drm/amdkfd: invalidate tables on page retry fault

GPU page tables are invalidated by unmapping prange directly at
the mmu notifier, when page fault retry is enabled through
amdgpu_noretry global parameter. The restore page table is
performed at the page fault handler.

If xnack is on, we update GPU mappings after migration to avoid
unnecessary GPUVM faults.

Signed-off-by: Alex Sierra <alex.sierra@amd.com>
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/amd/amdkfd/kfd_migrate.c

@@ -808,7 +808,11 @@ static vm_fault_t svm_migrate_to_ram(struct vm_fault *vmf)
 		pr_debug("failed %d migrate 0x%p [0x%lx 0x%lx] to ram\n", r,
 			 prange, prange->start, prange->last);
 
-	op = SVM_OP_UPDATE_RANGE_NOTIFIER;
+	/* xnack on, update mapping on GPUs with ACCESS_IN_PLACE */
+	if (p->xnack_enabled && parent == prange)
+		op = SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP;
+	else
+		op = SVM_OP_UPDATE_RANGE_NOTIFIER;
 	svm_range_add_list_work(&p->svms, parent, mm, op);
 	schedule_deferred_list_work(&p->svms);
 

drivers/gpu/drm/amd/amdkfd/kfd_svm.c

@@ -912,6 +912,13 @@ svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm,
 		svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE);
 	}
 
+	/* xnack on, update mapping on GPUs with ACCESS_IN_PLACE */
+	if (p->xnack_enabled && prange->work_item.op == SVM_OP_ADD_RANGE) {
+		prange->work_item.op = SVM_OP_ADD_RANGE_AND_MAP;
+		pr_debug("change prange 0x%p [0x%lx 0x%lx] op %d\n",
+			 prange, prange->start, prange->last,
+			 SVM_OP_ADD_RANGE_AND_MAP);
+	}
 	return 0;
 }
 
@@ -1452,25 +1459,52 @@ svm_range_evict(struct svm_range *prange, struct mm_struct *mm,
 		unsigned long start, unsigned long last)
 {
 	struct svm_range_list *svms = prange->svms;
-	int evicted_ranges;
+	struct kfd_process *p;
 	int r = 0;
 
-	atomic_inc(&prange->invalid);
+	p = container_of(svms, struct kfd_process, svms);
-	evicted_ranges = atomic_inc_return(&svms->evicted_ranges);
-	if (evicted_ranges != 1)
-		return r;
 
-	pr_debug("evicting svms 0x%p range [0x%lx 0x%lx]\n",
+	pr_debug("invalidate svms 0x%p prange [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
-		 prange->svms, prange->start, prange->last);
+		 svms, prange->start, prange->last, start, last);
 
-	/* First eviction, stop the queues */
+	if (!p->xnack_enabled) {
-	r = kgd2kfd_quiesce_mm(mm);
+		int evicted_ranges;
-	if (r)
-		pr_debug("failed to quiesce KFD\n");
 
-	pr_debug("schedule to restore svm %p ranges\n", svms);
+		atomic_inc(&prange->invalid);
-	schedule_delayed_work(&svms->restore_work,
+		evicted_ranges = atomic_inc_return(&svms->evicted_ranges);
-		msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
+		if (evicted_ranges != 1)
+			return r;
+
+		pr_debug("evicting svms 0x%p range [0x%lx 0x%lx]\n",
+			 prange->svms, prange->start, prange->last);
+
+		/* First eviction, stop the queues */
+		r = kgd2kfd_quiesce_mm(mm);
+		if (r)
+			pr_debug("failed to quiesce KFD\n");
+
+		pr_debug("schedule to restore svm %p ranges\n", svms);
+		schedule_delayed_work(&svms->restore_work,
+			msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
+	} else {
+		struct svm_range *pchild;
+		unsigned long s, l;
+
+		pr_debug("invalidate unmap svms 0x%p [0x%lx 0x%lx] from GPUs\n",
+			 prange->svms, start, last);
+		list_for_each_entry(pchild, &prange->child_list, child_list) {
+			mutex_lock_nested(&pchild->lock, 1);
+			s = max(start, pchild->start);
+			l = min(last, pchild->last);
+			if (l >= s)
+				svm_range_unmap_from_gpus(pchild, s, l);
+			mutex_unlock(&pchild->lock);
+		}
+		s = max(start, prange->start);
+		l = min(last, prange->last);
+		if (l >= s)
+			svm_range_unmap_from_gpus(prange, s, l);
+	}
 
 	return r;
 }
@@ -1673,12 +1707,25 @@ svm_range_handle_list_op(struct svm_range_list *svms, struct svm_range *prange)
 			 svms, prange, prange->start, prange->last);
 		svm_range_update_notifier_and_interval_tree(mm, prange);
 		break;
+	case SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP:
+		pr_debug("update and map 0x%p prange 0x%p [0x%lx 0x%lx]\n",
+			 svms, prange, prange->start, prange->last);
+		svm_range_update_notifier_and_interval_tree(mm, prange);
+		/* TODO: implement deferred validation and mapping */
+		break;
 	case SVM_OP_ADD_RANGE:
 		pr_debug("add 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms, prange,
 			 prange->start, prange->last);
 		svm_range_add_to_svms(prange);
 		svm_range_add_notifier_locked(mm, prange);
 		break;
+	case SVM_OP_ADD_RANGE_AND_MAP:
+		pr_debug("add and map 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms,
+			 prange, prange->start, prange->last);
+		svm_range_add_to_svms(prange);
+		svm_range_add_notifier_locked(mm, prange);
+		/* TODO: implement deferred validation and mapping */
+		break;
 	default:
 		WARN_ONCE(1, "Unknown prange 0x%p work op %d\n", prange,
 			 prange->work_item.op);
@@ -2301,7 +2348,7 @@ svm_range_set_attr(struct kfd_process *p, uint64_t start, uint64_t size,
 		if (r)
 			goto out_unlock_range;
 
-		if (migrated) {
+		if (migrated && !p->xnack_enabled) {
 			pr_debug("restore_work will update mappings of GPUs\n");
 			mutex_unlock(&prange->migrate_mutex);
 			continue;

drivers/gpu/drm/amd/amdkfd/kfd_svm.h

@@ -44,7 +44,9 @@ enum svm_work_list_ops {
 	SVM_OP_NULL,
 	SVM_OP_UNMAP_RANGE,
 	SVM_OP_UPDATE_RANGE_NOTIFIER,
-	SVM_OP_ADD_RANGE
+	SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP,
+	SVM_OP_ADD_RANGE,
+	SVM_OP_ADD_RANGE_AND_MAP
 };
 
 struct svm_work_list_item {