Skip to contentSkip to conversation

remoteproc: core: do pm relax when not first crash

PATCH In review Maria Yu 4 participants 36 messages Sep 9, 2022 – Dec 2, 2022

Subsystems (2)
lore
Share & tools
lore

Maria Yu

author
#0 lore raw

PATCH 1/1 remoteproc: core: do pm relax when not first crash

Even if it is not first crash, need to relax the pm wakelock otherwise the device will stay awake.

---
 drivers/remoteproc/remoteproc_core.c | 1 +
 1 file changed, 1 insertion(+)
drivers/remoteproc/remoteproc_core.c +1 -0
index e5279ed9a8d7..30078043e939 100644
@@ -1956,6 +1956,7 @@ static void rproc_crash_handler_work(struct work_struct *work)
 	if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
 		/* handle only the first crash detected */
 		mutex_unlock(&rproc->lock);
+		pm_relax(rproc->dev.parent);
 		return;
 	}
 
Signature · 1 line
2.7.4
maintainer
#1 lore raw

Hi Maria,

Quoted 3 lines · Maria Yu

Even if it is not first crash, need to relax the pm wakelock otherwise the device will stay awake.

The goal is exactly to keep the device awake...

Quoted 10 lines
---
 drivers/remoteproc/remoteproc_core.c | 1 +
 1 file changed, 1 insertion(+)
drivers/remoteproc/remoteproc_core.c +0 -0
index e5279ed9a8d7..30078043e939 100644

drivers/remoteproc/remoteproc_core.c static void rproc_crash_handler_work(struct work_struct *work) · line 1959

Quoted 4 lines
drivers/remoteproc/remoteproc_core.c +1 -0
@@ -1956,6 +1956,7 @@ static void rproc_crash_handler_work(struct work_struct *work)
 	if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
 		/* handle only the first crash detected */
 		mutex_unlock(&rproc->lock);
+		pm_relax(rproc->dev.parent);

If we are here it means that rproc_crash_handler_work() has already been called _and_ that a recovery is in process. When the first crash handler completes pm_relax() will be called and the device will go to sleep as expected.

Thanks,
Mathieu

Quoted 6 lines
drivers/remoteproc/remoteproc_core.c +0 -0
 		return;
 	}
 
Signature · 1 line
2.7.4
#2 lore raw

Hi Mathieu,

pm_awake and pm_relax needed to be used as a pair. There is chance that pm_relax is not being called, and make the device always in cannot suspend state.

Quoted 14 lines · Mathieu Poirier

Hi Maria,

Quoted 3 lines · Maria Yu

Even if it is not first crash, need to relax the pm wakelock otherwise the device will stay awake.

The goal is exactly to keep the device awake...

Quoted 5 lines
---
  drivers/remoteproc/remoteproc_core.c | 1 +
  1 file changed, 1 insertion(+)
Quoted 13 lines

drivers/remoteproc/remoteproc_core.c static void rproc_crash_handler_work(struct work_struct *work) · line 1959

Quoted 9 lines
drivers/remoteproc/remoteproc_core.c +1 -0
index e5279ed9a8d7..30078043e939 100644
@@ -1956,6 +1956,7 @@ static void rproc_crash_handler_work(struct work_struct *work)
  	if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
  		/* handle only the first crash detected */
  		mutex_unlock(&rproc->lock);
+		pm_relax(rproc->dev.parent);

If we are here it means that rproc_crash_handler_work() has already been called _and_ that a recovery is in process. When the first crash handler completes pm_relax() will be called and the device will go to sleep as expected.

If the rproc->state cannot be changed to running state, the device will always be awake from this return. Also APROC_OFFLINE state can be given in other path like an shutdown request is issued.

While this patch is not considering carefully as well, I think I need to upload a new patchset with an ordered workqueue to make each work have each pm_relax before return. what do you think?

Quoted 10 lines

Thanks,
Mathieu

Quoted 6 lines
drivers/remoteproc/remoteproc_core.c +0 -0
  		return;
  	}
 
Signature · 1 line
2.7.4
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
author
#3 lore raw

PATCH v2 1/1 remoteproc: core: do pm relax when in RPROC_OFFLINE

RPROC_OFFLINE state indicate there is no recovery process is in progress and no chance to do the pm_relax. Because when recovering from crash, rproc->lock is hold and state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING, and then unlock rproc->lock.
When the state is in RPROC_OFFLINE it means separate request of rproc_stop was done and no need to hold the wakeup source in crash handler to recover any more.

---
 drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
 1 file changed, 11 insertions(+)
drivers/remoteproc/remoteproc_core.c +11 -0
index e5279ed9a8d7..247ced6b0655 100644
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct work_struct *work)
 	if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
 		/* handle only the first crash detected */
 		mutex_unlock(&rproc->lock);
+		/*
+		 * RPROC_OFFLINE state indicate there is no recovery process
+		 * is in progress and no chance to have pm_relax in place.
+		 * Because when recovering from crash, rproc->lock is hold and
+		 * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+		 * and then unlock rproc->lock.
+		 * RPROC_OFFLINE is only an intermediate state in recovery
+		 * process.
+		 */
+		if (rproc->state == RPROC_OFFLINE)
+			pm_relax(rproc->dev.parent);
 		return;
 	}
 
Signature · 1 line
2.7.4
Reviewed-by on v2 1/1
#4 lore raw

Re: [PATCH v2] remoteproc: core: do pm relax when in RPROC_OFFLINE

Hi,

Quoted 3 lines · Maria Yu

RPROC_OFFLINE state indicate there is no recovery process is in progress and no chance to do the pm_relax. Because when recovering from crash, rproc->lock is hold and

s/hold/held ?

Quoted 16 lines

state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING, and then unlock rproc->lock.
When the state is in RPROC_OFFLINE it means separate request of rproc_stop was done and no need to hold the wakeup source in crash handler to recover any more.

---
  drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
  1 file changed, 11 insertions(+)
drivers/remoteproc/remoteproc_core.c +0 -0
index e5279ed9a8d7..247ced6b0655 100644

drivers/remoteproc/remoteproc_core.c static void rproc_crash_handler_work(struct work_struct *work) · line 1962

Quoted 7 lines
drivers/remoteproc/remoteproc_core.c +4 -0
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct work_struct *work)
  	if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
  		/* handle only the first crash detected */
  		mutex_unlock(&rproc->lock);
+		/*
+		 * RPROC_OFFLINE state indicate there is no recovery process
+		 * is in progress and no chance to have pm_relax in place.
+		 * Because when recovering from crash, rproc->lock is hold and

..ditto..

drivers/remoteproc/remoteproc_core.c static void rproc_crash_handler_work(struct work_struct *work) · line 1969

Quoted 7 lines
drivers/remoteproc/remoteproc_core.c +7 -0
+		 * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+		 * and then unlock rproc->lock.
+		 * RPROC_OFFLINE is only an intermediate state in recovery
+		 * process.
+		 */
+		if (rproc->state == RPROC_OFFLINE)
+			pm_relax(rproc->dev.parent);

Looks good, apart from minor nit.

-Mukesh

Quoted 3 lines
drivers/remoteproc/remoteproc_core.c +0 -0
  		return;
  	}
 
#5 lore raw

Re: [PATCH v2] remoteproc: core: do pm relax when in RPROC_OFFLINE

Quoted 9 lines · Mukesh Ojha

Hi,

Quoted 3 lines · Maria Yu

RPROC_OFFLINE state indicate there is no recovery process is in progress and no chance to do the pm_relax. Because when recovering from crash, rproc->lock is hold and

s/hold/held ?

Thanks.

Quoted 36 lines
Quoted 16 lines

state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING, and then unlock rproc->lock.
When the state is in RPROC_OFFLINE it means separate request of rproc_stop was done and no need to hold the wakeup source in crash handler to recover any more.

---
  drivers/remoteproc/remoteproc_core.c | 11 +++++++++++

  1 file changed, 11 insertions(+)
drivers/remoteproc/remoteproc_core.c +0 -0
index e5279ed9a8d7..247ced6b0655 100644

drivers/remoteproc/remoteproc_core.c static void rproc_crash_handler_work(struct · line 1963

Quoted 8 lines
drivers/remoteproc/remoteproc_core.c +4 -0
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct 
 work_struct *work)
       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
           /* handle only the first crash detected */
           mutex_unlock(&rproc->lock);
+        /*
+         * RPROC_OFFLINE state indicate there is no recovery process
+         * is in progress and no chance to have pm_relax in place.
+         * Because when recovering from crash, rproc->lock is hold and

..ditto..

drivers/remoteproc/remoteproc_core.c static void rproc_crash_handler_work(struct work_struct *work) · line 1969

Quoted 7 lines
drivers/remoteproc/remoteproc_core.c +7 -0
+         * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+         * and then unlock rproc->lock.
+         * RPROC_OFFLINE is only an intermediate state in recovery
+         * process.
+         */
+        if (rproc->state == RPROC_OFFLINE)
+            pm_relax(rproc->dev.parent);
Quoted 1 line

Looks good, apart from minor nit.

:)

Quoted 6 lines

-Mukesh

Quoted 2 lines
drivers/remoteproc/remoteproc_core.c +0 -0
           return;
       }
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
author
#6 lore raw

PATCH v3 1/1 remoteproc: core: do pm relax when in RPROC_OFFLINE

In reply to Aiqun(Maria) Yu

RPROC_OFFLINE state indicate there is no recovery process is in progress and no chance to do the pm_relax. Because when recovering from crash, rproc->lock is held and state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING, and then unlock rproc->lock.
When the state is in RPROC_OFFLINE it means separate request of rproc_stop was done and no need to hold the wakeup source in crash handler to recover any more.

---
 drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
 1 file changed, 11 insertions(+)
drivers/remoteproc/remoteproc_core.c +11 -0
index e5279ed9a8d7..247ced6b0655 100644
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct work_struct *work)
 	if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
 		/* handle only the first crash detected */
 		mutex_unlock(&rproc->lock);
+		/*
+		 * RPROC_OFFLINE state indicate there is no recovery process
+		 * is in progress and no chance to have pm_relax in place.
+		 * Because when recovering from crash, rproc->lock is hold and
+		 * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+		 * and then unlock rproc->lock.
+		 * RPROC_OFFLINE is only an intermediate state in recovery
+		 * process.
+		 */
+		if (rproc->state == RPROC_OFFLINE)
+			pm_relax(rproc->dev.parent);
 		return;
 	}
 
Signature · 1 line
2.7.4
author
#7 lore raw

PATCH v4 1/1 remoteproc: core: do pm relax when in RPROC_OFFLINE

In reply to Aiqun(Maria) Yu

RPROC_OFFLINE state indicate there is no recovery process is in progress and no chance to do the pm_relax. Because when recovering from crash, rproc->lock is held and state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING, and then unlock rproc->lock.
When the state is in RPROC_OFFLINE it means separate request of rproc_stop was done and no need to hold the wakeup source in crash handler to recover any more.

---
 drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
 1 file changed, 11 insertions(+)
drivers/remoteproc/remoteproc_core.c +11 -0
index e5279ed9a8d7..6bc7b8b7d01e 100644
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct work_struct *work)
 	if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
 		/* handle only the first crash detected */
 		mutex_unlock(&rproc->lock);
+		/*
+		 * RPROC_OFFLINE state indicate there is no recovery process
+		 * is in progress and no chance to have pm_relax in place.
+		 * Because when recovering from crash, rproc->lock is held and
+		 * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+		 * and then unlock rproc->lock.
+		 * RPROC_OFFLINE is only an intermediate state in recovery
+		 * process.
+		 */
+		if (rproc->state == RPROC_OFFLINE)
+			pm_relax(rproc->dev.parent);
 		return;
 	}
 
Signature · 1 line
2.7.4
Reviewed-by on v4 1/1
#8 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

Hi,

Quoted 23 lines · Maria Yu

RPROC_OFFLINE state indicate there is no recovery process is in progress and no chance to do the pm_relax. Because when recovering from crash, rproc->lock is held and state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING, and then unlock rproc->lock.
When the state is in RPROC_OFFLINE it means separate request of rproc_stop was done and no need to hold the wakeup source in crash handler to recover any more.

---
  drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
  1 file changed, 11 insertions(+)
drivers/remoteproc/remoteproc_core.c +1 -0
index e5279ed9a8d7..6bc7b8b7d01e 100644
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct work_struct *work)
  	if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
  		/* handle only the first crash detected */
  		mutex_unlock(&rproc->lock);
+		/*

drivers/remoteproc/remoteproc_core.c static void rproc_crash_handler_work(struct work_struct *work) · line 1969

Quoted 10 lines
drivers/remoteproc/remoteproc_core.c +10 -0
+		 * RPROC_OFFLINE state indicate there is no recovery process
+		 * is in progress and no chance to have pm_relax in place.
+		 * Because when recovering from crash, rproc->lock is held and
+		 * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+		 * and then unlock rproc->lock.
+		 * RPROC_OFFLINE is only an intermediate state in recovery
+		 * process.
+		 */
+		if (rproc->state == RPROC_OFFLINE)
+			pm_relax(rproc->dev.parent);

LGTM.

-Mukesh

Quoted 3 lines
drivers/remoteproc/remoteproc_core.c +0 -0
  		return;
  	}
 
maintainer
#9 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

In reply to Maria Yu

Please add what has changed from one version to another, either in a cover letter or after the "Signed-off-by". There are many examples on how to do that on the mailing list.

Quoted 5 lines · Maria Yu

RPROC_OFFLINE state indicate there is no recovery process is in progress and no chance to do the pm_relax. Because when recovering from crash, rproc->lock is held and state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING, and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to RPROC_RUNNING when rproc_trigger_recovery() returns. If that is not the case then something went wrong.

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before returning, so we know the remote processor was stopped. Therefore if rproc->state is set to RPROC_OFFLINE something went wrong in either request_firmware() or rproc_start(). Either way the remote processor is offline and the system probably in an unknown/unstable. As such I don't see how calling pm_relax() can help things along.

I suggest spending time understanding what leads to the failure when recovering from a crash and address that problem(s).

Thanks,
Mathieu

Quoted 34 lines

When the state is in RPROC_OFFLINE it means separate request of rproc_stop was done and no need to hold the wakeup source in crash handler to recover any more.

---
 drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
 1 file changed, 11 insertions(+)
drivers/remoteproc/remoteproc_core.c +11 -0
index e5279ed9a8d7..6bc7b8b7d01e 100644
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct work_struct *work)
 	if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
 		/* handle only the first crash detected */
 		mutex_unlock(&rproc->lock);
+		/*
+		 * RPROC_OFFLINE state indicate there is no recovery process
+		 * is in progress and no chance to have pm_relax in place.
+		 * Because when recovering from crash, rproc->lock is held and
+		 * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+		 * and then unlock rproc->lock.
+		 * RPROC_OFFLINE is only an intermediate state in recovery
+		 * process.
+		 */
+		if (rproc->state == RPROC_OFFLINE)
+			pm_relax(rproc->dev.parent);
 		return;
 	}
 
Signature · 1 line
2.7.4
#10 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

Hi Mathieu,

Quoted 4 lines · Mathieu Poirier

Please add what has changed from one version to another, either in a cover letter or after the "Signed-off-by". There are many examples on how to do that on the mailing list.

Thx for the information, will take a note and benefit for next time.

Quoted 11 lines · Maria Yu
Quoted 5 lines

RPROC_OFFLINE state indicate there is no recovery process is in progress and no chance to do the pm_relax. Because when recovering from crash, rproc->lock is held and state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING, and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to RPROC_RUNNING when rproc_trigger_recovery() returns. If that is not the case then something went wrong.

Quoted 7 lines

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before returning, so we know the remote processor was stopped. Therefore if rproc->state is set to RPROC_OFFLINE something went wrong in either request_firmware() or rproc_start(). Either way the remote processor is offline and the system probably in an unknown/unstable. As such I don't see how calling pm_relax() can help things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and last rproc_trigger_recovery bailed out with only rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair. Since the subsystem may still can be recovered with customer's next trigger of rproc_start, and we can make each error out path clean with pm resources.

Quoted 3 lines

I suggest spending time understanding what leads to the failure when recovering from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when rproc_shutdown is triggered at similar time. So not an issue from error out of rproc_trigger_recovery.

Quoted 38 lines

Thanks,
Mathieu

Quoted 34 lines

When the state is in RPROC_OFFLINE it means separate request of rproc_stop was done and no need to hold the wakeup source in crash handler to recover any more.

---
  drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
  1 file changed, 11 insertions(+)
drivers/remoteproc/remoteproc_core.c +11 -0
index e5279ed9a8d7..6bc7b8b7d01e 100644
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct work_struct *work)
  	if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
  		/* handle only the first crash detected */
  		mutex_unlock(&rproc->lock);
+		/*
+		 * RPROC_OFFLINE state indicate there is no recovery process
+		 * is in progress and no chance to have pm_relax in place.
+		 * Because when recovering from crash, rproc->lock is held and
+		 * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+		 * and then unlock rproc->lock.
+		 * RPROC_OFFLINE is only an intermediate state in recovery
+		 * process.
+		 */
+		if (rproc->state == RPROC_OFFLINE)
+			pm_relax(rproc->dev.parent);
  		return;
  	}
 
Signature · 1 line
2.7.4
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
maintainer
#11 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

Quoted 35 lines · Aiqun(Maria) Yu

Hi Mathieu,

Quoted 4 lines · Mathieu Poirier

Please add what has changed from one version to another, either in a cover letter or after the "Signed-off-by". There are many examples on how to do that on the mailing list.

Thx for the information, will take a note and benefit for next time.

Quoted 11 lines · Maria Yu
Quoted 5 lines

RPROC_OFFLINE state indicate there is no recovery process is in progress and no chance to do the pm_relax. Because when recovering from crash, rproc->lock is held and state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING, and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to RPROC_RUNNING when rproc_trigger_recovery() returns. If that is not the case then something went wrong.

Quoted 7 lines

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before returning, so we know the remote processor was stopped. Therefore if rproc->state is set to RPROC_OFFLINE something went wrong in either request_firmware() or rproc_start(). Either way the remote processor is offline and the system probably in an unknown/unstable. As such I don't see how calling pm_relax() can help things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and last rproc_trigger_recovery bailed out with only rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair. Since the subsystem may still can be recovered with customer's next trigger of rproc_start, and we can make each error out path clean with pm resources.

Quoted 6 lines
Quoted 3 lines

I suggest spending time understanding what leads to the failure when recovering from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when rproc_shutdown is triggered at similar time. So not an issue from error out of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned from the beginning. What I see happening is the following:

rproc_report_crash()
pm_stay_awake()
queue_work() // current thread is suspended

rproc_shutdown()

        rproc_stop()
                rproc->state = RPROC_OFFLINE;

rproc_crash_handler_work()
if (rproc->state == RPROC_OFFLINE)
return // pm_relax() is not called

The right way to fix this is to add a pm_relax() in rproc_shutdown() and rproc_detach(), along with a very descriptive comment as to why it is needed.

Quoted 42 lines
Quoted 37 lines

Thanks,
Mathieu

Quoted 33 lines

When the state is in RPROC_OFFLINE it means separate request of rproc_stop was done and no need to hold the wakeup source in crash handler to recover any more.

---
  drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
  1 file changed, 11 insertions(+)
drivers/remoteproc/remoteproc_core.c +11 -0
index e5279ed9a8d7..6bc7b8b7d01e 100644
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct work_struct *work)
  	if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
  		/* handle only the first crash detected */
  		mutex_unlock(&rproc->lock);
+		/*
+		 * RPROC_OFFLINE state indicate there is no recovery process
+		 * is in progress and no chance to have pm_relax in place.
+		 * Because when recovering from crash, rproc->lock is held and
+		 * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+		 * and then unlock rproc->lock.
+		 * RPROC_OFFLINE is only an intermediate state in recovery
+		 * process.
+		 */
+		if (rproc->state == RPROC_OFFLINE)
+			pm_relax(rproc->dev.parent);
  		return;
  	}
Signature · 1 line
2.7.4
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
maintainer
#12 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

Quoted 54 lines · Mathieu Poirier
Quoted 35 lines · Aiqun(Maria) Yu

Hi Mathieu,

Quoted 4 lines · Mathieu Poirier

Please add what has changed from one version to another, either in a cover letter or after the "Signed-off-by". There are many examples on how to do that on the mailing list.

Thx for the information, will take a note and benefit for next time.

Quoted 11 lines · Maria Yu
Quoted 5 lines

RPROC_OFFLINE state indicate there is no recovery process is in progress and no chance to do the pm_relax. Because when recovering from crash, rproc->lock is held and state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING, and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to RPROC_RUNNING when rproc_trigger_recovery() returns. If that is not the case then something went wrong.

Quoted 7 lines

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before returning, so we know the remote processor was stopped. Therefore if rproc->state is set to RPROC_OFFLINE something went wrong in either request_firmware() or rproc_start(). Either way the remote processor is offline and the system probably in an unknown/unstable. As such I don't see how calling pm_relax() can help things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and last rproc_trigger_recovery bailed out with only rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair. Since the subsystem may still can be recovered with customer's next trigger of rproc_start, and we can make each error out path clean with pm resources.

Quoted 6 lines
Quoted 3 lines

I suggest spending time understanding what leads to the failure when recovering from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when rproc_shutdown is triggered at similar time. So not an issue from error out of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned from the beginning. What I see happening is the following:

rproc_report_crash()
pm_stay_awake()
queue_work() // current thread is suspended

rproc_shutdown()

        rproc_stop()
                rproc->state = RPROC_OFFLINE;
Quoted 6 lines

rproc_crash_handler_work()
if (rproc->state == RPROC_OFFLINE)
return // pm_relax() is not called

The right way to fix this is to add a pm_relax() in rproc_shutdown() and rproc_detach(), along with a very descriptive comment as to why it is needed.

Thinking about this further there are more ramifications to consider. Please confirm the above scenario is what you are facing. I will advise on how to move forward if that is the case.

Quoted 44 lines
Quoted 44 lines
Quoted 39 lines

Thanks,
Mathieu

Quoted 33 lines

When the state is in RPROC_OFFLINE it means separate request of rproc_stop was done and no need to hold the wakeup source in crash handler to recover any more.

---
  drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
  1 file changed, 11 insertions(+)
drivers/remoteproc/remoteproc_core.c +11 -0
index e5279ed9a8d7..6bc7b8b7d01e 100644
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct work_struct *work)
  	if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
  		/* handle only the first crash detected */
  		mutex_unlock(&rproc->lock);
+		/*
+		 * RPROC_OFFLINE state indicate there is no recovery process
+		 * is in progress and no chance to have pm_relax in place.
+		 * Because when recovering from crash, rproc->lock is held and
+		 * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+		 * and then unlock rproc->lock.
+		 * RPROC_OFFLINE is only an intermediate state in recovery
+		 * process.
+		 */
+		if (rproc->state == RPROC_OFFLINE)
+			pm_relax(rproc->dev.parent);
  		return;
  	}
Signature · 1 line
2.7.4
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
#13 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

Quoted 59 lines · Mathieu Poirier
Quoted 58 lines · Mathieu Poirier
Quoted 35 lines · Aiqun(Maria) Yu

Hi Mathieu,

Quoted 4 lines · Mathieu Poirier

Please add what has changed from one version to another, either in a cover letter or after the "Signed-off-by". There are many examples on how to do that on the mailing list.

Thx for the information, will take a note and benefit for next time.

Quoted 11 lines · Maria Yu
Quoted 5 lines

RPROC_OFFLINE state indicate there is no recovery process is in progress and no chance to do the pm_relax. Because when recovering from crash, rproc->lock is held and state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING, and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to RPROC_RUNNING when rproc_trigger_recovery() returns. If that is not the case then something went wrong.

Quoted 7 lines

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before returning, so we know the remote processor was stopped. Therefore if rproc->state is set to RPROC_OFFLINE something went wrong in either request_firmware() or rproc_start(). Either way the remote processor is offline and the system probably in an unknown/unstable. As such I don't see how calling pm_relax() can help things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and last rproc_trigger_recovery bailed out with only rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair. Since the subsystem may still can be recovered with customer's next trigger of rproc_start, and we can make each error out path clean with pm resources.

Quoted 6 lines
Quoted 3 lines

I suggest spending time understanding what leads to the failure when recovering from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when rproc_shutdown is triggered at similar time. So not an issue from error out of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned from the beginning. What I see happening is the following:

rproc_report_crash()
pm_stay_awake()
queue_work() // current thread is suspended

rproc_shutdown()

         rproc_stop()
                 rproc->state = RPROC_OFFLINE;

rproc_crash_handler_work()
if (rproc->state == RPROC_OFFLINE)
return // pm_relax() is not called

Quoted 7 lines
Quoted 2 lines

The right way to fix this is to add a pm_relax() in rproc_shutdown() and rproc_detach(), along with a very descriptive comment as to why it is needed.

Thinking about this further there are more ramifications to consider. Please confirm the above scenario is what you are facing. I will advise on how to move forward if that is the case.

Not sure if the situation is clear or not. So resend the email again.

The above senario is what customer is facing. crash hanppened while at the same time shutdown is triggered. And the device cannto goes to suspend state after that.
the subsystem can still be start normally after this.

Quoted 44 lines
Quoted 44 lines
Quoted 44 lines
Quoted 39 lines

Thanks,
Mathieu

Quoted 33 lines

When the state is in RPROC_OFFLINE it means separate request of rproc_stop was done and no need to hold the wakeup source in crash handler to recover any more.

---
   drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
   1 file changed, 11 insertions(+)
drivers/remoteproc/remoteproc_core.c +11 -0
index e5279ed9a8d7..6bc7b8b7d01e 100644
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct work_struct *work)
   	if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
   		/* handle only the first crash detected */
   		mutex_unlock(&rproc->lock);
+		/*
+		 * RPROC_OFFLINE state indicate there is no recovery process
+		 * is in progress and no chance to have pm_relax in place.
+		 * Because when recovering from crash, rproc->lock is held and
+		 * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+		 * and then unlock rproc->lock.
+		 * RPROC_OFFLINE is only an intermediate state in recovery
+		 * process.
+		 */
+		if (rproc->state == RPROC_OFFLINE)
+			pm_relax(rproc->dev.parent);
   		return;
   	}
Signature · 1 line
2.7.4
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
maintainer
#14 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

Quoted 64 lines · Aiqun(Maria) Yu
Quoted 62 lines · Mathieu Poirier
Quoted 60 lines · Mathieu Poirier
Quoted 35 lines · Aiqun(Maria) Yu

Hi Mathieu,

Quoted 4 lines · Mathieu Poirier

Please add what has changed from one version to another, either in a cover letter or after the "Signed-off-by". There are many examples on how to do that on the mailing list.

Thx for the information, will take a note and benefit for next time.

Quoted 11 lines · Maria Yu
Quoted 5 lines

RPROC_OFFLINE state indicate there is no recovery process is in progress and no chance to do the pm_relax. Because when recovering from crash, rproc->lock is held and state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING, and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to RPROC_RUNNING when rproc_trigger_recovery() returns. If that is not the case then something went wrong.

Quoted 7 lines

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before returning, so we know the remote processor was stopped. Therefore if rproc->state is set to RPROC_OFFLINE something went wrong in either request_firmware() or rproc_start(). Either way the remote processor is offline and the system probably in an unknown/unstable. As such I don't see how calling pm_relax() can help things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and last rproc_trigger_recovery bailed out with only rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair. Since the subsystem may still can be recovered with customer's next trigger of rproc_start, and we can make each error out path clean with pm resources.

Quoted 6 lines
Quoted 3 lines

I suggest spending time understanding what leads to the failure when recovering from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when rproc_shutdown is triggered at similar time. So not an issue from error out of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned from the beginning. What I see happening is the following:

rproc_report_crash()
pm_stay_awake()
queue_work() // current thread is suspended

rproc_shutdown()

         rproc_stop()
                 rproc->state = RPROC_OFFLINE;

rproc_crash_handler_work()
if (rproc->state == RPROC_OFFLINE)
return // pm_relax() is not called

The right way to fix this is to add a pm_relax() in rproc_shutdown() and rproc_detach(), along with a very descriptive comment as to why it is needed.

Quoted 8 lines
Quoted 4 lines

Thinking about this further there are more ramifications to consider. Please confirm the above scenario is what you are facing. I will advise on how to move forward if that is the case.

Not sure if the situation is clear or not. So resend the email again.

The above senario is what customer is facing. crash hanppened while at the same time shutdown is triggered.

Unfortunately this is not enough details to address a problem as complex as this one.

Quoted 2 lines

And the device cannto goes to suspend state after that. the subsystem can still be start normally after this.

If the code flow I pasted above reflects the problem at hand, the current patch will not be sufficient to address the issue. If Arnaud confirms my suspicions we will have to think about a better solution.

Quoted 50 lines
Quoted 37 lines
Quoted 37 lines
Quoted 37 lines
Quoted 37 lines

Thanks,
Mathieu

Quoted 30 lines

When the state is in RPROC_OFFLINE it means separate request of rproc_stop was done and no need to hold the wakeup source in crash handler to recover any more.

---
   drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
   1 file changed, 11 insertions(+)
drivers/remoteproc/remoteproc_core.c +11 -0
index e5279ed9a8d7..6bc7b8b7d01e 100644
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct work_struct *work)
          if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
                  /* handle only the first crash detected */
                  mutex_unlock(&rproc->lock);
+         /*
+          * RPROC_OFFLINE state indicate there is no recovery process
+          * is in progress and no chance to have pm_relax in place.
+          * Because when recovering from crash, rproc->lock is held and
+          * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+          * and then unlock rproc->lock.
+          * RPROC_OFFLINE is only an intermediate state in recovery
+          * process.
+          */
+         if (rproc->state == RPROC_OFFLINE)
+                 pm_relax(rproc->dev.parent);
                  return;
          }
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines
drivers/remoteproc/remoteproc_core.c +0 -1
--

2.7.4

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

#15 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

Quoted 77 lines · Mathieu Poirier
Quoted 64 lines · Aiqun(Maria) Yu
Quoted 62 lines · Mathieu Poirier
Quoted 60 lines · Mathieu Poirier
Quoted 35 lines · Aiqun(Maria) Yu

Hi Mathieu,

Quoted 4 lines · Mathieu Poirier

Please add what has changed from one version to another, either in a cover letter or after the "Signed-off-by". There are many examples on how to do that on the mailing list.

Thx for the information, will take a note and benefit for next time.

Quoted 11 lines · Maria Yu
Quoted 5 lines

RPROC_OFFLINE state indicate there is no recovery process is in progress and no chance to do the pm_relax. Because when recovering from crash, rproc->lock is held and state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING, and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to RPROC_RUNNING when rproc_trigger_recovery() returns. If that is not the case then something went wrong.

Quoted 7 lines

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before returning, so we know the remote processor was stopped. Therefore if rproc->state is set to RPROC_OFFLINE something went wrong in either request_firmware() or rproc_start(). Either way the remote processor is offline and the system probably in an unknown/unstable. As such I don't see how calling pm_relax() can help things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and last rproc_trigger_recovery bailed out with only rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair. Since the subsystem may still can be recovered with customer's next trigger of rproc_start, and we can make each error out path clean with pm resources.

Quoted 6 lines
Quoted 3 lines

I suggest spending time understanding what leads to the failure when recovering from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when rproc_shutdown is triggered at similar time. So not an issue from error out of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned from the beginning. What I see happening is the following:

rproc_report_crash()
pm_stay_awake()
queue_work() // current thread is suspended

rproc_shutdown()

          rproc_stop()
                  rproc->state = RPROC_OFFLINE;

rproc_crash_handler_work()
if (rproc->state == RPROC_OFFLINE)
return // pm_relax() is not called

The right way to fix this is to add a pm_relax() in rproc_shutdown() and rproc_detach(), along with a very descriptive comment as to why it is needed.

Quoted 8 lines
Quoted 4 lines

Thinking about this further there are more ramifications to consider. Please confirm the above scenario is what you are facing. I will advise on how to move forward if that is the case.

Not sure if the situation is clear or not. So resend the email again.

The above senario is what customer is facing. crash hanppened while at the same time shutdown is triggered.

Unfortunately this is not enough details to address a problem as complex as this one.

Quoted 7 lines
Quoted 2 lines

And the device cannto goes to suspend state after that. the subsystem can still be start normally after this.

If the code flow I pasted above reflects the problem at hand, the current patch will not be sufficient to address the issue. If Arnaud confirms my suspicions we will have to think about a better solution.

Hi Mathiew,

Could you pls have more details of any side effects other then power issue of the current senario? Why the current patch is not sufficient pls?

Have the current senario in details with rproc->lock information in details:

| subsystem crashed interrupt issued      | user trigger shutdown
| rproc_report_crash()                    |
|          pm_stay_awake()                |
|          queue_work()                   |
|                                         |rproc_shutdown
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_stop()
|rproc_crash_handler_work()               |rproc->state = RPROC_OFFLINE;
|                                         |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);                |
|if (rproc->state == RPROC_OFFLINE)       |
|return // pm_relax() is not called       |rproc_boot
|mutex_unlock(&rproc->lock);              |
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_start()
|                                         |mutex_unlock(&rproc->lock);
Quoted 50 lines
Quoted 50 lines
Quoted 37 lines
Quoted 37 lines
Quoted 37 lines
Quoted 37 lines

Thanks,
Mathieu

Quoted 30 lines

When the state is in RPROC_OFFLINE it means separate request of rproc_stop was done and no need to hold the wakeup source in crash handler to recover any more.

---
    drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
    1 file changed, 11 insertions(+)
drivers/remoteproc/remoteproc_core.c +11 -0
index e5279ed9a8d7..6bc7b8b7d01e 100644
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct work_struct *work)
           if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
                   /* handle only the first crash detected */
                   mutex_unlock(&rproc->lock);
+         /*
+          * RPROC_OFFLINE state indicate there is no recovery process
+          * is in progress and no chance to have pm_relax in place.
+          * Because when recovering from crash, rproc->lock is held and
+          * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+          * and then unlock rproc->lock.
+          * RPROC_OFFLINE is only an intermediate state in recovery
+          * process.
+          */
+         if (rproc->state == RPROC_OFFLINE)
+                 pm_relax(rproc->dev.parent);
                   return;
           }
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines
drivers/remoteproc/remoteproc_core.c +0 -1
--

2.7.4

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
#16 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

Hi,

Quoted 104 lines · Aiqun(Maria) Yu
Quoted 86 lines · Mathieu Poirier
Quoted 77 lines · Aiqun(Maria) Yu
Quoted 67 lines · Mathieu Poirier
Quoted 66 lines · Mathieu Poirier
Quoted 41 lines · Aiqun(Maria) Yu

Hi Mathieu,

Quoted 5 lines · Mathieu Poirier

Please add what has changed from one version to another, either in a cover letter or after the "Signed-off-by".  There are many examples on how to do that
on the mailing list.

Thx for the information, will take a note and benefit for next time.

Quoted 13 lines · Maria Yu
Quoted 5 lines

RPROC_OFFLINE state indicate there is no recovery process is in progress and no chance to do the pm_relax. Because when recovering from crash, rproc->lock is held and state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING, and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to RPROC_RUNNING
when rproc_trigger_recovery() returns.  If that is not the case then something
went wrong.

Quoted 10 lines

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before returning,
so we know the remote processor was stopped.  Therefore if rproc->state is set
to RPROC_OFFLINE something went wrong in either request_firmware() or rproc_start().  Either way the remote processor is offline and the system probably
in an unknown/unstable.  As such I don't see how calling pm_relax() can help things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and last rproc_trigger_recovery bailed out with only rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair. Since the subsystem may still can be recovered with customer's next trigger of rproc_start, and we can make each error out path clean with pm resources.

Quoted 7 lines
Quoted 4 lines

I suggest spending time understanding what leads to the failure when recovering
from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when rproc_shutdown is triggered at similar time. So not an issue from error out of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned from
the beginning.  What I see happening is the following:

rproc_report_crash()
          pm_stay_awake()
          queue_work() // current thread is suspended

rproc_shutdown()
          rproc_stop()
                  rproc->state = RPROC_OFFLINE;

rproc_crash_handler_work()
          if (rproc->state == RPROC_OFFLINE) return // pm_relax() is not called

Quoted 8 lines
Quoted 2 lines

The right way to fix this is to add a pm_relax() in rproc_shutdown() and rproc_detach(), along with a very descriptive comment as to why it is needed.

Thinking about this further there are more ramifications to consider.  Please confirm the above scenario is what you are facing.  I will advise on how to move
forward if that is the case.

Quoted 4 lines

Not sure if the situation is clear or not. So resend the email again.

The above senario is what customer is facing. crash hanppened while at the same time shutdown is triggered.

Unfortunately this is not enough details to address a problem as complex as this one.

Quoted 7 lines
Quoted 2 lines

And the device cannto goes to suspend state after that. the subsystem can still be start normally after this.

If the code flow I pasted above reflects the problem at hand, the current patch will not be sufficient to address the issue.  If Arnaud confirms my suspicions we will have to think about a better solution.

Hi Mathiew,

Could you pls have more details of any side effects other then power issue of the current senario?
Why the current patch is not sufficient pls?

Have the current senario in details with rproc->lock information in details:

Quoted 18 lines
| subsystem crashed interrupt issued      | user trigger shutdown
| rproc_report_crash()                    |
|          pm_stay_awake()                |
|          queue_work()                   |
|                                         |rproc_shutdown
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_stop()
|rproc_crash_handler_work()               |rproc->state = RPROC_OFFLINE;
|                                         |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);                |
|if (rproc->state == RPROC_OFFLINE)       |
|return // pm_relax() is not called       |rproc_boot
|mutex_unlock(&rproc->lock);              |
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_start()
|                                         |mutex_unlock(&rproc->lock);

Agree with Mathieu, this is not so simple.

Here is my view hoping I haven't missed a point in your discussion or an other corner cases.

I tried to analyze the issues (in what follows, the term "condition" means the "if" condition in which Aiqun proposes to add the fix) :

I can see 4 use cases with race condition

1) crash report while already one is treated (rproc_boot_recovery called) => not a real use case as if the remote processor is crashed we should not have a second crash report

2) rproc_stop executed between the queuing of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_OFFLINE
=> we enter in the "condition" and the pm_relax has to be called => This commit fix should solve this use case

3) rproc_detach executed between the queue of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_DETACHED;
=> we don't go in "the condition" and issue because the recovery reattach
to the remote processor
=> but pm_relax is called
=> probably need an extra fix to avoid to re-attach

4) crash report while already one is treated (rproc_attach_recovery called) this one corresponds to an auto reboot of the remote processor, with a new crash
=> rproc->state = RPROC_CRASHED or rproc->state = RPROC_DETACHED; 4)a) rproc->state = RPROC_CRASHED if rproc->recovery_disabled = true

	=> should call pm_relax if rproc->recovery_disabled = true
	=> commit does not work for this use case

4)b) rproc->state = RPROC_DETACHED if recovery fails
=> error case with an unstable state
=> how to differentiate it from the use case 3) ? => introduce a RPROC_RECOVERY_FAIL state?

Then pm_stay_awake is called when the crash work is queued. It seems to me coherent to call the pm_relax in the work handler.

Here is a quick and dirty patch (not tested) that should take into account the main use cases ( except 1) and 4)b) )

@@ -2009,8 +2009,18 @@ static void rproc_crash_handler_work(struct work_struct *work)

mutex_lock(&rproc->lock);

-	if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
+	if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE ||

+ rproc->state == RPROC_DETACHED) {
/* handle only the first crash detected */
+
+ /*
+ * call pm-relax in following use cases:
+ * - the remote processor has been stopped by the user
+ * - the remote processor is detached
+ + - the remote proc has an autonomous reset but recovery_disabled is true.
+ */
+ if(rproc->state != RPROC_CRASHED || rproc->recovery_disabled)
+ pm_relax(rproc->dev.parent);
mutex_unlock(&rproc->lock); return;
}

Regards,
Arnaud

Quoted 55 lines
Quoted 53 lines
Quoted 53 lines
Quoted 48 lines
Quoted 48 lines
Quoted 48 lines
Quoted 43 lines

Thanks,
Mathieu

Quoted 36 lines

When the state is in RPROC_OFFLINE it means separate request of rproc_stop was done and no need to hold the wakeup source in crash handler to recover any more.

---
    drivers/remoteproc/remoteproc_core.c | 11 +++++++++++

    1 file changed, 11 insertions(+)
drivers/remoteproc/remoteproc_core.c +0 -0
index e5279ed9a8d7..6bc7b8b7d01e 100644
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct
 work_struct *work)

           if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
                   /* handle only the first crash detected */ mutex_unlock(&rproc->lock);
+         /*
+          * RPROC_OFFLINE state indicate there is no recovery process +          * is in progress and no chance to have pm_relax in place. +          * Because when recovering from crash, rproc->lock is held and +          * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING, +          * and then unlock rproc->lock.
+          * RPROC_OFFLINE is only an intermediate state in recovery +          * process.
+          */
+         if (rproc->state == RPROC_OFFLINE)
+                 pm_relax(rproc->dev.parent);
                   return;
           }

Signature · 1 line
2.7.4
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
#17 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

Hi,

Thx for the detailed comments, here is my updated comments in below as well:

For the contention of export function of rproc_detach and crash interrupt handler is never take condition and well handlered in current driver.

If this condition wanted to be take into consideration, what about we have an extra change to support rproc_detach and crash contention senarios?

other use cases should be already well handler per my understanding. Let me know if there is any mis-understanding.

Quoted 131 lines · Arnaud POULIQUEN

Hi,

Quoted 104 lines · Aiqun(Maria) Yu
Quoted 86 lines · Mathieu Poirier
Quoted 77 lines · Aiqun(Maria) Yu
Quoted 67 lines · Mathieu Poirier
Quoted 66 lines · Mathieu Poirier
Quoted 41 lines · Aiqun(Maria) Yu

Hi Mathieu,

Quoted 5 lines · Mathieu Poirier

Please add what has changed from one version to another, either in a cover letter or after the "Signed-off-by".  There are many examples on how to do that
on the mailing list.

Thx for the information, will take a note and benefit for next time.

Quoted 13 lines · Maria Yu
Quoted 5 lines

RPROC_OFFLINE state indicate there is no recovery process is in progress and no chance to do the pm_relax. Because when recovering from crash, rproc->lock is held and state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING, and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to RPROC_RUNNING
when rproc_trigger_recovery() returns.  If that is not the case then something
went wrong.

Quoted 10 lines

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before returning,
so we know the remote processor was stopped.  Therefore if rproc->state is set
to RPROC_OFFLINE something went wrong in either request_firmware() or rproc_start().  Either way the remote processor is offline and the system probably
in an unknown/unstable.  As such I don't see how calling pm_relax() can help things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and last rproc_trigger_recovery bailed out with only rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair. Since the subsystem may still can be recovered with customer's next trigger of rproc_start, and we can make each error out path clean with pm resources.

Quoted 7 lines
Quoted 4 lines

I suggest spending time understanding what leads to the failure when recovering
from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when rproc_shutdown is triggered at similar time. So not an issue from error out of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned from
the beginning.  What I see happening is the following:

rproc_report_crash()
          pm_stay_awake()
          queue_work() // current thread is suspended

rproc_shutdown()
          rproc_stop()
                  rproc->state = RPROC_OFFLINE;

rproc_crash_handler_work()
          if (rproc->state == RPROC_OFFLINE) return // pm_relax() is not called

Quoted 8 lines
Quoted 2 lines

The right way to fix this is to add a pm_relax() in rproc_shutdown() and rproc_detach(), along with a very descriptive comment as to why it is needed.

Thinking about this further there are more ramifications to consider.  Please confirm the above scenario is what you are facing.  I will advise on how to move
forward if that is the case.

Quoted 4 lines

Not sure if the situation is clear or not. So resend the email again.

The above senario is what customer is facing. crash hanppened while at the same time shutdown is triggered.

Unfortunately this is not enough details to address a problem as complex as this one.

Quoted 7 lines
Quoted 2 lines

And the device cannto goes to suspend state after that. the subsystem can still be start normally after this.

If the code flow I pasted above reflects the problem at hand, the current patch will not be sufficient to address the issue.  If Arnaud confirms my suspicions we will have to think about a better solution.

Hi Mathiew,

Could you pls have more details of any side effects other then power issue of the current senario?
Why the current patch is not sufficient pls?

Have the current senario in details with rproc->lock information in details:

Quoted 18 lines
| subsystem crashed interrupt issued      | user trigger shutdown
| rproc_report_crash()                    |
|          pm_stay_awake()                |
|          queue_work()                   |
|                                         |rproc_shutdown
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_stop()
|rproc_crash_handler_work()               |rproc->state = RPROC_OFFLINE;
|                                         |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);                |
|if (rproc->state == RPROC_OFFLINE)       |
|return // pm_relax() is not called       |rproc_boot
|mutex_unlock(&rproc->lock);              |
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_start()
|                                         |mutex_unlock(&rproc->lock);

Agree with Mathieu, this is not so simple.

Here is my view hoping I haven't missed a point in your discussion or an other corner cases.

Quoted 9 lines

I tried to analyze the issues (in what follows, the term "condition" means the "if" condition in which Aiqun proposes to add the fix) :

I can see 4 use cases with race condition

1) crash report while already one is treated (rproc_boot_recovery called) => not a real use case as if the remote processor is crashed we should not have a second crash report

The second crash report can be avoid via the current rproc->state check within the protection of rproc->lock.

Quoted 6 lines

2) rproc_stop executed between the queuing of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_OFFLINE
=> we enter in the "condition" and the pm_relax has to be called => This commit fix should solve this use case

Quoted 8 lines

3) rproc_detach executed between the queue of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_DETACHED;
=> we don't go in "the condition" and issue because the recovery reattach
to the remote processor
=> but pm_relax is called
=> probably need an extra fix to avoid to re-attach

The RPROC_DETACHED state is never considered with crash contention in current driver.
if the subsystem is in RPROC_DETACHED state while it still received a crash intterupt, should it ignore the current crash interrupt or continue to do a rproc_boot and continue to attach?

Quoted 8 lines

4) crash report while already one is treated (rproc_attach_recovery called)

    this one corresponds to an auto reboot of the remote processor, with a
    new crash
    => rproc->state = RPROC_CRASHED or rproc->state = RPROC_DETACHED;
    4)a) rproc->state = RPROC_CRASHED if rproc->recovery_disabled = true
	=> should call pm_relax if rproc->recovery_disabled = true
	=> commit does not work for this use case

In the recovery_disabled==true case rproc_crash_handler_work will call pm_relax directly. The senario was well handled.

Quoted 5 lines

4)b) rproc->state = RPROC_DETACHED if recovery fails
=> error case with an unstable state
=> how to differentiate it from the use case 3) ? => introduce a RPROC_RECOVERY_FAIL state?

For every recvery fails senario, function rproc_crash_handler_work will call pm_relax always. The senario was well handled.

Quoted 88 lines

Then pm_stay_awake is called when the crash work is queued. It seems to me coherent to call the pm_relax in the work handler.

Here is a quick and dirty patch (not tested) that should take into account the main use cases ( except 1) and 4)b) )

(quoted diff) +11 -1
@@ -2009,8 +2009,18 @@ static void rproc_crash_handler_work(struct work_struct *work)
 
  	mutex_lock(&rproc->lock);
 
-	if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
+	if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE ||
+	    rproc->state == RPROC_DETACHED) {
  		/* handle only the first crash detected */
+
+		/*
+		 * call pm-relax in following use cases:
+		 * - the remote processor has been stopped by the user
+		 * - the remote processor is detached
+		 + - the remote proc has an autonomous reset but recovery_disabled is true.
+		 */
+		if(rproc->state != RPROC_CRASHED || rproc->recovery_disabled)
+			pm_relax(rproc->dev.parent);
  		mutex_unlock(&rproc->lock);
  		return;
  	}

Regards,
Arnaud

Quoted 55 lines
Quoted 53 lines
Quoted 53 lines
Quoted 48 lines
Quoted 48 lines
Quoted 48 lines
Quoted 43 lines

Thanks,
Mathieu

Quoted 36 lines

When the state is in RPROC_OFFLINE it means separate request of rproc_stop was done and no need to hold the wakeup source in crash handler to recover any more.

---
     drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
     1 file changed, 11 insertions(+)
drivers/remoteproc/remoteproc_core.c +11 -0
index e5279ed9a8d7..6bc7b8b7d01e 100644
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct
 work_struct *work)
            if (rproc->state == RPROC_CRASHED || rproc->state ==
 RPROC_OFFLINE) {
                    /* handle only the first crash detected */
                    mutex_unlock(&rproc->lock);
+         /*
+          * RPROC_OFFLINE state indicate there is no recovery process
+          * is in progress and no chance to have pm_relax in place.
+          * Because when recovering from crash, rproc->lock is held and
+          * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+          * and then unlock rproc->lock.
+          * RPROC_OFFLINE is only an intermediate state in recovery
+          * process.
+          */
+         if (rproc->state == RPROC_OFFLINE)
+                 pm_relax(rproc->dev.parent);
                    return;

           }

Signature · 1 line
2.7.4
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
maintainer
#18 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

In reply to Arnaud POULIQUEN
Quoted 127 lines · Arnaud POULIQUEN

Hi,

Quoted 104 lines · Aiqun(Maria) Yu
Quoted 86 lines · Mathieu Poirier
Quoted 77 lines · Aiqun(Maria) Yu
Quoted 67 lines · Mathieu Poirier
Quoted 66 lines · Mathieu Poirier
Quoted 41 lines · Aiqun(Maria) Yu

Hi Mathieu,

Quoted 5 lines · Mathieu Poirier

Please add what has changed from one version to another, either in a cover letter or after the "Signed-off-by". There are many examples on how to do that
on the mailing list.

Thx for the information, will take a note and benefit for next time.

Quoted 13 lines · Maria Yu
Quoted 5 lines

RPROC_OFFLINE state indicate there is no recovery process is in progress and no chance to do the pm_relax. Because when recovering from crash, rproc->lock is held and state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING, and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to RPROC_RUNNING
when rproc_trigger_recovery() returns. If that is not the case then something
went wrong.

Quoted 10 lines

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before returning,
so we know the remote processor was stopped. Therefore if rproc->state is set
to RPROC_OFFLINE something went wrong in either request_firmware() or rproc_start(). Either way the remote processor is offline and the system probably
in an unknown/unstable. As such I don't see how calling pm_relax() can help things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and last rproc_trigger_recovery bailed out with only rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair. Since the subsystem may still can be recovered with customer's next trigger of rproc_start, and we can make each error out path clean with pm resources.

Quoted 7 lines
Quoted 4 lines

I suggest spending time understanding what leads to the failure when recovering
from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when rproc_shutdown is triggered at similar time. So not an issue from error out of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned from
the beginning. What I see happening is the following:

rproc_report_crash()
pm_stay_awake()
queue_work() // current thread is suspended

rproc_shutdown()

          rproc_stop()
                  rproc->state = RPROC_OFFLINE;

rproc_crash_handler_work()
if (rproc->state == RPROC_OFFLINE)
return // pm_relax() is not called

Quoted 8 lines
Quoted 2 lines

The right way to fix this is to add a pm_relax() in rproc_shutdown() and rproc_detach(), along with a very descriptive comment as to why it is needed.

Thinking about this further there are more ramifications to consider. Please confirm the above scenario is what you are facing. I will advise on how to move
forward if that is the case.

Quoted 4 lines

Not sure if the situation is clear or not. So resend the email again.

The above senario is what customer is facing. crash hanppened while at the same time shutdown is triggered.

Unfortunately this is not enough details to address a problem as complex as this one.

Quoted 7 lines
Quoted 2 lines

And the device cannto goes to suspend state after that. the subsystem can still be start normally after this.

If the code flow I pasted above reflects the problem at hand, the current patch will not be sufficient to address the issue. If Arnaud confirms my suspicions we will have to think about a better solution.

Hi Mathiew,

Could you pls have more details of any side effects other then power issue of the current senario?
Why the current patch is not sufficient pls?

Have the current senario in details with rproc->lock information in details:

Quoted 18 lines
| subsystem crashed interrupt issued      | user trigger shutdown
| rproc_report_crash()                    |
|          pm_stay_awake()                |
|          queue_work()                   |
|                                         |rproc_shutdown
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_stop()
|rproc_crash_handler_work()               |rproc->state = RPROC_OFFLINE;
|                                         |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);                |
|if (rproc->state == RPROC_OFFLINE)       |
|return // pm_relax() is not called       |rproc_boot
|mutex_unlock(&rproc->lock);              |
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_start()
|                                         |mutex_unlock(&rproc->lock);
Quoted 2 lines

Agree with Mathieu, this is not so simple.

Thanks for looking into this.

Quoted 12 lines

Here is my view hoping I haven't missed a point in your discussion or an other corner cases.

I tried to analyze the issues (in what follows, the term "condition" means the "if" condition in which Aiqun proposes to add the fix) :

I can see 4 use cases with race condition

1) crash report while already one is treated (rproc_boot_recovery called) => not a real use case as if the remote processor is crashed we should not have a second crash report

That part is of great concern to me. *Theoretically* we should not get a new crash report while one has already been dispatched but the current code accounts for this scenario and as such the possibility can't be dismissed. Therefore we need to expect rproc_report_crash() to be called multiple times before a single instance of rproc_boot_recovery() is scheduled.

Quoted 22 lines

2) rproc_stop executed between the queuing of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_OFFLINE
=> we enter in the "condition" and the pm_relax has to be called => This commit fix should solve this use case

3) rproc_detach executed between the queue of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_DETACHED;
=> we don't go in "the condition" and issue because the recovery reattach
to the remote processor
=> but pm_relax is called
=> probably need an extra fix to avoid to re-attach

4) crash report while already one is treated (rproc_attach_recovery called) this one corresponds to an auto reboot of the remote processor, with a new crash
=> rproc->state = RPROC_CRASHED or rproc->state = RPROC_DETACHED; 4)a) rproc->state = RPROC_CRASHED if rproc->recovery_disabled = true
=> should call pm_relax if rproc->recovery_disabled = true => commit does not work for this use case

Quoted 5 lines

4)b) rproc->state = RPROC_DETACHED if recovery fails
=> error case with an unstable state
=> how to differentiate it from the use case 3) ? => introduce a RPROC_RECOVERY_FAIL state?

The case where a recovery fails needs to be considered and is the reason the original patch doesn't work. Right now in rproc_crash_handler_work(), it is not possible to differentiate between a legitimate shutdown request (scenario #2 above) and a recovery that went wrong. I think introducing RPROC_RECOVERY_FAIL would greatly simplify things.

My initial evaluation had not considered the attach/detach scenarios - thanks for adding that in the mix.

Aiqun, please send a new patchset that adds a new remote processor state, i.e RPROC_RECOVERY_FAIL. There should also be another patch in that set that takes attach/detach scenarios into account. The code between the v6.0 and v6.1 cycle has changed a lot in that area so make sure to properly rebase.

Quoted 88 lines

Then pm_stay_awake is called when the crash work is queued. It seems to me coherent to call the pm_relax in the work handler.

Here is a quick and dirty patch (not tested) that should take into account the main use cases ( except 1) and 4)b) )

(quoted diff) +11 -1
@@ -2009,8 +2009,18 @@ static void rproc_crash_handler_work(struct work_struct *work)
 
        mutex_lock(&rproc->lock);
 
-       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
+       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE ||
+           rproc->state == RPROC_DETACHED) {
                /* handle only the first crash detected */
+
+               /*
+                * call pm-relax in following use cases:
+                * - the remote processor has been stopped by the user
+                * - the remote processor is detached
+                + - the remote proc has an autonomous reset but recovery_disabled is true.
+                */
+               if(rproc->state != RPROC_CRASHED || rproc->recovery_disabled)
+                       pm_relax(rproc->dev.parent);
                mutex_unlock(&rproc->lock);
                return;
        }

Regards,
Arnaud

Quoted 55 lines
Quoted 53 lines
Quoted 53 lines
Quoted 40 lines
Quoted 40 lines
Quoted 40 lines
Quoted 40 lines

Thanks,
Mathieu

Quoted 33 lines

When the state is in RPROC_OFFLINE it means separate request of rproc_stop was done and no need to hold the wakeup source in crash handler to recover any more.

---
    drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
    1 file changed, 11 insertions(+)
drivers/remoteproc/remoteproc_core.c +11 -0
index e5279ed9a8d7..6bc7b8b7d01e 100644
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct
 work_struct *work)
           if (rproc->state == RPROC_CRASHED || rproc->state ==
 RPROC_OFFLINE) {
                   /* handle only the first crash detected */
                   mutex_unlock(&rproc->lock);
+         /*
+          * RPROC_OFFLINE state indicate there is no recovery process
+          * is in progress and no chance to have pm_relax in place.
+          * Because when recovering from crash, rproc->lock is held and
+          * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+          * and then unlock rproc->lock.
+          * RPROC_OFFLINE is only an intermediate state in recovery
+          * process.
+          */
+         if (rproc->state == RPROC_OFFLINE)
+                 pm_relax(rproc->dev.parent);
                   return;
           }
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines

--
2.7.4

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

#19 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

Hi,

Let me think about this carefully.

When in RPROC_RECOVERY_FAIL case we want to re-do the recovery process again or just leave the pm_relax?

recovery fail case 1:

|                                      |firstcrash interrupt issued
| second crashed interrupt issued      | rproc_report_crash()
| rproc_report_crash()                 |          pm_stay_awake()
|          pm_stay_awake()             |          queue_work()
|          queue_work()                |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|rproc_crash_handler_work()            |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);             |pm_relax()
|if (rproc->state == RPROC_OFFLINE)    |
|return // shouldn't do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

recovery fail case 2:

|                                      |firstcrash interrupt issued
|                                      | rproc_report_crash()
|                                      |          pm_stay_awake()
|                                      |          queue_work()
|                                      |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|                                      |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|                                      |pm_relax()
|
| second crashed interrupt issued      |
| rproc_report_crash()                 |
|          pm_stay_awake()             |
|          queue_work()                |
|pm_stay_awake()
|mutex_lock(&rproc->lock);
|if (rproc->state == RPROC_OFFLINE)    |
|return // still need do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

Maybe I can have:
1. the pm_stay_awake and pm_relax with count based and call with paired for fix current concurency issue.
2. RPROC_RECOVERY_FAIL can be another patch for continue try to do recovery work.
3. handle RPROC_DETACHED case.

Quoted 191 lines · Mathieu Poirier
Quoted 127 lines · Arnaud POULIQUEN

Hi,

Quoted 104 lines · Aiqun(Maria) Yu
Quoted 86 lines · Mathieu Poirier
Quoted 77 lines · Aiqun(Maria) Yu
Quoted 67 lines · Mathieu Poirier
Quoted 66 lines · Mathieu Poirier
Quoted 41 lines · Aiqun(Maria) Yu

Hi Mathieu,

Quoted 5 lines · Mathieu Poirier

Please add what has changed from one version to another, either in a cover letter or after the "Signed-off-by". There are many examples on how to do that
on the mailing list.

Thx for the information, will take a note and benefit for next time.

Quoted 13 lines · Maria Yu
Quoted 5 lines

RPROC_OFFLINE state indicate there is no recovery process is in progress and no chance to do the pm_relax. Because when recovering from crash, rproc->lock is held and state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING, and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to RPROC_RUNNING
when rproc_trigger_recovery() returns. If that is not the case then something
went wrong.

Quoted 10 lines

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before returning,
so we know the remote processor was stopped. Therefore if rproc->state is set
to RPROC_OFFLINE something went wrong in either request_firmware() or rproc_start(). Either way the remote processor is offline and the system probably
in an unknown/unstable. As such I don't see how calling pm_relax() can help things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and last rproc_trigger_recovery bailed out with only rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair. Since the subsystem may still can be recovered with customer's next trigger of rproc_start, and we can make each error out path clean with pm resources.

Quoted 7 lines
Quoted 4 lines

I suggest spending time understanding what leads to the failure when recovering
from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when rproc_shutdown is triggered at similar time. So not an issue from error out of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned from
the beginning. What I see happening is the following:

rproc_report_crash()
pm_stay_awake()
queue_work() // current thread is suspended

rproc_shutdown()

           rproc_stop()
                   rproc->state = RPROC_OFFLINE;

rproc_crash_handler_work()
if (rproc->state == RPROC_OFFLINE)
return // pm_relax() is not called

Quoted 8 lines
Quoted 2 lines

The right way to fix this is to add a pm_relax() in rproc_shutdown() and rproc_detach(), along with a very descriptive comment as to why it is needed.

Thinking about this further there are more ramifications to consider. Please confirm the above scenario is what you are facing. I will advise on how to move
forward if that is the case.

Quoted 4 lines

Not sure if the situation is clear or not. So resend the email again.

The above senario is what customer is facing. crash hanppened while at the same time shutdown is triggered.

Unfortunately this is not enough details to address a problem as complex as this one.

Quoted 7 lines
Quoted 2 lines

And the device cannto goes to suspend state after that. the subsystem can still be start normally after this.

If the code flow I pasted above reflects the problem at hand, the current patch will not be sufficient to address the issue. If Arnaud confirms my suspicions we will have to think about a better solution.

Hi Mathiew,

Could you pls have more details of any side effects other then power issue of the current senario?
Why the current patch is not sufficient pls?

Have the current senario in details with rproc->lock information in details:

Quoted 18 lines
| subsystem crashed interrupt issued      | user trigger shutdown
| rproc_report_crash()                    |
|          pm_stay_awake()                |
|          queue_work()                   |
|                                         |rproc_shutdown
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_stop()
|rproc_crash_handler_work()               |rproc->state = RPROC_OFFLINE;
|                                         |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);                |
|if (rproc->state == RPROC_OFFLINE)       |
|return // pm_relax() is not called       |rproc_boot
|mutex_unlock(&rproc->lock);              |
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_start()
|                                         |mutex_unlock(&rproc->lock);
Quoted 2 lines

Agree with Mathieu, this is not so simple.

Thanks for looking into this.

Quoted 12 lines

Here is my view hoping I haven't missed a point in your discussion or an other corner cases.

I tried to analyze the issues (in what follows, the term "condition" means the "if" condition in which Aiqun proposes to add the fix) :

I can see 4 use cases with race condition

1) crash report while already one is treated (rproc_boot_recovery called) => not a real use case as if the remote processor is crashed we should not have a second crash report

That part is of great concern to me. *Theoretically* we should not get a new crash report while one has already been dispatched but the current code accounts for this scenario and as such the possibility can't be dismissed. Therefore we need to expect rproc_report_crash() to be called multiple times before a single instance of rproc_boot_recovery() is scheduled.

Quoted 22 lines

2) rproc_stop executed between the queuing of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_OFFLINE
=> we enter in the "condition" and the pm_relax has to be called => This commit fix should solve this use case

3) rproc_detach executed between the queue of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_DETACHED;
=> we don't go in "the condition" and issue because the recovery reattach
to the remote processor
=> but pm_relax is called
=> probably need an extra fix to avoid to re-attach

4) crash report while already one is treated (rproc_attach_recovery called) this one corresponds to an auto reboot of the remote processor, with a new crash
=> rproc->state = RPROC_CRASHED or rproc->state = RPROC_DETACHED; 4)a) rproc->state = RPROC_CRASHED if rproc->recovery_disabled = true
=> should call pm_relax if rproc->recovery_disabled = true => commit does not work for this use case

Quoted 5 lines

4)b) rproc->state = RPROC_DETACHED if recovery fails
=> error case with an unstable state
=> how to differentiate it from the use case 3) ? => introduce a RPROC_RECOVERY_FAIL state?

The case where a recovery fails needs to be considered and is the reason the original patch doesn't work. Right now in rproc_crash_handler_work(), it is not possible to differentiate between a legitimate shutdown request (scenario #2 above) and a recovery that went wrong. I think introducing RPROC_RECOVERY_FAIL would greatly simplify things.

Quoted 9 lines

My initial evaluation had not considered the attach/detach scenarios - thanks for adding that in the mix.

Aiqun, please send a new patchset that adds a new remote processor state, i.e RPROC_RECOVERY_FAIL. There should also be another patch in that set that takes attach/detach scenarios into account. The code between the v6.0 and v6.1 cycle has changed a lot in that area so make sure to properly rebase.

I will try.

Quoted 88 lines
Quoted 88 lines

Then pm_stay_awake is called when the crash work is queued. It seems to me coherent to call the pm_relax in the work handler.

Here is a quick and dirty patch (not tested) that should take into account the main use cases ( except 1) and 4)b) )

(quoted diff) +11 -1
@@ -2009,8 +2009,18 @@ static void rproc_crash_handler_work(struct work_struct *work)
 
         mutex_lock(&rproc->lock);
 
-       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
+       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE ||
+           rproc->state == RPROC_DETACHED) {
                 /* handle only the first crash detected */
+
+               /*
+                * call pm-relax in following use cases:
+                * - the remote processor has been stopped by the user
+                * - the remote processor is detached
+                + - the remote proc has an autonomous reset but recovery_disabled is true.
+                */
+               if(rproc->state != RPROC_CRASHED || rproc->recovery_disabled)
+                       pm_relax(rproc->dev.parent);
                 mutex_unlock(&rproc->lock);
                 return;
         }

Regards,
Arnaud

Quoted 55 lines
Quoted 53 lines
Quoted 53 lines
Quoted 40 lines
Quoted 40 lines
Quoted 40 lines
Quoted 40 lines

Thanks,
Mathieu

Quoted 33 lines

When the state is in RPROC_OFFLINE it means separate request of rproc_stop was done and no need to hold the wakeup source in crash handler to recover any more.

---
     drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
     1 file changed, 11 insertions(+)
drivers/remoteproc/remoteproc_core.c +11 -0
index e5279ed9a8d7..6bc7b8b7d01e 100644
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct
 work_struct *work)
            if (rproc->state == RPROC_CRASHED || rproc->state ==
 RPROC_OFFLINE) {
                    /* handle only the first crash detected */
                    mutex_unlock(&rproc->lock);
+         /*
+          * RPROC_OFFLINE state indicate there is no recovery process
+          * is in progress and no chance to have pm_relax in place.
+          * Because when recovering from crash, rproc->lock is held and
+          * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+          * and then unlock rproc->lock.
+          * RPROC_OFFLINE is only an intermediate state in recovery
+          * process.
+          */
+         if (rproc->state == RPROC_OFFLINE)
+                 pm_relax(rproc->dev.parent);
                    return;
            }
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines

--
2.7.4

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
maintainer
#20 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

Quoted 6 lines · Aiqun(Maria) Yu

Hi,

Let me think about this carefully.

When in RPROC_RECOVERY_FAIL case we want to re-do the recovery process again or just leave the pm_relax?

Neither.

When a recovery fail we don't want to call pm_relax(). The code in rproc_crash_handler_work() becomes:

	if (rproc->state == RPROC_OFFLINE) {
                /* We have raced with rproc_shutdown() */
                pm_relax()
		mutex_unlock(&rproc->lock);
		return;
        }

	if (rproc->state == RPROC_CRASHED ||
            rproc->state == RPROC_RECOVERY_FAILED) {
		/* handle only the first crash detected */
		mutex_unlock(&rproc->lock);
		return;
	}

RPROC_RECOVERY_FAILED gets set in rproc_boot_recovery() if request_firmware() or rproc_start() fail. Function rproc_trigger_recovery() needs to allow for the recovery the the remote processor is in RPROC_RECOVERY_FAILED state. As such the condition becomes:

        /* State could have changed before we got the mutex */
	if (rproc->state != RPROC_CRASHED &&
            rproc->state != RPROC_RECOVERY_FAILED)
		goto unlock_mutex;

Start with that and we can look at corner cases (if some exists) with a fresh patchset. Note that I have not addressed the attach/detach() scenario in the above.

Thanks,
Mathieu

Quoted 349 lines

recovery fail case 1:

|                                      |firstcrash interrupt issued
| second crashed interrupt issued      | rproc_report_crash()
| rproc_report_crash()                 |          pm_stay_awake()
|          pm_stay_awake()             |          queue_work()
|          queue_work()                |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|rproc_crash_handler_work()            |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);             |pm_relax()
|if (rproc->state == RPROC_OFFLINE)    |
|return // shouldn't do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

recovery fail case 2:

|                                      |firstcrash interrupt issued
|                                      | rproc_report_crash()
|                                      |          pm_stay_awake()
|                                      |          queue_work()
|                                      |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|                                      |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|                                      |pm_relax()
|
| second crashed interrupt issued      |
| rproc_report_crash()                 |
|          pm_stay_awake()             |
|          queue_work()                |
|pm_stay_awake()
|mutex_lock(&rproc->lock);
|if (rproc->state == RPROC_OFFLINE)    |
|return // still need do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

Maybe I can have:
1. the pm_stay_awake and pm_relax with count based and call with paired for fix current concurency issue.
2. RPROC_RECOVERY_FAIL can be another patch for continue try to do recovery work.
3. handle RPROC_DETACHED case.

Quoted 191 lines · Mathieu Poirier
Quoted 127 lines · Arnaud POULIQUEN

Hi,

Quoted 104 lines · Aiqun(Maria) Yu
Quoted 86 lines · Mathieu Poirier
Quoted 77 lines · Aiqun(Maria) Yu
Quoted 67 lines · Mathieu Poirier
Quoted 66 lines · Mathieu Poirier
Quoted 41 lines · Aiqun(Maria) Yu

Hi Mathieu,

Quoted 5 lines · Mathieu Poirier

Please add what has changed from one version to another, either in a cover letter or after the "Signed-off-by". There are many examples on how to do that
on the mailing list.

Thx for the information, will take a note and benefit for next time.

Quoted 13 lines · Maria Yu
Quoted 5 lines

RPROC_OFFLINE state indicate there is no recovery process is in progress and no chance to do the pm_relax. Because when recovering from crash, rproc->lock is held and state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING, and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to RPROC_RUNNING
when rproc_trigger_recovery() returns. If that is not the case then something
went wrong.

Quoted 10 lines

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before returning,
so we know the remote processor was stopped. Therefore if rproc->state is set
to RPROC_OFFLINE something went wrong in either request_firmware() or rproc_start(). Either way the remote processor is offline and the system probably
in an unknown/unstable. As such I don't see how calling pm_relax() can help things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and last rproc_trigger_recovery bailed out with only rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair. Since the subsystem may still can be recovered with customer's next trigger of rproc_start, and we can make each error out path clean with pm resources.

Quoted 7 lines
Quoted 4 lines

I suggest spending time understanding what leads to the failure when recovering
from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when rproc_shutdown is triggered at similar time. So not an issue from error out of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned from
the beginning. What I see happening is the following:

rproc_report_crash()
pm_stay_awake()
queue_work() // current thread is suspended

rproc_shutdown()

           rproc_stop()
                   rproc->state = RPROC_OFFLINE;

rproc_crash_handler_work()
if (rproc->state == RPROC_OFFLINE)
return // pm_relax() is not called

Quoted 8 lines
Quoted 2 lines

The right way to fix this is to add a pm_relax() in rproc_shutdown() and rproc_detach(), along with a very descriptive comment as to why it is needed.

Thinking about this further there are more ramifications to consider. Please confirm the above scenario is what you are facing. I will advise on how to move
forward if that is the case.

Quoted 4 lines

Not sure if the situation is clear or not. So resend the email again.

The above senario is what customer is facing. crash hanppened while at the same time shutdown is triggered.

Unfortunately this is not enough details to address a problem as complex as this one.

Quoted 7 lines
Quoted 2 lines

And the device cannto goes to suspend state after that. the subsystem can still be start normally after this.

If the code flow I pasted above reflects the problem at hand, the current patch will not be sufficient to address the issue. If Arnaud confirms my suspicions we will have to think about a better solution.

Hi Mathiew,

Could you pls have more details of any side effects other then power issue of the current senario?
Why the current patch is not sufficient pls?

Have the current senario in details with rproc->lock information in details:

Quoted 18 lines
| subsystem crashed interrupt issued      | user trigger shutdown
| rproc_report_crash()                    |
|          pm_stay_awake()                |
|          queue_work()                   |
|                                         |rproc_shutdown
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_stop()
|rproc_crash_handler_work()               |rproc->state = RPROC_OFFLINE;
|                                         |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);                |
|if (rproc->state == RPROC_OFFLINE)       |
|return // pm_relax() is not called       |rproc_boot
|mutex_unlock(&rproc->lock);              |
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_start()
|                                         |mutex_unlock(&rproc->lock);
Quoted 2 lines

Agree with Mathieu, this is not so simple.

Thanks for looking into this.

Quoted 12 lines

Here is my view hoping I haven't missed a point in your discussion or an other corner cases.

I tried to analyze the issues (in what follows, the term "condition" means the "if" condition in which Aiqun proposes to add the fix) :

I can see 4 use cases with race condition

1) crash report while already one is treated (rproc_boot_recovery called) => not a real use case as if the remote processor is crashed we should not have a second crash report

That part is of great concern to me. *Theoretically* we should not get a new crash report while one has already been dispatched but the current code accounts for this scenario and as such the possibility can't be dismissed. Therefore we need to expect rproc_report_crash() to be called multiple times before a single instance of rproc_boot_recovery() is scheduled.

Quoted 22 lines

2) rproc_stop executed between the queuing of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_OFFLINE
=> we enter in the "condition" and the pm_relax has to be called => This commit fix should solve this use case

3) rproc_detach executed between the queue of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_DETACHED;
=> we don't go in "the condition" and issue because the recovery reattach
to the remote processor
=> but pm_relax is called
=> probably need an extra fix to avoid to re-attach

4) crash report while already one is treated (rproc_attach_recovery called) this one corresponds to an auto reboot of the remote processor, with a new crash
=> rproc->state = RPROC_CRASHED or rproc->state = RPROC_DETACHED; 4)a) rproc->state = RPROC_CRASHED if rproc->recovery_disabled = true
=> should call pm_relax if rproc->recovery_disabled = true => commit does not work for this use case

Quoted 5 lines

4)b) rproc->state = RPROC_DETACHED if recovery fails
=> error case with an unstable state
=> how to differentiate it from the use case 3) ? => introduce a RPROC_RECOVERY_FAIL state?

The case where a recovery fails needs to be considered and is the reason the original patch doesn't work. Right now in rproc_crash_handler_work(), it is not possible to differentiate between a legitimate shutdown request (scenario #2 above) and a recovery that went wrong. I think introducing RPROC_RECOVERY_FAIL would greatly simplify things.

Quoted 9 lines

My initial evaluation had not considered the attach/detach scenarios - thanks for adding that in the mix.

Aiqun, please send a new patchset that adds a new remote processor state, i.e RPROC_RECOVERY_FAIL. There should also be another patch in that set that takes attach/detach scenarios into account. The code between the v6.0 and v6.1 cycle has changed a lot in that area so make sure to properly rebase.

I will try.

Quoted 88 lines
Quoted 88 lines

Then pm_stay_awake is called when the crash work is queued. It seems to me coherent to call the pm_relax in the work handler.

Here is a quick and dirty patch (not tested) that should take into account the main use cases ( except 1) and 4)b) )

(quoted diff) +11 -1
@@ -2009,8 +2009,18 @@ static void rproc_crash_handler_work(struct work_struct *work)
 
         mutex_lock(&rproc->lock);
 
-       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
+       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE ||
+           rproc->state == RPROC_DETACHED) {
                 /* handle only the first crash detected */
+
+               /*
+                * call pm-relax in following use cases:
+                * - the remote processor has been stopped by the user
+                * - the remote processor is detached
+                + - the remote proc has an autonomous reset but recovery_disabled is true.
+                */
+               if(rproc->state != RPROC_CRASHED || rproc->recovery_disabled)
+                       pm_relax(rproc->dev.parent);
                 mutex_unlock(&rproc->lock);
                 return;
         }

Regards,
Arnaud

Quoted 55 lines
Quoted 53 lines
Quoted 53 lines
Quoted 40 lines
Quoted 40 lines
Quoted 40 lines
Quoted 40 lines

Thanks,
Mathieu

Quoted 33 lines

When the state is in RPROC_OFFLINE it means separate request of rproc_stop was done and no need to hold the wakeup source in crash handler to recover any more.

---
     drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
     1 file changed, 11 insertions(+)
drivers/remoteproc/remoteproc_core.c +11 -0
index e5279ed9a8d7..6bc7b8b7d01e 100644
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct
 work_struct *work)
            if (rproc->state == RPROC_CRASHED || rproc->state ==
 RPROC_OFFLINE) {
                    /* handle only the first crash detected */
                    mutex_unlock(&rproc->lock);
+         /*
+          * RPROC_OFFLINE state indicate there is no recovery process
+          * is in progress and no chance to have pm_relax in place.
+          * Because when recovering from crash, rproc->lock is held and
+          * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+          * and then unlock rproc->lock.
+          * RPROC_OFFLINE is only an intermediate state in recovery
+          * process.
+          */
+         if (rproc->state == RPROC_OFFLINE)
+                 pm_relax(rproc->dev.parent);
                    return;
            }
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines

--
2.7.4

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
#21 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

Quoted 34 lines · Mathieu Poirier
Quoted 6 lines · Aiqun(Maria) Yu

Hi,

Let me think about this carefully.

When in RPROC_RECOVERY_FAIL case we want to re-do the recovery process again or just leave the pm_relax?

Neither.

When a recovery fail we don't want to call pm_relax(). The code in rproc_crash_handler_work() becomes:

	if (rproc->state == RPROC_OFFLINE) {
                 /* We have raced with rproc_shutdown() */
                 pm_relax()
		mutex_unlock(&rproc->lock);
		return;
         }

	if (rproc->state == RPROC_CRASHED ||
             rproc->state == RPROC_RECOVERY_FAILED) {
		/* handle only the first crash detected */
		mutex_unlock(&rproc->lock);
		return;
	}

RPROC_RECOVERY_FAILED gets set in rproc_boot_recovery() if request_firmware() or rproc_start() fail. Function rproc_trigger_recovery() needs to allow for the recovery the the remote processor is in RPROC_RECOVERY_FAILED state. As such the condition becomes:

Quoted 8 lines
         /* State could have changed before we got the mutex */
	if (rproc->state != RPROC_CRASHED &&
             rproc->state != RPROC_RECOVERY_FAILED)
		goto unlock_mutex;

Start with that and we can look at corner cases (if some exists) with a fresh patchset. Note that I have not addressed the attach/detach() scenario in the above.

If we didn't deal with the recovery failed case with correct pm_relax call, it may left the device in a state that cannot enter to suspend state. Because first PROC_RECOVERY_FAIL case cannot ensure it have pm_relax called before the second crash handler call pm_stay_awake or not.

So, What about the atomic count along with pm_relax and pm_stay_awake ?

struct rproc{
...
atomic_t wake_count;
...
}

rproc_pm_stay_awake()

{
	atomic_inc(&wake_count);
	pm_stay_awake();
}

rproc_pm_relax()

{
	if (atomic_dec_return(&wake_count) == 0)
		pm_stay_awake();
}

can refer code like:

rproc_report_crash()

{
	...
	rproc_pm_stay_awake();
	queue_work();
	...
}

rproc_crash_handler_work()

{
	...
	if (rproc->state == RPROC_OFFLINE || rproc->state == RPROC_CRASHED) {
                  /* We have raced with rproc_shutdown() */
                  rproc_pm_relax();
 		mutex_unlock(&rproc->lock);
 		return;
          }
	...
}
Quoted 353 lines

Thanks,
Mathieu

Quoted 349 lines

recovery fail case 1:

|                                      |firstcrash interrupt issued
| second crashed interrupt issued      | rproc_report_crash()
| rproc_report_crash()                 |          pm_stay_awake()
|          pm_stay_awake()             |          queue_work()
|          queue_work()                |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|rproc_crash_handler_work()            |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);             |pm_relax()
|if (rproc->state == RPROC_OFFLINE)    |
|return // shouldn't do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

recovery fail case 2:

|                                      |firstcrash interrupt issued
|                                      | rproc_report_crash()
|                                      |          pm_stay_awake()
|                                      |          queue_work()
|                                      |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|                                      |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|                                      |pm_relax()
|
| second crashed interrupt issued      |
| rproc_report_crash()                 |
|          pm_stay_awake()             |
|          queue_work()                |
|pm_stay_awake()
|mutex_lock(&rproc->lock);
|if (rproc->state == RPROC_OFFLINE)    |
|return // still need do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

Maybe I can have:
1. the pm_stay_awake and pm_relax with count based and call with paired for fix current concurency issue.
2. RPROC_RECOVERY_FAIL can be another patch for continue try to do recovery work.
3. handle RPROC_DETACHED case.

Quoted 191 lines · Mathieu Poirier
Quoted 127 lines · Arnaud POULIQUEN

Hi,

Quoted 104 lines · Aiqun(Maria) Yu
Quoted 86 lines · Mathieu Poirier
Quoted 77 lines · Aiqun(Maria) Yu
Quoted 67 lines · Mathieu Poirier
Quoted 66 lines · Mathieu Poirier
Quoted 41 lines · Aiqun(Maria) Yu

Hi Mathieu,

Quoted 5 lines · Mathieu Poirier

Please add what has changed from one version to another, either in a cover letter or after the "Signed-off-by". There are many examples on how to do that
on the mailing list.

Thx for the information, will take a note and benefit for next time.

Quoted 13 lines · Maria Yu
Quoted 5 lines

RPROC_OFFLINE state indicate there is no recovery process is in progress and no chance to do the pm_relax. Because when recovering from crash, rproc->lock is held and state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING, and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to RPROC_RUNNING
when rproc_trigger_recovery() returns. If that is not the case then something
went wrong.

Quoted 10 lines

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before returning,
so we know the remote processor was stopped. Therefore if rproc->state is set
to RPROC_OFFLINE something went wrong in either request_firmware() or rproc_start(). Either way the remote processor is offline and the system probably
in an unknown/unstable. As such I don't see how calling pm_relax() can help things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and last rproc_trigger_recovery bailed out with only rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair. Since the subsystem may still can be recovered with customer's next trigger of rproc_start, and we can make each error out path clean with pm resources.

Quoted 7 lines
Quoted 4 lines

I suggest spending time understanding what leads to the failure when recovering
from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when rproc_shutdown is triggered at similar time. So not an issue from error out of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned from
the beginning. What I see happening is the following:

rproc_report_crash()
pm_stay_awake()
queue_work() // current thread is suspended

rproc_shutdown()

            rproc_stop()
                    rproc->state = RPROC_OFFLINE;

rproc_crash_handler_work()
if (rproc->state == RPROC_OFFLINE)
return // pm_relax() is not called

Quoted 8 lines
Quoted 2 lines

The right way to fix this is to add a pm_relax() in rproc_shutdown() and rproc_detach(), along with a very descriptive comment as to why it is needed.

Thinking about this further there are more ramifications to consider. Please confirm the above scenario is what you are facing. I will advise on how to move
forward if that is the case.

Quoted 4 lines

Not sure if the situation is clear or not. So resend the email again.

The above senario is what customer is facing. crash hanppened while at the same time shutdown is triggered.

Unfortunately this is not enough details to address a problem as complex as this one.

Quoted 7 lines
Quoted 2 lines

And the device cannto goes to suspend state after that. the subsystem can still be start normally after this.

If the code flow I pasted above reflects the problem at hand, the current patch will not be sufficient to address the issue. If Arnaud confirms my suspicions we will have to think about a better solution.

Hi Mathiew,

Could you pls have more details of any side effects other then power issue of the current senario?
Why the current patch is not sufficient pls?

Have the current senario in details with rproc->lock information in details:

Quoted 18 lines
| subsystem crashed interrupt issued      | user trigger shutdown
| rproc_report_crash()                    |
|          pm_stay_awake()                |
|          queue_work()                   |
|                                         |rproc_shutdown
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_stop()
|rproc_crash_handler_work()               |rproc->state = RPROC_OFFLINE;
|                                         |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);                |
|if (rproc->state == RPROC_OFFLINE)       |
|return // pm_relax() is not called       |rproc_boot
|mutex_unlock(&rproc->lock);              |
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_start()
|                                         |mutex_unlock(&rproc->lock);
Quoted 2 lines

Agree with Mathieu, this is not so simple.

Thanks for looking into this.

Quoted 12 lines

Here is my view hoping I haven't missed a point in your discussion or an other corner cases.

I tried to analyze the issues (in what follows, the term "condition" means the "if" condition in which Aiqun proposes to add the fix) :

I can see 4 use cases with race condition

1) crash report while already one is treated (rproc_boot_recovery called) => not a real use case as if the remote processor is crashed we should not have a second crash report

That part is of great concern to me. *Theoretically* we should not get a new crash report while one has already been dispatched but the current code accounts for this scenario and as such the possibility can't be dismissed. Therefore we need to expect rproc_report_crash() to be called multiple times before a single instance of rproc_boot_recovery() is scheduled.

Quoted 22 lines

2) rproc_stop executed between the queuing of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_OFFLINE
=> we enter in the "condition" and the pm_relax has to be called => This commit fix should solve this use case

3) rproc_detach executed between the queue of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_DETACHED;
=> we don't go in "the condition" and issue because the recovery reattach
to the remote processor
=> but pm_relax is called
=> probably need an extra fix to avoid to re-attach

4) crash report while already one is treated (rproc_attach_recovery called) this one corresponds to an auto reboot of the remote processor, with a new crash
=> rproc->state = RPROC_CRASHED or rproc->state = RPROC_DETACHED; 4)a) rproc->state = RPROC_CRASHED if rproc->recovery_disabled = true
=> should call pm_relax if rproc->recovery_disabled = true => commit does not work for this use case

Quoted 5 lines

4)b) rproc->state = RPROC_DETACHED if recovery fails
=> error case with an unstable state
=> how to differentiate it from the use case 3) ? => introduce a RPROC_RECOVERY_FAIL state?

The case where a recovery fails needs to be considered and is the reason the original patch doesn't work. Right now in rproc_crash_handler_work(), it is not possible to differentiate between a legitimate shutdown request (scenario #2 above) and a recovery that went wrong. I think introducing RPROC_RECOVERY_FAIL would greatly simplify things.

Quoted 9 lines

My initial evaluation had not considered the attach/detach scenarios - thanks for adding that in the mix.

Aiqun, please send a new patchset that adds a new remote processor state, i.e RPROC_RECOVERY_FAIL. There should also be another patch in that set that takes attach/detach scenarios into account. The code between the v6.0 and v6.1 cycle has changed a lot in that area so make sure to properly rebase.

I will try.

Quoted 88 lines
Quoted 88 lines

Then pm_stay_awake is called when the crash work is queued. It seems to me coherent to call the pm_relax in the work handler.

Here is a quick and dirty patch (not tested) that should take into account the main use cases ( except 1) and 4)b) )

(quoted diff) +11 -1
@@ -2009,8 +2009,18 @@ static void rproc_crash_handler_work(struct work_struct *work)
 
          mutex_lock(&rproc->lock);
 
-       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
+       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE ||
+           rproc->state == RPROC_DETACHED) {
                  /* handle only the first crash detected */
+
+               /*
+                * call pm-relax in following use cases:
+                * - the remote processor has been stopped by the user
+                * - the remote processor is detached
+                + - the remote proc has an autonomous reset but recovery_disabled is true.
+                */
+               if(rproc->state != RPROC_CRASHED || rproc->recovery_disabled)
+                       pm_relax(rproc->dev.parent);
                  mutex_unlock(&rproc->lock);
                  return;
          }

Regards,
Arnaud

Quoted 55 lines
Quoted 53 lines
Quoted 53 lines
Quoted 40 lines
Quoted 40 lines
Quoted 40 lines
Quoted 40 lines

Thanks,
Mathieu

Quoted 33 lines

When the state is in RPROC_OFFLINE it means separate request of rproc_stop was done and no need to hold the wakeup source in crash handler to recover any more.

---
      drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
      1 file changed, 11 insertions(+)
drivers/remoteproc/remoteproc_core.c +11 -0
index e5279ed9a8d7..6bc7b8b7d01e 100644
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct
 work_struct *work)
             if (rproc->state == RPROC_CRASHED || rproc->state ==
 RPROC_OFFLINE) {
                     /* handle only the first crash detected */
                     mutex_unlock(&rproc->lock);
+         /*
+          * RPROC_OFFLINE state indicate there is no recovery process
+          * is in progress and no chance to have pm_relax in place.
+          * Because when recovering from crash, rproc->lock is held and
+          * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+          * and then unlock rproc->lock.
+          * RPROC_OFFLINE is only an intermediate state in recovery
+          * process.
+          */
+         if (rproc->state == RPROC_OFFLINE)
+                 pm_relax(rproc->dev.parent);
                     return;
             }
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines

--
2.7.4

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
maintainer
#22 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

Quoted 39 lines · Aiqun(Maria) Yu
Quoted 38 lines · Mathieu Poirier
Quoted 6 lines · Aiqun(Maria) Yu

Hi,

Let me think about this carefully.

When in RPROC_RECOVERY_FAIL case we want to re-do the recovery process again or just leave the pm_relax?

Neither.

When a recovery fail we don't want to call pm_relax(). The code in rproc_crash_handler_work() becomes:

	if (rproc->state == RPROC_OFFLINE) {
                 /* We have raced with rproc_shutdown() */
                 pm_relax()
		mutex_unlock(&rproc->lock);
		return;
         }

	if (rproc->state == RPROC_CRASHED ||
             rproc->state == RPROC_RECOVERY_FAILED) {
		/* handle only the first crash detected */
		mutex_unlock(&rproc->lock);
		return;
	}

RPROC_RECOVERY_FAILED gets set in rproc_boot_recovery() if request_firmware() or rproc_start() fail. Function rproc_trigger_recovery() needs to allow for the recovery the the remote processor is in RPROC_RECOVERY_FAILED state. As such the condition becomes:

         /* State could have changed before we got the mutex */
	if (rproc->state != RPROC_CRASHED &&
             rproc->state != RPROC_RECOVERY_FAILED)
		goto unlock_mutex;
Quoted 6 lines
Quoted 3 lines

Start with that and we can look at corner cases (if some exists) with a fresh patchset. Note that I have not addressed the attach/detach() scenario in the above.

If we didn't deal with the recovery failed case with correct pm_relax call, it may left the device in a state that cannot enter to suspend state.

That is what I am looking for. We don't want to give the impression that everything is fine by allowing the device to suspend. If the remote processor can't be recovered than it needs to be dealth with.

Quoted 3 lines

Because first PROC_RECOVERY_FAIL case cannot ensure it have pm_relax called before the second crash handler call pm_stay_awake or not.

I've been thinking about that part. I don't think adding a wake_count to control calls to pm_stay_awake()/pm_relax() is the best way to go. There is a similar count happening in the PM runtime subsystem and that is what we should be using. I have asked a power management expert at Linaro for guidance with this matter. I should be able to get back to you with a way forward by the end of next week.

Quoted 400 lines

So, What about the atomic count along with pm_relax and pm_stay_awake ?

struct rproc{
...
atomic_t wake_count;
...
}

rproc_pm_stay_awake()

{
	atomic_inc(&wake_count);
	pm_stay_awake();
}

rproc_pm_relax()

{
	if (atomic_dec_return(&wake_count) == 0)
		pm_stay_awake();
}

can refer code like:

rproc_report_crash()

{
	...
	rproc_pm_stay_awake();
	queue_work();
	...
}

rproc_crash_handler_work()

{
	...
	if (rproc->state == RPROC_OFFLINE || rproc->state == RPROC_CRASHED) {
                  /* We have raced with rproc_shutdown() */
                  rproc_pm_relax();
(quoted diff) +0 -0
 		mutex_unlock(&rproc->lock);
 		return;
          }
...

}

Quoted 353 lines

Thanks,
Mathieu

Quoted 349 lines

recovery fail case 1:

|                                      |firstcrash interrupt issued
| second crashed interrupt issued      | rproc_report_crash()
| rproc_report_crash()                 |          pm_stay_awake()
|          pm_stay_awake()             |          queue_work()
|          queue_work()                |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|rproc_crash_handler_work()            |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);             |pm_relax()
|if (rproc->state == RPROC_OFFLINE)    |
|return // shouldn't do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

recovery fail case 2:

|                                      |firstcrash interrupt issued
|                                      | rproc_report_crash()
|                                      |          pm_stay_awake()
|                                      |          queue_work()
|                                      |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|                                      |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|                                      |pm_relax()
|
| second crashed interrupt issued      |
| rproc_report_crash()                 |
|          pm_stay_awake()             |
|          queue_work()                |
|pm_stay_awake()
|mutex_lock(&rproc->lock);
|if (rproc->state == RPROC_OFFLINE)    |
|return // still need do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

Maybe I can have:
1. the pm_stay_awake and pm_relax with count based and call with paired for fix current concurency issue.
2. RPROC_RECOVERY_FAIL can be another patch for continue try to do recovery work.
3. handle RPROC_DETACHED case.

Quoted 191 lines · Mathieu Poirier
Quoted 127 lines · Arnaud POULIQUEN

Hi,

Quoted 104 lines · Aiqun(Maria) Yu
Quoted 86 lines · Mathieu Poirier
Quoted 77 lines · Aiqun(Maria) Yu
Quoted 67 lines · Mathieu Poirier
Quoted 66 lines · Mathieu Poirier
Quoted 41 lines · Aiqun(Maria) Yu

Hi Mathieu,

Quoted 5 lines · Mathieu Poirier

Please add what has changed from one version to another, either in a cover letter or after the "Signed-off-by". There are many examples on how to do that
on the mailing list.

Thx for the information, will take a note and benefit for next time.

Quoted 13 lines · Maria Yu
Quoted 5 lines

RPROC_OFFLINE state indicate there is no recovery process
is in progress and no chance to do the pm_relax.
Because when recovering from crash, rproc->lock is held and
state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to RPROC_RUNNING
when rproc_trigger_recovery() returns. If that is not the case then something
went wrong.

Quoted 10 lines

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before returning,
so we know the remote processor was stopped. Therefore if rproc->state is set
to RPROC_OFFLINE something went wrong in either request_firmware() or rproc_start(). Either way the remote processor is offline and the system probably
in an unknown/unstable. As such I don't see how calling pm_relax() can help things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and last rproc_trigger_recovery bailed out with only rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair. Since the subsystem may still can be recovered with customer's next trigger of rproc_start, and we can make each error out path clean with pm resources.

Quoted 7 lines
Quoted 4 lines

I suggest spending time understanding what leads to the failure when recovering
from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when rproc_shutdown is triggered at similar time. So not an issue from error out of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned from
the beginning. What I see happening is the following:

rproc_report_crash()
pm_stay_awake()
queue_work() // current thread is suspended

rproc_shutdown()

            rproc_stop()
                    rproc->state = RPROC_OFFLINE;

rproc_crash_handler_work()
if (rproc->state == RPROC_OFFLINE)
return // pm_relax() is not called

Quoted 8 lines
Quoted 2 lines

The right way to fix this is to add a pm_relax() in rproc_shutdown() and rproc_detach(), along with a very descriptive comment as to why it is needed.

Thinking about this further there are more ramifications to consider. Please confirm the above scenario is what you are facing. I will advise on how to move
forward if that is the case.

Quoted 4 lines

Not sure if the situation is clear or not. So resend the email again.

The above senario is what customer is facing. crash hanppened while at the same time shutdown is triggered.

Unfortunately this is not enough details to address a problem as complex as this one.

Quoted 7 lines
Quoted 2 lines

And the device cannto goes to suspend state after that. the subsystem can still be start normally after this.

If the code flow I pasted above reflects the problem at hand, the current patch will not be sufficient to address the issue. If Arnaud confirms my suspicions we will have to think about a better solution.

Hi Mathiew,

Could you pls have more details of any side effects other then power issue of the current senario?
Why the current patch is not sufficient pls?

Have the current senario in details with rproc->lock information in details:

Quoted 18 lines
| subsystem crashed interrupt issued      | user trigger shutdown
| rproc_report_crash()                    |
|          pm_stay_awake()                |
|          queue_work()                   |
|                                         |rproc_shutdown
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_stop()
|rproc_crash_handler_work()               |rproc->state = RPROC_OFFLINE;
|                                         |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);                |
|if (rproc->state == RPROC_OFFLINE)       |
|return // pm_relax() is not called       |rproc_boot
|mutex_unlock(&rproc->lock);              |
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_start()
|                                         |mutex_unlock(&rproc->lock);
Quoted 2 lines

Agree with Mathieu, this is not so simple.

Thanks for looking into this.

Quoted 12 lines

Here is my view hoping I haven't missed a point in your discussion or an other corner cases.

I tried to analyze the issues (in what follows, the term "condition" means the "if" condition in which Aiqun proposes to add the fix) :

I can see 4 use cases with race condition

1) crash report while already one is treated (rproc_boot_recovery called) => not a real use case as if the remote processor is crashed we should not have a second crash report

That part is of great concern to me. *Theoretically* we should not get a new crash report while one has already been dispatched but the current code accounts for this scenario and as such the possibility can't be dismissed. Therefore we need to expect rproc_report_crash() to be called multiple times before a single instance of rproc_boot_recovery() is scheduled.

Quoted 22 lines

2) rproc_stop executed between the queuing of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_OFFLINE
=> we enter in the "condition" and the pm_relax has to be called => This commit fix should solve this use case

3) rproc_detach executed between the queue of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_DETACHED;
=> we don't go in "the condition" and issue because the recovery reattach
to the remote processor
=> but pm_relax is called
=> probably need an extra fix to avoid to re-attach

4) crash report while already one is treated (rproc_attach_recovery called) this one corresponds to an auto reboot of the remote processor, with a new crash
=> rproc->state = RPROC_CRASHED or rproc->state = RPROC_DETACHED; 4)a) rproc->state = RPROC_CRASHED if rproc->recovery_disabled = true
=> should call pm_relax if rproc->recovery_disabled = true => commit does not work for this use case

Quoted 5 lines

4)b) rproc->state = RPROC_DETACHED if recovery fails
=> error case with an unstable state
=> how to differentiate it from the use case 3) ? => introduce a RPROC_RECOVERY_FAIL state?

The case where a recovery fails needs to be considered and is the reason the original patch doesn't work. Right now in rproc_crash_handler_work(), it is not possible to differentiate between a legitimate shutdown request (scenario #2 above) and a recovery that went wrong. I think introducing RPROC_RECOVERY_FAIL would greatly simplify things.

Quoted 9 lines

My initial evaluation had not considered the attach/detach scenarios - thanks for adding that in the mix.

Aiqun, please send a new patchset that adds a new remote processor state, i.e RPROC_RECOVERY_FAIL. There should also be another patch in that set that takes attach/detach scenarios into account. The code between the v6.0 and v6.1 cycle has changed a lot in that area so make sure to properly rebase.

I will try.

Quoted 88 lines
Quoted 88 lines

Then pm_stay_awake is called when the crash work is queued. It seems to me coherent to call the pm_relax in the work handler.

Here is a quick and dirty patch (not tested) that should take into account the main use cases ( except 1) and 4)b) )

(quoted diff) +11 -1
@@ -2009,8 +2009,18 @@ static void rproc_crash_handler_work(struct work_struct *work)
 
          mutex_lock(&rproc->lock);
 
-       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
+       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE ||
+           rproc->state == RPROC_DETACHED) {
                  /* handle only the first crash detected */
+
+               /*
+                * call pm-relax in following use cases:
+                * - the remote processor has been stopped by the user
+                * - the remote processor is detached
+                + - the remote proc has an autonomous reset but recovery_disabled is true.
+                */
+               if(rproc->state != RPROC_CRASHED || rproc->recovery_disabled)
+                       pm_relax(rproc->dev.parent);
                  mutex_unlock(&rproc->lock);
                  return;
          }

Regards,
Arnaud

Quoted 55 lines
Quoted 53 lines
Quoted 53 lines
Quoted 40 lines
Quoted 40 lines
Quoted 40 lines
Quoted 40 lines

Thanks,
Mathieu

Quoted 33 lines

When the state is in RPROC_OFFLINE it means separate request
of rproc_stop was done and no need to hold the wakeup source
in crash handler to recover any more.

Signed-off-by: Maria Yu <[email protected]>
---
drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
1 file changed, 11 insertions(+)

diff --git a/drivers/remoteproc/remoteproc_core.c
b/drivers/remoteproc/remoteproc_core.c
index e5279ed9a8d7..6bc7b8b7d01e 100644
--- a/drivers/remoteproc/remoteproc_core.c
+++ b/drivers/remoteproc/remoteproc_core.c
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct
work_struct *work)
if (rproc->state == RPROC_CRASHED || rproc->state ==
RPROC_OFFLINE) {
/* handle only the first crash detected */
mutex_unlock(&rproc->lock);
+ /*
+ * RPROC_OFFLINE state indicate there is no recovery process
+ * is in progress and no chance to have pm_relax in place.
+ * Because when recovering from crash, rproc->lock is held and
+ * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+ * and then unlock rproc->lock.
+ * RPROC_OFFLINE is only an intermediate state in recovery
+ * process.
+ */
+ if (rproc->state == RPROC_OFFLINE)
+ pm_relax(rproc->dev.parent);
return;
}

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines

--
2.7.4

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
#23 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

Hi,

Quoted 44 lines · Mathieu Poirier
Quoted 43 lines · Aiqun(Maria) Yu
Quoted 41 lines · Mathieu Poirier
Quoted 6 lines · Aiqun(Maria) Yu

Hi,

Let me think about this carefully.

When in RPROC_RECOVERY_FAIL case we want to re-do the recovery process again or just leave the pm_relax?

Neither.

When a recovery fail we don't want to call pm_relax(). The code in rproc_crash_handler_work() becomes:

	if (rproc->state == RPROC_OFFLINE) {
                  /* We have raced with rproc_shutdown() */
                  pm_relax()
		mutex_unlock(&rproc->lock);
		return;
          }

	if (rproc->state == RPROC_CRASHED ||
              rproc->state == RPROC_RECOVERY_FAILED) {
		/* handle only the first crash detected */
		mutex_unlock(&rproc->lock);
		return;
	}

RPROC_RECOVERY_FAILED gets set in rproc_boot_recovery() if request_firmware() or rproc_start() fail. Function rproc_trigger_recovery() needs to allow for the recovery the the remote processor is in RPROC_RECOVERY_FAILED state. As such the condition becomes:

          /* State could have changed before we got the mutex */
	if (rproc->state != RPROC_CRASHED &&
              rproc->state != RPROC_RECOVERY_FAILED)
		goto unlock_mutex;

Start with that and we can look at corner cases (if some exists) with a fresh patchset. Note that I have not addressed the attach/detach() scenario in the above.

Quoted 6 lines
Quoted 2 lines

If we didn't deal with the recovery failed case with correct pm_relax call, it may left the device in a state that cannot enter to suspend state.

That is what I am looking for. We don't want to give the impression that everything is fine by allowing the device to suspend. If the remote processor can't be recovered than it needs to be dealth with.

For the normal recovery failed case, it still need to do pm_relax to not prevent the device goes to suspend. It is what in normal recovery failed case we do in rproc_crash_handler_work as well. rproc_crash_handler_work will not check the result of the rproc_trigger_recovery return value, and will always do pm_relax.

For current conconrency cornor case as well, it is better to consistant with the current design of recovery fail senarios in normal cases.

I personally agree that we shouldn't do nothing when it is a RPROC_RECOVERY_FAILED senario when it is in rproc_crash_handler_work check, because it maybe crash happened when it is trying to do the recovery. So I suggested to do a continue try of trigger recovery again instead of doing nothing and bail out if it is a RPROC_RECOVERY_FAILED state.

Quoted 5 lines
Quoted 4 lines

Because first PROC_RECOVERY_FAIL case cannot ensure it have pm_relax called before the second crash handler call pm_stay_awake or not.

Quoted 7 lines

I've been thinking about that part. I don't think adding a wake_count to control calls to pm_stay_awake()/pm_relax() is the best way to go. There is a similar count happening in the PM runtime subsystem and that is what we should be using. I have asked a power management expert at Linaro for guidance with this matter. I should be able to get back to you with a way forward by the end of next week.

Thx for the specific date provided as well. I will wait until your reply for next patchset then.

Quoted 400 lines
Quoted 400 lines

So, What about the atomic count along with pm_relax and pm_stay_awake ?

struct rproc{
...
atomic_t wake_count;
...
}

rproc_pm_stay_awake()

{
	atomic_inc(&wake_count);
	pm_stay_awake();
}

rproc_pm_relax()

{
	if (atomic_dec_return(&wake_count) == 0)
		pm_stay_awake();
}

can refer code like:

rproc_report_crash()

{
	...
	rproc_pm_stay_awake();
	queue_work();
	...
}

rproc_crash_handler_work()

{
	...
	if (rproc->state == RPROC_OFFLINE || rproc->state == RPROC_CRASHED) {
                   /* We have raced with rproc_shutdown() */
                   rproc_pm_relax();
  		mutex_unlock(&rproc->lock);
  		return;
           }
	...
}
Quoted 353 lines

Thanks,
Mathieu

Quoted 349 lines

recovery fail case 1:

|                                      |firstcrash interrupt issued
| second crashed interrupt issued      | rproc_report_crash()
| rproc_report_crash()                 |          pm_stay_awake()
|          pm_stay_awake()             |          queue_work()
|          queue_work()                |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|rproc_crash_handler_work()            |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);             |pm_relax()
|if (rproc->state == RPROC_OFFLINE)    |
|return // shouldn't do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

recovery fail case 2:

|                                      |firstcrash interrupt issued
|                                      | rproc_report_crash()
|                                      |          pm_stay_awake()
|                                      |          queue_work()
|                                      |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|                                      |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|                                      |pm_relax()
|
| second crashed interrupt issued      |
| rproc_report_crash()                 |
|          pm_stay_awake()             |
|          queue_work()                |
|pm_stay_awake()
|mutex_lock(&rproc->lock);
|if (rproc->state == RPROC_OFFLINE)    |
|return // still need do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

Maybe I can have:
1. the pm_stay_awake and pm_relax with count based and call with paired for fix current concurency issue.
2. RPROC_RECOVERY_FAIL can be another patch for continue try to do recovery work.
3. handle RPROC_DETACHED case.

Quoted 191 lines · Mathieu Poirier
Quoted 127 lines · Arnaud POULIQUEN

Hi,

Quoted 104 lines · Aiqun(Maria) Yu
Quoted 86 lines · Mathieu Poirier
Quoted 77 lines · Aiqun(Maria) Yu
Quoted 67 lines · Mathieu Poirier
Quoted 66 lines · Mathieu Poirier
Quoted 41 lines · Aiqun(Maria) Yu

Hi Mathieu,

Quoted 5 lines · Mathieu Poirier

Please add what has changed from one version to another, either in a cover
letter or after the "Signed-off-by". There are many examples on how to
do that
on the mailing list.

Thx for the information, will take a note and benefit for next time.

Quoted 23 lines · Maria Yu

RPROC_OFFLINE state indicate there is no recovery process
is in progress and no chance to do the pm_relax.
Because when recovering from crash, rproc->lock is held and
state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to
RPROC_RUNNING
when rproc_trigger_recovery() returns. If that is not the case then
something
went wrong.

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before
returning,
so we know the remote processor was stopped. Therefore if rproc->state
is set
to RPROC_OFFLINE something went wrong in either request_firmware() or
rproc_start(). Either way the remote processor is offline and the system
probably
in an unknown/unstable. As such I don't see how calling pm_relax() can help
things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and last rproc_trigger_recovery bailed out with only rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair. Since the subsystem may still can be recovered with customer's next trigger of rproc_start, and we can make each error out path clean with pm resources.

Quoted 7 lines
Quoted 4 lines

I suggest spending time understanding what leads to the failure when
recovering
from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when rproc_shutdown is triggered at similar time. So not an issue from error out of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned from
the beginning. What I see happening is the following:

rproc_report_crash()
pm_stay_awake()
queue_work() // current thread is suspended

rproc_shutdown()

             rproc_stop()
                     rproc->state = RPROC_OFFLINE;

rproc_crash_handler_work()
if (rproc->state == RPROC_OFFLINE)
return // pm_relax() is not called

Quoted 8 lines
Quoted 2 lines

The right way to fix this is to add a pm_relax() in rproc_shutdown() and rproc_detach(), along with a very descriptive comment as to why it is needed.

Thinking about this further there are more ramifications to consider. Please confirm the above scenario is what you are facing. I will advise on how to move
forward if that is the case.

Quoted 4 lines

Not sure if the situation is clear or not. So resend the email again.

The above senario is what customer is facing. crash hanppened while at the same time shutdown is triggered.

Unfortunately this is not enough details to address a problem as complex as this one.

Quoted 7 lines
Quoted 2 lines

And the device cannto goes to suspend state after that. the subsystem can still be start normally after this.

If the code flow I pasted above reflects the problem at hand, the current patch will not be sufficient to address the issue. If Arnaud confirms my suspicions we will have to think about a better solution.

Hi Mathiew,

Could you pls have more details of any side effects other then power issue of the current senario?
Why the current patch is not sufficient pls?

Have the current senario in details with rproc->lock information in details:

Quoted 18 lines
| subsystem crashed interrupt issued      | user trigger shutdown
| rproc_report_crash()                    |
|          pm_stay_awake()                |
|          queue_work()                   |
|                                         |rproc_shutdown
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_stop()
|rproc_crash_handler_work()               |rproc->state = RPROC_OFFLINE;
|                                         |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);                |
|if (rproc->state == RPROC_OFFLINE)       |
|return // pm_relax() is not called       |rproc_boot
|mutex_unlock(&rproc->lock);              |
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_start()
|                                         |mutex_unlock(&rproc->lock);
Quoted 2 lines

Agree with Mathieu, this is not so simple.

Thanks for looking into this.

Quoted 12 lines

Here is my view hoping I haven't missed a point in your discussion or an other corner cases.

I tried to analyze the issues (in what follows, the term "condition" means the "if" condition in which Aiqun proposes to add the fix) :

I can see 4 use cases with race condition

1) crash report while already one is treated (rproc_boot_recovery called) => not a real use case as if the remote processor is crashed we should not have a second crash report

That part is of great concern to me. *Theoretically* we should not get a new crash report while one has already been dispatched but the current code accounts for this scenario and as such the possibility can't be dismissed. Therefore we need to expect rproc_report_crash() to be called multiple times before a single instance of rproc_boot_recovery() is scheduled.

Quoted 22 lines

2) rproc_stop executed between the queuing of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_OFFLINE
=> we enter in the "condition" and the pm_relax has to be called => This commit fix should solve this use case

3) rproc_detach executed between the queue of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_DETACHED;
=> we don't go in "the condition" and issue because the recovery reattach
to the remote processor
=> but pm_relax is called
=> probably need an extra fix to avoid to re-attach

4) crash report while already one is treated (rproc_attach_recovery called) this one corresponds to an auto reboot of the remote processor, with a new crash
=> rproc->state = RPROC_CRASHED or rproc->state = RPROC_DETACHED; 4)a) rproc->state = RPROC_CRASHED if rproc->recovery_disabled = true
=> should call pm_relax if rproc->recovery_disabled = true => commit does not work for this use case

Quoted 5 lines

4)b) rproc->state = RPROC_DETACHED if recovery fails
=> error case with an unstable state
=> how to differentiate it from the use case 3) ? => introduce a RPROC_RECOVERY_FAIL state?

The case where a recovery fails needs to be considered and is the reason the original patch doesn't work. Right now in rproc_crash_handler_work(), it is not possible to differentiate between a legitimate shutdown request (scenario #2 above) and a recovery that went wrong. I think introducing RPROC_RECOVERY_FAIL would greatly simplify things.

Quoted 9 lines

My initial evaluation had not considered the attach/detach scenarios - thanks for adding that in the mix.

Aiqun, please send a new patchset that adds a new remote processor state, i.e RPROC_RECOVERY_FAIL. There should also be another patch in that set that takes attach/detach scenarios into account. The code between the v6.0 and v6.1 cycle has changed a lot in that area so make sure to properly rebase.

I will try.

Quoted 88 lines
Quoted 88 lines

Then pm_stay_awake is called when the crash work is queued. It seems to me coherent to call the pm_relax in the work handler.

Here is a quick and dirty patch (not tested) that should take into account the main use cases ( except 1) and 4)b) )

(quoted diff) +11 -1
@@ -2009,8 +2009,18 @@ static void rproc_crash_handler_work(struct work_struct *work)
 
           mutex_lock(&rproc->lock);
 
-       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
+       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE ||
+           rproc->state == RPROC_DETACHED) {
                   /* handle only the first crash detected */
+
+               /*
+                * call pm-relax in following use cases:
+                * - the remote processor has been stopped by the user
+                * - the remote processor is detached
+                + - the remote proc has an autonomous reset but recovery_disabled is true.
+                */
+               if(rproc->state != RPROC_CRASHED || rproc->recovery_disabled)
+                       pm_relax(rproc->dev.parent);
                   mutex_unlock(&rproc->lock);
                   return;
           }

Regards,
Arnaud

Quoted 55 lines
Quoted 53 lines
Quoted 53 lines
Quoted 40 lines
Quoted 40 lines
Quoted 40 lines
Quoted 40 lines

Thanks,
Mathieu

When the state is in RPROC_OFFLINE it means separate request
of rproc_stop was done and no need to hold the wakeup source
in crash handler to recover any more.

Signed-off-by: Maria Yu <[email protected]>
---
drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
1 file changed, 11 insertions(+)

diff --git a/drivers/remoteproc/remoteproc_core.c
b/drivers/remoteproc/remoteproc_core.c
index e5279ed9a8d7..6bc7b8b7d01e 100644
--- a/drivers/remoteproc/remoteproc_core.c
+++ b/drivers/remoteproc/remoteproc_core.c
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct
work_struct *work)
if (rproc->state == RPROC_CRASHED || rproc->state ==
RPROC_OFFLINE) {
/* handle only the first crash detected */
mutex_unlock(&rproc->lock);
+ /*
+ * RPROC_OFFLINE state indicate there is no recovery process
+ * is in progress and no chance to have pm_relax in place.
+ * Because when recovering from crash, rproc->lock is held and
+ * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+ * and then unlock rproc->lock.
+ * RPROC_OFFLINE is only an intermediate state in recovery
+ * process.
+ */
+ if (rproc->state == RPROC_OFFLINE)
+ pm_relax(rproc->dev.parent);
return;
}

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 3 lines

--
2.7.4

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
maintainer
#24 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

I had a couple of good discussions with our power management expert and even then, the way forward isn't as clear as I would have liked. I am currently travelling and as such don't have the required time to go into greater details, something I will be doing next week.

Quoted 488 lines · Aiqun(Maria) Yu

Hi,

Quoted 44 lines · Mathieu Poirier
Quoted 43 lines · Aiqun(Maria) Yu
Quoted 41 lines · Mathieu Poirier
Quoted 6 lines · Aiqun(Maria) Yu

Hi,

Let me think about this carefully.

When in RPROC_RECOVERY_FAIL case we want to re-do the recovery process again or just leave the pm_relax?

Neither.

When a recovery fail we don't want to call pm_relax(). The code in rproc_crash_handler_work() becomes:

    if (rproc->state == RPROC_OFFLINE) {
                  /* We have raced with rproc_shutdown() */
                  pm_relax()
            mutex_unlock(&rproc->lock);
            return;
          }

    if (rproc->state == RPROC_CRASHED ||
              rproc->state == RPROC_RECOVERY_FAILED) {
            /* handle only the first crash detected */
            mutex_unlock(&rproc->lock);
            return;
    }

RPROC_RECOVERY_FAILED gets set in rproc_boot_recovery() if request_firmware() or rproc_start() fail. Function rproc_trigger_recovery() needs to allow for the recovery the the remote processor is in RPROC_RECOVERY_FAILED state. As such the condition becomes:

          /* State could have changed before we got the mutex */
    if (rproc->state != RPROC_CRASHED &&
              rproc->state != RPROC_RECOVERY_FAILED)
            goto unlock_mutex;

Start with that and we can look at corner cases (if some exists) with a fresh patchset. Note that I have not addressed the attach/detach() scenario in the above.

Quoted 6 lines
Quoted 2 lines

If we didn't deal with the recovery failed case with correct pm_relax call, it may left the device in a state that cannot enter to suspend state.

That is what I am looking for. We don't want to give the impression that everything is fine by allowing the device to suspend. If the remote processor can't be recovered than it needs to be dealth with.

For the normal recovery failed case, it still need to do pm_relax to not prevent the device goes to suspend. It is what in normal recovery failed case we do in rproc_crash_handler_work as well. rproc_crash_handler_work will not check the result of the rproc_trigger_recovery return value, and will always do pm_relax.

For current conconrency cornor case as well, it is better to consistant with the current design of recovery fail senarios in normal cases.

I personally agree that we shouldn't do nothing when it is a RPROC_RECOVERY_FAILED senario when it is in rproc_crash_handler_work check, because it maybe crash happened when it is trying to do the recovery. So I suggested to do a continue try of trigger recovery again instead of doing nothing and bail out if it is a RPROC_RECOVERY_FAILED state.

Quoted 5 lines
Quoted 4 lines

Because first PROC_RECOVERY_FAIL case cannot ensure it have pm_relax called before the second crash handler call pm_stay_awake or not.

Quoted 7 lines

I've been thinking about that part. I don't think adding a wake_count to control calls to pm_stay_awake()/pm_relax() is the best way to go. There is a similar count happening in the PM runtime subsystem and that is what we should be using. I have asked a power management expert at Linaro for guidance with this matter. I should be able to get back to you with a way forward by the end of next week.

Thx for the specific date provided as well. I will wait until your reply for next patchset then.

Quoted 392 lines
Quoted 392 lines

So, What about the atomic count along with pm_relax and pm_stay_awake ?

struct rproc{
...
atomic_t wake_count;
...
}

rproc_pm_stay_awake()

{
     atomic_inc(&wake_count);
     pm_stay_awake();
}

rproc_pm_relax()

{
     if (atomic_dec_return(&wake_count) == 0)
             pm_stay_awake();
}

can refer code like:

rproc_report_crash()

{
     ...
     rproc_pm_stay_awake();
     queue_work();
     ...
}

rproc_crash_handler_work()

{
     ...
     if (rproc->state == RPROC_OFFLINE || rproc->state == RPROC_CRASHED) {
                   /* We have raced with rproc_shutdown() */
                   rproc_pm_relax();
             mutex_unlock(&rproc->lock);
             return;
           }
     ...
}
Quoted 350 lines

Thanks,
Mathieu

Quoted 346 lines

recovery fail case 1:

|                                      |firstcrash interrupt issued
| second crashed interrupt issued      | rproc_report_crash()
| rproc_report_crash()                 |          pm_stay_awake()
|          pm_stay_awake()             |          queue_work()
|          queue_work()                |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|rproc_crash_handler_work()            |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);             |pm_relax()
|if (rproc->state == RPROC_OFFLINE)    |
|return // shouldn't do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

recovery fail case 2:

|                                      |firstcrash interrupt issued
|                                      | rproc_report_crash()
|                                      |          pm_stay_awake()
|                                      |          queue_work()
|                                      |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|                                      |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|                                      |pm_relax()
|
| second crashed interrupt issued      |
| rproc_report_crash()                 |
|          pm_stay_awake()             |
|          queue_work()                |
|pm_stay_awake()
|mutex_lock(&rproc->lock);
|if (rproc->state == RPROC_OFFLINE)    |
|return // still need do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

Maybe I can have:
1. the pm_stay_awake and pm_relax with count based and call with paired for fix current concurency issue.
2. RPROC_RECOVERY_FAIL can be another patch for continue try to do recovery work.
3. handle RPROC_DETACHED case.

Quoted 191 lines · Mathieu Poirier
Quoted 127 lines · Arnaud POULIQUEN

Hi,

Quoted 104 lines · Aiqun(Maria) Yu
Quoted 86 lines · Mathieu Poirier
Quoted 77 lines · Aiqun(Maria) Yu
Quoted 67 lines · Mathieu Poirier
Quoted 66 lines · Mathieu Poirier
Quoted 48 lines · Aiqun(Maria) Yu

Hi Mathieu,

On 10/13/2022 4:43 AM, Mathieu Poirier wrote:
Please add what has changed from one version to another, either in a cover
letter or after the "Signed-off-by". There are many examples on how to
do that
on the mailing list.

Thx for the information, will take a note and benefit for next time.

On Fri, Sep 16, 2022 at 03:12:31PM +0800, Maria Yu wrote:
RPROC_OFFLINE state indicate there is no recovery process
is in progress and no chance to do the pm_relax.
Because when recovering from crash, rproc->lock is held and
state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to
RPROC_RUNNING
when rproc_trigger_recovery() returns. If that is not the case then
something
went wrong.

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before
returning,
so we know the remote processor was stopped. Therefore if rproc->state
is set
to RPROC_OFFLINE something went wrong in either request_firmware() or
rproc_start(). Either way the remote processor is offline and the system
probably
in an unknown/unstable. As such I don't see how calling pm_relax() can help
things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully
finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and
last rproc_trigger_recovery bailed out with only
rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair.
Since the subsystem may still can be recovered with customer's next trigger
of rproc_start, and we can make each error out path clean with pm resources.

I suggest spending time understanding what leads to the failure when
recovering
from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when
rproc_shutdown is triggered at similar time. So not an issue from error out
of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned from
the beginning. What I see happening is the following:

rproc_report_crash()
pm_stay_awake()
queue_work() // current thread is suspended

rproc_shutdown()

             rproc_stop()
                     rproc->state = RPROC_OFFLINE;

rproc_crash_handler_work()
if (rproc->state == RPROC_OFFLINE)
return // pm_relax() is not called

Quoted 8 lines
Quoted 2 lines

The right way to fix this is to add a pm_relax() in rproc_shutdown() and rproc_detach(), along with a very descriptive comment as to why it is needed.

Thinking about this further there are more ramifications to consider. Please confirm the above scenario is what you are facing. I will advise on how to move
forward if that is the case.

Quoted 4 lines

Not sure if the situation is clear or not. So resend the email again.

The above senario is what customer is facing. crash hanppened while at the same time shutdown is triggered.

Unfortunately this is not enough details to address a problem as complex as this one.

Quoted 7 lines
Quoted 2 lines

And the device cannto goes to suspend state after that. the subsystem can still be start normally after this.

If the code flow I pasted above reflects the problem at hand, the current patch will not be sufficient to address the issue. If Arnaud confirms my suspicions we will have to think about a better solution.

Hi Mathiew,

Could you pls have more details of any side effects other then power issue of the current senario?
Why the current patch is not sufficient pls?

Have the current senario in details with rproc->lock information in details:

Quoted 18 lines
| subsystem crashed interrupt issued      | user trigger shutdown
| rproc_report_crash()                    |
|          pm_stay_awake()                |
|          queue_work()                   |
|                                         |rproc_shutdown
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_stop()
|rproc_crash_handler_work()               |rproc->state = RPROC_OFFLINE;
|                                         |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);                |
|if (rproc->state == RPROC_OFFLINE)       |
|return // pm_relax() is not called       |rproc_boot
|mutex_unlock(&rproc->lock);              |
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_start()
|                                         |mutex_unlock(&rproc->lock);
Quoted 2 lines

Agree with Mathieu, this is not so simple.

Thanks for looking into this.

Quoted 12 lines

Here is my view hoping I haven't missed a point in your discussion or an other corner cases.

I tried to analyze the issues (in what follows, the term "condition" means the "if" condition in which Aiqun proposes to add the fix) :

I can see 4 use cases with race condition

1) crash report while already one is treated (rproc_boot_recovery called) => not a real use case as if the remote processor is crashed we should not have a second crash report

That part is of great concern to me. *Theoretically* we should not get a new crash report while one has already been dispatched but the current code accounts for this scenario and as such the possibility can't be dismissed. Therefore we need to expect rproc_report_crash() to be called multiple times before a single instance of rproc_boot_recovery() is scheduled.

Quoted 22 lines

2) rproc_stop executed between the queuing of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_OFFLINE
=> we enter in the "condition" and the pm_relax has to be called => This commit fix should solve this use case

3) rproc_detach executed between the queue of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_DETACHED;
=> we don't go in "the condition" and issue because the recovery reattach
to the remote processor
=> but pm_relax is called
=> probably need an extra fix to avoid to re-attach

4) crash report while already one is treated (rproc_attach_recovery called) this one corresponds to an auto reboot of the remote processor, with a new crash
=> rproc->state = RPROC_CRASHED or rproc->state = RPROC_DETACHED; 4)a) rproc->state = RPROC_CRASHED if rproc->recovery_disabled = true
=> should call pm_relax if rproc->recovery_disabled = true => commit does not work for this use case

Quoted 5 lines

4)b) rproc->state = RPROC_DETACHED if recovery fails
=> error case with an unstable state
=> how to differentiate it from the use case 3) ? => introduce a RPROC_RECOVERY_FAIL state?

The case where a recovery fails needs to be considered and is the reason the original patch doesn't work. Right now in rproc_crash_handler_work(), it is not possible to differentiate between a legitimate shutdown request (scenario #2 above) and a recovery that went wrong. I think introducing RPROC_RECOVERY_FAIL would greatly simplify things.

Quoted 9 lines

My initial evaluation had not considered the attach/detach scenarios - thanks for adding that in the mix.

Aiqun, please send a new patchset that adds a new remote processor state, i.e RPROC_RECOVERY_FAIL. There should also be another patch in that set that takes attach/detach scenarios into account. The code between the v6.0 and v6.1 cycle has changed a lot in that area so make sure to properly rebase.

I will try.

Quoted 88 lines
Quoted 88 lines

Then pm_stay_awake is called when the crash work is queued. It seems to me coherent to call the pm_relax in the work handler.

Here is a quick and dirty patch (not tested) that should take into account the main use cases ( except 1) and 4)b) )

(quoted diff) +11 -1
@@ -2009,8 +2009,18 @@ static void rproc_crash_handler_work(struct work_struct *work)
 
           mutex_lock(&rproc->lock);
 
-       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
+       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE ||
+           rproc->state == RPROC_DETACHED) {
                   /* handle only the first crash detected */
+
+               /*
+                * call pm-relax in following use cases:
+                * - the remote processor has been stopped by the user
+                * - the remote processor is detached
+                + - the remote proc has an autonomous reset but recovery_disabled is true.
+                */
+               if(rproc->state != RPROC_CRASHED || rproc->recovery_disabled)
+                       pm_relax(rproc->dev.parent);
                   mutex_unlock(&rproc->lock);
                   return;
           }

Regards,
Arnaud

Quoted 55 lines
Quoted 53 lines
Quoted 53 lines
Quoted 40 lines
Quoted 40 lines
Quoted 40 lines

Thanks,
Mathieu

When the state is in RPROC_OFFLINE it means separate request
of rproc_stop was done and no need to hold the wakeup source
in crash handler to recover any more.

Signed-off-by: Maria Yu <[email protected]>
---
drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
1 file changed, 11 insertions(+)

diff --git a/drivers/remoteproc/remoteproc_core.c
b/drivers/remoteproc/remoteproc_core.c
index e5279ed9a8d7..6bc7b8b7d01e 100644
--- a/drivers/remoteproc/remoteproc_core.c
+++ b/drivers/remoteproc/remoteproc_core.c
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct
work_struct *work)
if (rproc->state == RPROC_CRASHED || rproc->state ==
RPROC_OFFLINE) {
/* handle only the first crash detected */
mutex_unlock(&rproc->lock);
+ /*
+ * RPROC_OFFLINE state indicate there is no recovery process
+ * is in progress and no chance to have pm_relax in place.
+ * Because when recovering from crash, rproc->lock is held and
+ * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+ * and then unlock rproc->lock.
+ * RPROC_OFFLINE is only an intermediate state in recovery
+ * process.
+ */
+ if (rproc->state == RPROC_OFFLINE)
+ pm_relax(rproc->dev.parent);
return;
}

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines

--
2.7.4

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Quoted 8 lines
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

#25 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

Quoted 5 lines · Mathieu Poirier

I had a couple of good discussions with our power management expert and even then, the way forward isn't as clear as I would have liked. I am currently travelling and as such don't have the required time to go into greater details, something I will be doing next week.

Thx Mathieu for the info updated.
I'll wait for your update next week then.
let me know any initial questions that you have, perhaps I can also discuss that with our power team at the same time.

Quoted 489 lines · Aiqun(Maria) Yu
Quoted 488 lines

Hi,

Quoted 44 lines · Mathieu Poirier
Quoted 43 lines · Aiqun(Maria) Yu
Quoted 41 lines · Mathieu Poirier
Quoted 6 lines · Aiqun(Maria) Yu

Hi,

Let me think about this carefully.

When in RPROC_RECOVERY_FAIL case we want to re-do the recovery process again or just leave the pm_relax?

Neither.

When a recovery fail we don't want to call pm_relax(). The code in rproc_crash_handler_work() becomes:

     if (rproc->state == RPROC_OFFLINE) {
                   /* We have raced with rproc_shutdown() */
                   pm_relax()
             mutex_unlock(&rproc->lock);
             return;
           }

     if (rproc->state == RPROC_CRASHED ||
               rproc->state == RPROC_RECOVERY_FAILED) {
             /* handle only the first crash detected */
             mutex_unlock(&rproc->lock);
             return;
     }

RPROC_RECOVERY_FAILED gets set in rproc_boot_recovery() if request_firmware() or rproc_start() fail. Function rproc_trigger_recovery() needs to allow for the recovery the the remote processor is in RPROC_RECOVERY_FAILED state. As such the condition becomes:

           /* State could have changed before we got the mutex */
     if (rproc->state != RPROC_CRASHED &&
               rproc->state != RPROC_RECOVERY_FAILED)
             goto unlock_mutex;

Start with that and we can look at corner cases (if some exists) with a fresh patchset. Note that I have not addressed the attach/detach() scenario in the above.

Quoted 6 lines
Quoted 2 lines

If we didn't deal with the recovery failed case with correct pm_relax call, it may left the device in a state that cannot enter to suspend state.

That is what I am looking for. We don't want to give the impression that everything is fine by allowing the device to suspend. If the remote processor can't be recovered than it needs to be dealth with.

For the normal recovery failed case, it still need to do pm_relax to not prevent the device goes to suspend. It is what in normal recovery failed case we do in rproc_crash_handler_work as well. rproc_crash_handler_work will not check the result of the rproc_trigger_recovery return value, and will always do pm_relax.

For current conconrency cornor case as well, it is better to consistant with the current design of recovery fail senarios in normal cases.

I personally agree that we shouldn't do nothing when it is a RPROC_RECOVERY_FAILED senario when it is in rproc_crash_handler_work check, because it maybe crash happened when it is trying to do the recovery. So I suggested to do a continue try of trigger recovery again instead of doing nothing and bail out if it is a RPROC_RECOVERY_FAILED state.

Quoted 5 lines
Quoted 4 lines

Because first PROC_RECOVERY_FAIL case cannot ensure it have pm_relax called before the second crash handler call pm_stay_awake or not.

Quoted 7 lines

I've been thinking about that part. I don't think adding a wake_count to control calls to pm_stay_awake()/pm_relax() is the best way to go. There is a similar count happening in the PM runtime subsystem and that is what we should be using. I have asked a power management expert at Linaro for guidance with this matter. I should be able to get back to you with a way forward by the end of next week.

Thx for the specific date provided as well. I will wait until your reply for next patchset then.

Quoted 392 lines
Quoted 392 lines

So, What about the atomic count along with pm_relax and pm_stay_awake ?

struct rproc{
...
atomic_t wake_count;
...
}

rproc_pm_stay_awake()

{
      atomic_inc(&wake_count);
      pm_stay_awake();
}

rproc_pm_relax()

{
      if (atomic_dec_return(&wake_count) == 0)
              pm_stay_awake();
}

can refer code like:

rproc_report_crash()

{
      ...
      rproc_pm_stay_awake();
      queue_work();
      ...
}

rproc_crash_handler_work()

{
      ...
      if (rproc->state == RPROC_OFFLINE || rproc->state == RPROC_CRASHED) {
                    /* We have raced with rproc_shutdown() */
                    rproc_pm_relax();
              mutex_unlock(&rproc->lock);
              return;
            }
      ...
}
Quoted 350 lines

Thanks,
Mathieu

Quoted 346 lines

recovery fail case 1:

|                                      |firstcrash interrupt issued
| second crashed interrupt issued      | rproc_report_crash()
| rproc_report_crash()                 |          pm_stay_awake()
|          pm_stay_awake()             |          queue_work()
|          queue_work()                |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|rproc_crash_handler_work()            |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);             |pm_relax()
|if (rproc->state == RPROC_OFFLINE)    |
|return // shouldn't do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

recovery fail case 2:

|                                      |firstcrash interrupt issued
|                                      | rproc_report_crash()
|                                      |          pm_stay_awake()
|                                      |          queue_work()
|                                      |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|                                      |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|                                      |pm_relax()
|
| second crashed interrupt issued      |
| rproc_report_crash()                 |
|          pm_stay_awake()             |
|          queue_work()                |
|pm_stay_awake()
|mutex_lock(&rproc->lock);
|if (rproc->state == RPROC_OFFLINE)    |
|return // still need do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

Maybe I can have:
1. the pm_stay_awake and pm_relax with count based and call with paired for fix current concurency issue.
2. RPROC_RECOVERY_FAIL can be another patch for continue try to do recovery work.
3. handle RPROC_DETACHED case.

Quoted 191 lines · Mathieu Poirier
Quoted 127 lines · Arnaud POULIQUEN

Hi,

Quoted 104 lines · Aiqun(Maria) Yu
Quoted 86 lines · Mathieu Poirier
Quoted 77 lines · Aiqun(Maria) Yu
Quoted 67 lines · Mathieu Poirier
Quoted 66 lines · Mathieu Poirier

On Thu, Oct 13, 2022 at 09:40:09AM +0800, Aiqun(Maria) Yu wrote:
Hi Mathieu,

On 10/13/2022 4:43 AM, Mathieu Poirier wrote:
Please add what has changed from one version to another, either in a cover
letter or after the "Signed-off-by". There are many examples on how to
do that
on the mailing list.

Thx for the information, will take a note and benefit for next time.

On Fri, Sep 16, 2022 at 03:12:31PM +0800, Maria Yu wrote:
RPROC_OFFLINE state indicate there is no recovery process
is in progress and no chance to do the pm_relax.
Because when recovering from crash, rproc->lock is held and
state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to
RPROC_RUNNING
when rproc_trigger_recovery() returns. If that is not the case then
something
went wrong.

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before
returning,
so we know the remote processor was stopped. Therefore if rproc->state
is set
to RPROC_OFFLINE something went wrong in either request_firmware() or
rproc_start(). Either way the remote processor is offline and the system
probably
in an unknown/unstable. As such I don't see how calling pm_relax() can help
things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully
finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and
last rproc_trigger_recovery bailed out with only
rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair.
Since the subsystem may still can be recovered with customer's next trigger
of rproc_start, and we can make each error out path clean with pm resources.

I suggest spending time understanding what leads to the failure when
recovering
from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when
rproc_shutdown is triggered at similar time. So not an issue from error out
of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned
from
the beginning. What I see happening is the following:

rproc_report_crash()
pm_stay_awake()
queue_work() // current thread is suspended

rproc_shutdown()
rproc_stop()
rproc->state = RPROC_OFFLINE;

rproc_crash_handler_work()
if (rproc->state == RPROC_OFFLINE)
return // pm_relax() is not called

Quoted 8 lines
Quoted 2 lines

The right way to fix this is to add a pm_relax() in rproc_shutdown() and
rproc_detach(), along with a very descriptive comment as to why it is needed.

Thinking about this further there are more ramifications to consider. Please confirm the above scenario is what you are facing. I will advise on how to move
forward if that is the case.

Quoted 4 lines

Not sure if the situation is clear or not. So resend the email again.

The above senario is what customer is facing. crash hanppened while at the same time shutdown is triggered.

Unfortunately this is not enough details to address a problem as complex as this one.

Quoted 7 lines
Quoted 2 lines

And the device cannto goes to suspend state after that. the subsystem can still be start normally after this.

If the code flow I pasted above reflects the problem at hand, the current patch will not be sufficient to address the issue. If Arnaud confirms my suspicions we will have to think about a better solution.

Hi Mathiew,

Could you pls have more details of any side effects other then power issue of the current senario?
Why the current patch is not sufficient pls?

Have the current senario in details with rproc->lock information in details:

Quoted 18 lines
| subsystem crashed interrupt issued      | user trigger shutdown
| rproc_report_crash()                    |
|          pm_stay_awake()                |
|          queue_work()                   |
|                                         |rproc_shutdown
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_stop()
|rproc_crash_handler_work()               |rproc->state = RPROC_OFFLINE;
|                                         |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);                |
|if (rproc->state == RPROC_OFFLINE)       |
|return // pm_relax() is not called       |rproc_boot
|mutex_unlock(&rproc->lock);              |
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_start()
|                                         |mutex_unlock(&rproc->lock);
Quoted 2 lines

Agree with Mathieu, this is not so simple.

Thanks for looking into this.

Quoted 12 lines

Here is my view hoping I haven't missed a point in your discussion or an other corner cases.

I tried to analyze the issues (in what follows, the term "condition" means the "if" condition in which Aiqun proposes to add the fix) :

I can see 4 use cases with race condition

1) crash report while already one is treated (rproc_boot_recovery called) => not a real use case as if the remote processor is crashed we should not have a second crash report

That part is of great concern to me. *Theoretically* we should not get a new crash report while one has already been dispatched but the current code accounts for this scenario and as such the possibility can't be dismissed. Therefore we need to expect rproc_report_crash() to be called multiple times before a single instance of rproc_boot_recovery() is scheduled.

Quoted 22 lines

2) rproc_stop executed between the queuing of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_OFFLINE
=> we enter in the "condition" and the pm_relax has to be called => This commit fix should solve this use case

3) rproc_detach executed between the queue of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_DETACHED;
=> we don't go in "the condition" and issue because the recovery reattach
to the remote processor
=> but pm_relax is called
=> probably need an extra fix to avoid to re-attach

4) crash report while already one is treated (rproc_attach_recovery called) this one corresponds to an auto reboot of the remote processor, with a new crash
=> rproc->state = RPROC_CRASHED or rproc->state = RPROC_DETACHED; 4)a) rproc->state = RPROC_CRASHED if rproc->recovery_disabled = true
=> should call pm_relax if rproc->recovery_disabled = true => commit does not work for this use case

Quoted 5 lines

4)b) rproc->state = RPROC_DETACHED if recovery fails
=> error case with an unstable state
=> how to differentiate it from the use case 3) ? => introduce a RPROC_RECOVERY_FAIL state?

The case where a recovery fails needs to be considered and is the reason the original patch doesn't work. Right now in rproc_crash_handler_work(), it is not possible to differentiate between a legitimate shutdown request (scenario #2 above) and a recovery that went wrong. I think introducing RPROC_RECOVERY_FAIL would greatly simplify things.

Quoted 9 lines

My initial evaluation had not considered the attach/detach scenarios - thanks for adding that in the mix.

Aiqun, please send a new patchset that adds a new remote processor state, i.e RPROC_RECOVERY_FAIL. There should also be another patch in that set that takes attach/detach scenarios into account. The code between the v6.0 and v6.1 cycle has changed a lot in that area so make sure to properly rebase.

I will try.

Quoted 88 lines
Quoted 88 lines

Then pm_stay_awake is called when the crash work is queued. It seems to me coherent to call the pm_relax in the work handler.

Here is a quick and dirty patch (not tested) that should take into account the main use cases ( except 1) and 4)b) )

(quoted diff) +11 -1
@@ -2009,8 +2009,18 @@ static void rproc_crash_handler_work(struct work_struct *work)
 
            mutex_lock(&rproc->lock);
 
-       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
+       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE ||
+           rproc->state == RPROC_DETACHED) {
                    /* handle only the first crash detected */
+
+               /*
+                * call pm-relax in following use cases:
+                * - the remote processor has been stopped by the user
+                * - the remote processor is detached
+                + - the remote proc has an autonomous reset but recovery_disabled is true.
+                */
+               if(rproc->state != RPROC_CRASHED || rproc->recovery_disabled)
+                       pm_relax(rproc->dev.parent);
                    mutex_unlock(&rproc->lock);
                    return;
            }

Regards,
Arnaud

Quoted 55 lines
Quoted 53 lines
Quoted 53 lines
Quoted 40 lines
Quoted 40 lines

Thanks,
Mathieu

When the state is in RPROC_OFFLINE it means separate request
of rproc_stop was done and no need to hold the wakeup source
in crash handler to recover any more.

Signed-off-by: Maria Yu <[email protected]>
---
drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
1 file changed, 11 insertions(+)

diff --git a/drivers/remoteproc/remoteproc_core.c
b/drivers/remoteproc/remoteproc_core.c
index e5279ed9a8d7..6bc7b8b7d01e 100644
--- a/drivers/remoteproc/remoteproc_core.c
+++ b/drivers/remoteproc/remoteproc_core.c
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct
work_struct *work)
if (rproc->state == RPROC_CRASHED || rproc->state ==
RPROC_OFFLINE) {
/* handle only the first crash detected */
mutex_unlock(&rproc->lock);
+ /*
+ * RPROC_OFFLINE state indicate there is no recovery process
+ * is in progress and no chance to have pm_relax in place.
+ * Because when recovering from crash, rproc->lock is held and
+ * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+ * and then unlock rproc->lock.
+ * RPROC_OFFLINE is only an intermediate state in recovery
+ * process.
+ */
+ if (rproc->state == RPROC_OFFLINE)
+ pm_relax(rproc->dev.parent);
return;
}

Quoted 8 lines
Quoted 8 lines

--
2.7.4

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Quoted 8 lines
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
maintainer
#26 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

Quoted 10 lines · Aiqun(Maria) Yu
Quoted 5 lines · Mathieu Poirier

I had a couple of good discussions with our power management expert and even then, the way forward isn't as clear as I would have liked. I am currently travelling and as such don't have the required time to go into greater details, something I will be doing next week.

Thx Mathieu for the info updated.
I'll wait for your update next week then.
let me know any initial questions that you have, perhaps I can also discuss that with our power team at the same time.

The problem is to determine exactly what the WQ_FREEZABLE flag does to the rproc_recovery_wq workueue. The documentation [1] indicate that work items on the WQ are drained before the system is suspended. What I understand from this is that if two work items are queued and one is executing at the time a system suspend is initiated, all three items will be executed before the system is allowed to be suspended. _If_ that is the case, there would not be a need to call pm_stay_awake() and pm_relax() at all.

On the other hand, the PM resource I spoke to thought that in reality things don't happen that way. Taking the same above scenario where 2 work items are queued and one is executing at the time of the suspend, only the work item that is executing will be allowed to execute to completion before the system is suspended. The remaining two items that were queued will not execute.

If that is the case then we do need to call pm_stay_awake() and pm_relax(), and find another strategy to fix this situation.

Until we have a clear view of how the WQ_FREEZABLE flag works, we won't be able to move forward with this patchset. Unfortunately, I currently do not have the time to look into this.

If you want to take on this investigation, keep in mind that any conclusion will need to be backed by a proof. That can be debug messages on a console output or a code reference in the workqueue core.

[1]. https://elixir.bootlin.com/linux/v6.1-rc4/source/Documentation/core-api/workqueue.rst#L184

Quoted 494 lines
Quoted 489 lines · Aiqun(Maria) Yu
Quoted 488 lines

Hi,

Quoted 44 lines · Mathieu Poirier
Quoted 43 lines · Aiqun(Maria) Yu
Quoted 41 lines · Mathieu Poirier
Quoted 6 lines · Aiqun(Maria) Yu

Hi,

Let me think about this carefully.

When in RPROC_RECOVERY_FAIL case we want to re-do the recovery process again or just leave the pm_relax?

Neither.

When a recovery fail we don't want to call pm_relax(). The code in rproc_crash_handler_work() becomes:

     if (rproc->state == RPROC_OFFLINE) {
                   /* We have raced with rproc_shutdown() */
                   pm_relax()
             mutex_unlock(&rproc->lock);
             return;
           }

     if (rproc->state == RPROC_CRASHED ||
               rproc->state == RPROC_RECOVERY_FAILED) {
             /* handle only the first crash detected */
             mutex_unlock(&rproc->lock);
             return;
     }

RPROC_RECOVERY_FAILED gets set in rproc_boot_recovery() if request_firmware() or rproc_start() fail. Function rproc_trigger_recovery() needs to allow for the recovery the the remote processor is in RPROC_RECOVERY_FAILED state. As such the condition becomes:

           /* State could have changed before we got the mutex */
     if (rproc->state != RPROC_CRASHED &&
               rproc->state != RPROC_RECOVERY_FAILED)
             goto unlock_mutex;

Start with that and we can look at corner cases (if some exists) with a fresh patchset. Note that I have not addressed the attach/detach() scenario in the above.

Quoted 6 lines
Quoted 2 lines

If we didn't deal with the recovery failed case with correct pm_relax call, it may left the device in a state that cannot enter to suspend state.

That is what I am looking for. We don't want to give the impression that everything is fine by allowing the device to suspend. If the remote processor can't be recovered than it needs to be dealth with.

For the normal recovery failed case, it still need to do pm_relax to not prevent the device goes to suspend. It is what in normal recovery failed case we do in rproc_crash_handler_work as well. rproc_crash_handler_work will not check the result of the rproc_trigger_recovery return value, and will always do pm_relax.

For current conconrency cornor case as well, it is better to consistant with the current design of recovery fail senarios in normal cases.

I personally agree that we shouldn't do nothing when it is a RPROC_RECOVERY_FAILED senario when it is in rproc_crash_handler_work check, because it maybe crash happened when it is trying to do the recovery. So I suggested to do a continue try of trigger recovery again instead of doing nothing and bail out if it is a RPROC_RECOVERY_FAILED state.

Quoted 5 lines
Quoted 4 lines

Because first PROC_RECOVERY_FAIL case cannot ensure it have pm_relax called before the second crash handler call pm_stay_awake or not.

Quoted 7 lines

I've been thinking about that part. I don't think adding a wake_count to control calls to pm_stay_awake()/pm_relax() is the best way to go. There is a similar count happening in the PM runtime subsystem and that is what we should be using. I have asked a power management expert at Linaro for guidance with this matter. I should be able to get back to you with a way forward by the end of next week.

Thx for the specific date provided as well. I will wait until your reply for next patchset then.

Quoted 392 lines
Quoted 392 lines

So, What about the atomic count along with pm_relax and pm_stay_awake ?

struct rproc{
...
atomic_t wake_count;
...
}

rproc_pm_stay_awake()

{
      atomic_inc(&wake_count);
      pm_stay_awake();
}

rproc_pm_relax()

{
      if (atomic_dec_return(&wake_count) == 0)
              pm_stay_awake();
}

can refer code like:

rproc_report_crash()

{
      ...
      rproc_pm_stay_awake();
      queue_work();
      ...
}

rproc_crash_handler_work()

{
      ...
      if (rproc->state == RPROC_OFFLINE || rproc->state == RPROC_CRASHED) {
                    /* We have raced with rproc_shutdown() */
                    rproc_pm_relax();
              mutex_unlock(&rproc->lock);
              return;
            }
      ...
}
Quoted 350 lines

Thanks,
Mathieu

Quoted 346 lines

recovery fail case 1:

|                                      |firstcrash interrupt issued
| second crashed interrupt issued      | rproc_report_crash()
| rproc_report_crash()                 |          pm_stay_awake()
|          pm_stay_awake()             |          queue_work()
|          queue_work()                |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|rproc_crash_handler_work()            |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);             |pm_relax()
|if (rproc->state == RPROC_OFFLINE)    |
|return // shouldn't do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

recovery fail case 2:

|                                      |firstcrash interrupt issued
|                                      | rproc_report_crash()
|                                      |          pm_stay_awake()
|                                      |          queue_work()
|                                      |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|                                      |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|                                      |pm_relax()
|
| second crashed interrupt issued      |
| rproc_report_crash()                 |
|          pm_stay_awake()             |
|          queue_work()                |
|pm_stay_awake()
|mutex_lock(&rproc->lock);
|if (rproc->state == RPROC_OFFLINE)    |
|return // still need do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

Maybe I can have:
1. the pm_stay_awake and pm_relax with count based and call with paired for fix current concurency issue.
2. RPROC_RECOVERY_FAIL can be another patch for continue try to do recovery work.
3. handle RPROC_DETACHED case.

Quoted 191 lines · Mathieu Poirier
Quoted 127 lines · Arnaud POULIQUEN

Hi,

Quoted 104 lines · Aiqun(Maria) Yu
Quoted 86 lines · Mathieu Poirier
Quoted 77 lines · Aiqun(Maria) Yu
Quoted 75 lines · Mathieu Poirier

On Thu, Oct 13, 2022 at 11:34:42AM -0600, Mathieu Poirier wrote:
On Thu, Oct 13, 2022 at 09:40:09AM +0800, Aiqun(Maria) Yu wrote:
Hi Mathieu,

On 10/13/2022 4:43 AM, Mathieu Poirier wrote:
Please add what has changed from one version to another, either in a cover
letter or after the "Signed-off-by". There are many examples on how to
do that
on the mailing list.

Thx for the information, will take a note and benefit for next time.

On Fri, Sep 16, 2022 at 03:12:31PM +0800, Maria Yu wrote:
RPROC_OFFLINE state indicate there is no recovery process
is in progress and no chance to do the pm_relax.
Because when recovering from crash, rproc->lock is held and
state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to
RPROC_RUNNING
when rproc_trigger_recovery() returns. If that is not the case then
something
went wrong.

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before
returning,
so we know the remote processor was stopped. Therefore if rproc->state
is set
to RPROC_OFFLINE something went wrong in either request_firmware() or
rproc_start(). Either way the remote processor is offline and the system
probably
in an unknown/unstable. As such I don't see how calling pm_relax() can help
things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully
finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and
last rproc_trigger_recovery bailed out with only
rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair.
Since the subsystem may still can be recovered with customer's next trigger
of rproc_start, and we can make each error out path clean with pm resources.

I suggest spending time understanding what leads to the failure when
recovering
from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when
rproc_shutdown is triggered at similar time. So not an issue from error out
of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned
from
the beginning. What I see happening is the following:

rproc_report_crash()
pm_stay_awake()
queue_work() // current thread is suspended

rproc_shutdown()
rproc_stop()
rproc->state = RPROC_OFFLINE;

rproc_crash_handler_work()
if (rproc->state == RPROC_OFFLINE)
return // pm_relax() is not called

The right way to fix this is to add a pm_relax() in rproc_shutdown() and
rproc_detach(), along with a very descriptive comment as to why it is needed.

Thinking about this further there are more ramifications to consider. Please
confirm the above scenario is what you are facing. I will advise on how to
move
forward if that is the case.

Quoted 4 lines

Not sure if the situation is clear or not. So resend the email again.

The above senario is what customer is facing. crash hanppened while at the same time shutdown is triggered.

Unfortunately this is not enough details to address a problem as complex as this one.

Quoted 7 lines
Quoted 2 lines

And the device cannto goes to suspend state after that. the subsystem can still be start normally after this.

If the code flow I pasted above reflects the problem at hand, the current patch will not be sufficient to address the issue. If Arnaud confirms my suspicions we will have to think about a better solution.

Hi Mathiew,

Could you pls have more details of any side effects other then power issue of the current senario?
Why the current patch is not sufficient pls?

Have the current senario in details with rproc->lock information in details:

Quoted 18 lines
| subsystem crashed interrupt issued      | user trigger shutdown
| rproc_report_crash()                    |
|          pm_stay_awake()                |
|          queue_work()                   |
|                                         |rproc_shutdown
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_stop()
|rproc_crash_handler_work()               |rproc->state = RPROC_OFFLINE;
|                                         |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);                |
|if (rproc->state == RPROC_OFFLINE)       |
|return // pm_relax() is not called       |rproc_boot
|mutex_unlock(&rproc->lock);              |
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_start()
|                                         |mutex_unlock(&rproc->lock);
Quoted 2 lines

Agree with Mathieu, this is not so simple.

Thanks for looking into this.

Quoted 12 lines

Here is my view hoping I haven't missed a point in your discussion or an other corner cases.

I tried to analyze the issues (in what follows, the term "condition" means the "if" condition in which Aiqun proposes to add the fix) :

I can see 4 use cases with race condition

1) crash report while already one is treated (rproc_boot_recovery called) => not a real use case as if the remote processor is crashed we should not have a second crash report

That part is of great concern to me. *Theoretically* we should not get a new crash report while one has already been dispatched but the current code accounts for this scenario and as such the possibility can't be dismissed. Therefore we need to expect rproc_report_crash() to be called multiple times before a single instance of rproc_boot_recovery() is scheduled.

Quoted 22 lines

2) rproc_stop executed between the queuing of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_OFFLINE
=> we enter in the "condition" and the pm_relax has to be called => This commit fix should solve this use case

3) rproc_detach executed between the queue of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_DETACHED;
=> we don't go in "the condition" and issue because the recovery reattach
to the remote processor
=> but pm_relax is called
=> probably need an extra fix to avoid to re-attach

4) crash report while already one is treated (rproc_attach_recovery called) this one corresponds to an auto reboot of the remote processor, with a new crash
=> rproc->state = RPROC_CRASHED or rproc->state = RPROC_DETACHED; 4)a) rproc->state = RPROC_CRASHED if rproc->recovery_disabled = true
=> should call pm_relax if rproc->recovery_disabled = true => commit does not work for this use case

Quoted 5 lines

4)b) rproc->state = RPROC_DETACHED if recovery fails
=> error case with an unstable state
=> how to differentiate it from the use case 3) ? => introduce a RPROC_RECOVERY_FAIL state?

The case where a recovery fails needs to be considered and is the reason the original patch doesn't work. Right now in rproc_crash_handler_work(), it is not possible to differentiate between a legitimate shutdown request (scenario #2 above) and a recovery that went wrong. I think introducing RPROC_RECOVERY_FAIL would greatly simplify things.

Quoted 9 lines

My initial evaluation had not considered the attach/detach scenarios - thanks for adding that in the mix.

Aiqun, please send a new patchset that adds a new remote processor state, i.e RPROC_RECOVERY_FAIL. There should also be another patch in that set that takes attach/detach scenarios into account. The code between the v6.0 and v6.1 cycle has changed a lot in that area so make sure to properly rebase.

I will try.

Quoted 88 lines
Quoted 88 lines

Then pm_stay_awake is called when the crash work is queued. It seems to me coherent to call the pm_relax in the work handler.

Here is a quick and dirty patch (not tested) that should take into account the main use cases ( except 1) and 4)b) )

(quoted diff) +11 -1
@@ -2009,8 +2009,18 @@ static void rproc_crash_handler_work(struct work_struct *work)
 
            mutex_lock(&rproc->lock);
 
-       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
+       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE ||
+           rproc->state == RPROC_DETACHED) {
                    /* handle only the first crash detected */
+
+               /*
+                * call pm-relax in following use cases:
+                * - the remote processor has been stopped by the user
+                * - the remote processor is detached
+                + - the remote proc has an autonomous reset but recovery_disabled is true.
+                */
+               if(rproc->state != RPROC_CRASHED || rproc->recovery_disabled)
+                       pm_relax(rproc->dev.parent);
                    mutex_unlock(&rproc->lock);
                    return;
            }

Regards,
Arnaud

Quoted 55 lines
Quoted 53 lines
Quoted 53 lines
Quoted 48 lines

Thanks,
Mathieu

When the state is in RPROC_OFFLINE it means separate request
of rproc_stop was done and no need to hold the wakeup source
in crash handler to recover any more.

Signed-off-by: Maria Yu <[email protected]>
---
drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
1 file changed, 11 insertions(+)

diff --git a/drivers/remoteproc/remoteproc_core.c
b/drivers/remoteproc/remoteproc_core.c
index e5279ed9a8d7..6bc7b8b7d01e 100644
--- a/drivers/remoteproc/remoteproc_core.c
+++ b/drivers/remoteproc/remoteproc_core.c
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct
work_struct *work)
if (rproc->state == RPROC_CRASHED || rproc->state ==
RPROC_OFFLINE) {
/* handle only the first crash detected */
mutex_unlock(&rproc->lock);
+ /*
+ * RPROC_OFFLINE state indicate there is no recovery process
+ * is in progress and no chance to have pm_relax in place.
+ * Because when recovering from crash, rproc->lock is held and
+ * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+ * and then unlock rproc->lock.
+ * RPROC_OFFLINE is only an intermediate state in recovery
+ * process.
+ */
+ if (rproc->state == RPROC_OFFLINE)
+ pm_relax(rproc->dev.parent);
return;
}
--
2.7.4

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Quoted 8 lines
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
#27 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

Hi,

Quoted 26 lines · Mathieu Poirier
Quoted 10 lines · Aiqun(Maria) Yu
Quoted 5 lines · Mathieu Poirier

I had a couple of good discussions with our power management expert and even then, the way forward isn't as clear as I would have liked. I am currently travelling and as such don't have the required time to go into greater details, something I will be doing next week.

Thx Mathieu for the info updated.
I'll wait for your update next week then.
let me know any initial questions that you have, perhaps I can also discuss that with our power team at the same time.

The problem is to determine exactly what the WQ_FREEZABLE flag does to the rproc_recovery_wq workueue. The documentation [1] indicate that work items on the WQ are drained before the system is suspended. What I understand from this is that if two work items are queued and one is executing at the time a system suspend is initiated, all three items will be executed before the system is allowed to be suspended. _If_ that is the case, there would not be a need to call pm_stay_awake() and pm_relax() at all.

On the other hand, the PM resource I spoke to thought that in reality things don't happen that way. Taking the same above scenario where 2 work items are queued and one is executing at the time of the suspend, only the work item that is executing will be allowed to execute to completion before the system is suspended. The remaining two items that were queued will not execute.

Quoted 6 lines

If that is the case then we do need to call pm_stay_awake() and pm_relax(), and find another strategy to fix this situation.

Until we have a clear view of how the WQ_FREEZABLE flag works, we won't be able to move forward with this patchset. Unfortunately, I currently do not have the time to look into this.

I had a check on the WQ_FREEZABLE flag, here is my understanding:

when the interrupt happened, it still need pm_stay_awake to make sure queue_work action can active the work instead of susepend the device.

1. If WQ_FREEZABLE, pwq->max_active = 0; // maximum number of in-flight work items is set to 0.

[1]. https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L3748

2. If WQ_FREEZABLE, will only check pwq->nr_active to see if there is still freeze_workqueues_busy. [2]. https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L5270

3. When in queue_work, if max_active is 0, when do queue_work it will not actually active the work. [3]. https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L1418

for the current issue, the work is already complete and forget to set pm_relax in some condition that make the system cannot be suspended.

Quoted 502 lines

If you want to take on this investigation, keep in mind that any conclusion will need to be backed by a proof. That can be debug messages on a console output or a code reference in the workqueue core.

[1]. https://elixir.bootlin.com/linux/v6.1-rc4/source/Documentation/core-api/workqueue.rst#L184

Quoted 494 lines
Quoted 489 lines · Aiqun(Maria) Yu
Quoted 488 lines

Hi,

Quoted 44 lines · Mathieu Poirier
Quoted 43 lines · Aiqun(Maria) Yu
Quoted 41 lines · Mathieu Poirier
Quoted 6 lines · Aiqun(Maria) Yu

Hi,

Let me think about this carefully.

When in RPROC_RECOVERY_FAIL case we want to re-do the recovery process again or just leave the pm_relax?

Neither.

When a recovery fail we don't want to call pm_relax(). The code in rproc_crash_handler_work() becomes:

      if (rproc->state == RPROC_OFFLINE) {
                    /* We have raced with rproc_shutdown() */
                    pm_relax()
              mutex_unlock(&rproc->lock);
              return;
            }

      if (rproc->state == RPROC_CRASHED ||
                rproc->state == RPROC_RECOVERY_FAILED) {
              /* handle only the first crash detected */
              mutex_unlock(&rproc->lock);
              return;
      }

RPROC_RECOVERY_FAILED gets set in rproc_boot_recovery() if request_firmware() or rproc_start() fail. Function rproc_trigger_recovery() needs to allow for the recovery the the remote processor is in RPROC_RECOVERY_FAILED state. As such the condition becomes:

            /* State could have changed before we got the mutex */
      if (rproc->state != RPROC_CRASHED &&
                rproc->state != RPROC_RECOVERY_FAILED)
              goto unlock_mutex;

Start with that and we can look at corner cases (if some exists) with a fresh patchset. Note that I have not addressed the attach/detach() scenario in the above.

Quoted 6 lines
Quoted 2 lines

If we didn't deal with the recovery failed case with correct pm_relax call, it may left the device in a state that cannot enter to suspend state.

That is what I am looking for. We don't want to give the impression that everything is fine by allowing the device to suspend. If the remote processor can't be recovered than it needs to be dealth with.

For the normal recovery failed case, it still need to do pm_relax to not prevent the device goes to suspend. It is what in normal recovery failed case we do in rproc_crash_handler_work as well. rproc_crash_handler_work will not check the result of the rproc_trigger_recovery return value, and will always do pm_relax.

For current conconrency cornor case as well, it is better to consistant with the current design of recovery fail senarios in normal cases.

I personally agree that we shouldn't do nothing when it is a RPROC_RECOVERY_FAILED senario when it is in rproc_crash_handler_work check, because it maybe crash happened when it is trying to do the recovery. So I suggested to do a continue try of trigger recovery again instead of doing nothing and bail out if it is a RPROC_RECOVERY_FAILED state.

Quoted 5 lines
Quoted 4 lines

Because first PROC_RECOVERY_FAIL case cannot ensure it have pm_relax called before the second crash handler call pm_stay_awake or not.

Quoted 7 lines

I've been thinking about that part. I don't think adding a wake_count to control calls to pm_stay_awake()/pm_relax() is the best way to go. There is a similar count happening in the PM runtime subsystem and that is what we should be using. I have asked a power management expert at Linaro for guidance with this matter. I should be able to get back to you with a way forward by the end of next week.

Thx for the specific date provided as well. I will wait until your reply for next patchset then.

Quoted 392 lines
Quoted 392 lines

So, What about the atomic count along with pm_relax and pm_stay_awake ?

struct rproc{
...
atomic_t wake_count;
...
}

rproc_pm_stay_awake()

{
       atomic_inc(&wake_count);
       pm_stay_awake();
}

rproc_pm_relax()

{
       if (atomic_dec_return(&wake_count) == 0)
               pm_stay_awake();
}

can refer code like:

rproc_report_crash()

{
       ...
       rproc_pm_stay_awake();
       queue_work();
       ...
}

rproc_crash_handler_work()

{
       ...
       if (rproc->state == RPROC_OFFLINE || rproc->state == RPROC_CRASHED) {
                     /* We have raced with rproc_shutdown() */
                     rproc_pm_relax();
               mutex_unlock(&rproc->lock);
               return;
             }
       ...
}
Quoted 350 lines

Thanks,
Mathieu

Quoted 346 lines

recovery fail case 1:

|                                      |firstcrash interrupt issued
| second crashed interrupt issued      | rproc_report_crash()
| rproc_report_crash()                 |          pm_stay_awake()
|          pm_stay_awake()             |          queue_work()
|          queue_work()                |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|rproc_crash_handler_work()            |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);             |pm_relax()
|if (rproc->state == RPROC_OFFLINE)    |
|return // shouldn't do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

recovery fail case 2:

|                                      |firstcrash interrupt issued
|                                      | rproc_report_crash()
|                                      |          pm_stay_awake()
|                                      |          queue_work()
|                                      |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|                                      |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|                                      |pm_relax()
|
| second crashed interrupt issued      |
| rproc_report_crash()                 |
|          pm_stay_awake()             |
|          queue_work()                |
|pm_stay_awake()
|mutex_lock(&rproc->lock);
|if (rproc->state == RPROC_OFFLINE)    |
|return // still need do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

Maybe I can have:
1. the pm_stay_awake and pm_relax with count based and call with paired for fix current concurency issue.
2. RPROC_RECOVERY_FAIL can be another patch for continue try to do recovery work.
3. handle RPROC_DETACHED case.

Quoted 191 lines · Mathieu Poirier
Quoted 127 lines · Arnaud POULIQUEN

Hi,

Quoted 104 lines · Aiqun(Maria) Yu
Quoted 86 lines · Mathieu Poirier
Quoted 81 lines · Aiqun(Maria) Yu

On 10/14/2022 2:03 AM, Mathieu Poirier wrote:
On Thu, Oct 13, 2022 at 11:34:42AM -0600, Mathieu Poirier wrote:
On Thu, Oct 13, 2022 at 09:40:09AM +0800, Aiqun(Maria) Yu wrote:
Hi Mathieu,

On 10/13/2022 4:43 AM, Mathieu Poirier wrote:
Please add what has changed from one version to another, either in a cover
letter or after the "Signed-off-by". There are many examples on how to
do that
on the mailing list.

Thx for the information, will take a note and benefit for next time.

On Fri, Sep 16, 2022 at 03:12:31PM +0800, Maria Yu wrote:
RPROC_OFFLINE state indicate there is no recovery process
is in progress and no chance to do the pm_relax.
Because when recovering from crash, rproc->lock is held and
state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to
RPROC_RUNNING
when rproc_trigger_recovery() returns. If that is not the case then
something
went wrong.

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before
returning,
so we know the remote processor was stopped. Therefore if rproc->state
is set
to RPROC_OFFLINE something went wrong in either request_firmware() or
rproc_start(). Either way the remote processor is offline and the system
probably
in an unknown/unstable. As such I don't see how calling pm_relax() can help
things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully
finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and
last rproc_trigger_recovery bailed out with only
rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair.
Since the subsystem may still can be recovered with customer's next trigger
of rproc_start, and we can make each error out path clean with pm resources.

I suggest spending time understanding what leads to the failure when
recovering
from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when
rproc_shutdown is triggered at similar time. So not an issue from error out
of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned
from
the beginning. What I see happening is the following:

rproc_report_crash()
pm_stay_awake()
queue_work() // current thread is suspended

rproc_shutdown()
rproc_stop()
rproc->state = RPROC_OFFLINE;

rproc_crash_handler_work()
if (rproc->state == RPROC_OFFLINE)
return // pm_relax() is not called

The right way to fix this is to add a pm_relax() in rproc_shutdown() and
rproc_detach(), along with a very descriptive comment as to why it is needed.

Thinking about this further there are more ramifications to consider. Please
confirm the above scenario is what you are facing. I will advise on how to
move
forward if that is the case.

Not sure if the situation is clear or not. So resend the email again.

The above senario is what customer is facing. crash hanppened while at
the same time shutdown is triggered.

Unfortunately this is not enough details to address a problem as complex as this one.

Quoted 7 lines
Quoted 2 lines

And the device cannto goes to suspend state after that.
the subsystem can still be start normally after this.

If the code flow I pasted above reflects the problem at hand, the current patch will not be sufficient to address the issue. If Arnaud confirms my suspicions we will have to think about a better solution.

Hi Mathiew,

Could you pls have more details of any side effects other then power issue of the current senario?
Why the current patch is not sufficient pls?

Have the current senario in details with rproc->lock information in details:

Quoted 18 lines
| subsystem crashed interrupt issued      | user trigger shutdown
| rproc_report_crash()                    |
|          pm_stay_awake()                |
|          queue_work()                   |
|                                         |rproc_shutdown
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_stop()
|rproc_crash_handler_work()               |rproc->state = RPROC_OFFLINE;
|                                         |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);                |
|if (rproc->state == RPROC_OFFLINE)       |
|return // pm_relax() is not called       |rproc_boot
|mutex_unlock(&rproc->lock);              |
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_start()
|                                         |mutex_unlock(&rproc->lock);
Quoted 2 lines

Agree with Mathieu, this is not so simple.

Thanks for looking into this.

Quoted 12 lines

Here is my view hoping I haven't missed a point in your discussion or an other corner cases.

I tried to analyze the issues (in what follows, the term "condition" means the "if" condition in which Aiqun proposes to add the fix) :

I can see 4 use cases with race condition

1) crash report while already one is treated (rproc_boot_recovery called) => not a real use case as if the remote processor is crashed we should not have a second crash report

That part is of great concern to me. *Theoretically* we should not get a new crash report while one has already been dispatched but the current code accounts for this scenario and as such the possibility can't be dismissed. Therefore we need to expect rproc_report_crash() to be called multiple times before a single instance of rproc_boot_recovery() is scheduled.

Quoted 22 lines

2) rproc_stop executed between the queuing of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_OFFLINE
=> we enter in the "condition" and the pm_relax has to be called => This commit fix should solve this use case

3) rproc_detach executed between the queue of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_DETACHED;
=> we don't go in "the condition" and issue because the recovery reattach
to the remote processor
=> but pm_relax is called
=> probably need an extra fix to avoid to re-attach

4) crash report while already one is treated (rproc_attach_recovery called) this one corresponds to an auto reboot of the remote processor, with a new crash
=> rproc->state = RPROC_CRASHED or rproc->state = RPROC_DETACHED; 4)a) rproc->state = RPROC_CRASHED if rproc->recovery_disabled = true
=> should call pm_relax if rproc->recovery_disabled = true => commit does not work for this use case

Quoted 5 lines

4)b) rproc->state = RPROC_DETACHED if recovery fails
=> error case with an unstable state
=> how to differentiate it from the use case 3) ? => introduce a RPROC_RECOVERY_FAIL state?

The case where a recovery fails needs to be considered and is the reason the original patch doesn't work. Right now in rproc_crash_handler_work(), it is not possible to differentiate between a legitimate shutdown request (scenario #2 above) and a recovery that went wrong. I think introducing RPROC_RECOVERY_FAIL would greatly simplify things.

Quoted 9 lines

My initial evaluation had not considered the attach/detach scenarios - thanks for adding that in the mix.

Aiqun, please send a new patchset that adds a new remote processor state, i.e RPROC_RECOVERY_FAIL. There should also be another patch in that set that takes attach/detach scenarios into account. The code between the v6.0 and v6.1 cycle has changed a lot in that area so make sure to properly rebase.

I will try.

Quoted 88 lines
Quoted 88 lines

Then pm_stay_awake is called when the crash work is queued. It seems to me coherent to call the pm_relax in the work handler.

Here is a quick and dirty patch (not tested) that should take into account the main use cases ( except 1) and 4)b) )

(quoted diff) +11 -1
@@ -2009,8 +2009,18 @@ static void rproc_crash_handler_work(struct work_struct *work)
 
             mutex_lock(&rproc->lock);
 
-       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
+       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE ||
+           rproc->state == RPROC_DETACHED) {
                     /* handle only the first crash detected */
+
+               /*
+                * call pm-relax in following use cases:
+                * - the remote processor has been stopped by the user
+                * - the remote processor is detached
+                + - the remote proc has an autonomous reset but recovery_disabled is true.
+                */
+               if(rproc->state != RPROC_CRASHED || rproc->recovery_disabled)
+                       pm_relax(rproc->dev.parent);
                     mutex_unlock(&rproc->lock);
                     return;
             }

Regards,
Arnaud

Quoted 55 lines
Quoted 53 lines
Quoted 53 lines

Thanks,
Mathieu

When the state is in RPROC_OFFLINE it means separate request
of rproc_stop was done and no need to hold the wakeup source
in crash handler to recover any more.

Signed-off-by: Maria Yu <[email protected]>
---
drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
1 file changed, 11 insertions(+)

diff --git a/drivers/remoteproc/remoteproc_core.c
b/drivers/remoteproc/remoteproc_core.c
index e5279ed9a8d7..6bc7b8b7d01e 100644
--- a/drivers/remoteproc/remoteproc_core.c
+++ b/drivers/remoteproc/remoteproc_core.c
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct
work_struct *work)
if (rproc->state == RPROC_CRASHED || rproc->state ==
RPROC_OFFLINE) {
/* handle only the first crash detected */
mutex_unlock(&rproc->lock);
+ /*
+ * RPROC_OFFLINE state indicate there is no recovery process
+ * is in progress and no chance to have pm_relax in place.
+ * Because when recovering from crash, rproc->lock is held and
+ * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+ * and then unlock rproc->lock.
+ * RPROC_OFFLINE is only an intermediate state in recovery
+ * process.
+ */
+ if (rproc->state == RPROC_OFFLINE)
+ pm_relax(rproc->dev.parent);
return;
}
--
2.7.4

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Quoted 8 lines
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
maintainer
#28 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

Quoted 51 lines · Aiqun(Maria) Yu

Hi,

Quoted 26 lines · Mathieu Poirier
Quoted 10 lines · Aiqun(Maria) Yu
Quoted 5 lines · Mathieu Poirier

I had a couple of good discussions with our power management expert and even then, the way forward isn't as clear as I would have liked. I am currently travelling and as such don't have the required time to go into greater details, something I will be doing next week.

Thx Mathieu for the info updated.
I'll wait for your update next week then.
let me know any initial questions that you have, perhaps I can also discuss that with our power team at the same time.

The problem is to determine exactly what the WQ_FREEZABLE flag does to the rproc_recovery_wq workueue. The documentation [1] indicate that work items on the WQ are drained before the system is suspended. What I understand from this is that if two work items are queued and one is executing at the time a system suspend is initiated, all three items will be executed before the system is allowed to be suspended. _If_ that is the case, there would not be a need to call pm_stay_awake() and pm_relax() at all.

On the other hand, the PM resource I spoke to thought that in reality things don't happen that way. Taking the same above scenario where 2 work items are queued and one is executing at the time of the suspend, only the work item that is executing will be allowed to execute to completion before the system is suspended. The remaining two items that were queued will not execute.

Quoted 6 lines

If that is the case then we do need to call pm_stay_awake() and pm_relax(), and find another strategy to fix this situation.

Until we have a clear view of how the WQ_FREEZABLE flag works, we won't be able to move forward with this patchset. Unfortunately, I currently do not have the time to look into this.

I had a check on the WQ_FREEZABLE flag, here is my understanding:

when the interrupt happened, it still need pm_stay_awake to make sure queue_work action can active the work instead of susepend the device.

1. If WQ_FREEZABLE, pwq->max_active = 0; // maximum number of in-flight work items is set to 0.

[1].
https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L3748

2. If WQ_FREEZABLE, will only check pwq->nr_active to see if there is still freeze_workqueues_busy.
[2].
https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L5270

Quoted 7 lines

3. When in queue_work, if max_active is 0, when do queue_work it will not actually active the work.
[3].
https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L1418

for the current issue, the work is already complete and forget to set pm_relax in some condition that make the system cannot be suspended.

What you have above does not describe what happens to work already queue to a WQ_FREEZABLE workqueue when the system gets suspended, and that is the most important thing.

I ended up doing that investigation myself and found the documentation on WQ_FREEZABLE seems to be accurate. I am still discussing this with the Linaro power management team and will get back to you when we reach a conclusion.

Quoted 508 lines
Quoted 503 lines

If you want to take on this investigation, keep in mind that any conclusion will need to be backed by a proof. That can be debug messages on a console output or a code reference in the workqueue core.

[1]. https://elixir.bootlin.com/linux/v6.1-rc4/source/Documentation/core-api/workqueue.rst#L184

Quoted 494 lines
Quoted 489 lines · Aiqun(Maria) Yu
Quoted 488 lines

Hi,

Quoted 44 lines · Mathieu Poirier
Quoted 43 lines · Aiqun(Maria) Yu
Quoted 41 lines · Mathieu Poirier
Quoted 6 lines · Aiqun(Maria) Yu

Hi,

Let me think about this carefully.

When in RPROC_RECOVERY_FAIL case we want to re-do the recovery process again or just leave the pm_relax?

Neither.

When a recovery fail we don't want to call pm_relax(). The code in rproc_crash_handler_work() becomes:

      if (rproc->state == RPROC_OFFLINE) {
                    /* We have raced with rproc_shutdown() */
                    pm_relax()
              mutex_unlock(&rproc->lock);
              return;
            }

      if (rproc->state == RPROC_CRASHED ||
                rproc->state == RPROC_RECOVERY_FAILED) {
              /* handle only the first crash detected */
              mutex_unlock(&rproc->lock);
              return;
      }

RPROC_RECOVERY_FAILED gets set in rproc_boot_recovery() if request_firmware() or rproc_start() fail. Function rproc_trigger_recovery() needs to allow for the recovery the the remote processor is in RPROC_RECOVERY_FAILED state. As such the condition becomes:

            /* State could have changed before we got the mutex */
      if (rproc->state != RPROC_CRASHED &&
                rproc->state != RPROC_RECOVERY_FAILED)
              goto unlock_mutex;

Start with that and we can look at corner cases (if some exists) with a fresh patchset. Note that I have not addressed the attach/detach() scenario in the above.

Quoted 6 lines
Quoted 2 lines

If we didn't deal with the recovery failed case with correct pm_relax call, it may left the device in a state that cannot enter to suspend state.

That is what I am looking for. We don't want to give the impression that everything is fine by allowing the device to suspend. If the remote processor can't be recovered than it needs to be dealth with.

For the normal recovery failed case, it still need to do pm_relax to not prevent the device goes to suspend. It is what in normal recovery failed case we do in rproc_crash_handler_work as well. rproc_crash_handler_work will not check the result of the rproc_trigger_recovery return value, and will always do pm_relax.

For current conconrency cornor case as well, it is better to consistant with the current design of recovery fail senarios in normal cases.

I personally agree that we shouldn't do nothing when it is a RPROC_RECOVERY_FAILED senario when it is in rproc_crash_handler_work check, because it maybe crash happened when it is trying to do the recovery. So I suggested to do a continue try of trigger recovery again instead of doing nothing and bail out if it is a RPROC_RECOVERY_FAILED state.

Quoted 5 lines
Quoted 4 lines

Because first PROC_RECOVERY_FAIL case cannot ensure it have pm_relax called before the second crash handler call pm_stay_awake or not.

Quoted 7 lines

I've been thinking about that part. I don't think adding a wake_count to control calls to pm_stay_awake()/pm_relax() is the best way to go. There is a similar count happening in the PM runtime subsystem and that is what we should be using. I have asked a power management expert at Linaro for guidance with this matter. I should be able to get back to you with a way forward by the end of next week.

Thx for the specific date provided as well. I will wait until your reply for next patchset then.

Quoted 392 lines
Quoted 392 lines

So, What about the atomic count along with pm_relax and pm_stay_awake ?

struct rproc{
...
atomic_t wake_count;
...
}

rproc_pm_stay_awake()

{
       atomic_inc(&wake_count);
       pm_stay_awake();
}

rproc_pm_relax()

{
       if (atomic_dec_return(&wake_count) == 0)
               pm_stay_awake();
}

can refer code like:

rproc_report_crash()

{
       ...
       rproc_pm_stay_awake();
       queue_work();
       ...
}

rproc_crash_handler_work()

{
       ...
       if (rproc->state == RPROC_OFFLINE || rproc->state == RPROC_CRASHED) {
                     /* We have raced with rproc_shutdown() */
                     rproc_pm_relax();
               mutex_unlock(&rproc->lock);
               return;
             }
       ...
}
Quoted 350 lines

Thanks,
Mathieu

Quoted 346 lines

recovery fail case 1:

|                                      |firstcrash interrupt issued
| second crashed interrupt issued      | rproc_report_crash()
| rproc_report_crash()                 |          pm_stay_awake()
|          pm_stay_awake()             |          queue_work()
|          queue_work()                |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|rproc_crash_handler_work()            |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);             |pm_relax()
|if (rproc->state == RPROC_OFFLINE)    |
|return // shouldn't do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

recovery fail case 2:

|                                      |firstcrash interrupt issued
|                                      | rproc_report_crash()
|                                      |          pm_stay_awake()
|                                      |          queue_work()
|                                      |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|                                      |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|                                      |pm_relax()
|
| second crashed interrupt issued      |
| rproc_report_crash()                 |
|          pm_stay_awake()             |
|          queue_work()                |
|pm_stay_awake()
|mutex_lock(&rproc->lock);
|if (rproc->state == RPROC_OFFLINE)    |
|return // still need do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

Maybe I can have:
1. the pm_stay_awake and pm_relax with count based and call with paired for fix current concurency issue.
2. RPROC_RECOVERY_FAIL can be another patch for continue try to do recovery work.
3. handle RPROC_DETACHED case.

Quoted 191 lines · Mathieu Poirier
Quoted 127 lines · Arnaud POULIQUEN

Hi,

Quoted 104 lines · Aiqun(Maria) Yu
Quoted 93 lines · Mathieu Poirier

On Wed, 19 Oct 2022 at 23:52, Aiqun(Maria) Yu <[email protected]> wrote:

On 10/14/2022 2:03 AM, Mathieu Poirier wrote:
On Thu, Oct 13, 2022 at 11:34:42AM -0600, Mathieu Poirier wrote:
On Thu, Oct 13, 2022 at 09:40:09AM +0800, Aiqun(Maria) Yu wrote:
Hi Mathieu,

On 10/13/2022 4:43 AM, Mathieu Poirier wrote:
Please add what has changed from one version to another, either in a cover
letter or after the "Signed-off-by". There are many examples on how to
do that
on the mailing list.

Thx for the information, will take a note and benefit for next time.

On Fri, Sep 16, 2022 at 03:12:31PM +0800, Maria Yu wrote:
RPROC_OFFLINE state indicate there is no recovery process
is in progress and no chance to do the pm_relax.
Because when recovering from crash, rproc->lock is held and
state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to
RPROC_RUNNING
when rproc_trigger_recovery() returns. If that is not the case then
something
went wrong.

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before
returning,
so we know the remote processor was stopped. Therefore if rproc->state
is set
to RPROC_OFFLINE something went wrong in either request_firmware() or
rproc_start(). Either way the remote processor is offline and the system
probably
in an unknown/unstable. As such I don't see how calling pm_relax() can help
things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully
finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and
last rproc_trigger_recovery bailed out with only
rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair.
Since the subsystem may still can be recovered with customer's next trigger
of rproc_start, and we can make each error out path clean with pm resources.

I suggest spending time understanding what leads to the failure when
recovering
from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when
rproc_shutdown is triggered at similar time. So not an issue from error out
of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned
from
the beginning. What I see happening is the following:

rproc_report_crash()
pm_stay_awake()
queue_work() // current thread is suspended

rproc_shutdown()
rproc_stop()
rproc->state = RPROC_OFFLINE;

rproc_crash_handler_work()
if (rproc->state == RPROC_OFFLINE)
return // pm_relax() is not called

The right way to fix this is to add a pm_relax() in rproc_shutdown() and
rproc_detach(), along with a very descriptive comment as to why it is needed.

Thinking about this further there are more ramifications to consider. Please
confirm the above scenario is what you are facing. I will advise on how to
move
forward if that is the case.

Not sure if the situation is clear or not. So resend the email again.

The above senario is what customer is facing. crash hanppened while at
the same time shutdown is triggered.

Unfortunately this is not enough details to address a problem as
complex as this one.

And the device cannto goes to suspend state after that.
the subsystem can still be start normally after this.

If the code flow I pasted above reflects the problem at hand, the
current patch will not be sufficient to address the issue. If Arnaud
confirms my suspicions we will have to think about a better solution.

Hi Mathiew,

Could you pls have more details of any side effects other then power issue of the current senario?
Why the current patch is not sufficient pls?

Have the current senario in details with rproc->lock information in details:

Quoted 18 lines
| subsystem crashed interrupt issued      | user trigger shutdown
| rproc_report_crash()                    |
|          pm_stay_awake()                |
|          queue_work()                   |
|                                         |rproc_shutdown
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_stop()
|rproc_crash_handler_work()               |rproc->state = RPROC_OFFLINE;
|                                         |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);                |
|if (rproc->state == RPROC_OFFLINE)       |
|return // pm_relax() is not called       |rproc_boot
|mutex_unlock(&rproc->lock);              |
|                                         |mutex_lock(&rproc->lock);
|                                         |rproc_start()
|                                         |mutex_unlock(&rproc->lock);
Quoted 2 lines

Agree with Mathieu, this is not so simple.

Thanks for looking into this.

Quoted 12 lines

Here is my view hoping I haven't missed a point in your discussion or an other corner cases.

I tried to analyze the issues (in what follows, the term "condition" means the "if" condition in which Aiqun proposes to add the fix) :

I can see 4 use cases with race condition

1) crash report while already one is treated (rproc_boot_recovery called) => not a real use case as if the remote processor is crashed we should not have a second crash report

That part is of great concern to me. *Theoretically* we should not get a new crash report while one has already been dispatched but the current code accounts for this scenario and as such the possibility can't be dismissed. Therefore we need to expect rproc_report_crash() to be called multiple times before a single instance of rproc_boot_recovery() is scheduled.

Quoted 22 lines

2) rproc_stop executed between the queuing of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_OFFLINE
=> we enter in the "condition" and the pm_relax has to be called => This commit fix should solve this use case

3) rproc_detach executed between the queue of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_DETACHED;
=> we don't go in "the condition" and issue because the recovery reattach
to the remote processor
=> but pm_relax is called
=> probably need an extra fix to avoid to re-attach

4) crash report while already one is treated (rproc_attach_recovery called) this one corresponds to an auto reboot of the remote processor, with a new crash
=> rproc->state = RPROC_CRASHED or rproc->state = RPROC_DETACHED; 4)a) rproc->state = RPROC_CRASHED if rproc->recovery_disabled = true
=> should call pm_relax if rproc->recovery_disabled = true => commit does not work for this use case

Quoted 5 lines

4)b) rproc->state = RPROC_DETACHED if recovery fails
=> error case with an unstable state
=> how to differentiate it from the use case 3) ? => introduce a RPROC_RECOVERY_FAIL state?

The case where a recovery fails needs to be considered and is the reason the original patch doesn't work. Right now in rproc_crash_handler_work(), it is not possible to differentiate between a legitimate shutdown request (scenario #2 above) and a recovery that went wrong. I think introducing RPROC_RECOVERY_FAIL would greatly simplify things.

Quoted 9 lines

My initial evaluation had not considered the attach/detach scenarios - thanks for adding that in the mix.

Aiqun, please send a new patchset that adds a new remote processor state, i.e RPROC_RECOVERY_FAIL. There should also be another patch in that set that takes attach/detach scenarios into account. The code between the v6.0 and v6.1 cycle has changed a lot in that area so make sure to properly rebase.

I will try.

Quoted 88 lines
Quoted 88 lines

Then pm_stay_awake is called when the crash work is queued. It seems to me coherent to call the pm_relax in the work handler.

Here is a quick and dirty patch (not tested) that should take into account the main use cases ( except 1) and 4)b) )

(quoted diff) +11 -1
@@ -2009,8 +2009,18 @@ static void rproc_crash_handler_work(struct work_struct *work)
 
             mutex_lock(&rproc->lock);
 
-       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
+       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE ||
+           rproc->state == RPROC_DETACHED) {
                     /* handle only the first crash detected */
+
+               /*
+                * call pm-relax in following use cases:
+                * - the remote processor has been stopped by the user
+                * - the remote processor is detached
+                + - the remote proc has an autonomous reset but recovery_disabled is true.
+                */
+               if(rproc->state != RPROC_CRASHED || rproc->recovery_disabled)
+                       pm_relax(rproc->dev.parent);
                     mutex_unlock(&rproc->lock);
                     return;
             }

Regards,
Arnaud

Quoted 55 lines
Quoted 53 lines

Thanks,
Mathieu

When the state is in RPROC_OFFLINE it means separate request
of rproc_stop was done and no need to hold the wakeup source
in crash handler to recover any more.

Signed-off-by: Maria Yu <[email protected]>
---
drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
1 file changed, 11 insertions(+)

diff --git a/drivers/remoteproc/remoteproc_core.c
b/drivers/remoteproc/remoteproc_core.c
index e5279ed9a8d7..6bc7b8b7d01e 100644
--- a/drivers/remoteproc/remoteproc_core.c
+++ b/drivers/remoteproc/remoteproc_core.c
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct
work_struct *work)
if (rproc->state == RPROC_CRASHED || rproc->state ==
RPROC_OFFLINE) {
/* handle only the first crash detected */
mutex_unlock(&rproc->lock);
+ /*
+ * RPROC_OFFLINE state indicate there is no recovery process
+ * is in progress and no chance to have pm_relax in place.
+ * Because when recovering from crash, rproc->lock is held and
+ * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+ * and then unlock rproc->lock.
+ * RPROC_OFFLINE is only an intermediate state in recovery
+ * process.
+ */
+ if (rproc->state == RPROC_OFFLINE)
+ pm_relax(rproc->dev.parent);
return;
}
--
2.7.4

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Quoted 8 lines
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
#29 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

Quoted 60 lines · Mathieu Poirier
Quoted 51 lines · Aiqun(Maria) Yu

Hi,

Quoted 26 lines · Mathieu Poirier
Quoted 10 lines · Aiqun(Maria) Yu
Quoted 5 lines · Mathieu Poirier

I had a couple of good discussions with our power management expert and even then, the way forward isn't as clear as I would have liked. I am currently travelling and as such don't have the required time to go into greater details, something I will be doing next week.

Thx Mathieu for the info updated.
I'll wait for your update next week then.
let me know any initial questions that you have, perhaps I can also discuss that with our power team at the same time.

The problem is to determine exactly what the WQ_FREEZABLE flag does to the rproc_recovery_wq workueue. The documentation [1] indicate that work items on the WQ are drained before the system is suspended. What I understand from this is that if two work items are queued and one is executing at the time a system suspend is initiated, all three items will be executed before the system is allowed to be suspended. _If_ that is the case, there would not be a need to call pm_stay_awake() and pm_relax() at all.

On the other hand, the PM resource I spoke to thought that in reality things don't happen that way. Taking the same above scenario where 2 work items are queued and one is executing at the time of the suspend, only the work item that is executing will be allowed to execute to completion before the system is suspended. The remaining two items that were queued will not execute.

Quoted 6 lines

If that is the case then we do need to call pm_stay_awake() and pm_relax(), and find another strategy to fix this situation.

Until we have a clear view of how the WQ_FREEZABLE flag works, we won't be able to move forward with this patchset. Unfortunately, I currently do not have the time to look into this.

I had a check on the WQ_FREEZABLE flag, here is my understanding:

when the interrupt happened, it still need pm_stay_awake to make sure queue_work action can active the work instead of susepend the device.

1. If WQ_FREEZABLE, pwq->max_active = 0; // maximum number of in-flight work items is set to 0.

[1].
https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L3748

2. If WQ_FREEZABLE, will only check pwq->nr_active to see if there is still freeze_workqueues_busy.
[2].
https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L5270

Quoted 7 lines

3. When in queue_work, if max_active is 0, when do queue_work it will not actually active the work.
[3].
https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L1418

for the current issue, the work is already complete and forget to set pm_relax in some condition that make the system cannot be suspended.

Quoted 8 lines

What you have above does not describe what happens to work already queue to a WQ_FREEZABLE workqueue when the system gets suspended, and that is the most important thing.

I ended up doing that investigation myself and found the documentation on WQ_FREEZABLE seems to be accurate. I am still discussing this with the Linaro power management team and will get back to you when we reach a conclusion.

Glad that you have it clear.
And wait for your update on this.

here is my understandings maybe you can reference when to discuss with power teams:

1. pm_stay_awake in interrupt handler is needed.
when there is a subsystem crash interrupt, the remote proc driver want to keep the system awake until the crash is handled with subysystem state transfered to shutdown/restarted.

2. as long as the subsystem state is changed to RPROC_CRASHED or RPROC_OFFLINE state, it will be safe to do pm_relax and release the driver's wakelock at that point.

because 2 conditions can be met:
a. inactive work exist:
if there is an inactive work triggered by other crash interrupt, even if it become active again, it will bail out because of the subsystem state and not try to do any actual recovery work. So no wakelock needed to be hold at this condition.

And if the subsystem resume again, and the inactive work become active and do the pm_relax, it will be fine if the device resumed with the state, the extra pm_relax won't make any difference. Or even if there is another crash handler happened at that time, the inactive work will try to do the recovery work directly.

b. active work exist:
WQ_FREEZABLE flag will ensure the active work being finished before going to suspend.

3. pm_relax can be called multi times. even if there is no pm_stay_awake called before this.

Feel free to comment if there is any understanding not correct.

Quoted 508 lines
Quoted 508 lines
Quoted 503 lines

If you want to take on this investigation, keep in mind that any conclusion will need to be backed by a proof. That can be debug messages on a console output or a code reference in the workqueue core.

[1]. https://elixir.bootlin.com/linux/v6.1-rc4/source/Documentation/core-api/workqueue.rst#L184

Quoted 494 lines
Quoted 489 lines · Aiqun(Maria) Yu
Quoted 488 lines

Hi,

Quoted 44 lines · Mathieu Poirier
Quoted 43 lines · Aiqun(Maria) Yu
Quoted 41 lines · Mathieu Poirier
Quoted 6 lines · Aiqun(Maria) Yu

Hi,

Let me think about this carefully.

When in RPROC_RECOVERY_FAIL case we want to re-do the recovery process again or just leave the pm_relax?

Neither.

When a recovery fail we don't want to call pm_relax(). The code in rproc_crash_handler_work() becomes:

       if (rproc->state == RPROC_OFFLINE) {
                     /* We have raced with rproc_shutdown() */
                     pm_relax()
               mutex_unlock(&rproc->lock);
               return;
             }

       if (rproc->state == RPROC_CRASHED ||
                 rproc->state == RPROC_RECOVERY_FAILED) {
               /* handle only the first crash detected */
               mutex_unlock(&rproc->lock);
               return;
       }

RPROC_RECOVERY_FAILED gets set in rproc_boot_recovery() if request_firmware() or rproc_start() fail. Function rproc_trigger_recovery() needs to allow for the recovery the the remote processor is in RPROC_RECOVERY_FAILED state. As such the condition becomes:

             /* State could have changed before we got the mutex */
       if (rproc->state != RPROC_CRASHED &&
                 rproc->state != RPROC_RECOVERY_FAILED)
               goto unlock_mutex;

Start with that and we can look at corner cases (if some exists) with a fresh patchset. Note that I have not addressed the attach/detach() scenario in the above.

Quoted 6 lines
Quoted 2 lines

If we didn't deal with the recovery failed case with correct pm_relax call, it may left the device in a state that cannot enter to suspend state.

That is what I am looking for. We don't want to give the impression that everything is fine by allowing the device to suspend. If the remote processor can't be recovered than it needs to be dealth with.

For the normal recovery failed case, it still need to do pm_relax to not prevent the device goes to suspend. It is what in normal recovery failed case we do in rproc_crash_handler_work as well. rproc_crash_handler_work will not check the result of the rproc_trigger_recovery return value, and will always do pm_relax.

For current conconrency cornor case as well, it is better to consistant with the current design of recovery fail senarios in normal cases.

I personally agree that we shouldn't do nothing when it is a RPROC_RECOVERY_FAILED senario when it is in rproc_crash_handler_work check, because it maybe crash happened when it is trying to do the recovery. So I suggested to do a continue try of trigger recovery again instead of doing nothing and bail out if it is a RPROC_RECOVERY_FAILED state.

Quoted 5 lines
Quoted 4 lines

Because first PROC_RECOVERY_FAIL case cannot ensure it have pm_relax called before the second crash handler call pm_stay_awake or not.

Quoted 7 lines

I've been thinking about that part. I don't think adding a wake_count to control calls to pm_stay_awake()/pm_relax() is the best way to go. There is a similar count happening in the PM runtime subsystem and that is what we should be using. I have asked a power management expert at Linaro for guidance with this matter. I should be able to get back to you with a way forward by the end of next week.

Thx for the specific date provided as well. I will wait until your reply for next patchset then.

Quoted 392 lines
Quoted 392 lines

So, What about the atomic count along with pm_relax and pm_stay_awake ?

struct rproc{
...
atomic_t wake_count;
...
}

rproc_pm_stay_awake()

{
        atomic_inc(&wake_count);
        pm_stay_awake();
}

rproc_pm_relax()

{
        if (atomic_dec_return(&wake_count) == 0)
                pm_stay_awake();
}

can refer code like:

rproc_report_crash()

{
        ...
        rproc_pm_stay_awake();
        queue_work();
        ...
}

rproc_crash_handler_work()

{
        ...
        if (rproc->state == RPROC_OFFLINE || rproc->state == RPROC_CRASHED) {
                      /* We have raced with rproc_shutdown() */
                      rproc_pm_relax();
                mutex_unlock(&rproc->lock);
                return;
              }
        ...
}
Quoted 350 lines

Thanks,
Mathieu

Quoted 346 lines

recovery fail case 1:

|                                      |firstcrash interrupt issued
| second crashed interrupt issued      | rproc_report_crash()
| rproc_report_crash()                 |          pm_stay_awake()
|          pm_stay_awake()             |          queue_work()
|          queue_work()                |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|rproc_crash_handler_work()            |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);             |pm_relax()
|if (rproc->state == RPROC_OFFLINE)    |
|return // shouldn't do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

recovery fail case 2:

|                                      |firstcrash interrupt issued
|                                      | rproc_report_crash()
|                                      |          pm_stay_awake()
|                                      |          queue_work()
|                                      |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|                                      |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|                                      |pm_relax()
|
| second crashed interrupt issued      |
| rproc_report_crash()                 |
|          pm_stay_awake()             |
|          queue_work()                |
|pm_stay_awake()
|mutex_lock(&rproc->lock);
|if (rproc->state == RPROC_OFFLINE)    |
|return // still need do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

Maybe I can have:
1. the pm_stay_awake and pm_relax with count based and call with paired for fix current concurency issue.
2. RPROC_RECOVERY_FAIL can be another patch for continue try to do recovery work.
3. handle RPROC_DETACHED case.

Quoted 191 lines · Mathieu Poirier
Quoted 127 lines · Arnaud POULIQUEN

Hi,

Quoted 122 lines · Aiqun(Maria) Yu

On 10/22/2022 3:34 AM, Mathieu Poirier wrote:
On Wed, 19 Oct 2022 at 23:52, Aiqun(Maria) Yu <[email protected]> wrote:

On 10/14/2022 2:03 AM, Mathieu Poirier wrote:
On Thu, Oct 13, 2022 at 11:34:42AM -0600, Mathieu Poirier wrote:
On Thu, Oct 13, 2022 at 09:40:09AM +0800, Aiqun(Maria) Yu wrote:
Hi Mathieu,

On 10/13/2022 4:43 AM, Mathieu Poirier wrote:
Please add what has changed from one version to another, either in a cover
letter or after the "Signed-off-by". There are many examples on how to
do that
on the mailing list.

Thx for the information, will take a note and benefit for next time.

On Fri, Sep 16, 2022 at 03:12:31PM +0800, Maria Yu wrote:
RPROC_OFFLINE state indicate there is no recovery process
is in progress and no chance to do the pm_relax.
Because when recovering from crash, rproc->lock is held and
state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to
RPROC_RUNNING
when rproc_trigger_recovery() returns. If that is not the case then
something
went wrong.

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before
returning,
so we know the remote processor was stopped. Therefore if rproc->state
is set
to RPROC_OFFLINE something went wrong in either request_firmware() or
rproc_start(). Either way the remote processor is offline and the system
probably
in an unknown/unstable. As such I don't see how calling pm_relax() can help
things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully
finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and
last rproc_trigger_recovery bailed out with only
rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair.
Since the subsystem may still can be recovered with customer's next trigger
of rproc_start, and we can make each error out path clean with pm resources.

I suggest spending time understanding what leads to the failure when
recovering
from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when
rproc_shutdown is triggered at similar time. So not an issue from error out
of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned
from
the beginning. What I see happening is the following:

rproc_report_crash()
pm_stay_awake()
queue_work() // current thread is suspended

rproc_shutdown()
rproc_stop()
rproc->state = RPROC_OFFLINE;

rproc_crash_handler_work()
if (rproc->state == RPROC_OFFLINE)
return // pm_relax() is not called

The right way to fix this is to add a pm_relax() in rproc_shutdown() and
rproc_detach(), along with a very descriptive comment as to why it is needed.

Thinking about this further there are more ramifications to consider. Please
confirm the above scenario is what you are facing. I will advise on how to
move
forward if that is the case.

Not sure if the situation is clear or not. So resend the email again.

The above senario is what customer is facing. crash hanppened while at
the same time shutdown is triggered.

Unfortunately this is not enough details to address a problem as
complex as this one.

And the device cannto goes to suspend state after that.
the subsystem can still be start normally after this.

If the code flow I pasted above reflects the problem at hand, the
current patch will not be sufficient to address the issue. If Arnaud
confirms my suspicions we will have to think about a better solution.

Hi Mathiew,

Could you pls have more details of any side effects other then power issue of
the current senario?
Why the current patch is not sufficient pls?

Have the current senario in details with rproc->lock information in details:

| subsystem crashed interrupt issued | user trigger shutdown
| rproc_report_crash() |
| pm_stay_awake() |
| queue_work() |
| |rproc_shutdown
| |mutex_lock(&rproc->lock);
| |rproc_stop()
|rproc_crash_handler_work() |rproc->state = RPROC_OFFLINE;
| |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock); |
|if (rproc->state == RPROC_OFFLINE) |
|return // pm_relax() is not called |rproc_boot
|mutex_unlock(&rproc->lock); |
| |mutex_lock(&rproc->lock);
| |rproc_start()
| |mutex_unlock(&rproc->lock);

Quoted 2 lines

Agree with Mathieu, this is not so simple.

Thanks for looking into this.

Quoted 12 lines

Here is my view hoping I haven't missed a point in your discussion or an other corner cases.

I tried to analyze the issues (in what follows, the term "condition" means the "if" condition in which Aiqun proposes to add the fix) :

I can see 4 use cases with race condition

1) crash report while already one is treated (rproc_boot_recovery called) => not a real use case as if the remote processor is crashed we should not have a second crash report

That part is of great concern to me. *Theoretically* we should not get a new crash report while one has already been dispatched but the current code accounts for this scenario and as such the possibility can't be dismissed. Therefore we need to expect rproc_report_crash() to be called multiple times before a single instance of rproc_boot_recovery() is scheduled.

Quoted 22 lines

2) rproc_stop executed between the queuing of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_OFFLINE
=> we enter in the "condition" and the pm_relax has to be called => This commit fix should solve this use case

3) rproc_detach executed between the queue of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_DETACHED;
=> we don't go in "the condition" and issue because the recovery reattach
to the remote processor
=> but pm_relax is called
=> probably need an extra fix to avoid to re-attach

4) crash report while already one is treated (rproc_attach_recovery called) this one corresponds to an auto reboot of the remote processor, with a new crash
=> rproc->state = RPROC_CRASHED or rproc->state = RPROC_DETACHED; 4)a) rproc->state = RPROC_CRASHED if rproc->recovery_disabled = true
=> should call pm_relax if rproc->recovery_disabled = true => commit does not work for this use case

Quoted 5 lines

4)b) rproc->state = RPROC_DETACHED if recovery fails
=> error case with an unstable state
=> how to differentiate it from the use case 3) ? => introduce a RPROC_RECOVERY_FAIL state?

The case where a recovery fails needs to be considered and is the reason the original patch doesn't work. Right now in rproc_crash_handler_work(), it is not possible to differentiate between a legitimate shutdown request (scenario #2 above) and a recovery that went wrong. I think introducing RPROC_RECOVERY_FAIL would greatly simplify things.

Quoted 9 lines

My initial evaluation had not considered the attach/detach scenarios - thanks for adding that in the mix.

Aiqun, please send a new patchset that adds a new remote processor state, i.e RPROC_RECOVERY_FAIL. There should also be another patch in that set that takes attach/detach scenarios into account. The code between the v6.0 and v6.1 cycle has changed a lot in that area so make sure to properly rebase.

I will try.

Quoted 88 lines
Quoted 88 lines

Then pm_stay_awake is called when the crash work is queued. It seems to me coherent to call the pm_relax in the work handler.

Here is a quick and dirty patch (not tested) that should take into account the main use cases ( except 1) and 4)b) )

(quoted diff) +11 -1
@@ -2009,8 +2009,18 @@ static void rproc_crash_handler_work(struct work_struct *work)
 
              mutex_lock(&rproc->lock);
 
-       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
+       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE ||
+           rproc->state == RPROC_DETACHED) {
                      /* handle only the first crash detected */
+
+               /*
+                * call pm-relax in following use cases:
+                * - the remote processor has been stopped by the user
+                * - the remote processor is detached
+                + - the remote proc has an autonomous reset but recovery_disabled is true.
+                */
+               if(rproc->state != RPROC_CRASHED || rproc->recovery_disabled)
+                       pm_relax(rproc->dev.parent);
                      mutex_unlock(&rproc->lock);
                      return;
              }

Regards,
Arnaud

Quoted 55 lines

Thanks,
Mathieu

When the state is in RPROC_OFFLINE it means separate request
of rproc_stop was done and no need to hold the wakeup source
in crash handler to recover any more.

Signed-off-by: Maria Yu <[email protected]>
---
drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
1 file changed, 11 insertions(+)

diff --git a/drivers/remoteproc/remoteproc_core.c
b/drivers/remoteproc/remoteproc_core.c
index e5279ed9a8d7..6bc7b8b7d01e 100644
--- a/drivers/remoteproc/remoteproc_core.c
+++ b/drivers/remoteproc/remoteproc_core.c
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct
work_struct *work)
if (rproc->state == RPROC_CRASHED || rproc->state ==
RPROC_OFFLINE) {
/* handle only the first crash detected */
mutex_unlock(&rproc->lock);
+ /*
+ * RPROC_OFFLINE state indicate there is no recovery process
+ * is in progress and no chance to have pm_relax in place.
+ * Because when recovering from crash, rproc->lock is held and
+ * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+ * and then unlock rproc->lock.
+ * RPROC_OFFLINE is only an intermediate state in recovery
+ * process.
+ */
+ if (rproc->state == RPROC_OFFLINE)
+ pm_relax(rproc->dev.parent);
return;
}
--
2.7.4

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Quoted 8 lines
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
maintainer
#30 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

In reply to Mathieu Poirier
Quoted 60 lines · Mathieu Poirier
Quoted 51 lines · Aiqun(Maria) Yu

Hi,

Quoted 26 lines · Mathieu Poirier
Quoted 10 lines · Aiqun(Maria) Yu
Quoted 5 lines · Mathieu Poirier

I had a couple of good discussions with our power management expert and even then, the way forward isn't as clear as I would have liked. I am currently travelling and as such don't have the required time to go into greater details, something I will be doing next week.

Thx Mathieu for the info updated.
I'll wait for your update next week then.
let me know any initial questions that you have, perhaps I can also discuss that with our power team at the same time.

The problem is to determine exactly what the WQ_FREEZABLE flag does to the rproc_recovery_wq workueue. The documentation [1] indicate that work items on the WQ are drained before the system is suspended. What I understand from this is that if two work items are queued and one is executing at the time a system suspend is initiated, all three items will be executed before the system is allowed to be suspended. _If_ that is the case, there would not be a need to call pm_stay_awake() and pm_relax() at all.

On the other hand, the PM resource I spoke to thought that in reality things don't happen that way. Taking the same above scenario where 2 work items are queued and one is executing at the time of the suspend, only the work item that is executing will be allowed to execute to completion before the system is suspended. The remaining two items that were queued will not execute.

Quoted 6 lines

If that is the case then we do need to call pm_stay_awake() and pm_relax(), and find another strategy to fix this situation.

Until we have a clear view of how the WQ_FREEZABLE flag works, we won't be able to move forward with this patchset. Unfortunately, I currently do not have the time to look into this.

I had a check on the WQ_FREEZABLE flag, here is my understanding:

when the interrupt happened, it still need pm_stay_awake to make sure queue_work action can active the work instead of susepend the device.

1. If WQ_FREEZABLE, pwq->max_active = 0; // maximum number of in-flight work items is set to 0.

[1].
https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L3748

2. If WQ_FREEZABLE, will only check pwq->nr_active to see if there is still freeze_workqueues_busy.
[2].
https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L5270

Quoted 7 lines

3. When in queue_work, if max_active is 0, when do queue_work it will not actually active the work.
[3].
https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L1418

for the current issue, the work is already complete and forget to set pm_relax in some condition that make the system cannot be suspended.

Quoted 8 lines

What you have above does not describe what happens to work already queue to a WQ_FREEZABLE workqueue when the system gets suspended, and that is the most important thing.

I ended up doing that investigation myself and found the documentation on WQ_FREEZABLE seems to be accurate. I am still discussing this with the Linaro power management team and will get back to you when we reach a conclusion.

I finally have time to get back to this problem...

As I reported above I went down to the bottom of how WQ_FREEZABLE workqueues work. The conclusion of that research is that _any_ work, queued or currently executing, will finish before the system is suspended. As such, as soon as queue_work() returns in rproc_report_crash(), you can be guaranteed what was just queued will be executed. Therefore, calling pm_stay_awake() in rproc_report_crash() and pm_relax() in rproc_crash_handler_work() should not be needed because the system will not suspend for as long as there is work to be done by the freezable worqueues.

I don't have details on what Rishabh - the author of the patch that added the calls - was facing but based on the my understanding of freezable workqueues, it seem to have been the wrong solution.

Until we have a clear vision on what Rishabh was trying to fix, nothing in that area will be changing. I suggest you touch base with him and sort it out. If that is not possible then calls to pm_stay_awake() and pm_relax() should be removed.

Quoted 508 lines
Quoted 508 lines
Quoted 503 lines

If you want to take on this investigation, keep in mind that any conclusion will need to be backed by a proof. That can be debug messages on a console output or a code reference in the workqueue core.

[1]. https://elixir.bootlin.com/linux/v6.1-rc4/source/Documentation/core-api/workqueue.rst#L184

Quoted 494 lines
Quoted 489 lines · Aiqun(Maria) Yu
Quoted 488 lines

Hi,

Quoted 44 lines · Mathieu Poirier
Quoted 43 lines · Aiqun(Maria) Yu
Quoted 41 lines · Mathieu Poirier
Quoted 6 lines · Aiqun(Maria) Yu

Hi,

Let me think about this carefully.

When in RPROC_RECOVERY_FAIL case we want to re-do the recovery process again or just leave the pm_relax?

Neither.

When a recovery fail we don't want to call pm_relax(). The code in rproc_crash_handler_work() becomes:

      if (rproc->state == RPROC_OFFLINE) {
                    /* We have raced with rproc_shutdown() */
                    pm_relax()
              mutex_unlock(&rproc->lock);
              return;
            }

      if (rproc->state == RPROC_CRASHED ||
                rproc->state == RPROC_RECOVERY_FAILED) {
              /* handle only the first crash detected */
              mutex_unlock(&rproc->lock);
              return;
      }

RPROC_RECOVERY_FAILED gets set in rproc_boot_recovery() if request_firmware() or rproc_start() fail. Function rproc_trigger_recovery() needs to allow for the recovery the the remote processor is in RPROC_RECOVERY_FAILED state. As such the condition becomes:

            /* State could have changed before we got the mutex */
      if (rproc->state != RPROC_CRASHED &&
                rproc->state != RPROC_RECOVERY_FAILED)
              goto unlock_mutex;

Start with that and we can look at corner cases (if some exists) with a fresh patchset. Note that I have not addressed the attach/detach() scenario in the above.

Quoted 6 lines
Quoted 2 lines

If we didn't deal with the recovery failed case with correct pm_relax call, it may left the device in a state that cannot enter to suspend state.

That is what I am looking for. We don't want to give the impression that everything is fine by allowing the device to suspend. If the remote processor can't be recovered than it needs to be dealth with.

For the normal recovery failed case, it still need to do pm_relax to not prevent the device goes to suspend. It is what in normal recovery failed case we do in rproc_crash_handler_work as well. rproc_crash_handler_work will not check the result of the rproc_trigger_recovery return value, and will always do pm_relax.

For current conconrency cornor case as well, it is better to consistant with the current design of recovery fail senarios in normal cases.

I personally agree that we shouldn't do nothing when it is a RPROC_RECOVERY_FAILED senario when it is in rproc_crash_handler_work check, because it maybe crash happened when it is trying to do the recovery. So I suggested to do a continue try of trigger recovery again instead of doing nothing and bail out if it is a RPROC_RECOVERY_FAILED state.

Quoted 5 lines
Quoted 4 lines

Because first PROC_RECOVERY_FAIL case cannot ensure it have pm_relax called before the second crash handler call pm_stay_awake or not.

Quoted 7 lines

I've been thinking about that part. I don't think adding a wake_count to control calls to pm_stay_awake()/pm_relax() is the best way to go. There is a similar count happening in the PM runtime subsystem and that is what we should be using. I have asked a power management expert at Linaro for guidance with this matter. I should be able to get back to you with a way forward by the end of next week.

Thx for the specific date provided as well. I will wait until your reply for next patchset then.

Quoted 392 lines
Quoted 392 lines

So, What about the atomic count along with pm_relax and pm_stay_awake ?

struct rproc{
...
atomic_t wake_count;
...
}

rproc_pm_stay_awake()

{
       atomic_inc(&wake_count);
       pm_stay_awake();
}

rproc_pm_relax()

{
       if (atomic_dec_return(&wake_count) == 0)
               pm_stay_awake();
}

can refer code like:

rproc_report_crash()

{
       ...
       rproc_pm_stay_awake();
       queue_work();
       ...
}

rproc_crash_handler_work()

{
       ...
       if (rproc->state == RPROC_OFFLINE || rproc->state == RPROC_CRASHED) {
                     /* We have raced with rproc_shutdown() */
                     rproc_pm_relax();
               mutex_unlock(&rproc->lock);
               return;
             }
       ...
}
Quoted 350 lines

Thanks,
Mathieu

Quoted 346 lines

recovery fail case 1:

|                                      |firstcrash interrupt issued
| second crashed interrupt issued      | rproc_report_crash()
| rproc_report_crash()                 |          pm_stay_awake()
|          pm_stay_awake()             |          queue_work()
|          queue_work()                |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|rproc_crash_handler_work()            |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);             |pm_relax()
|if (rproc->state == RPROC_OFFLINE)    |
|return // shouldn't do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

recovery fail case 2:

|                                      |firstcrash interrupt issued
|                                      | rproc_report_crash()
|                                      |          pm_stay_awake()
|                                      |          queue_work()
|                                      |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|                                      |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|                                      |pm_relax()
|
| second crashed interrupt issued      |
| rproc_report_crash()                 |
|          pm_stay_awake()             |
|          queue_work()                |
|pm_stay_awake()
|mutex_lock(&rproc->lock);
|if (rproc->state == RPROC_OFFLINE)    |
|return // still need do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

Maybe I can have:
1. the pm_stay_awake and pm_relax with count based and call with paired for fix current concurency issue.
2. RPROC_RECOVERY_FAIL can be another patch for continue try to do recovery work.
3. handle RPROC_DETACHED case.

Quoted 191 lines · Mathieu Poirier
Quoted 127 lines · Arnaud POULIQUEN

Hi,

Quoted 122 lines · Aiqun(Maria) Yu

On 10/22/2022 3:34 AM, Mathieu Poirier wrote:
On Wed, 19 Oct 2022 at 23:52, Aiqun(Maria) Yu <[email protected]> wrote:

On 10/14/2022 2:03 AM, Mathieu Poirier wrote:
On Thu, Oct 13, 2022 at 11:34:42AM -0600, Mathieu Poirier wrote:
On Thu, Oct 13, 2022 at 09:40:09AM +0800, Aiqun(Maria) Yu wrote:
Hi Mathieu,

On 10/13/2022 4:43 AM, Mathieu Poirier wrote:
Please add what has changed from one version to another, either in a cover
letter or after the "Signed-off-by". There are many examples on how to
do that
on the mailing list.

Thx for the information, will take a note and benefit for next time.

On Fri, Sep 16, 2022 at 03:12:31PM +0800, Maria Yu wrote:
RPROC_OFFLINE state indicate there is no recovery process
is in progress and no chance to do the pm_relax.
Because when recovering from crash, rproc->lock is held and
state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to
RPROC_RUNNING
when rproc_trigger_recovery() returns. If that is not the case then
something
went wrong.

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before
returning,
so we know the remote processor was stopped. Therefore if rproc->state
is set
to RPROC_OFFLINE something went wrong in either request_firmware() or
rproc_start(). Either way the remote processor is offline and the system
probably
in an unknown/unstable. As such I don't see how calling pm_relax() can help
things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully
finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and
last rproc_trigger_recovery bailed out with only
rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair.
Since the subsystem may still can be recovered with customer's next trigger
of rproc_start, and we can make each error out path clean with pm resources.

I suggest spending time understanding what leads to the failure when
recovering
from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when
rproc_shutdown is triggered at similar time. So not an issue from error out
of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned
from
the beginning. What I see happening is the following:

rproc_report_crash()
pm_stay_awake()
queue_work() // current thread is suspended

rproc_shutdown()
rproc_stop()
rproc->state = RPROC_OFFLINE;

rproc_crash_handler_work()
if (rproc->state == RPROC_OFFLINE)
return // pm_relax() is not called

The right way to fix this is to add a pm_relax() in rproc_shutdown() and
rproc_detach(), along with a very descriptive comment as to why it is needed.

Thinking about this further there are more ramifications to consider. Please
confirm the above scenario is what you are facing. I will advise on how to
move
forward if that is the case.

Not sure if the situation is clear or not. So resend the email again.

The above senario is what customer is facing. crash hanppened while at
the same time shutdown is triggered.

Unfortunately this is not enough details to address a problem as
complex as this one.

And the device cannto goes to suspend state after that.
the subsystem can still be start normally after this.

If the code flow I pasted above reflects the problem at hand, the
current patch will not be sufficient to address the issue. If Arnaud
confirms my suspicions we will have to think about a better solution.

Hi Mathiew,

Could you pls have more details of any side effects other then power issue of
the current senario?
Why the current patch is not sufficient pls?

Have the current senario in details with rproc->lock information in details:

| subsystem crashed interrupt issued | user trigger shutdown
| rproc_report_crash() |
| pm_stay_awake() |
| queue_work() |
| |rproc_shutdown
| |mutex_lock(&rproc->lock);
| |rproc_stop()
|rproc_crash_handler_work() |rproc->state = RPROC_OFFLINE;
| |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock); |
|if (rproc->state == RPROC_OFFLINE) |
|return // pm_relax() is not called |rproc_boot
|mutex_unlock(&rproc->lock); |
| |mutex_lock(&rproc->lock);
| |rproc_start()
| |mutex_unlock(&rproc->lock);

Quoted 2 lines

Agree with Mathieu, this is not so simple.

Thanks for looking into this.

Quoted 12 lines

Here is my view hoping I haven't missed a point in your discussion or an other corner cases.

I tried to analyze the issues (in what follows, the term "condition" means the "if" condition in which Aiqun proposes to add the fix) :

I can see 4 use cases with race condition

1) crash report while already one is treated (rproc_boot_recovery called) => not a real use case as if the remote processor is crashed we should not have a second crash report

That part is of great concern to me. *Theoretically* we should not get a new crash report while one has already been dispatched but the current code accounts for this scenario and as such the possibility can't be dismissed. Therefore we need to expect rproc_report_crash() to be called multiple times before a single instance of rproc_boot_recovery() is scheduled.

Quoted 22 lines

2) rproc_stop executed between the queuing of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_OFFLINE
=> we enter in the "condition" and the pm_relax has to be called => This commit fix should solve this use case

3) rproc_detach executed between the queue of the crash work and the call of rproc_crash_handler_work
=> rproc->state = RPROC_DETACHED;
=> we don't go in "the condition" and issue because the recovery reattach
to the remote processor
=> but pm_relax is called
=> probably need an extra fix to avoid to re-attach

4) crash report while already one is treated (rproc_attach_recovery called) this one corresponds to an auto reboot of the remote processor, with a new crash
=> rproc->state = RPROC_CRASHED or rproc->state = RPROC_DETACHED; 4)a) rproc->state = RPROC_CRASHED if rproc->recovery_disabled = true
=> should call pm_relax if rproc->recovery_disabled = true => commit does not work for this use case

Quoted 5 lines

4)b) rproc->state = RPROC_DETACHED if recovery fails
=> error case with an unstable state
=> how to differentiate it from the use case 3) ? => introduce a RPROC_RECOVERY_FAIL state?

The case where a recovery fails needs to be considered and is the reason the original patch doesn't work. Right now in rproc_crash_handler_work(), it is not possible to differentiate between a legitimate shutdown request (scenario #2 above) and a recovery that went wrong. I think introducing RPROC_RECOVERY_FAIL would greatly simplify things.

Quoted 9 lines

My initial evaluation had not considered the attach/detach scenarios - thanks for adding that in the mix.

Aiqun, please send a new patchset that adds a new remote processor state, i.e RPROC_RECOVERY_FAIL. There should also be another patch in that set that takes attach/detach scenarios into account. The code between the v6.0 and v6.1 cycle has changed a lot in that area so make sure to properly rebase.

I will try.

Quoted 88 lines
Quoted 88 lines

Then pm_stay_awake is called when the crash work is queued. It seems to me coherent to call the pm_relax in the work handler.

Here is a quick and dirty patch (not tested) that should take into account the main use cases ( except 1) and 4)b) )

(quoted diff) +11 -1
@@ -2009,8 +2009,18 @@ static void rproc_crash_handler_work(struct work_struct *work)
 
             mutex_lock(&rproc->lock);
 
-       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
+       if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE ||
+           rproc->state == RPROC_DETACHED) {
                     /* handle only the first crash detected */
+
+               /*
+                * call pm-relax in following use cases:
+                * - the remote processor has been stopped by the user
+                * - the remote processor is detached
+                + - the remote proc has an autonomous reset but recovery_disabled is true.
+                */
+               if(rproc->state != RPROC_CRASHED || rproc->recovery_disabled)
+                       pm_relax(rproc->dev.parent);
                     mutex_unlock(&rproc->lock);
                     return;
             }

Regards,
Arnaud

Quoted 55 lines

Thanks,
Mathieu

When the state is in RPROC_OFFLINE it means separate request
of rproc_stop was done and no need to hold the wakeup source
in crash handler to recover any more.

Signed-off-by: Maria Yu <[email protected]>
---
drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
1 file changed, 11 insertions(+)

diff --git a/drivers/remoteproc/remoteproc_core.c
b/drivers/remoteproc/remoteproc_core.c
index e5279ed9a8d7..6bc7b8b7d01e 100644
--- a/drivers/remoteproc/remoteproc_core.c
+++ b/drivers/remoteproc/remoteproc_core.c
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct
work_struct *work)
if (rproc->state == RPROC_CRASHED || rproc->state ==
RPROC_OFFLINE) {
/* handle only the first crash detected */
mutex_unlock(&rproc->lock);
+ /*
+ * RPROC_OFFLINE state indicate there is no recovery process
+ * is in progress and no chance to have pm_relax in place.
+ * Because when recovering from crash, rproc->lock is held and
+ * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+ * and then unlock rproc->lock.
+ * RPROC_OFFLINE is only an intermediate state in recovery
+ * process.
+ */
+ if (rproc->state == RPROC_OFFLINE)
+ pm_relax(rproc->dev.parent);
return;
}
--
2.7.4

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Quoted 8 lines
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
#31 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

Quoted 72 lines · Mathieu Poirier
Quoted 60 lines · Mathieu Poirier
Quoted 51 lines · Aiqun(Maria) Yu

Hi,

Quoted 26 lines · Mathieu Poirier
Quoted 10 lines · Aiqun(Maria) Yu
Quoted 5 lines · Mathieu Poirier

I had a couple of good discussions with our power management expert and even then, the way forward isn't as clear as I would have liked. I am currently travelling and as such don't have the required time to go into greater details, something I will be doing next week.

Thx Mathieu for the info updated.
I'll wait for your update next week then.
let me know any initial questions that you have, perhaps I can also discuss that with our power team at the same time.

The problem is to determine exactly what the WQ_FREEZABLE flag does to the rproc_recovery_wq workueue. The documentation [1] indicate that work items on the WQ are drained before the system is suspended. What I understand from this is that if two work items are queued and one is executing at the time a system suspend is initiated, all three items will be executed before the system is allowed to be suspended. _If_ that is the case, there would not be a need to call pm_stay_awake() and pm_relax() at all.

On the other hand, the PM resource I spoke to thought that in reality things don't happen that way. Taking the same above scenario where 2 work items are queued and one is executing at the time of the suspend, only the work item that is executing will be allowed to execute to completion before the system is suspended. The remaining two items that were queued will not execute.

Quoted 6 lines

If that is the case then we do need to call pm_stay_awake() and pm_relax(), and find another strategy to fix this situation.

Until we have a clear view of how the WQ_FREEZABLE flag works, we won't be able to move forward with this patchset. Unfortunately, I currently do not have the time to look into this.

I had a check on the WQ_FREEZABLE flag, here is my understanding:

when the interrupt happened, it still need pm_stay_awake to make sure queue_work action can active the work instead of susepend the device.

1. If WQ_FREEZABLE, pwq->max_active = 0; // maximum number of in-flight work items is set to 0.

[1].
https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L3748

2. If WQ_FREEZABLE, will only check pwq->nr_active to see if there is still freeze_workqueues_busy.
[2].
https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L5270

Quoted 7 lines

3. When in queue_work, if max_active is 0, when do queue_work it will not actually active the work.
[3].
https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L1418

for the current issue, the work is already complete and forget to set pm_relax in some condition that make the system cannot be suspended.

Quoted 8 lines

What you have above does not describe what happens to work already queue to a WQ_FREEZABLE workqueue when the system gets suspended, and that is the most important thing.

I ended up doing that investigation myself and found the documentation on WQ_FREEZABLE seems to be accurate. I am still discussing this with the Linaro power management team and will get back to you when we reach a conclusion.

I finally have time to get back to this problem...

Quoted 9 lines

As I reported above I went down to the bottom of how WQ_FREEZABLE workqueues work. The conclusion of that research is that _any_ work, queued or currently executing, will finish before the system is suspended. As such, as soon as queue_work() returns in rproc_report_crash(), you can be guaranteed what was just queued will be executed. Therefore, calling pm_stay_awake() in rproc_report_crash() and pm_relax() in rproc_crash_handler_work() should not be needed because the system will not suspend for as long as there is work to be done by the freezable worqueues.

The problem is when crash interrupt happened in a later point(after freeze_kernel_thread), the queue_work will result with the state of WORK_STRUCT_INACTIVE and insert to inactive_works. And this is the senario that pm_stay_awake wanted to address inside the interrupt handler call.

refer to:
[1] https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L1512

the current suspend simple flow can be:

        +-+-------------------+
        |   freeze_processes  |
        +----------+----------+
                   |
                   |
       +-----------|-----------+
       | freeze_kernel_thread  |
       +-----------+-----------+
                   |
        +-+--------|----------+
        |   device_suspend    |
        +----------|----------+
                   |
                   |
       +-----------+-----------+
       | pm_wakeup_pending     |
       +-----------|-----------+
                   |
                   |
 +-----------------+-------------------+
 | suspend_ops->enter (machine suspend)|
 +-------------------------------------+

refer to:
suspend_prepare -> suspend_freeze_processes -->freeze_processes and freeze_kernel_threads

refer to:

suspend_devices_and_enter -->
dpm_suspend_start -> dpm_suspend --> device_suspend --> device suspend callbacks suspend_enter -->dpm_suspend_late-->device_suspend_late

[2] https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/power/suspend.c#L582

Quoted 3 lines

I don't have details on what Rishabh - the author of the patch that added the calls - was facing but based on the my understanding of freezable workqueues, it seem to have been the wrong solution.

In my opinion, the freezable workqueues are in the right solution. It only protect for the system freeze state.
While if we want the crash interrupt to be handled to prevent current suspend process, it still need to deal with other suspend state to machin suspend senarios. Typical example can be before disable the device interrupt in device suspend callbacks, it will flush the necessary workqueues.

while in our case, since the crash handler will invoke a subsystem restart process which reply on user thread uevent firmware loaders, so it need to abort the suspend process with pm_stay_awake called.

So in a word, pm_stay_awake is necessary per my understanding.

Quoted 514 lines

Until we have a clear vision on what Rishabh was trying to fix, nothing in that area will be changing. I suggest you touch base with him and sort it out. If that is not possible then calls to pm_stay_awake() and pm_relax() should be removed.

Quoted 508 lines
Quoted 508 lines
Quoted 503 lines

If you want to take on this investigation, keep in mind that any conclusion will need to be backed by a proof. That can be debug messages on a console output or a code reference in the workqueue core.

[1]. https://elixir.bootlin.com/linux/v6.1-rc4/source/Documentation/core-api/workqueue.rst#L184

Quoted 494 lines
Quoted 489 lines · Aiqun(Maria) Yu
Quoted 488 lines

Hi,

Quoted 44 lines · Mathieu Poirier
Quoted 43 lines · Aiqun(Maria) Yu
Quoted 41 lines · Mathieu Poirier
Quoted 6 lines · Aiqun(Maria) Yu

Hi,

Let me think about this carefully.

When in RPROC_RECOVERY_FAIL case we want to re-do the recovery process again or just leave the pm_relax?

Neither.

When a recovery fail we don't want to call pm_relax(). The code in rproc_crash_handler_work() becomes:

       if (rproc->state == RPROC_OFFLINE) {
                     /* We have raced with rproc_shutdown() */
                     pm_relax()
               mutex_unlock(&rproc->lock);
               return;
             }

       if (rproc->state == RPROC_CRASHED ||
                 rproc->state == RPROC_RECOVERY_FAILED) {
               /* handle only the first crash detected */
               mutex_unlock(&rproc->lock);
               return;
       }

RPROC_RECOVERY_FAILED gets set in rproc_boot_recovery() if request_firmware() or rproc_start() fail. Function rproc_trigger_recovery() needs to allow for the recovery the the remote processor is in RPROC_RECOVERY_FAILED state. As such the condition becomes:

             /* State could have changed before we got the mutex */
       if (rproc->state != RPROC_CRASHED &&
                 rproc->state != RPROC_RECOVERY_FAILED)
               goto unlock_mutex;

Start with that and we can look at corner cases (if some exists) with a fresh patchset. Note that I have not addressed the attach/detach() scenario in the above.

Quoted 6 lines
Quoted 2 lines

If we didn't deal with the recovery failed case with correct pm_relax call, it may left the device in a state that cannot enter to suspend state.

That is what I am looking for. We don't want to give the impression that everything is fine by allowing the device to suspend. If the remote processor can't be recovered than it needs to be dealth with.

For the normal recovery failed case, it still need to do pm_relax to not prevent the device goes to suspend. It is what in normal recovery failed case we do in rproc_crash_handler_work as well. rproc_crash_handler_work will not check the result of the rproc_trigger_recovery return value, and will always do pm_relax.

For current conconrency cornor case as well, it is better to consistant with the current design of recovery fail senarios in normal cases.

I personally agree that we shouldn't do nothing when it is a RPROC_RECOVERY_FAILED senario when it is in rproc_crash_handler_work check, because it maybe crash happened when it is trying to do the recovery. So I suggested to do a continue try of trigger recovery again instead of doing nothing and bail out if it is a RPROC_RECOVERY_FAILED state.

Quoted 5 lines
Quoted 4 lines

Because first PROC_RECOVERY_FAIL case cannot ensure it have pm_relax called before the second crash handler call pm_stay_awake or not.

Quoted 7 lines

I've been thinking about that part. I don't think adding a wake_count to control calls to pm_stay_awake()/pm_relax() is the best way to go. There is a similar count happening in the PM runtime subsystem and that is what we should be using. I have asked a power management expert at Linaro for guidance with this matter. I should be able to get back to you with a way forward by the end of next week.

Thx for the specific date provided as well. I will wait until your reply for next patchset then.

Quoted 392 lines
Quoted 392 lines

So, What about the atomic count along with pm_relax and pm_stay_awake ?

struct rproc{
...
atomic_t wake_count;
...
}

rproc_pm_stay_awake()

{
        atomic_inc(&wake_count);
        pm_stay_awake();
}

rproc_pm_relax()

{
        if (atomic_dec_return(&wake_count) == 0)
                pm_stay_awake();
}

can refer code like:

rproc_report_crash()

{
        ...
        rproc_pm_stay_awake();
        queue_work();
        ...
}

rproc_crash_handler_work()

{
        ...
        if (rproc->state == RPROC_OFFLINE || rproc->state == RPROC_CRASHED) {
                      /* We have raced with rproc_shutdown() */
                      rproc_pm_relax();
                mutex_unlock(&rproc->lock);
                return;
              }
        ...
}
Quoted 350 lines

Thanks,
Mathieu

Quoted 346 lines

recovery fail case 1:

|                                      |firstcrash interrupt issued
| second crashed interrupt issued      | rproc_report_crash()
| rproc_report_crash()                 |          pm_stay_awake()
|          pm_stay_awake()             |          queue_work()
|          queue_work()                |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|rproc_crash_handler_work()            |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);             |pm_relax()
|if (rproc->state == RPROC_OFFLINE)    |
|return // shouldn't do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

recovery fail case 2:

|                                      |firstcrash interrupt issued
|                                      | rproc_report_crash()
|                                      |          pm_stay_awake()
|                                      |          queue_work()
|                                      |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|                                      |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|                                      |pm_relax()
|
| second crashed interrupt issued      |
| rproc_report_crash()                 |
|          pm_stay_awake()             |
|          queue_work()                |
|pm_stay_awake()
|mutex_lock(&rproc->lock);
|if (rproc->state == RPROC_OFFLINE)    |
|return // still need do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

Maybe I can have:
1. the pm_stay_awake and pm_relax with count based and call with paired for fix current concurency issue.
2. RPROC_RECOVERY_FAIL can be another patch for continue try to do recovery work.
3. handle RPROC_DETACHED case.

Quoted 191 lines · Mathieu Poirier
Quoted 129 lines · Arnaud POULIQUEN

Hi,

On 10/24/22 05:17, Aiqun(Maria) Yu wrote:
On 10/22/2022 3:34 AM, Mathieu Poirier wrote:
On Wed, 19 Oct 2022 at 23:52, Aiqun(Maria) Yu <[email protected]> wrote:

On 10/14/2022 2:03 AM, Mathieu Poirier wrote:
On Thu, Oct 13, 2022 at 11:34:42AM -0600, Mathieu Poirier wrote:
On Thu, Oct 13, 2022 at 09:40:09AM +0800, Aiqun(Maria) Yu wrote:
Hi Mathieu,

On 10/13/2022 4:43 AM, Mathieu Poirier wrote:
Please add what has changed from one version to another, either in a cover
letter or after the "Signed-off-by". There are many examples on how to
do that
on the mailing list.

Thx for the information, will take a note and benefit for next time.

On Fri, Sep 16, 2022 at 03:12:31PM +0800, Maria Yu wrote:
RPROC_OFFLINE state indicate there is no recovery process
is in progress and no chance to do the pm_relax.
Because when recovering from crash, rproc->lock is held and
state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to
RPROC_RUNNING
when rproc_trigger_recovery() returns. If that is not the case then
something
went wrong.

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before
returning,
so we know the remote processor was stopped. Therefore if rproc->state
is set
to RPROC_OFFLINE something went wrong in either request_firmware() or
rproc_start(). Either way the remote processor is offline and the system
probably
in an unknown/unstable. As such I don't see how calling pm_relax() can help
things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully
finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and
last rproc_trigger_recovery bailed out with only
rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair.
Since the subsystem may still can be recovered with customer's next trigger
of rproc_start, and we can make each error out path clean with pm resources.

I suggest spending time understanding what leads to the failure when
recovering
from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when
rproc_shutdown is triggered at similar time. So not an issue from error out
of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned
from
the beginning. What I see happening is the following:

rproc_report_crash()
pm_stay_awake()
queue_work() // current thread is suspended

rproc_shutdown()
rproc_stop()
rproc->state = RPROC_OFFLINE;

rproc_crash_handler_work()
if (rproc->state == RPROC_OFFLINE)
return // pm_relax() is not called

The right way to fix this is to add a pm_relax() in rproc_shutdown() and
rproc_detach(), along with a very descriptive comment as to why it is needed.

Thinking about this further there are more ramifications to consider. Please
confirm the above scenario is what you are facing. I will advise on how to
move
forward if that is the case.

Not sure if the situation is clear or not. So resend the email again.

The above senario is what customer is facing. crash hanppened while at
the same time shutdown is triggered.

Unfortunately this is not enough details to address a problem as
complex as this one.

And the device cannto goes to suspend state after that.
the subsystem can still be start normally after this.

If the code flow I pasted above reflects the problem at hand, the
current patch will not be sufficient to address the issue. If Arnaud
confirms my suspicions we will have to think about a better solution.

Hi Mathiew,

Could you pls have more details of any side effects other then power issue of
the current senario?
Why the current patch is not sufficient pls?

Have the current senario in details with rproc->lock information in details:

| subsystem crashed interrupt issued | user trigger shutdown
| rproc_report_crash() |
| pm_stay_awake() |
| queue_work() |
| |rproc_shutdown
| |mutex_lock(&rproc->lock);
| |rproc_stop()
|rproc_crash_handler_work() |rproc->state = RPROC_OFFLINE;
| |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock); |
|if (rproc->state == RPROC_OFFLINE) |
|return // pm_relax() is not called |rproc_boot
|mutex_unlock(&rproc->lock); |
| |mutex_lock(&rproc->lock);
| |rproc_start()
| |mutex_unlock(&rproc->lock);

Agree with Mathieu, this is not so simple.

Thanks for looking into this.

Quoted 12 lines

Here is my view hoping I haven't missed a point in your discussion or
an other corner cases.

I tried to analyze the issues (in what follows, the term "condition" means
the "if" condition in which Aiqun proposes to add the fix) :

I can see 4 use cases with race condition

1) crash report while already one is treated (rproc_boot_recovery called)
=> not a real use case as if the remote processor is crashed we
should not have a second crash report

That part is of great concern to me. *Theoretically* we should not get a new crash report while one has already been dispatched but the current code accounts for this scenario and as such the possibility can't be dismissed. Therefore we need to expect rproc_report_crash() to be called multiple times before a single instance of rproc_boot_recovery() is scheduled.

Quoted 27 lines

2) rproc_stop executed between the queuing of the crash work and the call of
rproc_crash_handler_work
=> rproc->state = RPROC_OFFLINE
=> we enter in the "condition" and the pm_relax has to be called
=> This commit fix should solve this use case

3) rproc_detach executed between the queue of the crash work and the call of
rproc_crash_handler_work
=> rproc->state = RPROC_DETACHED;
=> we don't go in "the condition" and issue because the recovery reattach
to the remote processor
=> but pm_relax is called
=> probably need an extra fix to avoid to re-attach

4) crash report while already one is treated (rproc_attach_recovery called)
this one corresponds to an auto reboot of the remote processor, with a
new crash
=> rproc->state = RPROC_CRASHED or rproc->state = RPROC_DETACHED;
4)a) rproc->state = RPROC_CRASHED if rproc->recovery_disabled = true
=> should call pm_relax if rproc->recovery_disabled = true
=> commit does not work for this use case

4)b) rproc->state = RPROC_DETACHED if recovery fails
=> error case with an unstable state
=> how to differentiate it from the use case 3) ?
=> introduce a RPROC_RECOVERY_FAIL state?

The case where a recovery fails needs to be considered and is the reason the original patch doesn't work. Right now in rproc_crash_handler_work(), it is not possible to differentiate between a legitimate shutdown request (scenario #2 above) and a recovery that went wrong. I think introducing RPROC_RECOVERY_FAIL would greatly simplify things.

Quoted 9 lines

My initial evaluation had not considered the attach/detach scenarios - thanks for adding that in the mix.

Aiqun, please send a new patchset that adds a new remote processor state, i.e RPROC_RECOVERY_FAIL. There should also be another patch in that set that takes attach/detach scenarios into account. The code between the v6.0 and v6.1 cycle has changed a lot in that area so make sure to properly rebase.

I will try.

Quoted 88 lines
Quoted 88 lines

Then pm_stay_awake is called when the crash work is queued.
It seems to me coherent to call the pm_relax in the work handler.

Here is a quick and dirty patch (not tested) that should take into account the
main use cases ( except 1) and 4)b) )

@@ -2009,8 +2009,18 @@ static void rproc_crash_handler_work(struct work_struct *work)

mutex_lock(&rproc->lock);

- if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
+ if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE ||
+ rproc->state == RPROC_DETACHED) {
/* handle only the first crash detected */
+
+ /*
+ * call pm-relax in following use cases:
+ * - the remote processor has been stopped by the user
+ * - the remote processor is detached
+ + - the remote proc has an autonomous reset but recovery_disabled is true.
+ */
+ if(rproc->state != RPROC_CRASHED || rproc->recovery_disabled)
+ pm_relax(rproc->dev.parent);
mutex_unlock(&rproc->lock);
return;
}

Regards,
Arnaud

Thanks,
Mathieu

When the state is in RPROC_OFFLINE it means separate request
of rproc_stop was done and no need to hold the wakeup source
in crash handler to recover any more.

Signed-off-by: Maria Yu <[email protected]>
---
drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
1 file changed, 11 insertions(+)

diff --git a/drivers/remoteproc/remoteproc_core.c
b/drivers/remoteproc/remoteproc_core.c
index e5279ed9a8d7..6bc7b8b7d01e 100644
--- a/drivers/remoteproc/remoteproc_core.c
+++ b/drivers/remoteproc/remoteproc_core.c
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct
work_struct *work)
if (rproc->state == RPROC_CRASHED || rproc->state ==
RPROC_OFFLINE) {
/* handle only the first crash detected */
mutex_unlock(&rproc->lock);
+ /*
+ * RPROC_OFFLINE state indicate there is no recovery process
+ * is in progress and no chance to have pm_relax in place.
+ * Because when recovering from crash, rproc->lock is held and
+ * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+ * and then unlock rproc->lock.
+ * RPROC_OFFLINE is only an intermediate state in recovery
+ * process.
+ */
+ if (rproc->state == RPROC_OFFLINE)
+ pm_relax(rproc->dev.parent);
return;
}
--
2.7.4

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Quoted 8 lines
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
maintainer
#32 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

Quoted 145 lines · Aiqun(Maria) Yu
Quoted 72 lines · Mathieu Poirier
Quoted 60 lines · Mathieu Poirier
Quoted 51 lines · Aiqun(Maria) Yu

Hi,

Quoted 26 lines · Mathieu Poirier
Quoted 10 lines · Aiqun(Maria) Yu
Quoted 5 lines · Mathieu Poirier

I had a couple of good discussions with our power management expert and even then, the way forward isn't as clear as I would have liked. I am currently travelling and as such don't have the required time to go into greater details, something I will be doing next week.

Thx Mathieu for the info updated.
I'll wait for your update next week then.
let me know any initial questions that you have, perhaps I can also discuss that with our power team at the same time.

The problem is to determine exactly what the WQ_FREEZABLE flag does to the rproc_recovery_wq workueue. The documentation [1] indicate that work items on the WQ are drained before the system is suspended. What I understand from this is that if two work items are queued and one is executing at the time a system suspend is initiated, all three items will be executed before the system is allowed to be suspended. _If_ that is the case, there would not be a need to call pm_stay_awake() and pm_relax() at all.

On the other hand, the PM resource I spoke to thought that in reality things don't happen that way. Taking the same above scenario where 2 work items are queued and one is executing at the time of the suspend, only the work item that is executing will be allowed to execute to completion before the system is suspended. The remaining two items that were queued will not execute.

Quoted 6 lines

If that is the case then we do need to call pm_stay_awake() and pm_relax(), and find another strategy to fix this situation.

Until we have a clear view of how the WQ_FREEZABLE flag works, we won't be able to move forward with this patchset. Unfortunately, I currently do not have the time to look into this.

I had a check on the WQ_FREEZABLE flag, here is my understanding:

when the interrupt happened, it still need pm_stay_awake to make sure queue_work action can active the work instead of susepend the device.

1. If WQ_FREEZABLE, pwq->max_active = 0; // maximum number of in-flight work items is set to 0.

[1].
https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L3748

2. If WQ_FREEZABLE, will only check pwq->nr_active to see if there is still freeze_workqueues_busy.
[2].
https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L5270

Quoted 7 lines

3. When in queue_work, if max_active is 0, when do queue_work it will not actually active the work.
[3].
https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L1418

for the current issue, the work is already complete and forget to set pm_relax in some condition that make the system cannot be suspended.

Quoted 8 lines

What you have above does not describe what happens to work already queue to a WQ_FREEZABLE workqueue when the system gets suspended, and that is the most important thing.

I ended up doing that investigation myself and found the documentation on WQ_FREEZABLE seems to be accurate. I am still discussing this with the Linaro power management team and will get back to you when we reach a conclusion.

I finally have time to get back to this problem...

Quoted 9 lines

As I reported above I went down to the bottom of how WQ_FREEZABLE workqueues work. The conclusion of that research is that _any_ work, queued or currently executing, will finish before the system is suspended. As such, as soon as queue_work() returns in rproc_report_crash(), you can be guaranteed what was just queued will be executed. Therefore, calling pm_stay_awake() in rproc_report_crash() and pm_relax() in rproc_crash_handler_work() should not be needed because the system will not suspend for as long as there is work to be done by the freezable worqueues.

The problem is when crash interrupt happened in a later point(after freeze_kernel_thread), the queue_work will result with the state of WORK_STRUCT_INACTIVE and insert to inactive_works.
And this is the senario that pm_stay_awake wanted to address inside the interrupt handler call.

refer to:
[1]
https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L1512

the current suspend simple flow can be:

        +-+-------------------+
        |   freeze_processes  |
        +----------+----------+
                   |
                   |
       +-----------|-----------+
       | freeze_kernel_thread  |
       +-----------+-----------+
                   |
        +-+--------|----------+
        |   device_suspend    |
        +----------|----------+
                   |
                   |
       +-----------+-----------+
       | pm_wakeup_pending     |
       +-----------|-----------+
                   |
                   |
 +-----------------+-------------------+
 | suspend_ops->enter (machine suspend)|
 +-------------------------------------+

refer to:
suspend_prepare -> suspend_freeze_processes -->freeze_processes and freeze_kernel_threads

refer to:

suspend_devices_and_enter -->
dpm_suspend_start -> dpm_suspend --> device_suspend --> device suspend callbacks
suspend_enter -->dpm_suspend_late-->device_suspend_late

[2]
https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/power/suspend.c#L582

Quoted 3 lines

I don't have details on what Rishabh - the author of the patch that added the calls - was facing but based on the my understanding of freezable workqueues, it seem to have been the wrong solution.

In my opinion, the freezable workqueues are in the right solution. It only protect for the system freeze state.
While if we want the crash interrupt to be handled to prevent current suspend process, it still need to deal with other suspend state to machin suspend senarios.
Typical example can be before disable the device interrupt in device suspend callbacks, it will flush the necessary workqueues.

Quoted 6 lines

while in our case, since the crash handler will invoke a subsystem restart process which reply on user thread uevent firmware loaders, so it need to abort the suspend process with pm_stay_awake called.

So in a word, pm_stay_awake is necessary per my understanding.

I see your point - thanks for being perseverant. What is important to understand is that pm_stay_awake() can stop and reverse the suspend process that is currently underway.

With the code you had for V4 in mind, I suggest to modify rproc_crash_handler_work() to be:

...
...
...

if (rproc->state == RPROC_CRASHED) {
        /* handle only the first crash detected */
        mutex_unlock(&rproc->lock);
        return;
}
if (rproc->state == RPROC_OFFLINE) {
        /* The remote processor is offline, no need to recover anything */
        mutex_unlock(&rproc->lock);
        goto out;
}

...
...
...

out:

        pm_relax();

Do you think that will work for you?

Quoted 519 lines
Quoted 514 lines

Until we have a clear vision on what Rishabh was trying to fix, nothing in that area will be changing. I suggest you touch base with him and sort it out. If that is not possible then calls to pm_stay_awake() and pm_relax() should be removed.

Quoted 508 lines
Quoted 508 lines
Quoted 503 lines

If you want to take on this investigation, keep in mind that any conclusion will need to be backed by a proof. That can be debug messages on a console output or a code reference in the workqueue core.

[1]. https://elixir.bootlin.com/linux/v6.1-rc4/source/Documentation/core-api/workqueue.rst#L184

Quoted 494 lines
Quoted 489 lines · Aiqun(Maria) Yu
Quoted 488 lines

Hi,

Quoted 44 lines · Mathieu Poirier
Quoted 43 lines · Aiqun(Maria) Yu
Quoted 41 lines · Mathieu Poirier
Quoted 6 lines · Aiqun(Maria) Yu

Hi,

Let me think about this carefully.

When in RPROC_RECOVERY_FAIL case we want to re-do the recovery process again or just leave the pm_relax?

Neither.

When a recovery fail we don't want to call pm_relax(). The code in rproc_crash_handler_work() becomes:

       if (rproc->state == RPROC_OFFLINE) {
                     /* We have raced with rproc_shutdown() */
                     pm_relax()
               mutex_unlock(&rproc->lock);
               return;
             }

       if (rproc->state == RPROC_CRASHED ||
                 rproc->state == RPROC_RECOVERY_FAILED) {
               /* handle only the first crash detected */
               mutex_unlock(&rproc->lock);
               return;
       }

RPROC_RECOVERY_FAILED gets set in rproc_boot_recovery() if request_firmware() or rproc_start() fail. Function rproc_trigger_recovery() needs to allow for the recovery the the remote processor is in RPROC_RECOVERY_FAILED state. As such the condition becomes:

             /* State could have changed before we got the mutex */
       if (rproc->state != RPROC_CRASHED &&
                 rproc->state != RPROC_RECOVERY_FAILED)
               goto unlock_mutex;

Start with that and we can look at corner cases (if some exists) with a fresh patchset. Note that I have not addressed the attach/detach() scenario in the above.

Quoted 6 lines
Quoted 2 lines

If we didn't deal with the recovery failed case with correct pm_relax call, it may left the device in a state that cannot enter to suspend state.

That is what I am looking for. We don't want to give the impression that everything is fine by allowing the device to suspend. If the remote processor can't be recovered than it needs to be dealth with.

For the normal recovery failed case, it still need to do pm_relax to not prevent the device goes to suspend. It is what in normal recovery failed case we do in rproc_crash_handler_work as well. rproc_crash_handler_work will not check the result of the rproc_trigger_recovery return value, and will always do pm_relax.

For current conconrency cornor case as well, it is better to consistant with the current design of recovery fail senarios in normal cases.

I personally agree that we shouldn't do nothing when it is a RPROC_RECOVERY_FAILED senario when it is in rproc_crash_handler_work check, because it maybe crash happened when it is trying to do the recovery. So I suggested to do a continue try of trigger recovery again instead of doing nothing and bail out if it is a RPROC_RECOVERY_FAILED state.

Quoted 5 lines
Quoted 4 lines

Because first PROC_RECOVERY_FAIL case cannot ensure it have pm_relax called before the second crash handler call pm_stay_awake or not.

Quoted 7 lines

I've been thinking about that part. I don't think adding a wake_count to control calls to pm_stay_awake()/pm_relax() is the best way to go. There is a similar count happening in the PM runtime subsystem and that is what we should be using. I have asked a power management expert at Linaro for guidance with this matter. I should be able to get back to you with a way forward by the end of next week.

Thx for the specific date provided as well. I will wait until your reply for next patchset then.

Quoted 392 lines
Quoted 392 lines

So, What about the atomic count along with pm_relax and pm_stay_awake ?

struct rproc{
...
atomic_t wake_count;
...
}

rproc_pm_stay_awake()

{
        atomic_inc(&wake_count);
        pm_stay_awake();
}

rproc_pm_relax()

{
        if (atomic_dec_return(&wake_count) == 0)
                pm_stay_awake();
}

can refer code like:

rproc_report_crash()

{
        ...
        rproc_pm_stay_awake();
        queue_work();
        ...
}

rproc_crash_handler_work()

{
        ...
        if (rproc->state == RPROC_OFFLINE || rproc->state == RPROC_CRASHED) {
                      /* We have raced with rproc_shutdown() */
                      rproc_pm_relax();
                mutex_unlock(&rproc->lock);
                return;
              }
        ...
}
Quoted 350 lines

Thanks,
Mathieu

Quoted 346 lines

recovery fail case 1:

|                                      |firstcrash interrupt issued
| second crashed interrupt issued      | rproc_report_crash()
| rproc_report_crash()                 |          pm_stay_awake()
|          pm_stay_awake()             |          queue_work()
|          queue_work()                |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|rproc_crash_handler_work()            |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock);             |pm_relax()
|if (rproc->state == RPROC_OFFLINE)    |
|return // shouldn't do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

recovery fail case 2:

|                                      |firstcrash interrupt issued
|                                      | rproc_report_crash()
|                                      |          pm_stay_awake()
|                                      |          queue_work()
|                                      |rproc_crash_handler_work()
|                                      |mutex_lock(&rproc->lock);
|                                      |rproc_stop()
|                                      |rproc->state = RPROC_OFFLINE;
|                                      |RPROC_RECOVERY_FAIL //new
|                                      |mutex_unlock(&rproc->lock);
|                                      |pm_relax()
|
| second crashed interrupt issued      |
| rproc_report_crash()                 |
|          pm_stay_awake()             |
|          queue_work()                |
|pm_stay_awake()
|mutex_lock(&rproc->lock);
|if (rproc->state == RPROC_OFFLINE)    |
|return // still need do pm_relax if RPROC_RECOVERY_FAIL?  |
|mutex_unlock(&rproc->lock);           |
|                                      |
|                                      |
|                                      |

Maybe I can have:
1. the pm_stay_awake and pm_relax with count based and call with paired for fix current concurency issue.
2. RPROC_RECOVERY_FAIL can be another patch for continue try to do recovery work.
3. handle RPROC_DETACHED case.

Quoted 200 lines · Mathieu Poirier

On Fri, 28 Oct 2022 at 09:31, Arnaud POULIQUEN
<[email protected]> wrote:

Hi,

On 10/24/22 05:17, Aiqun(Maria) Yu wrote:
On 10/22/2022 3:34 AM, Mathieu Poirier wrote:
On Wed, 19 Oct 2022 at 23:52, Aiqun(Maria) Yu <[email protected]> wrote:

On 10/14/2022 2:03 AM, Mathieu Poirier wrote:
On Thu, Oct 13, 2022 at 11:34:42AM -0600, Mathieu Poirier wrote:
On Thu, Oct 13, 2022 at 09:40:09AM +0800, Aiqun(Maria) Yu wrote:
Hi Mathieu,

On 10/13/2022 4:43 AM, Mathieu Poirier wrote:
Please add what has changed from one version to another, either in a cover
letter or after the "Signed-off-by". There are many examples on how to
do that
on the mailing list.

Thx for the information, will take a note and benefit for next time.

On Fri, Sep 16, 2022 at 03:12:31PM +0800, Maria Yu wrote:
RPROC_OFFLINE state indicate there is no recovery process
is in progress and no chance to do the pm_relax.
Because when recovering from crash, rproc->lock is held and
state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to
RPROC_RUNNING
when rproc_trigger_recovery() returns. If that is not the case then
something
went wrong.

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before
returning,
so we know the remote processor was stopped. Therefore if rproc->state
is set
to RPROC_OFFLINE something went wrong in either request_firmware() or
rproc_start(). Either way the remote processor is offline and the system
probably
in an unknown/unstable. As such I don't see how calling pm_relax() can help
things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully
finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and
last rproc_trigger_recovery bailed out with only
rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair.
Since the subsystem may still can be recovered with customer's next trigger
of rproc_start, and we can make each error out path clean with pm resources.

I suggest spending time understanding what leads to the failure when
recovering
from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when
rproc_shutdown is triggered at similar time. So not an issue from error out
of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned
from
the beginning. What I see happening is the following:

rproc_report_crash()
pm_stay_awake()
queue_work() // current thread is suspended

rproc_shutdown()
rproc_stop()
rproc->state = RPROC_OFFLINE;

rproc_crash_handler_work()
if (rproc->state == RPROC_OFFLINE)
return // pm_relax() is not called

The right way to fix this is to add a pm_relax() in rproc_shutdown() and
rproc_detach(), along with a very descriptive comment as to why it is needed.

Thinking about this further there are more ramifications to consider. Please
confirm the above scenario is what you are facing. I will advise on how to
move
forward if that is the case.

Not sure if the situation is clear or not. So resend the email again.

The above senario is what customer is facing. crash hanppened while at
the same time shutdown is triggered.

Unfortunately this is not enough details to address a problem as
complex as this one.

And the device cannto goes to suspend state after that.
the subsystem can still be start normally after this.

If the code flow I pasted above reflects the problem at hand, the
current patch will not be sufficient to address the issue. If Arnaud
confirms my suspicions we will have to think about a better solution.

Hi Mathiew,

Could you pls have more details of any side effects other then power issue of
the current senario?
Why the current patch is not sufficient pls?

Have the current senario in details with rproc->lock information in details:

| subsystem crashed interrupt issued | user trigger shutdown
| rproc_report_crash() |
| pm_stay_awake() |
| queue_work() |
| |rproc_shutdown
| |mutex_lock(&rproc->lock);
| |rproc_stop()
|rproc_crash_handler_work() |rproc->state = RPROC_OFFLINE;
| |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock); |
|if (rproc->state == RPROC_OFFLINE) |
|return // pm_relax() is not called |rproc_boot
|mutex_unlock(&rproc->lock); |
| |mutex_lock(&rproc->lock);
| |rproc_start()
| |mutex_unlock(&rproc->lock);

Agree with Mathieu, this is not so simple.

Thanks for looking into this.

Here is my view hoping I haven't missed a point in your discussion or
an other corner cases.

I tried to analyze the issues (in what follows, the term "condition" means
the "if" condition in which Aiqun proposes to add the fix) :

I can see 4 use cases with race condition

1) crash report while already one is treated (rproc_boot_recovery called)
=> not a real use case as if the remote processor is crashed we
should not have a second crash report

That part is of great concern to me. *Theoretically* we should not
get a new crash report while one has already been dispatched but the
current code accounts for this scenario and as such the possibility
can't be dismissed. Therefore we need to expect rproc_report_crash()
to be called multiple times before a single instance of
rproc_boot_recovery() is scheduled.

2) rproc_stop executed between the queuing of the crash work and the call of
rproc_crash_handler_work
=> rproc->state = RPROC_OFFLINE
=> we enter in the "condition" and the pm_relax has to be called
=> This commit fix should solve this use case

3) rproc_detach executed between the queue of the crash work and the call of
rproc_crash_handler_work
=> rproc->state = RPROC_DETACHED;
=> we don't go in "the condition" and issue because the recovery reattach
to the remote processor
=> but pm_relax is called
=> probably need an extra fix to avoid to re-attach

4) crash report while already one is treated (rproc_attach_recovery called)
this one corresponds to an auto reboot of the remote processor, with a
new crash
=> rproc->state = RPROC_CRASHED or rproc->state = RPROC_DETACHED;
4)a) rproc->state = RPROC_CRASHED if rproc->recovery_disabled = true
=> should call pm_relax if rproc->recovery_disabled = true
=> commit does not work for this use case

4)b) rproc->state = RPROC_DETACHED if recovery fails
=> error case with an unstable state
=> how to differentiate it from the use case 3) ?
=> introduce a RPROC_RECOVERY_FAIL state?

The case where a recovery fails needs to be considered and is the
reason the original patch doesn't work. Right now in
rproc_crash_handler_work(), it is not possible to differentiate
between a legitimate shutdown request (scenario #2 above) and a
recovery that went wrong. I think introducing RPROC_RECOVERY_FAIL
would greatly simplify things.

My initial evaluation had not considered the attach/detach scenarios -
thanks for adding that in the mix.

Aiqun, please send a new patchset that adds a new remote processor
state, i.e RPROC_RECOVERY_FAIL. There should also be another patch in
that set that takes attach/detach scenarios into account. The code
between the v6.0 and v6.1 cycle has changed a lot in that area so make
sure to properly rebase.

I will try.

Quoted 88 lines

Then pm_stay_awake is called when the crash work is queued.
It seems to me coherent to call the pm_relax in the work handler.

Here is a quick and dirty patch (not tested) that should take into account the
main use cases ( except 1) and 4)b) )

@@ -2009,8 +2009,18 @@ static void rproc_crash_handler_work(struct work_struct *work)

mutex_lock(&rproc->lock);

- if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
+ if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE ||
+ rproc->state == RPROC_DETACHED) {
/* handle only the first crash detected */
+
+ /*
+ * call pm-relax in following use cases:
+ * - the remote processor has been stopped by the user
+ * - the remote processor is detached
+ + - the remote proc has an autonomous reset but recovery_disabled is true.
+ */
+ if(rproc->state != RPROC_CRASHED || rproc->recovery_disabled)
+ pm_relax(rproc->dev.parent);
mutex_unlock(&rproc->lock);
return;
}

Regards,
Arnaud

Thanks,
Mathieu

When the state is in RPROC_OFFLINE it means separate request
of rproc_stop was done and no need to hold the wakeup source
in crash handler to recover any more.

Signed-off-by: Maria Yu <[email protected]>
---
drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
1 file changed, 11 insertions(+)

diff --git a/drivers/remoteproc/remoteproc_core.c
b/drivers/remoteproc/remoteproc_core.c
index e5279ed9a8d7..6bc7b8b7d01e 100644
--- a/drivers/remoteproc/remoteproc_core.c
+++ b/drivers/remoteproc/remoteproc_core.c
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct
work_struct *work)
if (rproc->state == RPROC_CRASHED || rproc->state ==
RPROC_OFFLINE) {
/* handle only the first crash detected */
mutex_unlock(&rproc->lock);
+ /*
+ * RPROC_OFFLINE state indicate there is no recovery process
+ * is in progress and no chance to have pm_relax in place.
+ * Because when recovering from crash, rproc->lock is held and
+ * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+ * and then unlock rproc->lock.
+ * RPROC_OFFLINE is only an intermediate state in recovery
+ * process.
+ */
+ if (rproc->state == RPROC_OFFLINE)
+ pm_relax(rproc->dev.parent);
return;
}
--
2.7.4

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Quoted 8 lines
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
#33 lore raw

Re: [PATCH v4] remoteproc: core: do pm relax when in RPROC_OFFLINE

Quoted 150 lines · Mathieu Poirier
Quoted 149 lines · Aiqun(Maria) Yu
Quoted 72 lines · Mathieu Poirier
Quoted 60 lines · Mathieu Poirier
Quoted 51 lines · Aiqun(Maria) Yu

Hi,

Quoted 26 lines · Mathieu Poirier
Quoted 10 lines · Aiqun(Maria) Yu
Quoted 5 lines · Mathieu Poirier

I had a couple of good discussions with our power management expert and even then, the way forward isn't as clear as I would have liked. I am currently travelling and as such don't have the required time to go into greater details, something I will be doing next week.

Thx Mathieu for the info updated.
I'll wait for your update next week then.
let me know any initial questions that you have, perhaps I can also discuss that with our power team at the same time.

The problem is to determine exactly what the WQ_FREEZABLE flag does to the rproc_recovery_wq workueue. The documentation [1] indicate that work items on the WQ are drained before the system is suspended. What I understand from this is that if two work items are queued and one is executing at the time a system suspend is initiated, all three items will be executed before the system is allowed to be suspended. _If_ that is the case, there would not be a need to call pm_stay_awake() and pm_relax() at all.

On the other hand, the PM resource I spoke to thought that in reality things don't happen that way. Taking the same above scenario where 2 work items are queued and one is executing at the time of the suspend, only the work item that is executing will be allowed to execute to completion before the system is suspended. The remaining two items that were queued will not execute.

Quoted 6 lines

If that is the case then we do need to call pm_stay_awake() and pm_relax(), and find another strategy to fix this situation.

Until we have a clear view of how the WQ_FREEZABLE flag works, we won't be able to move forward with this patchset. Unfortunately, I currently do not have the time to look into this.

I had a check on the WQ_FREEZABLE flag, here is my understanding:

when the interrupt happened, it still need pm_stay_awake to make sure queue_work action can active the work instead of susepend the device.

1. If WQ_FREEZABLE, pwq->max_active = 0; // maximum number of in-flight work items is set to 0.

[1].
https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L3748

2. If WQ_FREEZABLE, will only check pwq->nr_active to see if there is still freeze_workqueues_busy.
[2].
https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L5270

Quoted 7 lines

3. When in queue_work, if max_active is 0, when do queue_work it will not actually active the work.
[3].
https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L1418

for the current issue, the work is already complete and forget to set pm_relax in some condition that make the system cannot be suspended.

Quoted 8 lines

What you have above does not describe what happens to work already queue to a WQ_FREEZABLE workqueue when the system gets suspended, and that is the most important thing.

I ended up doing that investigation myself and found the documentation on WQ_FREEZABLE seems to be accurate. I am still discussing this with the Linaro power management team and will get back to you when we reach a conclusion.

I finally have time to get back to this problem...

Quoted 9 lines

As I reported above I went down to the bottom of how WQ_FREEZABLE workqueues work. The conclusion of that research is that _any_ work, queued or currently executing, will finish before the system is suspended. As such, as soon as queue_work() returns in rproc_report_crash(), you can be guaranteed what was just queued will be executed. Therefore, calling pm_stay_awake() in rproc_report_crash() and pm_relax() in rproc_crash_handler_work() should not be needed because the system will not suspend for as long as there is work to be done by the freezable worqueues.

The problem is when crash interrupt happened in a later point(after freeze_kernel_thread), the queue_work will result with the state of WORK_STRUCT_INACTIVE and insert to inactive_works.
And this is the senario that pm_stay_awake wanted to address inside the interrupt handler call.

refer to:
[1]
https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/workqueue.c#L1512

the current suspend simple flow can be:

         +-+-------------------+
         |   freeze_processes  |
         +----------+----------+
                    |
                    |
        +-----------|-----------+
        | freeze_kernel_thread  |
        +-----------+-----------+
                    |
         +-+--------|----------+
         |   device_suspend    |
         +----------|----------+
                    |
                    |
        +-----------+-----------+
        | pm_wakeup_pending     |
        +-----------|-----------+
                    |
                    |
  +-----------------+-------------------+
  | suspend_ops->enter (machine suspend)|
  +-------------------------------------+

refer to:
suspend_prepare -> suspend_freeze_processes -->freeze_processes and freeze_kernel_threads

refer to:

suspend_devices_and_enter -->
dpm_suspend_start -> dpm_suspend --> device_suspend --> device suspend callbacks
suspend_enter -->dpm_suspend_late-->device_suspend_late

[2]
https://elixir.bootlin.com/linux/v6.1-rc4/source/kernel/power/suspend.c#L582

Quoted 3 lines

I don't have details on what Rishabh - the author of the patch that added the calls - was facing but based on the my understanding of freezable workqueues, it seem to have been the wrong solution.

In my opinion, the freezable workqueues are in the right solution. It only protect for the system freeze state.
While if we want the crash interrupt to be handled to prevent current suspend process, it still need to deal with other suspend state to machin suspend senarios.
Typical example can be before disable the device interrupt in device suspend callbacks, it will flush the necessary workqueues.

while in our case, since the crash handler will invoke a subsystem restart process which reply on user thread uevent firmware loaders, so it need to abort the suspend process with pm_stay_awake called.

Quoted 6 lines
Quoted 2 lines

So in a word, pm_stay_awake is necessary per my understanding.

I see your point - thanks for being perseverant. What is important to understand is that pm_stay_awake() can stop and reverse the suspend process that is currently underway.

Thx for the understanding as well. Perseverant in most of time can be a win win result.

Quoted 20 lines

With the code you had for V4 in mind, I suggest to modify rproc_crash_handler_work() to be:

...
...
...

if (rproc->state == RPROC_CRASHED) {
         /* handle only the first crash detected */
         mutex_unlock(&rproc->lock);
         return;
}
if (rproc->state == RPROC_OFFLINE) {
         /* The remote processor is offline, no need to recover anything */
         mutex_unlock(&rproc->lock);
         goto out;
}
Quoted 8 lines

...
...
...

out:

         pm_relax();
Quoted 1 line

Do you think that will work for you?

This is good.

Also, alloc_ordered_workqueue instead of alloc_workqueue is suggested in the same patch.
Because with alloc_ordered_workqueue, only 1 work will be actively executing. It is to avoid unnecessary contentions which have multi crash work executing at the same time.

I will upload new patchset for discussion.

Quoted 520 lines
Quoted 520 lines
Quoted 515 lines

Until we have a clear vision on what Rishabh was trying to fix, nothing in that area will be changing. I suggest you touch base with him and sort it out. If that is not possible then calls to pm_stay_awake() and pm_relax() should be removed.

Quoted 508 lines
Quoted 508 lines
Quoted 503 lines

If you want to take on this investigation, keep in mind that any conclusion will need to be backed by a proof. That can be debug messages on a console output or a code reference in the workqueue core.

[1]. https://elixir.bootlin.com/linux/v6.1-rc4/source/Documentation/core-api/workqueue.rst#L184

Quoted 494 lines
Quoted 489 lines · Aiqun(Maria) Yu
Quoted 488 lines

Hi,

Quoted 44 lines · Mathieu Poirier
Quoted 43 lines · Aiqun(Maria) Yu
Quoted 41 lines · Mathieu Poirier
Quoted 6 lines · Aiqun(Maria) Yu

Hi,

Let me think about this carefully.

When in RPROC_RECOVERY_FAIL case we want to re-do the recovery process again
or just leave the pm_relax?

Neither.

When a recovery fail we don't want to call pm_relax(). The code in rproc_crash_handler_work() becomes:

        if (rproc->state == RPROC_OFFLINE) {
                      /* We have raced with rproc_shutdown() */
                      pm_relax()
                mutex_unlock(&rproc->lock);
                return;
              }

        if (rproc->state == RPROC_CRASHED ||
                  rproc->state == RPROC_RECOVERY_FAILED) {
                /* handle only the first crash detected */
                mutex_unlock(&rproc->lock);
                return;
        }

RPROC_RECOVERY_FAILED gets set in rproc_boot_recovery() if request_firmware() or rproc_start() fail. Function rproc_trigger_recovery() needs to allow for the recovery the the remote processor is in RPROC_RECOVERY_FAILED state. As such the condition becomes:

              /* State could have changed before we got the mutex */
        if (rproc->state != RPROC_CRASHED &&
                  rproc->state != RPROC_RECOVERY_FAILED)
                goto unlock_mutex;

Start with that and we can look at corner cases (if some exists) with a fresh patchset. Note that I have not addressed the attach/detach() scenario in the above.

Quoted 6 lines
Quoted 2 lines

If we didn't deal with the recovery failed case with correct pm_relax call, it may left the device in a state that cannot enter to suspend state.

That is what I am looking for. We don't want to give the impression that everything is fine by allowing the device to suspend. If the remote processor can't be recovered than it needs to be dealth with.

For the normal recovery failed case, it still need to do pm_relax to not prevent the device goes to suspend. It is what in normal recovery failed case we do in rproc_crash_handler_work as well. rproc_crash_handler_work will not check the result of the rproc_trigger_recovery return value, and will always do pm_relax.

For current conconrency cornor case as well, it is better to consistant with the current design of recovery fail senarios in normal cases.

I personally agree that we shouldn't do nothing when it is a RPROC_RECOVERY_FAILED senario when it is in rproc_crash_handler_work check, because it maybe crash happened when it is trying to do the recovery. So I suggested to do a continue try of trigger recovery again instead of doing nothing and bail out if it is a RPROC_RECOVERY_FAILED state.

Quoted 5 lines
Quoted 4 lines

Because first PROC_RECOVERY_FAIL case cannot ensure it have pm_relax called before the second crash handler call pm_stay_awake or not.

Quoted 7 lines

I've been thinking about that part. I don't think adding a wake_count to control calls to pm_stay_awake()/pm_relax() is the best way to go. There is a similar count happening in the PM runtime subsystem and that is what we should be using. I have asked a power management expert at Linaro for guidance with this matter. I should be able to get back to you with a way forward by the end of next week.

Thx for the specific date provided as well. I will wait until your reply for next patchset then.

Quoted 392 lines
Quoted 392 lines

So, What about the atomic count along with pm_relax and pm_stay_awake ?

struct rproc{
...
atomic_t wake_count;
...
}

rproc_pm_stay_awake()

{
         atomic_inc(&wake_count);
         pm_stay_awake();
}

rproc_pm_relax()

{
         if (atomic_dec_return(&wake_count) == 0)
                 pm_stay_awake();
}

can refer code like:

rproc_report_crash()

{
         ...
         rproc_pm_stay_awake();
         queue_work();
         ...
}

rproc_crash_handler_work()

{
         ...
         if (rproc->state == RPROC_OFFLINE || rproc->state == RPROC_CRASHED) {
                       /* We have raced with rproc_shutdown() */
                       rproc_pm_relax();
                 mutex_unlock(&rproc->lock);
                 return;
               }
         ...
}
Quoted 350 lines

Thanks,
Mathieu

Quoted 346 lines

recovery fail case 1:
| |firstcrash interrupt issued
| second crashed interrupt issued | rproc_report_crash()
| rproc_report_crash() | pm_stay_awake()
| pm_stay_awake() | queue_work()
| queue_work() |rproc_crash_handler_work()
| |mutex_lock(&rproc->lock);
| |rproc_stop()
|rproc_crash_handler_work() |rproc->state = RPROC_OFFLINE;
| |RPROC_RECOVERY_FAIL //new
| |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock); |pm_relax()
|if (rproc->state == RPROC_OFFLINE) |
|return // shouldn't do pm_relax if RPROC_RECOVERY_FAIL? |
|mutex_unlock(&rproc->lock); |
| |
| |
| |

recovery fail case 2:
| |firstcrash interrupt issued
| | rproc_report_crash()
| | pm_stay_awake()
| | queue_work()
| |rproc_crash_handler_work()
| |mutex_lock(&rproc->lock);
| |rproc_stop()
| |rproc->state = RPROC_OFFLINE;
| |RPROC_RECOVERY_FAIL //new
| |mutex_unlock(&rproc->lock);
| |pm_relax()
|
| second crashed interrupt issued |
| rproc_report_crash() |
| pm_stay_awake() |
| queue_work() |
|pm_stay_awake()
|mutex_lock(&rproc->lock);
|if (rproc->state == RPROC_OFFLINE) |
|return // still need do pm_relax if RPROC_RECOVERY_FAIL? |
|mutex_unlock(&rproc->lock); |
| |
| |
| |

Maybe I can have:
1. the pm_stay_awake and pm_relax with count based and call with paired for
fix current concurency issue.
2. RPROC_RECOVERY_FAIL can be another patch for continue try to do recovery
work.
3. handle RPROC_DETACHED case.

On 11/2/2022 4:11 AM, Mathieu Poirier wrote:
On Fri, 28 Oct 2022 at 09:31, Arnaud POULIQUEN
<[email protected]> wrote:

Hi,

On 10/24/22 05:17, Aiqun(Maria) Yu wrote:
On 10/22/2022 3:34 AM, Mathieu Poirier wrote:
On Wed, 19 Oct 2022 at 23:52, Aiqun(Maria) Yu <[email protected]> wrote:

On 10/14/2022 2:03 AM, Mathieu Poirier wrote:
On Thu, Oct 13, 2022 at 11:34:42AM -0600, Mathieu Poirier wrote:
On Thu, Oct 13, 2022 at 09:40:09AM +0800, Aiqun(Maria) Yu wrote:
Hi Mathieu,

On 10/13/2022 4:43 AM, Mathieu Poirier wrote:
Please add what has changed from one version to another, either in a cover
letter or after the "Signed-off-by". There are many examples on how to
do that
on the mailing list.

Thx for the information, will take a note and benefit for next time.

On Fri, Sep 16, 2022 at 03:12:31PM +0800, Maria Yu wrote:
RPROC_OFFLINE state indicate there is no recovery process
is in progress and no chance to do the pm_relax.
Because when recovering from crash, rproc->lock is held and
state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
and then unlock rproc->lock.

You are correct - because the lock is held rproc->state should be set to
RPROC_RUNNING
when rproc_trigger_recovery() returns. If that is not the case then
something
went wrong.

Function rproc_stop() sets rproc->state to RPROC_OFFLINE just before
returning,
so we know the remote processor was stopped. Therefore if rproc->state
is set
to RPROC_OFFLINE something went wrong in either request_firmware() or
rproc_start(). Either way the remote processor is offline and the system
probably
in an unknown/unstable. As such I don't see how calling pm_relax() can help
things along.

PROC_OFFLINE is possible that rproc_shutdown is triggered and successfully
finished.
Even if it is multi crash rproc_crash_handler_work contention issue, and
last rproc_trigger_recovery bailed out with only
rproc->state==RPROC_OFFLINE, it is still worth to do pm_relax in pair.
Since the subsystem may still can be recovered with customer's next trigger
of rproc_start, and we can make each error out path clean with pm resources.

I suggest spending time understanding what leads to the failure when
recovering
from a crash and address that problem(s).

In current case, the customer's information is that the issue happened when
rproc_shutdown is triggered at similar time. So not an issue from error out
of rproc_trigger_recovery.

That is a very important element to consider and should have been mentioned
from
the beginning. What I see happening is the following:

rproc_report_crash()
pm_stay_awake()
queue_work() // current thread is suspended

rproc_shutdown()
rproc_stop()
rproc->state = RPROC_OFFLINE;

rproc_crash_handler_work()
if (rproc->state == RPROC_OFFLINE)
return // pm_relax() is not called

The right way to fix this is to add a pm_relax() in rproc_shutdown() and
rproc_detach(), along with a very descriptive comment as to why it is needed.

Thinking about this further there are more ramifications to consider. Please
confirm the above scenario is what you are facing. I will advise on how to
move
forward if that is the case.

Not sure if the situation is clear or not. So resend the email again.

The above senario is what customer is facing. crash hanppened while at
the same time shutdown is triggered.

Unfortunately this is not enough details to address a problem as
complex as this one.

And the device cannto goes to suspend state after that.
the subsystem can still be start normally after this.

If the code flow I pasted above reflects the problem at hand, the
current patch will not be sufficient to address the issue. If Arnaud
confirms my suspicions we will have to think about a better solution.

Hi Mathiew,

Could you pls have more details of any side effects other then power issue of
the current senario?
Why the current patch is not sufficient pls?

Have the current senario in details with rproc->lock information in details:

| subsystem crashed interrupt issued | user trigger shutdown
| rproc_report_crash() |
| pm_stay_awake() |
| queue_work() |
| |rproc_shutdown
| |mutex_lock(&rproc->lock);
| |rproc_stop()
|rproc_crash_handler_work() |rproc->state = RPROC_OFFLINE;
| |mutex_unlock(&rproc->lock);
|mutex_lock(&rproc->lock); |
|if (rproc->state == RPROC_OFFLINE) |
|return // pm_relax() is not called |rproc_boot
|mutex_unlock(&rproc->lock); |
| |mutex_lock(&rproc->lock);
| |rproc_start()
| |mutex_unlock(&rproc->lock);

Agree with Mathieu, this is not so simple.

Thanks for looking into this.

Here is my view hoping I haven't missed a point in your discussion or
an other corner cases.

I tried to analyze the issues (in what follows, the term "condition" means
the "if" condition in which Aiqun proposes to add the fix) :

I can see 4 use cases with race condition

1) crash report while already one is treated (rproc_boot_recovery called)
=> not a real use case as if the remote processor is crashed we
should not have a second crash report

That part is of great concern to me. *Theoretically* we should not
get a new crash report while one has already been dispatched but the
current code accounts for this scenario and as such the possibility
can't be dismissed. Therefore we need to expect rproc_report_crash()
to be called multiple times before a single instance of
rproc_boot_recovery() is scheduled.

2) rproc_stop executed between the queuing of the crash work and the call of
rproc_crash_handler_work
=> rproc->state = RPROC_OFFLINE
=> we enter in the "condition" and the pm_relax has to be called
=> This commit fix should solve this use case

3) rproc_detach executed between the queue of the crash work and the call of
rproc_crash_handler_work
=> rproc->state = RPROC_DETACHED;
=> we don't go in "the condition" and issue because the recovery reattach
to the remote processor
=> but pm_relax is called
=> probably need an extra fix to avoid to re-attach

4) crash report while already one is treated (rproc_attach_recovery called)
this one corresponds to an auto reboot of the remote processor, with a
new crash
=> rproc->state = RPROC_CRASHED or rproc->state = RPROC_DETACHED;
4)a) rproc->state = RPROC_CRASHED if rproc->recovery_disabled = true
=> should call pm_relax if rproc->recovery_disabled = true
=> commit does not work for this use case

4)b) rproc->state = RPROC_DETACHED if recovery fails
=> error case with an unstable state
=> how to differentiate it from the use case 3) ?
=> introduce a RPROC_RECOVERY_FAIL state?

The case where a recovery fails needs to be considered and is the
reason the original patch doesn't work. Right now in
rproc_crash_handler_work(), it is not possible to differentiate
between a legitimate shutdown request (scenario #2 above) and a
recovery that went wrong. I think introducing RPROC_RECOVERY_FAIL
would greatly simplify things.

My initial evaluation had not considered the attach/detach scenarios -
thanks for adding that in the mix.

Aiqun, please send a new patchset that adds a new remote processor
state, i.e RPROC_RECOVERY_FAIL. There should also be another patch in
that set that takes attach/detach scenarios into account. The code
between the v6.0 and v6.1 cycle has changed a lot in that area so make
sure to properly rebase.

I will try.

Then pm_stay_awake is called when the crash work is queued.
It seems to me coherent to call the pm_relax in the work handler.

Here is a quick and dirty patch (not tested) that should take into account the
main use cases ( except 1) and 4)b) )

@@ -2009,8 +2009,18 @@ static void rproc_crash_handler_work(struct work_struct *work)

mutex_lock(&rproc->lock);

- if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
+ if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE ||
+ rproc->state == RPROC_DETACHED) {
/* handle only the first crash detected */
+
+ /*
+ * call pm-relax in following use cases:
+ * - the remote processor has been stopped by the user
+ * - the remote processor is detached
+ + - the remote proc has an autonomous reset but recovery_disabled is true.
+ */
+ if(rproc->state != RPROC_CRASHED || rproc->recovery_disabled)
+ pm_relax(rproc->dev.parent);
mutex_unlock(&rproc->lock);
return;
}

Regards,
Arnaud

Thanks,
Mathieu

When the state is in RPROC_OFFLINE it means separate request
of rproc_stop was done and no need to hold the wakeup source
in crash handler to recover any more.

Signed-off-by: Maria Yu <[email protected]>
---
drivers/remoteproc/remoteproc_core.c | 11 +++++++++++
1 file changed, 11 insertions(+)

diff --git a/drivers/remoteproc/remoteproc_core.c
b/drivers/remoteproc/remoteproc_core.c
index e5279ed9a8d7..6bc7b8b7d01e 100644
--- a/drivers/remoteproc/remoteproc_core.c
+++ b/drivers/remoteproc/remoteproc_core.c
@@ -1956,6 +1956,17 @@ static void rproc_crash_handler_work(struct
work_struct *work)
if (rproc->state == RPROC_CRASHED || rproc->state ==
RPROC_OFFLINE) {
/* handle only the first crash detected */
mutex_unlock(&rproc->lock);
+ /*
+ * RPROC_OFFLINE state indicate there is no recovery process
+ * is in progress and no chance to have pm_relax in place.
+ * Because when recovering from crash, rproc->lock is held and
+ * state is RPROC_CRASHED -> RPROC_OFFLINE -> RPROC_RUNNING,
+ * and then unlock rproc->lock.
+ * RPROC_OFFLINE is only an intermediate state in recovery
+ * process.
+ */
+ if (rproc->state == RPROC_OFFLINE)
+ pm_relax(rproc->dev.parent);
return;
}
--
2.7.4

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Quoted 8 lines
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

--
Thx and BRs,
Aiqun(Maria) Yu

Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu
maintainer
#34 lore raw
In reply to Aiqun(Maria) Yu
Quoted 40 lines · Aiqun(Maria) Yu

Hi Mathieu,

pm_awake and pm_relax needed to be used as a pair. There is chance that pm_relax is not being called, and make the device always in cannot suspend state.

Quoted 14 lines · Mathieu Poirier

Hi Maria,

Quoted 3 lines · Maria Yu

Even if it is not first crash, need to relax the pm wakelock otherwise the device will stay awake.

The goal is exactly to keep the device awake...

Quoted 5 lines
---
  drivers/remoteproc/remoteproc_core.c | 1 +
  1 file changed, 1 insertion(+)
Quoted 13 lines

drivers/remoteproc/remoteproc_core.c static void rproc_crash_handler_work(struct work_struct *work) · line 1959

Quoted 9 lines
drivers/remoteproc/remoteproc_core.c +1 -0
index e5279ed9a8d7..30078043e939 100644
@@ -1956,6 +1956,7 @@ static void rproc_crash_handler_work(struct work_struct *work)
     if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
             /* handle only the first crash detected */
             mutex_unlock(&rproc->lock);
+            pm_relax(rproc->dev.parent);

If we are here it means that rproc_crash_handler_work() has already been called _and_ that a recovery is in process. When the first crash handler completes pm_relax() will be called and the device will go to sleep as expected.

If the rproc->state cannot be changed to running state, the device will always be awake from this return.
Also APROC_OFFLINE state can be given in other path like an shutdown request is issued.

Quoted 4 lines

While this patch is not considering carefully as well, I think I need to upload a new patchset with an ordered workqueue to make each work have each pm_relax before return.
what do you think?

I was travelling this week and as such did not have the time to follow-up with this thread, something I will do next week.

Quoted 16 lines
Quoted 11 lines

Thanks,
Mathieu

Quoted 6 lines
drivers/remoteproc/remoteproc_core.c +0 -1
             return;
     }
 
--

2.7.4

--
Thx and BRs,
Aiqun(Maria) Yu

#35 lore raw
Quoted 41 lines · Mathieu Poirier
Quoted 40 lines · Aiqun(Maria) Yu

Hi Mathieu,

pm_awake and pm_relax needed to be used as a pair. There is chance that pm_relax is not being called, and make the device always in cannot suspend state.

Quoted 14 lines · Mathieu Poirier

Hi Maria,

Quoted 3 lines · Maria Yu

Even if it is not first crash, need to relax the pm wakelock otherwise the device will stay awake.

The goal is exactly to keep the device awake...

Quoted 5 lines
---
   drivers/remoteproc/remoteproc_core.c | 1 +
   1 file changed, 1 insertion(+)
Quoted 13 lines

drivers/remoteproc/remoteproc_core.c static void rproc_crash_handler_work(struct work_struct *work) · line 1959

Quoted 9 lines
drivers/remoteproc/remoteproc_core.c +1 -0
index e5279ed9a8d7..30078043e939 100644
@@ -1956,6 +1956,7 @@ static void rproc_crash_handler_work(struct work_struct *work)
      if (rproc->state == RPROC_CRASHED || rproc->state == RPROC_OFFLINE) {
              /* handle only the first crash detected */
              mutex_unlock(&rproc->lock);
+            pm_relax(rproc->dev.parent);

If we are here it means that rproc_crash_handler_work() has already been called _and_ that a recovery is in process. When the first crash handler completes pm_relax() will be called and the device will go to sleep as expected.

If the rproc->state cannot be changed to running state, the device will always be awake from this return.
Also APROC_OFFLINE state can be given in other path like an shutdown request is issued.

Quoted 8 lines
Quoted 4 lines

While this patch is not considering carefully as well, I think I need to upload a new patchset with an ordered workqueue to make each work have each pm_relax before return.
what do you think?

I was travelling this week and as such did not have the time to follow-up with this thread, something I will do next week.

Thx for follow up. I have new patchset posted on this thread.
After reconsideration, extra action can be done only for RPROC_OFFLINE state. Pls check the newest v4 patchset on this thread.

Quoted 16 lines
Quoted 16 lines
Quoted 11 lines

Thanks,
Mathieu

Quoted 6 lines
drivers/remoteproc/remoteproc_core.c +0 -1
              return;
      }
 
--

2.7.4

--
Thx and BRs,
Aiqun(Maria) Yu

Signature · 2 lines
Thx and BRs,
Aiqun(Maria) Yu