On Wed, Nov 27, 2024 at 10:16 PM Jens Axboe <axboe@xxxxxxxxx> wrote: > On 11/27/24 2:08 PM, Kent Overstreet wrote: > > On Wed, Nov 27, 2024 at 09:44:21PM +0100, Jann Horn wrote: > >> On Wed, Nov 27, 2024 at 9:25?PM Kent Overstreet > >> <kent.overstreet@xxxxxxxxx> wrote: > >>> On Wed, Nov 27, 2024 at 11:09:14AM -0700, Jens Axboe wrote: > >>>> On 11/27/24 9:57 AM, Jann Horn wrote: > >>>>> Hi! > >>>>> > >>>>> In fs/bcachefs/fs-io-direct.c, "struct dio_write" contains a pointer > >>>>> to an mm_struct. This pointer is grabbed in bch2_direct_write() > >>>>> (without any kind of refcount increment), and used in > >>>>> bch2_dio_write_continue() for kthread_use_mm()/kthread_unuse_mm() > >>>>> which are used to enable userspace memory access from kthread context. > >>>>> I believe kthread_use_mm()/kthread_unuse_mm() require that the caller > >>>>> guarantees that the MM hasn't gone through exit_mmap() yet (normally > >>>>> by holding an mmget() reference). > >>>>> > >>>>> If we reach this codepath via io_uring, do we have a guarantee that > >>>>> the mm_struct that called bch2_direct_write() is still alive and > >>>>> hasn't yet gone through exit_mmap() when it is accessed from > >>>>> bch2_dio_write_continue()? > >>>>> > >>>>> I don't know the async direct I/O codepath particularly well, so I > >>>>> cc'ed the uring maintainers, who probably know this better than me. > >>>> > >>>> I _think_ this is fine as-is, even if it does look dubious and bcachefs > >>>> arguably should grab an mm ref for this just for safety to avoid future > >>>> problems. The reason is that bcachefs doesn't set FMODE_NOWAIT, which > >>>> means that on the io_uring side it cannot do non-blocking issue of > >>>> requests. This is slower as it always punts to an io-wq thread, which > >>>> shares the same mm. Hence if the request is alive, there's always a > >>>> thread with the same mm alive as well. > >>>> > >>>> Now if FMODE_NOWAIT was set, then the original task could exit. I'd need > >>>> to dig a bit deeper to verify that would always be safe and there's not > >>>> a of time today with a few days off in the US looming, so I'll defer > >>>> that to next week. It certainly would be fine with an mm ref grabbed. > >>> > >>> Wouldn't delivery of completions be tied to an address space (not a > >>> process) like it is for aio? > >> > >> An io_uring instance is primarily exposed to userspace as a file > >> descriptor, so AFAIK it is possible for the io_uring instance to live > >> beyond when the last mmput() happens. io_uring initially only holds an > >> mmgrab() reference on the MM (a comment above that explains: "This is > >> just grabbed for accounting purposes"), which I think is not enough to > >> make it stable enough for kthread_use_mm(); I think in io_uring, only > >> the worker threads actually keep the MM fully alive (and AFAIK the > >> uring worker threads can exit before the uring instance itself is torn > >> down). > >> > >> To receive io_uring completions, there are multiple ways, but they > >> don't use a pointer from the io_uring instance to the MM to access > >> userspace memory. Instead, you can have a VMA that points to the > >> io_uring instance, created by calling mmap() on the io_uring fd; or > >> alternatively, with IORING_SETUP_NO_MMAP, you can have io_uring grab > >> references to userspace-provided pages. > >> > >> On top of that, I think it might currently be possible to use the > >> io_uring file descriptor from another task to submit work. (That would > >> probably be fairly nonsensical, but I think the kernel does not > >> currently prevent it.) > > > > Ok, that's a wrinkle. > > I'd argue the fact that you are using an mm from a different process > without grabbing a reference is the wrinkle. I just don't think it's a > problem right now, but it could be... aio is tied to the mm because of > how it does completions, potentially, and hence needs this exit_aio() > hack because of that. aio also doesn't care, because it doesn't care > about blocking - it'll happily block during issue. > > > Jens, is it really FMODE_NOWAIT that controls whether we can hit this? A > > very cursory glance leads me to suspect "no", it seems like this is a > > bug if io_uring is allowed on bcachefs at all. > > I just looked at bcachefs dio writes, which look to be the only case of > this. And yes, for writes, if FMODE_NOWAIT isn't set, then io-wq is > always involved for the IO. I guess it could be an issue if the iocb can outlive the io-wq thread? Like, a userspace task creates a uring instance and starts a write; the write will be punted to a uring worker because of missing FMODE_NOWAIT; then the uring worker enters io_write() and starts a write on a kiocb. Can this write initiated from the worker be async? And could the uring worker exit (for example, because the userspace task exited) while the kiocb is still in flight? Anyway, I see Kent just posted a patch to make this more robust.
