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

[Keith Busch]

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.