fs/btrfs/inode.c

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