On Sat, 27 Aug 2022, Trond Myklebust wrote: > On Fri, 2022-08-26 at 10:59 -0400, Benjamin Coddington wrote: > > On 16 May 2022, at 21:36, Trond Myklebust wrote: > > > So until you have a different solution that doesn't impact the > > > client's > > > ability to cache permissions, then the answer is going to be "no" > > > to > > > these patches. > > > > Hi Trond, > > > > We have some folks negatively impacted by this issue as well. Are > > you > > willing to consider this via a mount option? > > > > Ben > > > > I don't see how that answers my concern. Could you please spell out again what your concerns are? I still don't understand. The only performance impact is when a permission test fails. In what circumstance is permission failure expected on a fast-path? > > I'd rather see us set up an explicit trigger mechanism. It doesn't have > to be particularly sophisticated. I can imagine just having a global, > or more likely a per-container, cookie that has a control mechanism in > /sys/fs/nfs, and that can be used to order all the inodes to invalidate > their permissions caches when you believe there is a need to do so. I hope it would only invalidate negative cached permissions, not positive. Caches positive permissions aren't really a problem as we'll find out they were wrong as soon as we send the relevant request to the server. The problem with cached negative permissions is that we never even try to send a request to the server. The client doesn't *know* when the server changes it's understanding of group membership, so it cannot know when to write to this. So the best the client can do is invalidate negative cached permissions periodically. So if this /sys/fs/nfs/ tunable were to be added, I would like it to be a time interval after which they can expire (I would set it to zero of course). Thanks, NeilBrown > > i.e. you cache the value of the global cookie in the inode, and if you > notice a change, then that's the signal that you need to invalidate the > permissions cache before updating the cached value of the cookie. > > That way, you have a mechanism that serves all purposes: it can do an > immediate one-time only flush, or you can set up a userspace job that > issues a global flush once every so often, e.g. using a cron job. > > -- > Trond Myklebust > Linux NFS client maintainer, Hammerspace > trond.myklebust@xxxxxxxxxxxxxxx > > >
