2018-05-24 11:30:50

by jianchao.wang

Subject: [PATCH 1/2] nvme-pci: set nvmeq->cq_vector after alloc cq/sq

Currently nvmeq->cq_vector is set before alloc cq/sq. If the alloc
cq/sq command timeout, nvme_suspend_queue will invoke free_irq
for the nvmeq because the cq_vector is valid, this will cause warning
'Trying to free already-free IRQ xxx'. set nvmeq->cq_vector after
alloc cq/sq successes to fix this.

Signed-off-by: Jianchao Wang <[email protected]>
---
drivers/nvme/host/pci.c | 25 ++++++++++++++++---------
1 file changed, 16 insertions(+), 9 deletions(-)

diff --git a/drivers/nvme/host/pci.c b/drivers/nvme/host/pci.c
index 17a0190..70b79f2 100644
--- a/drivers/nvme/host/pci.c
+++ b/drivers/nvme/host/pci.c
@@ -1070,7 +1070,7 @@ static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
}

static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
- struct nvme_queue *nvmeq)
+ struct nvme_queue *nvmeq, int cq_vector)
{
struct nvme_command c;
int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
@@ -1085,7 +1085,7 @@ static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
c.create_cq.cqid = cpu_to_le16(qid);
c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
c.create_cq.cq_flags = cpu_to_le16(flags);
- c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
+ c.create_cq.irq_vector = cpu_to_le16(cq_vector);

return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
}
@@ -1459,6 +1459,7 @@ static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
{
struct nvme_dev *dev = nvmeq->dev;
int result;
+ int cq_vector;

if (dev->cmb && use_cmb_sqes && (dev->cmbsz & NVME_CMBSZ_SQS)) {
unsigned offset = (qid - 1) * roundup(SQ_SIZE(nvmeq->q_depth),
@@ -1471,15 +1472,21 @@ static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
* A queue's vector matches the queue identifier unless the controller
* has only one vector available.
*/
- nvmeq->cq_vector = dev->num_vecs == 1 ? 0 : qid;
- result = adapter_alloc_cq(dev, qid, nvmeq);
+ cq_vector = dev->num_vecs == 1 ? 0 : qid;
+ result = adapter_alloc_cq(dev, qid, nvmeq, cq_vector);
if (result < 0)
- goto release_vector;
+ goto out;

result = adapter_alloc_sq(dev, qid, nvmeq);
if (result < 0)
goto release_cq;

+ /*
+ * set cq_vector after alloc cq/sq, otherwise, if alloc cq/sq command
+ * timeout, nvme_suspend_queue will invoke free_irq for it and cause warning
+ * 'Trying to free already-free IRQ xxx'
+ */
+ nvmeq->cq_vector = cq_vector;
nvme_init_queue(nvmeq, qid);
result = queue_request_irq(nvmeq);
if (result < 0)
@@ -1487,13 +1494,13 @@ static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)

return result;

- release_sq:
+release_sq:
+ nvmeq->cq_vector = -1;
dev->online_queues--;
adapter_delete_sq(dev, qid);
- release_cq:
+release_cq:
adapter_delete_cq(dev, qid);
- release_vector:
- nvmeq->cq_vector = -1;
+out:
return result;
}

--
2.7.4

2018-05-24 11:29:30

by jianchao.wang

Subject: [PATCH 2/2] nvme-pci: don't send delete cq command when allocate sq return -EINTR

When allocate sq in nvme_create_queue return -EINTR, it means the
allocate sq command times out and is completed with NVME_REQ_CANCELLED.
At the moment, the controller has been disabled and admin request
queue has been quiesced. Don't send delete cq command, otherwise,
it will incur io hang.

Signed-off-by: Jianchao Wang <[email protected]>
---
drivers/nvme/host/pci.c | 10 +++++++++-
1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/drivers/nvme/host/pci.c b/drivers/nvme/host/pci.c
index 70b79f2..4ce1c89 100644
--- a/drivers/nvme/host/pci.c
+++ b/drivers/nvme/host/pci.c
@@ -1478,7 +1478,15 @@ static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
goto out;

result = adapter_alloc_sq(dev, qid, nvmeq);
- if (result < 0)
+ /*
+ * If return -EINTR, it means the allocate sq command times out and is completed
+ * with NVME_REQ_CANCELLED. At the time, the controller has been disabled
+ * and admin request queue has been quiesced. So don't try to send delete cq
+ * command any more.
+ */
+ if (result == -EINTR)
+ goto out;
+ else if (result < 0)
goto release_cq;

/*
--
2.7.4

2018-05-24 23:23:42

by Keith Busch

Subject: Re: [PATCH 1/2] nvme-pci: set nvmeq->cq_vector after alloc cq/sq

This looks fine.

Reviewed-by: Keith Busch <[email protected]>

2018-05-24 23:27:37

by Keith Busch

Subject: Re: [PATCH 2/2] nvme-pci: don't send delete cq command when allocate sq return -EINTR

On Thu, May 24, 2018 at 05:51:34PM +0800, Jianchao Wang wrote:
> result = adapter_alloc_sq(dev, qid, nvmeq);
> - if (result < 0)
> + /*
> + * If return -EINTR, it means the allocate sq command times out and is completed
> + * with NVME_REQ_CANCELLED. At the time, the controller has been disabled
> + * and admin request queue has been quiesced. So don't try to send delete cq
> + * command any more.
> + */
> + if (result == -EINTR)
> + goto out;
> + else if (result < 0)
> goto release_cq;

Since you're touching this part anyway, you'd really only want to goto
the release_cq if result is > 0.

2018-05-25 08:48:36

by Christoph Hellwig

Subject: Re: [PATCH 1/2] nvme-pci: set nvmeq->cq_vector after alloc cq/sq

Thanks,

applied to nvme-4.18.