On Tue, Jun 27, 2023 at 09:16:31PM -0600, Jens Axboe wrote: > On 6/27/23 2:15?PM, Kent Overstreet wrote: > >> to ktest/tests/xfstests/ and run it with -bcachefs, otherwise it kept > >> failing because it assumed it was XFS. > >> > >> I suspected this was just a timing issue, and it looks like that's > >> exactly what it is. Looking at the test case, it'll randomly kill -9 > >> fsstress, and if that happens while we have io_uring IO pending, then we > >> process completions inline (for a PF_EXITING current). This means they > >> get pushed to fallback work, which runs out of line. If we hit that case > >> AND the timing is such that it hasn't been processed yet, we'll still be > >> holding a file reference under the mount point and umount will -EBUSY > >> fail. > >> > >> As far as I can tell, this can happen with aio as well, it's just harder > >> to hit. If the fput happens while the task is exiting, then fput will > >> end up being delayed through a workqueue as well. The test case assumes > >> that once it's reaped the exit of the killed task that all files are > >> released, which isn't necessarily true if they are done out-of-line. > > > > Yeah, I traced it through to the delayed fput code as well. > > > > I'm not sure delayed fput is responsible here; what I learned when I was > > tracking this down has mostly fell out of my brain, so take anything I > > say with a large grain of salt. But I believe I tested with delayed_fput > > completely disabled, and found another thing in io_uring with the same > > effect as delayed_fput that wasn't being flushed. > > I'm not saying it's delayed_fput(), I'm saying it's the delayed putting > io_uring can end up doing. But yes, delayed_fput() is another candidate. Sorry - was just working through my recollections/initial thought process out loud > >> For io_uring specifically, it may make sense to wait on the fallback > >> work. The below patch does this, and should fix the issue. But I'm not > >> fully convinced that this is really needed, as I do think this can > >> happen without io_uring as well. It just doesn't right now as the test > >> does buffered IO, and aio will be fully sync with buffered IO. That > >> means there's either no gap where aio will hit it without O_DIRECT, or > >> it's just small enough that it hasn't been hit. > > > > I just tried your patch and I still have generic/388 failing - it > > might've taken a bit longer to pop this time. > > Yep see the same here. Didn't have time to look into it after sending > that email today, just took a quick stab at writing a reproducer and > ended up crashing bcachefs: You must have hit an error before we finished initializing the filesystem, the list head never got initialized. Patch for that will be in the testing branch momentarily. > > I wonder if there might be a better way of solving this though? For aio, > > when a process is exiting we just synchronously tear down the ioctx, > > including waiting for outstanding iocbs. > > aio is pretty trivial, because the only async it supports is O_DIRECT > on regular files which always completes in finite time. io_uring has to > cancel etc, so we need to do a lot more. ahh yes, buffered IO would complicate things > But the concept of my patch should be fine, but I think we must be > missing a case. Which is why I started writing a small reproducer > instead. I'll pick it up again tomorrow and see what is going on here. Ok. Soon as you've got a patch I'll throw it at my CI, or I can point my CI at your branch if you have one.
