Skip to contentSkip to conversation

drm/amdkfd: fix missed queue reset on queue destroy

PATCH In review Jonathan Kim 2 participants 3 messages Aug 21, 2024 – Aug 22, 2024

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/amdkfd: fix missed queue reset on queue destroy

If a queue is being destroyed but causes a HWS hang on removal, the KFD may issue an unnecessary gpu reset if the destroyed queue can be fixed by a queue reset.

This is because the queue has been removed from the KFD's queue list prior to the preemption action on destroy so the reset call will fail to match the HQD PQ reset information against the KFD's queue record to do the actual reset.

Since a queue destroy request is under the same device lock as any other preemption request (which subsumes queue reset calls), transiently store the destroyed queue's reference so that a potential subsequent queue reset call can check against this queue as well.

---
 drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 10 +++++++++-
 drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h |  1 +
 2 files changed, 10 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +9 -1
index 577d121cc6d1..09e39a72ca31 100644
@@ -1842,6 +1842,8 @@ static int start_cpsch(struct device_queue_manager *dqm)
 		goto fail_detect_hang_buffer;
 	}
 
+	dqm->cur_destroyed_queue = NULL;
+
 	dqm_unlock(dqm);
 
 	return 0;
@@ -2105,7 +2107,7 @@ static void set_queue_as_reset(struct device_queue_manager *dqm, struct queue *q
 		q->properties.queue_id, q->process->pasid);
 
 	pdd->has_reset_queue = true;
-	if (q->properties.is_active) {
+	if (q->properties.is_active && dqm->cur_destroyed_queue != q) {
 		q->properties.is_active = false;
 		decrement_queue_count(dqm, qpd, q);
 	}
@@ -2160,6 +2162,10 @@ static struct queue *find_queue_by_address(struct device_queue_manager *dqm, uin
 	struct qcm_process_device *qpd;
 	struct queue *q;
 
+	if (dqm->cur_destroyed_queue &&
+	    dqm->cur_destroyed_queue->properties.queue_address == queue_address)
+		return dqm->cur_destroyed_queue;
+
 	list_for_each_entry(cur, &dqm->queues, list) {
 		qpd = cur->qpd;
 		list_for_each_entry(q, &qpd->queues_list, list) {
@@ -2409,6 +2415,7 @@ static int destroy_queue_cpsch(struct device_queue_manager *dqm,
 
 	list_del(&q->list);
 	qpd->queue_count--;
+	dqm->cur_destroyed_queue = q;
 	if (q->properties.is_active) {
 		decrement_queue_count(dqm, qpd, q);
 		if (!dqm->dev->kfd->shared_resources.enable_mes) {
@@ -2421,6 +2428,7 @@ static int destroy_queue_cpsch(struct device_queue_manager *dqm,
 			retval = remove_queue_mes(dqm, q, qpd);
 		}
 	}
+	dqm->cur_destroyed_queue = NULL;
 
 	/*
 	 * Unconditionally decrement this counter, regardless of the queue's
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h +1 -0
index 08b40826ad1e..5425c1dd7924 100644
@@ -285,6 +285,7 @@ struct device_queue_manager {
 	struct dqm_detect_hang_info *detect_hang_info;
 	size_t detect_hang_info_size;
 	int detect_hang_count;
+	struct queue *cur_destroyed_queue;
 };
 
 void device_queue_manager_init_cik(
Signature · 1 line
2.34.1
maintainer
#1 lore raw
Quoted 9 lines · Jonathan Kim

If a queue is being destroyed but causes a HWS hang on removal, the KFD may issue an unnecessary gpu reset if the destroyed queue can be fixed by a queue reset.

This is because the queue has been removed from the KFD's queue list prior to the preemption action on destroy so the reset call will fail to match the HQD PQ reset information against the KFD's queue record to do the actual reset.

Quoted 4 lines

Since a queue destroy request is under the same device lock as any other preemption request (which subsumes queue reset calls), transiently store the destroyed queue's reference so that a potential subsequent queue reset call can check against this queue as well.

Maybe this could be simplified by disabling the queues before destroying it. That way the queue would still exist when it's being unmapped and you don't need to hack the special case "cur_destroyed_queue" into the queue reset code.

Regards,
  Felix

Quoted 68 lines
---
  drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 10 +++++++++-
  drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h |  1 +
  2 files changed, 10 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +9 -1
index 577d121cc6d1..09e39a72ca31 100644
@@ -1842,6 +1842,8 @@ static int start_cpsch(struct device_queue_manager *dqm)
  		goto fail_detect_hang_buffer;
  	}
 
+	dqm->cur_destroyed_queue = NULL;
+
  	dqm_unlock(dqm);
 
  	return 0;
@@ -2105,7 +2107,7 @@ static void set_queue_as_reset(struct device_queue_manager *dqm, struct queue *q
  		q->properties.queue_id, q->process->pasid);
 
  	pdd->has_reset_queue = true;
-	if (q->properties.is_active) {
+	if (q->properties.is_active && dqm->cur_destroyed_queue != q) {
  		q->properties.is_active = false;
  		decrement_queue_count(dqm, qpd, q);
  	}
@@ -2160,6 +2162,10 @@ static struct queue *find_queue_by_address(struct device_queue_manager *dqm, uin
  	struct qcm_process_device *qpd;
  	struct queue *q;
 
+	if (dqm->cur_destroyed_queue &&
+	    dqm->cur_destroyed_queue->properties.queue_address == queue_address)
+		return dqm->cur_destroyed_queue;
+
  	list_for_each_entry(cur, &dqm->queues, list) {
  		qpd = cur->qpd;
  		list_for_each_entry(q, &qpd->queues_list, list) {
@@ -2409,6 +2415,7 @@ static int destroy_queue_cpsch(struct device_queue_manager *dqm,
 
  	list_del(&q->list);
  	qpd->queue_count--;
+	dqm->cur_destroyed_queue = q;
  	if (q->properties.is_active) {
  		decrement_queue_count(dqm, qpd, q);
  		if (!dqm->dev->kfd->shared_resources.enable_mes) {
@@ -2421,6 +2428,7 @@ static int destroy_queue_cpsch(struct device_queue_manager *dqm,
  			retval = remove_queue_mes(dqm, q, qpd);
  		}
  	}
+	dqm->cur_destroyed_queue = NULL;
 
  	/*
  	 * Unconditionally decrement this counter, regardless of the queue's
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h +1 -0
index 08b40826ad1e..5425c1dd7924 100644
@@ -285,6 +285,7 @@ struct device_queue_manager {
  	struct dqm_detect_hang_info *detect_hang_info;
  	size_t detect_hang_info_size;
  	int detect_hang_count;
+	struct queue *cur_destroyed_queue;
  };
 
  void device_queue_manager_init_cik(
#2 lore raw

[Public]

Quoted 24 lines

-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Wednesday, August 21, 2024 5:51 PM
To: Kim, Jonathan <[email protected]>; [email protected]
Cc: Kasiviswanathan, Harish <[email protected]>; Deucher, Alexander <[email protected]>
Subject: Re: [PATCH] drm/amdkfd: fix missed queue reset on queue destroy

Quoted 9 lines · Jonathan Kim

If a queue is being destroyed but causes a HWS hang on removal, the KFD may issue an unnecessary gpu reset if the destroyed queue can be fixed by a queue reset.

This is because the queue has been removed from the KFD's queue list prior to the preemption action on destroy so the reset call will fail to match the HQD PQ reset information against the KFD's queue record to do the actual reset.

Quoted 4 lines

Since a queue destroy request is under the same device lock as any other preemption request (which subsumes queue reset calls), transiently store the destroyed queue's reference so that a potential subsequent queue reset call can check against this queue as well.

Quoted 4 lines

Maybe this could be simplified by disabling the queues before destroying it. That way the queue would still exist when it's being unmapped and you don't need to hack the special case "cur_destroyed_queue" into the queue reset code.

Thanks Felix. That's a much simpler fix.
Sending it out.

Jon

Quoted 82 lines

Regards,
Felix

Quoted 4 lines
---
  drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 10

+++++++++-

Review

Quoted 4 lines
  drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h |  1 +
  2 files changed, 10 insertions(+), 1 deletion(-)
(quoted diff) +0 -0

b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c

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

Quoted 4 lines

index 577d121cc6d1..09e39a72ca31 100644

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
@@ -1842,6 +1842,8 @@ static int start_cpsch(struct

device_queue_manager *dqm)

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static void set_queue_as_reset(struct

Quoted 9 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +2 -0
            goto fail_detect_hang_buffer;
    }
 
+   dqm->cur_destroyed_queue = NULL;
+
    dqm_unlock(dqm);
 
    return 0;
@@ -2105,7 +2107,7 @@ static void set_queue_as_reset(struct

device_queue_manager *dqm, struct queue *q

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static struct queue

Quoted 9 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -1
            q->properties.queue_id, q->process->pasid);
 
    pdd->has_reset_queue = true;
-   if (q->properties.is_active) {
+   if (q->properties.is_active && dqm->cur_destroyed_queue != q) {
            q->properties.is_active = false;
            decrement_queue_count(dqm, qpd, q);
    }
@@ -2160,6 +2162,10 @@ static struct queue

*find_queue_by_address(struct device_queue_manager *dqm, uin

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static struct queue *find_queue_by_address(struct device_queue_manager *dqm, uin

Quoted 5 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +2 -0
    struct qcm_process_device *qpd;
    struct queue *q;
 
+   if (dqm->cur_destroyed_queue &&
+       dqm->cur_destroyed_queue->properties.queue_address ==

queue_address)

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

Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +2 -0
+           return dqm->cur_destroyed_queue;
+
    list_for_each_entry(cur, &dqm->queues, list) {
            qpd = cur->qpd;
            list_for_each_entry(q, &qpd->queues_list, list) {
@@ -2409,6 +2415,7 @@ static int destroy_queue_cpsch(struct

device_queue_manager *dqm,

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

Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -0
    list_del(&q->list);
    qpd->queue_count--;
+   dqm->cur_destroyed_queue = q;
    if (q->properties.is_active) {
            decrement_queue_count(dqm, qpd, q);
            if (!dqm->dev->kfd->shared_resources.enable_mes) {
@@ -2421,6 +2428,7 @@ static int destroy_queue_cpsch(struct

device_queue_manager *dqm,

Review

Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -0
                    retval = remove_queue_mes(dqm, q, qpd);
            }
    }
+   dqm->cur_destroyed_queue = NULL;
 
    /*
     * Unconditionally decrement this counter, regardless of the queue's
(quoted diff) +0 -0

b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h

Quoted 11 lines

index 08b40826ad1e..5425c1dd7924 100644

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h +1 -0
@@ -285,6 +285,7 @@ struct device_queue_manager {
    struct dqm_detect_hang_info *detect_hang_info;
    size_t detect_hang_info_size;
    int detect_hang_count;
+   struct queue *cur_destroyed_queue;
  };
 
  void device_queue_manager_init_cik(