xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c (revision 24168c5e6dfbdd5b414f048f47f75d64533296ca)
1 /*
2  * Copyright 2019 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  *
22  */
23 
24 #include <linux/firmware.h>
25 #include <drm/drm_exec.h>
26 
27 #include "amdgpu_mes.h"
28 #include "amdgpu.h"
29 #include "soc15_common.h"
30 #include "amdgpu_mes_ctx.h"
31 
32 #define AMDGPU_MES_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
33 #define AMDGPU_ONE_DOORBELL_SIZE 8
34 
35 signed long amdgpu_mes_fence_wait_polling(u64 *fence,
36 					  u64 wait_seq,
37 					  signed long timeout)
38 {
39 
40 	while ((s64)(wait_seq - *fence) > 0 && timeout > 0) {
41 		udelay(2);
42 		timeout -= 2;
43 	}
44 	return timeout > 0 ? timeout : 0;
45 }
46 
47 int amdgpu_mes_doorbell_process_slice(struct amdgpu_device *adev)
48 {
49 	return roundup(AMDGPU_ONE_DOORBELL_SIZE *
50 		       AMDGPU_MES_MAX_NUM_OF_QUEUES_PER_PROCESS,
51 		       PAGE_SIZE);
52 }
53 
54 static int amdgpu_mes_kernel_doorbell_get(struct amdgpu_device *adev,
55 					 int ip_type, uint64_t *doorbell_index)
56 {
57 	unsigned int offset, found;
58 	struct amdgpu_mes *mes = &adev->mes;
59 
60 	if (ip_type == AMDGPU_RING_TYPE_SDMA)
61 		offset = adev->doorbell_index.sdma_engine[0];
62 	else
63 		offset = 0;
64 
65 	found = find_next_zero_bit(mes->doorbell_bitmap, mes->num_mes_dbs, offset);
66 	if (found >= mes->num_mes_dbs) {
67 		DRM_WARN("No doorbell available\n");
68 		return -ENOSPC;
69 	}
70 
71 	set_bit(found, mes->doorbell_bitmap);
72 
73 	/* Get the absolute doorbell index on BAR */
74 	*doorbell_index = mes->db_start_dw_offset + found * 2;
75 	return 0;
76 }
77 
78 static void amdgpu_mes_kernel_doorbell_free(struct amdgpu_device *adev,
79 					   uint32_t doorbell_index)
80 {
81 	unsigned int old, rel_index;
82 	struct amdgpu_mes *mes = &adev->mes;
83 
84 	/* Find the relative index of the doorbell in this object */
85 	rel_index = (doorbell_index - mes->db_start_dw_offset) / 2;
86 	old = test_and_clear_bit(rel_index, mes->doorbell_bitmap);
87 	WARN_ON(!old);
88 }
89 
90 static int amdgpu_mes_doorbell_init(struct amdgpu_device *adev)
91 {
92 	int i;
93 	struct amdgpu_mes *mes = &adev->mes;
94 
95 	/* Bitmap for dynamic allocation of kernel doorbells */
96 	mes->doorbell_bitmap = bitmap_zalloc(PAGE_SIZE / sizeof(u32), GFP_KERNEL);
97 	if (!mes->doorbell_bitmap) {
98 		DRM_ERROR("Failed to allocate MES doorbell bitmap\n");
99 		return -ENOMEM;
100 	}
101 
102 	mes->num_mes_dbs = PAGE_SIZE / AMDGPU_ONE_DOORBELL_SIZE;
103 	for (i = 0; i < AMDGPU_MES_PRIORITY_NUM_LEVELS; i++) {
104 		adev->mes.aggregated_doorbells[i] = mes->db_start_dw_offset + i * 2;
105 		set_bit(i, mes->doorbell_bitmap);
106 	}
107 
108 	return 0;
109 }
110 
111 static int amdgpu_mes_event_log_init(struct amdgpu_device *adev)
112 {
113 	int r;
114 
115 	if (!amdgpu_mes_log_enable)
116 		return 0;
117 
118 	r = amdgpu_bo_create_kernel(adev, AMDGPU_MES_LOG_BUFFER_SIZE, PAGE_SIZE,
119 				    AMDGPU_GEM_DOMAIN_GTT,
120 				    &adev->mes.event_log_gpu_obj,
121 				    &adev->mes.event_log_gpu_addr,
122 				    &adev->mes.event_log_cpu_addr);
123 	if (r) {
124 		dev_warn(adev->dev, "failed to create MES event log buffer (%d)", r);
125 		return r;
126 	}
127 
128 	memset(adev->mes.event_log_cpu_addr, 0, PAGE_SIZE);
129 
130 	return  0;
131 
132 }
133 
134 static void amdgpu_mes_doorbell_free(struct amdgpu_device *adev)
135 {
136 	bitmap_free(adev->mes.doorbell_bitmap);
137 }
138 
139 int amdgpu_mes_init(struct amdgpu_device *adev)
140 {
141 	int i, r;
142 
143 	adev->mes.adev = adev;
144 
145 	idr_init(&adev->mes.pasid_idr);
146 	idr_init(&adev->mes.gang_id_idr);
147 	idr_init(&adev->mes.queue_id_idr);
148 	ida_init(&adev->mes.doorbell_ida);
149 	spin_lock_init(&adev->mes.queue_id_lock);
150 	spin_lock_init(&adev->mes.ring_lock);
151 	mutex_init(&adev->mes.mutex_hidden);
152 
153 	adev->mes.total_max_queue = AMDGPU_FENCE_MES_QUEUE_ID_MASK;
154 	adev->mes.vmid_mask_mmhub = 0xffffff00;
155 	adev->mes.vmid_mask_gfxhub = 0xffffff00;
156 
157 	for (i = 0; i < AMDGPU_MES_MAX_COMPUTE_PIPES; i++) {
158 		/* use only 1st MEC pipes */
159 		if (i >= 4)
160 			continue;
161 		adev->mes.compute_hqd_mask[i] = 0xc;
162 	}
163 
164 	for (i = 0; i < AMDGPU_MES_MAX_GFX_PIPES; i++)
165 		adev->mes.gfx_hqd_mask[i] = i ? 0 : 0xfffffffe;
166 
167 	for (i = 0; i < AMDGPU_MES_MAX_SDMA_PIPES; i++) {
168 		if (amdgpu_ip_version(adev, SDMA0_HWIP, 0) <
169 		    IP_VERSION(6, 0, 0))
170 			adev->mes.sdma_hqd_mask[i] = i ? 0 : 0x3fc;
171 		/* zero sdma_hqd_mask for non-existent engine */
172 		else if (adev->sdma.num_instances == 1)
173 			adev->mes.sdma_hqd_mask[i] = i ? 0 : 0xfc;
174 		else
175 			adev->mes.sdma_hqd_mask[i] = 0xfc;
176 	}
177 
178 	r = amdgpu_device_wb_get(adev, &adev->mes.sch_ctx_offs);
179 	if (r) {
180 		dev_err(adev->dev,
181 			"(%d) ring trail_fence_offs wb alloc failed\n", r);
182 		goto error_ids;
183 	}
184 	adev->mes.sch_ctx_gpu_addr =
185 		adev->wb.gpu_addr + (adev->mes.sch_ctx_offs * 4);
186 	adev->mes.sch_ctx_ptr =
187 		(uint64_t *)&adev->wb.wb[adev->mes.sch_ctx_offs];
188 
189 	r = amdgpu_device_wb_get(adev, &adev->mes.query_status_fence_offs);
190 	if (r) {
191 		amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
192 		dev_err(adev->dev,
193 			"(%d) query_status_fence_offs wb alloc failed\n", r);
194 		goto error_ids;
195 	}
196 	adev->mes.query_status_fence_gpu_addr =
197 		adev->wb.gpu_addr + (adev->mes.query_status_fence_offs * 4);
198 	adev->mes.query_status_fence_ptr =
199 		(uint64_t *)&adev->wb.wb[adev->mes.query_status_fence_offs];
200 
201 	r = amdgpu_device_wb_get(adev, &adev->mes.read_val_offs);
202 	if (r) {
203 		amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
204 		amdgpu_device_wb_free(adev, adev->mes.query_status_fence_offs);
205 		dev_err(adev->dev,
206 			"(%d) read_val_offs alloc failed\n", r);
207 		goto error_ids;
208 	}
209 	adev->mes.read_val_gpu_addr =
210 		adev->wb.gpu_addr + (adev->mes.read_val_offs * 4);
211 	adev->mes.read_val_ptr =
212 		(uint32_t *)&adev->wb.wb[adev->mes.read_val_offs];
213 
214 	r = amdgpu_mes_doorbell_init(adev);
215 	if (r)
216 		goto error;
217 
218 	r = amdgpu_mes_event_log_init(adev);
219 	if (r)
220 		goto error_doorbell;
221 
222 	return 0;
223 
224 error_doorbell:
225 	amdgpu_mes_doorbell_free(adev);
226 error:
227 	amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
228 	amdgpu_device_wb_free(adev, adev->mes.query_status_fence_offs);
229 	amdgpu_device_wb_free(adev, adev->mes.read_val_offs);
230 error_ids:
231 	idr_destroy(&adev->mes.pasid_idr);
232 	idr_destroy(&adev->mes.gang_id_idr);
233 	idr_destroy(&adev->mes.queue_id_idr);
234 	ida_destroy(&adev->mes.doorbell_ida);
235 	mutex_destroy(&adev->mes.mutex_hidden);
236 	return r;
237 }
238 
239 void amdgpu_mes_fini(struct amdgpu_device *adev)
240 {
241 	amdgpu_bo_free_kernel(&adev->mes.event_log_gpu_obj,
242 			      &adev->mes.event_log_gpu_addr,
243 			      &adev->mes.event_log_cpu_addr);
244 
245 	amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
246 	amdgpu_device_wb_free(adev, adev->mes.query_status_fence_offs);
247 	amdgpu_device_wb_free(adev, adev->mes.read_val_offs);
248 	amdgpu_mes_doorbell_free(adev);
249 
250 	idr_destroy(&adev->mes.pasid_idr);
251 	idr_destroy(&adev->mes.gang_id_idr);
252 	idr_destroy(&adev->mes.queue_id_idr);
253 	ida_destroy(&adev->mes.doorbell_ida);
254 	mutex_destroy(&adev->mes.mutex_hidden);
255 }
256 
257 static void amdgpu_mes_queue_free_mqd(struct amdgpu_mes_queue *q)
258 {
259 	amdgpu_bo_free_kernel(&q->mqd_obj,
260 			      &q->mqd_gpu_addr,
261 			      &q->mqd_cpu_ptr);
262 }
263 
264 int amdgpu_mes_create_process(struct amdgpu_device *adev, int pasid,
265 			      struct amdgpu_vm *vm)
266 {
267 	struct amdgpu_mes_process *process;
268 	int r;
269 
270 	/* allocate the mes process buffer */
271 	process = kzalloc(sizeof(struct amdgpu_mes_process), GFP_KERNEL);
272 	if (!process) {
273 		DRM_ERROR("no more memory to create mes process\n");
274 		return -ENOMEM;
275 	}
276 
277 	/* allocate the process context bo and map it */
278 	r = amdgpu_bo_create_kernel(adev, AMDGPU_MES_PROC_CTX_SIZE, PAGE_SIZE,
279 				    AMDGPU_GEM_DOMAIN_GTT,
280 				    &process->proc_ctx_bo,
281 				    &process->proc_ctx_gpu_addr,
282 				    &process->proc_ctx_cpu_ptr);
283 	if (r) {
284 		DRM_ERROR("failed to allocate process context bo\n");
285 		goto clean_up_memory;
286 	}
287 	memset(process->proc_ctx_cpu_ptr, 0, AMDGPU_MES_PROC_CTX_SIZE);
288 
289 	/*
290 	 * Avoid taking any other locks under MES lock to avoid circular
291 	 * lock dependencies.
292 	 */
293 	amdgpu_mes_lock(&adev->mes);
294 
295 	/* add the mes process to idr list */
296 	r = idr_alloc(&adev->mes.pasid_idr, process, pasid, pasid + 1,
297 		      GFP_KERNEL);
298 	if (r < 0) {
299 		DRM_ERROR("failed to lock pasid=%d\n", pasid);
300 		goto clean_up_ctx;
301 	}
302 
303 	INIT_LIST_HEAD(&process->gang_list);
304 	process->vm = vm;
305 	process->pasid = pasid;
306 	process->process_quantum = adev->mes.default_process_quantum;
307 	process->pd_gpu_addr = amdgpu_bo_gpu_offset(vm->root.bo);
308 
309 	amdgpu_mes_unlock(&adev->mes);
310 	return 0;
311 
312 clean_up_ctx:
313 	amdgpu_mes_unlock(&adev->mes);
314 	amdgpu_bo_free_kernel(&process->proc_ctx_bo,
315 			      &process->proc_ctx_gpu_addr,
316 			      &process->proc_ctx_cpu_ptr);
317 clean_up_memory:
318 	kfree(process);
319 	return r;
320 }
321 
322 void amdgpu_mes_destroy_process(struct amdgpu_device *adev, int pasid)
323 {
324 	struct amdgpu_mes_process *process;
325 	struct amdgpu_mes_gang *gang, *tmp1;
326 	struct amdgpu_mes_queue *queue, *tmp2;
327 	struct mes_remove_queue_input queue_input;
328 	unsigned long flags;
329 	int r;
330 
331 	/*
332 	 * Avoid taking any other locks under MES lock to avoid circular
333 	 * lock dependencies.
334 	 */
335 	amdgpu_mes_lock(&adev->mes);
336 
337 	process = idr_find(&adev->mes.pasid_idr, pasid);
338 	if (!process) {
339 		DRM_WARN("pasid %d doesn't exist\n", pasid);
340 		amdgpu_mes_unlock(&adev->mes);
341 		return;
342 	}
343 
344 	/* Remove all queues from hardware */
345 	list_for_each_entry_safe(gang, tmp1, &process->gang_list, list) {
346 		list_for_each_entry_safe(queue, tmp2, &gang->queue_list, list) {
347 			spin_lock_irqsave(&adev->mes.queue_id_lock, flags);
348 			idr_remove(&adev->mes.queue_id_idr, queue->queue_id);
349 			spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
350 
351 			queue_input.doorbell_offset = queue->doorbell_off;
352 			queue_input.gang_context_addr = gang->gang_ctx_gpu_addr;
353 
354 			r = adev->mes.funcs->remove_hw_queue(&adev->mes,
355 							     &queue_input);
356 			if (r)
357 				DRM_WARN("failed to remove hardware queue\n");
358 		}
359 
360 		idr_remove(&adev->mes.gang_id_idr, gang->gang_id);
361 	}
362 
363 	idr_remove(&adev->mes.pasid_idr, pasid);
364 	amdgpu_mes_unlock(&adev->mes);
365 
366 	/* free all memory allocated by the process */
367 	list_for_each_entry_safe(gang, tmp1, &process->gang_list, list) {
368 		/* free all queues in the gang */
369 		list_for_each_entry_safe(queue, tmp2, &gang->queue_list, list) {
370 			amdgpu_mes_queue_free_mqd(queue);
371 			list_del(&queue->list);
372 			kfree(queue);
373 		}
374 		amdgpu_bo_free_kernel(&gang->gang_ctx_bo,
375 				      &gang->gang_ctx_gpu_addr,
376 				      &gang->gang_ctx_cpu_ptr);
377 		list_del(&gang->list);
378 		kfree(gang);
379 
380 	}
381 	amdgpu_bo_free_kernel(&process->proc_ctx_bo,
382 			      &process->proc_ctx_gpu_addr,
383 			      &process->proc_ctx_cpu_ptr);
384 	kfree(process);
385 }
386 
387 int amdgpu_mes_add_gang(struct amdgpu_device *adev, int pasid,
388 			struct amdgpu_mes_gang_properties *gprops,
389 			int *gang_id)
390 {
391 	struct amdgpu_mes_process *process;
392 	struct amdgpu_mes_gang *gang;
393 	int r;
394 
395 	/* allocate the mes gang buffer */
396 	gang = kzalloc(sizeof(struct amdgpu_mes_gang), GFP_KERNEL);
397 	if (!gang) {
398 		return -ENOMEM;
399 	}
400 
401 	/* allocate the gang context bo and map it to cpu space */
402 	r = amdgpu_bo_create_kernel(adev, AMDGPU_MES_GANG_CTX_SIZE, PAGE_SIZE,
403 				    AMDGPU_GEM_DOMAIN_GTT,
404 				    &gang->gang_ctx_bo,
405 				    &gang->gang_ctx_gpu_addr,
406 				    &gang->gang_ctx_cpu_ptr);
407 	if (r) {
408 		DRM_ERROR("failed to allocate process context bo\n");
409 		goto clean_up_mem;
410 	}
411 	memset(gang->gang_ctx_cpu_ptr, 0, AMDGPU_MES_GANG_CTX_SIZE);
412 
413 	/*
414 	 * Avoid taking any other locks under MES lock to avoid circular
415 	 * lock dependencies.
416 	 */
417 	amdgpu_mes_lock(&adev->mes);
418 
419 	process = idr_find(&adev->mes.pasid_idr, pasid);
420 	if (!process) {
421 		DRM_ERROR("pasid %d doesn't exist\n", pasid);
422 		r = -EINVAL;
423 		goto clean_up_ctx;
424 	}
425 
426 	/* add the mes gang to idr list */
427 	r = idr_alloc(&adev->mes.gang_id_idr, gang, 1, 0,
428 		      GFP_KERNEL);
429 	if (r < 0) {
430 		DRM_ERROR("failed to allocate idr for gang\n");
431 		goto clean_up_ctx;
432 	}
433 
434 	gang->gang_id = r;
435 	*gang_id = r;
436 
437 	INIT_LIST_HEAD(&gang->queue_list);
438 	gang->process = process;
439 	gang->priority = gprops->priority;
440 	gang->gang_quantum = gprops->gang_quantum ?
441 		gprops->gang_quantum : adev->mes.default_gang_quantum;
442 	gang->global_priority_level = gprops->global_priority_level;
443 	gang->inprocess_gang_priority = gprops->inprocess_gang_priority;
444 	list_add_tail(&gang->list, &process->gang_list);
445 
446 	amdgpu_mes_unlock(&adev->mes);
447 	return 0;
448 
449 clean_up_ctx:
450 	amdgpu_mes_unlock(&adev->mes);
451 	amdgpu_bo_free_kernel(&gang->gang_ctx_bo,
452 			      &gang->gang_ctx_gpu_addr,
453 			      &gang->gang_ctx_cpu_ptr);
454 clean_up_mem:
455 	kfree(gang);
456 	return r;
457 }
458 
459 int amdgpu_mes_remove_gang(struct amdgpu_device *adev, int gang_id)
460 {
461 	struct amdgpu_mes_gang *gang;
462 
463 	/*
464 	 * Avoid taking any other locks under MES lock to avoid circular
465 	 * lock dependencies.
466 	 */
467 	amdgpu_mes_lock(&adev->mes);
468 
469 	gang = idr_find(&adev->mes.gang_id_idr, gang_id);
470 	if (!gang) {
471 		DRM_ERROR("gang id %d doesn't exist\n", gang_id);
472 		amdgpu_mes_unlock(&adev->mes);
473 		return -EINVAL;
474 	}
475 
476 	if (!list_empty(&gang->queue_list)) {
477 		DRM_ERROR("queue list is not empty\n");
478 		amdgpu_mes_unlock(&adev->mes);
479 		return -EBUSY;
480 	}
481 
482 	idr_remove(&adev->mes.gang_id_idr, gang->gang_id);
483 	list_del(&gang->list);
484 	amdgpu_mes_unlock(&adev->mes);
485 
486 	amdgpu_bo_free_kernel(&gang->gang_ctx_bo,
487 			      &gang->gang_ctx_gpu_addr,
488 			      &gang->gang_ctx_cpu_ptr);
489 
490 	kfree(gang);
491 
492 	return 0;
493 }
494 
495 int amdgpu_mes_suspend(struct amdgpu_device *adev)
496 {
497 	struct idr *idp;
498 	struct amdgpu_mes_process *process;
499 	struct amdgpu_mes_gang *gang;
500 	struct mes_suspend_gang_input input;
501 	int r, pasid;
502 
503 	/*
504 	 * Avoid taking any other locks under MES lock to avoid circular
505 	 * lock dependencies.
506 	 */
507 	amdgpu_mes_lock(&adev->mes);
508 
509 	idp = &adev->mes.pasid_idr;
510 
511 	idr_for_each_entry(idp, process, pasid) {
512 		list_for_each_entry(gang, &process->gang_list, list) {
513 			r = adev->mes.funcs->suspend_gang(&adev->mes, &input);
514 			if (r)
515 				DRM_ERROR("failed to suspend pasid %d gangid %d",
516 					 pasid, gang->gang_id);
517 		}
518 	}
519 
520 	amdgpu_mes_unlock(&adev->mes);
521 	return 0;
522 }
523 
524 int amdgpu_mes_resume(struct amdgpu_device *adev)
525 {
526 	struct idr *idp;
527 	struct amdgpu_mes_process *process;
528 	struct amdgpu_mes_gang *gang;
529 	struct mes_resume_gang_input input;
530 	int r, pasid;
531 
532 	/*
533 	 * Avoid taking any other locks under MES lock to avoid circular
534 	 * lock dependencies.
535 	 */
536 	amdgpu_mes_lock(&adev->mes);
537 
538 	idp = &adev->mes.pasid_idr;
539 
540 	idr_for_each_entry(idp, process, pasid) {
541 		list_for_each_entry(gang, &process->gang_list, list) {
542 			r = adev->mes.funcs->resume_gang(&adev->mes, &input);
543 			if (r)
544 				DRM_ERROR("failed to resume pasid %d gangid %d",
545 					 pasid, gang->gang_id);
546 		}
547 	}
548 
549 	amdgpu_mes_unlock(&adev->mes);
550 	return 0;
551 }
552 
553 static int amdgpu_mes_queue_alloc_mqd(struct amdgpu_device *adev,
554 				     struct amdgpu_mes_queue *q,
555 				     struct amdgpu_mes_queue_properties *p)
556 {
557 	struct amdgpu_mqd *mqd_mgr = &adev->mqds[p->queue_type];
558 	u32 mqd_size = mqd_mgr->mqd_size;
559 	int r;
560 
561 	r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
562 				    AMDGPU_GEM_DOMAIN_GTT,
563 				    &q->mqd_obj,
564 				    &q->mqd_gpu_addr, &q->mqd_cpu_ptr);
565 	if (r) {
566 		dev_warn(adev->dev, "failed to create queue mqd bo (%d)", r);
567 		return r;
568 	}
569 	memset(q->mqd_cpu_ptr, 0, mqd_size);
570 
571 	r = amdgpu_bo_reserve(q->mqd_obj, false);
572 	if (unlikely(r != 0))
573 		goto clean_up;
574 
575 	return 0;
576 
577 clean_up:
578 	amdgpu_bo_free_kernel(&q->mqd_obj,
579 			      &q->mqd_gpu_addr,
580 			      &q->mqd_cpu_ptr);
581 	return r;
582 }
583 
584 static void amdgpu_mes_queue_init_mqd(struct amdgpu_device *adev,
585 				     struct amdgpu_mes_queue *q,
586 				     struct amdgpu_mes_queue_properties *p)
587 {
588 	struct amdgpu_mqd *mqd_mgr = &adev->mqds[p->queue_type];
589 	struct amdgpu_mqd_prop mqd_prop = {0};
590 
591 	mqd_prop.mqd_gpu_addr = q->mqd_gpu_addr;
592 	mqd_prop.hqd_base_gpu_addr = p->hqd_base_gpu_addr;
593 	mqd_prop.rptr_gpu_addr = p->rptr_gpu_addr;
594 	mqd_prop.wptr_gpu_addr = p->wptr_gpu_addr;
595 	mqd_prop.queue_size = p->queue_size;
596 	mqd_prop.use_doorbell = true;
597 	mqd_prop.doorbell_index = p->doorbell_off;
598 	mqd_prop.eop_gpu_addr = p->eop_gpu_addr;
599 	mqd_prop.hqd_pipe_priority = p->hqd_pipe_priority;
600 	mqd_prop.hqd_queue_priority = p->hqd_queue_priority;
601 	mqd_prop.hqd_active = false;
602 
603 	if (p->queue_type == AMDGPU_RING_TYPE_GFX ||
604 	    p->queue_type == AMDGPU_RING_TYPE_COMPUTE) {
605 		mutex_lock(&adev->srbm_mutex);
606 		amdgpu_gfx_select_me_pipe_q(adev, p->ring->me, p->ring->pipe, 0, 0, 0);
607 	}
608 
609 	mqd_mgr->init_mqd(adev, q->mqd_cpu_ptr, &mqd_prop);
610 
611 	if (p->queue_type == AMDGPU_RING_TYPE_GFX ||
612 	    p->queue_type == AMDGPU_RING_TYPE_COMPUTE) {
613 		amdgpu_gfx_select_me_pipe_q(adev, 0, 0, 0, 0, 0);
614 		mutex_unlock(&adev->srbm_mutex);
615 	}
616 
617 	amdgpu_bo_unreserve(q->mqd_obj);
618 }
619 
620 int amdgpu_mes_add_hw_queue(struct amdgpu_device *adev, int gang_id,
621 			    struct amdgpu_mes_queue_properties *qprops,
622 			    int *queue_id)
623 {
624 	struct amdgpu_mes_queue *queue;
625 	struct amdgpu_mes_gang *gang;
626 	struct mes_add_queue_input queue_input;
627 	unsigned long flags;
628 	int r;
629 
630 	memset(&queue_input, 0, sizeof(struct mes_add_queue_input));
631 
632 	/* allocate the mes queue buffer */
633 	queue = kzalloc(sizeof(struct amdgpu_mes_queue), GFP_KERNEL);
634 	if (!queue) {
635 		DRM_ERROR("Failed to allocate memory for queue\n");
636 		return -ENOMEM;
637 	}
638 
639 	/* Allocate the queue mqd */
640 	r = amdgpu_mes_queue_alloc_mqd(adev, queue, qprops);
641 	if (r)
642 		goto clean_up_memory;
643 
644 	/*
645 	 * Avoid taking any other locks under MES lock to avoid circular
646 	 * lock dependencies.
647 	 */
648 	amdgpu_mes_lock(&adev->mes);
649 
650 	gang = idr_find(&adev->mes.gang_id_idr, gang_id);
651 	if (!gang) {
652 		DRM_ERROR("gang id %d doesn't exist\n", gang_id);
653 		r = -EINVAL;
654 		goto clean_up_mqd;
655 	}
656 
657 	/* add the mes gang to idr list */
658 	spin_lock_irqsave(&adev->mes.queue_id_lock, flags);
659 	r = idr_alloc(&adev->mes.queue_id_idr, queue, 1, 0,
660 		      GFP_ATOMIC);
661 	if (r < 0) {
662 		spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
663 		goto clean_up_mqd;
664 	}
665 	spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
666 	*queue_id = queue->queue_id = r;
667 
668 	/* allocate a doorbell index for the queue */
669 	r = amdgpu_mes_kernel_doorbell_get(adev,
670 					  qprops->queue_type,
671 					  &qprops->doorbell_off);
672 	if (r)
673 		goto clean_up_queue_id;
674 
675 	/* initialize the queue mqd */
676 	amdgpu_mes_queue_init_mqd(adev, queue, qprops);
677 
678 	/* add hw queue to mes */
679 	queue_input.process_id = gang->process->pasid;
680 
681 	queue_input.page_table_base_addr =
682 		adev->vm_manager.vram_base_offset + gang->process->pd_gpu_addr -
683 		adev->gmc.vram_start;
684 
685 	queue_input.process_va_start = 0;
686 	queue_input.process_va_end =
687 		(adev->vm_manager.max_pfn - 1) << AMDGPU_GPU_PAGE_SHIFT;
688 	queue_input.process_quantum = gang->process->process_quantum;
689 	queue_input.process_context_addr = gang->process->proc_ctx_gpu_addr;
690 	queue_input.gang_quantum = gang->gang_quantum;
691 	queue_input.gang_context_addr = gang->gang_ctx_gpu_addr;
692 	queue_input.inprocess_gang_priority = gang->inprocess_gang_priority;
693 	queue_input.gang_global_priority_level = gang->global_priority_level;
694 	queue_input.doorbell_offset = qprops->doorbell_off;
695 	queue_input.mqd_addr = queue->mqd_gpu_addr;
696 	queue_input.wptr_addr = qprops->wptr_gpu_addr;
697 	queue_input.wptr_mc_addr = qprops->wptr_mc_addr;
698 	queue_input.queue_type = qprops->queue_type;
699 	queue_input.paging = qprops->paging;
700 	queue_input.is_kfd_process = 0;
701 
702 	r = adev->mes.funcs->add_hw_queue(&adev->mes, &queue_input);
703 	if (r) {
704 		DRM_ERROR("failed to add hardware queue to MES, doorbell=0x%llx\n",
705 			  qprops->doorbell_off);
706 		goto clean_up_doorbell;
707 	}
708 
709 	DRM_DEBUG("MES hw queue was added, pasid=%d, gang id=%d, "
710 		  "queue type=%d, doorbell=0x%llx\n",
711 		  gang->process->pasid, gang_id, qprops->queue_type,
712 		  qprops->doorbell_off);
713 
714 	queue->ring = qprops->ring;
715 	queue->doorbell_off = qprops->doorbell_off;
716 	queue->wptr_gpu_addr = qprops->wptr_gpu_addr;
717 	queue->queue_type = qprops->queue_type;
718 	queue->paging = qprops->paging;
719 	queue->gang = gang;
720 	queue->ring->mqd_ptr = queue->mqd_cpu_ptr;
721 	list_add_tail(&queue->list, &gang->queue_list);
722 
723 	amdgpu_mes_unlock(&adev->mes);
724 	return 0;
725 
726 clean_up_doorbell:
727 	amdgpu_mes_kernel_doorbell_free(adev, qprops->doorbell_off);
728 clean_up_queue_id:
729 	spin_lock_irqsave(&adev->mes.queue_id_lock, flags);
730 	idr_remove(&adev->mes.queue_id_idr, queue->queue_id);
731 	spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
732 clean_up_mqd:
733 	amdgpu_mes_unlock(&adev->mes);
734 	amdgpu_mes_queue_free_mqd(queue);
735 clean_up_memory:
736 	kfree(queue);
737 	return r;
738 }
739 
740 int amdgpu_mes_remove_hw_queue(struct amdgpu_device *adev, int queue_id)
741 {
742 	unsigned long flags;
743 	struct amdgpu_mes_queue *queue;
744 	struct amdgpu_mes_gang *gang;
745 	struct mes_remove_queue_input queue_input;
746 	int r;
747 
748 	/*
749 	 * Avoid taking any other locks under MES lock to avoid circular
750 	 * lock dependencies.
751 	 */
752 	amdgpu_mes_lock(&adev->mes);
753 
754 	/* remove the mes gang from idr list */
755 	spin_lock_irqsave(&adev->mes.queue_id_lock, flags);
756 
757 	queue = idr_find(&adev->mes.queue_id_idr, queue_id);
758 	if (!queue) {
759 		spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
760 		amdgpu_mes_unlock(&adev->mes);
761 		DRM_ERROR("queue id %d doesn't exist\n", queue_id);
762 		return -EINVAL;
763 	}
764 
765 	idr_remove(&adev->mes.queue_id_idr, queue_id);
766 	spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
767 
768 	DRM_DEBUG("try to remove queue, doorbell off = 0x%llx\n",
769 		  queue->doorbell_off);
770 
771 	gang = queue->gang;
772 	queue_input.doorbell_offset = queue->doorbell_off;
773 	queue_input.gang_context_addr = gang->gang_ctx_gpu_addr;
774 
775 	r = adev->mes.funcs->remove_hw_queue(&adev->mes, &queue_input);
776 	if (r)
777 		DRM_ERROR("failed to remove hardware queue, queue id = %d\n",
778 			  queue_id);
779 
780 	list_del(&queue->list);
781 	amdgpu_mes_kernel_doorbell_free(adev, queue->doorbell_off);
782 	amdgpu_mes_unlock(&adev->mes);
783 
784 	amdgpu_mes_queue_free_mqd(queue);
785 	kfree(queue);
786 	return 0;
787 }
788 
789 int amdgpu_mes_map_legacy_queue(struct amdgpu_device *adev,
790 				struct amdgpu_ring *ring)
791 {
792 	struct mes_map_legacy_queue_input queue_input;
793 	int r;
794 
795 	memset(&queue_input, 0, sizeof(queue_input));
796 
797 	queue_input.queue_type = ring->funcs->type;
798 	queue_input.doorbell_offset = ring->doorbell_index;
799 	queue_input.pipe_id = ring->pipe;
800 	queue_input.queue_id = ring->queue;
801 	queue_input.mqd_addr = amdgpu_bo_gpu_offset(ring->mqd_obj);
802 	queue_input.wptr_addr = ring->wptr_gpu_addr;
803 
804 	r = adev->mes.funcs->map_legacy_queue(&adev->mes, &queue_input);
805 	if (r)
806 		DRM_ERROR("failed to map legacy queue\n");
807 
808 	return r;
809 }
810 
811 int amdgpu_mes_unmap_legacy_queue(struct amdgpu_device *adev,
812 				  struct amdgpu_ring *ring,
813 				  enum amdgpu_unmap_queues_action action,
814 				  u64 gpu_addr, u64 seq)
815 {
816 	struct mes_unmap_legacy_queue_input queue_input;
817 	int r;
818 
819 	queue_input.action = action;
820 	queue_input.queue_type = ring->funcs->type;
821 	queue_input.doorbell_offset = ring->doorbell_index;
822 	queue_input.pipe_id = ring->pipe;
823 	queue_input.queue_id = ring->queue;
824 	queue_input.trail_fence_addr = gpu_addr;
825 	queue_input.trail_fence_data = seq;
826 
827 	r = adev->mes.funcs->unmap_legacy_queue(&adev->mes, &queue_input);
828 	if (r)
829 		DRM_ERROR("failed to unmap legacy queue\n");
830 
831 	return r;
832 }
833 
834 uint32_t amdgpu_mes_rreg(struct amdgpu_device *adev, uint32_t reg)
835 {
836 	struct mes_misc_op_input op_input;
837 	int r, val = 0;
838 
839 	op_input.op = MES_MISC_OP_READ_REG;
840 	op_input.read_reg.reg_offset = reg;
841 	op_input.read_reg.buffer_addr = adev->mes.read_val_gpu_addr;
842 
843 	if (!adev->mes.funcs->misc_op) {
844 		DRM_ERROR("mes rreg is not supported!\n");
845 		goto error;
846 	}
847 
848 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
849 	if (r)
850 		DRM_ERROR("failed to read reg (0x%x)\n", reg);
851 	else
852 		val = *(adev->mes.read_val_ptr);
853 
854 error:
855 	return val;
856 }
857 
858 int amdgpu_mes_wreg(struct amdgpu_device *adev,
859 		    uint32_t reg, uint32_t val)
860 {
861 	struct mes_misc_op_input op_input;
862 	int r;
863 
864 	op_input.op = MES_MISC_OP_WRITE_REG;
865 	op_input.write_reg.reg_offset = reg;
866 	op_input.write_reg.reg_value = val;
867 
868 	if (!adev->mes.funcs->misc_op) {
869 		DRM_ERROR("mes wreg is not supported!\n");
870 		r = -EINVAL;
871 		goto error;
872 	}
873 
874 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
875 	if (r)
876 		DRM_ERROR("failed to write reg (0x%x)\n", reg);
877 
878 error:
879 	return r;
880 }
881 
882 int amdgpu_mes_reg_write_reg_wait(struct amdgpu_device *adev,
883 				  uint32_t reg0, uint32_t reg1,
884 				  uint32_t ref, uint32_t mask)
885 {
886 	struct mes_misc_op_input op_input;
887 	int r;
888 
889 	op_input.op = MES_MISC_OP_WRM_REG_WR_WAIT;
890 	op_input.wrm_reg.reg0 = reg0;
891 	op_input.wrm_reg.reg1 = reg1;
892 	op_input.wrm_reg.ref = ref;
893 	op_input.wrm_reg.mask = mask;
894 
895 	if (!adev->mes.funcs->misc_op) {
896 		DRM_ERROR("mes reg_write_reg_wait is not supported!\n");
897 		r = -EINVAL;
898 		goto error;
899 	}
900 
901 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
902 	if (r)
903 		DRM_ERROR("failed to reg_write_reg_wait\n");
904 
905 error:
906 	return r;
907 }
908 
909 int amdgpu_mes_reg_wait(struct amdgpu_device *adev, uint32_t reg,
910 			uint32_t val, uint32_t mask)
911 {
912 	struct mes_misc_op_input op_input;
913 	int r;
914 
915 	op_input.op = MES_MISC_OP_WRM_REG_WAIT;
916 	op_input.wrm_reg.reg0 = reg;
917 	op_input.wrm_reg.ref = val;
918 	op_input.wrm_reg.mask = mask;
919 
920 	if (!adev->mes.funcs->misc_op) {
921 		DRM_ERROR("mes reg wait is not supported!\n");
922 		r = -EINVAL;
923 		goto error;
924 	}
925 
926 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
927 	if (r)
928 		DRM_ERROR("failed to reg_write_reg_wait\n");
929 
930 error:
931 	return r;
932 }
933 
934 int amdgpu_mes_set_shader_debugger(struct amdgpu_device *adev,
935 				uint64_t process_context_addr,
936 				uint32_t spi_gdbg_per_vmid_cntl,
937 				const uint32_t *tcp_watch_cntl,
938 				uint32_t flags,
939 				bool trap_en)
940 {
941 	struct mes_misc_op_input op_input = {0};
942 	int r;
943 
944 	if (!adev->mes.funcs->misc_op) {
945 		DRM_ERROR("mes set shader debugger is not supported!\n");
946 		return -EINVAL;
947 	}
948 
949 	op_input.op = MES_MISC_OP_SET_SHADER_DEBUGGER;
950 	op_input.set_shader_debugger.process_context_addr = process_context_addr;
951 	op_input.set_shader_debugger.flags.u32all = flags;
952 
953 	/* use amdgpu mes_flush_shader_debugger instead */
954 	if (op_input.set_shader_debugger.flags.process_ctx_flush)
955 		return -EINVAL;
956 
957 	op_input.set_shader_debugger.spi_gdbg_per_vmid_cntl = spi_gdbg_per_vmid_cntl;
958 	memcpy(op_input.set_shader_debugger.tcp_watch_cntl, tcp_watch_cntl,
959 			sizeof(op_input.set_shader_debugger.tcp_watch_cntl));
960 
961 	if (((adev->mes.sched_version & AMDGPU_MES_API_VERSION_MASK) >>
962 			AMDGPU_MES_API_VERSION_SHIFT) >= 14)
963 		op_input.set_shader_debugger.trap_en = trap_en;
964 
965 	amdgpu_mes_lock(&adev->mes);
966 
967 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
968 	if (r)
969 		DRM_ERROR("failed to set_shader_debugger\n");
970 
971 	amdgpu_mes_unlock(&adev->mes);
972 
973 	return r;
974 }
975 
976 int amdgpu_mes_flush_shader_debugger(struct amdgpu_device *adev,
977 				     uint64_t process_context_addr)
978 {
979 	struct mes_misc_op_input op_input = {0};
980 	int r;
981 
982 	if (!adev->mes.funcs->misc_op) {
983 		DRM_ERROR("mes flush shader debugger is not supported!\n");
984 		return -EINVAL;
985 	}
986 
987 	op_input.op = MES_MISC_OP_SET_SHADER_DEBUGGER;
988 	op_input.set_shader_debugger.process_context_addr = process_context_addr;
989 	op_input.set_shader_debugger.flags.process_ctx_flush = true;
990 
991 	amdgpu_mes_lock(&adev->mes);
992 
993 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
994 	if (r)
995 		DRM_ERROR("failed to set_shader_debugger\n");
996 
997 	amdgpu_mes_unlock(&adev->mes);
998 
999 	return r;
1000 }
1001 
1002 static void
1003 amdgpu_mes_ring_to_queue_props(struct amdgpu_device *adev,
1004 			       struct amdgpu_ring *ring,
1005 			       struct amdgpu_mes_queue_properties *props)
1006 {
1007 	props->queue_type = ring->funcs->type;
1008 	props->hqd_base_gpu_addr = ring->gpu_addr;
1009 	props->rptr_gpu_addr = ring->rptr_gpu_addr;
1010 	props->wptr_gpu_addr = ring->wptr_gpu_addr;
1011 	props->wptr_mc_addr =
1012 		ring->mes_ctx->meta_data_mc_addr + ring->wptr_offs;
1013 	props->queue_size = ring->ring_size;
1014 	props->eop_gpu_addr = ring->eop_gpu_addr;
1015 	props->hqd_pipe_priority = AMDGPU_GFX_PIPE_PRIO_NORMAL;
1016 	props->hqd_queue_priority = AMDGPU_GFX_QUEUE_PRIORITY_MINIMUM;
1017 	props->paging = false;
1018 	props->ring = ring;
1019 }
1020 
1021 #define DEFINE_AMDGPU_MES_CTX_GET_OFFS_ENG(_eng)			\
1022 do {									\
1023        if (id_offs < AMDGPU_MES_CTX_MAX_OFFS)				\
1024 		return offsetof(struct amdgpu_mes_ctx_meta_data,	\
1025 				_eng[ring->idx].slots[id_offs]);        \
1026        else if (id_offs == AMDGPU_MES_CTX_RING_OFFS)			\
1027 		return offsetof(struct amdgpu_mes_ctx_meta_data,        \
1028 				_eng[ring->idx].ring);                  \
1029        else if (id_offs == AMDGPU_MES_CTX_IB_OFFS)			\
1030 		return offsetof(struct amdgpu_mes_ctx_meta_data,        \
1031 				_eng[ring->idx].ib);                    \
1032        else if (id_offs == AMDGPU_MES_CTX_PADDING_OFFS)			\
1033 		return offsetof(struct amdgpu_mes_ctx_meta_data,        \
1034 				_eng[ring->idx].padding);               \
1035 } while(0)
1036 
1037 int amdgpu_mes_ctx_get_offs(struct amdgpu_ring *ring, unsigned int id_offs)
1038 {
1039 	switch (ring->funcs->type) {
1040 	case AMDGPU_RING_TYPE_GFX:
1041 		DEFINE_AMDGPU_MES_CTX_GET_OFFS_ENG(gfx);
1042 		break;
1043 	case AMDGPU_RING_TYPE_COMPUTE:
1044 		DEFINE_AMDGPU_MES_CTX_GET_OFFS_ENG(compute);
1045 		break;
1046 	case AMDGPU_RING_TYPE_SDMA:
1047 		DEFINE_AMDGPU_MES_CTX_GET_OFFS_ENG(sdma);
1048 		break;
1049 	default:
1050 		break;
1051 	}
1052 
1053 	WARN_ON(1);
1054 	return -EINVAL;
1055 }
1056 
1057 int amdgpu_mes_add_ring(struct amdgpu_device *adev, int gang_id,
1058 			int queue_type, int idx,
1059 			struct amdgpu_mes_ctx_data *ctx_data,
1060 			struct amdgpu_ring **out)
1061 {
1062 	struct amdgpu_ring *ring;
1063 	struct amdgpu_mes_gang *gang;
1064 	struct amdgpu_mes_queue_properties qprops = {0};
1065 	int r, queue_id, pasid;
1066 
1067 	/*
1068 	 * Avoid taking any other locks under MES lock to avoid circular
1069 	 * lock dependencies.
1070 	 */
1071 	amdgpu_mes_lock(&adev->mes);
1072 	gang = idr_find(&adev->mes.gang_id_idr, gang_id);
1073 	if (!gang) {
1074 		DRM_ERROR("gang id %d doesn't exist\n", gang_id);
1075 		amdgpu_mes_unlock(&adev->mes);
1076 		return -EINVAL;
1077 	}
1078 	pasid = gang->process->pasid;
1079 
1080 	ring = kzalloc(sizeof(struct amdgpu_ring), GFP_KERNEL);
1081 	if (!ring) {
1082 		amdgpu_mes_unlock(&adev->mes);
1083 		return -ENOMEM;
1084 	}
1085 
1086 	ring->ring_obj = NULL;
1087 	ring->use_doorbell = true;
1088 	ring->is_mes_queue = true;
1089 	ring->mes_ctx = ctx_data;
1090 	ring->idx = idx;
1091 	ring->no_scheduler = true;
1092 
1093 	if (queue_type == AMDGPU_RING_TYPE_COMPUTE) {
1094 		int offset = offsetof(struct amdgpu_mes_ctx_meta_data,
1095 				      compute[ring->idx].mec_hpd);
1096 		ring->eop_gpu_addr =
1097 			amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
1098 	}
1099 
1100 	switch (queue_type) {
1101 	case AMDGPU_RING_TYPE_GFX:
1102 		ring->funcs = adev->gfx.gfx_ring[0].funcs;
1103 		ring->me = adev->gfx.gfx_ring[0].me;
1104 		ring->pipe = adev->gfx.gfx_ring[0].pipe;
1105 		break;
1106 	case AMDGPU_RING_TYPE_COMPUTE:
1107 		ring->funcs = adev->gfx.compute_ring[0].funcs;
1108 		ring->me = adev->gfx.compute_ring[0].me;
1109 		ring->pipe = adev->gfx.compute_ring[0].pipe;
1110 		break;
1111 	case AMDGPU_RING_TYPE_SDMA:
1112 		ring->funcs = adev->sdma.instance[0].ring.funcs;
1113 		break;
1114 	default:
1115 		BUG();
1116 	}
1117 
1118 	r = amdgpu_ring_init(adev, ring, 1024, NULL, 0,
1119 			     AMDGPU_RING_PRIO_DEFAULT, NULL);
1120 	if (r)
1121 		goto clean_up_memory;
1122 
1123 	amdgpu_mes_ring_to_queue_props(adev, ring, &qprops);
1124 
1125 	dma_fence_wait(gang->process->vm->last_update, false);
1126 	dma_fence_wait(ctx_data->meta_data_va->last_pt_update, false);
1127 	amdgpu_mes_unlock(&adev->mes);
1128 
1129 	r = amdgpu_mes_add_hw_queue(adev, gang_id, &qprops, &queue_id);
1130 	if (r)
1131 		goto clean_up_ring;
1132 
1133 	ring->hw_queue_id = queue_id;
1134 	ring->doorbell_index = qprops.doorbell_off;
1135 
1136 	if (queue_type == AMDGPU_RING_TYPE_GFX)
1137 		sprintf(ring->name, "gfx_%d.%d.%d", pasid, gang_id, queue_id);
1138 	else if (queue_type == AMDGPU_RING_TYPE_COMPUTE)
1139 		sprintf(ring->name, "compute_%d.%d.%d", pasid, gang_id,
1140 			queue_id);
1141 	else if (queue_type == AMDGPU_RING_TYPE_SDMA)
1142 		sprintf(ring->name, "sdma_%d.%d.%d", pasid, gang_id,
1143 			queue_id);
1144 	else
1145 		BUG();
1146 
1147 	*out = ring;
1148 	return 0;
1149 
1150 clean_up_ring:
1151 	amdgpu_ring_fini(ring);
1152 clean_up_memory:
1153 	kfree(ring);
1154 	amdgpu_mes_unlock(&adev->mes);
1155 	return r;
1156 }
1157 
1158 void amdgpu_mes_remove_ring(struct amdgpu_device *adev,
1159 			    struct amdgpu_ring *ring)
1160 {
1161 	if (!ring)
1162 		return;
1163 
1164 	amdgpu_mes_remove_hw_queue(adev, ring->hw_queue_id);
1165 	del_timer_sync(&ring->fence_drv.fallback_timer);
1166 	amdgpu_ring_fini(ring);
1167 	kfree(ring);
1168 }
1169 
1170 uint32_t amdgpu_mes_get_aggregated_doorbell_index(struct amdgpu_device *adev,
1171 						   enum amdgpu_mes_priority_level prio)
1172 {
1173 	return adev->mes.aggregated_doorbells[prio];
1174 }
1175 
1176 int amdgpu_mes_ctx_alloc_meta_data(struct amdgpu_device *adev,
1177 				   struct amdgpu_mes_ctx_data *ctx_data)
1178 {
1179 	int r;
1180 
1181 	r = amdgpu_bo_create_kernel(adev,
1182 			    sizeof(struct amdgpu_mes_ctx_meta_data),
1183 			    PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
1184 			    &ctx_data->meta_data_obj,
1185 			    &ctx_data->meta_data_mc_addr,
1186 			    &ctx_data->meta_data_ptr);
1187 	if (r) {
1188 		dev_warn(adev->dev, "(%d) create CTX bo failed\n", r);
1189 		return r;
1190 	}
1191 
1192 	if (!ctx_data->meta_data_obj)
1193 		return -ENOMEM;
1194 
1195 	memset(ctx_data->meta_data_ptr, 0,
1196 	       sizeof(struct amdgpu_mes_ctx_meta_data));
1197 
1198 	return 0;
1199 }
1200 
1201 void amdgpu_mes_ctx_free_meta_data(struct amdgpu_mes_ctx_data *ctx_data)
1202 {
1203 	if (ctx_data->meta_data_obj)
1204 		amdgpu_bo_free_kernel(&ctx_data->meta_data_obj,
1205 				      &ctx_data->meta_data_mc_addr,
1206 				      &ctx_data->meta_data_ptr);
1207 }
1208 
1209 int amdgpu_mes_ctx_map_meta_data(struct amdgpu_device *adev,
1210 				 struct amdgpu_vm *vm,
1211 				 struct amdgpu_mes_ctx_data *ctx_data)
1212 {
1213 	struct amdgpu_bo_va *bo_va;
1214 	struct amdgpu_sync sync;
1215 	struct drm_exec exec;
1216 	int r;
1217 
1218 	amdgpu_sync_create(&sync);
1219 
1220 	drm_exec_init(&exec, 0, 0);
1221 	drm_exec_until_all_locked(&exec) {
1222 		r = drm_exec_lock_obj(&exec,
1223 				      &ctx_data->meta_data_obj->tbo.base);
1224 		drm_exec_retry_on_contention(&exec);
1225 		if (unlikely(r))
1226 			goto error_fini_exec;
1227 
1228 		r = amdgpu_vm_lock_pd(vm, &exec, 0);
1229 		drm_exec_retry_on_contention(&exec);
1230 		if (unlikely(r))
1231 			goto error_fini_exec;
1232 	}
1233 
1234 	bo_va = amdgpu_vm_bo_add(adev, vm, ctx_data->meta_data_obj);
1235 	if (!bo_va) {
1236 		DRM_ERROR("failed to create bo_va for meta data BO\n");
1237 		r = -ENOMEM;
1238 		goto error_fini_exec;
1239 	}
1240 
1241 	r = amdgpu_vm_bo_map(adev, bo_va, ctx_data->meta_data_gpu_addr, 0,
1242 			     sizeof(struct amdgpu_mes_ctx_meta_data),
1243 			     AMDGPU_PTE_READABLE | AMDGPU_PTE_WRITEABLE |
1244 			     AMDGPU_PTE_EXECUTABLE);
1245 
1246 	if (r) {
1247 		DRM_ERROR("failed to do bo_map on meta data, err=%d\n", r);
1248 		goto error_del_bo_va;
1249 	}
1250 
1251 	r = amdgpu_vm_bo_update(adev, bo_va, false);
1252 	if (r) {
1253 		DRM_ERROR("failed to do vm_bo_update on meta data\n");
1254 		goto error_del_bo_va;
1255 	}
1256 	amdgpu_sync_fence(&sync, bo_va->last_pt_update);
1257 
1258 	r = amdgpu_vm_update_pdes(adev, vm, false);
1259 	if (r) {
1260 		DRM_ERROR("failed to update pdes on meta data\n");
1261 		goto error_del_bo_va;
1262 	}
1263 	amdgpu_sync_fence(&sync, vm->last_update);
1264 
1265 	amdgpu_sync_wait(&sync, false);
1266 	drm_exec_fini(&exec);
1267 
1268 	amdgpu_sync_free(&sync);
1269 	ctx_data->meta_data_va = bo_va;
1270 	return 0;
1271 
1272 error_del_bo_va:
1273 	amdgpu_vm_bo_del(adev, bo_va);
1274 
1275 error_fini_exec:
1276 	drm_exec_fini(&exec);
1277 	amdgpu_sync_free(&sync);
1278 	return r;
1279 }
1280 
1281 int amdgpu_mes_ctx_unmap_meta_data(struct amdgpu_device *adev,
1282 				   struct amdgpu_mes_ctx_data *ctx_data)
1283 {
1284 	struct amdgpu_bo_va *bo_va = ctx_data->meta_data_va;
1285 	struct amdgpu_bo *bo = ctx_data->meta_data_obj;
1286 	struct amdgpu_vm *vm = bo_va->base.vm;
1287 	struct dma_fence *fence;
1288 	struct drm_exec exec;
1289 	long r;
1290 
1291 	drm_exec_init(&exec, 0, 0);
1292 	drm_exec_until_all_locked(&exec) {
1293 		r = drm_exec_lock_obj(&exec,
1294 				      &ctx_data->meta_data_obj->tbo.base);
1295 		drm_exec_retry_on_contention(&exec);
1296 		if (unlikely(r))
1297 			goto out_unlock;
1298 
1299 		r = amdgpu_vm_lock_pd(vm, &exec, 0);
1300 		drm_exec_retry_on_contention(&exec);
1301 		if (unlikely(r))
1302 			goto out_unlock;
1303 	}
1304 
1305 	amdgpu_vm_bo_del(adev, bo_va);
1306 	if (!amdgpu_vm_ready(vm))
1307 		goto out_unlock;
1308 
1309 	r = dma_resv_get_singleton(bo->tbo.base.resv, DMA_RESV_USAGE_BOOKKEEP,
1310 				   &fence);
1311 	if (r)
1312 		goto out_unlock;
1313 	if (fence) {
1314 		amdgpu_bo_fence(bo, fence, true);
1315 		fence = NULL;
1316 	}
1317 
1318 	r = amdgpu_vm_clear_freed(adev, vm, &fence);
1319 	if (r || !fence)
1320 		goto out_unlock;
1321 
1322 	dma_fence_wait(fence, false);
1323 	amdgpu_bo_fence(bo, fence, true);
1324 	dma_fence_put(fence);
1325 
1326 out_unlock:
1327 	if (unlikely(r < 0))
1328 		dev_err(adev->dev, "failed to clear page tables (%ld)\n", r);
1329 	drm_exec_fini(&exec);
1330 
1331 	return r;
1332 }
1333 
1334 static int amdgpu_mes_test_create_gang_and_queues(struct amdgpu_device *adev,
1335 					  int pasid, int *gang_id,
1336 					  int queue_type, int num_queue,
1337 					  struct amdgpu_ring **added_rings,
1338 					  struct amdgpu_mes_ctx_data *ctx_data)
1339 {
1340 	struct amdgpu_ring *ring;
1341 	struct amdgpu_mes_gang_properties gprops = {0};
1342 	int r, j;
1343 
1344 	/* create a gang for the process */
1345 	gprops.priority = AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
1346 	gprops.gang_quantum = adev->mes.default_gang_quantum;
1347 	gprops.inprocess_gang_priority = AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
1348 	gprops.priority_level = AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
1349 	gprops.global_priority_level = AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
1350 
1351 	r = amdgpu_mes_add_gang(adev, pasid, &gprops, gang_id);
1352 	if (r) {
1353 		DRM_ERROR("failed to add gang\n");
1354 		return r;
1355 	}
1356 
1357 	/* create queues for the gang */
1358 	for (j = 0; j < num_queue; j++) {
1359 		r = amdgpu_mes_add_ring(adev, *gang_id, queue_type, j,
1360 					ctx_data, &ring);
1361 		if (r) {
1362 			DRM_ERROR("failed to add ring\n");
1363 			break;
1364 		}
1365 
1366 		DRM_INFO("ring %s was added\n", ring->name);
1367 		added_rings[j] = ring;
1368 	}
1369 
1370 	return 0;
1371 }
1372 
1373 static int amdgpu_mes_test_queues(struct amdgpu_ring **added_rings)
1374 {
1375 	struct amdgpu_ring *ring;
1376 	int i, r;
1377 
1378 	for (i = 0; i < AMDGPU_MES_CTX_MAX_RINGS; i++) {
1379 		ring = added_rings[i];
1380 		if (!ring)
1381 			continue;
1382 
1383 		r = amdgpu_ring_test_helper(ring);
1384 		if (r)
1385 			return r;
1386 
1387 		r = amdgpu_ring_test_ib(ring, 1000 * 10);
1388 		if (r) {
1389 			DRM_DEV_ERROR(ring->adev->dev,
1390 				      "ring %s ib test failed (%d)\n",
1391 				      ring->name, r);
1392 			return r;
1393 		} else
1394 			DRM_INFO("ring %s ib test pass\n", ring->name);
1395 	}
1396 
1397 	return 0;
1398 }
1399 
1400 int amdgpu_mes_self_test(struct amdgpu_device *adev)
1401 {
1402 	struct amdgpu_vm *vm = NULL;
1403 	struct amdgpu_mes_ctx_data ctx_data = {0};
1404 	struct amdgpu_ring *added_rings[AMDGPU_MES_CTX_MAX_RINGS] = { NULL };
1405 	int gang_ids[3] = {0};
1406 	int queue_types[][2] = { { AMDGPU_RING_TYPE_GFX, 1 },
1407 				 { AMDGPU_RING_TYPE_COMPUTE, 1 },
1408 				 { AMDGPU_RING_TYPE_SDMA, 1} };
1409 	int i, r, pasid, k = 0;
1410 
1411 	pasid = amdgpu_pasid_alloc(16);
1412 	if (pasid < 0) {
1413 		dev_warn(adev->dev, "No more PASIDs available!");
1414 		pasid = 0;
1415 	}
1416 
1417 	vm = kzalloc(sizeof(*vm), GFP_KERNEL);
1418 	if (!vm) {
1419 		r = -ENOMEM;
1420 		goto error_pasid;
1421 	}
1422 
1423 	r = amdgpu_vm_init(adev, vm, -1);
1424 	if (r) {
1425 		DRM_ERROR("failed to initialize vm\n");
1426 		goto error_pasid;
1427 	}
1428 
1429 	r = amdgpu_mes_ctx_alloc_meta_data(adev, &ctx_data);
1430 	if (r) {
1431 		DRM_ERROR("failed to alloc ctx meta data\n");
1432 		goto error_fini;
1433 	}
1434 
1435 	ctx_data.meta_data_gpu_addr = AMDGPU_VA_RESERVED_BOTTOM;
1436 	r = amdgpu_mes_ctx_map_meta_data(adev, vm, &ctx_data);
1437 	if (r) {
1438 		DRM_ERROR("failed to map ctx meta data\n");
1439 		goto error_vm;
1440 	}
1441 
1442 	r = amdgpu_mes_create_process(adev, pasid, vm);
1443 	if (r) {
1444 		DRM_ERROR("failed to create MES process\n");
1445 		goto error_vm;
1446 	}
1447 
1448 	for (i = 0; i < ARRAY_SIZE(queue_types); i++) {
1449 		/* On GFX v10.3, fw hasn't supported to map sdma queue. */
1450 		if (amdgpu_ip_version(adev, GC_HWIP, 0) >=
1451 			    IP_VERSION(10, 3, 0) &&
1452 		    amdgpu_ip_version(adev, GC_HWIP, 0) <
1453 			    IP_VERSION(11, 0, 0) &&
1454 		    queue_types[i][0] == AMDGPU_RING_TYPE_SDMA)
1455 			continue;
1456 
1457 		r = amdgpu_mes_test_create_gang_and_queues(adev, pasid,
1458 							   &gang_ids[i],
1459 							   queue_types[i][0],
1460 							   queue_types[i][1],
1461 							   &added_rings[k],
1462 							   &ctx_data);
1463 		if (r)
1464 			goto error_queues;
1465 
1466 		k += queue_types[i][1];
1467 	}
1468 
1469 	/* start ring test and ib test for MES queues */
1470 	amdgpu_mes_test_queues(added_rings);
1471 
1472 error_queues:
1473 	/* remove all queues */
1474 	for (i = 0; i < ARRAY_SIZE(added_rings); i++) {
1475 		if (!added_rings[i])
1476 			continue;
1477 		amdgpu_mes_remove_ring(adev, added_rings[i]);
1478 	}
1479 
1480 	for (i = 0; i < ARRAY_SIZE(gang_ids); i++) {
1481 		if (!gang_ids[i])
1482 			continue;
1483 		amdgpu_mes_remove_gang(adev, gang_ids[i]);
1484 	}
1485 
1486 	amdgpu_mes_destroy_process(adev, pasid);
1487 
1488 error_vm:
1489 	amdgpu_mes_ctx_unmap_meta_data(adev, &ctx_data);
1490 
1491 error_fini:
1492 	amdgpu_vm_fini(adev, vm);
1493 
1494 error_pasid:
1495 	if (pasid)
1496 		amdgpu_pasid_free(pasid);
1497 
1498 	amdgpu_mes_ctx_free_meta_data(&ctx_data);
1499 	kfree(vm);
1500 	return 0;
1501 }
1502 
1503 int amdgpu_mes_init_microcode(struct amdgpu_device *adev, int pipe)
1504 {
1505 	const struct mes_firmware_header_v1_0 *mes_hdr;
1506 	struct amdgpu_firmware_info *info;
1507 	char ucode_prefix[30];
1508 	char fw_name[50];
1509 	bool need_retry = false;
1510 	int r;
1511 
1512 	amdgpu_ucode_ip_version_decode(adev, GC_HWIP, ucode_prefix,
1513 				       sizeof(ucode_prefix));
1514 	if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(11, 0, 0)) {
1515 		snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes%s.bin",
1516 			 ucode_prefix,
1517 			 pipe == AMDGPU_MES_SCHED_PIPE ? "_2" : "1");
1518 		need_retry = true;
1519 	} else {
1520 		snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes%s.bin",
1521 			 ucode_prefix,
1522 			 pipe == AMDGPU_MES_SCHED_PIPE ? "" : "1");
1523 	}
1524 
1525 	r = amdgpu_ucode_request(adev, &adev->mes.fw[pipe], fw_name);
1526 	if (r && need_retry && pipe == AMDGPU_MES_SCHED_PIPE) {
1527 		snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes.bin",
1528 			 ucode_prefix);
1529 		DRM_INFO("try to fall back to %s\n", fw_name);
1530 		r = amdgpu_ucode_request(adev, &adev->mes.fw[pipe],
1531 					 fw_name);
1532 	}
1533 
1534 	if (r)
1535 		goto out;
1536 
1537 	mes_hdr = (const struct mes_firmware_header_v1_0 *)
1538 		adev->mes.fw[pipe]->data;
1539 	adev->mes.uc_start_addr[pipe] =
1540 		le32_to_cpu(mes_hdr->mes_uc_start_addr_lo) |
1541 		((uint64_t)(le32_to_cpu(mes_hdr->mes_uc_start_addr_hi)) << 32);
1542 	adev->mes.data_start_addr[pipe] =
1543 		le32_to_cpu(mes_hdr->mes_data_start_addr_lo) |
1544 		((uint64_t)(le32_to_cpu(mes_hdr->mes_data_start_addr_hi)) << 32);
1545 
1546 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
1547 		int ucode, ucode_data;
1548 
1549 		if (pipe == AMDGPU_MES_SCHED_PIPE) {
1550 			ucode = AMDGPU_UCODE_ID_CP_MES;
1551 			ucode_data = AMDGPU_UCODE_ID_CP_MES_DATA;
1552 		} else {
1553 			ucode = AMDGPU_UCODE_ID_CP_MES1;
1554 			ucode_data = AMDGPU_UCODE_ID_CP_MES1_DATA;
1555 		}
1556 
1557 		info = &adev->firmware.ucode[ucode];
1558 		info->ucode_id = ucode;
1559 		info->fw = adev->mes.fw[pipe];
1560 		adev->firmware.fw_size +=
1561 			ALIGN(le32_to_cpu(mes_hdr->mes_ucode_size_bytes),
1562 			      PAGE_SIZE);
1563 
1564 		info = &adev->firmware.ucode[ucode_data];
1565 		info->ucode_id = ucode_data;
1566 		info->fw = adev->mes.fw[pipe];
1567 		adev->firmware.fw_size +=
1568 			ALIGN(le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes),
1569 			      PAGE_SIZE);
1570 	}
1571 
1572 	return 0;
1573 out:
1574 	amdgpu_ucode_release(&adev->mes.fw[pipe]);
1575 	return r;
1576 }
1577 
1578 #if defined(CONFIG_DEBUG_FS)
1579 
1580 static int amdgpu_debugfs_mes_event_log_show(struct seq_file *m, void *unused)
1581 {
1582 	struct amdgpu_device *adev = m->private;
1583 	uint32_t *mem = (uint32_t *)(adev->mes.event_log_cpu_addr);
1584 
1585 	seq_hex_dump(m, "", DUMP_PREFIX_OFFSET, 32, 4,
1586 		     mem, AMDGPU_MES_LOG_BUFFER_SIZE, false);
1587 
1588 	return 0;
1589 }
1590 
1591 DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_mes_event_log);
1592 
1593 #endif
1594 
1595 void amdgpu_debugfs_mes_event_log_init(struct amdgpu_device *adev)
1596 {
1597 
1598 #if defined(CONFIG_DEBUG_FS)
1599 	struct drm_minor *minor = adev_to_drm(adev)->primary;
1600 	struct dentry *root = minor->debugfs_root;
1601 	if (adev->enable_mes && amdgpu_mes_log_enable)
1602 		debugfs_create_file("amdgpu_mes_event_log", 0444, root,
1603 				    adev, &amdgpu_debugfs_mes_event_log_fops);
1604 
1605 #endif
1606 }
1607