On Tue, May 29, 2018 at 03:43:32PM -0700, Darrick J. Wong wrote:
> On Wed, May 30, 2018 at 08:24:28AM +1000, Dave Chinner wrote:
> > On Tue, May 29, 2018 at 03:07:16PM -0700, Darrick J. Wong wrote:
> > > On Tue, May 29, 2018 at 01:28:10PM +1000, Dave Chinner wrote:
> > > > On Thu, May 17, 2018 at 08:56:23PM -0700, Darrick J. Wong wrote:
> > > > > + /*
> > > > > + * Reinitialize the counters. The on-disk and in-core counters differ
> > > > > + * by the number of inodes/blocks reserved by the admin, the per-AG
> > > > > + * reservation, and any transactions in progress, so we have to
> > > > > + * account for that. First we take the sb lock and update its
> > > > > + * counters...
> > > > > + */
> > > > > + spin_lock(&mp->m_sb_lock);
> > > > > + delta_icount = (int64_t)mp->m_sb.sb_icount - icount;
> > > > > + delta_ifree = (int64_t)mp->m_sb.sb_ifree - ifree;
> > > > > + delta_fdblocks = (int64_t)mp->m_sb.sb_fdblocks - fdblocks;
> > > > > + mp->m_sb.sb_icount = icount;
> > > > > + mp->m_sb.sb_ifree = ifree;
> > > > > + mp->m_sb.sb_fdblocks = fdblocks;
> > > > > + spin_unlock(&mp->m_sb_lock);
> > > >
> > > > This seems racy to me ? i.e. the per-ag counters can change while
> > > > we are summing them, and once we've summed them then sb counters
> > > > can change while we are waiting for the m_sb_lock. It's looks to me
> > > > like the summed per-ag counters are not in any way coherent
> > > > wit the superblock or the in-core per-CPU counters, so I'm
> > > > struggling to understand why this is safe?
> > >
> > > Hmm, yes, I think this is racy too. The purpose of this code is to
> > > recompute the global counters from the AG counters after any operation
> > > that modifies anything that would affect the icount/ifreecount/fdblocks
> > > counters...
> >
> > *nod*
> >
> > > > We can do this sort of summation at mount time (in
> > > > xfs_initialize_perag_data()) because the filesystem is running
> > > > single threaded while the summation is taking place and so nothing
> > > > is changing during th summation. The filesystem is active in this
> > > > case, so I don't think we can do the same thing here.
> > >
> > > ...however, you're correct to point out that the fs must be quiesced
> > > before we can actually do this. In other words, I think the filesystem
> > > has to be completely frozen before we can do this. Perhaps it's better
> > > to have the per-ag rebuilders fix only the per-ag counters and leave the
> > > global counters alone. Then add a new scrubber that checks the summary
> > > counters and fixes them if necessary.
> >
> > So the question here is whether we actually need to accurately
> > correct the global superblock counters?
>
> I think so, because what happens if the superblock counter is
> artificially high but the AGs do not actually have the free space?
> xfs_trans_reserve won't ENOSPC like it should, so we could end up
> blowing out of transactions and shutting down because some allocation
> that has to succeed ("because trans_reserve said there was space!")
> fails...
Yes, but I would have thought the reset will get us close enough
that this wouldn't be an issue for the vast majority of people.
And the other side of it is that if we get close enough to ENOSPC
that it matters, we could freeze/sum/thaw to be fully accurate on
demand in xfs_trans_reserve(), right? We already slow down greatly
at ENOSPC, so at that point the perf overhead fo a freeze/thaw cycle
just doesn't matter...
> > We know that if we have a dirty unmount, the counters will we
> > re-initialised on mount from the AG header information, so perhaps
> > what we need here is a flag to tell unmount to dirty the log again
> > after it has written the unmount record (like we currently do for
> > quiesce).
>
> ...but now that we've repaired the filesystem, it could potentially run
> for a very long time until the next unmount. During that run, we'd be
> misleading users about the real amount of free space and risking a hard
> shutdown. I prefer that online repair try not to leave any weird state
> around after xfs_scrub exits.
Sure, but user's may not want a freeze/read-all-ag-headers/thaw
cycle as part of repair if it can be avoided. If there are thousands
of AGs, this could take many seconds....
> > That was we can do a racy "near enough" update here to get us out of
> > the worst of the space accounting mismatches, knowing that on the
> > next mount it will be accurately rebuilt.
> >
> > Thoughts?
>
> Well, I think the best solution is to have the AGF/AGI/inobt rebuilders
> adjust the global counters by the same amount that they're adjusting the
> counters in the AGF/AGI, then add a new scrubber that runs at the end to
> freeze the fs and check/repair the global counter state. :)
I'm just not convinced that we can get away with a global freeze to
do this summation without having noticable impact on applications.
Cheers,
Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx
--
To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
