disable FB

hc

2024-01-04 1543e317f1da31b75942316931e8f491a8920811

 kernel/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c
..
..
@@ -41,14 +41,15 @@
41
41
 				unsigned int *rlib_size,
42
42
 				bool *over_subscription)
43
43
 {
44
-	unsigned int process_count, queue_count, compute_queue_count;
44
+	unsigned int process_count, queue_count, compute_queue_count, gws_queue_count;
45
45
 	unsigned int map_queue_size;
46
46
 	unsigned int max_proc_per_quantum = 1;
47
47
 	struct kfd_dev *dev = pm->dqm->dev;
48
48
 
49
49
 	process_count = pm->dqm->processes_count;
50
-	queue_count = pm->dqm->queue_count;
51
-	compute_queue_count = queue_count - pm->dqm->sdma_queue_count;
50
+	queue_count = pm->dqm->active_queue_count;
51
+	compute_queue_count = pm->dqm->active_cp_queue_count;
52
+	gws_queue_count = pm->dqm->gws_queue_count;
52
53
 
53
54
 	/* check if there is over subscription
54
55
 	 * Note: the arbitration between the number of VMIDs and
..
..
@@ -61,7 +62,8 @@
61
62
 		max_proc_per_quantum = dev->max_proc_per_quantum;
62
63
 
63
64
 	if ((process_count > max_proc_per_quantum) ||
64
-	    compute_queue_count > get_queues_num(pm->dqm)) {
65
+	    compute_queue_count > get_cp_queues_num(pm->dqm) ||
66
+	    gws_queue_count > 1) {
65
67
 		*over_subscription = true;
66
68
 		pr_debug("Over subscribed runlist\n");
67
69
 	}
..
..
@@ -140,7 +142,7 @@
140
142
 	pm->ib_size_bytes = alloc_size_bytes;
141
143
 
142
144
 	pr_debug("Building runlist ib process count: %d queues count %d\n",
143
-		pm->dqm->processes_count, pm->dqm->queue_count);
145
+		pm->dqm->processes_count, pm->dqm->active_queue_count);
144
146
 
145
147
 	/* build the run list ib packet */
146
148
 	list_for_each_entry(cur, queues, list) {
..
..
@@ -202,11 +204,15 @@
202
204
 
203
205
 	pr_debug("Finished map process and queues to runlist\n");
204
206
 
205
-	if (is_over_subscription)
207
+	if (is_over_subscription) {
208
+		if (!pm->is_over_subscription)
209
+			pr_warn("Runlist is getting oversubscribed. Expect reduced ROCm performance.\n");
206
210
 		retval = pm->pmf->runlist(pm, &rl_buffer[rl_wptr],
207
211
 					*rl_gpu_addr,
208
212
 					alloc_size_bytes / sizeof(uint32_t),
209
213
 					true);
214
+	}
215
+	pm->is_over_subscription = is_over_subscription;
210
216
 
211
217
 	for (i = 0; i < alloc_size_bytes / sizeof(uint32_t); i++)
212
218
 		pr_debug("0x%2X ", rl_buffer[i]);
..
..
@@ -226,10 +232,21 @@
226
232
 	case CHIP_FIJI:
227
233
 	case CHIP_POLARIS10:
228
234
 	case CHIP_POLARIS11:
235
+	case CHIP_POLARIS12:
236
+	case CHIP_VEGAM:
229
237
 		pm->pmf = &kfd_vi_pm_funcs;
230
238
 		break;
231
239
 	case CHIP_VEGA10:
240
+	case CHIP_VEGA12:
241
+	case CHIP_VEGA20:
232
242
 	case CHIP_RAVEN:
243
+	case CHIP_RENOIR:
244
+	case CHIP_ARCTURUS:
245
+	case CHIP_NAVI10:
246
+	case CHIP_NAVI12:
247
+	case CHIP_NAVI14:
248
+	case CHIP_SIENNA_CICHLID:
249
+	case CHIP_NAVY_FLOUNDER:
233
250
 		pm->pmf = &kfd_v9_pm_funcs;
234
251
 		break;
235
252
 	default:
..
..
@@ -250,10 +267,10 @@
250
267
 	return 0;
251
268
 }
252
269
 
253
-void pm_uninit(struct packet_manager *pm)
270
+void pm_uninit(struct packet_manager *pm, bool hanging)
254
271
 {
255
272
 	mutex_destroy(&pm->lock);
256
-	kernel_queue_uninit(pm->priv_queue);
273
+	kernel_queue_uninit(pm->priv_queue, hanging);
257
274
 }
258
275
 
259
276
 int pm_send_set_resources(struct packet_manager *pm,
..
..
@@ -264,7 +281,7 @@
264
281
 
265
282
 	size = pm->pmf->set_resources_size;
266
283
 	mutex_lock(&pm->lock);
267
-	pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
284
+	kq_acquire_packet_buffer(pm->priv_queue,
268
285
 					size / sizeof(uint32_t),
269
286
 					(unsigned int **)&buffer);
270
287
 	if (!buffer) {
..
..
@@ -275,9 +292,9 @@
275
292
 
276
293
 	retval = pm->pmf->set_resources(pm, buffer, res);
277
294
 	if (!retval)
278
-		pm->priv_queue->ops.submit_packet(pm->priv_queue);
295
+		kq_submit_packet(pm->priv_queue);
279
296
 	else
280
-		pm->priv_queue->ops.rollback_packet(pm->priv_queue);
297
+		kq_rollback_packet(pm->priv_queue);
281
298
 
282
299
 out:
283
300
 	mutex_unlock(&pm->lock);
..
..
@@ -302,7 +319,7 @@
302
319
 	packet_size_dwords = pm->pmf->runlist_size / sizeof(uint32_t);
303
320
 	mutex_lock(&pm->lock);
304
321
 
305
-	retval = pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
322
+	retval = kq_acquire_packet_buffer(pm->priv_queue,
306
323
 					packet_size_dwords, &rl_buffer);
307
324
 	if (retval)
308
325
 		goto fail_acquire_packet_buffer;
..
..
@@ -312,14 +329,14 @@
312
329
 	if (retval)
313
330
 		goto fail_create_runlist;
314
331
 
315
-	pm->priv_queue->ops.submit_packet(pm->priv_queue);
332
+	kq_submit_packet(pm->priv_queue);
316
333
 
317
334
 	mutex_unlock(&pm->lock);
318
335
 
319
336
 	return retval;
320
337
 
321
338
 fail_create_runlist:
322
-	pm->priv_queue->ops.rollback_packet(pm->priv_queue);
339
+	kq_rollback_packet(pm->priv_queue);
323
340
 fail_acquire_packet_buffer:
324
341
 	mutex_unlock(&pm->lock);
325
342
 fail_create_runlist_ib:
..
..
@@ -328,7 +345,7 @@
328
345
 }
329
346
 
330
347
 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
331
-			uint32_t fence_value)
348
+			uint64_t fence_value)
332
349
 {
333
350
 	uint32_t *buffer, size;
334
351
 	int retval = 0;
..
..
@@ -338,7 +355,7 @@
338
355
 
339
356
 	size = pm->pmf->query_status_size;
340
357
 	mutex_lock(&pm->lock);
341
-	pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
358
+	kq_acquire_packet_buffer(pm->priv_queue,
342
359
 			size / sizeof(uint32_t), (unsigned int **)&buffer);
343
360
 	if (!buffer) {
344
361
 		pr_err("Failed to allocate buffer on kernel queue\n");
..
..
@@ -348,9 +365,9 @@
348
365
 
349
366
 	retval = pm->pmf->query_status(pm, buffer, fence_address, fence_value);
350
367
 	if (!retval)
351
-		pm->priv_queue->ops.submit_packet(pm->priv_queue);
368
+		kq_submit_packet(pm->priv_queue);
352
369
 	else
353
-		pm->priv_queue->ops.rollback_packet(pm->priv_queue);
370
+		kq_rollback_packet(pm->priv_queue);
354
371
 
355
372
 out:
356
373
 	mutex_unlock(&pm->lock);
..
..
@@ -367,7 +384,7 @@
367
384
 
368
385
 	size = pm->pmf->unmap_queues_size;
369
386
 	mutex_lock(&pm->lock);
370
-	pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
387
+	kq_acquire_packet_buffer(pm->priv_queue,
371
388
 			size / sizeof(uint32_t), (unsigned int **)&buffer);
372
389
 	if (!buffer) {
373
390
 		pr_err("Failed to allocate buffer on kernel queue\n");
..
..
@@ -378,9 +395,9 @@
378
395
 	retval = pm->pmf->unmap_queues(pm, buffer, type, filter, filter_param,
379
396
 				       reset, sdma_engine);
380
397
 	if (!retval)
381
-		pm->priv_queue->ops.submit_packet(pm->priv_queue);
398
+		kq_submit_packet(pm->priv_queue);
382
399
 	else
383
-		pm->priv_queue->ops.rollback_packet(pm->priv_queue);
400
+		kq_rollback_packet(pm->priv_queue);
384
401
 
385
402
 out:
386
403
 	mutex_unlock(&pm->lock);
..
..
@@ -425,7 +442,7 @@
425
442
 
426
443
 	size = pm->pmf->query_status_size;
427
444
 	mutex_lock(&pm->lock);
428
-	pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
445
+	kq_acquire_packet_buffer(pm->priv_queue,
429
446
 			size / sizeof(uint32_t), (unsigned int **)&buffer);
430
447
 	if (!buffer) {
431
448
 		pr_err("Failed to allocate buffer on kernel queue\n");
..
..
@@ -433,7 +450,7 @@
433
450
 		goto out;
434
451
 	}
435
452
 	memset(buffer, 0x55, size);
436
-	pm->priv_queue->ops.submit_packet(pm->priv_queue);
453
+	kq_submit_packet(pm->priv_queue);
437
454
 
438
455
 	pr_info("Submitting %x %x %x %x %x %x %x to HIQ to hang the HWS.",
439
456
 		buffer[0], buffer[1], buffer[2], buffer[3],