On 2020/10/15 15:50, Ming Lei wrote:
On Thu, Oct 15, 2020 at 02:05:01PM +0800, Chao Leng wrote:
On 2020/10/14 17:56, Ming Lei wrote:
On Wed, Oct 14, 2020 at 05:39:12PM +0800, Chao Leng wrote:
On 2020/10/14 11:34, Ming Lei wrote:
On Wed, Oct 14, 2020 at 09:08:28AM +0800, Ming Lei wrote:
On Tue, Oct 13, 2020 at 03:36:08PM -0700, Sagi Grimberg wrote:
This may just reduce the probability. The concurrency of timeout
and teardown will cause the same request
be treated repeatly, this is not we expected.
That is right, not like SCSI, NVME doesn't apply atomic request
completion, so
request may be completed/freed from both timeout & nvme_cancel_request().
.teardown_lock still may cover the race with Sagi's patch because
teardown
actually cancels requests in sync style.
In extreme scenarios, the request may be already retry success(rq state
change to inflight).
Timeout processing may wrongly stop the queue and abort the request.
teardown_lock serialize the process of timeout and teardown, but do not
avoid the race.
It might not be safe.
Not sure I understand the scenario you are describing.
what do you mean by "In extreme scenarios, the request may be already retry
success(rq state change to inflight)"?
What will retry the request? only when the host will reconnect
the request will be retried.
We can call nvme_sync_queues in the last part of the teardown, but
I still don't understand the race here.
Not like SCSI, NVME doesn't complete request atomically, so double
completion/free can be done from both timeout & nvme_cancel_request()(via teardown).
Given request is completed remotely or asynchronously in the two code paths,
the teardown_lock can't protect the case.
Thinking of the issue further, the race shouldn't be between timeout and
teardown.
Both nvme_cancel_request() and nvme_tcp_complete_timed_out() are called
with .teardown_lock, and both check if the request is completed before
calling blk_mq_complete_request() which marks the request as COMPLETE state.
So the request shouldn't be double-freed in the two code paths.
Another possible reason is that between timeout and normal completion(fail
fast pending requests after ctrl state is updated to CONNECTING).
Yi, can you try the following patch and see if the issue is fixed?
diff --git a/drivers/nvme/host/tcp.c b/drivers/nvme/host/tcp.c
index d6a3e1487354..fab9220196bd 100644
--- a/drivers/nvme/host/tcp.c
+++ b/drivers/nvme/host/tcp.c
@@ -1886,7 +1886,6 @@ static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new)
static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
bool remove)
{
- mutex_lock(&to_tcp_ctrl(ctrl)->teardown_lock);
blk_mq_quiesce_queue(ctrl->admin_q);
nvme_tcp_stop_queue(ctrl, 0);
if (ctrl->admin_tagset) {
@@ -1897,15 +1896,13 @@ static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
if (remove)
blk_mq_unquiesce_queue(ctrl->admin_q);
nvme_tcp_destroy_admin_queue(ctrl, remove);
- mutex_unlock(&to_tcp_ctrl(ctrl)->teardown_lock);
}
static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
bool remove)
{
- mutex_lock(&to_tcp_ctrl(ctrl)->teardown_lock);
if (ctrl->queue_count <= 1)
- goto out;
+ return;
blk_mq_quiesce_queue(ctrl->admin_q);
nvme_start_freeze(ctrl);
nvme_stop_queues(ctrl);
@@ -1918,8 +1915,6 @@ static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
if (remove)
nvme_start_queues(ctrl);
nvme_tcp_destroy_io_queues(ctrl, remove);
-out:
- mutex_unlock(&to_tcp_ctrl(ctrl)->teardown_lock);
}
static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl)
@@ -2030,11 +2025,11 @@ static void nvme_tcp_error_recovery_work(struct work_struct *work)
struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
nvme_stop_keep_alive(ctrl);
+
+ mutex_lock(&tcp_ctrl->teardown_lock);
nvme_tcp_teardown_io_queues(ctrl, false);
- /* unquiesce to fail fast pending requests */
- nvme_start_queues(ctrl);
nvme_tcp_teardown_admin_queue(ctrl, false);
- blk_mq_unquiesce_queue(ctrl->admin_q);
Delete blk_mq_unquiesce_queue will cause a bug which may cause reconnect failed.
Delete nvme_start_queues may cause another bug.
nvme_tcp_setup_ctrl() will re-start io and admin queue, and only .connect_q
and .fabrics_q are required during reconnect.I check the code. Unquiesce the admin queue in nvme_tcp_configure_admin_queue, so reconnect can work well.
So can you explain in detail about the bug?
First if reconnect failed, quiesce the io queue and admin queue will cause IO pause long time.
Any normal IO can't make progress until reconnect is successful, so this
change won't increase IO pause. This way is exactly what NVMe PCI takes,
see nvme_start_queues() called from nvme_reset_work().
now is ok. Now the patch which fix the long pause time is discussing.
Second if reconnect failed more than max_reconnects, delete ctrl will hang.
No, delete ctrl won't hang, because 'shutdown' parameter is true in case
of deleting ctrl, which will unquiesce both admin_q and io queues in
nvme_tcp_teardown_io_queues() and nvme_tcp_teardown_admin_queue().
No, now nvme_remove_namespaces is before tear down queues.
tear down queues is in ctrl->ops->delete_ctrl.
static void nvme_do_delete_ctrl(struct nvme_ctrl *ctrl)
{
dev_info(ctrl->device,
"Removing ctrl: NQN \"%s\"\n", ctrl->opts->subsysnqn);
flush_work(&ctrl->reset_work);
nvme_stop_ctrl(ctrl);
nvme_remove_namespaces(ctrl);
ctrl->ops->delete_ctrl(ctrl);
nvme_uninit_ctrl(ctrl);
}
Thanks,
Ming
_______________________________________________
Linux-nvme mailing list
Linux-nvme@xxxxxxxxxxxxxxxxxxx
http://lists.infradead.org/mailman/listinfo/linux-nvme
.
