Re: [PATCH 5/5] nvme: use __blk_mq_complete_request in timeout path

From: Sagi Grimberg
Date: Sun Jun 24 2018 - 14:08:06 EST



Hi Christoph

Thanks for your kindly response.

On 06/20/2018 10:39 PM, Christoph Hellwig wrote:
diff --git a/drivers/nvme/host/pci.c b/drivers/nvme/host/pci.c
index 73a97fc..2a161f6 100644
--- a/drivers/nvme/host/pci.c
+++ b/drivers/nvme/host/pci.c
@@ -1203,6 +1203,7 @@ static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
nvme_warn_reset(dev, csts);
nvme_dev_disable(dev, false);
nvme_reset_ctrl(&dev->ctrl);
+ __blk_mq_complete_request(req);
return BLK_EH_DONE;
}
@@ -1213,6 +1214,11 @@ static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
dev_warn(dev->ctrl.device,
"I/O %d QID %d timeout, completion polled\n",
req->tag, nvmeq->qid);
+ /*
+ * nvme_end_request will invoke blk_mq_complete_request,
+ * it will do nothing for this timed out request.
+ */
+ __blk_mq_complete_request(req);

And this clearly is bogus. We want to iterate over the tagetset
and cancel all requests, not do that manually here.

That was the whole point of the original change.


For nvme-pci, we indeed have an issue that when nvme_reset_work->nvme_dev_disable returns, timeout path maybe still
running and the nvme_dev_disable invoked by timeout path will race with the nvme_reset_work.
However, the hole is still there right now w/o my changes, but just narrower.

Given the amount of fixes (and fixes of fixes) we had in the timeout handler, maybe it'd be a good idea to step back and take a another look?

Won't it be better to avoid disabling the device and return
BLK_EH_RESET_TIMER if we are not aborting in the timeout handler?