Recent FITRIM work, namely bbbf7243 ("btrfs: combine device update
operations during transaction commit") combined the way certain
operations are recoded in a transaction. As a result an ASSERT was added
in dev_replace_finish to ensure the new code works correctly.
Unfortunately I got reports that it's possible to trigger the assert,
meaning that during a device replace it's possible to have an unfinished
chunk allocation on the source device.

This is supposed to be prevented by the fact that a transaction is
committed before finishing the replace oepration and alter acquiring the
chunk mutex. This is not sufficient since by the time the transaction is
committed and the chunk mutex acquired it's possible to allocate a chunk
depending on the workload being executed on the replaced device. This
bug has been present ever since device replace was introduced but there
was never code which checks for it.

The correct way to fix is to ensure that there is no pending device
modification operation when the chunk mutex is acquire and if there is
repeat transaction commit. Unfortunately it's not possible to just
exclude the source device from btrfs_fs_devices::dev_alloc_list since
this causes ENOSPC to be hit in transaction commit.

Fixing that in another way would need to add special cases to handle the
last writes and forbid new ones. The looped transaction fix is more
obvious, and can be easily backported. The runtime of dev-replace is
long so there's no noticeable delay caused by that.
Reported-by: David Sterba <dsterba@suse.com>
Fixes: 391cd9df ("Btrfs: fix unprotected alloc list insertion during the finishing procedure of replace")
CC: stable@vger.kernel.org # 4.4+
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>