Re: [PATCH v5] conversion to blk-mq

[Keith Busch]

On Wed, 4 Jun 2014, Matias Bjørling wrote:
> On 06/04/2014 12:27 AM, Keith Busch wrote:
> > > On Tue, 3 Jun 2014, Matias Bjorling wrote:
> > > > Keith, will you take the nvmemq_wip_v6 branch for a spin? Thanks!
> >
> > BTW, if you want to test this out yourself, it's pretty simple to
> > recreate. I just run a simple user admin program sending nvme passthrough
> > commands in a tight loop, then run:
> >
> >   # echo 1 > /sys/bus/pci/devices/<bdf>/remove
>
> I can't recreate- I use the nvme_get_feature program to continuously hit the 
> ioctl path, testing using your nvme qemu branch.

Okay, I'll try to fix it.

I think there are multiple problems, but the first is that since there
is no gendisk associated with the admin_q, the QUEUE_FLAG_INIT_DONE flag
is never set, and blk_mq_queue_enter returns successful whenever this
flag is not set even though this queue is dying, so we enter with all
its invalid pointers.

Here's a couple diff's. The first fixes the kernel oops by not entering a
dying queue. The second is just a few unrelated clean-ups in nvme-core.c.

I still can't complete my current hot-removal test, though; something
appears hung, but haven't nailed that down yet.

Please let me know what you think! Thanks.

diff --git a/block/blk-mq.c b/block/blk-mq.c
index d10013b..5a9ae8a 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -105,6 +105,10 @@ static int blk_mq_queue_enter(struct request_queue *q)
 	__percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
 	smp_wmb();
 	/* we have problems to freeze the queue if it's initializing */
+	if (blk_queue_dying(q)) {
+		__percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
+		ret = -ENODEV;
+	}
 	if (!blk_queue_bypass(q) || !blk_queue_init_done(q))
 		return 0;

diff --git a/drivers/block/nvme-core.c b/drivers/block/nvme-core.c
index 243a5e6..22e9c82 100644
--- a/drivers/block/nvme-core.c
+++ b/drivers/block/nvme-core.c
@@ -98,7 +98,6 @@ struct nvme_queue {
 	u8 cq_phase;
 	u8 cqe_seen;
 	u8 q_suspended;
-	cpumask_var_t cpu_mask;
 	struct async_cmd_info cmdinfo;
 	struct blk_mq_hw_ctx *hctx;
 };
@@ -1055,8 +1054,6 @@ static void nvme_free_queue(struct nvme_queue *nvmeq)
 				(void *)nvmeq->cqes, nvmeq->cq_dma_addr);
 	dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
 					nvmeq->sq_cmds, nvmeq->sq_dma_addr);
-	if (nvmeq->qid)
-		free_cpumask_var(nvmeq->cpu_mask);
 	kfree(nvmeq);
 }

@@ -1066,9 +1063,9 @@ static void nvme_free_queues(struct nvme_dev *dev, int lowest)

 	for (i = dev->queue_count - 1; i >= lowest; i--) {
 		struct nvme_queue *nvmeq = dev->queues[i];
-		nvme_free_queue(nvmeq);
 		dev->queue_count--;
 		dev->queues[i] = NULL;
+		nvme_free_queue(nvmeq);
 	}
 }

@@ -1142,9 +1139,6 @@ static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
 	if (!nvmeq->sq_cmds)
 		goto free_cqdma;

-	if (qid && !zalloc_cpumask_var(&nvmeq->cpu_mask, GFP_KERNEL))
-		goto free_sqdma;
-
 	nvmeq->q_dmadev = dmadev;
 	nvmeq->dev = dev;
 	snprintf(nvmeq->irqname, sizeof(nvmeq->irqname), "nvme%dq%d",
@@ -1162,9 +1156,6 @@ static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,

 	return nvmeq;

- free_sqdma:
-	dma_free_coherent(dmadev, SQ_SIZE(depth), (void *)nvmeq->sq_cmds,
-							nvmeq->sq_dma_addr);
  free_cqdma:
 	dma_free_coherent(dmadev, CQ_SIZE(depth), (void *)nvmeq->cqes,
 							nvmeq->cq_dma_addr);