nvme-pci: fix timeout request state check
Subsystems (2)
PATCH v2 1/1 nvme-pci: fix timeout request state check
From: Keith Busch <[email protected]>
Polling the completion can progress the request state to IDLE, either inline with the completion, or through softirq. Either way, the state may not be COMPLETED, so don't check for that. We only care if the state isn't STARTED.
This is fixing an issue where the driver aborts an IO that we just completed. Seeing the "aborting" message instead of "polled" is very misleading as to where the timeout problem resides.
---v1->v2: Fixed spelling mistake in subject
drivers/nvme/host/pci.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/nvme/host/pci.c +1 -1
@@ -1362,7 +1362,7 @@ static enum blk_eh_timer_return nvme_timeout(struct request *req)else nvme_poll_irqdisable (nvmeq); -if (blk_mq_request_completed (req)) { +if (!blk_mq_request_started (req)) {dev_warn (dev->ctrl.device,"I/O %d QID %d timeout, completion polled\n" , req->tag, nvmeq->qid);
Signature · 1 line
2.30.2
Christoph Hellwig
maintainerQuoted 10 lines · Keith Busch
From: Keith Busch <[email protected]>
Polling the completion can progress the request state to IDLE, either inline with the completion, or through softirq. Either way, the state may not be COMPLETED, so don't check for that. We only care if the state isn't STARTED.
This is fixing an issue where the driver aborts an IO that we just completed. Seeing the "aborting" message instead of "polled" is very misleading as to where the timeout problem resides.
Hmm. Using a started helper for something that by definition is started doesn't really make much sense. I guess the problem here is that blk_mq_end_request_batch sets the state to MQ_RQ_IDLE afte calling blk_complete_request? Maybe we just need an explicit check for MQ_RQ_IDLE here as started seems like the wrong implication here.
Keith Busch
maintainerQuoted 12 lines · Christoph Hellwig
Quoted 10 lines · Keith Busch
From: Keith Busch <[email protected]>
Polling the completion can progress the request state to IDLE, either inline with the completion, or through softirq. Either way, the state may not be COMPLETED, so don't check for that. We only care if the state isn't STARTED.
This is fixing an issue where the driver aborts an IO that we just completed. Seeing the "aborting" message instead of "polled" is very misleading as to where the timeout problem resides.
Quoted 5 lines
Hmm. Using a started helper for something that by definition is started doesn't really make much sense. I guess the problem here is that
blk_mq_end_request_batchsets the state toMQ_RQ_IDLEafte callingblk_complete_request? Maybe we just need an explicit check forMQ_RQ_IDLEhere as started seems like the wrong implication here.
We're actually not batching here (no IOB in the timeout context), so we are either:
a. calling nvme_pci_complete_rq() inline with the cqe b. racing with smp ipi or softirq
If case (a), we will always see IDLE. If (b), we are racing and may see either COMPLETED or IDLE, so we have to check that it's not either of those. Since there's only one other state (STARTED) that was guaranteed prior to entering the timeout handler, we can just make sure it's not that one after the poll to know if abort escalation is needed.
Christoph Hellwig
maintainerQuoted 11 lines · Keith Busch
We're actually not batching here (no IOB in the timeout context), so we are either:
a. calling
nvme_pci_complete_rq()inline with the cqe b. racing with smp ipi or softirqIf case (a), we will always see IDLE. If (b), we are racing and may see either COMPLETED or IDLE, so we have to check that it's not either of those. Since there's only one other state (STARTED) that was guaranteed prior to entering the timeout handler, we can just make sure it's not that one after the poll to know if abort escalation is needed.
The point is still that "started" is the wrong check here and relies on an implementation detail. I think we're better off with an explicit IDLE check and a big fat comment.
Keith Busch
maintainerQuoted 13 lines · Christoph Hellwig
Quoted 11 lines · Keith Busch
We're actually not batching here (no IOB in the timeout context), so we are either:
a. calling
nvme_pci_complete_rq()inline with the cqe b. racing with smp ipi or softirqIf case (a), we will always see IDLE. If (b), we are racing and may see either COMPLETED or IDLE, so we have to check that it's not either of those. Since there's only one other state (STARTED) that was guaranteed prior to entering the timeout handler, we can just make sure it's not that one after the poll to know if abort escalation is needed.
Quoted 3 lines
The point is still that "started" is the wrong check here and relies on an implementation detail. I think we're better off with an explicit IDLE check and a big fat comment.
So you want the check to look like this instead?
---
@@ -1362,7 +1362,8 @@ static enum blk_eh_timer_return nvme_timeout(struct request *req)
else
nvme_poll_irqdisable(nvmeq);- if (blk_mq_request_completed(req)) {
+ if (blk_mq_request_completed(req) ||
+ blk_mq_rq_state(req) == MQ_RQ_IDLE) {
dev_warn(dev->ctrl.device,
"I/O %d QID %d timeout, completion polled\n",
req->tag, nvmeq->qid);--
That's essentially a more complicated equivalent to what I have, but fine with me if you think it's more clear.
Alternatively, I also considered moving the IDLE state setting to when the request is actually freed, which might make more sense and works without changing the nvme driver:
---
block/blk-mq.c +1 -1
@@ -713,6 +713,7 @@ static void __blk_mq_free_request(struct request *rq)struct blk_mq_hw_ctx *hctx = rq->mq_hctx;const int sched_tag = rq->internal_tag; +WRITE_ONCE (rq->state,MQ_RQ_IDLE );blk_crypto_free_request (rq);blk_pm_mark_last_busy (rq); rq->mq_hctx =NULL ;
@@ -741,7 +742,6 @@ void blk_mq_free_request(struct request *rq)rq_qos_done (q, rq); -WRITE_ONCE (rq->state,MQ_RQ_IDLE );if (req_ref_put_and_test (rq))__blk_mq_free_request (rq); }
--
Christoph Hellwig
maintainerQuoted 8 lines · Keith Busch
Quoted 3 lines
The point is still that "started" is the wrong check here and relies on an implementation detail. I think we're better off with an explicit IDLE check and a big fat comment.
So you want the check to look like this instead?
---
Review static enum blk_eh_timer_return nvme_timeout(struct request *req)
Quoted 10 lines
(quoted diff) +2
-2@@ -1362,7 +1362,8 @@ static enum blk_eh_timer_return nvme_timeout(struct request *req) else nvme_poll_irqdisable(nvmeq); - if (blk_mq_request_completed(req)) { + if (blk_mq_request_completed(req) || + blk_mq_rq_state(req) == MQ_RQ_IDLE) { dev_warn(dev->ctrl.device, "I/O %d QID %d timeout, completion polled\n", req->tag, nvmeq->qid); --
Or maybe just:
if (blk_mq_rq_state(req) != MQ_RQ_IN_FLIGHT) {
..Quoted 3 lines
Alternatively, I also considered moving the IDLE state setting to when the request is actually freed, which might make more sense and works without changing the nvme driver:
I like that idea, but this touches gnarly code, so it'll need good testing.