When we abort a transaction we iterate over all the ranges marked as dirty
in fs_info->freed_extents[0] and fs_info->freed_extents[1], clear them
from those trees, add them back (unpin) to the free space caches and, if
the fs was mounted with "-o discard", perform a discard on those regions.
Also, after adding the regions to the free space caches, a fitrim ioctl call
can see those ranges in a block group's free space cache and perform a discard
on the ranges, so the same issue can happen without "-o discard" as well.

This causes corruption, affecting one or multiple btree nodes (in the worst
case leaving the fs unmountable) because some of those ranges (the ones in
the fs_info->pinned_extents tree) correspond to btree nodes/leafs that are
referred by the last committed super block - breaking the rule that anything
that was committed by a transaction is untouched until the next transaction
commits successfully.

I ran into this while running in a loop (for several hours) the fstest that
I recently submitted:

  [PATCH] fstests: add btrfs test to stress chunk allocation/removal and fstrim

The corruption always happened when a transaction aborted and then fsck complained
like this:

   _check_btrfs_filesystem: filesystem on /dev/sdc is inconsistent
   *** fsck.btrfs output ***
   Check tree block failed, want=94945280, have=0
   Check tree block failed, want=94945280, have=0
   Check tree block failed, want=94945280, have=0
   Check tree block failed, want=94945280, have=0
   Check tree block failed, want=94945280, have=0
   read block failed check_tree_block
   Couldn't open file system

In this case 94945280 corresponded to the root of a tree.
Using frace what I observed was the following sequence of steps happened:

   1) transaction N started, fs_info->pinned_extents pointed to
      fs_info->freed_extents[0];

   2) node/eb 94945280 is created;

   3) eb is persisted to disk;

   4) transaction N commit starts, fs_info->pinned_extents now points to
      fs_info->freed_extents[1], and transaction N completes;

   5) transaction N + 1 starts;

   6) eb is COWed, and btrfs_free_tree_block() called for this eb;

   7) eb range (94945280 to 94945280 + 16Kb) is added to
      fs_info->pinned_extents (fs_info->freed_extents[1]);

   8) Something goes wrong in transaction N + 1, like hitting ENOSPC
      for example, and the transaction is aborted, turning the fs into
      readonly mode. The stack trace I got for example:

      [112065.253935]  [<ffffffff8140c7b6>] dump_stack+0x4d/0x66
      [112065.254271]  [<ffffffff81042984>] warn_slowpath_common+0x7f/0x98
      [112065.254567]  [<ffffffffa0325990>] ? __btrfs_abort_transaction+0x50/0x10b [btrfs]
      [112065.261674]  [<ffffffff810429e5>] warn_slowpath_fmt+0x48/0x50
      [112065.261922]  [<ffffffffa032949e>] ? btrfs_free_path+0x26/0x29 [btrfs]
      [112065.262211]  [<ffffffffa0325990>] __btrfs_abort_transaction+0x50/0x10b [btrfs]
      [112065.262545]  [<ffffffffa036b1d6>] btrfs_remove_chunk+0x537/0x58b [btrfs]
      [112065.262771]  [<ffffffffa033840f>] btrfs_delete_unused_bgs+0x1de/0x21b [btrfs]
      [112065.263105]  [<ffffffffa0343106>] cleaner_kthread+0x100/0x12f [btrfs]
      (...)
      [112065.264493] ---[ end trace dd7903a975a31a08 ]---
      [112065.264673] BTRFS: error (device sdc) in btrfs_remove_chunk:2625: errno=-28 No space left
      [112065.264997] BTRFS info (device sdc): forced readonly

   9) The clear kthread sees that the BTRFS_FS_STATE_ERROR bit is set in
      fs_info->fs_state and calls btrfs_cleanup_transaction(), which in
      turn calls btrfs_destroy_pinned_extent();

   10) Then btrfs_destroy_pinned_extent() iterates over all the ranges
       marked as dirty in fs_info->freed_extents[], and for each one
       it calls discard, if the fs was mounted with "-o discard", and
       adds the range to the free space cache of the respective block
       group;

   11) btrfs_trim_block_group(), invoked from the fitrim ioctl code path,
       sees the free space entries and performs a discard;

   12) After an umount and mount (or fsck), our eb's location on disk was full
       of zeroes, and it should have been untouched, because it was marked as
       dirty in the fs_info->pinned_extents tree, and therefore used by the
       trees that the last committed superblock points to.

Fix this by not performing a discard and not adding the ranges to the free space
caches - it's useless from this point since the fs is now in readonly mode and
we won't write free space caches to disk anymore (otherwise we would leak space)
nor any new superblock. By not adding the ranges to the free space caches, it
prevents other code paths from allocating that space and write to it as well,
therefore being safer and simpler.

This isn't a new problem, as it's been present since 2011 (git commit
acce952b).

Cc: stable@vger.kernel.org  # any kernel released after 2011-01-06
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>