Skip to contentSkip to conversation

drm/amdkfd: Use mmget_not_zero in MMU notifier

PATCH In review Philip Yang 2 participants 3 messages May 30, 2022

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

PATCH 1/2 drm/amdkfd: Use mmget_not_zero in MMU notifier

MMU notifier callback may pass in mm with mm->mm_users==0 when process is exiting, use mmget_no_zero to avoid accessing invalid mm in deferred list work after mm is gone.

---
 drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +3 -0
index 835b5187f0b8..2ba3de0fb8aa 100644
@@ -2307,6 +2307,8 @@ svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
 
 	if (range->event == MMU_NOTIFY_RELEASE)
 		return true;
+	if (!mmget_not_zero(mni->mm))
+		return true;
 
 	start = mni->interval_tree.start;
 	last = mni->interval_tree.last;
@@ -2333,6 +2335,7 @@ svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
 	}
 
 	svm_range_unlock(prange);
+	mmput(mni->mm);
 
 	return true;
 }
Signature · 1 line
2.35.1
#1 lore raw

PATCH 2/2 drm/amdgpu: schedule fault recover work to another CPU

Flooding GPU vm fault interrupt may hold CPU and delay the scheduled recover work to handle vm fault, schedule recover work to another CPU of same NUMA node.

---
 drivers/gpu/drm/amd/amdgpu/vega10_ih.c | 13 +++++++++++--
 drivers/gpu/drm/amd/amdgpu/vega20_ih.c | 13 +++++++++++--
 2 files changed, 22 insertions(+), 4 deletions(-)
drivers/gpu/drm/amd/amdgpu/vega10_ih.c +11 -2
index cdd599a08125..3a47107737a3 100644
@@ -435,12 +435,21 @@ static int vega10_ih_self_irq(struct amdgpu_device *adev,
 			      struct amdgpu_irq_src *source,
 			      struct amdgpu_iv_entry *entry)
 {
+	int cpu, new_cpu;
+
+	cpu = new_cpu = smp_processor_id();
+	do {
+		new_cpu = cpumask_next(new_cpu, cpu_online_mask) % nr_cpu_ids;
+		if (cpu_to_node(new_cpu) == numa_node_id())
+			break;
+	} while (cpu != new_cpu);
+
 	switch (entry->ring_id) {
 	case 1:
-		schedule_work(&adev->irq.ih1_work);
+		schedule_work_on(new_cpu, &adev->irq.ih1_work);
 		break;
 	case 2:
-		schedule_work(&adev->irq.ih2_work);
+		schedule_work_on(new_cpu, &adev->irq.ih2_work);
 		break;
 	default: break;
 	}
drivers/gpu/drm/amd/amdgpu/vega20_ih.c +11 -2
index 3b4eb8285943..a9465f0d8fbe 100644
@@ -487,12 +487,21 @@ static int vega20_ih_self_irq(struct amdgpu_device *adev,
 			      struct amdgpu_irq_src *source,
 			      struct amdgpu_iv_entry *entry)
 {
+	int cpu, new_cpu;
+
+	cpu = new_cpu = smp_processor_id();
+	do {
+		new_cpu = cpumask_next(new_cpu, cpu_online_mask) % nr_cpu_ids;
+		if (cpu_to_node(new_cpu) == numa_node_id())
+			break;
+	} while (cpu != new_cpu);
+
 	switch (entry->ring_id) {
 	case 1:
-		schedule_work(&adev->irq.ih1_work);
+		schedule_work_on(new_cpu, &adev->irq.ih1_work);
 		break;
 	case 2:
-		schedule_work(&adev->irq.ih2_work);
+		schedule_work_on(new_cpu, &adev->irq.ih2_work);
 		break;
 	default: break;
 	}
Signature · 1 line
2.35.1
maintainer Reviewed-by on v1 1/2
#2 lore raw
In reply to Philip Yang

Am 2022-05-30 um 10:57 schrieb Philip Yang:

Quoted 5 lines

MMU notifier callback may pass in mm with mm->mm_users==0 when process is exiting, use mmget_no_zero to avoid accessing invalid mm in deferred list work after mm is gone.

Quoted 25 lines
---
  drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 3 +++
  1 file changed, 3 insertions(+)
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +3 -0
index 835b5187f0b8..2ba3de0fb8aa 100644
@@ -2307,6 +2307,8 @@ svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
 
  	if (range->event == MMU_NOTIFY_RELEASE)
  		return true;
+	if (!mmget_not_zero(mni->mm))
+		return true;
 
  	start = mni->interval_tree.start;
  	last = mni->interval_tree.last;
@@ -2333,6 +2335,7 @@ svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
  	}
 
  	svm_range_unlock(prange);
+	mmput(mni->mm);
 
  	return true;
  }