On Sat, May 18, 2013 at 01:07:28PM +0800, Jeff Liu wrote:
> On 05/18/2013 11:25 AM, Dave Chinner wrote:
> > On Fri, May 17, 2013 at 04:54:47PM -0400, Michael L. Semon wrote:
> >> On 05/17/2013 07:12 AM, Dave Chinner wrote:
> >>> Hi Folks,
> >>> This is the first real "works ok" CRC patchset for xfsprogs. It
> >>> provides full support for mkfs.xfs and xfs_repair, and partial
> >>> read-only support for xfs_db.
> >>> For mkfs.xfs, it does everything properly, and filesystems that are
> >>> freshly made also run cleanly through xfs_repair and mount and run
> >>> just fine.
> >>> For xfs_repair, it reads and writes all metadata with CRC checks,
> >>> calculations and validation just like the kernel code does, but it
> >>> currently silently ignores the validation done in the IO layer.
> >>> Enabling that is future work - it involves adding buffer error checking to
> >>> every libxfs_readbuf() call that is made, and we do none of that
> >>> right now. It does, however, fully validate all the non-CRC format
> >>> metadata just as it does for non-CRC filesystems, and so the
> >>> coverage it has is the same for both CRC and non-CRC filesystems.
> >>> For xfs_db, there is read-only support for looking at the filesystem
> >>> as the xfs_db IO stack does not support CRCs at all. We need to
> >>> convert xfs_db to use the libxfs infrastructure to enable that.
> >>> Apart from that, xfs_db has partial support for the extended
> >>> metadata fields - the directory/attribute blocks don't have extended
> >>> support yet, but everything else does.
> >>> xfs_check is made special. It currently detects a version 5
> >>> superblock, and immediately exits with success. Hence it always says
> >>> CRC enabled filesystems are OK. This is a temporary change that
> >>> enables running xfstests without full support in xfs_db for all the
> >>> new metadata structures (like headers in remote symlink and
> >>> attribute blocks). Depending on if we want to keep xfs-check useful
> >>> for xfstests, we can revisit this bypass hack once xfs_db has been
> >>> converted to use the libxfs IO engine.
> >>> Overall, xfstests is now running enough to start to find bugs in the
> >>> kernel CRC code - I'm mainly hitting remote attribute block bugs
> >>> right now (generic/117!) but there's certainly less problems being
> >>> reported than I expected.
> >>> Oh, and I've tested it with external log devices and real time
> >>> devices, too.
> >>> Comments, thoughts, flames, and testing all welcome!
> >>> Cheers,
> >>> Dave.
> >> OK. The basics look good so far. The patchset applied without need
> >> for additional work with vi and patch. Whitespace errors were
> >> reported for Patches 8, 14, 16, 17, 24, 25, and 27. xfsprogs built
> >> with no additional errors over a normal xfsprogs build.
> > Can you send me the output indicating where the whitespace errors
> > are? I don't get any warnings from guilt about them when I apply the
> > patchset here...
> >> That all stated, the `tar -xvf qt-source.tar.xz` still fails on a
> >> CRC-enabled filesystem.
> > Not surprising - I haven't got a crc enabled filesystem all the way
> > through xfstests yet. remote attributes are the current piece I'm
> > working on getting fixed.
> >> Worse, until I return home, I won't be able
> >> to do serial-console capture of hard oopses. However, the initial
> >> oops I got was a soft one, so it is included after my closing. The
> >> kernel is this...
> >> last night's kernel git
> >> last night's xfs-oss/master
> >> some of your recent patches (didn't apply your 6_5 patch yet)
> >> J. Liu's most recent patchset + 2 older bitness patches
> >> Chandra's v8 pquota/gquota patchset + one E-mail fix
> >> Shaggy's JFS patch to make it through the old xfstests #068 on JFS
> >> an NILFS2 patch to address broken bmap handling, lurked from the
> >> NILFS2 mailing list
> >> one local removed assert to make it through the old xfstests #111
> >> maybe one or two XFS patches beyond this
> >> ...all on a 32-bit Pentium 4.
> > And reporting bugs :)
> >> What I'm trying to state is that a lot is in there, but the PC is
> >> spinning like a top, and xfstests results are really good right now.
> >> However, if I feel the need to provide a fresh environment, patch
> >> management is taking some time.
> > How are you managing patches right now? When taking in a new
> > patchset from a mailing list, I save them all in a mbox file,
> > then use git-am to apply them to a temporary git branch. I then move
> > to my real working branch, and do a 'guilt import-commit x..y' to
> > convert the commits in the temporary branch to a set of guilt
> > patches, and then go from there....
> > The worst step for me is, by far, the git-am step. Resolving patch
> > conflicts is painful because you have to manually apply the patch,
> > then remember to git add all the files modified by the patch, etc.
> > It'd be really cool if guilt could do the import directly from the
> > mbox file without applying the patches, so the normal guilt
> > force-push-fix-and-refresh method of solving patch conflicts could
> > be used instead of git-am.
> > /me wonders if #jeffpc is listening here....
> Ah? #jeffpc == me ? #jeffpc is up and listening... : just ignore;
No, #jeffpc is Josef Sipek. Author of guilt and many other useful
> Looks our test for 32-bit system is insufficient. There has another bug
> reports regarding 32-bit yesterday:
> So I'm going to setup a 32-bit test environment for such tests together
> with Michael.
Sounds good to me ;)