On Tue, Aug 14, 2018 at 02:03:58PM +0800, jianchao.wang wrote:
>
>
> On 08/14/2018 01:56 PM, jianchao.wang wrote:
> > Hi Ming
> >
> > On 08/11/2018 03:12 PM, Ming Lei wrote:
> >> +static blk_status_t scsi_admin_queue_rq(struct blk_mq_hw_ctx *hctx,
> >> + const struct blk_mq_queue_data *bd)
> >> +{
> >> + struct scsi_device *sdev = scsi_req(bd->rq)->sdev;
> >> +
> >> + WARN_ON_ONCE(hctx->queue == sdev->request_queue);
> >> +
> >> + if (!__scsi_mq_get_budget(hctx, sdev))
> >> + return BLK_STS_RESOURCE;
> >> +
> >> + return __scsi_queue_rq(hctx, bd, sdev);
> >> +}
> >
> > Why does it return BLK_STS_RESOURCE here ?
> >
> > Do you want to use the following code branch in blk_mq_dispatch_rq_list to get the admin queue rerun ?
> >
> > else if (needs_restart && (ret == BLK_STS_RESOURCE))
> > blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
> >
> > In theory, BLK_STS_DEV_RESOURCE should be returned here. However, I guess it will not work. Because
> > the rr fashion loop restart has gone, the completion of normal queue will not restart the admin queue.
> > If yes, please add some comment here to describe it. :)
> >
>
> Looks like the patch 10 could avoid the admin queue hung.
Yes, this one is patch 12, :-)
Also not necessary to add comment on returning BLK_STS_RESOURCE because
it is always the safe say, given blk-mq will handle the queue rerun.
However, if driver returns BLK_STS_DEV_RESOURCE, some comment have to
be added, please see comment on BLK_STS_DEV_RESOURCE.
Thanks,
Ming
