btrfs: release correct delalloc amount in direct IO write path
authorNaohiro Aota <naohiro.aota@wdc.com>
Mon, 28 Mar 2022 12:32:05 +0000 (21:32 +0900)
committerDavid Sterba <dsterba@suse.com>
Tue, 5 Apr 2022 22:49:35 +0000 (00:49 +0200)
Running generic/406 causes the following WARNING in btrfs_destroy_inode()
which tells there are outstanding extents left.

In btrfs_get_blocks_direct_write(), we reserve a temporary outstanding
extents with btrfs_delalloc_reserve_metadata() (or indirectly from
btrfs_delalloc_reserve_space(()). We then release the outstanding extents
with btrfs_delalloc_release_extents(). However, the "len" can be modified
in the COW case, which releases fewer outstanding extents than expected.

Fix it by calling btrfs_delalloc_release_extents() for the original length.

To reproduce the warning, the filesystem should be 1 GiB.  It's
triggering a short-write, due to not being able to allocate a large
extent and instead allocating a smaller one.

  WARNING: CPU: 0 PID: 757 at fs/btrfs/inode.c:8848 btrfs_destroy_inode+0x1e6/0x210 [btrfs]
  Modules linked in: btrfs blake2b_generic xor lzo_compress
  lzo_decompress raid6_pq zstd zstd_decompress zstd_compress xxhash zram
  zsmalloc
  CPU: 0 PID: 757 Comm: umount Not tainted 5.17.0-rc8+ #101
  Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS d55cb5a 04/01/2014
  RIP: 0010:btrfs_destroy_inode+0x1e6/0x210 [btrfs]
  RSP: 0018:ffffc9000327bda8 EFLAGS: 00010206
  RAX: 0000000000000000 RBX: ffff888100548b78 RCX: 0000000000000000
  RDX: 0000000000026900 RSI: 0000000000000000 RDI: ffff888100548b78
  RBP: ffff888100548940 R08: 0000000000000000 R09: ffff88810b48aba8
  R10: 0000000000000001 R11: ffff8881004eb240 R12: ffff88810b48a800
  R13: ffff88810b48ec08 R14: ffff88810b48ed00 R15: ffff888100490c68
  FS:  00007f8549ea0b80(0000) GS:ffff888237c00000(0000) knlGS:0000000000000000
  CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  CR2: 00007f854a09e733 CR3: 000000010a2e9003 CR4: 0000000000370eb0
  DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
  DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
  Call Trace:
   <TASK>
   destroy_inode+0x33/0x70
   dispose_list+0x43/0x60
   evict_inodes+0x161/0x1b0
   generic_shutdown_super+0x2d/0x110
   kill_anon_super+0xf/0x20
   btrfs_kill_super+0xd/0x20 [btrfs]
   deactivate_locked_super+0x27/0x90
   cleanup_mnt+0x12c/0x180
   task_work_run+0x54/0x80
   exit_to_user_mode_prepare+0x152/0x160
   syscall_exit_to_user_mode+0x12/0x30
   do_syscall_64+0x42/0x80
   entry_SYSCALL_64_after_hwframe+0x44/0xae
   RIP: 0033:0x7f854a000fb7

Fixes: f0bfa76a11e9 ("btrfs: fix ENOSPC failure when attempting direct IO write into NOCOW range")
CC: stable@vger.kernel.org # 5.17
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Tested-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com>
Signed-off-by: David Sterba <dsterba@suse.com>
fs/btrfs/inode.c

index 286ab7c27db089989cd0bb567e0755a12da45841..53a3f5e5ae899d5d9c59ef646cdc2f5687661aa8 100644 (file)
@@ -7442,6 +7442,7 @@ static int btrfs_get_blocks_direct_write(struct extent_map **map,
        u64 block_start, orig_start, orig_block_len, ram_bytes;
        bool can_nocow = false;
        bool space_reserved = false;
+       u64 prev_len;
        int ret = 0;
 
        /*
@@ -7469,6 +7470,7 @@ static int btrfs_get_blocks_direct_write(struct extent_map **map,
                        can_nocow = true;
        }
 
+       prev_len = len;
        if (can_nocow) {
                struct extent_map *em2;
 
@@ -7498,8 +7500,6 @@ static int btrfs_get_blocks_direct_write(struct extent_map **map,
                        goto out;
                }
        } else {
-               const u64 prev_len = len;
-
                /* Our caller expects us to free the input extent map. */
                free_extent_map(em);
                *map = NULL;
@@ -7530,7 +7530,7 @@ static int btrfs_get_blocks_direct_write(struct extent_map **map,
         * We have created our ordered extent, so we can now release our reservation
         * for an outstanding extent.
         */
-       btrfs_delalloc_release_extents(BTRFS_I(inode), len);
+       btrfs_delalloc_release_extents(BTRFS_I(inode), prev_len);
 
        /*
         * Need to update the i_size under the extent lock so buffered