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

Date: Tue Mar 26 2019 - 22:28:23 EST

On 3/27/19 10:15 AM, Keith Busch wrote:
> On Wed, Mar 27, 2019 at 10:03:26AM +0800, wrote:
>> Hi Keith
>> On 3/27/19 7:57 AM, Keith Busch wrote:
>>> On Mon, Mar 25, 2019 at 08:05:53PM -0700, 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.

As the comment above, the stable request maybe something that has been freed due to following case,
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 freed request could be allocated by others which may change the field of request->state.