drm/amdkfd: queue kfd interrupt work to different CPU

Because queue_work schedule the work on the same CPU the interrupt
handler is running, if there are many interrupts pending, it takes
longer time for work queue to start, or even worse system will hang.

v2: queue work to same NUMA node for better cache locality
v3: handle cpumask_next wraparound case
Signed-off-by: NPhilip Yang <Philip.Yang@amd.com>
Reviewed-by: NEric Huang <JinhuiEric.Huang@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

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

@@ -822,6 +822,21 @@ static int kfd_resume(struct kfd_dev *kfd)
 	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)
 {
@@ -844,7 +859,7 @@ void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
 				   patched_ihre, &is_patched)
 	    && enqueue_ih_ring_entry(kfd,
 				     is_patched ? patched_ihre : ih_ring_entry))
-		queue_work(kfd->ih_wq, &kfd->interrupt_work);
+		kfd_queue_work(kfd->ih_wq, &kfd->interrupt_work);
 
 	spin_unlock_irqrestore(&kfd->interrupt_lock, flags);
 }