On Thu, Jun 13, 2024 at 10:56:39PM +0100, Theodore Ts'o wrote: > On Wed, Jun 12, 2024 at 12:41:36PM -0700, Darrick J. Wong wrote: > > > > I don't see this problem; if you apply this to fstests to turn off > > io_uring: > > https://lore.kernel.org/fstests/169335095953.3534600.16325849760213190849.stgit@frogsfrogsfrogs/#r > > > > do the problems go away? > > Thanks for pointing out the mail thread; I had a vague memory that > this had been raised as a problem before. Looking at the discussion > (from August 2023, so over 9 months ago), this is a bug that has been > acknowledged as an io_uring bug, but it still hasn't been fixed. > > Using Zorro's sugestion of adding "-f uring_read=0 -f uring_write=0" > to the fsstress options makes the EBUSY umount failures go away. I've > also created a new test which relaibly reproduces the "fsstress ; > umount" EBUSY bug (as opposed to the existing test failures which only > fail 1-10% of the time). So with that I can with a clean conscience > suggest that we omit io_uring calls from those tests using fsstress to > thest some non-io_uring related bug if they run into the umount EBUSY > bug, since there is now a new bug which reliably shows off the > problem.... Amusingly enough, I still have that patch (and generic/1220) in my fstests branch, and I haven't seen this problem happen on g/1220 in quite a while. --D > > - Ted
