On Thu, Jul 26, 2018 at 10:44 AM, Miklos Szeredi <miklos@xxxxxxxxxx> wrote: > On Wed, Jul 25, 2018 at 11:12 AM, Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote: >> On Tue, Jul 24, 2018 at 5:17 PM, Miklos Szeredi <miklos@xxxxxxxxxx> wrote: >> Maybe more waits in fuse need to be interruptible? E.g. request_wait_answer? > > That's an interesting aspect. Making request_wait_answer always be > killable would help with the issue you raise (killing set of processes > taking part in deadlock should resolve deadlock), but it breaks > another aspect of the interface. > > Namely that userspace filesystems expect some serialization from > kernel when performing operations. If we allow killing of a process > in the middle of an fs operation, then that serialization is no longer > there, which can break the server. > > One solution to that is to duplicate all locking in the server > (libfuse normally), but it would not solve the issue for legacy > libfuse or legacy non-libfuse servers. It would also be difficult to > test. Also it doesn't solve the problem of killing the server, as > that alone doesn't resolve the deadlock. Umm, we can actually do better. Duplicate all vfs locking in the fuse kernel implementation: when killing a task that has an outstanding request, return immediately (which results in releasing the VFS level lock and hence the deadlock) but hold onto our own lock until the reply from the userspace server comes back. Need to think about the details; this might not be easy to do this properly. Notably memory management locks (page->lock, mmap_sem, etc) are notoriously tricky. Thanks, MIklos
