On Tue, 2023-10-24 at 09:40 -1000, Linus Torvalds wrote: > On Tue, 24 Oct 2023 at 09:07, Jeff Layton <jlayton@xxxxxxxxxx> wrote: > > > > The new flag idea is a good one. The catch though is that there are no > > readers of i_version in-kernel other than NFSD and IMA, so there would > > be no in-kernel users of I_VERSION_QUERIED_STRICT. > > I actually see that as an absolute positive. > > I think we should *conceptually* do those two flags, but then realize > that there are no users of the STRICT version, and just skip it. > > So practically speaking, we'd end up with just a weaker version of > I_VERSION_QUERIED that is that "I don't care about atime" case. > To be clear, this is not kernel-wide behavior. Most filesystems already don't bump their i_version on atime updates. XFS is the only one that does. ext4 used to do that too, but we fixed that several months ago. I did try to just fix XFS in the same way, but the patch was NAK'ed. > I really can't find any use that would *want* to see i_version updates > for any atime updates. Ever. > > We may have had historical user interfaces for i_version, but I can't > find any currently. > > But to be very very clear: I've only done some random grepping, and I > may have missed something. I'm not dismissing Dave's worries, and he > may well be entirely correct. > > Somebody would need to do a much more careful check than my "I can't > find anything". Exactly. I'm not really an XFS guy, so I took those folks at their word that this was behavior that they just can't trivially change. None of the in-kernel callers that look at i_version want it to be incremented on atime-onlt updates, however. So IIRC, the objection was due to offline repair/analysis tools that depend this the value being incremented in a specific way. -- Jeff Layton <jlayton@xxxxxxxxxx>
