4293         struct inode *inode = d_inode(dentry);
4294         int err = 0;
4295         struct btrfs_root *root = BTRFS_I(dir)->root;
4296         struct btrfs_trans_handle *trans;
4297         u64 last_unlink_trans;
4298
4299         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4300                 return -ENOTEMPTY;
4301         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4302                 return -EPERM;
4303
4304         trans = __unlink_start_trans(dir);
4305         if (IS_ERR(trans))
4306                 return PTR_ERR(trans);
4307
4308         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4309                 err = btrfs_unlink_subvol(trans, root, dir,
4310                                           BTRFS_I(inode)->location.objectid,
4311                                           dentry->d_name.name,
4312                                           dentry->d_name.len);
4313                 goto out;
4314         }
4315
4316         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4317         if (err)
4318                 goto out;
4319
4320         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4321
4322         /* now the directory is empty */
4323         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4324                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4325                         dentry->d_name.len);
4326         if (!err) {
4327                 btrfs_i_size_write(BTRFS_I(inode), 0);
4328                 /*
4329                  * Propagate the last_unlink_trans value of the deleted dir to
4330                  * its parent directory. This is to prevent an unrecoverable
4331                  * log tree in the case we do something like this:
4332                  * 1) create dir foo
4333                  * 2) create snapshot under dir foo
4334                  * 3) delete the snapshot
4335                  * 4) rmdir foo
4336                  * 5) mkdir foo
4337                  * 6) fsync foo or some file inside foo
4338                  */
4339                 if (last_unlink_trans >= trans->transid)
4340                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4341         }
4342 out:
4343         btrfs_end_transaction(trans);
4344         btrfs_btree_balance_dirty(root->fs_info);
4345
4346         return err;
4347 }
4348
4349 static int truncate_space_check(struct btrfs_trans_handle *trans,
4350                                 struct btrfs_root *root,
4351                                 u64 bytes_deleted)
4352 {
4353         struct btrfs_fs_info *fs_info = root->fs_info;
4354         int ret;
4355
4356         /*
4357          * This is only used to apply pressure to the enospc system, we don't
4358          * intend to use this reservation at all.
4359          */
4360         bytes_deleted = btrfs_csum_bytes_to_leaves(fs_info, bytes_deleted);
4361         bytes_deleted *= fs_info->nodesize;
4362         ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
4363                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4364         if (!ret) {
4365                 trace_btrfs_space_reservation(fs_info, "transaction",
4366                                               trans->transid,
4367                                               bytes_deleted, 1);
4368                 trans->bytes_reserved += bytes_deleted;
4369         }
4370         return ret;
4371
4372 }
4373
4374 /*
4375  * Return this if we need to call truncate_block for the last bit of the
4376  * truncate.
4377  */
4378 #define NEED_TRUNCATE_BLOCK 1
4379
4380 /*
4381  * this can truncate away extent items, csum items and directory items.
4382  * It starts at a high offset and removes keys until it can't find
4383  * any higher than new_size
4384  *
4385  * csum items that cross the new i_size are truncated to the new size
4386  * as well.
4387  *
4388  * min_type is the minimum key type to truncate down to.  If set to 0, this
4389  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4390  */
4391 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4392                                struct btrfs_root *root,
4393                                struct inode *inode,
4394                                u64 new_size, u32 min_type)
4395 {
4396         struct btrfs_fs_info *fs_info = root->fs_info;
4397         struct btrfs_path *path;
4398         struct extent_buffer *leaf;
4399         struct btrfs_file_extent_item *fi;
4400         struct btrfs_key key;
4401         struct btrfs_key found_key;
4402         u64 extent_start = 0;
4403         u64 extent_num_bytes = 0;
4404         u64 extent_offset = 0;
4405         u64 item_end = 0;
4406         u64 last_size = new_size;
4407         u32 found_type = (u8)-1;
4408         int found_extent;
4409         int del_item;
4410         int pending_del_nr = 0;
4411         int pending_del_slot = 0;
4412         int extent_type = -1;
4413         int ret;
4414         int err = 0;
4415         u64 ino = btrfs_ino(BTRFS_I(inode));
4416         u64 bytes_deleted = 0;
4417         bool be_nice = false;
4418         bool should_throttle = false;
4419         bool should_end = false;
4420
4421         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4422
4423         /*
4424          * for non-free space inodes and ref cows, we want to back off from
4425          * time to time
4426          */
4427         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4428             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4429                 be_nice = true;
4430
4431         path = btrfs_alloc_path();
4432         if (!path)
4433                 return -ENOMEM;
4434         path->reada = READA_BACK;
4435
4436         /*
4437          * We want to drop from the next block forward in case this new size is
4438          * not block aligned since we will be keeping the last block of the
4439          * extent just the way it is.
4440          */
4441         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4442             root == fs_info->tree_root)
4443                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4444                                         fs_info->sectorsize),
4445                                         (u64)-1, 0);
4446
4447         /*
4448          * This function is also used to drop the items in the log tree before
4449          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4450          * it is used to drop the loged items. So we shouldn't kill the delayed
4451          * items.
4452          */
4453         if (min_type == 0 && root == BTRFS_I(inode)->root)
4454                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4455
4456         key.objectid = ino;
4457         key.offset = (u64)-1;
4458         key.type = (u8)-1;
4459
4460 search_again:
4461         /*
4462          * with a 16K leaf size and 128MB extents, you can actually queue
4463          * up a huge file in a single leaf.  Most of the time that
4464          * bytes_deleted is > 0, it will be huge by the time we get here
4465          */
4466         if (be_nice && bytes_deleted > SZ_32M) {
4467                 if (btrfs_should_end_transaction(trans)) {
4468                         err = -EAGAIN;
4469                         goto error;
4470                 }
4471         }
4472
4473
4474         path->leave_spinning = 1;
4475         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4476         if (ret < 0) {
4477                 err = ret;
4478                 goto out;
4479         }
4480
4481         if (ret > 0) {
4482                 /* there are no items in the tree for us to truncate, we're
4483                  * done
4484                  */
4485                 if (path->slots[0] == 0)
4486                         goto out;
4487                 path->slots[0]--;
4488         }
4489
4490         while (1) {
4491                 fi = NULL;
4492                 leaf = path->nodes[0];
4493                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4494                 found_type = found_key.type;
4495
4496                 if (found_key.objectid != ino)
4497                         break;
4498
4499                 if (found_type < min_type)
4500                         break;
4501
4502                 item_end = found_key.offset;
4503                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4504                         fi = btrfs_item_ptr(leaf, path->slots[0],
4505                                             struct btrfs_file_extent_item);
4506                         extent_type = btrfs_file_extent_type(leaf, fi);
4507                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4508                                 item_end +=
4509                                     btrfs_file_extent_num_bytes(leaf, fi);
4510
4511                                 trace_btrfs_truncate_show_fi_regular(
4512                                         BTRFS_I(inode), leaf, fi,
4513                                         found_key.offset);
4514                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4515                                 item_end += btrfs_file_extent_inline_len(leaf,
4516                                                          path->slots[0], fi);
4517
4518                                 trace_btrfs_truncate_show_fi_inline(
4519                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4520                                         found_key.offset);
4521                         }
4522                         item_end--;
4523                 }
4524                 if (found_type > min_type) {
4525                         del_item = 1;
4526                 } else {
4527                         if (item_end < new_size)
4528                                 break;
4529                         if (found_key.offset >= new_size)
4530                                 del_item = 1;
4531                         else
4532                                 del_item = 0;
4533                 }
4534                 found_extent = 0;
4535                 /* FIXME, shrink the extent if the ref count is only 1 */
4536                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4537                         goto delete;
4538
4539                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4540                         u64 num_dec;
4541                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4542                         if (!del_item) {
4543                                 u64 orig_num_bytes =
4544                                         btrfs_file_extent_num_bytes(leaf, fi);
4545                                 extent_num_bytes = ALIGN(new_size -
4546                                                 found_key.offset,
4547                                                 fs_info->sectorsize);
4548                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4549                                                          extent_num_bytes);
4550                                 num_dec = (orig_num_bytes -
4551                                            extent_num_bytes);
4552                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4553                                              &root->state) &&
4554                                     extent_start != 0)
4555                                         inode_sub_bytes(inode, num_dec);
4556                                 btrfs_mark_buffer_dirty(leaf);
4557                         } else {
4558                                 extent_num_bytes =
4559                                         btrfs_file_extent_disk_num_bytes(leaf,
4560                                                                          fi);
4561                                 extent_offset = found_key.offset -
4562                                         btrfs_file_extent_offset(leaf, fi);
4563
4564                                 /* FIXME blocksize != 4096 */
4565                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4566                                 if (extent_start != 0) {
4567                                         found_extent = 1;
4568                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4569                                                      &root->state))
4570                                                 inode_sub_bytes(inode, num_dec);
4571                                 }
4572                         }
4573                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4574                         /*
4575                          * we can't truncate inline items that have had
4576                          * special encodings
4577                          */
4578                         if (!del_item &&
4579                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4580                             btrfs_file_extent_other_encoding(leaf, fi) == 0 &&
4581                             btrfs_file_extent_compression(leaf, fi) == 0) {
4582                                 u32 size = (u32)(new_size - found_key.offset);
4583
4584                                 btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4585                                 size = btrfs_file_extent_calc_inline_size(size);
4586                                 btrfs_truncate_item(root->fs_info, path, size, 1);
4587                         } else if (!del_item) {
4588                                 /*
4589                                  * We have to bail so the last_size is set to
4590                                  * just before this extent.
4591                                  */
4592                                 err = NEED_TRUNCATE_BLOCK;
4593                                 break;
4594                         }
4595
4596                         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4597                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4598                 }
4599 delete:
4600                 if (del_item)
4601                         last_size = found_key.offset;
4602                 else
4603                         last_size = new_size;
4604                 if (del_item) {
4605                         if (!pending_del_nr) {
4606                                 /* no pending yet, add ourselves */
4607                                 pending_del_slot = path->slots[0];
4608                                 pending_del_nr = 1;
4609                         } else if (pending_del_nr &&
4610                                    path->slots[0] + 1 == pending_del_slot) {
4611                                 /* hop on the pending chunk */
4612                                 pending_del_nr++;
4613                                 pending_del_slot = path->slots[0];
4614                         } else {
4615                                 BUG();
4616                         }
4617                 } else {
4618                         break;
4619                 }
4620                 should_throttle = false;
4621
4622                 if (found_extent &&
4623                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4624                      root == fs_info->tree_root)) {
4625                         btrfs_set_path_blocking(path);
4626                         bytes_deleted += extent_num_bytes;
4627                         ret = btrfs_free_extent(trans, root, extent_start,
4628                                                 extent_num_bytes, 0,
4629                                                 btrfs_header_owner(leaf),
4630                                                 ino, extent_offset);
4631                         BUG_ON(ret);
4632                         if (btrfs_should_throttle_delayed_refs(trans, fs_info))
4633                                 btrfs_async_run_delayed_refs(fs_info,
4634                                         trans->delayed_ref_updates * 2,
4635                                         trans->transid, 0);
4636                         if (be_nice) {
4637                                 if (truncate_space_check(trans, root,
4638                                                          extent_num_bytes)) {
4639                                         should_end = true;
4640                                 }
4641                                 if (btrfs_should_throttle_delayed_refs(trans,
4642                                                                        fs_info))
4643                                         should_throttle = true;
4644                         }
4645                 }
4646
4647                 if (found_type == BTRFS_INODE_ITEM_KEY)
4648                         break;
4649
4650                 if (path->slots[0] == 0 ||
4651                     path->slots[0] != pending_del_slot ||
4652                     should_throttle || should_end) {
4653                         if (pending_del_nr) {
4654                                 ret = btrfs_del_items(trans, root, path,
4655                                                 pending_del_slot,
4656                                                 pending_del_nr);
4657                                 if (ret) {
4658                                         btrfs_abort_transaction(trans, ret);
4659                                         goto error;
4660                                 }
4661                                 pending_del_nr = 0;
4662                         }
4663                         btrfs_release_path(path);
4664                         if (should_throttle) {
4665                                 unsigned long updates = trans->delayed_ref_updates;
4666                                 if (updates) {
4667                                         trans->delayed_ref_updates = 0;
4668                                         ret = btrfs_run_delayed_refs(trans,
4669                                                                    fs_info,
4670                                                                    updates * 2);
4671                                         if (ret && !err)
4672                                                 err = ret;
4673                                 }
4674                         }
4675                         /*
4676                          * if we failed to refill our space rsv, bail out
4677                          * and let the transaction restart
4678                          */
4679                         if (should_end) {
4680                                 err = -EAGAIN;
4681                                 goto error;
4682                         }
4683                         goto search_again;
4684                 } else {
4685                         path->slots[0]--;
4686                 }
4687         }
4688 out:
4689         if (pending_del_nr) {
4690                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4691                                       pending_del_nr);
4692                 if (ret)
4693                         btrfs_abort_transaction(trans, ret);
4694         }
4695 error:
4696         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4697                 ASSERT(last_size >= new_size);
4698                 if (!err && last_size > new_size)
4699                         last_size = new_size;
4700                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4701         }
4702
4703         btrfs_free_path(path);
4704
4705         if (be_nice && bytes_deleted > SZ_32M) {
4706                 unsigned long updates = trans->delayed_ref_updates;
4707                 if (updates) {
4708                         trans->delayed_ref_updates = 0;
4709                         ret = btrfs_run_delayed_refs(trans, fs_info,
4710                                                      updates * 2);
4711                         if (ret && !err)
4712                                 err = ret;
4713                 }
4714         }
4715         return err;
4716 }
4717
4718 /*
4719  * btrfs_truncate_block - read, zero a chunk and write a block
4720  * @inode - inode that we're zeroing
4721  * @from - the offset to start zeroing
4722  * @len - the length to zero, 0 to zero the entire range respective to the
4723  *      offset
4724  * @front - zero up to the offset instead of from the offset on
4725  *
4726  * This will find the block for the "from" offset and cow the block and zero the
4727  * part we want to zero.  This is used with truncate and hole punching.
4728  */
4729 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4730                         int front)
4731 {
4732         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4733         struct address_space *mapping = inode->i_mapping;
4734         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4735         struct btrfs_ordered_extent *ordered;
4736         struct extent_state *cached_state = NULL;
4737         struct extent_changeset *data_reserved = NULL;
4738         char *kaddr;
4739         u32 blocksize = fs_info->sectorsize;
4740         pgoff_t index = from >> PAGE_SHIFT;
4741         unsigned offset = from & (blocksize - 1);
4742         struct page *page;
4743         gfp_t mask = btrfs_alloc_write_mask(mapping);
4744         int ret = 0;
4745         u64 block_start;
4746         u64 block_end;
4747
4748         if ((offset & (blocksize - 1)) == 0 &&
4749             (!len || ((len & (blocksize - 1)) == 0)))
4750                 goto out;
4751
4752         block_start = round_down(from, blocksize);
4753         block_end = block_start + blocksize - 1;
4754
4755         ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
4756                                            block_start, blocksize);
4757         if (ret)
4758                 goto out;
4759
4760 again:
4761         page = find_or_create_page(mapping, index, mask);
4762         if (!page) {
4763                 btrfs_delalloc_release_space(inode, data_reserved,
4764                                              block_start, blocksize);
4765                 btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
4766                 ret = -ENOMEM;
4767                 goto out;
4768         }
4769
4770         if (!PageUptodate(page)) {
4771                 ret = btrfs_readpage(NULL, page);
4772                 lock_page(page);
4773                 if (page->mapping != mapping) {
4774                         unlock_page(page);
4775                         put_page(page);
4776                         goto again;
4777                 }
4778                 if (!PageUptodate(page)) {
4779                         ret = -EIO;
4780                         goto out_unlock;
4781                 }
4782         }
4783         wait_on_page_writeback(page);
4784
4785         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4786         set_page_extent_mapped(page);
4787
4788         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4789         if (ordered) {
4790                 unlock_extent_cached(io_tree, block_start, block_end,
4791                                      &cached_state, GFP_NOFS);
4792                 unlock_page(page);
4793                 put_page(page);
4794                 btrfs_start_ordered_extent(inode, ordered, 1);
4795                 btrfs_put_ordered_extent(ordered);
4796                 goto again;
4797         }
4798
4799         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4800                           EXTENT_DIRTY | EXTENT_DELALLOC |
4801                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4802                           0, 0, &cached_state, GFP_NOFS);
4803
4804         ret = btrfs_set_extent_delalloc(inode, block_start, block_end, 0,
4805                                         &cached_state, 0);
4806         if (ret) {
4807                 unlock_extent_cached(io_tree, block_start, block_end,
4808                                      &cached_state, GFP_NOFS);
4809                 goto out_unlock;
4810         }
4811
4812         if (offset != blocksize) {
4813                 if (!len)
4814                         len = blocksize - offset;
4815                 kaddr = kmap(page);
4816                 if (front)
4817                         memset(kaddr + (block_start - page_offset(page)),
4818                                 0, offset);
4819                 else
4820                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4821                                 0, len);
4822                 flush_dcache_page(page);
4823                 kunmap(page);
4824         }
4825         ClearPageChecked(page);
4826         set_page_dirty(page);
4827         unlock_extent_cached(io_tree, block_start, block_end, &cached_state,
4828                              GFP_NOFS);
4829
4830 out_unlock:
4831         if (ret)
4832                 btrfs_delalloc_release_space(inode, data_reserved, block_start,
4833                                              blocksize);
4834         btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
4835         unlock_page(page);
4836         put_page(page);
4837 out:
4838         extent_changeset_free(data_reserved);
4839         return ret;
4840 }
4841
4842 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4843                              u64 offset, u64 len)
4844 {
4845         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4846         struct btrfs_trans_handle *trans;
4847         int ret;
4848
4849         /*
4850          * Still need to make sure the inode looks like it's been updated so
4851          * that any holes get logged if we fsync.
4852          */
4853         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
4854                 BTRFS_I(inode)->last_trans = fs_info->generation;
4855                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4856                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4857                 return 0;
4858         }
4859
4860         /*
4861          * 1 - for the one we're dropping
4862          * 1 - for the one we're adding
4863          * 1 - for updating the inode.
4864          */
4865         trans = btrfs_start_transaction(root, 3);
4866         if (IS_ERR(trans))
4867                 return PTR_ERR(trans);
4868
4869         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4870         if (ret) {
4871                 btrfs_abort_transaction(trans, ret);
4872                 btrfs_end_transaction(trans);
4873                 return ret;
4874         }
4875
4876         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
4877                         offset, 0, 0, len, 0, len, 0, 0, 0);
4878         if (ret)
4879                 btrfs_abort_transaction(trans, ret);
4880         else
4881                 btrfs_update_inode(trans, root, inode);
4882         btrfs_end_transaction(trans);
4883         return ret;
4884 }
4885
4886 /*
4887  * This function puts in dummy file extents for the area we're creating a hole
4888  * for.  So if we are truncating this file to a larger size we need to insert
4889  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4890  * the range between oldsize and size
4891  */
4892 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4893 {
4894         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4895         struct btrfs_root *root = BTRFS_I(inode)->root;
4896         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4897         struct extent_map *em = NULL;
4898         struct extent_state *cached_state = NULL;
4899         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4900         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
4901         u64 block_end = ALIGN(size, fs_info->sectorsize);
4902         u64 last_byte;
4903         u64 cur_offset;
4904         u64 hole_size;
4905         int err = 0;
4906
4907         /*
4908          * If our size started in the middle of a block we need to zero out the
4909          * rest of the block before we expand the i_size, otherwise we could
4910          * expose stale data.
4911          */
4912         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4913         if (err)
4914                 return err;
4915
4916         if (size <= hole_start)
4917                 return 0;
4918
4919         while (1) {
4920                 struct btrfs_ordered_extent *ordered;
4921
4922                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4923                                  &cached_state);
4924                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), hole_start,
4925                                                      block_end - hole_start);
4926                 if (!ordered)
4927                         break;
4928                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4929                                      &cached_state, GFP_NOFS);
4930                 btrfs_start_ordered_extent(inode, ordered, 1);
4931                 btrfs_put_ordered_extent(ordered);
4932         }
4933
4934         cur_offset = hole_start;
4935         while (1) {
4936                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
4937                                 block_end - cur_offset, 0);
4938                 if (IS_ERR(em)) {
4939                         err = PTR_ERR(em);
4940                         em = NULL;
4941                         break;
4942                 }
4943                 last_byte = min(extent_map_end(em), block_end);
4944                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
4945                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4946                         struct extent_map *hole_em;
4947                         hole_size = last_byte - cur_offset;
4948
4949                         err = maybe_insert_hole(root, inode, cur_offset,
4950                                                 hole_size);
4951                         if (err)
4952                                 break;
4953                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
4954                                                 cur_offset + hole_size - 1, 0);
4955                         hole_em = alloc_extent_map();
4956                         if (!hole_em) {
4957                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4958                                         &BTRFS_I(inode)->runtime_flags);
4959                                 goto next;
4960                         }
4961                         hole_em->start = cur_offset;
4962                         hole_em->len = hole_size;
4963                         hole_em->orig_start = cur_offset;
4964
4965                         hole_em->block_start = EXTENT_MAP_HOLE;
4966                         hole_em->block_len = 0;
4967                         hole_em->orig_block_len = 0;
4968                         hole_em->ram_bytes = hole_size;
4969                         hole_em->bdev = fs_info->fs_devices->latest_bdev;
4970                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4971                         hole_em->generation = fs_info->generation;
4972
4973                         while (1) {
4974                                 write_lock(&em_tree->lock);
4975                                 err = add_extent_mapping(em_tree, hole_em, 1);
4976                                 write_unlock(&em_tree->lock);
4977                                 if (err != -EEXIST)
4978                                         break;
4979                                 btrfs_drop_extent_cache(BTRFS_I(inode),
4980                                                         cur_offset,
4981                                                         cur_offset +
4982                                                         hole_size - 1, 0);
4983                         }
4984                         free_extent_map(hole_em);
4985                 }
4986 next:
4987                 free_extent_map(em);
4988                 em = NULL;
4989                 cur_offset = last_byte;
4990                 if (cur_offset >= block_end)
4991                         break;
4992         }
4993         free_extent_map(em);
4994         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4995                              GFP_NOFS);
4996         return err;
4997 }
4998
4999 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
5000 {
5001         struct btrfs_root *root = BTRFS_I(inode)->root;
5002         struct btrfs_trans_handle *trans;
5003         loff_t oldsize = i_size_read(inode);
5004         loff_t newsize = attr->ia_size;
5005         int mask = attr->ia_valid;
5006         int ret;
5007
5008         /*
5009          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
5010          * special case where we need to update the times despite not having
5011          * these flags set.  For all other operations the VFS set these flags
5012          * explicitly if it wants a timestamp update.
5013          */
5014         if (newsize != oldsize) {
5015                 inode_inc_iversion(inode);
5016                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
5017                         inode->i_ctime = inode->i_mtime =
5018                                 current_time(inode);
5019         }
5020
5021         if (newsize > oldsize) {
5022                 /*
5023                  * Don't do an expanding truncate while snapshotting is ongoing.
5024                  * This is to ensure the snapshot captures a fully consistent
5025                  * state of this file - if the snapshot captures this expanding
5026                  * truncation, it must capture all writes that happened before
5027                  * this truncation.
5028                  */
5029                 btrfs_wait_for_snapshot_creation(root);
5030                 ret = btrfs_cont_expand(inode, oldsize, newsize);
5031                 if (ret) {
5032                         btrfs_end_write_no_snapshotting(root);
5033                         return ret;
5034                 }
5035
5036                 trans = btrfs_start_transaction(root, 1);
5037                 if (IS_ERR(trans)) {
5038                         btrfs_end_write_no_snapshotting(root);
5039                         return PTR_ERR(trans);
5040                 }
5041
5042                 i_size_write(inode, newsize);
5043                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
5044                 pagecache_isize_extended(inode, oldsize, newsize);
5045                 ret = btrfs_update_inode(trans, root, inode);
5046                 btrfs_end_write_no_snapshotting(root);
5047                 btrfs_end_transaction(trans);
5048         } else {
5049
5050                 /*
5051                  * We're truncating a file that used to have good data down to
5052                  * zero. Make sure it gets into the ordered flush list so that
5053                  * any new writes get down to disk quickly.
5054                  */
5055                 if (newsize == 0)
5056                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5057                                 &BTRFS_I(inode)->runtime_flags);
5058
5059                 /*
5060                  * 1 for the orphan item we're going to add
5061                  * 1 for the orphan item deletion.
5062                  */
5063                 trans = btrfs_start_transaction(root, 2);
5064                 if (IS_ERR(trans))
5065                         return PTR_ERR(trans);
5066
5067                 /*
5068                  * We need to do this in case we fail at _any_ point during the
5069                  * actual truncate.  Once we do the truncate_setsize we could
5070                  * invalidate pages which forces any outstanding ordered io to
5071                  * be instantly completed which will give us extents that need
5072                  * to be truncated.  If we fail to get an orphan inode down we
5073                  * could have left over extents that were never meant to live,
5074                  * so we need to guarantee from this point on that everything
5075                  * will be consistent.
5076                  */
5077                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
5078                 btrfs_end_transaction(trans);
5079                 if (ret)
5080                         return ret;
5081
5082                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
5083                 truncate_setsize(inode, newsize);
5084
5085                 /* Disable nonlocked read DIO to avoid the end less truncate */
5086                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
5087                 inode_dio_wait(inode);
5088                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
5089
5090                 ret = btrfs_truncate(inode);
5091                 if (ret && inode->i_nlink) {
5092                         int err;
5093
5094                         /* To get a stable disk_i_size */
5095                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
5096                         if (err) {
5097                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5098                                 return err;
5099                         }
5100
5101                         /*
5102                          * failed to truncate, disk_i_size is only adjusted down
5103                          * as we remove extents, so it should represent the true
5104                          * size of the inode, so reset the in memory size and
5105                          * delete our orphan entry.
5106                          */
5107                         trans = btrfs_join_transaction(root);
5108                         if (IS_ERR(trans)) {
5109                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5110                                 return ret;
5111                         }
5112                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5113                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
5114                         if (err)
5115                                 btrfs_abort_transaction(trans, err);
5116                         btrfs_end_transaction(trans);
5117                 }
5118         }
5119
5120         return ret;
5121 }
5122
5123 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5124 {
5125         struct inode *inode = d_inode(dentry);
5126         struct btrfs_root *root = BTRFS_I(inode)->root;
5127         int err;
5128
5129         if (btrfs_root_readonly(root))
5130                 return -EROFS;
5131
5132         err = setattr_prepare(dentry, attr);
5133         if (err)
5134                 return err;
5135
5136         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5137                 err = btrfs_setsize(inode, attr);
5138                 if (err)
5139                         return err;
5140         }
5141
5142         if (attr->ia_valid) {
5143                 setattr_copy(inode, attr);
5144                 inode_inc_iversion(inode);
5145                 err = btrfs_dirty_inode(inode);
5146
5147                 if (!err && attr->ia_valid & ATTR_MODE)
5148                         err = posix_acl_chmod(inode, inode->i_mode);
5149         }
5150
5151         return err;
5152 }
5153
5154 /*
5155  * While truncating the inode pages during eviction, we get the VFS calling
5156  * btrfs_invalidatepage() against each page of the inode. This is slow because
5157  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5158  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5159  * extent_state structures over and over, wasting lots of time.
5160  *
5161  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5162  * those expensive operations on a per page basis and do only the ordered io
5163  * finishing, while we release here the extent_map and extent_state structures,
5164  * without the excessive merging and splitting.
5165  */
5166 static void evict_inode_truncate_pages(struct inode *inode)
5167 {
5168         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5169         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5170         struct rb_node *node;
5171
5172         ASSERT(inode->i_state & I_FREEING);
5173         truncate_inode_pages_final(&inode->i_data);
5174
5175         write_lock(&map_tree->lock);
5176         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5177                 struct extent_map *em;
5178
5179                 node = rb_first(&map_tree->map);
5180                 em = rb_entry(node, struct extent_map, rb_node);
5181                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5182                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5183                 remove_extent_mapping(map_tree, em);
5184                 free_extent_map(em);
5185                 if (need_resched()) {
5186                         write_unlock(&map_tree->lock);
5187                         cond_resched();
5188                         write_lock(&map_tree->lock);
5189                 }
5190         }
5191         write_unlock(&map_tree->lock);
5192
5193         /*
5194          * Keep looping until we have no more ranges in the io tree.
5195          * We can have ongoing bios started by readpages (called from readahead)
5196          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5197          * still in progress (unlocked the pages in the bio but did not yet
5198          * unlocked the ranges in the io tree). Therefore this means some
5199          * ranges can still be locked and eviction started because before
5200          * submitting those bios, which are executed by a separate task (work
5201          * queue kthread), inode references (inode->i_count) were not taken
5202          * (which would be dropped in the end io callback of each bio).
5203          * Therefore here we effectively end up waiting for those bios and
5204          * anyone else holding locked ranges without having bumped the inode's
5205          * reference count - if we don't do it, when they access the inode's
5206          * io_tree to unlock a range it may be too late, leading to an
5207          * use-after-free issue.
5208          */
5209         spin_lock(&io_tree->lock);
5210         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5211                 struct extent_state *state;
5212                 struct extent_state *cached_state = NULL;
5213                 u64 start;
5214                 u64 end;
5215
5216                 node = rb_first(&io_tree->state);
5217                 state = rb_entry(node, struct extent_state, rb_node);
5218                 start = state->start;
5219                 end = state->end;
5220                 spin_unlock(&io_tree->lock);
5221
5222                 lock_extent_bits(io_tree, start, end, &cached_state);
5223
5224                 /*
5225                  * If still has DELALLOC flag, the extent didn't reach disk,
5226                  * and its reserved space won't be freed by delayed_ref.
5227                  * So we need to free its reserved space here.
5228                  * (Refer to comment in btrfs_invalidatepage, case 2)
5229                  *
5230                  * Note, end is the bytenr of last byte, so we need + 1 here.
5231                  */
5232                 if (state->state & EXTENT_DELALLOC)
5233                         btrfs_qgroup_free_data(inode, NULL, start, end - start + 1);
5234
5235                 clear_extent_bit(io_tree, start, end,
5236                                  EXTENT_LOCKED | EXTENT_DIRTY |
5237                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5238                                  EXTENT_DEFRAG, 1, 1,
5239                                  &cached_state, GFP_NOFS);
5240
5241                 cond_resched();
5242                 spin_lock(&io_tree->lock);
5243         }
5244         spin_unlock(&io_tree->lock);
5245 }
5246
5247 void btrfs_evict_inode(struct inode *inode)
5248 {
5249         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5250         struct btrfs_trans_handle *trans;
5251         struct btrfs_root *root = BTRFS_I(inode)->root;
5252         struct btrfs_block_rsv *rsv, *global_rsv;
5253         int steal_from_global = 0;
5254         u64 min_size;
5255         int ret;
5256
5257         trace_btrfs_inode_evict(inode);
5258
5259         if (!root) {
5260                 kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
5261                 return;
5262         }
5263
5264         min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
5265
5266         evict_inode_truncate_pages(inode);
5267
5268         if (inode->i_nlink &&
5269             ((btrfs_root_refs(&root->root_item) != 0 &&
5270               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5271              btrfs_is_free_space_inode(BTRFS_I(inode))))
5272                 goto no_delete;
5273
5274         if (is_bad_inode(inode)) {
5275                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5276                 goto no_delete;
5277         }
5278         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5279         if (!special_file(inode->i_mode))
5280                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5281
5282         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5283
5284         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
5285                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5286                                  &BTRFS_I(inode)->runtime_flags));
5287                 goto no_delete;
5288         }
5289
5290         if (inode->i_nlink > 0) {
5291                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5292                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5293                 goto no_delete;
5294         }
5295
5296         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5297         if (ret) {
5298                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5299                 goto no_delete;
5300         }
5301
5302         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5303         if (!rsv) {
5304                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5305                 goto no_delete;
5306         }
5307         rsv->size = min_size;
5308         rsv->failfast = 1;
5309         global_rsv = &fs_info->global_block_rsv;
5310
5311         btrfs_i_size_write(BTRFS_I(inode), 0);
5312
5313         /*
5314          * This is a bit simpler than btrfs_truncate since we've already
5315          * reserved our space for our orphan item in the unlink, so we just
5316          * need to reserve some slack space in case we add bytes and update
5317          * inode item when doing the truncate.
5318          */
5319         while (1) {
5320                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5321                                              BTRFS_RESERVE_FLUSH_LIMIT);
5322
5323                 /*
5324                  * Try and steal from the global reserve since we will
5325                  * likely not use this space anyway, we want to try as
5326                  * hard as possible to get this to work.
5327                  */
5328                 if (ret)
5329                         steal_from_global++;
5330                 else
5331                         steal_from_global = 0;
5332                 ret = 0;
5333
5334                 /*
5335                  * steal_from_global == 0: we reserved stuff, hooray!
5336                  * steal_from_global == 1: we didn't reserve stuff, boo!
5337                  * steal_from_global == 2: we've committed, still not a lot of
5338                  * room but maybe we'll have room in the global reserve this
5339                  * time.
5340                  * steal_from_global == 3: abandon all hope!
5341                  */
5342                 if (steal_from_global > 2) {
5343                         btrfs_warn(fs_info,
5344                                    "Could not get space for a delete, will truncate on mount %d",
5345                                    ret);
5346                         btrfs_orphan_del(NULL, BTRFS_I(inode));
5347                         btrfs_free_block_rsv(fs_info, rsv);
5348                         goto no_delete;
5349                 }
5350
5351                 trans = btrfs_join_transaction(root);
5352                 if (IS_ERR(trans)) {
5353                         btrfs_orphan_del(NULL, BTRFS_I(inode));
5354                         btrfs_free_block_rsv(fs_info, rsv);
5355                         goto no_delete;
5356                 }
5357
5358                 /*
5359                  * We can't just steal from the global reserve, we need to make
5360                  * sure there is room to do it, if not we need to commit and try
5361                  * again.
5362                  */
5363                 if (steal_from_global) {
5364                         if (!btrfs_check_space_for_delayed_refs(trans, fs_info))
5365                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5366                                                               min_size, 0);
5367                         else
5368                                 ret = -ENOSPC;
5369                 }
5370
5371                 /*
5372                  * Couldn't steal from the global reserve, we have too much
5373                  * pending stuff built up, commit the transaction and try it
5374                  * again.
5375                  */
5376                 if (ret) {
5377                         ret = btrfs_commit_transaction(trans);
5378                         if (ret) {
5379                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5380                                 btrfs_free_block_rsv(fs_info, rsv);
5381                                 goto no_delete;
5382                         }
5383                         continue;
5384                 } else {
5385                         steal_from_global = 0;
5386                 }
5387
5388                 trans->block_rsv = rsv;
5389
5390                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5391                 if (ret != -ENOSPC && ret != -EAGAIN)
5392                         break;
5393
5394                 trans->block_rsv = &fs_info->trans_block_rsv;
5395                 btrfs_end_transaction(trans);
5396                 trans = NULL;
5397                 btrfs_btree_balance_dirty(fs_info);
5398         }
5399
5400         btrfs_free_block_rsv(fs_info, rsv);
5401
5402         /*
5403          * Errors here aren't a big deal, it just means we leave orphan items
5404          * in the tree.  They will be cleaned up on the next mount.
5405          */
5406         if (ret == 0) {
5407                 trans->block_rsv = root->orphan_block_rsv;
5408                 btrfs_orphan_del(trans, BTRFS_I(inode));
5409         } else {
5410                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5411         }
5412
5413         trans->block_rsv = &fs_info->trans_block_rsv;
5414         if (!(root == fs_info->tree_root ||
5415               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5416                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5417
5418         btrfs_end_transaction(trans);
5419         btrfs_btree_balance_dirty(fs_info);
5420 no_delete:
5421         btrfs_remove_delayed_node(BTRFS_I(inode));
5422         clear_inode(inode);
5423 }
5424
5425 /*
5426  * this returns the key found in the dir entry in the location pointer.
5427  * If no dir entries were found, location->objectid is 0.
5428  */
5429 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5430                                struct btrfs_key *location)
5431 {
5432         const char *name = dentry->d_name.name;
5433         int namelen = dentry->d_name.len;
5434         struct btrfs_dir_item *di;
5435         struct btrfs_path *path;
5436         struct btrfs_root *root = BTRFS_I(dir)->root;
5437         int ret = 0;
5438
5439         path = btrfs_alloc_path();
5440         if (!path)
5441                 return -ENOMEM;
5442
5443         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5444                         name, namelen, 0);
5445         if (IS_ERR(di))
5446                 ret = PTR_ERR(di);
5447
5448         if (IS_ERR_OR_NULL(di))
5449                 goto out_err;
5450
5451         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5452         if (location->type != BTRFS_INODE_ITEM_KEY &&
5453             location->type != BTRFS_ROOT_ITEM_KEY) {
5454                 btrfs_warn(root->fs_info,
5455 "%s gets something invalid in DIR_ITEM (name %s, directory ino %llu, location(%llu %u %llu))",
5456                            __func__, name, btrfs_ino(BTRFS_I(dir)),
5457                            location->objectid, location->type, location->offset);
5458                 goto out_err;
5459         }
5460 out:
5461         btrfs_free_path(path);
5462         return ret;
5463 out_err:
5464         location->objectid = 0;
5465         goto out;
5466 }
5467
5468 /*
5469  * when we hit a tree root in a directory, the btrfs part of the inode
5470  * needs to be changed to reflect the root directory of the tree root.  This
5471  * is kind of like crossing a mount point.
5472  */
5473 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5474                                     struct inode *dir,
5475                                     struct dentry *dentry,
5476                                     struct btrfs_key *location,
5477                                     struct btrfs_root **sub_root)
5478 {
5479         struct btrfs_path *path;
5480         struct btrfs_root *new_root;
5481         struct btrfs_root_ref *ref;
5482         struct extent_buffer *leaf;
5483         struct btrfs_key key;
5484         int ret;
5485         int err = 0;
5486
5487         path = btrfs_alloc_path();
5488         if (!path) {
5489                 err = -ENOMEM;
5490                 goto out;
5491         }
5492
5493         err = -ENOENT;
5494         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5495         key.type = BTRFS_ROOT_REF_KEY;
5496         key.offset = location->objectid;
5497
5498         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5499         if (ret) {
5500                 if (ret < 0)
5501                         err = ret;
5502                 goto out;
5503         }
5504
5505         leaf = path->nodes[0];
5506         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5507         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5508             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5509                 goto out;
5510
5511         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5512                                    (unsigned long)(ref + 1),
5513                                    dentry->d_name.len);
5514         if (ret)
5515                 goto out;
5516
5517         btrfs_release_path(path);
5518
5519         new_root = btrfs_read_fs_root_no_name(fs_info, location);
5520         if (IS_ERR(new_root)) {
5521                 err = PTR_ERR(new_root);
5522                 goto out;
5523         }
5524
5525         *sub_root = new_root;
5526         location->objectid = btrfs_root_dirid(&new_root->root_item);
5527         location->type = BTRFS_INODE_ITEM_KEY;
5528         location->offset = 0;
5529         err = 0;
5530 out:
5531         btrfs_free_path(path);
5532         return err;
5533 }
5534
5535 static void inode_tree_add(struct inode *inode)
5536 {
5537         struct btrfs_root *root = BTRFS_I(inode)->root;
5538         struct btrfs_inode *entry;
5539         struct rb_node **p;
5540         struct rb_node *parent;
5541         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5542         u64 ino = btrfs_ino(BTRFS_I(inode));
5543
5544         if (inode_unhashed(inode))
5545                 return;
5546         parent = NULL;
5547         spin_lock(&root->inode_lock);
5548         p = &root->inode_tree.rb_node;
5549         while (*p) {
5550                 parent = *p;
5551                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5552
5553                 if (ino < btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5554                         p = &parent->rb_left;
5555                 else if (ino > btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5556                         p = &parent->rb_right;
5557                 else {
5558                         WARN_ON(!(entry->vfs_inode.i_state &
5559                                   (I_WILL_FREE | I_FREEING)));
5560                         rb_replace_node(parent, new, &root->inode_tree);
5561                         RB_CLEAR_NODE(parent);
5562                         spin_unlock(&root->inode_lock);
5563                         return;
5564                 }
5565         }
5566         rb_link_node(new, parent, p);
5567         rb_insert_color(new, &root->inode_tree);
5568         spin_unlock(&root->inode_lock);
5569 }
5570
5571 static void inode_tree_del(struct inode *inode)
5572 {
5573         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5574         struct btrfs_root *root = BTRFS_I(inode)->root;
5575         int empty = 0;
5576
5577         spin_lock(&root->inode_lock);
5578         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5579                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5580                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5581                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5582         }
5583         spin_unlock(&root->inode_lock);
5584
5585         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5586                 synchronize_srcu(&fs_info->subvol_srcu);
5587                 spin_lock(&root->inode_lock);
5588                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5589                 spin_unlock(&root->inode_lock);
5590                 if (empty)
5591                         btrfs_add_dead_root(root);
5592         }
5593 }
5594
5595 void btrfs_invalidate_inodes(struct btrfs_root *root)
5596 {
5597         struct btrfs_fs_info *fs_info = root->fs_info;
5598         struct rb_node *node;
5599         struct rb_node *prev;
5600         struct btrfs_inode *entry;
5601         struct inode *inode;
5602         u64 objectid = 0;
5603
5604         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
5605                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5606
5607         spin_lock(&root->inode_lock);
5608 again:
5609         node = root->inode_tree.rb_node;
5610         prev = NULL;
5611         while (node) {
5612                 prev = node;
5613                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5614
5615                 if (objectid < btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5616                         node = node->rb_left;
5617                 else if (objectid > btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5618                         node = node->rb_right;
5619                 else
5620                         break;
5621         }
5622         if (!node) {
5623                 while (prev) {
5624                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5625                         if (objectid <= btrfs_ino(BTRFS_I(&entry->vfs_inode))) {
5626                                 node = prev;
5627                                 break;
5628                         }
5629                         prev = rb_next(prev);
5630                 }
5631         }
5632         while (node) {
5633                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5634                 objectid = btrfs_ino(BTRFS_I(&entry->vfs_inode)) + 1;
5635                 inode = igrab(&entry->vfs_inode);
5636                 if (inode) {
5637                         spin_unlock(&root->inode_lock);
5638                         if (atomic_read(&inode->i_count) > 1)
5639                                 d_prune_aliases(inode);
5640                         /*
5641                          * btrfs_drop_inode will have it removed from
5642                          * the inode cache when its usage count
5643                          * hits zero.
5644                          */
5645                         iput(inode);
5646                         cond_resched();
5647                         spin_lock(&root->inode_lock);
5648                         goto again;
5649                 }
5650
5651                 if (cond_resched_lock(&root->inode_lock))
5652                         goto again;
5653
5654                 node = rb_next(node);
5655         }
5656         spin_unlock(&root->inode_lock);
5657 }
5658
5659 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5660 {
5661         struct btrfs_iget_args *args = p;
5662         inode->i_ino = args->location->objectid;
5663         memcpy(&BTRFS_I(inode)->location, args->location,
5664                sizeof(*args->location));
5665         BTRFS_I(inode)->root = args->root;
5666         return 0;
5667 }
5668
5669 static int btrfs_find_actor(struct inode *inode, void *opaque)
5670 {
5671         struct btrfs_iget_args *args = opaque;
5672         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5673                 args->root == BTRFS_I(inode)->root;
5674 }
5675
5676 static struct inode *btrfs_iget_locked(struct super_block *s,
5677                                        struct btrfs_key *location,
5678                                        struct btrfs_root *root)
5679 {
5680         struct inode *inode;
5681         struct btrfs_iget_args args;
5682         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5683
5684         args.location = location;
5685         args.root = root;
5686
5687         inode = iget5_locked(s, hashval, btrfs_find_actor,
5688                              btrfs_init_locked_inode,
5689                              (void *)&args);
5690         return inode;
5691 }
5692
5693 /* Get an inode object given its location and corresponding root.
5694  * Returns in *is_new if the inode was read from disk
5695  */
5696 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5697                          struct btrfs_root *root, int *new)
5698 {
5699         struct inode *inode;
5700
5701         inode = btrfs_iget_locked(s, location, root);
5702         if (!inode)
5703                 return ERR_PTR(-ENOMEM);
5704
5705         if (inode->i_state & I_NEW) {
5706                 int ret;
5707
5708                 ret = btrfs_read_locked_inode(inode);
5709                 if (!is_bad_inode(inode)) {
5710                         inode_tree_add(inode);
5711                         unlock_new_inode(inode);
5712                         if (new)
5713                                 *new = 1;
5714                 } else {
5715                         unlock_new_inode(inode);
5716                         iput(inode);
5717                         ASSERT(ret < 0);
5718                         inode = ERR_PTR(ret < 0 ? ret : -ESTALE);
5719                 }
5720         }
5721
5722         return inode;
5723 }
5724
5725 static struct inode *new_simple_dir(struct super_block *s,
5726                                     struct btrfs_key *key,
5727                                     struct btrfs_root *root)
5728 {
5729         struct inode *inode = new_inode(s);
5730
5731         if (!inode)
5732                 return ERR_PTR(-ENOMEM);
5733
5734         BTRFS_I(inode)->root = root;
5735         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5736         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5737
5738         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5739         inode->i_op = &btrfs_dir_ro_inode_operations;
5740         inode->i_opflags &= ~IOP_XATTR;
5741         inode->i_fop = &simple_dir_operations;
5742         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5743         inode->i_mtime = current_time(inode);
5744         inode->i_atime = inode->i_mtime;
5745         inode->i_ctime = inode->i_mtime;
5746         BTRFS_I(inode)->i_otime = inode->i_mtime;
5747
5748         return inode;
5749 }
5750
5751 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5752 {
5753         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5754         struct inode *inode;
5755         struct btrfs_root *root = BTRFS_I(dir)->root;
5756         struct btrfs_root *sub_root = root;
5757         struct btrfs_key location;
5758         int index;
5759         int ret = 0;
5760
5761         if (dentry->d_name.len > BTRFS_NAME_LEN)
5762                 return ERR_PTR(-ENAMETOOLONG);
5763
5764         ret = btrfs_inode_by_name(dir, dentry, &location);
5765         if (ret < 0)
5766                 return ERR_PTR(ret);
5767
5768         if (location.objectid == 0)
5769                 return ERR_PTR(-ENOENT);
5770
5771         if (location.type == BTRFS_INODE_ITEM_KEY) {
5772                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5773                 return inode;
5774         }
5775
5776         index = srcu_read_lock(&fs_info->subvol_srcu);
5777         ret = fixup_tree_root_location(fs_info, dir, dentry,
5778                                        &location, &sub_root);
5779         if (ret < 0) {
5780                 if (ret != -ENOENT)
5781                         inode = ERR_PTR(ret);
5782                 else
5783                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5784         } else {
5785                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5786         }
5787         srcu_read_unlock(&fs_info->subvol_srcu, index);
5788
5789         if (!IS_ERR(inode) && root != sub_root) {
5790                 down_read(&fs_info->cleanup_work_sem);
5791                 if (!sb_rdonly(inode->i_sb))
5792                         ret = btrfs_orphan_cleanup(sub_root);
5793                 up_read(&fs_info->cleanup_work_sem);
5794                 if (ret) {
5795                         iput(inode);
5796                         inode = ERR_PTR(ret);
5797                 }
5798         }
5799
5800         return inode;
5801 }
5802
5803 static int btrfs_dentry_delete(const struct dentry *dentry)
5804 {
5805         struct btrfs_root *root;
5806         struct inode *inode = d_inode(dentry);
5807
5808         if (!inode && !IS_ROOT(dentry))
5809                 inode = d_inode(dentry->d_parent);
5810
5811         if (inode) {
5812                 root = BTRFS_I(inode)->root;
5813                 if (btrfs_root_refs(&root->root_item) == 0)
5814                         return 1;
5815
5816                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5817                         return 1;
5818         }
5819         return 0;
5820 }
5821
5822 static void btrfs_dentry_release(struct dentry *dentry)
5823 {
5824         kfree(dentry->d_fsdata);
5825 }
5826
5827 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5828                                    unsigned int flags)
5829 {
5830         struct inode *inode;
5831
5832         inode = btrfs_lookup_dentry(dir, dentry);
5833         if (IS_ERR(inode)) {
5834                 if (PTR_ERR(inode) == -ENOENT)
5835                         inode = NULL;
5836                 else
5837                         return ERR_CAST(inode);
5838         }
5839
5840         return d_splice_alias(inode, dentry);
5841 }
5842
5843 unsigned char btrfs_filetype_table[] = {
5844         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5845 };
5846
5847 /*
5848  * All this infrastructure exists because dir_emit can fault, and we are holding
5849  * the tree lock when doing readdir.  For now just allocate a buffer and copy
5850  * our information into that, and then dir_emit from the buffer.  This is
5851  * similar to what NFS does, only we don't keep the buffer around in pagecache
5852  * because I'm afraid I'll mess that up.  Long term we need to make filldir do
5853  * copy_to_user_inatomic so we don't have to worry about page faulting under the
5854  * tree lock.
5855  */
5856 static int btrfs_opendir(struct inode *inode, struct file *file)
5857 {
5858         struct btrfs_file_private *private;
5859
5860         private = kzalloc(sizeof(struct btrfs_file_private), GFP_KERNEL);
5861         if (!private)
5862                 return -ENOMEM;
5863         private->filldir_buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
5864         if (!private->filldir_buf) {
5865                 kfree(private);
5866                 return -ENOMEM;
5867         }
5868         file->private_data = private;
5869         return 0;
5870 }
5871
5872 struct dir_entry {
5873         u64 ino;
5874         u64 offset;
5875         unsigned type;
5876         int name_len;
5877 };
5878
5879 static int btrfs_filldir(void *addr, int entries, struct dir_context *ctx)
5880 {
5881         while (entries--) {
5882                 struct dir_entry *entry = addr;
5883                 char *name = (char *)(entry + 1);
5884
5885                 ctx->pos = entry->offset;
5886                 if (!dir_emit(ctx, name, entry->name_len, entry->ino,
5887                               entry->type))
5888                         return 1;
5889                 addr += sizeof(struct dir_entry) + entry->name_len;
5890                 ctx->pos++;
5891         }
5892         return 0;
5893 }
5894
5895 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5896 {
5897         struct inode *inode = file_inode(file);
5898         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5899         struct btrfs_root *root = BTRFS_I(inode)->root;
5900         struct btrfs_file_private *private = file->private_data;
5901         struct btrfs_dir_item *di;
5902         struct btrfs_key key;
5903         struct btrfs_key found_key;
5904         struct btrfs_path *path;
5905         void *addr;
5906         struct list_head ins_list;
5907         struct list_head del_list;
5908         int ret;
5909         struct extent_buffer *leaf;
5910         int slot;
5911         char *name_ptr;
5912         int name_len;
5913         int entries = 0;
5914         int total_len = 0;
5915         bool put = false;
5916         struct btrfs_key location;
5917
5918         if (!dir_emit_dots(file, ctx))
5919                 return 0;
5920
5921         path = btrfs_alloc_path();
5922         if (!path)
5923                 return -ENOMEM;
5924
5925         addr = private->filldir_buf;
5926         path->reada = READA_FORWARD;
5927
5928         INIT_LIST_HEAD(&ins_list);
5929         INIT_LIST_HEAD(&del_list);
5930         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
5931
5932 again:
5933         key.type = BTRFS_DIR_INDEX_KEY;
5934         key.offset = ctx->pos;
5935         key.objectid = btrfs_ino(BTRFS_I(inode));
5936
5937         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5938         if (ret < 0)
5939                 goto err;
5940
5941         while (1) {
5942                 struct dir_entry *entry;
5943
5944                 leaf = path->nodes[0];
5945                 slot = path->slots[0];
5946                 if (slot >= btrfs_header_nritems(leaf)) {
5947                         ret = btrfs_next_leaf(root, path);
5948                         if (ret < 0)
5949                                 goto err;
5950                         else if (ret > 0)
5951                                 break;
5952                         continue;
5953                 }
5954
5955                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5956
5957                 if (found_key.objectid != key.objectid)
5958                         break;
5959                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
5960                         break;
5961                 if (found_key.offset < ctx->pos)
5962                         goto next;
5963                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
5964                         goto next;
5965                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5966                 if (verify_dir_item(fs_info, leaf, slot, di))
5967                         goto next;
5968
5969                 name_len = btrfs_dir_name_len(leaf, di);
5970                 if ((total_len + sizeof(struct dir_entry) + name_len) >=
5971                     PAGE_SIZE) {
5972                         btrfs_release_path(path);
5973                         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
5974                         if (ret)
5975                                 goto nopos;
5976                         addr = private->filldir_buf;
5977                         entries = 0;
5978                         total_len = 0;
5979                         goto again;
5980                 }
5981
5982                 entry = addr;
5983                 entry->name_len = name_len;
5984                 name_ptr = (char *)(entry + 1);
5985                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
5986                                    name_len);
5987                 entry->type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5988                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
5989                 entry->ino = location.objectid;
5990                 entry->offset = found_key.offset;
5991                 entries++;
5992                 addr += sizeof(struct dir_entry) + name_len;
5993                 total_len += sizeof(struct dir_entry) + name_len;
5994 next:
5995                 path->slots[0]++;
5996         }
5997         btrfs_release_path(path);
5998
5999         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6000         if (ret)
6001                 goto nopos;
6002
6003         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
6004         if (ret)
6005                 goto nopos;
6006
6007         /*
6008          * Stop new entries from being returned after we return the last
6009          * entry.
6010          *
6011          * New directory entries are assigned a strictly increasing
6012          * offset.  This means that new entries created during readdir
6013          * are *guaranteed* to be seen in the future by that readdir.
6014          * This has broken buggy programs which operate on names as
6015          * they're returned by readdir.  Until we re-use freed offsets
6016          * we have this hack to stop new entries from being returned
6017          * under the assumption that they'll never reach this huge
6018          * offset.
6019          *
6020          * This is being careful not to overflow 32bit loff_t unless the
6021          * last entry requires it because doing so has broken 32bit apps
6022          * in the past.
6023          */
6024         if (ctx->pos >= INT_MAX)
6025                 ctx->pos = LLONG_MAX;
6026         else
6027                 ctx->pos = INT_MAX;
6028 nopos:
6029         ret = 0;
6030 err:
6031         if (put)
6032                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
6033         btrfs_free_path(path);
6034         return ret;
6035 }
6036
6037 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
6038 {
6039         struct btrfs_root *root = BTRFS_I(inode)->root;
6040         struct btrfs_trans_handle *trans;
6041         int ret = 0;
6042         bool nolock = false;
6043
6044         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6045                 return 0;
6046
6047         if (btrfs_fs_closing(root->fs_info) &&
6048                         btrfs_is_free_space_inode(BTRFS_I(inode)))
6049                 nolock = true;
6050
6051         if (wbc->sync_mode == WB_SYNC_ALL) {
6052                 if (nolock)
6053                         trans = btrfs_join_transaction_nolock(root);
6054                 else
6055                         trans = btrfs_join_transaction(root);
6056                 if (IS_ERR(trans))
6057                         return PTR_ERR(trans);
6058                 ret = btrfs_commit_transaction(trans);
6059         }
6060         return ret;
6061 }
6062
6063 /*
6064  * This is somewhat expensive, updating the tree every time the
6065  * inode changes.  But, it is most likely to find the inode in cache.
6066  * FIXME, needs more benchmarking...there are no reasons other than performance
6067  * to keep or drop this code.
6068  */
6069 static int btrfs_dirty_inode(struct inode *inode)
6070 {
6071         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6072         struct btrfs_root *root = BTRFS_I(inode)->root;
6073         struct btrfs_trans_handle *trans;
6074         int ret;
6075
6076         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6077                 return 0;
6078
6079         trans = btrfs_join_transaction(root);
6080         if (IS_ERR(trans))
6081                 return PTR_ERR(trans);
6082
6083         ret = btrfs_update_inode(trans, root, inode);
6084         if (ret && ret == -ENOSPC) {
6085                 /* whoops, lets try again with the full transaction */
6086                 btrfs_end_transaction(trans);
6087                 trans = btrfs_start_transaction(root, 1);
6088                 if (IS_ERR(trans))
6089                         return PTR_ERR(trans);
6090
6091                 ret = btrfs_update_inode(trans, root, inode);
6092         }
6093         btrfs_end_transaction(trans);
6094         if (BTRFS_I(inode)->delayed_node)
6095                 btrfs_balance_delayed_items(fs_info);
6096
6097         return ret;
6098 }
6099
6100 /*
6101  * This is a copy of file_update_time.  We need this so we can return error on
6102  * ENOSPC for updating the inode in the case of file write and mmap writes.
6103  */
6104 static int btrfs_update_time(struct inode *inode, struct timespec *now,
6105                              int flags)
6106 {
6107         struct btrfs_root *root = BTRFS_I(inode)->root;
6108
6109         if (btrfs_root_readonly(root))
6110                 return -EROFS;
6111
6112         if (flags & S_VERSION)
6113                 inode_inc_iversion(inode);
6114         if (flags & S_CTIME)
6115                 inode->i_ctime = *now;
6116         if (flags & S_MTIME)
6117                 inode->i_mtime = *now;
6118         if (flags & S_ATIME)
6119                 inode->i_atime = *now;
6120         return btrfs_dirty_inode(inode);
6121 }
6122
6123 /*
6124  * find the highest existing sequence number in a directory
6125  * and then set the in-memory index_cnt variable to reflect
6126  * free sequence numbers
6127  */
6128 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
6129 {
6130         struct btrfs_root *root = inode->root;
6131         struct btrfs_key key, found_key;
6132         struct btrfs_path *path;
6133         struct extent_buffer *leaf;
6134         int ret;
6135
6136         key.objectid = btrfs_ino(inode);
6137         key.type = BTRFS_DIR_INDEX_KEY;
6138         key.offset = (u64)-1;
6139
6140         path = btrfs_alloc_path();
6141         if (!path)
6142                 return -ENOMEM;
6143
6144         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6145         if (ret < 0)
6146                 goto out;
6147         /* FIXME: we should be able to handle this */
6148         if (ret == 0)
6149                 goto out;
6150         ret = 0;
6151
6152         /*
6153          * MAGIC NUMBER EXPLANATION:
6154          * since we search a directory based on f_pos we have to start at 2
6155          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6156          * else has to start at 2
6157          */
6158         if (path->slots[0] == 0) {
6159                 inode->index_cnt = 2;
6160                 goto out;
6161         }
6162
6163         path->slots[0]--;
6164
6165         leaf = path->nodes[0];
6166         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6167
6168         if (found_key.objectid != btrfs_ino(inode) ||
6169             found_key.type != BTRFS_DIR_INDEX_KEY) {
6170                 inode->index_cnt = 2;
6171                 goto out;
6172         }
6173
6174         inode->index_cnt = found_key.offset + 1;
6175 out:
6176         btrfs_free_path(path);
6177         return ret;
6178 }
6179
6180 /*
6181  * helper to find a free sequence number in a given directory.  This current
6182  * code is very simple, later versions will do smarter things in the btree
6183  */
6184 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
6185 {
6186         int ret = 0;
6187
6188         if (dir->index_cnt == (u64)-1) {
6189                 ret = btrfs_inode_delayed_dir_index_count(dir);
6190                 if (ret) {
6191                         ret = btrfs_set_inode_index_count(dir);
6192                         if (ret)
6193                                 return ret;
6194                 }
6195         }
6196
6197         *index = dir->index_cnt;
6198         dir->index_cnt++;
6199
6200         return ret;
6201 }
6202
6203 static int btrfs_insert_inode_locked(struct inode *inode)
6204 {
6205         struct btrfs_iget_args args;
6206         args.location = &BTRFS_I(inode)->location;
6207         args.root = BTRFS_I(inode)->root;
6208
6209         return insert_inode_locked4(inode,
6210                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6211                    btrfs_find_actor, &args);
6212 }
6213
6214 /*
6215  * Inherit flags from the parent inode.
6216  *
6217  * Currently only the compression flags and the cow flags are inherited.
6218  */
6219 static void btrfs_inherit_iflags(struct inode *inode, struct inode *dir)
6220 {
6221         unsigned int flags;
6222
6223         if (!dir)
6224                 return;
6225
6226         flags = BTRFS_I(dir)->flags;
6227
6228         if (flags & BTRFS_INODE_NOCOMPRESS) {
6229                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_COMPRESS;
6230                 BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
6231         } else if (flags & BTRFS_INODE_COMPRESS) {
6232                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NOCOMPRESS;
6233                 BTRFS_I(inode)->flags |= BTRFS_INODE_COMPRESS;
6234         }
6235
6236         if (flags & BTRFS_INODE_NODATACOW) {
6237                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
6238                 if (S_ISREG(inode->i_mode))
6239                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6240         }
6241
6242         btrfs_update_iflags(inode);
6243 }
6244
6245 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6246                                      struct btrfs_root *root,
6247                                      struct inode *dir,
6248                                      const char *name, int name_len,
6249                                      u64 ref_objectid, u64 objectid,
6250                                      umode_t mode, u64 *index)
6251 {
6252         struct btrfs_fs_info *fs_info = root->fs_info;
6253         struct inode *inode;
6254         struct btrfs_inode_item *inode_item;
6255         struct btrfs_key *location;
6256         struct btrfs_path *path;
6257         struct btrfs_inode_ref *ref;
6258         struct btrfs_key key[2];
6259         u32 sizes[2];
6260         int nitems = name ? 2 : 1;
6261         unsigned long ptr;
6262         int ret;
6263
6264         path = btrfs_alloc_path();
6265         if (!path)
6266                 return ERR_PTR(-ENOMEM);
6267
6268         inode = new_inode(fs_info->sb);
6269         if (!inode) {
6270                 btrfs_free_path(path);
6271                 return ERR_PTR(-ENOMEM);
6272         }
6273
6274         /*
6275          * O_TMPFILE, set link count to 0, so that after this point,
6276          * we fill in an inode item with the correct link count.
6277          */
6278         if (!name)
6279                 set_nlink(inode, 0);
6280
6281         /*
6282          * we have to initialize this early, so we can reclaim the inode
6283          * number if we fail afterwards in this function.
6284          */
6285         inode->i_ino = objectid;
6286
6287         if (dir && name) {
6288                 trace_btrfs_inode_request(dir);
6289
6290                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
6291                 if (ret) {
6292                         btrfs_free_path(path);
6293                         iput(inode);
6294                         return ERR_PTR(ret);
6295                 }
6296         } else if (dir) {
6297                 *index = 0;
6298         }
6299         /*
6300          * index_cnt is ignored for everything but a dir,
6301          * btrfs_get_inode_index_count has an explanation for the magic
6302          * number
6303          */
6304         BTRFS_I(inode)->index_cnt = 2;
6305         BTRFS_I(inode)->dir_index = *index;
6306         BTRFS_I(inode)->root = root;
6307         BTRFS_I(inode)->generation = trans->transid;
6308         inode->i_generation = BTRFS_I(inode)->generation;
6309
6310         /*
6311          * We could have gotten an inode number from somebody who was fsynced
6312          * and then removed in this same transaction, so let's just set full
6313          * sync since it will be a full sync anyway and this will blow away the
6314          * old info in the log.
6315          */
6316         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6317
6318         key[0].objectid = objectid;
6319         key[0].type = BTRFS_INODE_ITEM_KEY;
6320         key[0].offset = 0;
6321
6322         sizes[0] = sizeof(struct btrfs_inode_item);
6323
6324         if (name) {
6325                 /*
6326                  * Start new inodes with an inode_ref. This is slightly more
6327                  * efficient for small numbers of hard links since they will
6328                  * be packed into one item. Extended refs will kick in if we
6329                  * add more hard links than can fit in the ref item.
6330                  */
6331                 key[1].objectid = objectid;
6332                 key[1].type = BTRFS_INODE_REF_KEY;
6333                 key[1].offset = ref_objectid;
6334
6335                 sizes[1] = name_len + sizeof(*ref);
6336         }
6337
6338         location = &BTRFS_I(inode)->location;
6339         location->objectid = objectid;
6340         location->offset = 0;
6341         location->type = BTRFS_INODE_ITEM_KEY;
6342
6343         ret = btrfs_insert_inode_locked(inode);
6344         if (ret < 0)
6345                 goto fail;
6346
6347         path->leave_spinning = 1;
6348         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6349         if (ret != 0)
6350                 goto fail_unlock;
6351
6352         inode_init_owner(inode, dir, mode);
6353         inode_set_bytes(inode, 0);
6354
6355         inode->i_mtime = current_time(inode);
6356         inode->i_atime = inode->i_mtime;
6357         inode->i_ctime = inode->i_mtime;
6358         BTRFS_I(inode)->i_otime = inode->i_mtime;
6359
6360         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6361                                   struct btrfs_inode_item);
6362         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6363                              sizeof(*inode_item));
6364         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6365
6366         if (name) {
6367                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6368                                      struct btrfs_inode_ref);
6369                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6370                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6371                 ptr = (unsigned long)(ref + 1);
6372                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6373         }
6374
6375         btrfs_mark_buffer_dirty(path->nodes[0]);
6376         btrfs_free_path(path);
6377
6378         btrfs_inherit_iflags(inode, dir);
6379
6380         if (S_ISREG(mode)) {
6381                 if (btrfs_test_opt(fs_info, NODATASUM))
6382                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6383                 if (btrfs_test_opt(fs_info, NODATACOW))
6384                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6385                                 BTRFS_INODE_NODATASUM;
6386         }
6387
6388         inode_tree_add(inode);
6389
6390         trace_btrfs_inode_new(inode);
6391         btrfs_set_inode_last_trans(trans, inode);
6392
6393         btrfs_update_root_times(trans, root);
6394
6395         ret = btrfs_inode_inherit_props(trans, inode, dir);
6396         if (ret)
6397                 btrfs_err(fs_info,
6398                           "error inheriting props for ino %llu (root %llu): %d",
6399                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6400
6401         return inode;
6402
6403 fail_unlock:
6404         unlock_new_inode(inode);
6405 fail:
6406         if (dir && name)
6407                 BTRFS_I(dir)->index_cnt--;
6408         btrfs_free_path(path);
6409         iput(inode);
6410         return ERR_PTR(ret);
6411 }
6412
6413 static inline u8 btrfs_inode_type(struct inode *inode)
6414 {
6415         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6416 }
6417
6418 /*
6419  * utility function to add 'inode' into 'parent_inode' with
6420  * a give name and a given sequence number.
6421  * if 'add_backref' is true, also insert a backref from the
6422  * inode to the parent directory.
6423  */
6424 int btrfs_add_link(struct btrfs_trans_handle *trans,
6425                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6426                    const char *name, int name_len, int add_backref, u64 index)
6427 {
6428         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6429         int ret = 0;
6430         struct btrfs_key key;
6431         struct btrfs_root *root = parent_inode->root;
6432         u64 ino = btrfs_ino(inode);
6433         u64 parent_ino = btrfs_ino(parent_inode);
6434
6435         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6436                 memcpy(&key, &inode->root->root_key, sizeof(key));
6437         } else {
6438                 key.objectid = ino;
6439                 key.type = BTRFS_INODE_ITEM_KEY;
6440                 key.offset = 0;
6441         }
6442
6443         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6444                 ret = btrfs_add_root_ref(trans, fs_info, key.objectid,
6445                                          root->root_key.objectid, parent_ino,
6446                                          index, name, name_len);
6447         } else if (add_backref) {
6448                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6449                                              parent_ino, index);
6450         }
6451
6452         /* Nothing to clean up yet */
6453         if (ret)
6454                 return ret;
6455
6456         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6457                                     parent_inode, &key,
6458                                     btrfs_inode_type(&inode->vfs_inode), index);
6459         if (ret == -EEXIST || ret == -EOVERFLOW)
6460                 goto fail_dir_item;
6461         else if (ret) {
6462                 btrfs_abort_transaction(trans, ret);
6463                 return ret;
6464         }
6465
6466         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6467                            name_len * 2);
6468         inode_inc_iversion(&parent_inode->vfs_inode);
6469         parent_inode->vfs_inode.i_mtime = parent_inode->vfs_inode.i_ctime =
6470                 current_time(&parent_inode->vfs_inode);
6471         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6472         if (ret)
6473                 btrfs_abort_transaction(trans, ret);
6474         return ret;
6475
6476 fail_dir_item:
6477         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6478                 u64 local_index;
6479                 int err;
6480                 err = btrfs_del_root_ref(trans, fs_info, key.objectid,
6481                                          root->root_key.objectid, parent_ino,
6482                                          &local_index, name, name_len);
6483
6484         } else if (add_backref) {
6485                 u64 local_index;
6486                 int err;
6487
6488                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6489                                           ino, parent_ino, &local_index);
6490         }
6491         return ret;
6492 }
6493
6494 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6495                             struct btrfs_inode *dir, struct dentry *dentry,
6496                             struct btrfs_inode *inode, int backref, u64 index)
6497 {
6498         int err = btrfs_add_link(trans, dir, inode,
6499                                  dentry->d_name.name, dentry->d_name.len,
6500                                  backref, index);
6501         if (err > 0)
6502                 err = -EEXIST;
6503         return err;
6504 }
6505
6506 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6507                         umode_t mode, dev_t rdev)
6508 {
6509         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6510         struct btrfs_trans_handle *trans;
6511         struct btrfs_root *root = BTRFS_I(dir)->root;
6512         struct inode *inode = NULL;
6513         int err;
6514         int drop_inode = 0;
6515         u64 objectid;
6516         u64 index = 0;
6517
6518         /*
6519          * 2 for inode item and ref
6520          * 2 for dir items
6521          * 1 for xattr if selinux is on
6522          */
6523         trans = btrfs_start_transaction(root, 5);
6524         if (IS_ERR(trans))
6525                 return PTR_ERR(trans);
6526
6527         err = btrfs_find_free_ino(root, &objectid);
6528         if (err)
6529                 goto out_unlock;
6530
6531         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6532                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6533                         mode, &index);
6534         if (IS_ERR(inode)) {
6535                 err = PTR_ERR(inode);
6536                 goto out_unlock;
6537         }
6538
6539         /*
6540         * If the active LSM wants to access the inode during
6541         * d_instantiate it needs these. Smack checks to see
6542         * if the filesystem supports xattrs by looking at the
6543         * ops vector.
6544         */
6545         inode->i_op = &btrfs_special_inode_operations;
6546         init_special_inode(inode, inode->i_mode, rdev);
6547
6548         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6549         if (err)
6550                 goto out_unlock_inode;
6551
6552         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6553                         0, index);
6554         if (err) {
6555                 goto out_unlock_inode;
6556         } else {
6557                 btrfs_update_inode(trans, root, inode);
6558                 unlock_new_inode(inode);
6559                 d_instantiate(dentry, inode);
6560         }
6561
6562 out_unlock:
6563         btrfs_end_transaction(trans);
6564         btrfs_balance_delayed_items(fs_info);
6565         btrfs_btree_balance_dirty(fs_info);
6566         if (drop_inode) {
6567                 inode_dec_link_count(inode);
6568                 iput(inode);
6569         }
6570         return err;
6571
6572 out_unlock_inode:
6573         drop_inode = 1;
6574         unlock_new_inode(inode);
6575         goto out_unlock;
6576
6577 }
6578
6579 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6580                         umode_t mode, bool excl)
6581 {
6582         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6583         struct btrfs_trans_handle *trans;
6584         struct btrfs_root *root = BTRFS_I(dir)->root;
6585         struct inode *inode = NULL;
6586         int drop_inode_on_err = 0;
6587         int err;
6588         u64 objectid;
6589         u64 index = 0;
6590
6591         /*
6592          * 2 for inode item and ref
6593          * 2 for dir items
6594          * 1 for xattr if selinux is on
6595          */
6596         trans = btrfs_start_transaction(root, 5);
6597         if (IS_ERR(trans))
6598                 return PTR_ERR(trans);
6599
6600         err = btrfs_find_free_ino(root, &objectid);
6601         if (err)
6602                 goto out_unlock;
6603
6604         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6605                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6606                         mode, &index);
6607         if (IS_ERR(inode)) {
6608                 err = PTR_ERR(inode);
6609                 goto out_unlock;
6610         }
6611         drop_inode_on_err = 1;
6612         /*
6613         * If the active LSM wants to access the inode during
6614         * d_instantiate it needs these. Smack checks to see
6615         * if the filesystem supports xattrs by looking at the
6616         * ops vector.
6617         */
6618         inode->i_fop = &btrfs_file_operations;
6619         inode->i_op = &btrfs_file_inode_operations;
6620         inode->i_mapping->a_ops = &btrfs_aops;
6621
6622         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6623         if (err)
6624                 goto out_unlock_inode;
6625
6626         err = btrfs_update_inode(trans, root, inode);
6627         if (err)
6628                 goto out_unlock_inode;
6629
6630         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6631                         0, index);
6632         if (err)
6633                 goto out_unlock_inode;
6634
6635         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6636         unlock_new_inode(inode);
6637         d_instantiate(dentry, inode);
6638
6639 out_unlock:
6640         btrfs_end_transaction(trans);
6641         if (err && drop_inode_on_err) {
6642                 inode_dec_link_count(inode);
6643                 iput(inode);
6644         }
6645         btrfs_balance_delayed_items(fs_info);
6646         btrfs_btree_balance_dirty(fs_info);
6647         return err;
6648
6649 out_unlock_inode:
6650         unlock_new_inode(inode);
6651         goto out_unlock;
6652
6653 }
6654
6655 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6656                       struct dentry *dentry)
6657 {
6658         struct btrfs_trans_handle *trans = NULL;
6659         struct btrfs_root *root = BTRFS_I(dir)->root;
6660         struct inode *inode = d_inode(old_dentry);
6661         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6662         u64 index;
6663         int err;
6664         int drop_inode = 0;
6665
6666         /* do not allow sys_link's with other subvols of the same device */
6667         if (root->objectid != BTRFS_I(inode)->root->objectid)
6668                 return -EXDEV;
6669
6670         if (inode->i_nlink >= BTRFS_LINK_MAX)
6671                 return -EMLINK;
6672
6673         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6674         if (err)
6675                 goto fail;
6676
6677         /*
6678          * 2 items for inode and inode ref
6679          * 2 items for dir items
6680          * 1 item for parent inode
6681          */
6682         trans = btrfs_start_transaction(root, 5);
6683         if (IS_ERR(trans)) {
6684                 err = PTR_ERR(trans);
6685                 trans = NULL;
6686                 goto fail;
6687         }
6688
6689         /* There are several dir indexes for this inode, clear the cache. */
6690         BTRFS_I(inode)->dir_index = 0ULL;
6691         inc_nlink(inode);
6692         inode_inc_iversion(inode);
6693         inode->i_ctime = current_time(inode);
6694         ihold(inode);
6695         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6696
6697         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6698                         1, index);
6699
6700         if (err) {
6701                 drop_inode = 1;
6702         } else {
6703                 struct dentry *parent = dentry->d_parent;
6704                 err = btrfs_update_inode(trans, root, inode);
6705                 if (err)
6706                         goto fail;
6707                 if (inode->i_nlink == 1) {
6708                         /*
6709                          * If new hard link count is 1, it's a file created
6710                          * with open(2) O_TMPFILE flag.
6711                          */
6712                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6713                         if (err)
6714                                 goto fail;
6715                 }
6716                 d_instantiate(dentry, inode);
6717                 btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent);
6718         }
6719
6720         btrfs_balance_delayed_items(fs_info);
6721 fail:
6722         if (trans)
6723                 btrfs_end_transaction(trans);
6724         if (drop_inode) {
6725                 inode_dec_link_count(inode);
6726                 iput(inode);
6727         }
6728         btrfs_btree_balance_dirty(fs_info);
6729         return err;
6730 }
6731
6732 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6733 {
6734         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6735         struct inode *inode = NULL;
6736         struct btrfs_trans_handle *trans;
6737         struct btrfs_root *root = BTRFS_I(dir)->root;
6738         int err = 0;
6739         int drop_on_err = 0;
6740         u64 objectid = 0;
6741         u64 index = 0;
6742
6743         /*
6744          * 2 items for inode and ref
6745          * 2 items for dir items
6746          * 1 for xattr if selinux is on
6747          */
6748         trans = btrfs_start_transaction(root, 5);
6749         if (IS_ERR(trans))
6750                 return PTR_ERR(trans);
6751
6752         err = btrfs_find_free_ino(root, &objectid);
6753         if (err)
6754                 goto out_fail;
6755
6756         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6757                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6758                         S_IFDIR | mode, &index);
6759         if (IS_ERR(inode)) {
6760                 err = PTR_ERR(inode);
6761                 goto out_fail;
6762         }
6763
6764         drop_on_err = 1;
6765         /* these must be set before we unlock the inode */
6766         inode->i_op = &btrfs_dir_inode_operations;
6767         inode->i_fop = &btrfs_dir_file_operations;
6768
6769         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6770         if (err)
6771                 goto out_fail_inode;
6772
6773         btrfs_i_size_write(BTRFS_I(inode), 0);
6774         err = btrfs_update_inode(trans, root, inode);
6775         if (err)
6776                 goto out_fail_inode;
6777
6778         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6779                         dentry->d_name.name,
6780                         dentry->d_name.len, 0, index);
6781         if (err)
6782                 goto out_fail_inode;
6783
6784         d_instantiate(dentry, inode);
6785         /*
6786          * mkdir is special.  We're unlocking after we call d_instantiate
6787          * to avoid a race with nfsd calling d_instantiate.
6788          */
6789         unlock_new_inode(inode);
6790         drop_on_err = 0;
6791
6792 out_fail:
6793         btrfs_end_transaction(trans);
6794         if (drop_on_err) {
6795                 inode_dec_link_count(inode);
6796                 iput(inode);
6797         }
6798         btrfs_balance_delayed_items(fs_info);
6799         btrfs_btree_balance_dirty(fs_info);
6800         return err;
6801
6802 out_fail_inode:
6803         unlock_new_inode(inode);
6804         goto out_fail;
6805 }
6806
6807 /* Find next extent map of a given extent map, caller needs to ensure locks */
6808 static struct extent_map *next_extent_map(struct extent_map *em)
6809 {
6810         struct rb_node *next;
6811
6812         next = rb_next(&em->rb_node);
6813         if (!next)
6814                 return NULL;
6815         return container_of(next, struct extent_map, rb_node);
6816 }
6817
6818 static struct extent_map *prev_extent_map(struct extent_map *em)
6819 {
6820         struct rb_node *prev;
6821
6822         prev = rb_prev(&em->rb_node);
6823         if (!prev)
6824                 return NULL;
6825         return container_of(prev, struct extent_map, rb_node);
6826 }
6827
6828 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6829  * the existing extent is the nearest extent to map_start,
6830  * and an extent that you want to insert, deal with overlap and insert
6831  * the best fitted new extent into the tree.
6832  */
6833 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6834                                 struct extent_map *existing,
6835                                 struct extent_map *em,
6836                                 u64 map_start)
6837 {
6838         struct extent_map *prev;
6839         struct extent_map *next;
6840         u64 start;
6841         u64 end;
6842         u64 start_diff;
6843
6844         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6845
6846         if (existing->start > map_start) {
6847                 next = existing;
6848                 prev = prev_extent_map(next);
6849         } else {
6850                 prev = existing;
6851                 next = next_extent_map(prev);
6852         }
6853
6854         start = prev ? extent_map_end(prev) : em->start;
6855         start = max_t(u64, start, em->start);
6856         end = next ? next->start : extent_map_end(em);
6857         end = min_t(u64, end, extent_map_end(em));
6858         start_diff = start - em->start;
6859         em->start = start;
6860         em->len = end - start;
6861         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6862             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6863                 em->block_start += start_diff;
6864                 em->block_len -= start_diff;
6865         }
6866         return add_extent_mapping(em_tree, em, 0);
6867 }
6868
6869 static noinline int uncompress_inline(struct btrfs_path *path,
6870                                       struct page *page,
6871                                       size_t pg_offset, u64 extent_offset,
6872                                       struct btrfs_file_extent_item *item)
6873 {
6874         int ret;
6875         struct extent_buffer *leaf = path->nodes[0];
6876         char *tmp;
6877         size_t max_size;
6878         unsigned long inline_size;
6879         unsigned long ptr;
6880         int compress_type;
6881
6882         WARN_ON(pg_offset != 0);
6883         compress_type = btrfs_file_extent_compression(leaf, item);
6884         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6885         inline_size = btrfs_file_extent_inline_item_len(leaf,
6886                                         btrfs_item_nr(path->slots[0]));
6887         tmp = kmalloc(inline_size, GFP_NOFS);
6888         if (!tmp)
6889                 return -ENOMEM;
6890         ptr = btrfs_file_extent_inline_start(item);
6891
6892         read_extent_buffer(leaf, tmp, ptr, inline_size);
6893
6894         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6895         ret = btrfs_decompress(compress_type, tmp, page,
6896                                extent_offset, inline_size, max_size);
6897
6898         /*
6899          * decompression code contains a memset to fill in any space between the end
6900          * of the uncompressed data and the end of max_size in case the decompressed
6901          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6902          * the end of an inline extent and the beginning of the next block, so we
6903          * cover that region here.
6904          */
6905
6906         if (max_size + pg_offset < PAGE_SIZE) {
6907                 char *map = kmap(page);
6908                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6909                 kunmap(page);
6910         }
6911         kfree(tmp);
6912         return ret;
6913 }
6914
6915 /*
6916  * a bit scary, this does extent mapping from logical file offset to the disk.
6917  * the ugly parts come from merging extents from the disk with the in-ram
6918  * representation.  This gets more complex because of the data=ordered code,
6919  * where the in-ram extents might be locked pending data=ordered completion.
6920  *
6921  * This also copies inline extents directly into the page.
6922  */
6923 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6924                 struct page *page,
6925             size_t pg_offset, u64 start, u64 len,
6926                 int create)
6927 {
6928         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6929         int ret;
6930         int err = 0;
6931         u64 extent_start = 0;
6932         u64 extent_end = 0;
6933         u64 objectid = btrfs_ino(inode);
6934         u32 found_type;
6935         struct btrfs_path *path = NULL;
6936         struct btrfs_root *root = inode->root;
6937         struct btrfs_file_extent_item *item;
6938         struct extent_buffer *leaf;
6939         struct btrfs_key found_key;
6940         struct extent_map *em = NULL;
6941         struct extent_map_tree *em_tree = &inode->extent_tree;
6942         struct extent_io_tree *io_tree = &inode->io_tree;
6943         struct btrfs_trans_handle *trans = NULL;
6944         const bool new_inline = !page || create;
6945
6946 again:
6947         read_lock(&em_tree->lock);
6948         em = lookup_extent_mapping(em_tree, start, len);
6949         if (em)
6950                 em->bdev = fs_info->fs_devices->latest_bdev;
6951         read_unlock(&em_tree->lock);
6952
6953         if (em) {
6954                 if (em->start > start || em->start + em->len <= start)
6955                         free_extent_map(em);
6956                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6957                         free_extent_map(em);
6958                 else
6959                         goto out;
6960         }
6961         em = alloc_extent_map();
6962         if (!em) {
6963                 err = -ENOMEM;
6964                 goto out;
6965         }
6966         em->bdev = fs_info->fs_devices->latest_bdev;
6967         em->start = EXTENT_MAP_HOLE;
6968         em->orig_start = EXTENT_MAP_HOLE;
6969         em->len = (u64)-1;
6970         em->block_len = (u64)-1;
6971
6972         if (!path) {
6973                 path = btrfs_alloc_path();
6974                 if (!path) {
6975                         err = -ENOMEM;
6976                         goto out;
6977                 }
6978                 /*
6979                  * Chances are we'll be called again, so go ahead and do
6980                  * readahead
6981                  */
6982                 path->reada = READA_FORWARD;
6983         }
6984
6985         ret = btrfs_lookup_file_extent(trans, root, path,
6986                                        objectid, start, trans != NULL);
6987         if (ret < 0) {
6988                 err = ret;
6989                 goto out;
6990         }
6991
6992         if (ret != 0) {
6993                 if (path->slots[0] == 0)
6994                         goto not_found;
6995                 path->slots[0]--;
6996         }
6997
6998         leaf = path->nodes[0];
6999         item = btrfs_item_ptr(leaf, path->slots[0],
7000                               struct btrfs_file_extent_item);
7001         /* are we inside the extent that was found? */
7002         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7003         found_type = found_key.type;
7004         if (found_key.objectid != objectid ||
7005             found_type != BTRFS_EXTENT_DATA_KEY) {
7006                 /*
7007                  * If we backup past the first extent we want to move forward
7008                  * and see if there is an extent in front of us, otherwise we'll
7009                  * say there is a hole for our whole search range which can
7010                  * cause problems.
7011                  */
7012                 extent_end = start;
7013                 goto next;
7014         }
7015
7016         found_type = btrfs_file_extent_type(leaf, item);
7017         extent_start = found_key.offset;
7018         if (found_type == BTRFS_FILE_EXTENT_REG ||
7019             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7020                 extent_end = extent_start +
7021                        btrfs_file_extent_num_bytes(leaf, item);
7022
7023                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
7024                                                        extent_start);
7025         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
7026                 size_t size;
7027                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
7028                 extent_end = ALIGN(extent_start + size,
7029                                    fs_info->sectorsize);
7030
7031                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
7032                                                       path->slots[0],
7033                                                       extent_start);
7034         }
7035 next:
7036         if (start >= extent_end) {
7037                 path->slots[0]++;
7038                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
7039                         ret = btrfs_next_leaf(root, path);
7040                         if (ret < 0) {
7041                                 err = ret;
7042                                 goto out;
7043                         }
7044                         if (ret > 0)
7045                                 goto not_found;
7046                         leaf = path->nodes[0];
7047                 }
7048                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7049                 if (found_key.objectid != objectid ||
7050                     found_key.type != BTRFS_EXTENT_DATA_KEY)
7051                         goto not_found;
7052                 if (start + len <= found_key.offset)
7053                         goto not_found;
7054                 if (start > found_key.offset)
7055                         goto next;
7056                 em->start = start;
7057                 em->orig_start = start;
7058                 em->len = found_key.offset - start;
7059                 goto not_found_em;
7060         }
7061
7062         btrfs_extent_item_to_extent_map(inode, path, item,
7063                         new_inline, em);
7064
7065         if (found_type == BTRFS_FILE_EXTENT_REG ||
7066             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7067                 goto insert;
7068         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
7069                 unsigned long ptr;
7070                 char *map;
7071                 size_t size;
7072                 size_t extent_offset;
7073                 size_t copy_size;
7074
7075                 if (new_inline)
7076                         goto out;
7077
7078                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
7079                 extent_offset = page_offset(page) + pg_offset - extent_start;
7080                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
7081                                   size - extent_offset);
7082                 em->start = extent_start + extent_offset;
7083                 em->len = ALIGN(copy_size, fs_info->sectorsize);
7084                 em->orig_block_len = em->len;
7085                 em->orig_start = em->start;
7086                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
7087                 if (create == 0 && !PageUptodate(page)) {
7088                         if (btrfs_file_extent_compression(leaf, item) !=
7089                             BTRFS_COMPRESS_NONE) {
7090                                 ret = uncompress_inline(path, page, pg_offset,
7091                                                         extent_offset, item);
7092                                 if (ret) {
7093                                         err = ret;
7094                                         goto out;
7095                                 }
7096                         } else {
7097                                 map = kmap(page);
7098                                 read_extent_buffer(leaf, map + pg_offset, ptr,
7099                                                    copy_size);
7100                                 if (pg_offset + copy_size < PAGE_SIZE) {
7101                                         memset(map + pg_offset + copy_size, 0,
7102                                                PAGE_SIZE - pg_offset -
7103                                                copy_size);
7104                                 }
7105                                 kunmap(page);
7106                         }
7107                         flush_dcache_page(page);
7108                 } else if (create && PageUptodate(page)) {
7109                         BUG();
7110                         if (!trans) {
7111                                 kunmap(page);
7112                                 free_extent_map(em);
7113                                 em = NULL;
7114
7115                                 btrfs_release_path(path);
7116                                 trans = btrfs_join_transaction(root);
7117
7118                                 if (IS_ERR(trans))
7119                                         return ERR_CAST(trans);
7120                                 goto again;
7121                         }
7122                         map = kmap(page);
7123                         write_extent_buffer(leaf, map + pg_offset, ptr,
7124                                             copy_size);
7125                         kunmap(page);
7126                         btrfs_mark_buffer_dirty(leaf);
7127                 }
7128                 set_extent_uptodate(io_tree, em->start,
7129                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
7130                 goto insert;
7131         }
7132 not_found:
7133         em->start = start;
7134         em->orig_start = start;
7135         em->len = len;
7136 not_found_em:
7137         em->block_start = EXTENT_MAP_HOLE;
7138         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
7139 insert:
7140         btrfs_release_path(path);
7141         if (em->start > start || extent_map_end(em) <= start) {
7142                 btrfs_err(fs_info,
7143                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
7144                           em->start, em->len, start, len);
7145                 err = -EIO;
7146                 goto out;
7147         }
7148
7149         err = 0;
7150         write_lock(&em_tree->lock);
7151         ret = add_extent_mapping(em_tree, em, 0);
7152         /* it is possible that someone inserted the extent into the tree
7153          * while we had the lock dropped.  It is also possible that
7154          * an overlapping map exists in the tree
7155          */
7156         if (ret == -EEXIST) {
7157                 struct extent_map *existing;
7158
7159                 ret = 0;
7160
7161                 existing = search_extent_mapping(em_tree, start, len);
7162                 /*
7163                  * existing will always be non-NULL, since there must be
7164                  * extent causing the -EEXIST.
7165                  */
7166                 if (existing->start == em->start &&
7167                     extent_map_end(existing) >= extent_map_end(em) &&
7168                     em->block_start == existing->block_start) {
7169                         /*
7170                          * The existing extent map already encompasses the
7171                          * entire extent map we tried to add.
7172                          */
7173                         free_extent_map(em);
7174                         em = existing;
7175                         err = 0;
7176
7177                 } else if (start >= extent_map_end(existing) ||
7178                     start <= existing->start) {
7179                         /*
7180                          * The existing extent map is the one nearest to
7181                          * the [start, start + len) range which overlaps
7182                          */
7183                         err = merge_extent_mapping(em_tree, existing,
7184                                                    em, start);
7185                         free_extent_map(existing);
7186                         if (err) {
7187                                 free_extent_map(em);
7188                                 em = NULL;
7189                         }
7190                 } else {
7191                         free_extent_map(em);
7192                         em = existing;
7193                         err = 0;
7194                 }
7195         }
7196         write_unlock(&em_tree->lock);
7197 out:
7198
7199         trace_btrfs_get_extent(root, inode, em);
7200
7201         btrfs_free_path(path);
7202         if (trans) {
7203                 ret = btrfs_end_transaction(trans);
7204                 if (!err)
7205                         err = ret;
7206         }
7207         if (err) {
7208                 free_extent_map(em);
7209                 return ERR_PTR(err);
7210         }
7211         BUG_ON(!em); /* Error is always set */
7212         return em;
7213 }
7214
7215 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
7216                 struct page *page,
7217                 size_t pg_offset, u64 start, u64 len,
7218                 int create)
7219 {
7220         struct extent_map *em;
7221         struct extent_map *hole_em = NULL;
7222         u64 range_start = start;
7223         u64 end;
7224         u64 found;
7225         u64 found_end;
7226         int err = 0;
7227
7228         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7229         if (IS_ERR(em))
7230                 return em;
7231         /*
7232          * If our em maps to:
7233          * - a hole or
7234          * - a pre-alloc extent,
7235          * there might actually be delalloc bytes behind it.
7236          */
7237         if (em->block_start != EXTENT_MAP_HOLE &&
7238             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7239                 return em;
7240         else
7241                 hole_em = em;
7242
7243         /* check to see if we've wrapped (len == -1 or similar) */
7244         end = start + len;
7245         if (end < start)
7246                 end = (u64)-1;
7247         else
7248                 end -= 1;
7249
7250         em = NULL;
7251
7252         /* ok, we didn't find anything, lets look for delalloc */
7253         found = count_range_bits(&inode->io_tree, &range_start,
7254                                  end, len, EXTENT_DELALLOC, 1);
7255         found_end = range_start + found;
7256         if (found_end < range_start)
7257                 found_end = (u64)-1;
7258
7259         /*
7260          * we didn't find anything useful, return
7261          * the original results from get_extent()
7262          */
7263         if (range_start > end || found_end <= start) {
7264                 em = hole_em;
7265                 hole_em = NULL;
7266                 goto out;
7267         }
7268
7269         /* adjust the range_start to make sure it doesn't
7270          * go backwards from the start they passed in
7271          */
7272         range_start = max(start, range_start);
7273         found = found_end - range_start;
7274
7275         if (found > 0) {
7276                 u64 hole_start = start;
7277                 u64 hole_len = len;
7278
7279                 em = alloc_extent_map();
7280                 if (!em) {
7281                         err = -ENOMEM;
7282                         goto out;
7283                 }
7284                 /*
7285                  * when btrfs_get_extent can't find anything it
7286                  * returns one huge hole
7287                  *
7288                  * make sure what it found really fits our range, and
7289                  * adjust to make sure it is based on the start from
7290                  * the caller
7291                  */
7292                 if (hole_em) {
7293                         u64 calc_end = extent_map_end(hole_em);
7294
7295                         if (calc_end <= start || (hole_em->start > end)) {
7296                                 free_extent_map(hole_em);
7297                                 hole_em = NULL;
7298                         } else {
7299                                 hole_start = max(hole_em->start, start);
7300                                 hole_len = calc_end - hole_start;
7301                         }
7302                 }
7303                 em->bdev = NULL;
7304                 if (hole_em && range_start > hole_start) {
7305                         /* our hole starts before our delalloc, so we
7306                          * have to return just the parts of the hole
7307                          * that go until  the delalloc starts
7308                          */
7309                         em->len = min(hole_len,
7310                                       range_start - hole_start);
7311                         em->start = hole_start;
7312                         em->orig_start = hole_start;
7313                         /*
7314                          * don't adjust block start at all,
7315                          * it is fixed at EXTENT_MAP_HOLE
7316                          */
7317                         em->block_start = hole_em->block_start;
7318                         em->block_len = hole_len;
7319                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7320                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7321                 } else {
7322                         em->start = range_start;
7323                         em->len = found;
7324                         em->orig_start = range_start;
7325                         em->block_start = EXTENT_MAP_DELALLOC;
7326                         em->block_len = found;
7327                 }
7328         } else if (hole_em) {
7329                 return hole_em;
7330         }
7331 out:
7332
7333         free_extent_map(hole_em);
7334         if (err) {
7335                 free_extent_map(em);
7336                 return ERR_PTR(err);
7337         }
7338         return em;
7339 }
7340
7341 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7342                                                   const u64 start,
7343                                                   const u64 len,
7344                                                   const u64 orig_start,
7345                                                   const u64 block_start,
7346                                                   const u64 block_len,
7347                                                   const u64 orig_block_len,
7348                                                   const u64 ram_bytes,
7349                                                   const int type)
7350 {
7351         struct extent_map *em = NULL;
7352         int ret;
7353
7354         if (type != BTRFS_ORDERED_NOCOW) {
7355                 em = create_io_em(inode, start, len, orig_start,
7356                                   block_start, block_len, orig_block_len,
7357                                   ram_bytes,
7358                                   BTRFS_COMPRESS_NONE, /* compress_type */
7359                                   type);
7360                 if (IS_ERR(em))
7361                         goto out;
7362         }
7363         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7364                                            len, block_len, type);
7365         if (ret) {
7366                 if (em) {
7367                         free_extent_map(em);
7368                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
7369                                                 start + len - 1, 0);
7370                 }
7371                 em = ERR_PTR(ret);
7372         }
7373  out:
7374
7375         return em;
7376 }
7377
7378 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7379                                                   u64 start, u64 len)
7380 {
7381         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7382         struct btrfs_root *root = BTRFS_I(inode)->root;
7383         struct extent_map *em;
7384         struct btrfs_key ins;
7385         u64 alloc_hint;
7386         int ret;
7387
7388         alloc_hint = get_extent_allocation_hint(inode, start, len);
7389         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
7390                                    0, alloc_hint, &ins, 1, 1);
7391         if (ret)
7392                 return ERR_PTR(ret);
7393
7394         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7395                                      ins.objectid, ins.offset, ins.offset,
7396                                      ins.offset, BTRFS_ORDERED_REGULAR);
7397         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
7398         if (IS_ERR(em))
7399                 btrfs_free_reserved_extent(fs_info, ins.objectid,
7400                                            ins.offset, 1);
7401
7402         return em;
7403 }
7404
7405 /*
7406  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7407  * block must be cow'd
7408  */
7409 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7410                               u64 *orig_start, u64 *orig_block_len,
7411                               u64 *ram_bytes)
7412 {
7413         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7414         struct btrfs_path *path;
7415         int ret;
7416         struct extent_buffer *leaf;
7417         struct btrfs_root *root = BTRFS_I(inode)->root;
7418         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7419         struct btrfs_file_extent_item *fi;
7420         struct btrfs_key key;
7421         u64 disk_bytenr;
7422         u64 backref_offset;
7423         u64 extent_end;
7424         u64 num_bytes;
7425         int slot;
7426         int found_type;
7427         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7428
7429         path = btrfs_alloc_path();
7430         if (!path)
7431                 return -ENOMEM;
7432
7433         ret = btrfs_lookup_file_extent(NULL, root, path,
7434                         btrfs_ino(BTRFS_I(inode)), offset, 0);
7435         if (ret < 0)
7436                 goto out;
7437
7438         slot = path->slots[0];
7439         if (ret == 1) {
7440                 if (slot == 0) {
7441                         /* can't find the item, must cow */
7442                         ret = 0;
7443                         goto out;
7444                 }
7445                 slot--;
7446         }
7447         ret = 0;
7448         leaf = path->nodes[0];
7449         btrfs_item_key_to_cpu(leaf, &key, slot);
7450         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
7451             key.type != BTRFS_EXTENT_DATA_KEY) {
7452                 /* not our file or wrong item type, must cow */
7453                 goto out;
7454         }
7455
7456         if (key.offset > offset) {
7457                 /* Wrong offset, must cow */
7458                 goto out;
7459         }
7460
7461         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7462         found_type = btrfs_file_extent_type(leaf, fi);
7463         if (found_type != BTRFS_FILE_EXTENT_REG &&
7464             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7465                 /* not a regular extent, must cow */
7466                 goto out;
7467         }
7468
7469         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7470                 goto out;
7471
7472         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7473         if (extent_end <= offset)
7474                 goto out;
7475
7476         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7477         if (disk_bytenr == 0)
7478                 goto out;
7479
7480         if (btrfs_file_extent_compression(leaf, fi) ||
7481             btrfs_file_extent_encryption(leaf, fi) ||
7482             btrfs_file_extent_other_encoding(leaf, fi))
7483                 goto out;
7484
7485         backref_offset = btrfs_file_extent_offset(leaf, fi);
7486
7487         if (orig_start) {
7488                 *orig_start = key.offset - backref_offset;
7489                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7490                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7491         }
7492
7493         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7494                 goto out;
7495
7496         num_bytes = min(offset + *len, extent_end) - offset;
7497         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7498                 u64 range_end;
7499
7500                 range_end = round_up(offset + num_bytes,
7501                                      root->fs_info->sectorsize) - 1;
7502                 ret = test_range_bit(io_tree, offset, range_end,
7503                                      EXTENT_DELALLOC, 0, NULL);
7504                 if (ret) {
7505                         ret = -EAGAIN;
7506                         goto out;
7507                 }
7508         }
7509
7510         btrfs_release_path(path);
7511
7512         /*
7513          * look for other files referencing this extent, if we
7514          * find any we must cow
7515          */
7516
7517         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7518                                     key.offset - backref_offset, disk_bytenr);
7519         if (ret) {
7520                 ret = 0;
7521                 goto out;
7522         }
7523
7524         /*
7525          * adjust disk_bytenr and num_bytes to cover just the bytes
7526          * in this extent we are about to write.  If there
7527          * are any csums in that range we have to cow in order
7528          * to keep the csums correct
7529          */
7530         disk_bytenr += backref_offset;
7531         disk_bytenr += offset - key.offset;
7532         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7533                 goto out;
7534         /*
7535          * all of the above have passed, it is safe to overwrite this extent
7536          * without cow
7537          */
7538         *len = num_bytes;
7539         ret = 1;
7540 out:
7541         btrfs_free_path(path);
7542         return ret;
7543 }
7544
7545 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7546 {
7547         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7548         bool found = false;
7549         void **pagep = NULL;
7550         struct page *page = NULL;
7551         unsigned long start_idx;
7552         unsigned long end_idx;
7553
7554         start_idx = start >> PAGE_SHIFT;
7555
7556         /*
7557          * end is the last byte in the last page.  end == start is legal
7558          */
7559         end_idx = end >> PAGE_SHIFT;
7560
7561         rcu_read_lock();
7562
7563         /* Most of the code in this while loop is lifted from
7564          * find_get_page.  It's been modified to begin searching from a
7565          * page and return just the first page found in that range.  If the
7566          * found idx is less than or equal to the end idx then we know that
7567          * a page exists.  If no pages are found or if those pages are
7568          * outside of the range then we're fine (yay!) */
7569         while (page == NULL &&
7570                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7571                 page = radix_tree_deref_slot(pagep);
7572                 if (unlikely(!page))
7573                         break;
7574
7575                 if (radix_tree_exception(page)) {
7576                         if (radix_tree_deref_retry(page)) {
7577                                 page = NULL;
7578                                 continue;
7579                         }
7580                         /*
7581                          * Otherwise, shmem/tmpfs must be storing a swap entry
7582                          * here as an exceptional entry: so return it without
7583                          * attempting to raise page count.
7584                          */
7585                         page = NULL;
7586                         break; /* TODO: Is this relevant for this use case? */
7587                 }
7588
7589                 if (!page_cache_get_speculative(page)) {
7590                         page = NULL;
7591                         continue;
7592                 }
7593
7594                 /*
7595                  * Has the page moved?
7596                  * This is part of the lockless pagecache protocol. See
7597                  * include/linux/pagemap.h for details.
7598                  */
7599                 if (unlikely(page != *pagep)) {
7600                         put_page(page);
7601                         page = NULL;
7602                 }
7603         }
7604
7605         if (page) {
7606                 if (page->index <= end_idx)
7607                         found = true;
7608                 put_page(page);
7609         }
7610
7611         rcu_read_unlock();
7612         return found;
7613 }
7614
7615 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7616                               struct extent_state **cached_state, int writing)
7617 {
7618         struct btrfs_ordered_extent *ordered;
7619         int ret = 0;
7620
7621         while (1) {
7622                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7623                                  cached_state);
7624                 /*
7625                  * We're concerned with the entire range that we're going to be
7626                  * doing DIO to, so we need to make sure there's no ordered
7627                  * extents in this range.
7628                  */
7629                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7630                                                      lockend - lockstart + 1);
7631
7632                 /*
7633                  * We need to make sure there are no buffered pages in this
7634                  * range either, we could have raced between the invalidate in
7635                  * generic_file_direct_write and locking the extent.  The
7636                  * invalidate needs to happen so that reads after a write do not
7637                  * get stale data.
7638                  */
7639                 if (!ordered &&
7640                     (!writing ||
7641                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7642                         break;
7643
7644                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7645                                      cached_state, GFP_NOFS);
7646
7647                 if (ordered) {
7648                         /*
7649                          * If we are doing a DIO read and the ordered extent we
7650                          * found is for a buffered write, we can not wait for it
7651                          * to complete and retry, because if we do so we can
7652                          * deadlock with concurrent buffered writes on page
7653                          * locks. This happens only if our DIO read covers more
7654                          * than one extent map, if at this point has already
7655                          * created an ordered extent for a previous extent map
7656                          * and locked its range in the inode's io tree, and a
7657                          * concurrent write against that previous extent map's
7658                          * range and this range started (we unlock the ranges
7659                          * in the io tree only when the bios complete and
7660                          * buffered writes always lock pages before attempting
7661                          * to lock range in the io tree).
7662                          */
7663                         if (writing ||
7664                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7665                                 btrfs_start_ordered_extent(inode, ordered, 1);
7666                         else
7667                                 ret = -ENOTBLK;
7668                         btrfs_put_ordered_extent(ordered);
7669                 } else {
7670                         /*
7671                          * We could trigger writeback for this range (and wait
7672                          * for it to complete) and then invalidate the pages for
7673                          * this range (through invalidate_inode_pages2_range()),
7674                          * but that can lead us to a deadlock with a concurrent
7675                          * call to readpages() (a buffered read or a defrag call
7676                          * triggered a readahead) on a page lock due to an
7677                          * ordered dio extent we created before but did not have
7678                          * yet a corresponding bio submitted (whence it can not
7679                          * complete), which makes readpages() wait for that
7680                          * ordered extent to complete while holding a lock on
7681                          * that page.
7682                          */
7683                         ret = -ENOTBLK;
7684                 }
7685
7686                 if (ret)
7687                         break;
7688
7689                 cond_resched();
7690         }
7691
7692         return ret;
7693 }
7694
7695 /* The callers of this must take lock_extent() */
7696 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7697                                        u64 orig_start, u64 block_start,
7698                                        u64 block_len, u64 orig_block_len,
7699                                        u64 ram_bytes, int compress_type,
7700                                        int type)
7701 {
7702         struct extent_map_tree *em_tree;
7703         struct extent_map *em;
7704         struct btrfs_root *root = BTRFS_I(inode)->root;
7705         int ret;
7706
7707         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7708                type == BTRFS_ORDERED_COMPRESSED ||
7709                type == BTRFS_ORDERED_NOCOW ||
7710                type == BTRFS_ORDERED_REGULAR);
7711
7712         em_tree = &BTRFS_I(inode)->extent_tree;
7713         em = alloc_extent_map();
7714         if (!em)
7715                 return ERR_PTR(-ENOMEM);
7716
7717         em->start = start;
7718         em->orig_start = orig_start;
7719         em->len = len;
7720         em->block_len = block_len;
7721         em->block_start = block_start;
7722         em->bdev = root->fs_info->fs_devices->latest_bdev;
7723         em->orig_block_len = orig_block_len;
7724         em->ram_bytes = ram_bytes;
7725         em->generation = -1;
7726         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7727         if (type == BTRFS_ORDERED_PREALLOC) {
7728                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7729         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7730                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7731                 em->compress_type = compress_type;
7732         }
7733
7734         do {
7735                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7736                                 em->start + em->len - 1, 0);
7737                 write_lock(&em_tree->lock);
7738                 ret = add_extent_mapping(em_tree, em, 1);
7739                 write_unlock(&em_tree->lock);
7740                 /*
7741                  * The caller has taken lock_extent(), who could race with us
7742                  * to add em?
7743                  */
7744         } while (ret == -EEXIST);
7745
7746         if (ret) {
7747                 free_extent_map(em);
7748                 return ERR_PTR(ret);
7749         }
7750
7751         /* em got 2 refs now, callers needs to do free_extent_map once. */
7752         return em;
7753 }
7754
7755 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7756                                    struct buffer_head *bh_result, int create)
7757 {
7758         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7759         struct extent_map *em;
7760         struct extent_state *cached_state = NULL;
7761         struct btrfs_dio_data *dio_data = NULL;
7762         u64 start = iblock << inode->i_blkbits;
7763         u64 lockstart, lockend;
7764         u64 len = bh_result->b_size;
7765         int unlock_bits = EXTENT_LOCKED;
7766         int ret = 0;
7767
7768         if (create)
7769                 unlock_bits |= EXTENT_DIRTY;
7770         else
7771                 len = min_t(u64, len, fs_info->sectorsize);
7772
7773         lockstart = start;
7774         lockend = start + len - 1;
7775
7776         if (current->journal_info) {
7777                 /*
7778                  * Need to pull our outstanding extents and set journal_info to NULL so
7779                  * that anything that needs to check if there's a transaction doesn't get
7780                  * confused.
7781                  */
7782                 dio_data = current->journal_info;
7783                 current->journal_info = NULL;
7784         }
7785
7786         /*
7787          * If this errors out it's because we couldn't invalidate pagecache for
7788          * this range and we need to fallback to buffered.
7789          */
7790         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7791                                create)) {
7792                 ret = -ENOTBLK;
7793                 goto err;
7794         }
7795
7796         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
7797         if (IS_ERR(em)) {
7798                 ret = PTR_ERR(em);
7799                 goto unlock_err;
7800         }
7801
7802         /*
7803          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7804          * io.  INLINE is special, and we could probably kludge it in here, but
7805          * it's still buffered so for safety lets just fall back to the generic
7806          * buffered path.
7807          *
7808          * For COMPRESSED we _have_ to read the entire extent in so we can
7809          * decompress it, so there will be buffering required no matter what we
7810          * do, so go ahead and fallback to buffered.
7811          *
7812          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7813          * to buffered IO.  Don't blame me, this is the price we pay for using
7814          * the generic code.
7815          */
7816         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7817             em->block_start == EXTENT_MAP_INLINE) {
7818                 free_extent_map(em);
7819                 ret = -ENOTBLK;
7820                 goto unlock_err;
7821         }
7822
7823         /* Just a good old fashioned hole, return */
7824         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7825                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7826                 free_extent_map(em);
7827                 goto unlock_err;
7828         }
7829
7830         /*
7831          * We don't allocate a new extent in the following cases
7832          *
7833          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7834          * existing extent.
7835          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7836          * just use the extent.
7837          *
7838          */
7839         if (!create) {
7840                 len = min(len, em->len - (start - em->start));
7841                 lockstart = start + len;
7842                 goto unlock;
7843         }
7844
7845         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7846             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7847              em->block_start != EXTENT_MAP_HOLE)) {
7848                 int type;
7849                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7850
7851                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7852                         type = BTRFS_ORDERED_PREALLOC;
7853                 else
7854                         type = BTRFS_ORDERED_NOCOW;
7855                 len = min(len, em->len - (start - em->start));
7856                 block_start = em->block_start + (start - em->start);
7857
7858                 if (can_nocow_extent(inode, start, &len, &orig_start,
7859                                      &orig_block_len, &ram_bytes) == 1 &&
7860                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7861                         struct extent_map *em2;
7862
7863                         em2 = btrfs_create_dio_extent(inode, start, len,
7864                                                       orig_start, block_start,
7865                                                       len, orig_block_len,
7866                                                       ram_bytes, type);
7867                         btrfs_dec_nocow_writers(fs_info, block_start);
7868                         if (type == BTRFS_ORDERED_PREALLOC) {
7869                                 free_extent_map(em);
7870                                 em = em2;
7871                         }
7872                         if (em2 && IS_ERR(em2)) {
7873                                 ret = PTR_ERR(em2);
7874                                 goto unlock_err;
7875                         }
7876                         /*
7877                          * For inode marked NODATACOW or extent marked PREALLOC,
7878                          * use the existing or preallocated extent, so does not
7879                          * need to adjust btrfs_space_info's bytes_may_use.
7880                          */
7881                         btrfs_free_reserved_data_space_noquota(inode,
7882                                         start, len);
7883                         goto unlock;
7884                 }
7885         }
7886
7887         /*
7888          * this will cow the extent, reset the len in case we changed
7889          * it above
7890          */
7891         len = bh_result->b_size;
7892         free_extent_map(em);
7893         em = btrfs_new_extent_direct(inode, start, len);
7894         if (IS_ERR(em)) {
7895                 ret = PTR_ERR(em);
7896                 goto unlock_err;
7897         }
7898         len = min(len, em->len - (start - em->start));
7899 unlock:
7900         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7901                 inode->i_blkbits;
7902         bh_result->b_size = len;
7903         bh_result->b_bdev = em->bdev;
7904         set_buffer_mapped(bh_result);
7905         if (create) {
7906                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7907                         set_buffer_new(bh_result);
7908
7909                 /*
7910                  * Need to update the i_size under the extent lock so buffered
7911                  * readers will get the updated i_size when we unlock.
7912                  */
7913                 if (!dio_data->overwrite && start + len > i_size_read(inode))
7914                         i_size_write(inode, start + len);
7915
7916                 WARN_ON(dio_data->reserve < len);
7917                 dio_data->reserve -= len;
7918                 dio_data->unsubmitted_oe_range_end = start + len;
7919                 current->journal_info = dio_data;
7920         }
7921
7922         /*
7923          * In the case of write we need to clear and unlock the entire range,
7924          * in the case of read we need to unlock only the end area that we
7925          * aren't using if there is any left over space.
7926          */
7927         if (lockstart < lockend) {
7928                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7929                                  lockend, unlock_bits, 1, 0,
7930                                  &cached_state, GFP_NOFS);
7931         } else {
7932                 free_extent_state(cached_state);
7933         }
7934
7935         free_extent_map(em);
7936
7937         return 0;
7938
7939 unlock_err:
7940         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7941                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
7942 err:
7943         if (dio_data)
7944                 current->journal_info = dio_data;
7945         return ret;
7946 }
7947
7948 static inline blk_status_t submit_dio_repair_bio(struct inode *inode,
7949                                                  struct bio *bio,
7950                                                  int mirror_num)
7951 {
7952         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7953         blk_status_t ret;
7954
7955         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7956
7957         bio_get(bio);
7958
7959         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
7960         if (ret)
7961                 goto err;
7962
7963         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
7964 err:
7965         bio_put(bio);
7966         return ret;
7967 }
7968
7969 static int btrfs_check_dio_repairable(struct inode *inode,
7970                                       struct bio *failed_bio,
7971                                       struct io_failure_record *failrec,
7972                                       int failed_mirror)
7973 {
7974         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7975         int num_copies;
7976
7977         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
7978         if (num_copies == 1) {
7979                 /*
7980                  * we only have a single copy of the data, so don't bother with
7981                  * all the retry and error correction code that follows. no
7982                  * matter what the error is, it is very likely to persist.
7983                  */
7984                 btrfs_debug(fs_info,
7985                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
7986                         num_copies, failrec->this_mirror, failed_mirror);
7987                 return 0;
7988         }
7989
7990         failrec->failed_mirror = failed_mirror;
7991         failrec->this_mirror++;
7992         if (failrec->this_mirror == failed_mirror)
7993                 failrec->this_mirror++;
7994
7995         if (failrec->this_mirror > num_copies) {
7996                 btrfs_debug(fs_info,
7997                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
7998                         num_copies, failrec->this_mirror, failed_mirror);
7999                 return 0;
8000         }
8001
8002         return 1;
8003 }
8004
8005 static blk_status_t dio_read_error(struct inode *inode, struct bio *failed_bio,
8006                                    struct page *page, unsigned int pgoff,
8007                                    u64 start, u64 end, int failed_mirror,
8008                                    bio_end_io_t *repair_endio, void *repair_arg)
8009 {
8010         struct io_failure_record *failrec;
8011         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8012         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
8013         struct bio *bio;
8014         int isector;
8015         unsigned int read_mode = 0;
8016         int segs;
8017         int ret;
8018         blk_status_t status;
8019
8020         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
8021
8022         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
8023         if (ret)
8024                 return errno_to_blk_status(ret);
8025
8026         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
8027                                          failed_mirror);
8028         if (!ret) {
8029                 free_io_failure(failure_tree, io_tree, failrec);
8030                 return BLK_STS_IOERR;
8031         }
8032
8033         segs = bio_segments(failed_bio);
8034         if (segs > 1 ||
8035             (failed_bio->bi_io_vec->bv_len > btrfs_inode_sectorsize(inode)))
8036                 read_mode |= REQ_FAILFAST_DEV;
8037
8038         isector = start - btrfs_io_bio(failed_bio)->logical;
8039         isector >>= inode->i_sb->s_blocksize_bits;
8040         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
8041                                 pgoff, isector, repair_endio, repair_arg);
8042         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
8043
8044         btrfs_debug(BTRFS_I(inode)->root->fs_info,
8045                     "repair DIO read error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d",
8046                     read_mode, failrec->this_mirror, failrec->in_validation);
8047
8048         status = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
8049         if (status) {
8050                 free_io_failure(failure_tree, io_tree, failrec);
8051                 bio_put(bio);
8052         }
8053
8054         return status;
8055 }
8056
8057 struct btrfs_retry_complete {
8058         struct completion done;
8059         struct inode *inode;
8060         u64 start;
8061         int uptodate;
8062 };
8063
8064 static void btrfs_retry_endio_nocsum(struct bio *bio)
8065 {
8066         struct btrfs_retry_complete *done = bio->bi_private;
8067         struct inode *inode = done->inode;
8068         struct bio_vec *bvec;
8069         struct extent_io_tree *io_tree, *failure_tree;
8070         int i;
8071
8072         if (bio->bi_status)
8073                 goto end;
8074
8075         ASSERT(bio->bi_vcnt == 1);
8076         io_tree = &BTRFS_I(inode)->io_tree;
8077         failure_tree = &BTRFS_I(inode)->io_failure_tree;
8078         ASSERT(bio->bi_io_vec->bv_len == btrfs_inode_sectorsize(inode));
8079
8080         done->uptodate = 1;
8081         ASSERT(!bio_flagged(bio, BIO_CLONED));
8082         bio_for_each_segment_all(bvec, bio, i)
8083                 clean_io_failure(BTRFS_I(inode)->root->fs_info, failure_tree,
8084                                  io_tree, done->start, bvec->bv_page,
8085                                  btrfs_ino(BTRFS_I(inode)), 0);
8086 end:
8087         complete(&done->done);
8088         bio_put(bio);
8089 }
8090
8091 static blk_status_t __btrfs_correct_data_nocsum(struct inode *inode,
8092                                                 struct btrfs_io_bio *io_bio)
8093 {
8094         struct btrfs_fs_info *fs_info;
8095         struct bio_vec bvec;
8096         struct bvec_iter iter;
8097         struct btrfs_retry_complete done;
8098         u64 start;
8099         unsigned int pgoff;
8100         u32 sectorsize;
8101         int nr_sectors;
8102         blk_status_t ret;
8103         blk_status_t err = BLK_STS_OK;
8104
8105         fs_info = BTRFS_I(inode)->root->fs_info;
8106         sectorsize = fs_info->sectorsize;
8107
8108         start = io_bio->logical;
8109         done.inode = inode;
8110         io_bio->bio.bi_iter = io_bio->iter;
8111
8112         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8113                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8114                 pgoff = bvec.bv_offset;
8115
8116 next_block_or_try_again:
8117                 done.uptodate = 0;
8118                 done.start = start;
8119                 init_completion(&done.done);
8120
8121                 ret = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8122                                 pgoff, start, start + sectorsize - 1,
8123                                 io_bio->mirror_num,
8124                                 btrfs_retry_endio_nocsum, &done);
8125                 if (ret) {
8126                         err = ret;
8127                         goto next;
8128                 }
8129
8130                 wait_for_completion_io(&done.done);
8131
8132                 if (!done.uptodate) {
8133                         /* We might have another mirror, so try again */
8134                         goto next_block_or_try_again;
8135                 }
8136
8137 next:
8138                 start += sectorsize;
8139
8140                 nr_sectors--;
8141                 if (nr_sectors) {
8142                         pgoff += sectorsize;
8143                         ASSERT(pgoff < PAGE_SIZE);
8144                         goto next_block_or_try_again;
8145                 }
8146         }
8147
8148         return err;
8149 }
8150
8151 static void btrfs_retry_endio(struct bio *bio)
8152 {
8153         struct btrfs_retry_complete *done = bio->bi_private;
8154         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8155         struct extent_io_tree *io_tree, *failure_tree;
8156         struct inode *inode = done->inode;
8157         struct bio_vec *bvec;
8158         int uptodate;
8159         int ret;
8160         int i;
8161
8162         if (bio->bi_status)
8163                 goto end;
8164
8165         uptodate = 1;
8166
8167         ASSERT(bio->bi_vcnt == 1);
8168         ASSERT(bio->bi_io_vec->bv_len == btrfs_inode_sectorsize(done->inode));
8169
8170         io_tree = &BTRFS_I(inode)->io_tree;
8171         failure_tree = &BTRFS_I(inode)->io_failure_tree;
8172
8173         ASSERT(!bio_flagged(bio, BIO_CLONED));
8174         bio_for_each_segment_all(bvec, bio, i) {
8175                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
8176                                              bvec->bv_offset, done->start,
8177                                              bvec->bv_len);
8178                 if (!ret)
8179                         clean_io_failure(BTRFS_I(inode)->root->fs_info,
8180                                          failure_tree, io_tree, done->start,
8181                                          bvec->bv_page,
8182                                          btrfs_ino(BTRFS_I(inode)),
8183                                          bvec->bv_offset);
8184                 else
8185                         uptodate = 0;
8186         }
8187
8188         done->uptodate = uptodate;
8189 end:
8190         complete(&done->done);
8191         bio_put(bio);
8192 }
8193
8194 static blk_status_t __btrfs_subio_endio_read(struct inode *inode,
8195                 struct btrfs_io_bio *io_bio, blk_status_t err)
8196 {
8197         struct btrfs_fs_info *fs_info;
8198         struct bio_vec bvec;
8199         struct bvec_iter iter;
8200         struct btrfs_retry_complete done;
8201         u64 start;
8202         u64 offset = 0;
8203         u32 sectorsize;
8204         int nr_sectors;
8205         unsigned int pgoff;
8206         int csum_pos;
8207         bool uptodate = (err == 0);
8208         int ret;
8209         blk_status_t status;
8210
8211         fs_info = BTRFS_I(inode)->root->fs_info;
8212         sectorsize = fs_info->sectorsize;
8213
8214         err = BLK_STS_OK;
8215         start = io_bio->logical;
8216         done.inode = inode;
8217         io_bio->bio.bi_iter = io_bio->iter;
8218
8219         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8220                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8221
8222                 pgoff = bvec.bv_offset;
8223 next_block:
8224                 if (uptodate) {
8225                         csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8226                         ret = __readpage_endio_check(inode, io_bio, csum_pos,
8227                                         bvec.bv_page, pgoff, start, sectorsize);
8228                         if (likely(!ret))
8229                                 goto next;
8230                 }
8231 try_again:
8232                 done.uptodate = 0;
8233                 done.start = start;
8234                 init_completion(&done.done);
8235
8236                 status = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8237                                         pgoff, start, start + sectorsize - 1,
8238                                         io_bio->mirror_num, btrfs_retry_endio,
8239                                         &done);
8240                 if (status) {
8241                         err = status;
8242                         goto next;
8243                 }
8244
8245                 wait_for_completion_io(&done.done);
8246
8247                 if (!done.uptodate) {
8248                         /* We might have another mirror, so try again */
8249                         goto try_again;
8250                 }
8251 next:
8252                 offset += sectorsize;
8253                 start += sectorsize;
8254
8255                 ASSERT(nr_sectors);
8256
8257                 nr_sectors--;
8258                 if (nr_sectors) {
8259                         pgoff += sectorsize;
8260                         ASSERT(pgoff < PAGE_SIZE);
8261                         goto next_block;
8262                 }
8263         }
8264
8265         return err;
8266 }
8267
8268 static blk_status_t btrfs_subio_endio_read(struct inode *inode,
8269                 struct btrfs_io_bio *io_bio, blk_status_t err)
8270 {
8271         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8272
8273         if (skip_csum) {
8274                 if (unlikely(err))
8275                         return __btrfs_correct_data_nocsum(inode, io_bio);
8276                 else
8277                         return BLK_STS_OK;
8278         } else {
8279                 return __btrfs_subio_endio_read(inode, io_bio, err);
8280         }
8281 }
8282
8283 static void btrfs_endio_direct_read(struct bio *bio)
8284 {
8285         struct btrfs_dio_private *dip = bio->bi_private;
8286         struct inode *inode = dip->inode;
8287         struct bio *dio_bio;
8288         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8289         blk_status_t err = bio->bi_status;
8290
8291         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8292                 err = btrfs_subio_endio_read(inode, io_bio, err);
8293
8294         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8295                       dip->logical_offset + dip->bytes - 1);
8296         dio_bio = dip->dio_bio;
8297
8298         kfree(dip);
8299
8300         dio_bio->bi_status = err;
8301         dio_end_io(dio_bio);
8302
8303         if (io_bio->end_io)
8304                 io_bio->end_io(io_bio, blk_status_to_errno(err));
8305         bio_put(bio);
8306 }
8307
8308 static void __endio_write_update_ordered(struct inode *inode,
8309                                          const u64 offset, const u64 bytes,
8310                                          const bool uptodate)
8311 {
8312         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8313         struct btrfs_ordered_extent *ordered = NULL;
8314         struct btrfs_workqueue *wq;
8315         btrfs_work_func_t func;
8316         u64 ordered_offset = offset;
8317         u64 ordered_bytes = bytes;
8318         u64 last_offset;
8319         int ret;
8320
8321         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
8322                 wq = fs_info->endio_freespace_worker;
8323                 func = btrfs_freespace_write_helper;
8324         } else {
8325                 wq = fs_info->endio_write_workers;
8326                 func = btrfs_endio_write_helper;
8327         }
8328
8329 again:
8330         last_offset = ordered_offset;
8331         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8332                                                    &ordered_offset,
8333                                                    ordered_bytes,
8334                                                    uptodate);
8335         if (!ret)
8336                 goto out_test;
8337
8338         btrfs_init_work(&ordered->work, func, finish_ordered_fn, NULL, NULL);
8339         btrfs_queue_work(wq, &ordered->work);
8340 out_test:
8341         /*
8342          * If btrfs_dec_test_ordered_pending does not find any ordered extent
8343          * in the range, we can exit.
8344          */
8345         if (ordered_offset == last_offset)
8346                 return;
8347         /*
8348          * our bio might span multiple ordered extents.  If we haven't
8349          * completed the accounting for the whole dio, go back and try again
8350          */
8351         if (ordered_offset < offset + bytes) {
8352                 ordered_bytes = offset + bytes - ordered_offset;
8353                 ordered = NULL;
8354                 goto again;
8355         }
8356 }
8357
8358 static void btrfs_endio_direct_write(struct bio *bio)
8359 {
8360         struct btrfs_dio_private *dip = bio->bi_private;
8361         struct bio *dio_bio = dip->dio_bio;
8362
8363         __endio_write_update_ordered(dip->inode, dip->logical_offset,
8364                                      dip->bytes, !bio->bi_status);
8365
8366         kfree(dip);
8367
8368         dio_bio->bi_status = bio->bi_status;
8369         dio_end_io(dio_bio);
8370         bio_put(bio);
8371 }
8372
8373 static blk_status_t __btrfs_submit_bio_start_direct_io(void *private_data,
8374                                     struct bio *bio, int mirror_num,
8375                                     unsigned long bio_flags, u64 offset)
8376 {
8377         struct inode *inode = private_data;
8378         blk_status_t ret;
8379         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
8380         BUG_ON(ret); /* -ENOMEM */
8381         return 0;
8382 }
8383
8384 static void btrfs_end_dio_bio(struct bio *bio)
8385 {
8386         struct btrfs_dio_private *dip = bio->bi_private;
8387         blk_status_t err = bio->bi_status;
8388
8389         if (err)
8390                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8391                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8392                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
8393                            bio->bi_opf,
8394                            (unsigned long long)bio->bi_iter.bi_sector,
8395                            bio->bi_iter.bi_size, err);
8396
8397         if (dip->subio_endio)
8398                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8399
8400         if (err) {
8401                 dip->errors = 1;
8402
8403                 /*
8404                  * before atomic variable goto zero, we must make sure
8405                  * dip->errors is perceived to be set.
8406                  */
8407                 smp_mb__before_atomic();
8408         }
8409
8410         /* if there are more bios still pending for this dio, just exit */
8411         if (!atomic_dec_and_test(&dip->pending_bios))
8412                 goto out;
8413
8414         if (dip->errors) {
8415                 bio_io_error(dip->orig_bio);
8416         } else {
8417                 dip->dio_bio->bi_status = BLK_STS_OK;
8418                 bio_endio(dip->orig_bio);
8419         }
8420 out:
8421         bio_put(bio);
8422 }
8423
8424 static inline blk_status_t btrfs_lookup_and_bind_dio_csum(struct inode *inode,
8425                                                  struct btrfs_dio_private *dip,
8426                                                  struct bio *bio,
8427                                                  u64 file_offset)
8428 {
8429         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8430         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8431         blk_status_t ret;
8432
8433         /*
8434          * We load all the csum data we need when we submit
8435          * the first bio to reduce the csum tree search and
8436          * contention.
8437          */
8438         if (dip->logical_offset == file_offset) {
8439                 ret = btrfs_lookup_bio_sums_dio(inode, dip->orig_bio,
8440                                                 file_offset);
8441                 if (ret)
8442                         return ret;
8443         }
8444
8445         if (bio == dip->orig_bio)
8446                 return 0;
8447
8448         file_offset -= dip->logical_offset;
8449         file_offset >>= inode->i_sb->s_blocksize_bits;
8450         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8451
8452         return 0;
8453 }
8454
8455 static inline blk_status_t
8456 __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode, u64 file_offset,
8457                        int async_submit)
8458 {
8459         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8460         struct btrfs_dio_private *dip = bio->bi_private;
8461         bool write = bio_op(bio) == REQ_OP_WRITE;
8462         blk_status_t ret;
8463
8464         if (async_submit)
8465                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8466
8467         bio_get(bio);
8468
8469         if (!write) {
8470                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
8471                 if (ret)
8472                         goto err;
8473         }
8474
8475         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
8476                 goto map;
8477
8478         if (write && async_submit) {
8479                 ret = btrfs_wq_submit_bio(fs_info, bio, 0, 0,
8480                                           file_offset, inode,
8481                                           __btrfs_submit_bio_start_direct_io,
8482                                           __btrfs_submit_bio_done);
8483                 goto err;
8484         } else if (write) {
8485                 /*
8486                  * If we aren't doing async submit, calculate the csum of the
8487                  * bio now.
8488                  */
8489                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
8490                 if (ret)
8491                         goto err;
8492         } else {
8493                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
8494                                                      file_offset);
8495                 if (ret)
8496                         goto err;
8497         }
8498 map:
8499         ret = btrfs_map_bio(fs_info, bio, 0, 0);
8500 err:
8501         bio_put(bio);
8502         return ret;
8503 }
8504
8505 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip)
8506 {
8507         struct inode *inode = dip->inode;
8508         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8509         struct bio *bio;
8510         struct bio *orig_bio = dip->orig_bio;
8511         u64 start_sector = orig_bio->bi_iter.bi_sector;
8512         u64 file_offset = dip->logical_offset;
8513         u64 map_length;
8514         int async_submit = 0;
8515         u64 submit_len;
8516         int clone_offset = 0;
8517         int clone_len;
8518         int ret;
8519         blk_status_t status;
8520
8521         map_length = orig_bio->bi_iter.bi_size;
8522         submit_len = map_length;
8523         ret = btrfs_map_block(fs_info, btrfs_op(orig_bio), start_sector << 9,
8524                               &map_length, NULL, 0);
8525         if (ret)
8526                 return -EIO;
8527
8528         if (map_length >= submit_len) {
8529                 bio = orig_bio;
8530                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8531                 goto submit;
8532         }
8533
8534         /* async crcs make it difficult to collect full stripe writes. */
8535         if (btrfs_data_alloc_profile(fs_info) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8536                 async_submit = 0;
8537         else
8538                 async_submit = 1;
8539
8540         /* bio split */
8541         ASSERT(map_length <= INT_MAX);
8542         atomic_inc(&dip->pending_bios);
8543         do {
8544                 clone_len = min_t(int, submit_len, map_length);
8545
8546                 /*
8547                  * This will never fail as it's passing GPF_NOFS and
8548                  * the allocation is backed by btrfs_bioset.
8549                  */
8550                 bio = btrfs_bio_clone_partial(orig_bio, clone_offset,
8551                                               clone_len);
8552                 bio->bi_private = dip;
8553                 bio->bi_end_io = btrfs_end_dio_bio;
8554                 btrfs_io_bio(bio)->logical = file_offset;
8555
8556                 ASSERT(submit_len >= clone_len);
8557                 submit_len -= clone_len;
8558                 if (submit_len == 0)
8559                         break;
8560
8561                 /*
8562                  * Increase the count before we submit the bio so we know
8563                  * the end IO handler won't happen before we increase the
8564                  * count. Otherwise, the dip might get freed before we're
8565                  * done setting it up.
8566                  */
8567                 atomic_inc(&dip->pending_bios);
8568
8569                 status = __btrfs_submit_dio_bio(bio, inode, file_offset,
8570                                                 async_submit);
8571                 if (status) {
8572                         bio_put(bio);
8573                         atomic_dec(&dip->pending_bios);
8574                         goto out_err;
8575                 }
8576
8577                 clone_offset += clone_len;
8578                 start_sector += clone_len >> 9;
8579                 file_offset += clone_len;
8580
8581                 map_length = submit_len;
8582                 ret = btrfs_map_block(fs_info, btrfs_op(orig_bio),
8583                                       start_sector << 9, &map_length, NULL, 0);
8584                 if (ret)
8585                         goto out_err;
8586         } while (submit_len > 0);
8587
8588 submit:
8589         status = __btrfs_submit_dio_bio(bio, inode, file_offset, async_submit);
8590         if (!status)
8591                 return 0;
8592
8593         bio_put(bio);
8594 out_err:
8595         dip->errors = 1;
8596         /*
8597          * before atomic variable goto zero, we must
8598          * make sure dip->errors is perceived to be set.
8599          */
8600         smp_mb__before_atomic();
8601         if (atomic_dec_and_test(&dip->pending_bios))
8602                 bio_io_error(dip->orig_bio);
8603
8604         /* bio_end_io() will handle error, so we needn't return it */
8605         return 0;
8606 }
8607
8608 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8609                                 loff_t file_offset)
8610 {
8611         struct btrfs_dio_private *dip = NULL;
8612         struct bio *bio = NULL;
8613         struct btrfs_io_bio *io_bio;
8614         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8615         int ret = 0;
8616
8617         bio = btrfs_bio_clone(dio_bio);
8618
8619         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8620         if (!dip) {
8621                 ret = -ENOMEM;
8622                 goto free_ordered;
8623         }
8624
8625         dip->private = dio_bio->bi_private;
8626         dip->inode = inode;
8627         dip->logical_offset = file_offset;
8628         dip->bytes = dio_bio->bi_iter.bi_size;
8629         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8630         bio->bi_private = dip;
8631         dip->orig_bio = bio;
8632         dip->dio_bio = dio_bio;
8633         atomic_set(&dip->pending_bios, 0);
8634         io_bio = btrfs_io_bio(bio);
8635         io_bio->logical = file_offset;
8636
8637         if (write) {
8638                 bio->bi_end_io = btrfs_endio_direct_write;
8639         } else {
8640                 bio->bi_end_io = btrfs_endio_direct_read;
8641                 dip->subio_endio = btrfs_subio_endio_read;
8642         }
8643
8644         /*
8645          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8646          * even if we fail to submit a bio, because in such case we do the
8647          * corresponding error handling below and it must not be done a second
8648          * time by btrfs_direct_IO().
8649          */
8650         if (write) {
8651                 struct btrfs_dio_data *dio_data = current->journal_info;
8652
8653                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8654                         dip->bytes;
8655                 dio_data->unsubmitted_oe_range_start =
8656                         dio_data->unsubmitted_oe_range_end;
8657         }
8658
8659         ret = btrfs_submit_direct_hook(dip);
8660         if (!ret)
8661                 return;
8662
8663         if (io_bio->end_io)
8664                 io_bio->end_io(io_bio, ret);
8665
8666 free_ordered:
8667         /*
8668          * If we arrived here it means either we failed to submit the dip
8669          * or we either failed to clone the dio_bio or failed to allocate the
8670          * dip. If we cloned the dio_bio and allocated the dip, we can just
8671          * call bio_endio against our io_bio so that we get proper resource
8672          * cleanup if we fail to submit the dip, otherwise, we must do the
8673          * same as btrfs_endio_direct_[write|read] because we can't call these
8674          * callbacks - they require an allocated dip and a clone of dio_bio.
8675          */
8676         if (bio && dip) {
8677                 bio_io_error(bio);
8678                 /*
8679                  * The end io callbacks free our dip, do the final put on bio
8680                  * and all the cleanup and final put for dio_bio (through
8681                  * dio_end_io()).
8682                  */
8683                 dip = NULL;
8684                 bio = NULL;
8685         } else {
8686                 if (write)
8687                         __endio_write_update_ordered(inode,
8688                                                 file_offset,
8689                                                 dio_bio->bi_iter.bi_size,
8690                                                 false);
8691                 else
8692                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8693                               file_offset + dio_bio->bi_iter.bi_size - 1);
8694
8695                 dio_bio->bi_status = BLK_STS_IOERR;
8696                 /*
8697                  * Releases and cleans up our dio_bio, no need to bio_put()
8698                  * nor bio_endio()/bio_io_error() against dio_bio.
8699                  */
8700                 dio_end_io(dio_bio);
8701         }
8702         if (bio)
8703                 bio_put(bio);
8704         kfree(dip);
8705 }
8706
8707 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8708                                const struct iov_iter *iter, loff_t offset)
8709 {
8710         int seg;
8711         int i;
8712         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8713         ssize_t retval = -EINVAL;
8714
8715         if (offset & blocksize_mask)
8716                 goto out;
8717
8718         if (iov_iter_alignment(iter) & blocksize_mask)
8719                 goto out;
8720
8721         /* If this is a write we don't need to check anymore */
8722         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8723                 return 0;
8724         /*
8725          * Check to make sure we don't have duplicate iov_base's in this
8726          * iovec, if so return EINVAL, otherwise we'll get csum errors
8727          * when reading back.
8728          */
8729         for (seg = 0; seg < iter->nr_segs; seg++) {
8730                 for (i = seg + 1; i < iter->nr_segs; i++) {
8731                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8732                                 goto out;
8733                 }
8734         }
8735         retval = 0;
8736 out:
8737         return retval;
8738 }
8739
8740 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8741 {
8742         struct file *file = iocb->ki_filp;
8743         struct inode *inode = file->f_mapping->host;
8744         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8745         struct btrfs_dio_data dio_data = { 0 };
8746         struct extent_changeset *data_reserved = NULL;
8747         loff_t offset = iocb->ki_pos;
8748         size_t count = 0;
8749         int flags = 0;
8750         bool wakeup = true;
8751         bool relock = false;
8752         ssize_t ret;
8753
8754         if (check_direct_IO(fs_info, iter, offset))
8755                 return 0;
8756
8757         inode_dio_begin(inode);
8758
8759         /*
8760          * The generic stuff only does filemap_write_and_wait_range, which
8761          * isn't enough if we've written compressed pages to this area, so
8762          * we need to flush the dirty pages again to make absolutely sure
8763          * that any outstanding dirty pages are on disk.
8764          */
8765         count = iov_iter_count(iter);
8766         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8767                      &BTRFS_I(inode)->runtime_flags))
8768                 filemap_fdatawrite_range(inode->i_mapping, offset,
8769                                          offset + count - 1);
8770
8771         if (iov_iter_rw(iter) == WRITE) {
8772                 /*
8773                  * If the write DIO is beyond the EOF, we need update
8774                  * the isize, but it is protected by i_mutex. So we can
8775                  * not unlock the i_mutex at this case.
8776                  */
8777                 if (offset + count <= inode->i_size) {
8778                         dio_data.overwrite = 1;
8779                         inode_unlock(inode);
8780                         relock = true;
8781                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
8782                         ret = -EAGAIN;
8783                         goto out;
8784                 }
8785                 ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
8786                                                    offset, count);
8787                 if (ret)
8788                         goto out;
8789
8790                 /*
8791                  * We need to know how many extents we reserved so that we can
8792                  * do the accounting properly if we go over the number we
8793                  * originally calculated.  Abuse current->journal_info for this.
8794                  */
8795                 dio_data.reserve = round_up(count,
8796                                             fs_info->sectorsize);
8797                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8798                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8799                 current->journal_info = &dio_data;
8800                 down_read(&BTRFS_I(inode)->dio_sem);
8801         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8802                                      &BTRFS_I(inode)->runtime_flags)) {
8803                 inode_dio_end(inode);
8804                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8805                 wakeup = false;
8806         }
8807
8808         ret = __blockdev_direct_IO(iocb, inode,
8809                                    fs_info->fs_devices->latest_bdev,
8810                                    iter, btrfs_get_blocks_direct, NULL,
8811                                    btrfs_submit_direct, flags);
8812         if (iov_iter_rw(iter) == WRITE) {
8813                 up_read(&BTRFS_I(inode)->dio_sem);
8814                 current->journal_info = NULL;
8815                 if (ret < 0 && ret != -EIOCBQUEUED) {
8816                         if (dio_data.reserve)
8817                                 btrfs_delalloc_release_space(inode, data_reserved,
8818                                         offset, dio_data.reserve);
8819                         /*
8820                          * On error we might have left some ordered extents
8821                          * without submitting corresponding bios for them, so
8822                          * cleanup them up to avoid other tasks getting them
8823                          * and waiting for them to complete forever.
8824                          */
8825                         if (dio_data.unsubmitted_oe_range_start <
8826                             dio_data.unsubmitted_oe_range_end)
8827                                 __endio_write_update_ordered(inode,
8828                                         dio_data.unsubmitted_oe_range_start,
8829                                         dio_data.unsubmitted_oe_range_end -
8830                                         dio_data.unsubmitted_oe_range_start,
8831                                         false);
8832                 } else if (ret >= 0 && (size_t)ret < count)
8833                         btrfs_delalloc_release_space(inode, data_reserved,
8834                                         offset, count - (size_t)ret);
8835                 btrfs_delalloc_release_extents(BTRFS_I(inode), count);
8836         }
8837 out:
8838         if (wakeup)
8839                 inode_dio_end(inode);
8840         if (relock)
8841                 inode_lock(inode);
8842
8843         extent_changeset_free(data_reserved);
8844         return ret;
8845 }
8846
8847 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8848
8849 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8850                 __u64 start, __u64 len)
8851 {
8852         int     ret;
8853
8854         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8855         if (ret)
8856                 return ret;
8857
8858         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8859 }
8860
8861 int btrfs_readpage(struct file *file, struct page *page)
8862 {
8863         struct extent_io_tree *tree;
8864         tree = &BTRFS_I(page->mapping->host)->io_tree;
8865         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8866 }
8867
8868 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8869 {
8870         struct extent_io_tree *tree;
8871         struct inode *inode = page->mapping->host;
8872         int ret;
8873
8874         if (current->flags & PF_MEMALLOC) {
8875                 redirty_page_for_writepage(wbc, page);
8876                 unlock_page(page);
8877                 return 0;
8878         }
8879
8880         /*
8881          * If we are under memory pressure we will call this directly from the
8882          * VM, we need to make sure we have the inode referenced for the ordered
8883          * extent.  If not just return like we didn't do anything.
8884          */
8885         if (!igrab(inode)) {
8886                 redirty_page_for_writepage(wbc, page);
8887                 return AOP_WRITEPAGE_ACTIVATE;
8888         }
8889         tree = &BTRFS_I(page->mapping->host)->io_tree;
8890         ret = extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8891         btrfs_add_delayed_iput(inode);
8892         return ret;
8893 }
8894
8895 static int btrfs_writepages(struct address_space *mapping,
8896                             struct writeback_control *wbc)
8897 {
8898         struct extent_io_tree *tree;
8899
8900         tree = &BTRFS_I(mapping->host)->io_tree;
8901         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8902 }
8903
8904 static int
8905 btrfs_readpages(struct file *file, struct address_space *mapping,
8906                 struct list_head *pages, unsigned nr_pages)
8907 {
8908         struct extent_io_tree *tree;
8909         tree = &BTRFS_I(mapping->host)->io_tree;
8910         return extent_readpages(tree, mapping, pages, nr_pages,
8911                                 btrfs_get_extent);
8912 }
8913 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8914 {
8915         struct extent_io_tree *tree;
8916         struct extent_map_tree *map;
8917         int ret;
8918
8919         tree = &BTRFS_I(page->mapping->host)->io_tree;
8920         map = &BTRFS_I(page->mapping->host)->extent_tree;
8921         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8922         if (ret == 1) {
8923                 ClearPagePrivate(page);
8924                 set_page_private(page, 0);
8925                 put_page(page);
8926         }
8927         return ret;
8928 }
8929
8930 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8931 {
8932         if (PageWriteback(page) || PageDirty(page))
8933                 return 0;
8934         return __btrfs_releasepage(page, gfp_flags);
8935 }
8936
8937 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8938                                  unsigned int length)
8939 {
8940         struct inode *inode = page->mapping->host;
8941         struct extent_io_tree *tree;
8942         struct btrfs_ordered_extent *ordered;
8943         struct extent_state *cached_state = NULL;
8944         u64 page_start = page_offset(page);
8945         u64 page_end = page_start + PAGE_SIZE - 1;
8946         u64 start;
8947         u64 end;
8948         int inode_evicting = inode->i_state & I_FREEING;
8949
8950         /*
8951          * we have the page locked, so new writeback can't start,
8952          * and the dirty bit won't be cleared while we are here.
8953          *
8954          * Wait for IO on this page so that we can safely clear
8955          * the PagePrivate2 bit and do ordered accounting
8956          */
8957         wait_on_page_writeback(page);
8958
8959         tree = &BTRFS_I(inode)->io_tree;
8960         if (offset) {
8961                 btrfs_releasepage(page, GFP_NOFS);
8962                 return;
8963         }
8964
8965         if (!inode_evicting)
8966                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8967 again:
8968         start = page_start;
8969         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
8970                                         page_end - start + 1);
8971         if (ordered) {
8972                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8973                 /*
8974                  * IO on this page will never be started, so we need
8975                  * to account for any ordered extents now
8976                  */
8977                 if (!inode_evicting)
8978                         clear_extent_bit(tree, start, end,
8979                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8980                                          EXTENT_DELALLOC_NEW |
8981                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8982                                          EXTENT_DEFRAG, 1, 0, &cached_state,
8983                                          GFP_NOFS);
8984                 /*
8985                  * whoever cleared the private bit is responsible
8986                  * for the finish_ordered_io
8987                  */
8988                 if (TestClearPagePrivate2(page)) {
8989                         struct btrfs_ordered_inode_tree *tree;
8990                         u64 new_len;
8991
8992                         tree = &BTRFS_I(inode)->ordered_tree;
8993
8994                         spin_lock_irq(&tree->lock);
8995                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8996                         new_len = start - ordered->file_offset;
8997                         if (new_len < ordered->truncated_len)
8998                                 ordered->truncated_len = new_len;
8999                         spin_unlock_irq(&tree->lock);
9000
9001                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
9002                                                            start,
9003                                                            end - start + 1, 1))
9004                                 btrfs_finish_ordered_io(ordered);
9005                 }
9006                 btrfs_put_ordered_extent(ordered);
9007                 if (!inode_evicting) {
9008                         cached_state = NULL;
9009                         lock_extent_bits(tree, start, end,
9010                                          &cached_state);
9011                 }
9012
9013                 start = end + 1;
9014                 if (start < page_end)
9015                         goto again;
9016         }
9017
9018         /*
9019          * Qgroup reserved space handler
9020          * Page here will be either
9021          * 1) Already written to disk
9022          *    In this case, its reserved space is released from data rsv map
9023          *    and will be freed by delayed_ref handler finally.
9024          *    So even we call qgroup_free_data(), it won't decrease reserved
9025          *    space.
9026          * 2) Not written to disk
9027          *    This means the reserved space should be freed here. However,
9028          *    if a truncate invalidates the page (by clearing PageDirty)
9029          *    and the page is accounted for while allocating extent
9030          *    in btrfs_check_data_free_space() we let delayed_ref to
9031          *    free the entire extent.
9032          */
9033         if (PageDirty(page))
9034                 btrfs_qgroup_free_data(inode, NULL, page_start, PAGE_SIZE);
9035         if (!inode_evicting) {
9036                 clear_extent_bit(tree, page_start, page_end,
9037                                  EXTENT_LOCKED | EXTENT_DIRTY |
9038                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
9039                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
9040                                  &cached_state, GFP_NOFS);
9041
9042                 __btrfs_releasepage(page, GFP_NOFS);
9043         }
9044
9045         ClearPageChecked(page);
9046         if (PagePrivate(page)) {
9047                 ClearPagePrivate(page);
9048                 set_page_private(page, 0);
9049                 put_page(page);
9050         }
9051 }
9052
9053 /*
9054  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
9055  * called from a page fault handler when a page is first dirtied. Hence we must
9056  * be careful to check for EOF conditions here. We set the page up correctly
9057  * for a written page which means we get ENOSPC checking when writing into
9058  * holes and correct delalloc and unwritten extent mapping on filesystems that
9059  * support these features.
9060  *
9061  * We are not allowed to take the i_mutex here so we have to play games to
9062  * protect against truncate races as the page could now be beyond EOF.  Because
9063  * vmtruncate() writes the inode size before removing pages, once we have the
9064  * page lock we can determine safely if the page is beyond EOF. If it is not
9065  * beyond EOF, then the page is guaranteed safe against truncation until we
9066  * unlock the page.
9067  */
9068 int btrfs_page_mkwrite(struct vm_fault *vmf)
9069 {
9070         struct page *page = vmf->page;
9071         struct inode *inode = file_inode(vmf->vma->vm_file);
9072         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9073         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
9074         struct btrfs_ordered_extent *ordered;
9075         struct extent_state *cached_state = NULL;
9076         struct extent_changeset *data_reserved = NULL;
9077         char *kaddr;
9078         unsigned long zero_start;
9079         loff_t size;
9080         int ret;
9081         int reserved = 0;
9082         u64 reserved_space;
9083         u64 page_start;
9084         u64 page_end;
9085         u64 end;
9086
9087         reserved_space = PAGE_SIZE;
9088
9089         sb_start_pagefault(inode->i_sb);
9090         page_start = page_offset(page);
9091         page_end = page_start + PAGE_SIZE - 1;
9092         end = page_end;
9093
9094         /*
9095          * Reserving delalloc space after obtaining the page lock can lead to
9096          * deadlock. For example, if a dirty page is locked by this function
9097          * and the call to btrfs_delalloc_reserve_space() ends up triggering
9098          * dirty page write out, then the btrfs_writepage() function could
9099          * end up waiting indefinitely to get a lock on the page currently
9100          * being processed by btrfs_page_mkwrite() function.
9101          */
9102         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
9103                                            reserved_space);
9104         if (!ret) {
9105                 ret = file_update_time(vmf->vma->vm_file);
9106                 reserved = 1;
9107         }
9108         if (ret) {
9109                 if (ret == -ENOMEM)
9110                         ret = VM_FAULT_OOM;
9111                 else /* -ENOSPC, -EIO, etc */
9112                         ret = VM_FAULT_SIGBUS;
9113                 if (reserved)
9114                         goto out;
9115                 goto out_noreserve;
9116         }
9117
9118         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
9119 again:
9120         lock_page(page);
9121         size = i_size_read(inode);
9122
9123         if ((page->mapping != inode->i_mapping) ||
9124             (page_start >= size)) {
9125                 /* page got truncated out from underneath us */
9126                 goto out_unlock;
9127         }
9128         wait_on_page_writeback(page);
9129
9130         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
9131         set_page_extent_mapped(page);
9132
9133         /*
9134          * we can't set the delalloc bits if there are pending ordered
9135          * extents.  Drop our locks and wait for them to finish
9136          */
9137         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
9138                         PAGE_SIZE);
9139         if (ordered) {
9140                 unlock_extent_cached(io_tree, page_start, page_end,
9141                                      &cached_state, GFP_NOFS);
9142                 unlock_page(page);
9143                 btrfs_start_ordered_extent(inode, ordered, 1);
9144                 btrfs_put_ordered_extent(ordered);
9145                 goto again;
9146         }
9147
9148         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
9149                 reserved_space = round_up(size - page_start,
9150                                           fs_info->sectorsize);
9151                 if (reserved_space < PAGE_SIZE) {
9152                         end = page_start + reserved_space - 1;
9153                         btrfs_delalloc_release_space(inode, data_reserved,
9154                                         page_start, PAGE_SIZE - reserved_space);
9155                 }
9156         }
9157
9158         /*
9159          * page_mkwrite gets called when the page is firstly dirtied after it's
9160          * faulted in, but write(2) could also dirty a page and set delalloc
9161          * bits, thus in this case for space account reason, we still need to
9162          * clear any delalloc bits within this page range since we have to
9163          * reserve data&meta space before lock_page() (see above comments).
9164          */
9165         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9166                           EXTENT_DIRTY | EXTENT_DELALLOC |
9167                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
9168                           0, 0, &cached_state, GFP_NOFS);
9169
9170         ret = btrfs_set_extent_delalloc(inode, page_start, end, 0,
9171                                         &cached_state, 0);
9172         if (ret) {
9173                 unlock_extent_cached(io_tree, page_start, page_end,
9174                                      &cached_state, GFP_NOFS);
9175                 ret = VM_FAULT_SIGBUS;
9176                 goto out_unlock;
9177         }
9178         ret = 0;
9179
9180         /* page is wholly or partially inside EOF */
9181         if (page_start + PAGE_SIZE > size)
9182                 zero_start = size & ~PAGE_MASK;
9183         else
9184                 zero_start = PAGE_SIZE;
9185
9186         if (zero_start != PAGE_SIZE) {
9187                 kaddr = kmap(page);
9188                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9189                 flush_dcache_page(page);
9190                 kunmap(page);
9191         }
9192         ClearPageChecked(page);
9193         set_page_dirty(page);
9194         SetPageUptodate(page);
9195
9196         BTRFS_I(inode)->last_trans = fs_info->generation;
9197         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9198         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9199
9200         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
9201
9202 out_unlock:
9203         if (!ret) {
9204                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
9205                 sb_end_pagefault(inode->i_sb);
9206                 extent_changeset_free(data_reserved);
9207                 return VM_FAULT_LOCKED;
9208         }
9209         unlock_page(page);
9210 out:
9211         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
9212         btrfs_delalloc_release_space(inode, data_reserved, page_start,
9213                                      reserved_space);
9214 out_noreserve:
9215         sb_end_pagefault(inode->i_sb);
9216         extent_changeset_free(data_reserved);
9217         return ret;
9218 }
9219
9220 static int btrfs_truncate(struct inode *inode)
9221 {
9222         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9223         struct btrfs_root *root = BTRFS_I(inode)->root;
9224         struct btrfs_block_rsv *rsv;
9225         int ret = 0;
9226         int err = 0;
9227         struct btrfs_trans_handle *trans;
9228         u64 mask = fs_info->sectorsize - 1;
9229         u64 min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
9230
9231         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9232                                        (u64)-1);
9233         if (ret)
9234                 return ret;
9235
9236         /*
9237          * Yes ladies and gentlemen, this is indeed ugly.  The fact is we have
9238          * 3 things going on here
9239          *
9240          * 1) We need to reserve space for our orphan item and the space to
9241          * delete our orphan item.  Lord knows we don't want to have a dangling
9242          * orphan item because we didn't reserve space to remove it.
9243          *
9244          * 2) We need to reserve space to update our inode.
9245          *
9246          * 3) We need to have something to cache all the space that is going to
9247          * be free'd up by the truncate operation, but also have some slack
9248          * space reserved in case it uses space during the truncate (thank you
9249          * very much snapshotting).
9250          *
9251          * And we need these to all be separate.  The fact is we can use a lot of
9252          * space doing the truncate, and we have no earthly idea how much space
9253          * we will use, so we need the truncate reservation to be separate so it
9254          * doesn't end up using space reserved for updating the inode or
9255          * removing the orphan item.  We also need to be able to stop the
9256          * transaction and start a new one, which means we need to be able to
9257          * update the inode several times, and we have no idea of knowing how
9258          * many times that will be, so we can't just reserve 1 item for the
9259          * entirety of the operation, so that has to be done separately as well.
9260          * Then there is the orphan item, which does indeed need to be held on
9261          * to for the whole operation, and we need nobody to touch this reserved
9262          * space except the orphan code.
9263          *
9264          * So that leaves us with
9265          *
9266          * 1) root->orphan_block_rsv - for the orphan deletion.
9267          * 2) rsv - for the truncate reservation, which we will steal from the
9268          * transaction reservation.
9269          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
9270          * updating the inode.
9271          */
9272         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
9273         if (!rsv)
9274                 return -ENOMEM;
9275         rsv->size = min_size;
9276         rsv->failfast = 1;
9277
9278         /*
9279          * 1 for the truncate slack space
9280          * 1 for updating the inode.
9281          */
9282         trans = btrfs_start_transaction(root, 2);
9283         if (IS_ERR(trans)) {
9284                 err = PTR_ERR(trans);
9285                 goto out;
9286         }
9287
9288         /* Migrate the slack space for the truncate to our reserve */
9289         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
9290                                       min_size, 0);
9291         BUG_ON(ret);
9292
9293         /*
9294          * So if we truncate and then write and fsync we normally would just
9295          * write the extents that changed, which is a problem if we need to
9296          * first truncate that entire inode.  So set this flag so we write out
9297          * all of the extents in the inode to the sync log so we're completely
9298          * safe.
9299          */
9300         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9301         trans->block_rsv = rsv;
9302
9303         while (1) {
9304                 ret = btrfs_truncate_inode_items(trans, root, inode,
9305                                                  inode->i_size,
9306                                                  BTRFS_EXTENT_DATA_KEY);
9307                 trans->block_rsv = &fs_info->trans_block_rsv;
9308                 if (ret != -ENOSPC && ret != -EAGAIN) {
9309                         err = ret;
9310                         break;
9311                 }
9312
9313                 ret = btrfs_update_inode(trans, root, inode);
9314                 if (ret) {
9315                         err = ret;
9316                         break;
9317                 }
9318
9319                 btrfs_end_transaction(trans);
9320                 btrfs_btree_balance_dirty(fs_info);
9321
9322                 trans = btrfs_start_transaction(root, 2);
9323                 if (IS_ERR(trans)) {
9324                         ret = err = PTR_ERR(trans);
9325                         trans = NULL;
9326                         break;
9327                 }
9328
9329                 btrfs_block_rsv_release(fs_info, rsv, -1);
9330                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
9331                                               rsv, min_size, 0);
9332                 BUG_ON(ret);    /* shouldn't happen */
9333                 trans->block_rsv = rsv;
9334         }
9335
9336         /*
9337          * We can't call btrfs_truncate_block inside a trans handle as we could
9338          * deadlock with freeze, if we got NEED_TRUNCATE_BLOCK then we know
9339          * we've truncated everything except the last little bit, and can do
9340          * btrfs_truncate_block and then update the disk_i_size.
9341          */
9342         if (ret == NEED_TRUNCATE_BLOCK) {
9343                 btrfs_end_transaction(trans);
9344                 btrfs_btree_balance_dirty(fs_info);
9345
9346                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
9347                 if (ret)
9348                         goto out;
9349                 trans = btrfs_start_transaction(root, 1);
9350                 if (IS_ERR(trans)) {
9351                         ret = PTR_ERR(trans);
9352                         goto out;
9353                 }
9354                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
9355         }
9356
9357         if (ret == 0 && inode->i_nlink > 0) {
9358                 trans->block_rsv = root->orphan_block_rsv;
9359                 ret = btrfs_orphan_del(trans, BTRFS_I(inode));
9360                 if (ret)
9361                         err = ret;
9362         }
9363
9364         if (trans) {
9365                 trans->block_rsv = &fs_info->trans_block_rsv;
9366                 ret = btrfs_update_inode(trans, root, inode);
9367                 if (ret && !err)
9368                         err = ret;
9369
9370                 ret = btrfs_end_transaction(trans);
9371                 btrfs_btree_balance_dirty(fs_info);
9372         }
9373 out:
9374         btrfs_free_block_rsv(fs_info, rsv);
9375
9376         if (ret && !err)
9377                 err = ret;
9378
9379         return err;
9380 }
9381
9382 /*
9383  * create a new subvolume directory/inode (helper for the ioctl).
9384  */
9385 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9386                              struct btrfs_root *new_root,
9387                              struct btrfs_root *parent_root,
9388                              u64 new_dirid)
9389 {
9390         struct inode *inode;
9391         int err;
9392         u64 index = 0;
9393
9394         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9395                                 new_dirid, new_dirid,
9396                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9397                                 &index);
9398         if (IS_ERR(inode))
9399                 return PTR_ERR(inode);
9400         inode->i_op = &btrfs_dir_inode_operations;
9401         inode->i_fop = &btrfs_dir_file_operations;
9402
9403         set_nlink(inode, 1);
9404         btrfs_i_size_write(BTRFS_I(inode), 0);
9405         unlock_new_inode(inode);
9406
9407         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9408         if (err)
9409                 btrfs_err(new_root->fs_info,
9410                           "error inheriting subvolume %llu properties: %d",
9411                           new_root->root_key.objectid, err);
9412
9413         err = btrfs_update_inode(trans, new_root, inode);
9414
9415         iput(inode);
9416         return err;
9417 }
9418
9419 struct inode *btrfs_alloc_inode(struct super_block *sb)
9420 {
9421         struct btrfs_fs_info *fs_info = btrfs_sb(sb);
9422         struct btrfs_inode *ei;
9423         struct inode *inode;
9424
9425         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9426         if (!ei)
9427                 return NULL;
9428
9429         ei->root = NULL;
9430         ei->generation = 0;
9431         ei->last_trans = 0;
9432         ei->last_sub_trans = 0;
9433         ei->logged_trans = 0;
9434         ei->delalloc_bytes = 0;
9435         ei->new_delalloc_bytes = 0;
9436         ei->defrag_bytes = 0;
9437         ei->disk_i_size = 0;
9438         ei->flags = 0;
9439         ei->csum_bytes = 0;
9440         ei->index_cnt = (u64)-1;
9441         ei->dir_index = 0;
9442         ei->last_unlink_trans = 0;
9443         ei->last_log_commit = 0;
9444         ei->delayed_iput_count = 0;
9445
9446         spin_lock_init(&ei->lock);
9447         ei->outstanding_extents = 0;
9448         if (sb->s_magic != BTRFS_TEST_MAGIC)
9449                 btrfs_init_metadata_block_rsv(fs_info, &ei->block_rsv,
9450                                               BTRFS_BLOCK_RSV_DELALLOC);
9451         ei->runtime_flags = 0;
9452         ei->prop_compress = BTRFS_COMPRESS_NONE;
9453         ei->defrag_compress = BTRFS_COMPRESS_NONE;
9454
9455         ei->delayed_node = NULL;
9456
9457         ei->i_otime.tv_sec = 0;
9458         ei->i_otime.tv_nsec = 0;
9459
9460         inode = &ei->vfs_inode;
9461         extent_map_tree_init(&ei->extent_tree);
9462         extent_io_tree_init(&ei->io_tree, inode);
9463         extent_io_tree_init(&ei->io_failure_tree, inode);
9464         ei->io_tree.track_uptodate = 1;
9465         ei->io_failure_tree.track_uptodate = 1;
9466         atomic_set(&ei->sync_writers, 0);
9467         mutex_init(&ei->log_mutex);
9468         mutex_init(&ei->delalloc_mutex);
9469         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9470         INIT_LIST_HEAD(&ei->delalloc_inodes);
9471         INIT_LIST_HEAD(&ei->delayed_iput);
9472         RB_CLEAR_NODE(&ei->rb_node);
9473         init_rwsem(&ei->dio_sem);
9474
9475         return inode;
9476 }
9477
9478 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9479 void btrfs_test_destroy_inode(struct inode *inode)
9480 {
9481         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9482         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9483 }
9484 #endif
9485
9486 static void btrfs_i_callback(struct rcu_head *head)
9487 {
9488         struct inode *inode = container_of(head, struct inode, i_rcu);
9489         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9490 }
9491
9492 void btrfs_destroy_inode(struct inode *inode)
9493 {
9494         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9495         struct btrfs_ordered_extent *ordered;
9496         struct btrfs_root *root = BTRFS_I(inode)->root;
9497
9498         WARN_ON(!hlist_empty(&inode->i_dentry));
9499         WARN_ON(inode->i_data.nrpages);
9500         WARN_ON(BTRFS_I(inode)->block_rsv.reserved);
9501         WARN_ON(BTRFS_I(inode)->block_rsv.size);
9502         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9503         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9504         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
9505         WARN_ON(BTRFS_I(inode)->csum_bytes);
9506         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9507
9508         /*
9509          * This can happen where we create an inode, but somebody else also
9510          * created the same inode and we need to destroy the one we already
9511          * created.
9512          */
9513         if (!root)
9514                 goto free;
9515
9516         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9517                      &BTRFS_I(inode)->runtime_flags)) {
9518                 btrfs_info(fs_info, "inode %llu still on the orphan list",
9519                            btrfs_ino(BTRFS_I(inode)));
9520                 atomic_dec(&root->orphan_inodes);
9521         }
9522
9523         while (1) {
9524                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9525                 if (!ordered)
9526                         break;
9527                 else {
9528                         btrfs_err(fs_info,
9529                                   "found ordered extent %llu %llu on inode cleanup",
9530                                   ordered->file_offset, ordered->len);
9531                         btrfs_remove_ordered_extent(inode, ordered);
9532                         btrfs_put_ordered_extent(ordered);
9533                         btrfs_put_ordered_extent(ordered);
9534                 }
9535         }
9536         btrfs_qgroup_check_reserved_leak(inode);
9537         inode_tree_del(inode);
9538         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9539 free:
9540         call_rcu(&inode->i_rcu, btrfs_i_callback);
9541 }
9542
9543 int btrfs_drop_inode(struct inode *inode)
9544 {
9545         struct btrfs_root *root = BTRFS_I(inode)->root;
9546
9547         if (root == NULL)
9548                 return 1;
9549
9550         /* the snap/subvol tree is on deleting */
9551         if (btrfs_root_refs(&root->root_item) == 0)
9552                 return 1;
9553         else
9554                 return generic_drop_inode(inode);
9555 }
9556
9557 static void init_once(void *foo)
9558 {
9559         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9560
9561         inode_init_once(&ei->vfs_inode);
9562 }
9563
9564 void btrfs_destroy_cachep(void)
9565 {
9566         /*
9567          * Make sure all delayed rcu free inodes are flushed before we
9568          * destroy cache.
9569          */
9570         rcu_barrier();
9571         kmem_cache_destroy(btrfs_inode_cachep);
9572         kmem_cache_destroy(btrfs_trans_handle_cachep);
9573         kmem_cache_destroy(btrfs_path_cachep);
9574         kmem_cache_destroy(btrfs_free_space_cachep);
9575 }
9576
9577 int btrfs_init_cachep(void)
9578 {
9579         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9580                         sizeof(struct btrfs_inode), 0,
9581                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9582                         init_once);
9583         if (!btrfs_inode_cachep)
9584                 goto fail;
9585
9586         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9587                         sizeof(struct btrfs_trans_handle), 0,
9588                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9589         if (!btrfs_trans_handle_cachep)
9590                 goto fail;
9591
9592         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9593                         sizeof(struct btrfs_path), 0,
9594                         SLAB_MEM_SPREAD, NULL);
9595         if (!btrfs_path_cachep)
9596                 goto fail;
9597
9598         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9599                         sizeof(struct btrfs_free_space), 0,
9600                         SLAB_MEM_SPREAD, NULL);
9601         if (!btrfs_free_space_cachep)
9602                 goto fail;
9603
9604         return 0;
9605 fail:
9606         btrfs_destroy_cachep();
9607         return -ENOMEM;
9608 }
9609
9610 static int btrfs_getattr(const struct path *path, struct kstat *stat,
9611                          u32 request_mask, unsigned int flags)
9612 {
9613         u64 delalloc_bytes;
9614         struct inode *inode = d_inode(path->dentry);
9615         u32 blocksize = inode->i_sb->s_blocksize;
9616         u32 bi_flags = BTRFS_I(inode)->flags;
9617
9618         stat->result_mask |= STATX_BTIME;
9619         stat->btime.tv_sec = BTRFS_I(inode)->i_otime.tv_sec;
9620         stat->btime.tv_nsec = BTRFS_I(inode)->i_otime.tv_nsec;
9621         if (bi_flags & BTRFS_INODE_APPEND)
9622                 stat->attributes |= STATX_ATTR_APPEND;
9623         if (bi_flags & BTRFS_INODE_COMPRESS)
9624                 stat->attributes |= STATX_ATTR_COMPRESSED;
9625         if (bi_flags & BTRFS_INODE_IMMUTABLE)
9626                 stat->attributes |= STATX_ATTR_IMMUTABLE;
9627         if (bi_flags & BTRFS_INODE_NODUMP)
9628                 stat->attributes |= STATX_ATTR_NODUMP;
9629
9630         stat->attributes_mask |= (STATX_ATTR_APPEND |
9631                                   STATX_ATTR_COMPRESSED |
9632                                   STATX_ATTR_IMMUTABLE |
9633                                   STATX_ATTR_NODUMP);
9634
9635         generic_fillattr(inode, stat);
9636         stat->dev = BTRFS_I(inode)->root->anon_dev;
9637
9638         spin_lock(&BTRFS_I(inode)->lock);
9639         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
9640         spin_unlock(&BTRFS_I(inode)->lock);
9641         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9642                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9643         return 0;
9644 }
9645
9646 static int btrfs_rename_exchange(struct inode *old_dir,
9647                               struct dentry *old_dentry,
9648                               struct inode *new_dir,
9649                               struct dentry *new_dentry)
9650 {
9651         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9652         struct btrfs_trans_handle *trans;
9653         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9654         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9655         struct inode *new_inode = new_dentry->d_inode;
9656         struct inode *old_inode = old_dentry->d_inode;
9657         struct timespec ctime = current_time(old_inode);
9658         struct dentry *parent;
9659         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9660         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
9661         u64 old_idx = 0;
9662         u64 new_idx = 0;
9663         u64 root_objectid;
9664         int ret;
9665         bool root_log_pinned = false;
9666         bool dest_log_pinned = false;
9667
9668         /* we only allow rename subvolume link between subvolumes */
9669         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9670                 return -EXDEV;
9671
9672         /* close the race window with snapshot create/destroy ioctl */
9673         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9674                 down_read(&fs_info->subvol_sem);
9675         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9676                 down_read(&fs_info->subvol_sem);
9677
9678         /*
9679          * We want to reserve the absolute worst case amount of items.  So if
9680          * both inodes are subvols and we need to unlink them then that would
9681          * require 4 item modifications, but if they are both normal inodes it
9682          * would require 5 item modifications, so we'll assume their normal
9683          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9684          * should cover the worst case number of items we'll modify.
9685          */
9686         trans = btrfs_start_transaction(root, 12);
9687         if (IS_ERR(trans)) {
9688                 ret = PTR_ERR(trans);
9689                 goto out_notrans;
9690         }
9691
9692         /*
9693          * We need to find a free sequence number both in the source and
9694          * in the destination directory for the exchange.
9695          */
9696         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
9697         if (ret)
9698                 goto out_fail;
9699         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
9700         if (ret)
9701                 goto out_fail;
9702
9703         BTRFS_I(old_inode)->dir_index = 0ULL;
9704         BTRFS_I(new_inode)->dir_index = 0ULL;
9705
9706         /* Reference for the source. */
9707         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9708                 /* force full log commit if subvolume involved. */
9709                 btrfs_set_log_full_commit(fs_info, trans);
9710         } else {
9711                 btrfs_pin_log_trans(root);
9712                 root_log_pinned = true;
9713                 ret = btrfs_insert_inode_ref(trans, dest,
9714                                              new_dentry->d_name.name,
9715                                              new_dentry->d_name.len,
9716                                              old_ino,
9717                                              btrfs_ino(BTRFS_I(new_dir)),
9718                                              old_idx);
9719                 if (ret)
9720                         goto out_fail;
9721         }
9722
9723         /* And now for the dest. */
9724         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9725                 /* force full log commit if subvolume involved. */
9726                 btrfs_set_log_full_commit(fs_info, trans);
9727         } else {
9728                 btrfs_pin_log_trans(dest);
9729                 dest_log_pinned = true;
9730                 ret = btrfs_insert_inode_ref(trans, root,
9731                                              old_dentry->d_name.name,
9732                                              old_dentry->d_name.len,
9733                                              new_ino,
9734                                              btrfs_ino(BTRFS_I(old_dir)),
9735                                              new_idx);
9736                 if (ret)
9737                         goto out_fail;
9738         }
9739
9740         /* Update inode version and ctime/mtime. */
9741         inode_inc_iversion(old_dir);
9742         inode_inc_iversion(new_dir);
9743         inode_inc_iversion(old_inode);
9744         inode_inc_iversion(new_inode);
9745         old_dir->i_ctime = old_dir->i_mtime = ctime;
9746         new_dir->i_ctime = new_dir->i_mtime = ctime;
9747         old_inode->i_ctime = ctime;
9748         new_inode->i_ctime = ctime;
9749
9750         if (old_dentry->d_parent != new_dentry->d_parent) {
9751                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9752                                 BTRFS_I(old_inode), 1);
9753                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
9754                                 BTRFS_I(new_inode), 1);
9755         }
9756
9757         /* src is a subvolume */
9758         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9759                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9760                 ret = btrfs_unlink_subvol(trans, root, old_dir,
9761                                           root_objectid,
9762                                           old_dentry->d_name.name,
9763                                           old_dentry->d_name.len);
9764         } else { /* src is an inode */
9765                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9766                                            BTRFS_I(old_dentry->d_inode),
9767                                            old_dentry->d_name.name,
9768                                            old_dentry->d_name.len);
9769                 if (!ret)
9770                         ret = btrfs_update_inode(trans, root, old_inode);
9771         }
9772         if (ret) {
9773                 btrfs_abort_transaction(trans, ret);
9774                 goto out_fail;
9775         }
9776
9777         /* dest is a subvolume */
9778         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9779                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9780                 ret = btrfs_unlink_subvol(trans, dest, new_dir,
9781                                           root_objectid,
9782                                           new_dentry->d_name.name,
9783                                           new_dentry->d_name.len);
9784         } else { /* dest is an inode */
9785                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9786                                            BTRFS_I(new_dentry->d_inode),
9787                                            new_dentry->d_name.name,
9788                                            new_dentry->d_name.len);
9789                 if (!ret)
9790                         ret = btrfs_update_inode(trans, dest, new_inode);
9791         }
9792         if (ret) {
9793                 btrfs_abort_transaction(trans, ret);
9794                 goto out_fail;
9795         }
9796
9797         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9798                              new_dentry->d_name.name,
9799                              new_dentry->d_name.len, 0, old_idx);
9800         if (ret) {
9801                 btrfs_abort_transaction(trans, ret);
9802                 goto out_fail;
9803         }
9804
9805         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
9806                              old_dentry->d_name.name,
9807                              old_dentry->d_name.len, 0, new_idx);
9808         if (ret) {
9809                 btrfs_abort_transaction(trans, ret);
9810                 goto out_fail;
9811         }
9812
9813         if (old_inode->i_nlink == 1)
9814                 BTRFS_I(old_inode)->dir_index = old_idx;
9815         if (new_inode->i_nlink == 1)
9816                 BTRFS_I(new_inode)->dir_index = new_idx;
9817
9818         if (root_log_pinned) {
9819                 parent = new_dentry->d_parent;
9820                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
9821                                 parent);
9822                 btrfs_end_log_trans(root);
9823                 root_log_pinned = false;
9824         }
9825         if (dest_log_pinned) {
9826                 parent = old_dentry->d_parent;
9827                 btrfs_log_new_name(trans, BTRFS_I(new_inode), BTRFS_I(new_dir),
9828                                 parent);
9829                 btrfs_end_log_trans(dest);
9830                 dest_log_pinned = false;
9831         }
9832 out_fail:
9833         /*
9834          * If we have pinned a log and an error happened, we unpin tasks
9835          * trying to sync the log and force them to fallback to a transaction
9836          * commit if the log currently contains any of the inodes involved in
9837          * this rename operation (to ensure we do not persist a log with an
9838          * inconsistent state for any of these inodes or leading to any
9839          * inconsistencies when replayed). If the transaction was aborted, the
9840          * abortion reason is propagated to userspace when attempting to commit
9841          * the transaction. If the log does not contain any of these inodes, we
9842          * allow the tasks to sync it.
9843          */
9844         if (ret && (root_log_pinned || dest_log_pinned)) {
9845                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9846                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9847                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9848                     (new_inode &&
9849                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9850                         btrfs_set_log_full_commit(fs_info, trans);
9851
9852                 if (root_log_pinned) {
9853                         btrfs_end_log_trans(root);
9854                         root_log_pinned = false;
9855                 }
9856                 if (dest_log_pinned) {
9857                         btrfs_end_log_trans(dest);
9858                         dest_log_pinned = false;
9859                 }
9860         }
9861         ret = btrfs_end_transaction(trans);
9862 out_notrans:
9863         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9864                 up_read(&fs_info->subvol_sem);
9865         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9866                 up_read(&fs_info->subvol_sem);
9867
9868         return ret;
9869 }
9870
9871 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9872                                      struct btrfs_root *root,
9873                                      struct inode *dir,
9874                                      struct dentry *dentry)
9875 {
9876         int ret;
9877         struct inode *inode;
9878         u64 objectid;
9879         u64 index;
9880
9881         ret = btrfs_find_free_ino(root, &objectid);
9882         if (ret)
9883                 return ret;
9884
9885         inode = btrfs_new_inode(trans, root, dir,
9886                                 dentry->d_name.name,
9887                                 dentry->d_name.len,
9888                                 btrfs_ino(BTRFS_I(dir)),
9889                                 objectid,
9890                                 S_IFCHR | WHITEOUT_MODE,
9891                                 &index);
9892
9893         if (IS_ERR(inode)) {
9894                 ret = PTR_ERR(inode);
9895                 return ret;
9896         }
9897
9898         inode->i_op = &btrfs_special_inode_operations;
9899         init_special_inode(inode, inode->i_mode,
9900                 WHITEOUT_DEV);
9901
9902         ret = btrfs_init_inode_security(trans, inode, dir,
9903                                 &dentry->d_name);
9904         if (ret)
9905                 goto out;
9906
9907         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9908                                 BTRFS_I(inode), 0, index);
9909         if (ret)
9910                 goto out;
9911
9912         ret = btrfs_update_inode(trans, root, inode);
9913 out:
9914         unlock_new_inode(inode);
9915         if (ret)
9916                 inode_dec_link_count(inode);
9917         iput(inode);
9918
9919         return ret;
9920 }
9921
9922 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9923                            struct inode *new_dir, struct dentry *new_dentry,
9924                            unsigned int flags)
9925 {
9926         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9927         struct btrfs_trans_handle *trans;
9928         unsigned int trans_num_items;
9929         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9930         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9931         struct inode *new_inode = d_inode(new_dentry);
9932         struct inode *old_inode = d_inode(old_dentry);
9933         u64 index = 0;
9934         u64 root_objectid;
9935         int ret;
9936         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9937         bool log_pinned = false;
9938
9939         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9940                 return -EPERM;
9941
9942         /* we only allow rename subvolume link between subvolumes */
9943         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9944                 return -EXDEV;
9945
9946         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9947             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
9948                 return -ENOTEMPTY;
9949
9950         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9951             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9952                 return -ENOTEMPTY;
9953
9954
9955         /* check for collisions, even if the  name isn't there */
9956         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9957                              new_dentry->d_name.name,
9958                              new_dentry->d_name.len);
9959
9960         if (ret) {
9961                 if (ret == -EEXIST) {
9962                         /* we shouldn't get
9963                          * eexist without a new_inode */
9964                         if (WARN_ON(!new_inode)) {
9965                                 return ret;
9966                         }
9967                 } else {
9968                         /* maybe -EOVERFLOW */
9969                         return ret;
9970                 }
9971         }
9972         ret = 0;
9973
9974         /*
9975          * we're using rename to replace one file with another.  Start IO on it
9976          * now so  we don't add too much work to the end of the transaction
9977          */
9978         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9979                 filemap_flush(old_inode->i_mapping);
9980
9981         /* close the racy window with snapshot create/destroy ioctl */
9982         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9983                 down_read(&fs_info->subvol_sem);
9984         /*
9985          * We want to reserve the absolute worst case amount of items.  So if
9986          * both inodes are subvols and we need to unlink them then that would
9987          * require 4 item modifications, but if they are both normal inodes it
9988          * would require 5 item modifications, so we'll assume they are normal
9989          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9990          * should cover the worst case number of items we'll modify.
9991          * If our rename has the whiteout flag, we need more 5 units for the
9992          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9993          * when selinux is enabled).
9994          */
9995         trans_num_items = 11;
9996         if (flags & RENAME_WHITEOUT)
9997                 trans_num_items += 5;
9998         trans = btrfs_start_transaction(root, trans_num_items);
9999         if (IS_ERR(trans)) {
10000                 ret = PTR_ERR(trans);
10001                 goto out_notrans;
10002         }
10003
10004         if (dest != root)
10005                 btrfs_record_root_in_trans(trans, dest);
10006
10007         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
10008         if (ret)
10009                 goto out_fail;
10010
10011         BTRFS_I(old_inode)->dir_index = 0ULL;
10012         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
10013                 /* force full log commit if subvolume involved. */
10014                 btrfs_set_log_full_commit(fs_info, trans);
10015         } else {
10016                 btrfs_pin_log_trans(root);
10017                 log_pinned = true;
10018                 ret = btrfs_insert_inode_ref(trans, dest,
10019                                              new_dentry->d_name.name,
10020                                              new_dentry->d_name.len,
10021                                              old_ino,
10022                                              btrfs_ino(BTRFS_I(new_dir)), index);
10023                 if (ret)
10024                         goto out_fail;
10025         }
10026
10027         inode_inc_iversion(old_dir);
10028         inode_inc_iversion(new_dir);
10029         inode_inc_iversion(old_inode);
10030         old_dir->i_ctime = old_dir->i_mtime =
10031         new_dir->i_ctime = new_dir->i_mtime =
10032         old_inode->i_ctime = current_time(old_dir);
10033
10034         if (old_dentry->d_parent != new_dentry->d_parent)
10035                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
10036                                 BTRFS_I(old_inode), 1);
10037
10038         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
10039                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
10040                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
10041                                         old_dentry->d_name.name,
10042                                         old_dentry->d_name.len);
10043         } else {
10044                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
10045                                         BTRFS_I(d_inode(old_dentry)),
10046                                         old_dentry->d_name.name,
10047                                         old_dentry->d_name.len);
10048                 if (!ret)
10049                         ret = btrfs_update_inode(trans, root, old_inode);
10050         }
10051         if (ret) {
10052                 btrfs_abort_transaction(trans, ret);
10053                 goto out_fail;
10054         }
10055
10056         if (new_inode) {
10057                 inode_inc_iversion(new_inode);
10058                 new_inode->i_ctime = current_time(new_inode);
10059                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
10060                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
10061                         root_objectid = BTRFS_I(new_inode)->location.objectid;
10062                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
10063                                                 root_objectid,
10064                                                 new_dentry->d_name.name,
10065                                                 new_dentry->d_name.len);
10066                         BUG_ON(new_inode->i_nlink == 0);
10067                 } else {
10068                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
10069                                                  BTRFS_I(d_inode(new_dentry)),
10070                                                  new_dentry->d_name.name,
10071                                                  new_dentry->d_name.len);
10072                 }
10073                 if (!ret && new_inode->i_nlink == 0)
10074                         ret = btrfs_orphan_add(trans,
10075                                         BTRFS_I(d_inode(new_dentry)));
10076                 if (ret) {
10077                         btrfs_abort_transaction(trans, ret);
10078                         goto out_fail;
10079                 }
10080         }
10081
10082         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
10083                              new_dentry->d_name.name,
10084                              new_dentry->d_name.len, 0, index);
10085         if (ret) {
10086                 btrfs_abort_transaction(trans, ret);
10087                 goto out_fail;
10088         }
10089
10090         if (old_inode->i_nlink == 1)
10091                 BTRFS_I(old_inode)->dir_index = index;
10092
10093         if (log_pinned) {
10094                 struct dentry *parent = new_dentry->d_parent;
10095
10096                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
10097                                 parent);
10098                 btrfs_end_log_trans(root);
10099                 log_pinned = false;
10100         }
10101
10102         if (flags & RENAME_WHITEOUT) {
10103                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
10104                                                 old_dentry);
10105
10106                 if (ret) {
10107                         btrfs_abort_transaction(trans, ret);
10108                         goto out_fail;
10109                 }
10110         }
10111 out_fail:
10112         /*
10113          * If we have pinned the log and an error happened, we unpin tasks
10114          * trying to sync the log and force them to fallback to a transaction
10115          * commit if the log currently contains any of the inodes involved in
10116          * this rename operation (to ensure we do not persist a log with an
10117          * inconsistent state for any of these inodes or leading to any
10118          * inconsistencies when replayed). If the transaction was aborted, the
10119          * abortion reason is propagated to userspace when attempting to commit
10120          * the transaction. If the log does not contain any of these inodes, we
10121          * allow the tasks to sync it.
10122          */
10123         if (ret && log_pinned) {
10124                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
10125                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
10126                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
10127                     (new_inode &&
10128                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
10129                         btrfs_set_log_full_commit(fs_info, trans);
10130
10131                 btrfs_end_log_trans(root);
10132                 log_pinned = false;
10133         }
10134         btrfs_end_transaction(trans);
10135 out_notrans:
10136         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
10137                 up_read(&fs_info->subvol_sem);
10138
10139         return ret;
10140 }
10141
10142 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
10143                          struct inode *new_dir, struct dentry *new_dentry,
10144                          unsigned int flags)
10145 {
10146         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
10147                 return -EINVAL;
10148
10149         if (flags & RENAME_EXCHANGE)
10150                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
10151                                           new_dentry);
10152
10153         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
10154 }
10155
10156 static void btrfs_run_delalloc_work(struct btrfs_work *work)
10157 {
10158         struct btrfs_delalloc_work *delalloc_work;
10159         struct inode *inode;
10160
10161         delalloc_work = container_of(work, struct btrfs_delalloc_work,
10162                                      work);
10163         inode = delalloc_work->inode;
10164         filemap_flush(inode->i_mapping);
10165         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
10166                                 &BTRFS_I(inode)->runtime_flags))
10167                 filemap_flush(inode->i_mapping);
10168
10169         if (delalloc_work->delay_iput)
10170                 btrfs_add_delayed_iput(inode);
10171         else
10172                 iput(inode);
10173         complete(&delalloc_work->completion);
10174 }
10175
10176 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
10177                                                     int delay_iput)
10178 {
10179         struct btrfs_delalloc_work *work;
10180
10181         work = kmalloc(sizeof(*work), GFP_NOFS);
10182         if (!work)
10183                 return NULL;
10184
10185         init_completion(&work->completion);
10186         INIT_LIST_HEAD(&work->list);
10187         work->inode = inode;
10188         work->delay_iput = delay_iput;
10189         WARN_ON_ONCE(!inode);
10190         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
10191                         btrfs_run_delalloc_work, NULL, NULL);
10192
10193         return work;
10194 }
10195
10196 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
10197 {
10198         wait_for_completion(&work->completion);
10199         kfree(work);
10200 }
10201
10202 /*
10203  * some fairly slow code that needs optimization. This walks the list
10204  * of all the inodes with pending delalloc and forces them to disk.
10205  */
10206 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
10207                                    int nr)
10208 {
10209         struct btrfs_inode *binode;
10210         struct inode *inode;
10211         struct btrfs_delalloc_work *work, *next;
10212         struct list_head works;
10213         struct list_head splice;
10214         int ret = 0;
10215
10216         INIT_LIST_HEAD(&works);
10217         INIT_LIST_HEAD(&splice);
10218
10219         mutex_lock(&root->delalloc_mutex);
10220         spin_lock(&root->delalloc_lock);
10221         list_splice_init(&root->delalloc_inodes, &splice);
10222         while (!list_empty(&splice)) {
10223                 binode = list_entry(splice.next, struct btrfs_inode,
10224                                     delalloc_inodes);
10225
10226                 list_move_tail(&binode->delalloc_inodes,
10227                                &root->delalloc_inodes);
10228                 inode = igrab(&binode->vfs_inode);
10229                 if (!inode) {
10230                         cond_resched_lock(&root->delalloc_lock);
10231                         continue;
10232                 }
10233                 spin_unlock(&root->delalloc_lock);
10234
10235                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
10236                 if (!work) {
10237                         if (delay_iput)
10238                                 btrfs_add_delayed_iput(inode);
10239                         else
10240                                 iput(inode);
10241                         ret = -ENOMEM;
10242                         goto out;
10243                 }
10244                 list_add_tail(&work->list, &works);
10245                 btrfs_queue_work(root->fs_info->flush_workers,
10246                                  &work->work);
10247                 ret++;
10248                 if (nr != -1 && ret >= nr)
10249                         goto out;
10250                 cond_resched();
10251                 spin_lock(&root->delalloc_lock);
10252         }
10253         spin_unlock(&root->delalloc_lock);
10254
10255 out:
10256         list_for_each_entry_safe(work, next, &works, list) {
10257                 list_del_init(&work->list);
10258                 btrfs_wait_and_free_delalloc_work(work);
10259         }
10260
10261         if (!list_empty_careful(&splice)) {
10262                 spin_lock(&root->delalloc_lock);
10263                 list_splice_tail(&splice, &root->delalloc_inodes);
10264                 spin_unlock(&root->delalloc_lock);
10265         }
10266         mutex_unlock(&root->delalloc_mutex);
10267         return ret;
10268 }
10269
10270 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
10271 {
10272         struct btrfs_fs_info *fs_info = root->fs_info;
10273         int ret;
10274
10275         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10276                 return -EROFS;
10277
10278         ret = __start_delalloc_inodes(root, delay_iput, -1);
10279         if (ret > 0)
10280                 ret = 0;
10281         return ret;
10282 }
10283
10284 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
10285                                int nr)
10286 {
10287         struct btrfs_root *root;
10288         struct list_head splice;
10289         int ret;
10290
10291         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10292                 return -EROFS;
10293
10294         INIT_LIST_HEAD(&splice);
10295
10296         mutex_lock(&fs_info->delalloc_root_mutex);
10297         spin_lock(&fs_info->delalloc_root_lock);
10298         list_splice_init(&fs_info->delalloc_roots, &splice);
10299         while (!list_empty(&splice) && nr) {
10300                 root = list_first_entry(&splice, struct btrfs_root,
10301                                         delalloc_root);
10302                 root = btrfs_grab_fs_root(root);
10303                 BUG_ON(!root);
10304                 list_move_tail(&root->delalloc_root,
10305                                &fs_info->delalloc_roots);
10306                 spin_unlock(&fs_info->delalloc_root_lock);
10307
10308                 ret = __start_delalloc_inodes(root, delay_iput, nr);
10309                 btrfs_put_fs_root(root);
10310                 if (ret < 0)
10311                         goto out;
10312
10313                 if (nr != -1) {
10314                         nr -= ret;
10315                         WARN_ON(nr < 0);
10316                 }
10317                 spin_lock(&fs_info->delalloc_root_lock);
10318         }
10319         spin_unlock(&fs_info->delalloc_root_lock);
10320
10321         ret = 0;
10322 out:
10323         if (!list_empty_careful(&splice)) {
10324                 spin_lock(&fs_info->delalloc_root_lock);
10325                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10326                 spin_unlock(&fs_info->delalloc_root_lock);
10327         }
10328         mutex_unlock(&fs_info->delalloc_root_mutex);
10329         return ret;
10330 }
10331
10332 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10333                          const char *symname)
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 btrfs_path *path;
10339         struct btrfs_key key;
10340         struct inode *inode = NULL;
10341         int err;
10342         int drop_inode = 0;
10343         u64 objectid;
10344         u64 index = 0;
10345         int name_len;
10346         int datasize;
10347         unsigned long ptr;
10348         struct btrfs_file_extent_item *ei;
10349         struct extent_buffer *leaf;
10350
10351         name_len = strlen(symname);
10352         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
10353                 return -ENAMETOOLONG;
10354
10355         /*
10356          * 2 items for inode item and ref
10357          * 2 items for dir items
10358          * 1 item for updating parent inode item
10359          * 1 item for the inline extent item
10360          * 1 item for xattr if selinux is on
10361          */
10362         trans = btrfs_start_transaction(root, 7);
10363         if (IS_ERR(trans))
10364                 return PTR_ERR(trans);
10365
10366         err = btrfs_find_free_ino(root, &objectid);
10367         if (err)
10368                 goto out_unlock;
10369
10370         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10371                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
10372                                 objectid, S_IFLNK|S_IRWXUGO, &index);
10373         if (IS_ERR(inode)) {
10374                 err = PTR_ERR(inode);
10375                 goto out_unlock;
10376         }
10377
10378         /*
10379         * If the active LSM wants to access the inode during
10380         * d_instantiate it needs these. Smack checks to see
10381         * if the filesystem supports xattrs by looking at the
10382         * ops vector.
10383         */
10384         inode->i_fop = &btrfs_file_operations;
10385         inode->i_op = &btrfs_file_inode_operations;
10386         inode->i_mapping->a_ops = &btrfs_aops;
10387         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10388
10389         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10390         if (err)
10391                 goto out_unlock_inode;
10392
10393         path = btrfs_alloc_path();
10394         if (!path) {
10395                 err = -ENOMEM;
10396                 goto out_unlock_inode;
10397         }
10398         key.objectid = btrfs_ino(BTRFS_I(inode));
10399         key.offset = 0;
10400         key.type = BTRFS_EXTENT_DATA_KEY;
10401         datasize = btrfs_file_extent_calc_inline_size(name_len);
10402         err = btrfs_insert_empty_item(trans, root, path, &key,
10403                                       datasize);
10404         if (err) {
10405                 btrfs_free_path(path);
10406                 goto out_unlock_inode;
10407         }
10408         leaf = path->nodes[0];
10409         ei = btrfs_item_ptr(leaf, path->slots[0],
10410                             struct btrfs_file_extent_item);
10411         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10412         btrfs_set_file_extent_type(leaf, ei,
10413                                    BTRFS_FILE_EXTENT_INLINE);
10414         btrfs_set_file_extent_encryption(leaf, ei, 0);
10415         btrfs_set_file_extent_compression(leaf, ei, 0);
10416         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10417         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10418
10419         ptr = btrfs_file_extent_inline_start(ei);
10420         write_extent_buffer(leaf, symname, ptr, name_len);
10421         btrfs_mark_buffer_dirty(leaf);
10422         btrfs_free_path(path);
10423
10424         inode->i_op = &btrfs_symlink_inode_operations;
10425         inode_nohighmem(inode);
10426         inode->i_mapping->a_ops = &btrfs_symlink_aops;
10427         inode_set_bytes(inode, name_len);
10428         btrfs_i_size_write(BTRFS_I(inode), name_len);
10429         err = btrfs_update_inode(trans, root, inode);
10430         /*
10431          * Last step, add directory indexes for our symlink inode. This is the
10432          * last step to avoid extra cleanup of these indexes if an error happens
10433          * elsewhere above.
10434          */
10435         if (!err)
10436                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
10437                                 BTRFS_I(inode), 0, index);
10438         if (err) {
10439                 drop_inode = 1;
10440                 goto out_unlock_inode;
10441         }
10442
10443         unlock_new_inode(inode);
10444         d_instantiate(dentry, inode);
10445
10446 out_unlock:
10447         btrfs_end_transaction(trans);
10448         if (drop_inode) {
10449                 inode_dec_link_count(inode);
10450                 iput(inode);
10451         }
10452         btrfs_btree_balance_dirty(fs_info);
10453         return err;
10454
10455 out_unlock_inode:
10456         drop_inode = 1;
10457         unlock_new_inode(inode);
10458         goto out_unlock;
10459 }
10460
10461 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10462                                        u64 start, u64 num_bytes, u64 min_size,
10463                                        loff_t actual_len, u64 *alloc_hint,
10464                                        struct btrfs_trans_handle *trans)
10465 {
10466         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
10467         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10468         struct extent_map *em;
10469         struct btrfs_root *root = BTRFS_I(inode)->root;
10470         struct btrfs_key ins;
10471         u64 cur_offset = start;
10472         u64 i_size;
10473         u64 cur_bytes;
10474         u64 last_alloc = (u64)-1;
10475         int ret = 0;
10476         bool own_trans = true;
10477         u64 end = start + num_bytes - 1;
10478
10479         if (trans)
10480                 own_trans = false;
10481         while (num_bytes > 0) {
10482                 if (own_trans) {
10483                         trans = btrfs_start_transaction(root, 3);
10484                         if (IS_ERR(trans)) {
10485                                 ret = PTR_ERR(trans);
10486                                 break;
10487                         }
10488                 }
10489
10490                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10491                 cur_bytes = max(cur_bytes, min_size);
10492                 /*
10493                  * If we are severely fragmented we could end up with really
10494                  * small allocations, so if the allocator is returning small
10495                  * chunks lets make its job easier by only searching for those
10496                  * sized chunks.
10497                  */
10498                 cur_bytes = min(cur_bytes, last_alloc);
10499                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10500                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10501                 if (ret) {
10502                         if (own_trans)
10503                                 btrfs_end_transaction(trans);
10504                         break;
10505                 }
10506                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
10507
10508                 last_alloc = ins.offset;
10509                 ret = insert_reserved_file_extent(trans, inode,
10510                                                   cur_offset, ins.objectid,
10511                                                   ins.offset, ins.offset,
10512                                                   ins.offset, 0, 0, 0,
10513                                                   BTRFS_FILE_EXTENT_PREALLOC);
10514                 if (ret) {
10515                         btrfs_free_reserved_extent(fs_info, ins.objectid,
10516                                                    ins.offset, 0);
10517                         btrfs_abort_transaction(trans, ret);
10518                         if (own_trans)
10519                                 btrfs_end_transaction(trans);
10520                         break;
10521                 }
10522
10523                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10524                                         cur_offset + ins.offset -1, 0);
10525
10526                 em = alloc_extent_map();
10527                 if (!em) {
10528                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10529                                 &BTRFS_I(inode)->runtime_flags);
10530                         goto next;
10531                 }
10532
10533                 em->start = cur_offset;
10534                 em->orig_start = cur_offset;
10535                 em->len = ins.offset;
10536                 em->block_start = ins.objectid;
10537                 em->block_len = ins.offset;
10538                 em->orig_block_len = ins.offset;
10539                 em->ram_bytes = ins.offset;
10540                 em->bdev = fs_info->fs_devices->latest_bdev;
10541                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10542                 em->generation = trans->transid;
10543
10544                 while (1) {
10545                         write_lock(&em_tree->lock);
10546                         ret = add_extent_mapping(em_tree, em, 1);
10547                         write_unlock(&em_tree->lock);
10548                         if (ret != -EEXIST)
10549                                 break;
10550                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10551                                                 cur_offset + ins.offset - 1,
10552                                                 0);
10553                 }
10554                 free_extent_map(em);
10555 next:
10556                 num_bytes -= ins.offset;
10557                 cur_offset += ins.offset;
10558                 *alloc_hint = ins.objectid + ins.offset;
10559
10560                 inode_inc_iversion(inode);
10561                 inode->i_ctime = current_time(inode);
10562                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10563                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10564                     (actual_len > inode->i_size) &&
10565                     (cur_offset > inode->i_size)) {
10566                         if (cur_offset > actual_len)
10567                                 i_size = actual_len;
10568                         else
10569                                 i_size = cur_offset;
10570                         i_size_write(inode, i_size);
10571                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10572                 }
10573
10574                 ret = btrfs_update_inode(trans, root, inode);
10575
10576                 if (ret) {
10577                         btrfs_abort_transaction(trans, ret);
10578                         if (own_trans)
10579                                 btrfs_end_transaction(trans);
10580                         break;
10581                 }
10582
10583                 if (own_trans)
10584                         btrfs_end_transaction(trans);
10585         }
10586         if (cur_offset < end)
10587                 btrfs_free_reserved_data_space(inode, NULL, cur_offset,
10588                         end - cur_offset + 1);
10589         return ret;
10590 }
10591
10592 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10593                               u64 start, u64 num_bytes, u64 min_size,
10594                               loff_t actual_len, u64 *alloc_hint)
10595 {
10596         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10597                                            min_size, actual_len, alloc_hint,
10598                                            NULL);
10599 }
10600
10601 int btrfs_prealloc_file_range_trans(struct inode *inode,
10602                                     struct btrfs_trans_handle *trans, int mode,
10603                                     u64 start, u64 num_bytes, u64 min_size,
10604                                     loff_t actual_len, u64 *alloc_hint)
10605 {
10606         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10607                                            min_size, actual_len, alloc_hint, trans);
10608 }
10609
10610 static int btrfs_set_page_dirty(struct page *page)
10611 {
10612         return __set_page_dirty_nobuffers(page);
10613 }
10614
10615 static int btrfs_permission(struct inode *inode, int mask)
10616 {
10617         struct btrfs_root *root = BTRFS_I(inode)->root;
10618         umode_t mode = inode->i_mode;
10619
10620         if (mask & MAY_WRITE &&
10621             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10622                 if (btrfs_root_readonly(root))
10623                         return -EROFS;
10624                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10625                         return -EACCES;
10626         }
10627         return generic_permission(inode, mask);
10628 }
10629
10630 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10631 {
10632         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10633         struct btrfs_trans_handle *trans;
10634         struct btrfs_root *root = BTRFS_I(dir)->root;
10635         struct inode *inode = NULL;
10636         u64 objectid;
10637         u64 index;
10638         int ret = 0;
10639
10640         /*
10641          * 5 units required for adding orphan entry
10642          */
10643         trans = btrfs_start_transaction(root, 5);
10644         if (IS_ERR(trans))
10645                 return PTR_ERR(trans);
10646
10647         ret = btrfs_find_free_ino(root, &objectid);
10648         if (ret)
10649                 goto out;
10650
10651         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10652                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
10653         if (IS_ERR(inode)) {
10654                 ret = PTR_ERR(inode);
10655                 inode = NULL;
10656                 goto out;
10657         }
10658
10659         inode->i_fop = &btrfs_file_operations;
10660         inode->i_op = &btrfs_file_inode_operations;
10661
10662         inode->i_mapping->a_ops = &btrfs_aops;
10663         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10664
10665         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10666         if (ret)
10667                 goto out_inode;
10668
10669         ret = btrfs_update_inode(trans, root, inode);
10670         if (ret)
10671                 goto out_inode;
10672         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10673         if (ret)
10674                 goto out_inode;
10675
10676         /*
10677          * We set number of links to 0 in btrfs_new_inode(), and here we set
10678          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10679          * through:
10680          *
10681          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10682          */
10683         set_nlink(inode, 1);
10684         unlock_new_inode(inode);
10685         d_tmpfile(dentry, inode);
10686         mark_inode_dirty(inode);
10687
10688 out:
10689         btrfs_end_transaction(trans);
10690         if (ret)
10691                 iput(inode);
10692         btrfs_balance_delayed_items(fs_info);
10693         btrfs_btree_balance_dirty(fs_info);
10694         return ret;
10695
10696 out_inode:
10697         unlock_new_inode(inode);
10698         goto out;
10699
10700 }
10701
10702 __attribute__((const))
10703 static int btrfs_readpage_io_failed_hook(struct page *page, int failed_mirror)
10704 {
10705         return -EAGAIN;
10706 }
10707
10708 static struct btrfs_fs_info *iotree_fs_info(void *private_data)
10709 {
10710         struct inode *inode = private_data;
10711         return btrfs_sb(inode->i_sb);
10712 }
10713
10714 static void btrfs_check_extent_io_range(void *private_data, const char *caller,
10715                                         u64 start, u64 end)
10716 {
10717         struct inode *inode = private_data;
10718         u64 isize;
10719
10720         isize = i_size_read(inode);
10721         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
10722                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
10723                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
10724                         caller, btrfs_ino(BTRFS_I(inode)), isize, start, end);
10725         }
10726 }
10727
10728 void btrfs_set_range_writeback(void *private_data, u64 start, u64 end)
10729 {
10730         struct inode *inode = private_data;
10731         unsigned long index = start >> PAGE_SHIFT;
10732         unsigned long end_index = end >> PAGE_SHIFT;
10733         struct page *page;
10734
10735         while (index <= end_index) {
10736                 page = find_get_page(inode->i_mapping, index);
10737                 ASSERT(page); /* Pages should be in the extent_io_tree */
10738                 set_page_writeback(page);
10739                 put_page(page);
10740                 index++;
10741         }
10742 }
10743
10744 static const struct inode_operations btrfs_dir_inode_operations = {
10745         .getattr        = btrfs_getattr,
10746         .lookup         = btrfs_lookup,
10747         .create         = btrfs_create,
10748         .unlink         = btrfs_unlink,
10749         .link           = btrfs_link,
10750         .mkdir          = btrfs_mkdir,
10751         .rmdir          = btrfs_rmdir,
10752         .rename         = btrfs_rename2,
10753         .symlink        = btrfs_symlink,
10754         .setattr        = btrfs_setattr,
10755         .mknod          = btrfs_mknod,
10756         .listxattr      = btrfs_listxattr,
10757         .permission     = btrfs_permission,
10758         .get_acl        = btrfs_get_acl,
10759         .set_acl        = btrfs_set_acl,
10760         .update_time    = btrfs_update_time,
10761         .tmpfile        = btrfs_tmpfile,
10762 };
10763 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10764         .lookup         = btrfs_lookup,
10765         .permission     = btrfs_permission,
10766         .update_time    = btrfs_update_time,
10767 };
10768
10769 static const struct file_operations btrfs_dir_file_operations = {
10770         .llseek         = generic_file_llseek,
10771         .read           = generic_read_dir,
10772         .iterate_shared = btrfs_real_readdir,
10773         .open           = btrfs_opendir,
10774         .unlocked_ioctl = btrfs_ioctl,
10775 #ifdef CONFIG_COMPAT
10776         .compat_ioctl   = btrfs_compat_ioctl,
10777 #endif
10778         .release        = btrfs_release_file,
10779         .fsync          = btrfs_sync_file,
10780 };
10781
10782 static const struct extent_io_ops btrfs_extent_io_ops = {
10783         /* mandatory callbacks */
10784         .submit_bio_hook = btrfs_submit_bio_hook,
10785         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10786         .merge_bio_hook = btrfs_merge_bio_hook,
10787         .readpage_io_failed_hook = btrfs_readpage_io_failed_hook,
10788         .tree_fs_info = iotree_fs_info,
10789         .set_range_writeback = btrfs_set_range_writeback,
10790
10791         /* optional callbacks */
10792         .fill_delalloc = run_delalloc_range,
10793         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10794         .writepage_start_hook = btrfs_writepage_start_hook,
10795         .set_bit_hook = btrfs_set_bit_hook,
10796         .clear_bit_hook = btrfs_clear_bit_hook,
10797         .merge_extent_hook = btrfs_merge_extent_hook,
10798         .split_extent_hook = btrfs_split_extent_hook,
10799         .check_extent_io_range = btrfs_check_extent_io_range,
10800 };
10801
10802 /*
10803  * btrfs doesn't support the bmap operation because swapfiles
10804  * use bmap to make a mapping of extents in the file.  They assume
10805  * these extents won't change over the life of the file and they
10806  * use the bmap result to do IO directly to the drive.
10807  *
10808  * the btrfs bmap call would return logical addresses that aren't
10809  * suitable for IO and they also will change frequently as COW
10810  * operations happen.  So, swapfile + btrfs == corruption.
10811  *
10812  * For now we're avoiding this by dropping bmap.
10813  */
10814 static const struct address_space_operations btrfs_aops = {
10815         .readpage       = btrfs_readpage,
10816         .writepage      = btrfs_writepage,
10817         .writepages     = btrfs_writepages,
10818         .readpages      = btrfs_readpages,
10819         .direct_IO      = btrfs_direct_IO,
10820         .invalidatepage = btrfs_invalidatepage,
10821         .releasepage    = btrfs_releasepage,
10822         .set_page_dirty = btrfs_set_page_dirty,
10823         .error_remove_page = generic_error_remove_page,
10824 };
10825
10826 static const struct address_space_operations btrfs_symlink_aops = {
10827         .readpage       = btrfs_readpage,
10828         .writepage      = btrfs_writepage,
10829         .invalidatepage = btrfs_invalidatepage,
10830         .releasepage    = btrfs_releasepage,
10831 };
10832
10833 static const struct inode_operations btrfs_file_inode_operations = {
10834         .getattr        = btrfs_getattr,
10835         .setattr        = btrfs_setattr,
10836         .listxattr      = btrfs_listxattr,
10837         .permission     = btrfs_permission,
10838         .fiemap         = btrfs_fiemap,
10839         .get_acl        = btrfs_get_acl,
10840         .set_acl        = btrfs_set_acl,
10841         .update_time    = btrfs_update_time,
10842 };
10843 static const struct inode_operations btrfs_special_inode_operations = {
10844         .getattr        = btrfs_getattr,
10845         .setattr        = btrfs_setattr,
10846         .permission     = btrfs_permission,
10847         .listxattr      = btrfs_listxattr,
10848         .get_acl        = btrfs_get_acl,
10849         .set_acl        = btrfs_set_acl,
10850         .update_time    = btrfs_update_time,
10851 };
10852 static const struct inode_operations btrfs_symlink_inode_operations = {
10853         .get_link       = page_get_link,
10854         .getattr        = btrfs_getattr,
10855         .setattr        = btrfs_setattr,
10856         .permission     = btrfs_permission,
10857         .listxattr      = btrfs_listxattr,
10858         .update_time    = btrfs_update_time,
10859 };
10860
10861 const struct dentry_operations btrfs_dentry_operations = {
10862         .d_delete       = btrfs_dentry_delete,
10863         .d_release      = btrfs_dentry_release,
10864 };