Re: [PATCH V2 7/8] nvme: use blk_mq_queue_tag_inflight_iter

From: Keith Busch
Date: Tue Mar 26 2019 - 22:14:16 EST


On Wed, Mar 27, 2019 at 10:03:26AM +0800, jianchao.wang wrote:
> Hi Keith
>
> On 3/27/19 7:57 AM, Keith Busch wrote:
> > On Mon, Mar 25, 2019 at 08:05:53PM -0700, jianchao.wang wrote:
> >> What if there used to be a io scheduler and leave some stale requests of sched tags ?
> >> Or the nr_hw_queues was decreased and leave the hctx->fq->flush_rq ?
> >
> > Requests internally queued in scheduler or block layer are not eligible
> > for the nvme driver's iterator callback. We only use it to reclaim
> > dispatched requests that the target can't return, which only applies to
> > requests that must have a valid rq->tag value from hctx->tags.
> >
> >> The stable request could be some tings freed and used
> >> by others and the state field happen to be overwritten to non-zero...
> >
> > I am not sure I follow what this means. At least for nvme, every queue
> > sharing the same tagset is quiesced and frozen, there should be no
> > request state in flux at the time we iterate.
> >
>
> In nvme_dev_disable, when we try to reclaim the in-flight requests with blk_mq_tagset_busy_iter,
> the request_queues are quiesced but just start-freeze.
> We will try to _drain_ the in-flight requests for the _shutdown_ case when controller is not dead.
> For the reset case, there still could be someone escapes the checking of queue freezing and enters
> blk_mq_make_request and tries to allocate tag, then we may get,
>
> generic_make_request nvme_dev_disable
> -> blk_queue_enter
> -> nvme_start_freeze (just start freeze, no drain)
> -> nvme_stop_queues
> -> blk_mq_make_request
> - > blk_mq_get_request -> blk_mq_tagset_busy_iter
> -> blk_mq_get_tag
> ãã -> bt_tags_for_each
> ããã -> bt_tags_iter
> ãã -> rq = tags->rqs[] ---> [1]
> -> blk_mq_rq_ctx_init
> -> data->hctx->tags->rqs[rq->tag] = rq;
>
> The rq got on position [1] could be a stale request that has been freed due to,
> 1. a hctx->fq.flush_rq of dead request_queue that shares the same tagset
> 2. a removed io scheduler's sched request
>
> And this stale request may have been used by others and the request->state is changed to a non-zero
> value and passes the checking of blk_mq_request_started and then it will be handled by nvme_cancel_request.

How is that request state going to be anyting other than IDLE? A freed
request state is IDLE, and continues to be IDLE until dispatched. But
dispatch is blocked for the entire tagset, so request states can't be
started during an nvme reset.