drm/amdkfd: Queue interrupt work to different CPU

For CPX mode, each KFD node has interrupt worker to process ih_fifo to
send events to user space. Currently all interrupt workers of same adev
queue to same CPU, all workers execution are actually serialized and
this cause KFD ih_fifo overflow when CPU usage is high.

Use per-GPU unbounded highpri queue with number of workers equals to
number of partitions, let queue_work select the next CPU round robin
among the local CPUs of same NUMA.

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

3 files changed, 20 insertions, 33 deletions

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
index 9b49563f2c42..a29374c86405 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
@@ -649,6 +649,14 @@ static void kfd_cleanup_nodes(struct kfd_dev *kfd, unsigned int num_nodes)
 	struct kfd_node *knode;
 	unsigned int i;
 
+	/*
+	 * flush_work ensures that there are no outstanding
+	 * work-queue items that will access interrupt_ring. New work items
+	 * can't be created because we stopped interrupt handling above.
+	 */
+	flush_workqueue(kfd->ih_wq);
+	destroy_workqueue(kfd->ih_wq);
+
 	for (i = 0; i < num_nodes; i++) {
 		knode = kfd->nodes[i];
 		device_queue_manager_uninit(knode->dqm);
@@ -1066,21 +1074,6 @@ static int kfd_resume(struct kfd_node *node)
 	return err;
 }
 
-static inline void kfd_queue_work(struct workqueue_struct *wq,
-				  struct work_struct *work)
-{
-	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);
-
-	queue_work_on(new_cpu, wq, work);
-}
-
 /* This is called directly from KGD at ISR. */
 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
 {
@@ -1106,7 +1099,7 @@ void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
 			    	patched_ihre, &is_patched)
 		    && enqueue_ih_ring_entry(node,
 			    	is_patched ? patched_ihre : ih_ring_entry)) {
-			kfd_queue_work(node->ih_wq, &node->interrupt_work);
+			queue_work(node->kfd->ih_wq, &node->interrupt_work);
 			spin_unlock_irqrestore(&node->interrupt_lock, flags);
 			return;
 		}
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_interrupt.c b/drivers/gpu/drm/amd/amdkfd/kfd_interrupt.c
index 8e00800f3207..6beb786c582a 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_interrupt.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_interrupt.c
@@ -62,11 +62,14 @@ int kfd_interrupt_init(struct kfd_node *node)
 		return r;
 	}
 
-	node->ih_wq = alloc_workqueue("KFD IH", WQ_HIGHPRI, 1);
-	if (unlikely(!node->ih_wq)) {
-		kfifo_free(&node->ih_fifo);
-		dev_err(node->adev->dev, "Failed to allocate KFD IH workqueue\n");
-		return -ENOMEM;
+	if (!node->kfd->ih_wq) {
+		node->kfd->ih_wq = alloc_workqueue("KFD IH", WQ_HIGHPRI | WQ_UNBOUND,
+						   node->kfd->num_nodes);
+		if (unlikely(!node->kfd->ih_wq)) {
+			kfifo_free(&node->ih_fifo);
+			dev_err(node->adev->dev, "Failed to allocate KFD IH workqueue\n");
+			return -ENOMEM;
+		}
 	}
 	spin_lock_init(&node->interrupt_lock);
 
@@ -96,16 +99,6 @@ void kfd_interrupt_exit(struct kfd_node *node)
 	spin_lock_irqsave(&node->interrupt_lock, flags);
 	node->interrupts_active = false;
 	spin_unlock_irqrestore(&node->interrupt_lock, flags);
-
-	/*
-	 * flush_work ensures that there are no outstanding
-	 * work-queue items that will access interrupt_ring. New work items
-	 * can't be created because we stopped interrupt handling above.
-	 */
-	flush_workqueue(node->ih_wq);
-
-	destroy_workqueue(node->ih_wq);
-
 	kfifo_free(&node->ih_fifo);
 }
 
@@ -155,7 +148,7 @@ static void interrupt_wq(struct work_struct *work)
 			/* If we spent more than a second processing signals,
 			 * reschedule the worker to avoid soft-lockup warnings
 			 */
-			queue_work(dev->ih_wq, &dev->interrupt_work);
+			queue_work(dev->kfd->ih_wq, &dev->interrupt_work);
 			break;
 		}
 	}
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index e529fdc1b422..c32b255c0eb2 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
@@ -274,7 +274,6 @@ struct kfd_node {
 
 	/* Interrupts */
 	struct kfifo ih_fifo;
-	struct workqueue_struct *ih_wq;
 	struct work_struct interrupt_work;
 	spinlock_t interrupt_lock;
 
@@ -367,6 +366,8 @@ struct kfd_dev {
 	struct kfd_node *nodes[MAX_KFD_NODES];
 	unsigned int num_nodes;
 
+	struct workqueue_struct *ih_wq;
+
 	/* Kernel doorbells for KFD device */
 	struct amdgpu_bo *doorbells;