Btrfs: fix unprotected list move from unused_bgs to deleted_bgs list
authorFilipe Manana <fdmanana@suse.com>
Fri, 27 Nov 2015 12:16:16 +0000 (12:16 +0000)
committerFilipe Manana <fdmanana@suse.com>
Thu, 10 Dec 2015 11:22:38 +0000 (11:22 +0000)
commit348a0013d54acec35c22958480af054b97b5e4fe
treeff187ba3c2f0b4658f695fce69a8f121ae1be549
parentdba72cb30b6a4811038128c8a98b268d18ca60fe
Btrfs: fix unprotected list move from unused_bgs to deleted_bgs list

As of my previous change titled "Btrfs: fix scrub preventing unused block
groups from being deleted", the following warning at
extent-tree.c:btrfs_delete_unused_bgs() can be hit when we mount the a
filesysten with "-o discard":

 10263  void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
 10264  {
 (...)
 10405                  if (trimming) {
 10406                          WARN_ON(!list_empty(&block_group->bg_list));
 10407                          spin_lock(&trans->transaction->deleted_bgs_lock);
 10408                          list_move(&block_group->bg_list,
 10409                                    &trans->transaction->deleted_bgs);
 10410                          spin_unlock(&trans->transaction->deleted_bgs_lock);
 10411                          btrfs_get_block_group(block_group);
 10412                  }
 (...)

This happens because scrub can now add back the block group to the list of
unused block groups (fs_info->unused_bgs). This is dangerous because we
are moving the block group from the unused block groups list to the list
of deleted block groups without holding the lock that protects the source
list (fs_info->unused_bgs_lock).

The following diagram illustrates how this happens:

            CPU 1                                     CPU 2

 cleaner_kthread()
   btrfs_delete_unused_bgs()

     sees bg X in list
      fs_info->unused_bgs

     deletes bg X from list
      fs_info->unused_bgs

                                            scrub_enumerate_chunks()

                                              searches device tree using
                                              its commit root

                                              finds device extent for
                                              block group X

                                              gets block group X from the tree
                                              fs_info->block_group_cache_tree
                                              (via btrfs_lookup_block_group())

                                              sets bg X to RO (again)

                                              scrub_chunk(bg X)

                                              sets bg X back to RW mode

                                              adds bg X to the list
                                              fs_info->unused_bgs again,
                                              since it's still unused and
                                              currently not in that list

     sets bg X to RO mode

     btrfs_remove_chunk(bg X)

     --> discard is enabled and bg X
         is in the fs_info->unused_bgs
         list again so the warning is
         triggered
     --> we move it from that list into
         the transaction's delete_bgs
         list, but we can have another
         task currently manipulating
         the first list (fs_info->unused_bgs)

Fix this by using the same lock (fs_info->unused_bgs_lock) to protect both
the list of unused block groups and the list of deleted block groups. This
makes it safe and there's not much worry for more lock contention, as this
lock is seldom used and only the cleaner kthread adds elements to the list
of deleted block groups. The warning goes away too, as this was previously
an impossible case (and would have been better a BUG_ON/ASSERT) but it's
not impossible anymore.
Reproduced with fstest btrfs/073 (using MOUNT_OPTIONS="-o discard").

Signed-off-by: Filipe Manana <fdmanana@suse.com>
fs/btrfs/extent-tree.c
fs/btrfs/transaction.c
fs/btrfs/transaction.h