On Tue, Jan 24, 2023 at 07:56:09AM -0500, Jeff Layton wrote: > A few months ago, I posted a patch to make xfs not bump its i_version > counter on atime updates. Dave Chinner NAK'ed that patch, mentioning > that xfs would need to replace it with an entirely new field as the > existing counter is used for other purposes and its semantics are set in > stone. > > Has anything been done toward that end? No, because we don't have official specification of the behaviour the nfsd subsystem requires merged into the kernel yet. > Should I file a bug report or something? There's nothing we can really do until the new specification is set in stone. Filing a bug report won't change anything material. As it is, I'm guessing that you desire the behaviour to be as you described in the iversion patchset you just posted. That is effectively: * The change attribute (i_version) is mandated by NFSv4 and is mostly for * knfsd, but is also used for other purposes (e.g. IMA). The i_version must - * appear different to observers if there was a change to the inode's data or - * metadata since it was last queried. + * appear larger to observers if there was an explicit change to the inode's + * data or metadata since it was last queried. i.e. the definition is changing from *any* metadata or data change to *explicit* metadata/data changes, right? i.e. it should only change when ctime changes? IIUC the rest of the justification for i_version is that ctime might lack the timestamp granularity to disambiguate sub-timestamp granularity changes, so i_version is needed to bridge that gap. Given that XFS has nanosecond timestamp resolution in the on-disk format, both i_version and ctime changes are journalled, and ctime/i_version will always change at exactly the same time in the same transactions, there are no inherent sub-timestamp granularity problems with ctime within XFS. Any deficiency in ctime resolution comes solely from the granularity of the VFS inode timestamp functions. And so if current_time() was to provide fine-grained nanosecond timestamp resolution for exported XFS filesystems (i.e. use ktime_get_real_ts64() conditionally), then it seems to me that the nfsd i_version function becomes completely redundant. i.e. we are pretty much guaranteed that ctime on exported filesystems will always be different for explicit modifications to the same inode, and hence we can just use ctime as the version change identifier without needing any on-disk format changes at all. And we can optimise away that overhead when the filesystem is not exported by just using the coarse timestamps because there is no need for sub-timer-tick disambiguation of single file modifications.... Hence it appears to me that with the new i_version specification that there's an avenue out of this problem entirely that is "nfsd needs to use ctime, not i_version". This solution seems generic enough that filesystems with existing on-disk nanosecond timestamp granularity would no longer need explicit on-disk support for the nfsd i_version functionality, yes? Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx
