> On Jan 8, 2021, at 5:54 PM, Trond Myklebust <trondmy@xxxxxxxxxxxxxxx> wrote: > > On Fri, 2021-01-08 at 13:08 -0500, bfields@xxxxxxxxxxxx wrote: >> On Fri, Jan 08, 2021 at 04:35:50PM +0000, Trond Myklebust wrote: >>> Just ignore generic/465. As far as NFS is concerned, the test has >>> utterly borked assumptions about O_DIRECT ordering. >> >> Thanks, adding to my list of tests to skip. Should we report it as >> an >> xfstests bug? >> >> (Is the test just wrong, or is this some non-standard but documented >> NFS >> behavior, or something else?) >> >> --b. > > I'm not sure who decided the ordering requirements for O_DIRECT, but in > order to fix the generic/465 case, I'd either have to order all reads > with all outstanding writes or implement some kind of range locking to > do it in a more fine-grained way. > > We do order buffered I/O and O_DIRECT, so that backup programs can do > their thing on databases that use O_DIRECT. However we do assume that > anyone using O_DIRECT for I/O is doing their own synchronisation. Perhaps the best approach would be to add generic/465 to the exempt-list for NFS. -- Chuck Lever
