On Wed, Dec 06, 2017 at 12:28:25AM +0800, Ming Lei wrote:
> On Tue, Dec 05, 2017 at 04:08:20PM +0000, Bart Van Assche wrote:
> > On Tue, 2017-12-05 at 15:52 +0800, Ming Lei wrote:
> > > diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
> > > index db9556662e27..1816dd8259b3 100644
> > > --- a/drivers/scsi/scsi_lib.c
> > > +++ b/drivers/scsi/scsi_lib.c
> > > @@ -1967,6 +1967,8 @@ static bool scsi_mq_get_budget(struct blk_mq_hw_ctx *hctx)
> > > out_put_device:
> > > put_device(&sdev->sdev_gendev);
> > > out:
> > > + if (atomic_read(&sdev->device_busy) == 0 && !scsi_device_blocked(sdev))
> > > + blk_mq_delay_run_hw_queue(hctx, SCSI_QUEUE_DELAY);
> > > return false;
> > > }
> >
> > This cannot work since multiple threads can call scsi_mq_get_budget()
>
> That is exactly the way we are handling these cases before 0df21c86bdbf(scsi:
> implement .get_budget and .put_budget for blk-mq), so if it can't work,
> that is not fault of commit 0df21c86bdbf.
>
> > concurrently and hence it can happen that none of them sees
> > atomic_read(&sdev->device_busy) == 0. BTW, the above patch is something I
>
> If there is concurrent .get_budget(), one of them will see the counter
> becoming zero finally because each sdev->device_busy is inc/dec
> atomically. Or scsi_dev_queue_ready() return true.
>
> Anyway, we need this patch to avoid possible regression. If you think
> there are bugs in blk-mq RESTART, just root cause and and fix it.
>
> > consider as a workaround. Have you considered to fix the root cause and to
> > add blk_mq_sched_mark_restart_hctx() where these are missing, e.g. in
> > blk_mq_sched_dispatch_requests()? The patch below is not a full solution
> > but resulted in a significant improvement in my tests:
> >
> > diff --git a/block/blk-mq-sched.c b/block/blk-mq-sched.c
> > index 69e3226e66ca..9d86876ec503 100644
> > --- a/block/blk-mq-sched.c
> > +++ b/block/blk-mq-sched.c
> > @@ -226,6 +226,7 @@ void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
> > * TODO: get more budgets, and dequeue more requests in
> > * one time.
> > */
> > + blk_mq_sched_mark_restart_hctx(hctx);
> > blk_mq_do_dispatch_ctx(hctx);
> > } else {
> > blk_mq_flush_busy_ctxs(hctx, &rq_list);
>
BTW, this kind of change can't cover scsi_set_blocked() which is
triggered by timeout, scsi dispatch failure. You will see that
easily if you run the SCSI test script I provided in the commit log.
--
Ming
