On Fri, 2023-10-20 at 13:06 -0700, Linus Torvalds wrote: > On Fri, 20 Oct 2023 at 05:12, Jeff Layton <jlayton@xxxxxxxxxx> wrote:. > > > > I'd _really_ like to see a proper change counter added before it's > > merged, or at least space in the on-disk inode reserved for one until we > > can get it plumbed in. > > Hmm. Can we not perhaps just do an in-memory change counter, and try > to initialize it to a random value when instantiating an inode? Do we > even *require* on-disk format changes? > > So on reboot, the inode would count as "changed" as far any remote > user is concerned. It would flush client caches, but isn't that what > you'd want anyway? I'd hate to waste lots of memory, but maybe people > would be ok with just a 32-bit random value. And if not... > > But I actually came into this whole discussion purely through the > inode timestamp side, so I may *entirely* miss what the change counter > requirements for NFSd actually are. If it needs to be stable across > reboots, my idea is clearly complete garbage. > > You can now all jump on me and point out my severe intellectual > limitations. Please use small words when you do ;) > Much like inode timestamps, we do depend on the change attribute persisting across reboots. Having to invalidate all of your cached data just because the server rebooted is particularly awful. That usually results in the server being hammered with reads from all of the clients at once, soon after rebooting. -- Jeff Layton <jlayton@xxxxxxxxxx>
