On Wed, May 29, 2013 at 09:54:24AM +1000, Dave Chinner wrote:
> On Tue, May 28, 2013 at 12:56:27PM -0500, Ben Myers wrote:
> > On Mon, May 27, 2013 at 04:38:18PM +1000, Dave Chinner wrote:
> > > This is an update for all the fixes I have for 3,10-rc4.
> > >
> > > All the comments on the previous patches have been updated in this
> > > series, and there are two new patches. The first patch adds support
> > > for indicating that the filesystem has CRC enabled to userspace, and
> > > the second fixes the missing CRC updates on the unlinked inode list
> > > manipulations.
> > This seems to be getting a bit out of hand for a release candidate pull
> > request:
> You don't need to send them all at once. The ones you committed from
> my initial -rc2 series could just be sent to Linus immediately, and
> that halves the current outstanding queue. Then there's only 8-9
> patches remaining...
I am more concerned about the content of the patches than the number of
> > Fixes for xfs without crcs:
> > f648167f xfs: avoid nesting transactions in xfs_qm_scall_setqlim()
> > xfs: fix split buffer vector log recovery support
> > xfs: kill suid/sgid through the truncate path.
> And the log recovery ordering fix I just sent out.
> > Fixes for xfs with crcs:
> > 90253cf1 xfs: remote attribute allocation may be contiguous
> > 913e96bc xfs: remote attribute read too short
> > 4af3644c xfs: remote attribute tail zeroing does too much
> > 6863ef84 xfs: correctly map remote attr buffers during removal
> > 8517de2a xfs: fully initialise temp leaf in xfs_attr3_leaf_unbalance
> > d4c712bc xfs: fully initialise temp leaf in xfs_attr3_leaf_compact
> All of which fix filesystem corruption or - in the case of the
> buffer cache coherency problems - can lead to stale data exposure.
> > ad1858d7 xfs: rework remote attr CRCs
> Fixes cache coherency issues via a change in the on-disk format for
> remote attributes. We need to get that included before a full
> release is made, otherwise we need feature bits to distinguish
> between the original and the new formats.
> > 9a5b6dee xfs: fix incorrect remote symlink block count
> > xfs: rework dquot CRCs
> More filesystem corruption fixes.
> > xfs: inode unlinked list needs to recalculate the inode CRC *
> > replaced with "2-3 patches"
> Another corruption fix, now with an addition log recovery ordering
> fix which prevents log recovery from potentially causing silent
> filesystem corruption.
> > xfs: fix dir3 freespace block corruption * maybe
> > this effects filesystems w/o crcs?
> Again, a filesystem corruption, including fixing a problem in the
> on-disk format definition. While it probably doesn't affect
> non-crc filesystems, it's a landmine that should be removed....
> > Dev work for xfs with crcs:
> > xfs: increase number of ACL entries for V5 superblocks
> On disk format change. We need to get that out before release,
> otherwise it needs a feature bit.
> > xfs: don't emit v5 superblock warnings on write
> Fixes a nasty log spamming problem. Certainly not development code.
> > xfs: add fsgeom flag for v5 superblock support.
> Needed for userspace support of CRC filesystems, trivial patch. if
> we expect anyone to test this kernel code, we need this patch so
> userspace can correctly identify crc-enabled filesystems.
> > xfs: disable swap extents ioctl on CRC enabled filesystems
> Filesystem corruption fix, caused simply by running xfs_fsr.
> But, really, this entire series is not that much change in terms of
> $ git diff --stat b38958d..bf6331a -- fs/xfs
> fs/xfs/xfs_acl.c | 31 ++--
> fs/xfs/xfs_acl.h | 30 +++-
> fs/xfs/xfs_attr_leaf.c | 74 ++++++---
> fs/xfs/xfs_attr_remote.c | 408
> fs/xfs/xfs_attr_remote.h | 10 ++
> fs/xfs/xfs_buf.c | 1 +
> fs/xfs/xfs_buf_item.c | 7 +-
> fs/xfs/xfs_dfrag.c | 8 +
> fs/xfs/xfs_dir2_format.h | 1 +
> fs/xfs/xfs_dir2_node.c | 13 +-
> fs/xfs/xfs_dquot.c | 37 ++---
> fs/xfs/xfs_fs.h | 1 +
> fs/xfs/xfs_fsops.c | 4 +-
> fs/xfs/xfs_inode.c | 42 ++++-
> fs/xfs/xfs_iops.c | 47 ++++--
> fs/xfs/xfs_log_recover.c | 95 ++++++++++-
> fs/xfs/xfs_mount.c | 18 ++-
> fs/xfs/xfs_qm.c | 36 +++--
> fs/xfs/xfs_qm_syscalls.c | 40 +++--
> fs/xfs/xfs_quota.h | 2 +
> fs/xfs/xfs_symlink.c | 20 +--
> 21 files changed, 612 insertions(+), 313 deletions(-)
> And a large proportion of that change is the single attr CRC rework
> > We're representing 3 patches which are clearly approprate for a release
> > candidate, 11 fixes for CRCs + 2 more for the rework of unlinked lists
> > patch,
> > and 4 patches which are development work.
> Most file filesystem corruptions. Without those fixes, CRC enabled
> filesystems simple cannot be used without all these patches being
> applied. It's not an experimental feature at this point - it's just
If as an experimental feature "it's just broken" in -rc1, maybe I should
never have merged it.
> > We pulled in the CRC work with the understanding that it is an
> > experimental feature that might not be perfect, and with a focus
> > upon preventing regression for existing users. This did not imply
> > that we're going after perfection for CRCs during the 3.10 release
> > candidate series.
> Right, but there was also the understanding that the CRC code would
> be usable by release time, and so there would be fixes committed
> throughout the -rc series to fix problems with the CRC code so that
> it is usable.
> As it is, the result of all these patches is that xfstests is almost
> entirely passing on a 4k block size filesystem. The only tests that
> aren't passing are those that already fail or require userspace
> support that isn't currently available (e.g. xfs-db write support,
> metadump). And 1k block size filesystems mostly pass xfstests, too,
> though there are random assert failures from time to time.
> IOWs after this patch set is applied, the CRC functionality fits the
> definition of the 'experimental' tag. It mostly works, but still
> has some lurking issues that need to be flushed out. I'm unlikely to
> be posting 2-3 bug fix patches a day from this point onwards....
> Indeed, the last two days I've been doing 3.11 development work
> because xfstests on CRC enabled filesystems is now usable for
> regression testing. i.e. the majority of the "easy-to-hit" problems
> have been flushed out by this patchset.
> > I'll be happy to review the 17 CRC related patches for inclusion
> > in the master branch ASAP,
> That's a good start, but...
> > but I'm afraid 17 patches is a bit much
> > to ask for a release candidate, even if it were -rc2.
> ... I think you're being way too conservative here.
> BTRFS in january for an -rc merge:
> That was 30 commits and +300/-100 lines. There was a another pull
> request 2 weeks later with another 9 commits. The same thing
> happened with the RAID5/6 integration into btrfs in 3.9 - the -rc
> merges following it had 13, 7 and 15 commits. And for 3.10-rc, the
> first post-rc1 merge was 25 commits (+350/-250 lines).
You have announced publicly on several occasions that you feel BTRFS is
not ready for production use. If it is true that BTRFS is not ready for
production, then the concerns for BTRFS are different than those for XFS
in this context.
> And if you want another example, the ext4 extent status cache fixes
> that went into 3.9-rc4:
> that was 21 commits and +540/-81 lines.
Was extent status cache experimental code when this was merged?
> Yup, new filesystem code has bugs, and they get fixed in -rc
> releases. What makes new XFS code so special we can't do this?
> The amount of change we are talking about here is not unreasonable
> even for an -rc4 release, as evidenced by other filesystems under
> development. Hence I don't think "it's too late" is really a viable
> reason for not fixing all these filesystem corruption issues in the
> CRC code...
The 6-8 weeks after -rc1 is not the time to be changing the ondisk
format or rewriting swaths of code we didn't get right the first time,
unless absolutely necessary. Issues with an experimental feature that
are only relevant to the few developers and testers who have the
temerity to apply the userspace CRC patches do not merit the level of
risk which some of these patches represent to the much larger number of
XFS users who will not be using CRCs in 3.10.
> > Here we are in 3.11/feature-bit territory.
> I'd much prefer that we don't have to add code to 3.11 to reject any
> CRC-enabled filesystem without any feature bits set because we don't
> support a broken remote attr format that was fixed weeks before 3.10
> released but was not allowed to be fixed in 3.10. That's just crazy
> from any release management perspective you care to look at it from.
So would I.
> Ben, if the problem is that you can't review all the fixes in a timely
> manner, then we can fix that. I'm sure that Mark, Eric and Brian can
> help review the code if this is the sticking point.
Reviews are always welcome...
> There are also
> people like Michael who are actively testing the CRC code and
> reporting bugs to me all the time, so there's demand for these fixes
> to be integrated into the 3.10 tree...
I bet Michael can grab sources from the master branch.
> Everyone knew there would be fixes for the CRC code coming along in
> the -rc period - we discussed it on the weekly concall before the the
> CRC code was merged into 3.10. Hence I'm a bit surprised to get what
> appears to be a blanket rejection for these fixes to the CRC code this
> early in the 3.10-rc cycle.
Three (or is it 4 now?) on-disk changes, one of which is completely unrelated
to CRCs, along with a bunch of the changes deep in the remote attribute, quota,
and inode unlinked list code is not worth the risk in the middle of release
candidates. I'm ok with lower risk changes.
> Let's not waste all the effort that we put into getting the CRC code into
> 3.10 by not fixing known corruptions and on-disk format deficiencies.
I don't consider it to be a waste of effort if CRCs weren't quite perfect for
3.10. There is plenty to show for it, such as not having introduced
regressions for everyone else. People who want to test CRCs will be coming to
the list for the userspace patches anyway. It's easy enough to point them to
the master branch to get the kernel fixes too.
> Releasing CRC support in a seriously broken state that
> we need to add permanent workarounds for in 3.11 is about the worst
> possible outcome that could occur right now....
A worse outcome is that I pull in this code and something goes very
wrong for the thousands of users of 3.10 with existing non-crc XFS
filesystems. A feature bit and some inconvenience for a few XFS
developers and testers is a safer choice.