Re: [PATCH] xfs: drain inodegc before quota teardown on mount failure

[Cen Zhang]

Hi Christoph,

> Can you come up with an xfstest for this?  I guess you'd need a new
> error injection point, but that should not be horrible.

I tried to turn this into an xfstest.

In my local attempts, a mount-failure errortag after
xfs_qm_mount_quotas() was not enough by itself, because ordinary test
setup did not reliably leave an inodegc item that would enter the
dqattach/eofblocks path during the mount-failure unwind.

The draft I have locally uses two DEBUG-only errortags:

  - mount_fail_after_qm_mount_quotas

    This injects -EIO immediately after xfs_qm_mount_quotas().

  - force_inodegc_eofblocks

    This makes a regular inode released after quotacheck scanning go
    through inodegc/eofblocks cleanup, and delays that inodegc work so
    the mount-failure unwind can reach quota teardown first.

The xfstest creates one regular scratch file without quotas enabled,
unmounts, and then remounts with:

  -o uquota,errortag=force_inodegc_eofblocks,errortag=mount_fail_after_qm_mount_quotas

On a kernel with the errortags but without the inodegc drain fix, this
reproduces the same teardown ordering problem:

  KASAN: null-ptr-deref
  Workqueue: xfs-inodegc/vdc xfs_inodegc_worker
  xfs_qm_dqget_cache_lookup()
  xfs_qm_dqget_inode()
  xfs_qm_dqattach_locked()
  xfs_qm_dqattach()
  xfs_free_eofblocks()
  xfs_inactive()
  xfs_inodegc_worker()

With the drain fix applied, the same xfstest passes and the filesystem
can be mounted again cleanly afterwards.

I do not have much experience with XFS-specific fstests, so I wanted to
check whether this is the right shape before posting it formally.  If
this approach looks acceptable, I can send a small kernel series adding
the DEBUG errortags and the mount-failure fix, followed by a separate
xfstests patch.

Thanks,
Zhang Cen