Re: [PATCH V3] io_uring: fix IO hang in io_wq_put_and_exit from do_exit()

[Stefan Hajnoczi <stefanha@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 7756 bytes --]

On Tue, Sep 26, 2023 at 09:28:15AM +0800, Ming Lei wrote:
> On Mon, Sep 25, 2023 at 05:17:10PM -0400, Stefan Hajnoczi wrote:
> > On Fri, Sep 15, 2023 at 03:04:05PM +0800, Jason Wang wrote:
> > > On Fri, Sep 8, 2023 at 11:25 PM Ming Lei <ming.lei@redhat.com> wrote:
> > > >
> > > > On Fri, Sep 08, 2023 at 08:44:45AM -0600, Jens Axboe wrote:
> > > > > On 9/8/23 8:34 AM, Ming Lei wrote:
> > > > > > On Fri, Sep 08, 2023 at 07:49:53AM -0600, Jens Axboe wrote:
> > > > > >> On 9/8/23 3:30 AM, Ming Lei wrote:
> > > > > >>> diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
> > > > > >>> index ad636954abae..95a3d31a1ef1 100644
> > > > > >>> --- a/io_uring/io_uring.c
> > > > > >>> +++ b/io_uring/io_uring.c
> > > > > >>> @@ -1930,6 +1930,10 @@ void io_wq_submit_work(struct io_wq_work *work)
> > > > > >>>           }
> > > > > >>>   }
> > > > > >>>
> > > > > >>> + /* It is fragile to block POLLED IO, so switch to NON_BLOCK */
> > > > > >>> + if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
> > > > > >>> +         issue_flags |= IO_URING_F_NONBLOCK;
> > > > > >>> +
> > > > > >>
> > > > > >> I think this comment deserves to be more descriptive. Normally we
> > > > > >> absolutely cannot block for polled IO, it's only OK here because io-wq
> > > > > >
> > > > > > Yeah, we don't do that until commit 2bc057692599 ("block: don't make REQ_POLLED
> > > > > > imply REQ_NOWAIT") which actually push the responsibility/risk up to
> > > > > > io_uring.
> > > > > >
> > > > > >> is the issuer and not necessarily the poller of it. That generally falls
> > > > > >> upon the original issuer to poll these requests.
> > > > > >>
> > > > > >> I think this should be a separate commit, coming before the main fix
> > > > > >> which is below.
> > > > > >
> > > > > > Looks fine, actually IO_URING_F_NONBLOCK change isn't a must, and the
> > > > > > approach in V2 doesn't need this change.
> > > > > >
> > > > > >>
> > > > > >>> @@ -3363,6 +3367,12 @@ __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
> > > > > >>>           finish_wait(&tctx->wait, &wait);
> > > > > >>>   } while (1);
> > > > > >>>
> > > > > >>> + /*
> > > > > >>> +  * Reap events from each ctx, otherwise these requests may take
> > > > > >>> +  * resources and prevent other contexts from being moved on.
> > > > > >>> +  */
> > > > > >>> + xa_for_each(&tctx->xa, index, node)
> > > > > >>> +         io_iopoll_try_reap_events(node->ctx);
> > > > > >>
> > > > > >> The main issue here is that if someone isn't polling for them, then we
> > > > > >
> > > > > > That is actually what this patch is addressing, :-)
> > > > >
> > > > > Right, that part is obvious :)
> > > > >
> > > > > >> get to wait for a timeout before they complete. This can delay exit, for
> > > > > >> example, as we're now just waiting 30 seconds (or whatever the timeout
> > > > > >> is on the underlying device) for them to get timed out before exit can
> > > > > >> finish.
> > > > > >
> > > > > > For the issue on null_blk, device timeout handler provides
> > > > > > forward-progress, such as requests are released, so new IO can be
> > > > > > handled.
> > > > > >
> > > > > > However, not all devices support timeout, such as virtio device.
> > > > >
> > > > > That's a bug in the driver, you cannot sanely support polled IO and not
> > > > > be able to deal with timeouts. Someone HAS to reap the requests and
> > > > > there are only two things that can do that - the application doing the
> > > > > polled IO, or if that doesn't happen, a timeout.
> > > >
> > > > OK, then device driver timeout handler has new responsibility of covering
> > > > userspace accident, :-)
> > 
> > Sorry, I don't have enough context so this is probably a silly question:
> > 
> > When an application doesn't reap a polled request, why doesn't the block
> > layer take care of this in a generic way? I don't see anything
> > driver-specific about this.
> 
> block layer doesn't have knowledge to handle that, io_uring knows the
> application is exiting, and can help to reap the events.

I thought the discussion was about I/O timeouts in general but here
you're only mentioning application exit. Are we talking about I/O
timeouts or purely about cleaning up I/O requests when an application
exits?

> 
> But the big question is that if there is really IO timeout for virtio-blk.
> If there is, the reap done in io_uring may never return and cause other
> issue, so if it is done in io_uring, that can be just thought as sort of
> improvement.

virtio-blk drivers have no way of specifying timeouts on the device or
aborting/canceling requests.

virtio-blk devices may fail requests if they implement an internal
timeout mechanism (e.g. the host kernel fails requests after a host
timeout), but this is not controlled by the driver and there is no
guarantee that the device has an internal timeout. The driver will not
treat these timed out requests in a special way - the application will
see EIO errors.

> 
> The real bug fix is still in device driver, usually only the driver timeout
> handler can provide forward progress guarantee.

The only recourse for hung I/O on a virtio-blk device is device reset,
but that is often implemented as a synchronous operation and is likely
to block until in-flight I/O finishes.

An admin virtqueue could be added to virtio-blk along with an abort
command, but existing devices will not support the new hardware
interface.

However, I'm not sure a new abort command would solve the problem.
virtio-blk devices are often implemented as userspace processes and are
limited by the availability of I/O cancellation APIs. Maybe my
understanding is outdated, but I believe userspace processes cannot
force I/O to abort. For example, the man page says the following for
IORING_OP_ASYNC_CANCEL:

  In general, requests that are interruptible (like socket IO) will get
  canceled, while disk IO requests cannot be canceled if already
  started.

Even if an abort command is added to virtio-blk, won't we just end up in
this situation:
1. The guest kernel invokes ->timeout() on virtio_blk.ko.
2. virtio_blk.ko sends an abort command to the device and resets the
   timeout.
3. The device submits IORING_OP_ASYNC_CANCEL but it cannot cancel an
   in-flight disk I/O request.
4. ...time passes...
5. The guest kernel invokes ->timeout() again and virtio_blk.ko decides
   abort was ineffective. The entire device must be reset.
?

(I based this on the ->timeout() logic in the nvme driver.)

If we're effectively just going to wait for twice the timeout duration
and then reset the device, then why go through the trouble of sending
the abort command? I'm hoping you'll tell me that IORING_OP_ASYNC_CANCEL
is in fact able to cancel disk I/O nowadays :).

> 
> > 
> > Driver-specific behavior would be sending an abort/cancel upon timeout.
> > virtio-blk cannot do that because there is no such command in the device
> > specification at the moment. So simply waiting for the polled request to
> > complete is the only thing that can be done (aside from resetting the
> > device), and it's generic behavior.
> 
> Then looks not safe to support IO polling for virtio-blk, maybe disable it
> at default now until the virtio-blk spec starts to support IO abort?

The virtio_blk.ko poll_queues module parameter is already set to 0 by
default. Poll queues are only available when the user has explicitly set
the module parameter.

I have added Suwan Kim to the email thread. Suwan Kim added poll queue
support to the virtio-blk driver and may have a preference for how to
proceed.

Stefan

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]