On Wed, Dec 21, 2016 at 12:03:17PM -0500, Jeff Layton wrote: > tl;dr: I think we can greatly reduce the cost of the inode->i_version > counter, by exploiting the fact that we don't need to increment it > if no one is looking at it. We can also clean up the code to prepare > to eventually expose this value via statx(). > > The inode->i_version field is supposed to be a value that changes > whenever there is any data or metadata change to the inode. Some > filesystems use it internally to detect directory changes during > readdir. knfsd will use it if the filesystem has MS_I_VERSION > set. IMA will also use it (though it's not clear to me that that > works 100% -- no check for MS_I_VERSION there). I'm a little confused about the internal uses for directories. Is it necessarily kept on disk? In cases it's not, then there's not the same performance issue, right? Is there any risk these patches make performance slightly worse in that case? > Only btrfs, ext4, and xfs implement it for data changes. Because of > this, these filesystems must log the inode to disk whenever the > i_version counter changes. On those filesystems that's done for both files and directories, right? --b. > That has a non-zero performance impact, > especially on write-heavy workloads, because we end up dirtying the > inode metadata on every write, not just when the times change. [1] > > It turns out though that none of these users of i_version require that > i_version change on every change to the file. The only real requirement > is that it be different if _something_ changed since the last time we > queried for it. [2] > > So, if we simply keep track of when something queries the value, we > can avoid bumping the counter and that metadata update. > > This patchset implements this: > > It starts with some small cleanup patches to just remove any mention of > the i_version counter in filesystems that don't actually use it. > > Then, we add a new set of static inlines for managing the counter. The > initial version should work identically to what we have now. Then, all > of the remaining filesystems that use i_version are converted to the new > inlines. > > Once that's in place, we switch to a new implementation that allows us > to track readers of i_version counter, and only bump it when it's > necessary or convenient (i.e. we're going to disk anyway). > > The final patch switches from a scheme that uses the i_lock to serialize > the counter updates during write to an atomic64_t. That's a wash > performance-wise in my testing, but I like not having to take the i_lock > down where it could end up nested inside other locks. > > With this, we reduce inode metadata updates across all 3 filesystems > down to roughly the frequency of the timestamp granularity, particularly > when it's not being queried (the vastly common case). > > The pessimal workload here is 1 byte writes, and it helps that > significantly. Of course, that's not a real-world workload. > > A tiobench-example.fio workload also shows some modest performance > gains, and I've gotten mails from the kernel test robot that show some > significant performance gains on some microbenchmarks (case-msync-mt in > the vm-scalability testsuite to be specific). > > I'm happy to run other workloads if anyone can suggest them. > > At this point, the patchset works and does what it's expected to do in > my own testing. It seems like it's at least a modest performance win > across all 3 major disk-based filesystems. It may also encourage others > to implement i_version as well since it reduces that cost. > > Is this an avenue that's worthwhile to pursue? > > Note that I think we may have other changes coming in the future that > will make this sort of cleanup necessary anyway. I'd like to plug in the > Ceph change attribute here eventually, and that will require something > like this anyway. > > Thoughts, comments and suggestions are welcome... > > --- > > [1]: On ext4 it must be turned on with the i_version mount option, > mostly due to fears of incurring this impact, AFAICT. > > [2]: NFS also recommends that it appear to increase in value over time, so > that clients can discard metadata updates that are older than ones > they've already seen. > > Jeff Layton (30): > lustre: don't set f_version in ll_readdir > ecryptfs: remove unnecessary i_version bump > ceph: remove the bump of i_version > f2fs: don't bother setting i_version > hpfs: don't bother with the i_version counter > jfs: remove initialization of i_version counter > nilfs2: remove inode->i_version initialization > orangefs: remove initialization of i_version > reiserfs: remove unneeded i_version bump > ntfs: remove i_version handling > fs: new API for handling i_version > fat: convert to new i_version API > affs: convert to new i_version API > afs: convert to new i_version API > btrfs: convert to new i_version API > exofs: switch to new i_version API > ext2: convert to new i_version API > ext4: convert to new i_version API > nfs: convert to new i_version API > nfsd: convert to new i_version API > ocfs2: convert to new i_version API > ufs: use new i_version API > xfs: convert to new i_version API > IMA: switch IMA over to new i_version API > fs: add a "force" parameter to inode_inc_iversion > fs: only set S_VERSION when updating times if it has been queried > xfs: avoid setting XFS_ILOG_CORE if i_version doesn't need > incrementing > btrfs: only dirty the inode in btrfs_update_time if something was > changed > fs: track whether the i_version has been queried with an i_state flag > fs: convert i_version counter over to an atomic64_t > > drivers/staging/lustre/lustre/llite/dir.c | 3 - > fs/affs/amigaffs.c | 4 +- > fs/affs/dir.c | 4 +- > fs/affs/super.c | 2 +- > fs/afs/fsclient.c | 2 +- > fs/afs/inode.c | 4 +- > fs/btrfs/delayed-inode.c | 4 +- > fs/btrfs/file.c | 4 +- > fs/btrfs/inode.c | 41 ++++---- > fs/btrfs/ioctl.c | 4 +- > fs/btrfs/tree-log.c | 2 +- > fs/btrfs/xattr.c | 2 +- > fs/ceph/inode.c | 1 - > fs/ecryptfs/inode.c | 1 - > fs/exofs/dir.c | 8 +- > fs/exofs/super.c | 2 +- > fs/ext2/dir.c | 8 +- > fs/ext2/super.c | 4 +- > fs/ext4/dir.c | 8 +- > fs/ext4/inline.c | 6 +- > fs/ext4/inode.c | 16 ++-- > fs/ext4/ioctl.c | 2 +- > fs/ext4/namei.c | 8 +- > fs/ext4/super.c | 2 +- > fs/f2fs/super.c | 1 - > fs/fat/dir.c | 2 +- > fs/fat/inode.c | 8 +- > fs/fat/namei_msdos.c | 6 +- > fs/fat/namei_vfat.c | 20 ++-- > fs/hpfs/dir.c | 1 - > fs/hpfs/dnode.c | 2 - > fs/hpfs/super.c | 1 - > fs/inode.c | 9 +- > fs/jfs/super.c | 1 - > fs/nfs/delegation.c | 2 +- > fs/nfs/fscache-index.c | 4 +- > fs/nfs/inode.c | 16 ++-- > fs/nfs/nfs4proc.c | 4 +- > fs/nfs/nfstrace.h | 4 +- > fs/nfs/write.c | 2 +- > fs/nfsd/nfs3xdr.c | 2 +- > fs/nfsd/nfs4xdr.c | 2 +- > fs/nfsd/nfsfh.h | 2 +- > fs/nilfs2/super.c | 1 - > fs/ntfs/inode.c | 9 -- > fs/ntfs/mft.c | 6 -- > fs/ocfs2/dir.c | 14 +-- > fs/ocfs2/inode.c | 2 +- > fs/ocfs2/namei.c | 2 +- > fs/ocfs2/quota_global.c | 2 +- > fs/orangefs/super.c | 2 - > fs/reiserfs/super.c | 1 - > fs/ufs/dir.c | 8 +- > fs/ufs/inode.c | 2 +- > fs/ufs/super.c | 2 +- > fs/xfs/libxfs/xfs_inode_buf.c | 4 +- > fs/xfs/xfs_icache.c | 4 +- > fs/xfs/xfs_inode.c | 2 +- > fs/xfs/xfs_inode_item.c | 2 +- > fs/xfs/xfs_trans_inode.c | 12 +-- > include/linux/fs.h | 151 ++++++++++++++++++++++++++++-- > security/integrity/ima/ima_api.c | 2 +- > security/integrity/ima/ima_main.c | 2 +- > 63 files changed, 288 insertions(+), 173 deletions(-) > > -- > 2.7.4 > > -- > To unsubscribe from this list: send the line "unsubscribe linux-nfs" in > the body of a message to majordomo@xxxxxxxxxxxxxxx > More majordomo info at http://vger.kernel.org/majordomo-info.html
