On Tue, May 17, 2022 at 06:00:57PM +0800, Ziyang Zhang wrote:
> On 2022/5/17 13:53, Ming Lei wrote:
>
> > +
> > +static void ubd_cancel_queue(struct ubd_queue *ubq)
> > +{
> > + int i;
> > +
> > + for (i = 0; i < ubq->q_depth; i++) {
> > + struct ubd_io *io = &ubq->ios[i];
> > +
> > + if (io->flags & UBD_IO_FLAG_ACTIVE) {
> > + io->flags &= ~UBD_IO_FLAG_ACTIVE;
> > + io_uring_cmd_done(io->cmd, UBD_IO_RES_ABORT, 0);
> > + }
> > + }
> > +}
>
> Hi Ming,
>
> When ubdsrv sends STOP_DEV and all active IOs in ubd_drv are done(UBD_IO_RES_ABORT),
> there may be still some IOs handled by ubdsrv(UBD_IO_FLAG_ACTIVE not set).
> When these IOs complete and return to ubd_drv, how to handle them?
Either UBD_IO_COMMIT_AND_FETCH_REQ or UBD_IO_COMMIT_REQ will be sent to ubd_drv
for completing these IOs. And finally ubd_cancel_dev() in ubd driver will
cancel all pending io commands, so io_uring can be exited. I guess
UBD_IO_COMMIT_REQ can be removed too.
> I find that UBD_IO_FETCH_REQ are still set,
> so will these IOs be issued to ubdsrv again or canceled?
> (I see ubd_drv fails IOs when the daemon is dying
> but maybe here the daemon is still alive)
If daemon is alive, ubd_drv will rely on ubq_daemon for completing
all inflight IOs. Otherwise, the monitor work will be triggered for
completing/failing inflight IOs. The mechanism is actually very simple:
static void ubd_stop_dev(struct ubd_device *ub)
{
mutex_lock(&ub->mutex);
if (!disk_live(ub->ub_disk))
goto unlock;
del_gendisk(ub->ub_disk); // drain & wait in-flight IOs
ub->dev_info.state = UBD_S_DEV_DEAD;
ub->dev_info.ubdsrv_pid = -1;
ubd_cancel_dev(ub); //No IO is possible now, so cancel pending io commands
unlock:
mutex_unlock(&ub->mutex);
cancel_delayed_work_sync(&ub->monitor_work);
}
When waiting for IO completion in del_gendisk(), in case that ubq_daemon
is exiting/dying, monitor work will be triggered to call ubd_abort_queue() to
fail in-flight requests for making forward progress. ubd_abort_queue() may
looks a bit tricky to try using task work for aborting request, that
is just for sync with ubd_rq_task_work_fn().
Thanks,
Ming
