移去rt

hc

2023-12-08 01573e231f18eb2d99162747186f59511f56b64d

 kernel/drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c
..
..
@@ -26,6 +26,7 @@
26
26
 #include "nbio/nbio_6_1_sh_mask.h"
27
27
 #include "gc/gc_9_0_offset.h"
28
28
 #include "gc/gc_9_0_sh_mask.h"
29
+#include "mp/mp_9_0_offset.h"
29
30
 #include "soc15.h"
30
31
 #include "vega10_ih.h"
31
32
 #include "soc15_common.h"
..
..
@@ -237,19 +238,15 @@
237
238
 	struct amdgpu_virt *virt = container_of(work, struct amdgpu_virt, flr_work);
238
239
 	struct amdgpu_device *adev = container_of(virt, struct amdgpu_device, virt);
239
240
 	int timeout = AI_MAILBOX_POLL_FLR_TIMEDOUT;
240
-	int locked;
241
241
 
242
242
 	/* block amdgpu_gpu_recover till msg FLR COMPLETE received,
243
243
 	 * otherwise the mailbox msg will be ruined/reseted by
244
244
 	 * the VF FLR.
245
-	 *
246
-	 * we can unlock the lock_reset to allow "amdgpu_job_timedout"
247
-	 * to run gpu_recover() after FLR_NOTIFICATION_CMPL received
248
-	 * which means host side had finished this VF's FLR.
249
245
 	 */
250
-	locked = mutex_trylock(&adev->lock_reset);
251
-	if (locked)
252
-		adev->in_gpu_reset = 1;
246
+	if (!down_read_trylock(&adev->reset_sem))
247
+		return;
248
+
249
+	atomic_set(&adev->in_gpu_reset, 1);
253
250
 
254
251
 	do {
255
252
 		if (xgpu_ai_mailbox_peek_msg(adev) == IDH_FLR_NOTIFICATION_CMPL)
..
..
@@ -260,14 +257,14 @@
260
257
 	} while (timeout > 1);
261
258
 
262
259
 flr_done:
263
-	if (locked) {
264
-		adev->in_gpu_reset = 0;
265
-		mutex_unlock(&adev->lock_reset);
266
-	}
260
+	atomic_set(&adev->in_gpu_reset, 0);
261
+	up_read(&adev->reset_sem);
267
262
 
268
263
 	/* Trigger recovery for world switch failure if no TDR */
269
-	if (amdgpu_lockup_timeout == 0)
270
-		amdgpu_device_gpu_recover(adev, NULL, true);
264
+	if (amdgpu_device_should_recover_gpu(adev)
265
+		&& (!amdgpu_device_has_job_running(adev) ||
266
+		adev->sdma_timeout == MAX_SCHEDULE_TIMEOUT))
267
+		amdgpu_device_gpu_recover(adev, NULL);
271
268
 }
272
269
 
273
270
 static int xgpu_ai_set_mailbox_rcv_irq(struct amdgpu_device *adev,
..
..
@@ -295,6 +292,9 @@
295
292
 		if (amdgpu_sriov_runtime(adev))
296
293
 			schedule_work(&adev->virt.flr_work);
297
294
 		break;
295
+		case IDH_QUERY_ALIVE:
296
+			xgpu_ai_mailbox_send_ack(adev);
297
+			break;
298
298
 		/* READY_TO_ACCESS_GPU is fetched by kernel polling, IRQ can ignore
299
299
 		 * it byfar since that polling thread will handle it,
300
300
 		 * other msg like flr complete is not handled here.