On Mon, Feb 17, 2020 at 10:29:15AM -0500, Brian Foster wrote:
> On Mon, Feb 17, 2020 at 05:33:14AM -0800, Christoph Hellwig wrote:
> > On Fri, Feb 14, 2020 at 01:15:28PM -0500, Brian Foster wrote:
> > > Prior to commit df732b29c8 ("xfs: call xlog_state_release_iclog with
> > > l_icloglock held"), xlog_state_release_iclog() always performed a
> > > locked check of the iclog error state before proceeding into the
> > > sync state processing code. As of this commit, part of
> > > xlog_state_release_iclog() was open-coded into
> > > xfs_log_release_iclog() and as a result the locked error state check
> > > was lost.
> > >
> > > The lockless check still exists, but this doesn't account for the
> > > possibility of a race with a shutdown being performed by another
> > > task causing the iclog state to change while the original task waits
> > > on ->l_icloglock. This has reproduced very rarely via generic/475
> > > and manifests as an assert failure in __xlog_state_release_iclog()
> > > due to an unexpected iclog state.
> > >
> > > Restore the locked error state check in xlog_state_release_iclog()
> > > to ensure that an iclog state update via shutdown doesn't race with
> > > the iclog release state processing code.
> > >
> > > Reported-by: Zorro Lang <zlang@xxxxxxxxxx>
> > > Signed-off-by: Brian Foster <bfoster@xxxxxxxxxx>
> > > ---
> > > fs/xfs/xfs_log.c | 4 ++++
> > > 1 file changed, 4 insertions(+)
> > >
> > > diff --git a/fs/xfs/xfs_log.c b/fs/xfs/xfs_log.c
> > > index f6006d94a581..f38fc492a14d 100644
> > > --- a/fs/xfs/xfs_log.c
> > > +++ b/fs/xfs/xfs_log.c
> > > @@ -611,6 +611,10 @@ xfs_log_release_iclog(
> > > }
> > >
> > > if (atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock)) {
> > > + if (iclog->ic_state == XLOG_STATE_IOERROR) {
> > > + spin_unlock(&log->l_icloglock);
> > > + return -EIO;
> > > + }
> >
> > So the check just above also shuts the file system down. Any reason to
> > do that in one case and not the other?
> >
>
> The initial check (with the shutdown) was originally associated with the
> return from xlog_state_release_iclog(). That covers both state checks,
> as they were both originally within that function. My impression was
> there isn't a need to shutdown in the second check because the only way
> the iclog state changes to IOERROR across that lock cycle is due to a
> shutdown already in progress.
The original code did the force shutdown for both cases. So unless we
have a good reason to do it differently I'd just add a goto label and
merge the two cases to restore the old behavior.
