[xfs-masters] INFO: possible circular locking dependency detected

[Christoph Hellwig]

This has been around for a long time, the VM calls into fput and thus
->release with the mmap_sem held, which makes it really hard for a
filesystem to avoid a lock inversion if it uses a lock both in ->release
and the I/O path.

I have a workaround for this to only acquire the XFS iolock with a
trylock in the release path and leave cleaning up stale preallocations
until the final iput.  It's not pretty, but given that we're unlikely
to see the VM fixed I might aswell finally send it for inclusion.