On Wed, 2022-08-24 at 09:42 +1000, NeilBrown wrote: > On Wed, 24 Aug 2022, Dave Chinner wrote: > > On Wed, Aug 24, 2022 at 08:24:47AM +1000, NeilBrown wrote: > > > On Tue, 23 Aug 2022, Jeff Layton wrote: > > > > On Tue, 2022-08-23 at 21:38 +1000, NeilBrown wrote: > > > > > On Tue, 23 Aug 2022, Jeff Layton wrote: > > > > > > So, we can refer to that and simply say: > > > > > > > > > > > > "If the function updates the mtime or ctime on the inode, > > > > > > then the > > > > > > i_version should be incremented. If only the atime is being > > > > > > updated, > > > > > > then the i_version should not be incremented. The exception > > > > > > to this rule > > > > > > is explicit atime updates via utimes() or similar > > > > > > mechanism, which > > > > > > should result in the i_version being incremented." > > > > > > > > > > Is that exception needed? utimes() updates ctime. > > > > > > > > > > https://man7.org/linux/man-pages/man2/utimes.2.html > > > > > > > > > > doesn't say that, but > > > > > > > > > > https://pubs.opengroup.org/onlinepubs/007904875/functions/utimes.html > > > > > > > > > > does, as does the code. > > > > > > > > > > > > > Oh, good point! I think we can leave that out. Even better! > > > > > > Further, implicit mtime updates (file_update_time()) also update > > > ctime. > > > So all you need is > > > If the function updates the ctime, then i_version should be > > > incremented. > > > > > > and I have to ask - why not just use the ctime? Why have another > > > number > > > that is parallel? > > > > > > Timestamps are updated at HZ (ktime_get_course) which is at most > > > every > > > millisecond. > > > > Kernel time, and therefore timestamps, can go backwards. > > Yes, and when that happens you get to keep both halves... > > For NFSv4 I really don't think that matters. If it happened every > day, > that might be a problem. Even if it happens as a consequence of > normal > operations it might be a problem. But it can only happen if > something > goes wrong. > > Mostly, NFSv4 only needs to changeid to change. If the kernel time > goes > backwards it is possible that a changeid will repeat, though > unlikely. > It is even possible that a client will see the first and second > instances of that repeat, and assume there is no change in between - > but > that is astronomically unlikely. "touch"ing the file or remounting > will > fix that. > > When a write delegation is in force (which Linux doesn't currently > offer > and no-one seems to care about, but maybe one day), the client is > allowed to update the changeid, and when the delegation is returned, > the > server is supposed to ensure the new changeid is at least the last > one > assigned by the client. This is the only reason that it is defined > as > being monotonic (rather than just "non-repeating") - so the client > and > server can change it in the same way. > Sort of... Monotonicity of the change id is not a requirement, even for that case. The exact NFSv4 requirement is that if the server uses a callback to ask the client for the file attribute information, then the client is supposed to bump the server-supplied change id value by 1 unit if it is caching writes. The intention is simply to ensure that there is a notification mechanism to allow the server to know that writes are being cached (Note: I've no idea why we didn't just add a separate flag for that in the callback reply). > So while kernel time going backwards is theoretically less than > ideal, > it is not practically a problem. > I agree with that. As long as this results in few collisions, so that value uniqueness is guaranteed then applications (NFS or other) will have a way to determine if the filesystem object has changed since they last looked at it. IOW: I'm not saying this is the perfect way to implement an application-visible change attribute, but it would be a massive improvement over ctime, and might be a practical way to go about it. -- Trond Myklebust Linux NFS client maintainer, Hammerspace trond.myklebust@xxxxxxxxxxxxxxx
