[PATCH v3 0/3] keepalive bugfixes

[PATCH v3 0/3] keepalive bugfixes

[Uday Shankar]

While reviewing the Linux KATO implementation in an attempt to better
understand the current NVMe Keep Alive specification, we found a
few issues in the host implementation that could contribute to spurious
Keep Alive timeouts being detected by controllers.

Changes v2-v3 (https://lore.kernel.org/linux-nvme/20230424232225.1975793-1-ushankar@purestorage.com/):
- Patch 3: add a warning log line for long keepalive RTT

Changes v1-v2 (https://lore.kernel.org/linux-nvme/20230417225558.2890062-1-ushankar@purestorage.com/):
- Patch 2: fix indentation, set start_time even when stats disabled


Uday Shankar (3):
  nvme: double KA polling frequency to avoid KATO with TBKAS on
  nvme: check IO start time when deciding to defer KA
  nvme: improve handling of long keep alives

 drivers/nvme/host/core.c | 41 +++++++++++++++++++++++++++++++++++++---
 drivers/nvme/host/nvme.h |  4 ++--
 2 files changed, 40 insertions(+), 5 deletions(-)

-- 
2.25.1

[PATCH v3 1/3] nvme: double KA polling frequency to avoid KATO with TBKAS on

[Uday Shankar]

With TBKAS on, the completion of one command can defer sending a
keep alive for up to twice the delay between successive runs of
nvme_keep_alive_work. The current delay of KATO / 2 thus makes it
possible for one command to defer sending a keep alive for up to
KATO, which can result in the controller detecting a KATO. The following
trace demonstrates the issue, taking KATO = 8 for simplicity:

1. t = 0: run nvme_keep_alive_work, no keep-alive sent
2. t = ε: I/O completion seen, set comp_seen = true
3. t = 4: run nvme_keep_alive_work, see comp_seen == true,
          skip sending keep-alive, set comp_seen = false
4. t = 8: run nvme_keep_alive_work, see comp_seen == false,
          send a keep-alive command.

Here, there is a delay of 8 - ε between receiving a command completion
and sending the next command. With ε small, the controller is likely to
detect a keep alive timeout.

Fix this by running nvme_keep_alive_work with a delay of KATO / 4
whenever TBKAS is on. Going through the above trace now gives us a
worst-case delay of 4 - ε, which is in line with the recommendation of
sending a command every KATO / 2 in the NVMe specification.

Reported-by: Costa Sapuntzakis <costa@purestorage.com>
Reported-by: Randy Jennings <randyj@purestorage.com>
Signed-off-by: Uday Shankar <ushankar@purestorage.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
---
 drivers/nvme/host/core.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/drivers/nvme/host/core.c b/drivers/nvme/host/core.c
index f586a4808e6e..7726277b951a 100644
--- a/drivers/nvme/host/core.c
+++ b/drivers/nvme/host/core.c
@@ -1160,10 +1160,16 @@ EXPORT_SYMBOL_NS_GPL(nvme_passthru_end, NVME_TARGET_PASSTHRU);
  * 
  *   The host should send Keep Alive commands at half of the Keep Alive Timeout
  *   accounting for transport roundtrip times [..].
+ * 
+ * When TBKAS is on, we need to run nvme_keep_alive_work at twice this
+ * frequency, as one command completion can postpone sending a keep alive
+ * command by up to twice the delay between runs.
  */
 static void nvme_queue_keep_alive_work(struct nvme_ctrl *ctrl)
 {
-	queue_delayed_work(nvme_wq, &ctrl->ka_work, ctrl->kato * HZ / 2);
+	unsigned long delay = (ctrl->ctratt & NVME_CTRL_ATTR_TBKAS) ?
+		ctrl->kato * HZ / 4 : ctrl->kato * HZ / 2;
+	queue_delayed_work(nvme_wq, &ctrl->ka_work, delay);
 }
 
 static enum rq_end_io_ret nvme_keep_alive_end_io(struct request *rq,
-- 
2.25.1

Re: [PATCH v3 1/3] nvme: double KA polling frequency to avoid KATO with TBKAS on

[Christoph Hellwig]

On Thu, May 18, 2023 at 12:33:09PM -0600, Uday Shankar wrote:
> + * 
> + * When TBKAS is on, we need to run nvme_keep_alive_work at twice this
> + * frequency, as one command completion can postpone sending a keep alive
> + * command by up to twice the delay between runs.
>   */
>  static void nvme_queue_keep_alive_work(struct nvme_ctrl *ctrl)
>  {
> -	queue_delayed_work(nvme_wq, &ctrl->ka_work, ctrl->kato * HZ / 2);
> +	unsigned long delay = (ctrl->ctratt & NVME_CTRL_ATTR_TBKAS) ?
> +		ctrl->kato * HZ / 4 : ctrl->kato * HZ / 2;
> +	queue_delayed_work(nvme_wq, &ctrl->ka_work, delay);

Kernel coding style wants an empty line after the variable declarations.

I also find the style rather hard to read.  Why not:

	unsigned long delay = ctrl->kato * HZ / 2;

	/*
	 * When using Traffic Based Keep Alive, we need to run
	 * nvme_keep_alive_work at twice the normal frequency, as one command
	 * completion can postpone sending a keep alive command by up to
	 * twice the delay between runs.
	 */
	if (ctrl->ctratt & NVME_CTRL_ATTR_TBKAS)
		delay /= 2;

	queue_delayed_work(nvme_wq, &ctrl->ka_work, delay);

That's not quite as dense, but relaly gets the point across to the
reader much better.

[PATCH v3 2/3] nvme: check IO start time when deciding to defer KA

[Uday Shankar]

When a command completes, we set a flag which will skip sending a
keep alive at the next run of nvme_keep_alive_work when TBKAS is on.
However, if the command was submitted long ago, it's possible that
the controller may have also restarted its keep alive timer (as a
result of receiving the command) long ago. The following trace
demonstrates the issue, assuming TBKAS is on and KATO = 8 for
simplicity:

1. t = 0: submit I/O commands A, B, C, D, E
2. t = 0.5: commands A, B, C, D, E reach controller, restart its keep
            alive timer
3. t = 1: A completes
4. t = 2: run nvme_keep_alive_work, see recent completion, do nothing
5. t = 3: B completes
6. t = 4: run nvme_keep_alive_work, see recent completion, do nothing
7. t = 5: C completes
8. t = 6: run nvme_keep_alive_work, see recent completion, do nothing
9. t = 7: D completes
10. t = 8: run nvme_keep_alive_work, see recent completion, do nothing
11. t = 9: E completes

At this point, 8.5 seconds have passed without restarting the
controller's keep alive timer, so the controller will detect a keep
alive timeout.

Fix this by checking the IO start time when deciding to defer sending a
keep alive command. Only set comp_seen if the command started after the
most recent run of nvme_keep_alive_work. With this change, the
completions of B, C, and D will not set comp_seen and the run of
nvme_keep_alive_work at t = 4 will send a keep alive.

Reported-by: Costa Sapuntzakis <costa@purestorage.com>
Reported-by: Randy Jennings <randyj@purestorage.com>
Signed-off-by: Uday Shankar <ushankar@purestorage.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
---
 drivers/nvme/host/core.c | 14 +++++++++++++-
 drivers/nvme/host/nvme.h |  4 ++--
 2 files changed, 15 insertions(+), 3 deletions(-)

diff --git a/drivers/nvme/host/core.c b/drivers/nvme/host/core.c
index 7726277b951a..a31c04b5f849 100644
--- a/drivers/nvme/host/core.c
+++ b/drivers/nvme/host/core.c
@@ -397,7 +397,14 @@ void nvme_complete_rq(struct request *req)
 	trace_nvme_complete_rq(req);
 	nvme_cleanup_cmd(req);
 
-	if (ctrl->kas)
+	/*
+	 * Completions of long-running commands should not be able to
+	 * defer sending of periodic keep alives, since the controller
+	 * may have completed processing such commands a long time ago
+	 * (arbitrarily close to command submission time).
+	 */
+	if (ctrl->kas && !ctrl->comp_seen &&
+	    nvme_req(req)->start_time >= ctrl->ka_last_check_time)
 		ctrl->comp_seen = true;
 
 	switch (nvme_decide_disposition(req)) {
@@ -1188,6 +1195,8 @@ static enum rq_end_io_ret nvme_keep_alive_end_io(struct request *rq,
 		return RQ_END_IO_NONE;
 	}
 
+	WRITE_ONCE(ctrl->ka_last_check_time, jiffies);
+	smp_wmb();
 	ctrl->comp_seen = false;
 	spin_lock_irqsave(&ctrl->lock, flags);
 	if (ctrl->state == NVME_CTRL_LIVE ||
@@ -1206,6 +1215,9 @@ static void nvme_keep_alive_work(struct work_struct *work)
 	bool comp_seen = ctrl->comp_seen;
 	struct request *rq;
 
+	WRITE_ONCE(ctrl->ka_last_check_time, jiffies);
+	smp_wmb();
+
 	if ((ctrl->ctratt & NVME_CTRL_ATTR_TBKAS) && comp_seen) {
 		dev_dbg(ctrl->device,
 			"reschedule traffic based keep-alive timer\n");
diff --git a/drivers/nvme/host/nvme.h b/drivers/nvme/host/nvme.h
index f0a84e390a55..07ed39ea43e3 100644
--- a/drivers/nvme/host/nvme.h
+++ b/drivers/nvme/host/nvme.h
@@ -162,9 +162,7 @@ struct nvme_request {
 	u8			retries;
 	u8			flags;
 	u16			status;
-#ifdef CONFIG_NVME_MULTIPATH
 	unsigned long		start_time;
-#endif
 	struct nvme_ctrl	*ctrl;
 };
 
@@ -323,6 +321,7 @@ struct nvme_ctrl {
 	struct delayed_work ka_work;
 	struct delayed_work failfast_work;
 	struct nvme_command ka_cmd;
+	unsigned long ka_last_check_time;
 	struct work_struct fw_act_work;
 	unsigned long events;
 
@@ -1032,6 +1031,7 @@ static inline void nvme_start_request(struct request *rq)
 {
 	if (rq->cmd_flags & REQ_NVME_MPATH)
 		nvme_mpath_start_request(rq);
+	nvme_req(rq)->start_time = jiffies;
 	blk_mq_start_request(rq);
 }
 
-- 
2.25.1

Re: [PATCH v3 2/3] nvme: check IO start time when deciding to defer KA

[Christoph Hellwig]

On Thu, May 18, 2023 at 12:33:10PM -0600, Uday Shankar wrote:
> +	WRITE_ONCE(ctrl->ka_last_check_time, jiffies);
> +	smp_wmb();
>  	ctrl->comp_seen = false;
>  	spin_lock_irqsave(&ctrl->lock, flags);
>  	if (ctrl->state == NVME_CTRL_LIVE ||
> @@ -1206,6 +1215,9 @@ static void nvme_keep_alive_work(struct work_struct *work)
>  	bool comp_seen = ctrl->comp_seen;
>  	struct request *rq;
>  
> +	WRITE_ONCE(ctrl->ka_last_check_time, jiffies);
> +	smp_wmb();

WRITE_ONCE only works when paired with READ_ONCE on the read size,
which I don't ee here.

smp_wmb always pairs with smp_rmb, which I don't see here either.  We
also require documentation explaining the pairing because it tends to
be non-obvious (unlike READ_ONCE/WRITE_ONCE that are paired to a
variable).  I also can't really see a reason for the barrier here,
but the comment would be the thing to convince the reviewer.

Re: [PATCH v3 2/3] nvme: check IO start time when deciding to defer KA

[Uday Shankar]

On Sat, May 20, 2023 at 06:34:22AM +0200, Christoph Hellwig wrote:
> I also can't really see a reason for the barrier here

In this version of the patch, I have this bit in nvme_complete_rq:

+	if (ctrl->kas && !ctrl->comp_seen &&
+	    nvme_req(req)->start_time >= ctrl->ka_last_check_time)

and the code that updates ka_last_check_time looks like this:

+	WRITE_ONCE(ctrl->ka_last_check_time, jiffies);
+	smp_wmb();
 	ctrl->comp_seen = false;

I used WRITE_ONCE and smp_wmb here to try to avoid reordering of writes
here, as I didn't want nvme_complete_rq to be able to observe comp_seen
as false while ka_last_check_time still has its pre-update value. I
agree if I wanted to do this properly, I need to add a smb_rmb and
READ_ONCE(ka_last_check_time) in nvme_complete_rq.

However, the read of comp_seen here is a (probably premature) attempt at
optimization, which may well be completely unnoticeable. I removed it in
v4, and as a result the READ/WRITE_ONCE and barriers are no longer
needed.

[PATCH v3 3/3] nvme: improve handling of long keep alives

[Uday Shankar]

Upon keep alive completion, nvme_keep_alive_work is scheduled with the
same delay every time. If keep alive commands are completing slowly,
this may cause a keep alive timeout. The following trace illustrates the
issue, taking KATO = 8 and TBKAS off for simplicity:

1. t = 0: run nvme_keep_alive_work, send keep alive
2. t = ε: keep alive reaches controller, controller restarts its keep
          alive timer
3. t = 4: host receives keep alive completion, schedules
          nvme_keep_alive_work with delay 4
4. t = 8: run nvme_keep_alive_work, send keep alive

Here, a keep alive having RTT of 4 causes a delay of at least 8 - ε
between the controller receiving successive keep alives. With ε small,
the controller is likely to detect a keep alive timeout.

Fix this by calculating the RTT of the keep alive command, and adjusting
the scheduling delay of the next keep alive work accordingly.

Reported-by: Costa Sapuntzakis <costa@purestorage.com>
Reported-by: Randy Jennings <randyj@purestorage.com>
Signed-off-by: Uday Shankar <ushankar@purestorage.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
---
 drivers/nvme/host/core.c | 23 ++++++++++++++++++++---
 1 file changed, 20 insertions(+), 3 deletions(-)

diff --git a/drivers/nvme/host/core.c b/drivers/nvme/host/core.c
index a31c04b5f849..ce07218fc6eb 100644
--- a/drivers/nvme/host/core.c
+++ b/drivers/nvme/host/core.c
@@ -1172,10 +1172,15 @@ EXPORT_SYMBOL_NS_GPL(nvme_passthru_end, NVME_TARGET_PASSTHRU);
  * frequency, as one command completion can postpone sending a keep alive
  * command by up to twice the delay between runs.
  */
+static unsigned long nvme_keep_alive_work_period(struct nvme_ctrl *ctrl)
+{
+	return (ctrl->ctratt & NVME_CTRL_ATTR_TBKAS) ?
+		(ctrl->kato * HZ / 4) : (ctrl->kato * HZ / 2);
+}
+
 static void nvme_queue_keep_alive_work(struct nvme_ctrl *ctrl)
 {
-	unsigned long delay = (ctrl->ctratt & NVME_CTRL_ATTR_TBKAS) ?
-		ctrl->kato * HZ / 4 : ctrl->kato * HZ / 2;
+	unsigned long delay = nvme_keep_alive_work_period(ctrl);
 	queue_delayed_work(nvme_wq, &ctrl->ka_work, delay);
 }
 
@@ -1185,6 +1190,18 @@ static enum rq_end_io_ret nvme_keep_alive_end_io(struct request *rq,
 	struct nvme_ctrl *ctrl = rq->end_io_data;
 	unsigned long flags;
 	bool startka = false;
+	unsigned long rtt = jiffies - nvme_req(rq)->start_time;
+	unsigned long delay = nvme_keep_alive_work_period(ctrl);
+
+	/* Subtract off the keepalive RTT so nvme_keep_alive_work runs
+	 * at the desired frequency. */
+	if (rtt <= delay) {
+		delay -= rtt;
+	} else {
+		dev_warn(ctrl->device, "long keepalive RTT (%u ms)\n",
+			 jiffies_to_msecs(rtt));
+		delay = 0;
+	}
 
 	blk_mq_free_request(rq);
 
@@ -1204,7 +1221,7 @@ static enum rq_end_io_ret nvme_keep_alive_end_io(struct request *rq,
 		startka = true;
 	spin_unlock_irqrestore(&ctrl->lock, flags);
 	if (startka)
-		nvme_queue_keep_alive_work(ctrl);
+		queue_delayed_work(nvme_wq, &ctrl->ka_work, delay);
 	return RQ_END_IO_NONE;
 }
 
-- 
2.25.1

Re: [PATCH v3 3/3] nvme: improve handling of long keep alives

[Christoph Hellwig]

On Thu, May 18, 2023 at 12:33:11PM -0600, Uday Shankar wrote:
> Upon keep alive completion, nvme_keep_alive_work is scheduled with the
> same delay every time. If keep alive commands are completing slowly,
> this may cause a keep alive timeout. The following trace illustrates the
> issue, taking KATO = 8 and TBKAS off for simplicity:
> 
> 1. t = 0: run nvme_keep_alive_work, send keep alive
> 2. t = ε: keep alive reaches controller, controller restarts its keep
>           alive timer
> 3. t = 4: host receives keep alive completion, schedules
>           nvme_keep_alive_work with delay 4
> 4. t = 8: run nvme_keep_alive_work, send keep alive
> 
> Here, a keep alive having RTT of 4 causes a delay of at least 8 - ε
> between the controller receiving successive keep alives. With ε small,
> the controller is likely to detect a keep alive timeout.
> 
> Fix this by calculating the RTT of the keep alive command, and adjusting
> the scheduling delay of the next keep alive work accordingly.
> 
> Reported-by: Costa Sapuntzakis <costa@purestorage.com>
> Reported-by: Randy Jennings <randyj@purestorage.com>
> Signed-off-by: Uday Shankar <ushankar@purestorage.com>
> Reviewed-by: Hannes Reinecke <hare@suse.de>
> ---
>  drivers/nvme/host/core.c | 23 ++++++++++++++++++++---
>  1 file changed, 20 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/nvme/host/core.c b/drivers/nvme/host/core.c
> index a31c04b5f849..ce07218fc6eb 100644
> --- a/drivers/nvme/host/core.c
> +++ b/drivers/nvme/host/core.c
> @@ -1172,10 +1172,15 @@ EXPORT_SYMBOL_NS_GPL(nvme_passthru_end, NVME_TARGET_PASSTHRU);
>   * frequency, as one command completion can postpone sending a keep alive
>   * command by up to twice the delay between runs.
>   */
> +static unsigned long nvme_keep_alive_work_period(struct nvme_ctrl *ctrl)
> +{
> +	return (ctrl->ctratt & NVME_CTRL_ATTR_TBKAS) ?
> +		(ctrl->kato * HZ / 4) : (ctrl->kato * HZ / 2);
> +}

Please add this separate helper in patch 1 already instead of moving
things around.  Preferably in the style I mentioned there.

>  static void nvme_queue_keep_alive_work(struct nvme_ctrl *ctrl)
>  {
> -	unsigned long delay = (ctrl->ctratt & NVME_CTRL_ATTR_TBKAS) ?
> -		ctrl->kato * HZ / 4 : ctrl->kato * HZ / 2;
> +	unsigned long delay = nvme_keep_alive_work_period(ctrl);
>  	queue_delayed_work(nvme_wq, &ctrl->ka_work, delay);

.. and with this helper we don't really need the local variable here.

> +	/* Subtract off the keepalive RTT so nvme_keep_alive_work runs
> +	 * at the desired frequency. */

The kernel comment style is:

	/*
	 * Subtract off the keepalive RTT so nvme_keep_alive_work runs
	 * at the desired frequency.
	 */