| Message ID | 1499154912-10420-2-git-send-email-sagi@grimberg.me (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Tue, Jul 04, 2017 at 10:55:05AM +0300, Sagi Grimberg wrote: > unlike blk_mq_stop_hw_queues and blk_mq_start_stopped_hw_queues > quiescing/unquiescing respects the submission path rcu grace. > Also make sure to kick the requeue list when appropriate. > > Signed-off-by: Sagi Grimberg <sagi@grimberg.me> > --- > drivers/nvme/host/rdma.c | 7 ++++--- > 1 file changed, 4 insertions(+), 3 deletions(-) > > diff --git a/drivers/nvme/host/rdma.c b/drivers/nvme/host/rdma.c > index cfb22531fc16..cec2c89cc8da 100644 > --- a/drivers/nvme/host/rdma.c > +++ b/drivers/nvme/host/rdma.c > @@ -778,7 +778,7 @@ static void nvme_rdma_error_recovery_work(struct work_struct *work) > > if (ctrl->ctrl.queue_count > 1) > nvme_stop_queues(&ctrl->ctrl); > - blk_mq_stop_hw_queues(ctrl->ctrl.admin_q); > + blk_mq_quiesce_queue(ctrl->ctrl.admin_q); > > /* We must take care of fastfail/requeue all our inflight requests */ > if (ctrl->ctrl.queue_count > 1) > @@ -791,7 +791,8 @@ static void nvme_rdma_error_recovery_work(struct work_struct *work) > * queues are not a live anymore, so restart the queues to fail fast > * new IO > */ > - blk_mq_start_stopped_hw_queues(ctrl->ctrl.admin_q, true); > + blk_mq_unquiesce_queue(ctrl->ctrl.admin_q); > + blk_mq_kick_requeue_list(ctrl->ctrl.admin_q); Now the queue won't be stopped via blk_mq_quiesce_queue(), so why do you add blk_mq_kick_requeue_list() here? Thanks, Ming
>> @@ -791,7 +791,8 @@ static void nvme_rdma_error_recovery_work(struct work_struct *work) >> * queues are not a live anymore, so restart the queues to fail fast >> * new IO >> */ >> - blk_mq_start_stopped_hw_queues(ctrl->ctrl.admin_q, true); >> + blk_mq_unquiesce_queue(ctrl->ctrl.admin_q); >> + blk_mq_kick_requeue_list(ctrl->ctrl.admin_q); > > Now the queue won't be stopped via blk_mq_quiesce_queue(), so why do > you add blk_mq_kick_requeue_list() here? I think you're right. We now quiesce the queue and fast fail inflight io, in nvme_complete_rq we call blk_mq_requeue_request with !blk_mq_queue_stopped(req->q) which is now true. So the requeue_work is triggered and requeue the request, and when we unquiesce we simply run the hw queues again. If we were to call it with !blk_queue_quiesced(req->q) I think it would be needed though...
>>> @@ -791,7 +791,8 @@ static void nvme_rdma_error_recovery_work(struct >>> work_struct *work) >>> * queues are not a live anymore, so restart the queues to fail >>> fast >>> * new IO >>> */ >>> - blk_mq_start_stopped_hw_queues(ctrl->ctrl.admin_q, true); >>> + blk_mq_unquiesce_queue(ctrl->ctrl.admin_q); >>> + blk_mq_kick_requeue_list(ctrl->ctrl.admin_q); >> >> Now the queue won't be stopped via blk_mq_quiesce_queue(), so why do >> you add blk_mq_kick_requeue_list() here? > > I think you're right. > > We now quiesce the queue and fast fail inflight io, in > nvme_complete_rq we call blk_mq_requeue_request with > !blk_mq_queue_stopped(req->q) which is now true. > > So the requeue_work is triggered and requeue the request, > and when we unquiesce we simply run the hw queues again. > > If we were to call it with !blk_queue_quiesced(req->q) > I think it would be needed though... If you look at nvme_start_queues, it also kicks the requeue work. I think that the proper fix for this is _keep_ the requeue kick and in nvme_complete_rq call: blk_mq_requeue_request(req, !blk_queue_quiesced(req->q)); Thoughts?
On Tue, Jul 04, 2017 at 12:07:38PM +0300, Sagi Grimberg wrote: > > > > > @@ -791,7 +791,8 @@ static void > > > > nvme_rdma_error_recovery_work(struct work_struct *work) > > > > * queues are not a live anymore, so restart the queues to > > > > fail fast > > > > * new IO > > > > */ > > > > - blk_mq_start_stopped_hw_queues(ctrl->ctrl.admin_q, true); > > > > + blk_mq_unquiesce_queue(ctrl->ctrl.admin_q); > > > > + blk_mq_kick_requeue_list(ctrl->ctrl.admin_q); > > > > > > Now the queue won't be stopped via blk_mq_quiesce_queue(), so why do > > > you add blk_mq_kick_requeue_list() here? > > > > I think you're right. > > > > We now quiesce the queue and fast fail inflight io, in > > nvme_complete_rq we call blk_mq_requeue_request with > > !blk_mq_queue_stopped(req->q) which is now true. > > > > So the requeue_work is triggered and requeue the request, > > and when we unquiesce we simply run the hw queues again. > > > > If we were to call it with !blk_queue_quiesced(req->q) > > I think it would be needed though... > > If you look at nvme_start_queues, it also kicks the requeue > work. I think that the proper fix for this is _keep_ the Then the kick can be removed from nvme_start_queues() > requeue kick and in nvme_complete_rq call: > > blk_mq_requeue_request(req, !blk_queue_quiesced(req->q)); > > Thoughts? I think we can always to kick the requeue work even when queue is stopped. It is OK to put the requeue req into sw queue/scheduler queue when queue is stopped.
> Then the kick can be removed from nvme_start_queues() > >> requeue kick and in nvme_complete_rq call: >> >> blk_mq_requeue_request(req, !blk_queue_quiesced(req->q)); >> >> Thoughts? > > I think we can always to kick the requeue work even when queue > is stopped. It is OK to put the requeue req into sw queue/scheduler > queue when queue is stopped. > Agreed.
diff --git a/drivers/nvme/host/rdma.c b/drivers/nvme/host/rdma.c index cfb22531fc16..cec2c89cc8da 100644 --- a/drivers/nvme/host/rdma.c +++ b/drivers/nvme/host/rdma.c @@ -778,7 +778,7 @@ static void nvme_rdma_error_recovery_work(struct work_struct *work) if (ctrl->ctrl.queue_count > 1) nvme_stop_queues(&ctrl->ctrl); - blk_mq_stop_hw_queues(ctrl->ctrl.admin_q); + blk_mq_quiesce_queue(ctrl->ctrl.admin_q); /* We must take care of fastfail/requeue all our inflight requests */ if (ctrl->ctrl.queue_count > 1) @@ -791,7 +791,8 @@ static void nvme_rdma_error_recovery_work(struct work_struct *work) * queues are not a live anymore, so restart the queues to fail fast * new IO */ - blk_mq_start_stopped_hw_queues(ctrl->ctrl.admin_q, true); + blk_mq_unquiesce_queue(ctrl->ctrl.admin_q); + blk_mq_kick_requeue_list(ctrl->ctrl.admin_q); nvme_start_queues(&ctrl->ctrl); nvme_rdma_reconnect_or_remove(ctrl); @@ -1636,7 +1637,7 @@ static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl) if (test_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[0].flags)) nvme_shutdown_ctrl(&ctrl->ctrl); - blk_mq_stop_hw_queues(ctrl->ctrl.admin_q); + blk_mq_quiesce_queue(ctrl->ctrl.admin_q); blk_mq_tagset_busy_iter(&ctrl->admin_tag_set, nvme_cancel_request, &ctrl->ctrl); nvme_rdma_destroy_admin_queue(ctrl);
unlike blk_mq_stop_hw_queues and blk_mq_start_stopped_hw_queues quiescing/unquiescing respects the submission path rcu grace. Also make sure to kick the requeue list when appropriate. Signed-off-by: Sagi Grimberg <sagi@grimberg.me> --- drivers/nvme/host/rdma.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-)