fs/btrfs/inode.c

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