Skip to contentSkip to conversation

drm/amdkfd: Fix GWS queue count

PATCH Superseded David Yat Sin 2 participants 7 messages Apr 13, 2022 – Apr 14, 2022

Subsystems (3)

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/2 drm/amdkfd: Fix GWS queue count

Queue can be inactive during process termination. This would cause dqm->gws_queue_count to not be decremented. There can only be 1 GWS queue per device process so moving the logic out of loop.

---
 .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c    | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +6 -6
index acf4f7975850..7c107b88d944 100644
@@ -1945,17 +1945,17 @@ static int process_termination_cpsch(struct device_queue_manager *dqm,
 		else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
 			deallocate_sdma_queue(dqm, q);
 
-		if (q->properties.is_active) {
+		if (q->properties.is_active)
 			decrement_queue_count(dqm, q->properties.type);
-			if (q->properties.is_gws) {
-				dqm->gws_queue_count--;
-				qpd->mapped_gws_queue = false;
-			}
-		}
 
 		dqm->total_queue_count--;
 	}
 
+	if (qpd->mapped_gws_queue) {
+		dqm->gws_queue_count--;
+		qpd->mapped_gws_queue = false;
+	}
+
 	/* Unregister process */
 	list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) {
 		if (qpd == cur->qpd) {
Signature · 1 line
2.30.2
#1 lore raw

PATCH 2/2 drm/amdkfd: CRIU add support for GWS queues

Adding support to checkpoint/restore GWS(Global Wave Sync) queues.

---
 drivers/gpu/drm/amd/amdkfd/kfd_chardev.c      |  2 +-
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h         |  4 ++--
 .../amd/amdkfd/kfd_process_queue_manager.c    | 22 ++++++++++++++-----
 3 files changed, 19 insertions(+), 9 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1 -1
index e1e2362841f8..b2c72ddd9c1c 100644
@@ -1377,7 +1377,7 @@ static int kfd_ioctl_alloc_queue_gws(struct file *filep,
 		goto out_unlock;
 	}
 
-	retval = pqm_set_gws(&p->pqm, args->queue_id, args->num_gws ? dev->gws : NULL);
+	retval = pqm_set_gws(&p->pqm, args->queue_id, args->num_gws ? dev->gws : NULL, false);
 	mutex_unlock(&p->mutex);
 
 	args->first_gws = 0;
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +2 -2
index f36062be9ca8..3ec32675210c 100644
@@ -1102,7 +1102,7 @@ struct kfd_criu_queue_priv_data {
 	uint32_t priority;
 	uint32_t q_percent;
 	uint32_t doorbell_id;
-	uint32_t is_gws;
+	uint32_t gws;
 	uint32_t sdma_id;
 	uint32_t eop_ring_buffer_size;
 	uint32_t ctx_save_restore_area_size;
@@ -1199,7 +1199,7 @@ int pqm_update_queue_properties(struct process_queue_manager *pqm, unsigned int
 int pqm_update_mqd(struct process_queue_manager *pqm, unsigned int qid,
 			struct mqd_update_info *minfo);
 int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
-			void *gws);
+			void *gws, bool criu_restore);
 struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm,
 						unsigned int qid);
 struct queue *pqm_get_user_queue(struct process_queue_manager *pqm,
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +16 -6
index 6eca9509f2e3..3eb9d43fdaac 100644
@@ -91,7 +91,7 @@ void kfd_process_dequeue_from_device(struct kfd_process_device *pdd)
 }
 
 int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
-			void *gws)
+		void *gws, bool criu_restore)
 {
 	struct kfd_dev *dev = NULL;
 	struct process_queue_node *pqn;
@@ -135,8 +135,14 @@ int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
 	pqn->q->gws = mem;
 	pdd->qpd.num_gws = gws ? dev->adev->gds.gws_size : 0;
 
-	return pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
-							pqn->q, NULL);
+	ret = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
+						    pqn->q, NULL);
+
+	if (!ret && criu_restore) {
+		dev->dqm->gws_queue_count++;
+		pdd->qpd.mapped_gws_queue = true;
+	}
+	return ret;
 }
 
 void kfd_process_dequeue_from_all_devices(struct kfd_process *p)
@@ -636,6 +642,8 @@ static int criu_checkpoint_queue(struct kfd_process_device *pdd,
 	q_data->ctx_save_restore_area_size =
 		q->properties.ctx_save_restore_area_size;
 
+	q_data->gws = !!q->gws;
+
 	ret = pqm_checkpoint_mqd(&pdd->process->pqm, q->properties.queue_id, mqd, ctl_stack);
 	if (ret) {
 		pr_err("Failed checkpoint queue_mqd (%d)\n", ret);
@@ -743,7 +751,6 @@ static void set_queue_properties_from_criu(struct queue_properties *qp,
 					  struct kfd_criu_queue_priv_data *q_data)
 {
 	qp->is_interop = false;
-	qp->is_gws = q_data->is_gws;
 	qp->queue_percent = q_data->q_percent;
 	qp->priority = q_data->priority;
 	qp->queue_address = q_data->q_address;
@@ -826,12 +833,15 @@ int kfd_criu_restore_queue(struct kfd_process *p,
 				NULL);
 	if (ret) {
 		pr_err("Failed to create new queue err:%d\n", ret);
-		ret = -EINVAL;
+		goto exit;
 	}
 
+	if (q_data->gws)
+		ret = pqm_set_gws(&p->pqm, q_data->q_id, pdd->dev->gws, true);
+
 exit:
 	if (ret)
-		pr_err("Failed to create queue (%d)\n", ret);
+		pr_err("Failed to restore queue (%d)\n", ret);
 	else
 		pr_debug("Queue id %d was restored successfully\n", queue_id);
 
Signature · 1 line
2.30.2
maintainer
#2 lore raw
In reply to David Yat Sin

Am 2022-04-13 um 11:51 schrieb David Yat Sin:

Quoted 15 lines

Queue can be inactive during process termination. This would cause dqm->gws_queue_count to not be decremented. There can only be 1 GWS queue per device process so moving the logic out of loop.

---
  .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c    | 12 ++++++------
  1 file changed, 6 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
index acf4f7975850..7c107b88d944 100644
@@ -1945,17 +1945,17 @@ static int process_termination_cpsch(struct device_queue_manager *dqm,
  		else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int process_termination_cpsch(struct device_queue_manager *dqm, · line 1954

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -6
  			deallocate_sdma_queue(dqm, q);
 
-		if (q->properties.is_active) {
+		if (q->properties.is_active)
  			decrement_queue_count(dqm, q->properties.type);
-			if (q->properties.is_gws) {
-				dqm->gws_queue_count--;
-				qpd->mapped_gws_queue = false;
-			}
-		}

This looks like the original intention was to decrement the counter for inactive GWS queues. This makes sense because this counter is used to determine whether the runlist is oversubscribed. Inactive queues are not in the runlist, so they should not be counted.

I see that the counter is updated when queues are activated and deactivated in update_queue. So IMO this patch is both incorrect and unnecessary. Did you see an actual problem with the GWS counter during process termination? If so, I'd like to know more to understand the root cause.

Regards,
  Felix

Quoted 12 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +5 -0
  		dqm->total_queue_count--;
  	}
 
+	if (qpd->mapped_gws_queue) {
+		dqm->gws_queue_count--;
+		qpd->mapped_gws_queue = false;
+	}
+
  	/* Unregister process */
  	list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) {
  		if (qpd == cur->qpd) {
#3 lore raw
Quoted 45 lines

-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Thursday, April 14, 2022 10:52 AM
To: Yat Sin, David <[email protected]>; [email protected]
Subject: Re: [PATCH 1/2] drm/amdkfd: Fix GWS queue count

Am 2022-04-13 um 11:51 schrieb David Yat Sin:

Quoted 7 lines

Queue can be inactive during process termination. This would cause dqm->gws_queue_count to not be decremented. There can only be 1 GWS queue per device process so moving the logic out of loop.

---
  .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c    | 12 ++++++-----

-

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int process_termination_cpsch(struct

Quoted 8 lines
  1 file changed, 6 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
index acf4f7975850..7c107b88d944 100644
@@ -1945,17 +1945,17 @@ static int process_termination_cpsch(struct

device_queue_manager *dqm,

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int process_termination_cpsch(struct · line 1945

Quoted 1 line
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
  		else if (q->properties.type ==

KFD_QUEUE_TYPE_SDMA_XGMI)

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int process_termination_cpsch(struct device_queue_manager *dqm, · line 1954

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -6
  			deallocate_sdma_queue(dqm, q);
 
-		if (q->properties.is_active) {
+		if (q->properties.is_active)
  			decrement_queue_count(dqm, q->properties.type);
-			if (q->properties.is_gws) {
-				dqm->gws_queue_count--;
-				qpd->mapped_gws_queue = false;
-			}
-		}

This looks like the original intention was to decrement the counter for inactive GWS queues. This makes sense because this counter is used to determine whether the runlist is oversubscribed. Inactive queues are not in the runlist, so they should not be counted.

Quoted 7 lines

I see that the counter is updated when queues are activated and deactivated in update_queue. So IMO this patch is both incorrect and unnecessary. Did you see an actual problem with the GWS counter during process termination? If so, I'd like to know more to understand the root cause.

Regards,
  Felix

Yes, when using a unit test (for example KFDGWSTest.Semaphore),
1. Put a sleep in the middle of the application (after calling hsaKmtAllocQueueGWS)
2. Run application and kill the application which it is in sleep (application never calls queue.Destroy()).

Then inside kernel dqm->gws_queue_count keeps incrementing each time the experiment is repeated and never goes back to 0. This seems to be because the sleep causes the process to be evicted and q->properties.is_active is false so code does not enter that if statement.

Regards,
David

Quoted 14 lines
Quoted 14 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +5 -0
  		dqm->total_queue_count--;
  	}
 
+	if (qpd->mapped_gws_queue) {
+		dqm->gws_queue_count--;
+		qpd->mapped_gws_queue = false;
+	}
+
  	/* Unregister process */
  	list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) {
  		if (qpd == cur->qpd) {
maintainer
#4 lore raw

Am 2022-04-14 um 11:08 schrieb Yat Sin, David:

Quoted 47 lines
Quoted 47 lines

-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Thursday, April 14, 2022 10:52 AM
To: Yat Sin, David <[email protected]>; [email protected]
Subject: Re: [PATCH 1/2] drm/amdkfd: Fix GWS queue count

Am 2022-04-13 um 11:51 schrieb David Yat Sin:

Quoted 7 lines

Queue can be inactive during process termination. This would cause dqm->gws_queue_count to not be decremented. There can only be 1 GWS queue per device process so moving the logic out of loop.

---
   .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c    | 12 ++++++-----

-

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int process_termination_cpsch(struct

Quoted 8 lines
   1 file changed, 6 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
index acf4f7975850..7c107b88d944 100644
@@ -1945,17 +1945,17 @@ static int process_termination_cpsch(struct

device_queue_manager *dqm,

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int process_termination_cpsch(struct · line 1945

Quoted 1 line
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
   		else if (q->properties.type ==

KFD_QUEUE_TYPE_SDMA_XGMI)

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int process_termination_cpsch(struct device_queue_manager *dqm, · line 1954

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -6
   			deallocate_sdma_queue(dqm, q);
 
-		if (q->properties.is_active) {
+		if (q->properties.is_active)
   			decrement_queue_count(dqm, q->properties.type);
-			if (q->properties.is_gws) {
-				dqm->gws_queue_count--;
-				qpd->mapped_gws_queue = false;
-			}
-		}

This looks like the original intention was to decrement the counter for inactive GWS queues. This makes sense because this counter is used to determine whether the runlist is oversubscribed. Inactive queues are not in the runlist, so they should not be counted.

I see that the counter is updated when queues are activated and deactivated in update_queue. So IMO this patch is both incorrect and unnecessary. Did you see an actual problem with the GWS counter during process termination? If so, I'd like to know more to understand the root cause.

Quoted 7 lines
Quoted 2 lines

Regards,
  Felix

Yes, when using a unit test (for example KFDGWSTest.Semaphore),
1. Put a sleep in the middle of the application (after calling hsaKmtAllocQueueGWS)
2. Run application and kill the application which it is in sleep (application never calls queue.Destroy()).

Then inside kernel dqm->gws_queue_count keeps incrementing each time the experiment is repeated and never goes back to 0. This seems to be because the sleep causes the process to be evicted and q->properties.is_active is false so code does not enter that if statement.

That's weird. Can you find out why it's not getting there? In the test you describe, I would expect the queue to be active, so the counter should be decremented by the existing code.

Does the test evict the queues for some reason? is_active gets set to false when a queue is evicted. Looks like we're missing code to update the gws_queue_count in evict/restore_process_queues_cpsch (it is present in evict/restore_process_queues_nocpsch).

Maybe the management of this counter should be moved into increment/decrement_queue_count, so we don't need to duplicate it in many places.

Regards,
  Felix

Quoted 15 lines

Regards,
David

Quoted 11 lines
Quoted 11 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +5 -0
   		dqm->total_queue_count--;
   	}
 
+	if (qpd->mapped_gws_queue) {
+		dqm->gws_queue_count--;
+		qpd->mapped_gws_queue = false;
+	}
+
   	/* Unregister process */
   	list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) {
   		if (qpd == cur->qpd) {
#5 lore raw
Quoted 71 lines

-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Thursday, April 14, 2022 11:42 AM
To: Yat Sin, David <[email protected]>; [email protected]
Subject: Re: [PATCH 1/2] drm/amdkfd: Fix GWS queue count

Am 2022-04-14 um 11:08 schrieb Yat Sin, David:

Quoted 11 lines
Quoted 11 lines

-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Thursday, April 14, 2022 10:52 AM
To: Yat Sin, David <[email protected]>; [email protected]
Subject: Re: [PATCH 1/2] drm/amdkfd: Fix GWS queue count

Am 2022-04-13 um 11:51 schrieb David Yat Sin:

Quoted 2 lines

Queue can be inactive during process termination. This would cause dqm->gws_queue_count to not be decremented. There can only be 1

GWS

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines

queue per device process so moving the logic out of loop.

---
   .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c    | 12 ++++++-

----

Quoted 27 lines
Quoted 27 lines

-

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int process_termination_cpsch(struct

Quoted 8 lines
   1 file changed, 6 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
index acf4f7975850..7c107b88d944 100644
@@ -1945,17 +1945,17 @@ static int process_termination_cpsch(struct

device_queue_manager *dqm,

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int process_termination_cpsch(struct · line 1945

Quoted 1 line
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
   		else if (q->properties.type ==

KFD_QUEUE_TYPE_SDMA_XGMI)

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int process_termination_cpsch(struct device_queue_manager *dqm, · line 1954

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -6
   			deallocate_sdma_queue(dqm, q);
 
-		if (q->properties.is_active) {
+		if (q->properties.is_active)
   			decrement_queue_count(dqm, q->properties.type);
-			if (q->properties.is_gws) {
-				dqm->gws_queue_count--;
-				qpd->mapped_gws_queue = false;
-			}
-		}

This looks like the original intention was to decrement the counter for inactive GWS queues. This makes sense because this counter is used to determine whether the runlist is oversubscribed. Inactive queues are not in the runlist, so they should not be counted.

Quoted 3 lines
Quoted 3 lines

I see that the counter is updated when queues are activated and deactivated in update_queue. So IMO this patch is both incorrect and unnecessary. Did you see an actual problem with the GWS counter during

process termination?

Quoted 7 lines
Quoted 4 lines

If so, I'd like to know more to understand the root cause.

Regards,
  Felix

Yes, when using a unit test (for example KFDGWSTest.Semaphore), 1. Put a sleep in the middle of the application (after calling hsaKmtAllocQueueGWS) 2. Run application and kill the application which it

is in sleep (application never calls queue.Destroy()).

Quoted 2 lines

Then inside kernel dqm->gws_queue_count keeps incrementing each time

the experiment is repeated and never goes back to 0. This seems to be because the sleep causes the process to be evicted and q-

Quoted 8 lines
Quoted 1 line

properties.is_active is false so code does not enter that if statement.

That's weird. Can you find out why it's not getting there? In the test you describe, I would expect the queue to be active, so the counter should be decremented by the existing code.

Does the test evict the queues for some reason? is_active gets set to false when a queue is evicted.

Yes, the queue is evicted because of the sleep() call in userspace.

Quoted 3 lines

Looks like we're missing code to update the
gws_queue_count in evict/restore_process_queues_cpsch (it is present in evict/restore_process_queues_nocpsch).

I think this is the actual problem and the increment/decrement should be done there. I did not realize the dqm->gws_queue_count only counts not-evicted queues. I will post new patch with this change instead.

Quoted 4 lines

Maybe the management of this counter should be moved into increment/decrement_queue_count, so we don't need to duplicate it in many places.

ACK

Regards,
David

Quoted 20 lines

Regards,
  Felix

Quoted 15 lines

Regards,
David

Quoted 11 lines
Quoted 11 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +5 -0
   		dqm->total_queue_count--;
   	}
 
+	if (qpd->mapped_gws_queue) {
+		dqm->gws_queue_count--;
+		qpd->mapped_gws_queue = false;
+	}
+
   	/* Unregister process */
   	list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) {
   		if (qpd == cur->qpd) {
maintainer
#6 lore raw

Am 2022-04-14 um 13:56 schrieb Yat Sin, David:

Quoted 68 lines
Quoted 68 lines

-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Thursday, April 14, 2022 11:42 AM
To: Yat Sin, David <[email protected]>; [email protected]
Subject: Re: [PATCH 1/2] drm/amdkfd: Fix GWS queue count

Am 2022-04-14 um 11:08 schrieb Yat Sin, David:

Quoted 11 lines
Quoted 11 lines

-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Thursday, April 14, 2022 10:52 AM
To: Yat Sin, David <[email protected]>; [email protected]
Subject: Re: [PATCH 1/2] drm/amdkfd: Fix GWS queue count

Am 2022-04-13 um 11:51 schrieb David Yat Sin:

Quoted 2 lines

Queue can be inactive during process termination. This would cause dqm->gws_queue_count to not be decremented. There can only be 1

GWS

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines

queue per device process so moving the logic out of loop.

---
    .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c    | 12 ++++++-

----

Quoted 27 lines
Quoted 27 lines

-

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int process_termination_cpsch(struct

Quoted 8 lines
    1 file changed, 6 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
index acf4f7975850..7c107b88d944 100644
@@ -1945,17 +1945,17 @@ static int process_termination_cpsch(struct

device_queue_manager *dqm,

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int process_termination_cpsch(struct · line 1945

Quoted 1 line
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
    		else if (q->properties.type ==

KFD_QUEUE_TYPE_SDMA_XGMI)

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int process_termination_cpsch(struct device_queue_manager *dqm, · line 1954

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -6
    			deallocate_sdma_queue(dqm, q);
 
-		if (q->properties.is_active) {
+		if (q->properties.is_active)
    			decrement_queue_count(dqm, q->properties.type);
-			if (q->properties.is_gws) {
-				dqm->gws_queue_count--;
-				qpd->mapped_gws_queue = false;
-			}
-		}

This looks like the original intention was to decrement the counter for inactive GWS queues. This makes sense because this counter is used to determine whether the runlist is oversubscribed. Inactive queues are not in the runlist, so they should not be counted.

Quoted 3 lines
Quoted 3 lines

I see that the counter is updated when queues are activated and deactivated in update_queue. So IMO this patch is both incorrect and unnecessary. Did you see an actual problem with the GWS counter during

process termination?

Quoted 7 lines
Quoted 4 lines

If so, I'd like to know more to understand the root cause.

Regards,
  Felix

Yes, when using a unit test (for example KFDGWSTest.Semaphore), 1. Put a sleep in the middle of the application (after calling hsaKmtAllocQueueGWS) 2. Run application and kill the application which it

is in sleep (application never calls queue.Destroy()).

Quoted 1 line

Then inside kernel dqm->gws_queue_count keeps incrementing each time

the experiment is repeated and never goes back to 0. This seems to be because the sleep causes the process to be evicted and q-

Quoted 8 lines
Quoted 7 lines
Quoted 1 line

properties.is_active is false so code does not enter that if statement.

That's weird. Can you find out why it's not getting there? In the test you describe, I would expect the queue to be active, so the counter should be decremented by the existing code.

Does the test evict the queues for some reason? is_active gets set to false when a queue is evicted.

Yes, the queue is evicted because of the sleep() call in userspace.

Sleep() shouldn't cause queue evictions. Evictions are usually temporary due to memory manager events (memory evictions or MMU notifiers). More permanent evictions can happen after VM faults. Does the test cause a VM fault before the sleep?

Quoted 5 lines
Quoted 4 lines

Looks like we're missing code to update the
gws_queue_count in evict/restore_process_queues_cpsch (it is present in evict/restore_process_queues_nocpsch).

I think this is the actual problem and the increment/decrement should be done there. I did not realize the dqm->gws_queue_count only counts not-evicted queues. I will post new patch with this change instead.

Thanks,
  Felix

Quoted 27 lines
Quoted 4 lines

Maybe the management of this counter should be moved into increment/decrement_queue_count, so we don't need to duplicate it in many places.

ACK

Regards,
David

Quoted 18 lines

Regards,
  Felix

Quoted 14 lines

Regards,
David

Quoted 11 lines
Quoted 11 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +5 -0
    		dqm->total_queue_count--;
    	}
 
+	if (qpd->mapped_gws_queue) {
+		dqm->gws_queue_count--;
+		qpd->mapped_gws_queue = false;
+	}
+
    	/* Unregister process */
    	list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) {
    		if (qpd == cur->qpd) {