From: James Smart <jsmart2...@gmail.com>
If:
- A successful connect has occurred with an io queue count greater than
zero and namespaces detected and running.
- An error or something occurs which causes a termination of the prior
association and then starts a reconnect,
- The reconnect then creates a new controller, but for whatever reason,
nvme_set_queue_count() results in io queue count set to zero. This
will skip io queue and tag set changes.
- But... the controller will transition to live, calling
nvme_start_ctrl, which calls nvme_start_queues(), which then releases
I/Os into the transport which then sends them to the driver.
As there are no queues, things eventually hit the driver looking for a
handle, which was cleared when the original controller was reset, and it
can't proceed. Worst case, things progress, but everything fails.
In the failing scenario, the nvme_set_features(NVME_FEAT_NUM_QUEUES)
command actually failed with a NVME_SC_INTERNAL error. For some reason,
although nvme_set_queue_count() saw the error and set io queue count to
zero, it doesn't return a failure status to the transport, which allows
the transport to continue using the controller.
Fix the problem by simply rejecting the new association if at least 1
I/O queue can't be created. The association reject will fail the
reconnect attempt and fall into the reconnect retry policy.
Signed-off-by: James Smart <jsmart2...@gmail.com>
Reviewed-by: Sagi Grimberg <s...@grimberg.me>
Signed-off-by: Christoph Hellwig <h...@lst.de>
---
drivers/nvme/host/fc.c | 12 ++++++++++++
1 file changed, 12 insertions(+)
diff --git a/drivers/nvme/host/fc.c b/drivers/nvme/host/fc.c
index 23f6bad19274..f3b9d91ba0df 100644
--- a/drivers/nvme/host/fc.c
+++ b/drivers/nvme/host/fc.c
@@ -2475,6 +2475,7 @@ static int
nvme_fc_recreate_io_queues(struct nvme_fc_ctrl *ctrl)
{
struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
+ u32 prior_ioq_cnt = ctrl->ctrl.queue_count - 1;
unsigned int nr_io_queues;
int ret;
@@ -2487,6 +2488,13 @@ nvme_fc_recreate_io_queues(struct nvme_fc_ctrl *ctrl)
return ret;
}
+ if (!nr_io_queues && prior_ioq_cnt) {
+ dev_info(ctrl->ctrl.device,
+ "Fail Reconnect: At least 1 io queue "
+ "required (was %d)\n", prior_ioq_cnt);
+ return -ENOSPC;
+ }
+
ctrl->ctrl.queue_count = nr_io_queues + 1;
/* check for io queues existing */
if (ctrl->ctrl.queue_count == 1)
@@ -2500,6 +2508,10 @@ nvme_fc_recreate_io_queues(struct nvme_fc_ctrl *ctrl)
if (ret)
goto out_delete_hw_queues;
+ if (prior_ioq_cnt != nr_io_queues)
+ dev_info(ctrl->ctrl.device,
+ "reconnect: revising io queue count from %d to %d\n",
+ prior_ioq_cnt, nr_io_queues);
blk_mq_update_nr_hw_queues(&ctrl->tag_set, nr_io_queues);
return 0;
--
2.20.1
