On Tue, 2022-03-15 at 09:06 +1100, Dave Chinner wrote: > This is a patchset inspired by the performance regressions that were > seen from logging 64k xattrs with Allison's delayed attribute > patchset and trying to work out how to minimise the impact of > logging xattrs. Most of that is explained in the "xfs: intent item > whiteouts" patch, so I won't repeat it here. > > The whiteouts massively reduce the journal write overhead of logging > xattrs - with this patchset I've reduced 2.5GB/s of log traffic (16 > way file create w/64k xattr workload) down to approximately 220MB of > log traffic, and performance has increased from 9k creates/s to 36k > creates/s. The workload still writes to disk at 2.5GB/s, but that's > what writing 35k x 64k xattrs to disk does. > > This is still short of the non-logged attribute mechanism, which > runs at 40-45k creates a second and 3.5-4GB/s to disk, but it brings > logged attrs to within roughly 5-15% of non-logged attrs across the > full range of attribute sizes. > > So, while this patchset was clearly insired and has major positive > impact on Allison's delayed attribute work, it also applies > generically to all other intent/intent done pairs that already > exist. Hence I've created this patchset as a stand-alone patchset > that isn't dependent on the delayed attributes being committed, nor > does the delayed attribute patchset need this to function properly. > IOWs, they can be merged in parallel and then the attribute log item > implementation be updated to support whiteouts after the fact. > > This patchset is separate to the attr code, though, because > intent whiteouts are not specific to the attr code. They are a > generic mechanism that can be applied to all the intent/intent done > item pairs we already have. This patch set modifies all those > intents to use whiteouts, and so there is benefits from the patch > set for all operations that use these intents. > > With respect to the delayed attribute patchset, it can be merged > without whiteout support and still work correctly with/without this > patchset in place. Once both intent whiteouts and delayed attrs are > merged, we can add whiteout support to delayed attributes with only > a few lines of extra code. > I applied this set to the for-next tag: 01728b44ef1b (tag: xfs-5.18-merge-2), and ran into a few merge conflicts. I just went through and fixed them to try and get things working, but I'm getting some failed assert checks on generic/466: [ 155.219780] XFS: Assertion failed: reg->i_len % sizeof(int32_t) == 0, file: fs/xfs/xfs_log.c, line: 2536 So maybe still some bugglies to work out. I'll go over the conflicts in the review. Also, I'll push out the branch I have so far. Since I've already gone through the rebase effort, I may as well share it: https://github.com/allisonhenderson/xfs_work/tree/whiteouts Allison > Changelog: > > Version 3: > - rebased on 5.17-rc4 + xlog-write-rework > - no longer dependent on xfs-cil-scalability, so there's some porting > changes > that was needed to remove all the per-cpu CIL dependencies. > > Version 2: > - not published > - rebased on 5.15-rc2 + xfs-cil-scalability > - dropped the kvmalloc changes for CIL shadow buffers as that's a > separate perf problem and not something related to intent > whiteouts. > - dropped all the delayed attribute modifications so that the > patchset is not dependent on Allison's dev tree. > - Thanks to Allison for an initial quick review pass - I haven't > included those RVB tags because every patch in the series has > changed since the original RFC posting. > > RFC: > - > https://lore.kernel.org/linux-xfs/20210909212133.GE2361455@xxxxxxxxxxxxxxxxxxx/ >
