On Thu, Mar 29, 2018 at 10:53:05AM +0200, Christoph Hellwig wrote: > On Wed, Mar 28, 2018 at 05:35:26PM +0100, Al Viro wrote: > > > ret = vfs_fsync(req->file, req->datasync); > > > - fput(req->file); > > > - aio_complete(container_of(req, struct aio_kiocb, fsync), ret, 0); > > > + if (aio_complete(iocb, ret, 0, 0)) > > > + fput(file); > > > > IDGI. > > 1) can aio_complete() ever return false here? > > It won't. But sometimes checking the return value and sometimes not > seems like a bad pattern. > > > 2) do we ever have aio_kiocb that would not have an associated > > struct file * that needs to be dropped on successful aio_complete()? > > AFAICS, > > rw, fsync and poll variants all have one, and I'm not sure what kind of > > async IO *could* be done without an opened file. > > All have a file assoiated at least right now. As mentioned last time > finding a struct to pass that file would be rather annoying, so we'd either > have to pass it explicitly, or do something nasty like duplicating the > pointer in the aio_kiocb in addition to struct kiocb. Which might not > be that bad after all, as it would only bloat the aio_kiocb and not > struct kiocb used on stack all over.
OK. Let's leave that alone for now. Re deferred cancels - AFAICS, we *must* remove the sucker from ctx->active_reqs before dropping ->ctx_lock. As it is, you are creating a io_cancel()/io_cancel() race leading to double fput(). It's not that hard to fix; I can do that myself while applying your series (as described in previous posting - kiocb_cancel_locked() returning NULL or ERR_PTR() in non-deferred case and pointer to aio_kiocb removed from ->active_reqs in deferred one) or you could fix it in some other way and update your branch. As it is, the race is user-exploitable and not that hard to trigger - AIO_POLL, then have two threads try and cancel it at the same time.
