xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c (revision 173b0b5b0e865348684c02bd9cb1d22b5d46e458)
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 int amdgpu_mes_doorbell_process_slice(struct amdgpu_device *adev)
36 {
37 	return roundup(AMDGPU_ONE_DOORBELL_SIZE *
38 		       AMDGPU_MES_MAX_NUM_OF_QUEUES_PER_PROCESS,
39 		       PAGE_SIZE);
40 }
41 
42 static int amdgpu_mes_kernel_doorbell_get(struct amdgpu_device *adev,
43 					 struct amdgpu_mes_process *process,
44 					 int ip_type, uint64_t *doorbell_index)
45 {
46 	unsigned int offset, found;
47 	struct amdgpu_mes *mes = &adev->mes;
48 
49 	if (ip_type == AMDGPU_RING_TYPE_SDMA)
50 		offset = adev->doorbell_index.sdma_engine[0];
51 	else
52 		offset = 0;
53 
54 	found = find_next_zero_bit(mes->doorbell_bitmap, mes->num_mes_dbs, offset);
55 	if (found >= mes->num_mes_dbs) {
56 		DRM_WARN("No doorbell available\n");
57 		return -ENOSPC;
58 	}
59 
60 	set_bit(found, mes->doorbell_bitmap);
61 
62 	/* Get the absolute doorbell index on BAR */
63 	*doorbell_index = mes->db_start_dw_offset + found * 2;
64 	return 0;
65 }
66 
67 static void amdgpu_mes_kernel_doorbell_free(struct amdgpu_device *adev,
68 					   struct amdgpu_mes_process *process,
69 					   uint32_t doorbell_index)
70 {
71 	unsigned int old, rel_index;
72 	struct amdgpu_mes *mes = &adev->mes;
73 
74 	/* Find the relative index of the doorbell in this object */
75 	rel_index = (doorbell_index - mes->db_start_dw_offset) / 2;
76 	old = test_and_clear_bit(rel_index, mes->doorbell_bitmap);
77 	WARN_ON(!old);
78 }
79 
80 static int amdgpu_mes_doorbell_init(struct amdgpu_device *adev)
81 {
82 	int i;
83 	struct amdgpu_mes *mes = &adev->mes;
84 
85 	/* Bitmap for dynamic allocation of kernel doorbells */
86 	mes->doorbell_bitmap = bitmap_zalloc(PAGE_SIZE / sizeof(u32), GFP_KERNEL);
87 	if (!mes->doorbell_bitmap) {
88 		DRM_ERROR("Failed to allocate MES doorbell bitmap\n");
89 		return -ENOMEM;
90 	}
91 
92 	mes->num_mes_dbs = PAGE_SIZE / AMDGPU_ONE_DOORBELL_SIZE;
93 	for (i = 0; i < AMDGPU_MES_PRIORITY_NUM_LEVELS; i++) {
94 		adev->mes.aggregated_doorbells[i] = mes->db_start_dw_offset + i * 2;
95 		set_bit(i, mes->doorbell_bitmap);
96 	}
97 
98 	return 0;
99 }
100 
101 static int amdgpu_mes_event_log_init(struct amdgpu_device *adev)
102 {
103 	int r;
104 
105 	if (!amdgpu_mes_log_enable)
106 		return 0;
107 
108 	r = amdgpu_bo_create_kernel(adev, AMDGPU_MES_LOG_BUFFER_SIZE, PAGE_SIZE,
109 				    AMDGPU_GEM_DOMAIN_GTT,
110 				    &adev->mes.event_log_gpu_obj,
111 				    &adev->mes.event_log_gpu_addr,
112 				    &adev->mes.event_log_cpu_addr);
113 	if (r) {
114 		dev_warn(adev->dev, "failed to create MES event log buffer (%d)", r);
115 		return r;
116 	}
117 
118 	memset(adev->mes.event_log_cpu_addr, 0, PAGE_SIZE);
119 
120 	return  0;
121 
122 }
123 
124 static void amdgpu_mes_doorbell_free(struct amdgpu_device *adev)
125 {
126 	bitmap_free(adev->mes.doorbell_bitmap);
127 }
128 
129 int amdgpu_mes_init(struct amdgpu_device *adev)
130 {
131 	int i, r;
132 
133 	adev->mes.adev = adev;
134 
135 	idr_init(&adev->mes.pasid_idr);
136 	idr_init(&adev->mes.gang_id_idr);
137 	idr_init(&adev->mes.queue_id_idr);
138 	ida_init(&adev->mes.doorbell_ida);
139 	spin_lock_init(&adev->mes.queue_id_lock);
140 	spin_lock_init(&adev->mes.ring_lock);
141 	mutex_init(&adev->mes.mutex_hidden);
142 
143 	adev->mes.total_max_queue = AMDGPU_FENCE_MES_QUEUE_ID_MASK;
144 	adev->mes.vmid_mask_mmhub = 0xffffff00;
145 	adev->mes.vmid_mask_gfxhub = 0xffffff00;
146 
147 	for (i = 0; i < AMDGPU_MES_MAX_COMPUTE_PIPES; i++) {
148 		/* use only 1st MEC pipes */
149 		if (i >= 4)
150 			continue;
151 		adev->mes.compute_hqd_mask[i] = 0xc;
152 	}
153 
154 	for (i = 0; i < AMDGPU_MES_MAX_GFX_PIPES; i++)
155 		adev->mes.gfx_hqd_mask[i] = i ? 0 : 0xfffffffe;
156 
157 	for (i = 0; i < AMDGPU_MES_MAX_SDMA_PIPES; i++) {
158 		if (amdgpu_ip_version(adev, SDMA0_HWIP, 0) <
159 		    IP_VERSION(6, 0, 0))
160 			adev->mes.sdma_hqd_mask[i] = i ? 0 : 0x3fc;
161 		/* zero sdma_hqd_mask for non-existent engine */
162 		else if (adev->sdma.num_instances == 1)
163 			adev->mes.sdma_hqd_mask[i] = i ? 0 : 0xfc;
164 		else
165 			adev->mes.sdma_hqd_mask[i] = 0xfc;
166 	}
167 
168 	r = amdgpu_device_wb_get(adev, &adev->mes.sch_ctx_offs);
169 	if (r) {
170 		dev_err(adev->dev,
171 			"(%d) ring trail_fence_offs wb alloc failed\n", r);
172 		goto error_ids;
173 	}
174 	adev->mes.sch_ctx_gpu_addr =
175 		adev->wb.gpu_addr + (adev->mes.sch_ctx_offs * 4);
176 	adev->mes.sch_ctx_ptr =
177 		(uint64_t *)&adev->wb.wb[adev->mes.sch_ctx_offs];
178 
179 	r = amdgpu_device_wb_get(adev, &adev->mes.query_status_fence_offs);
180 	if (r) {
181 		amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
182 		dev_err(adev->dev,
183 			"(%d) query_status_fence_offs wb alloc failed\n", r);
184 		goto error_ids;
185 	}
186 	adev->mes.query_status_fence_gpu_addr =
187 		adev->wb.gpu_addr + (adev->mes.query_status_fence_offs * 4);
188 	adev->mes.query_status_fence_ptr =
189 		(uint64_t *)&adev->wb.wb[adev->mes.query_status_fence_offs];
190 
191 	r = amdgpu_device_wb_get(adev, &adev->mes.read_val_offs);
192 	if (r) {
193 		amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
194 		amdgpu_device_wb_free(adev, adev->mes.query_status_fence_offs);
195 		dev_err(adev->dev,
196 			"(%d) read_val_offs alloc failed\n", r);
197 		goto error_ids;
198 	}
199 	adev->mes.read_val_gpu_addr =
200 		adev->wb.gpu_addr + (adev->mes.read_val_offs * 4);
201 	adev->mes.read_val_ptr =
202 		(uint32_t *)&adev->wb.wb[adev->mes.read_val_offs];
203 
204 	r = amdgpu_mes_doorbell_init(adev);
205 	if (r)
206 		goto error;
207 
208 	r = amdgpu_mes_event_log_init(adev);
209 	if (r)
210 		goto error_doorbell;
211 
212 	return 0;
213 
214 error_doorbell:
215 	amdgpu_mes_doorbell_free(adev);
216 error:
217 	amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
218 	amdgpu_device_wb_free(adev, adev->mes.query_status_fence_offs);
219 	amdgpu_device_wb_free(adev, adev->mes.read_val_offs);
220 error_ids:
221 	idr_destroy(&adev->mes.pasid_idr);
222 	idr_destroy(&adev->mes.gang_id_idr);
223 	idr_destroy(&adev->mes.queue_id_idr);
224 	ida_destroy(&adev->mes.doorbell_ida);
225 	mutex_destroy(&adev->mes.mutex_hidden);
226 	return r;
227 }
228 
229 void amdgpu_mes_fini(struct amdgpu_device *adev)
230 {
231 	amdgpu_bo_free_kernel(&adev->mes.event_log_gpu_obj,
232 			      &adev->mes.event_log_gpu_addr,
233 			      &adev->mes.event_log_cpu_addr);
234 
235 	amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
236 	amdgpu_device_wb_free(adev, adev->mes.query_status_fence_offs);
237 	amdgpu_device_wb_free(adev, adev->mes.read_val_offs);
238 	amdgpu_mes_doorbell_free(adev);
239 
240 	idr_destroy(&adev->mes.pasid_idr);
241 	idr_destroy(&adev->mes.gang_id_idr);
242 	idr_destroy(&adev->mes.queue_id_idr);
243 	ida_destroy(&adev->mes.doorbell_ida);
244 	mutex_destroy(&adev->mes.mutex_hidden);
245 }
246 
247 static void amdgpu_mes_queue_free_mqd(struct amdgpu_mes_queue *q)
248 {
249 	amdgpu_bo_free_kernel(&q->mqd_obj,
250 			      &q->mqd_gpu_addr,
251 			      &q->mqd_cpu_ptr);
252 }
253 
254 int amdgpu_mes_create_process(struct amdgpu_device *adev, int pasid,
255 			      struct amdgpu_vm *vm)
256 {
257 	struct amdgpu_mes_process *process;
258 	int r;
259 
260 	/* allocate the mes process buffer */
261 	process = kzalloc(sizeof(struct amdgpu_mes_process), GFP_KERNEL);
262 	if (!process) {
263 		DRM_ERROR("no more memory to create mes process\n");
264 		return -ENOMEM;
265 	}
266 
267 	/* allocate the process context bo and map it */
268 	r = amdgpu_bo_create_kernel(adev, AMDGPU_MES_PROC_CTX_SIZE, PAGE_SIZE,
269 				    AMDGPU_GEM_DOMAIN_GTT,
270 				    &process->proc_ctx_bo,
271 				    &process->proc_ctx_gpu_addr,
272 				    &process->proc_ctx_cpu_ptr);
273 	if (r) {
274 		DRM_ERROR("failed to allocate process context bo\n");
275 		goto clean_up_memory;
276 	}
277 	memset(process->proc_ctx_cpu_ptr, 0, AMDGPU_MES_PROC_CTX_SIZE);
278 
279 	/*
280 	 * Avoid taking any other locks under MES lock to avoid circular
281 	 * lock dependencies.
282 	 */
283 	amdgpu_mes_lock(&adev->mes);
284 
285 	/* add the mes process to idr list */
286 	r = idr_alloc(&adev->mes.pasid_idr, process, pasid, pasid + 1,
287 		      GFP_KERNEL);
288 	if (r < 0) {
289 		DRM_ERROR("failed to lock pasid=%d\n", pasid);
290 		goto clean_up_ctx;
291 	}
292 
293 	INIT_LIST_HEAD(&process->gang_list);
294 	process->vm = vm;
295 	process->pasid = pasid;
296 	process->process_quantum = adev->mes.default_process_quantum;
297 	process->pd_gpu_addr = amdgpu_bo_gpu_offset(vm->root.bo);
298 
299 	amdgpu_mes_unlock(&adev->mes);
300 	return 0;
301 
302 clean_up_ctx:
303 	amdgpu_mes_unlock(&adev->mes);
304 	amdgpu_bo_free_kernel(&process->proc_ctx_bo,
305 			      &process->proc_ctx_gpu_addr,
306 			      &process->proc_ctx_cpu_ptr);
307 clean_up_memory:
308 	kfree(process);
309 	return r;
310 }
311 
312 void amdgpu_mes_destroy_process(struct amdgpu_device *adev, int pasid)
313 {
314 	struct amdgpu_mes_process *process;
315 	struct amdgpu_mes_gang *gang, *tmp1;
316 	struct amdgpu_mes_queue *queue, *tmp2;
317 	struct mes_remove_queue_input queue_input;
318 	unsigned long flags;
319 	int r;
320 
321 	/*
322 	 * Avoid taking any other locks under MES lock to avoid circular
323 	 * lock dependencies.
324 	 */
325 	amdgpu_mes_lock(&adev->mes);
326 
327 	process = idr_find(&adev->mes.pasid_idr, pasid);
328 	if (!process) {
329 		DRM_WARN("pasid %d doesn't exist\n", pasid);
330 		amdgpu_mes_unlock(&adev->mes);
331 		return;
332 	}
333 
334 	/* Remove all queues from hardware */
335 	list_for_each_entry_safe(gang, tmp1, &process->gang_list, list) {
336 		list_for_each_entry_safe(queue, tmp2, &gang->queue_list, list) {
337 			spin_lock_irqsave(&adev->mes.queue_id_lock, flags);
338 			idr_remove(&adev->mes.queue_id_idr, queue->queue_id);
339 			spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
340 
341 			queue_input.doorbell_offset = queue->doorbell_off;
342 			queue_input.gang_context_addr = gang->gang_ctx_gpu_addr;
343 
344 			r = adev->mes.funcs->remove_hw_queue(&adev->mes,
345 							     &queue_input);
346 			if (r)
347 				DRM_WARN("failed to remove hardware queue\n");
348 		}
349 
350 		idr_remove(&adev->mes.gang_id_idr, gang->gang_id);
351 	}
352 
353 	idr_remove(&adev->mes.pasid_idr, pasid);
354 	amdgpu_mes_unlock(&adev->mes);
355 
356 	/* free all memory allocated by the process */
357 	list_for_each_entry_safe(gang, tmp1, &process->gang_list, list) {
358 		/* free all queues in the gang */
359 		list_for_each_entry_safe(queue, tmp2, &gang->queue_list, list) {
360 			amdgpu_mes_queue_free_mqd(queue);
361 			list_del(&queue->list);
362 			kfree(queue);
363 		}
364 		amdgpu_bo_free_kernel(&gang->gang_ctx_bo,
365 				      &gang->gang_ctx_gpu_addr,
366 				      &gang->gang_ctx_cpu_ptr);
367 		list_del(&gang->list);
368 		kfree(gang);
369 
370 	}
371 	amdgpu_bo_free_kernel(&process->proc_ctx_bo,
372 			      &process->proc_ctx_gpu_addr,
373 			      &process->proc_ctx_cpu_ptr);
374 	kfree(process);
375 }
376 
377 int amdgpu_mes_add_gang(struct amdgpu_device *adev, int pasid,
378 			struct amdgpu_mes_gang_properties *gprops,
379 			int *gang_id)
380 {
381 	struct amdgpu_mes_process *process;
382 	struct amdgpu_mes_gang *gang;
383 	int r;
384 
385 	/* allocate the mes gang buffer */
386 	gang = kzalloc(sizeof(struct amdgpu_mes_gang), GFP_KERNEL);
387 	if (!gang) {
388 		return -ENOMEM;
389 	}
390 
391 	/* allocate the gang context bo and map it to cpu space */
392 	r = amdgpu_bo_create_kernel(adev, AMDGPU_MES_GANG_CTX_SIZE, PAGE_SIZE,
393 				    AMDGPU_GEM_DOMAIN_GTT,
394 				    &gang->gang_ctx_bo,
395 				    &gang->gang_ctx_gpu_addr,
396 				    &gang->gang_ctx_cpu_ptr);
397 	if (r) {
398 		DRM_ERROR("failed to allocate process context bo\n");
399 		goto clean_up_mem;
400 	}
401 	memset(gang->gang_ctx_cpu_ptr, 0, AMDGPU_MES_GANG_CTX_SIZE);
402 
403 	/*
404 	 * Avoid taking any other locks under MES lock to avoid circular
405 	 * lock dependencies.
406 	 */
407 	amdgpu_mes_lock(&adev->mes);
408 
409 	process = idr_find(&adev->mes.pasid_idr, pasid);
410 	if (!process) {
411 		DRM_ERROR("pasid %d doesn't exist\n", pasid);
412 		r = -EINVAL;
413 		goto clean_up_ctx;
414 	}
415 
416 	/* add the mes gang to idr list */
417 	r = idr_alloc(&adev->mes.gang_id_idr, gang, 1, 0,
418 		      GFP_KERNEL);
419 	if (r < 0) {
420 		DRM_ERROR("failed to allocate idr for gang\n");
421 		goto clean_up_ctx;
422 	}
423 
424 	gang->gang_id = r;
425 	*gang_id = r;
426 
427 	INIT_LIST_HEAD(&gang->queue_list);
428 	gang->process = process;
429 	gang->priority = gprops->priority;
430 	gang->gang_quantum = gprops->gang_quantum ?
431 		gprops->gang_quantum : adev->mes.default_gang_quantum;
432 	gang->global_priority_level = gprops->global_priority_level;
433 	gang->inprocess_gang_priority = gprops->inprocess_gang_priority;
434 	list_add_tail(&gang->list, &process->gang_list);
435 
436 	amdgpu_mes_unlock(&adev->mes);
437 	return 0;
438 
439 clean_up_ctx:
440 	amdgpu_mes_unlock(&adev->mes);
441 	amdgpu_bo_free_kernel(&gang->gang_ctx_bo,
442 			      &gang->gang_ctx_gpu_addr,
443 			      &gang->gang_ctx_cpu_ptr);
444 clean_up_mem:
445 	kfree(gang);
446 	return r;
447 }
448 
449 int amdgpu_mes_remove_gang(struct amdgpu_device *adev, int gang_id)
450 {
451 	struct amdgpu_mes_gang *gang;
452 
453 	/*
454 	 * Avoid taking any other locks under MES lock to avoid circular
455 	 * lock dependencies.
456 	 */
457 	amdgpu_mes_lock(&adev->mes);
458 
459 	gang = idr_find(&adev->mes.gang_id_idr, gang_id);
460 	if (!gang) {
461 		DRM_ERROR("gang id %d doesn't exist\n", gang_id);
462 		amdgpu_mes_unlock(&adev->mes);
463 		return -EINVAL;
464 	}
465 
466 	if (!list_empty(&gang->queue_list)) {
467 		DRM_ERROR("queue list is not empty\n");
468 		amdgpu_mes_unlock(&adev->mes);
469 		return -EBUSY;
470 	}
471 
472 	idr_remove(&adev->mes.gang_id_idr, gang->gang_id);
473 	list_del(&gang->list);
474 	amdgpu_mes_unlock(&adev->mes);
475 
476 	amdgpu_bo_free_kernel(&gang->gang_ctx_bo,
477 			      &gang->gang_ctx_gpu_addr,
478 			      &gang->gang_ctx_cpu_ptr);
479 
480 	kfree(gang);
481 
482 	return 0;
483 }
484 
485 int amdgpu_mes_suspend(struct amdgpu_device *adev)
486 {
487 	struct idr *idp;
488 	struct amdgpu_mes_process *process;
489 	struct amdgpu_mes_gang *gang;
490 	struct mes_suspend_gang_input input;
491 	int r, pasid;
492 
493 	/*
494 	 * Avoid taking any other locks under MES lock to avoid circular
495 	 * lock dependencies.
496 	 */
497 	amdgpu_mes_lock(&adev->mes);
498 
499 	idp = &adev->mes.pasid_idr;
500 
501 	idr_for_each_entry(idp, process, pasid) {
502 		list_for_each_entry(gang, &process->gang_list, list) {
503 			r = adev->mes.funcs->suspend_gang(&adev->mes, &input);
504 			if (r)
505 				DRM_ERROR("failed to suspend pasid %d gangid %d",
506 					 pasid, gang->gang_id);
507 		}
508 	}
509 
510 	amdgpu_mes_unlock(&adev->mes);
511 	return 0;
512 }
513 
514 int amdgpu_mes_resume(struct amdgpu_device *adev)
515 {
516 	struct idr *idp;
517 	struct amdgpu_mes_process *process;
518 	struct amdgpu_mes_gang *gang;
519 	struct mes_resume_gang_input input;
520 	int r, pasid;
521 
522 	/*
523 	 * Avoid taking any other locks under MES lock to avoid circular
524 	 * lock dependencies.
525 	 */
526 	amdgpu_mes_lock(&adev->mes);
527 
528 	idp = &adev->mes.pasid_idr;
529 
530 	idr_for_each_entry(idp, process, pasid) {
531 		list_for_each_entry(gang, &process->gang_list, list) {
532 			r = adev->mes.funcs->resume_gang(&adev->mes, &input);
533 			if (r)
534 				DRM_ERROR("failed to resume pasid %d gangid %d",
535 					 pasid, gang->gang_id);
536 		}
537 	}
538 
539 	amdgpu_mes_unlock(&adev->mes);
540 	return 0;
541 }
542 
543 static int amdgpu_mes_queue_alloc_mqd(struct amdgpu_device *adev,
544 				     struct amdgpu_mes_queue *q,
545 				     struct amdgpu_mes_queue_properties *p)
546 {
547 	struct amdgpu_mqd *mqd_mgr = &adev->mqds[p->queue_type];
548 	u32 mqd_size = mqd_mgr->mqd_size;
549 	int r;
550 
551 	r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
552 				    AMDGPU_GEM_DOMAIN_GTT,
553 				    &q->mqd_obj,
554 				    &q->mqd_gpu_addr, &q->mqd_cpu_ptr);
555 	if (r) {
556 		dev_warn(adev->dev, "failed to create queue mqd bo (%d)", r);
557 		return r;
558 	}
559 	memset(q->mqd_cpu_ptr, 0, mqd_size);
560 
561 	r = amdgpu_bo_reserve(q->mqd_obj, false);
562 	if (unlikely(r != 0))
563 		goto clean_up;
564 
565 	return 0;
566 
567 clean_up:
568 	amdgpu_bo_free_kernel(&q->mqd_obj,
569 			      &q->mqd_gpu_addr,
570 			      &q->mqd_cpu_ptr);
571 	return r;
572 }
573 
574 static void amdgpu_mes_queue_init_mqd(struct amdgpu_device *adev,
575 				     struct amdgpu_mes_queue *q,
576 				     struct amdgpu_mes_queue_properties *p)
577 {
578 	struct amdgpu_mqd *mqd_mgr = &adev->mqds[p->queue_type];
579 	struct amdgpu_mqd_prop mqd_prop = {0};
580 
581 	mqd_prop.mqd_gpu_addr = q->mqd_gpu_addr;
582 	mqd_prop.hqd_base_gpu_addr = p->hqd_base_gpu_addr;
583 	mqd_prop.rptr_gpu_addr = p->rptr_gpu_addr;
584 	mqd_prop.wptr_gpu_addr = p->wptr_gpu_addr;
585 	mqd_prop.queue_size = p->queue_size;
586 	mqd_prop.use_doorbell = true;
587 	mqd_prop.doorbell_index = p->doorbell_off;
588 	mqd_prop.eop_gpu_addr = p->eop_gpu_addr;
589 	mqd_prop.hqd_pipe_priority = p->hqd_pipe_priority;
590 	mqd_prop.hqd_queue_priority = p->hqd_queue_priority;
591 	mqd_prop.hqd_active = false;
592 
593 	if (p->queue_type == AMDGPU_RING_TYPE_GFX ||
594 	    p->queue_type == AMDGPU_RING_TYPE_COMPUTE) {
595 		mutex_lock(&adev->srbm_mutex);
596 		amdgpu_gfx_select_me_pipe_q(adev, p->ring->me, p->ring->pipe, 0, 0, 0);
597 	}
598 
599 	mqd_mgr->init_mqd(adev, q->mqd_cpu_ptr, &mqd_prop);
600 
601 	if (p->queue_type == AMDGPU_RING_TYPE_GFX ||
602 	    p->queue_type == AMDGPU_RING_TYPE_COMPUTE) {
603 		amdgpu_gfx_select_me_pipe_q(adev, 0, 0, 0, 0, 0);
604 		mutex_unlock(&adev->srbm_mutex);
605 	}
606 
607 	amdgpu_bo_unreserve(q->mqd_obj);
608 }
609 
610 int amdgpu_mes_add_hw_queue(struct amdgpu_device *adev, int gang_id,
611 			    struct amdgpu_mes_queue_properties *qprops,
612 			    int *queue_id)
613 {
614 	struct amdgpu_mes_queue *queue;
615 	struct amdgpu_mes_gang *gang;
616 	struct mes_add_queue_input queue_input;
617 	unsigned long flags;
618 	int r;
619 
620 	memset(&queue_input, 0, sizeof(struct mes_add_queue_input));
621 
622 	/* allocate the mes queue buffer */
623 	queue = kzalloc(sizeof(struct amdgpu_mes_queue), GFP_KERNEL);
624 	if (!queue) {
625 		DRM_ERROR("Failed to allocate memory for queue\n");
626 		return -ENOMEM;
627 	}
628 
629 	/* Allocate the queue mqd */
630 	r = amdgpu_mes_queue_alloc_mqd(adev, queue, qprops);
631 	if (r)
632 		goto clean_up_memory;
633 
634 	/*
635 	 * Avoid taking any other locks under MES lock to avoid circular
636 	 * lock dependencies.
637 	 */
638 	amdgpu_mes_lock(&adev->mes);
639 
640 	gang = idr_find(&adev->mes.gang_id_idr, gang_id);
641 	if (!gang) {
642 		DRM_ERROR("gang id %d doesn't exist\n", gang_id);
643 		r = -EINVAL;
644 		goto clean_up_mqd;
645 	}
646 
647 	/* add the mes gang to idr list */
648 	spin_lock_irqsave(&adev->mes.queue_id_lock, flags);
649 	r = idr_alloc(&adev->mes.queue_id_idr, queue, 1, 0,
650 		      GFP_ATOMIC);
651 	if (r < 0) {
652 		spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
653 		goto clean_up_mqd;
654 	}
655 	spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
656 	*queue_id = queue->queue_id = r;
657 
658 	/* allocate a doorbell index for the queue */
659 	r = amdgpu_mes_kernel_doorbell_get(adev, gang->process,
660 					  qprops->queue_type,
661 					  &qprops->doorbell_off);
662 	if (r)
663 		goto clean_up_queue_id;
664 
665 	/* initialize the queue mqd */
666 	amdgpu_mes_queue_init_mqd(adev, queue, qprops);
667 
668 	/* add hw queue to mes */
669 	queue_input.process_id = gang->process->pasid;
670 
671 	queue_input.page_table_base_addr =
672 		adev->vm_manager.vram_base_offset + gang->process->pd_gpu_addr -
673 		adev->gmc.vram_start;
674 
675 	queue_input.process_va_start = 0;
676 	queue_input.process_va_end =
677 		(adev->vm_manager.max_pfn - 1) << AMDGPU_GPU_PAGE_SHIFT;
678 	queue_input.process_quantum = gang->process->process_quantum;
679 	queue_input.process_context_addr = gang->process->proc_ctx_gpu_addr;
680 	queue_input.gang_quantum = gang->gang_quantum;
681 	queue_input.gang_context_addr = gang->gang_ctx_gpu_addr;
682 	queue_input.inprocess_gang_priority = gang->inprocess_gang_priority;
683 	queue_input.gang_global_priority_level = gang->global_priority_level;
684 	queue_input.doorbell_offset = qprops->doorbell_off;
685 	queue_input.mqd_addr = queue->mqd_gpu_addr;
686 	queue_input.wptr_addr = qprops->wptr_gpu_addr;
687 	queue_input.wptr_mc_addr = qprops->wptr_mc_addr;
688 	queue_input.queue_type = qprops->queue_type;
689 	queue_input.paging = qprops->paging;
690 	queue_input.is_kfd_process = 0;
691 
692 	r = adev->mes.funcs->add_hw_queue(&adev->mes, &queue_input);
693 	if (r) {
694 		DRM_ERROR("failed to add hardware queue to MES, doorbell=0x%llx\n",
695 			  qprops->doorbell_off);
696 		goto clean_up_doorbell;
697 	}
698 
699 	DRM_DEBUG("MES hw queue was added, pasid=%d, gang id=%d, "
700 		  "queue type=%d, doorbell=0x%llx\n",
701 		  gang->process->pasid, gang_id, qprops->queue_type,
702 		  qprops->doorbell_off);
703 
704 	queue->ring = qprops->ring;
705 	queue->doorbell_off = qprops->doorbell_off;
706 	queue->wptr_gpu_addr = qprops->wptr_gpu_addr;
707 	queue->queue_type = qprops->queue_type;
708 	queue->paging = qprops->paging;
709 	queue->gang = gang;
710 	queue->ring->mqd_ptr = queue->mqd_cpu_ptr;
711 	list_add_tail(&queue->list, &gang->queue_list);
712 
713 	amdgpu_mes_unlock(&adev->mes);
714 	return 0;
715 
716 clean_up_doorbell:
717 	amdgpu_mes_kernel_doorbell_free(adev, gang->process,
718 				       qprops->doorbell_off);
719 clean_up_queue_id:
720 	spin_lock_irqsave(&adev->mes.queue_id_lock, flags);
721 	idr_remove(&adev->mes.queue_id_idr, queue->queue_id);
722 	spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
723 clean_up_mqd:
724 	amdgpu_mes_unlock(&adev->mes);
725 	amdgpu_mes_queue_free_mqd(queue);
726 clean_up_memory:
727 	kfree(queue);
728 	return r;
729 }
730 
731 int amdgpu_mes_remove_hw_queue(struct amdgpu_device *adev, int queue_id)
732 {
733 	unsigned long flags;
734 	struct amdgpu_mes_queue *queue;
735 	struct amdgpu_mes_gang *gang;
736 	struct mes_remove_queue_input queue_input;
737 	int r;
738 
739 	/*
740 	 * Avoid taking any other locks under MES lock to avoid circular
741 	 * lock dependencies.
742 	 */
743 	amdgpu_mes_lock(&adev->mes);
744 
745 	/* remove the mes gang from idr list */
746 	spin_lock_irqsave(&adev->mes.queue_id_lock, flags);
747 
748 	queue = idr_find(&adev->mes.queue_id_idr, queue_id);
749 	if (!queue) {
750 		spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
751 		amdgpu_mes_unlock(&adev->mes);
752 		DRM_ERROR("queue id %d doesn't exist\n", queue_id);
753 		return -EINVAL;
754 	}
755 
756 	idr_remove(&adev->mes.queue_id_idr, queue_id);
757 	spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
758 
759 	DRM_DEBUG("try to remove queue, doorbell off = 0x%llx\n",
760 		  queue->doorbell_off);
761 
762 	gang = queue->gang;
763 	queue_input.doorbell_offset = queue->doorbell_off;
764 	queue_input.gang_context_addr = gang->gang_ctx_gpu_addr;
765 
766 	r = adev->mes.funcs->remove_hw_queue(&adev->mes, &queue_input);
767 	if (r)
768 		DRM_ERROR("failed to remove hardware queue, queue id = %d\n",
769 			  queue_id);
770 
771 	list_del(&queue->list);
772 	amdgpu_mes_kernel_doorbell_free(adev, gang->process,
773 				       queue->doorbell_off);
774 	amdgpu_mes_unlock(&adev->mes);
775 
776 	amdgpu_mes_queue_free_mqd(queue);
777 	kfree(queue);
778 	return 0;
779 }
780 
781 int amdgpu_mes_unmap_legacy_queue(struct amdgpu_device *adev,
782 				  struct amdgpu_ring *ring,
783 				  enum amdgpu_unmap_queues_action action,
784 				  u64 gpu_addr, u64 seq)
785 {
786 	struct mes_unmap_legacy_queue_input queue_input;
787 	int r;
788 
789 	queue_input.action = action;
790 	queue_input.queue_type = ring->funcs->type;
791 	queue_input.doorbell_offset = ring->doorbell_index;
792 	queue_input.pipe_id = ring->pipe;
793 	queue_input.queue_id = ring->queue;
794 	queue_input.trail_fence_addr = gpu_addr;
795 	queue_input.trail_fence_data = seq;
796 
797 	r = adev->mes.funcs->unmap_legacy_queue(&adev->mes, &queue_input);
798 	if (r)
799 		DRM_ERROR("failed to unmap legacy queue\n");
800 
801 	return r;
802 }
803 
804 uint32_t amdgpu_mes_rreg(struct amdgpu_device *adev, uint32_t reg)
805 {
806 	struct mes_misc_op_input op_input;
807 	int r, val = 0;
808 
809 	op_input.op = MES_MISC_OP_READ_REG;
810 	op_input.read_reg.reg_offset = reg;
811 	op_input.read_reg.buffer_addr = adev->mes.read_val_gpu_addr;
812 
813 	if (!adev->mes.funcs->misc_op) {
814 		DRM_ERROR("mes rreg is not supported!\n");
815 		goto error;
816 	}
817 
818 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
819 	if (r)
820 		DRM_ERROR("failed to read reg (0x%x)\n", reg);
821 	else
822 		val = *(adev->mes.read_val_ptr);
823 
824 error:
825 	return val;
826 }
827 
828 int amdgpu_mes_wreg(struct amdgpu_device *adev,
829 		    uint32_t reg, uint32_t val)
830 {
831 	struct mes_misc_op_input op_input;
832 	int r;
833 
834 	op_input.op = MES_MISC_OP_WRITE_REG;
835 	op_input.write_reg.reg_offset = reg;
836 	op_input.write_reg.reg_value = val;
837 
838 	if (!adev->mes.funcs->misc_op) {
839 		DRM_ERROR("mes wreg is not supported!\n");
840 		r = -EINVAL;
841 		goto error;
842 	}
843 
844 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
845 	if (r)
846 		DRM_ERROR("failed to write reg (0x%x)\n", reg);
847 
848 error:
849 	return r;
850 }
851 
852 int amdgpu_mes_reg_write_reg_wait(struct amdgpu_device *adev,
853 				  uint32_t reg0, uint32_t reg1,
854 				  uint32_t ref, uint32_t mask)
855 {
856 	struct mes_misc_op_input op_input;
857 	int r;
858 
859 	op_input.op = MES_MISC_OP_WRM_REG_WR_WAIT;
860 	op_input.wrm_reg.reg0 = reg0;
861 	op_input.wrm_reg.reg1 = reg1;
862 	op_input.wrm_reg.ref = ref;
863 	op_input.wrm_reg.mask = mask;
864 
865 	if (!adev->mes.funcs->misc_op) {
866 		DRM_ERROR("mes reg_write_reg_wait is not supported!\n");
867 		r = -EINVAL;
868 		goto error;
869 	}
870 
871 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
872 	if (r)
873 		DRM_ERROR("failed to reg_write_reg_wait\n");
874 
875 error:
876 	return r;
877 }
878 
879 int amdgpu_mes_reg_wait(struct amdgpu_device *adev, uint32_t reg,
880 			uint32_t val, uint32_t mask)
881 {
882 	struct mes_misc_op_input op_input;
883 	int r;
884 
885 	op_input.op = MES_MISC_OP_WRM_REG_WAIT;
886 	op_input.wrm_reg.reg0 = reg;
887 	op_input.wrm_reg.ref = val;
888 	op_input.wrm_reg.mask = mask;
889 
890 	if (!adev->mes.funcs->misc_op) {
891 		DRM_ERROR("mes reg wait is not supported!\n");
892 		r = -EINVAL;
893 		goto error;
894 	}
895 
896 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
897 	if (r)
898 		DRM_ERROR("failed to reg_write_reg_wait\n");
899 
900 error:
901 	return r;
902 }
903 
904 int amdgpu_mes_set_shader_debugger(struct amdgpu_device *adev,
905 				uint64_t process_context_addr,
906 				uint32_t spi_gdbg_per_vmid_cntl,
907 				const uint32_t *tcp_watch_cntl,
908 				uint32_t flags,
909 				bool trap_en)
910 {
911 	struct mes_misc_op_input op_input = {0};
912 	int r;
913 
914 	if (!adev->mes.funcs->misc_op) {
915 		DRM_ERROR("mes set shader debugger is not supported!\n");
916 		return -EINVAL;
917 	}
918 
919 	op_input.op = MES_MISC_OP_SET_SHADER_DEBUGGER;
920 	op_input.set_shader_debugger.process_context_addr = process_context_addr;
921 	op_input.set_shader_debugger.flags.u32all = flags;
922 
923 	/* use amdgpu mes_flush_shader_debugger instead */
924 	if (op_input.set_shader_debugger.flags.process_ctx_flush)
925 		return -EINVAL;
926 
927 	op_input.set_shader_debugger.spi_gdbg_per_vmid_cntl = spi_gdbg_per_vmid_cntl;
928 	memcpy(op_input.set_shader_debugger.tcp_watch_cntl, tcp_watch_cntl,
929 			sizeof(op_input.set_shader_debugger.tcp_watch_cntl));
930 
931 	if (((adev->mes.sched_version & AMDGPU_MES_API_VERSION_MASK) >>
932 			AMDGPU_MES_API_VERSION_SHIFT) >= 14)
933 		op_input.set_shader_debugger.trap_en = trap_en;
934 
935 	amdgpu_mes_lock(&adev->mes);
936 
937 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
938 	if (r)
939 		DRM_ERROR("failed to set_shader_debugger\n");
940 
941 	amdgpu_mes_unlock(&adev->mes);
942 
943 	return r;
944 }
945 
946 int amdgpu_mes_flush_shader_debugger(struct amdgpu_device *adev,
947 				     uint64_t process_context_addr)
948 {
949 	struct mes_misc_op_input op_input = {0};
950 	int r;
951 
952 	if (!adev->mes.funcs->misc_op) {
953 		DRM_ERROR("mes flush shader debugger is not supported!\n");
954 		return -EINVAL;
955 	}
956 
957 	op_input.op = MES_MISC_OP_SET_SHADER_DEBUGGER;
958 	op_input.set_shader_debugger.process_context_addr = process_context_addr;
959 	op_input.set_shader_debugger.flags.process_ctx_flush = true;
960 
961 	amdgpu_mes_lock(&adev->mes);
962 
963 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
964 	if (r)
965 		DRM_ERROR("failed to set_shader_debugger\n");
966 
967 	amdgpu_mes_unlock(&adev->mes);
968 
969 	return r;
970 }
971 
972 static void
973 amdgpu_mes_ring_to_queue_props(struct amdgpu_device *adev,
974 			       struct amdgpu_ring *ring,
975 			       struct amdgpu_mes_queue_properties *props)
976 {
977 	props->queue_type = ring->funcs->type;
978 	props->hqd_base_gpu_addr = ring->gpu_addr;
979 	props->rptr_gpu_addr = ring->rptr_gpu_addr;
980 	props->wptr_gpu_addr = ring->wptr_gpu_addr;
981 	props->wptr_mc_addr =
982 		ring->mes_ctx->meta_data_mc_addr + ring->wptr_offs;
983 	props->queue_size = ring->ring_size;
984 	props->eop_gpu_addr = ring->eop_gpu_addr;
985 	props->hqd_pipe_priority = AMDGPU_GFX_PIPE_PRIO_NORMAL;
986 	props->hqd_queue_priority = AMDGPU_GFX_QUEUE_PRIORITY_MINIMUM;
987 	props->paging = false;
988 	props->ring = ring;
989 }
990 
991 #define DEFINE_AMDGPU_MES_CTX_GET_OFFS_ENG(_eng)			\
992 do {									\
993        if (id_offs < AMDGPU_MES_CTX_MAX_OFFS)				\
994 		return offsetof(struct amdgpu_mes_ctx_meta_data,	\
995 				_eng[ring->idx].slots[id_offs]);        \
996        else if (id_offs == AMDGPU_MES_CTX_RING_OFFS)			\
997 		return offsetof(struct amdgpu_mes_ctx_meta_data,        \
998 				_eng[ring->idx].ring);                  \
999        else if (id_offs == AMDGPU_MES_CTX_IB_OFFS)			\
1000 		return offsetof(struct amdgpu_mes_ctx_meta_data,        \
1001 				_eng[ring->idx].ib);                    \
1002        else if (id_offs == AMDGPU_MES_CTX_PADDING_OFFS)			\
1003 		return offsetof(struct amdgpu_mes_ctx_meta_data,        \
1004 				_eng[ring->idx].padding);               \
1005 } while(0)
1006 
1007 int amdgpu_mes_ctx_get_offs(struct amdgpu_ring *ring, unsigned int id_offs)
1008 {
1009 	switch (ring->funcs->type) {
1010 	case AMDGPU_RING_TYPE_GFX:
1011 		DEFINE_AMDGPU_MES_CTX_GET_OFFS_ENG(gfx);
1012 		break;
1013 	case AMDGPU_RING_TYPE_COMPUTE:
1014 		DEFINE_AMDGPU_MES_CTX_GET_OFFS_ENG(compute);
1015 		break;
1016 	case AMDGPU_RING_TYPE_SDMA:
1017 		DEFINE_AMDGPU_MES_CTX_GET_OFFS_ENG(sdma);
1018 		break;
1019 	default:
1020 		break;
1021 	}
1022 
1023 	WARN_ON(1);
1024 	return -EINVAL;
1025 }
1026 
1027 int amdgpu_mes_add_ring(struct amdgpu_device *adev, int gang_id,
1028 			int queue_type, int idx,
1029 			struct amdgpu_mes_ctx_data *ctx_data,
1030 			struct amdgpu_ring **out)
1031 {
1032 	struct amdgpu_ring *ring;
1033 	struct amdgpu_mes_gang *gang;
1034 	struct amdgpu_mes_queue_properties qprops = {0};
1035 	int r, queue_id, pasid;
1036 
1037 	/*
1038 	 * Avoid taking any other locks under MES lock to avoid circular
1039 	 * lock dependencies.
1040 	 */
1041 	amdgpu_mes_lock(&adev->mes);
1042 	gang = idr_find(&adev->mes.gang_id_idr, gang_id);
1043 	if (!gang) {
1044 		DRM_ERROR("gang id %d doesn't exist\n", gang_id);
1045 		amdgpu_mes_unlock(&adev->mes);
1046 		return -EINVAL;
1047 	}
1048 	pasid = gang->process->pasid;
1049 
1050 	ring = kzalloc(sizeof(struct amdgpu_ring), GFP_KERNEL);
1051 	if (!ring) {
1052 		amdgpu_mes_unlock(&adev->mes);
1053 		return -ENOMEM;
1054 	}
1055 
1056 	ring->ring_obj = NULL;
1057 	ring->use_doorbell = true;
1058 	ring->is_mes_queue = true;
1059 	ring->mes_ctx = ctx_data;
1060 	ring->idx = idx;
1061 	ring->no_scheduler = true;
1062 
1063 	if (queue_type == AMDGPU_RING_TYPE_COMPUTE) {
1064 		int offset = offsetof(struct amdgpu_mes_ctx_meta_data,
1065 				      compute[ring->idx].mec_hpd);
1066 		ring->eop_gpu_addr =
1067 			amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
1068 	}
1069 
1070 	switch (queue_type) {
1071 	case AMDGPU_RING_TYPE_GFX:
1072 		ring->funcs = adev->gfx.gfx_ring[0].funcs;
1073 		ring->me = adev->gfx.gfx_ring[0].me;
1074 		ring->pipe = adev->gfx.gfx_ring[0].pipe;
1075 		break;
1076 	case AMDGPU_RING_TYPE_COMPUTE:
1077 		ring->funcs = adev->gfx.compute_ring[0].funcs;
1078 		ring->me = adev->gfx.compute_ring[0].me;
1079 		ring->pipe = adev->gfx.compute_ring[0].pipe;
1080 		break;
1081 	case AMDGPU_RING_TYPE_SDMA:
1082 		ring->funcs = adev->sdma.instance[0].ring.funcs;
1083 		break;
1084 	default:
1085 		BUG();
1086 	}
1087 
1088 	r = amdgpu_ring_init(adev, ring, 1024, NULL, 0,
1089 			     AMDGPU_RING_PRIO_DEFAULT, NULL);
1090 	if (r)
1091 		goto clean_up_memory;
1092 
1093 	amdgpu_mes_ring_to_queue_props(adev, ring, &qprops);
1094 
1095 	dma_fence_wait(gang->process->vm->last_update, false);
1096 	dma_fence_wait(ctx_data->meta_data_va->last_pt_update, false);
1097 	amdgpu_mes_unlock(&adev->mes);
1098 
1099 	r = amdgpu_mes_add_hw_queue(adev, gang_id, &qprops, &queue_id);
1100 	if (r)
1101 		goto clean_up_ring;
1102 
1103 	ring->hw_queue_id = queue_id;
1104 	ring->doorbell_index = qprops.doorbell_off;
1105 
1106 	if (queue_type == AMDGPU_RING_TYPE_GFX)
1107 		sprintf(ring->name, "gfx_%d.%d.%d", pasid, gang_id, queue_id);
1108 	else if (queue_type == AMDGPU_RING_TYPE_COMPUTE)
1109 		sprintf(ring->name, "compute_%d.%d.%d", pasid, gang_id,
1110 			queue_id);
1111 	else if (queue_type == AMDGPU_RING_TYPE_SDMA)
1112 		sprintf(ring->name, "sdma_%d.%d.%d", pasid, gang_id,
1113 			queue_id);
1114 	else
1115 		BUG();
1116 
1117 	*out = ring;
1118 	return 0;
1119 
1120 clean_up_ring:
1121 	amdgpu_ring_fini(ring);
1122 clean_up_memory:
1123 	kfree(ring);
1124 	amdgpu_mes_unlock(&adev->mes);
1125 	return r;
1126 }
1127 
1128 void amdgpu_mes_remove_ring(struct amdgpu_device *adev,
1129 			    struct amdgpu_ring *ring)
1130 {
1131 	if (!ring)
1132 		return;
1133 
1134 	amdgpu_mes_remove_hw_queue(adev, ring->hw_queue_id);
1135 	amdgpu_ring_fini(ring);
1136 	kfree(ring);
1137 }
1138 
1139 uint32_t amdgpu_mes_get_aggregated_doorbell_index(struct amdgpu_device *adev,
1140 						   enum amdgpu_mes_priority_level prio)
1141 {
1142 	return adev->mes.aggregated_doorbells[prio];
1143 }
1144 
1145 int amdgpu_mes_ctx_alloc_meta_data(struct amdgpu_device *adev,
1146 				   struct amdgpu_mes_ctx_data *ctx_data)
1147 {
1148 	int r;
1149 
1150 	r = amdgpu_bo_create_kernel(adev,
1151 			    sizeof(struct amdgpu_mes_ctx_meta_data),
1152 			    PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
1153 			    &ctx_data->meta_data_obj,
1154 			    &ctx_data->meta_data_mc_addr,
1155 			    &ctx_data->meta_data_ptr);
1156 	if (r) {
1157 		dev_warn(adev->dev, "(%d) create CTX bo failed\n", r);
1158 		return r;
1159 	}
1160 
1161 	if (!ctx_data->meta_data_obj)
1162 		return -ENOMEM;
1163 
1164 	memset(ctx_data->meta_data_ptr, 0,
1165 	       sizeof(struct amdgpu_mes_ctx_meta_data));
1166 
1167 	return 0;
1168 }
1169 
1170 void amdgpu_mes_ctx_free_meta_data(struct amdgpu_mes_ctx_data *ctx_data)
1171 {
1172 	if (ctx_data->meta_data_obj)
1173 		amdgpu_bo_free_kernel(&ctx_data->meta_data_obj,
1174 				      &ctx_data->meta_data_mc_addr,
1175 				      &ctx_data->meta_data_ptr);
1176 }
1177 
1178 int amdgpu_mes_ctx_map_meta_data(struct amdgpu_device *adev,
1179 				 struct amdgpu_vm *vm,
1180 				 struct amdgpu_mes_ctx_data *ctx_data)
1181 {
1182 	struct amdgpu_bo_va *bo_va;
1183 	struct amdgpu_sync sync;
1184 	struct drm_exec exec;
1185 	int r;
1186 
1187 	amdgpu_sync_create(&sync);
1188 
1189 	drm_exec_init(&exec, 0, 0);
1190 	drm_exec_until_all_locked(&exec) {
1191 		r = drm_exec_lock_obj(&exec,
1192 				      &ctx_data->meta_data_obj->tbo.base);
1193 		drm_exec_retry_on_contention(&exec);
1194 		if (unlikely(r))
1195 			goto error_fini_exec;
1196 
1197 		r = amdgpu_vm_lock_pd(vm, &exec, 0);
1198 		drm_exec_retry_on_contention(&exec);
1199 		if (unlikely(r))
1200 			goto error_fini_exec;
1201 	}
1202 
1203 	bo_va = amdgpu_vm_bo_add(adev, vm, ctx_data->meta_data_obj);
1204 	if (!bo_va) {
1205 		DRM_ERROR("failed to create bo_va for meta data BO\n");
1206 		r = -ENOMEM;
1207 		goto error_fini_exec;
1208 	}
1209 
1210 	r = amdgpu_vm_bo_map(adev, bo_va, ctx_data->meta_data_gpu_addr, 0,
1211 			     sizeof(struct amdgpu_mes_ctx_meta_data),
1212 			     AMDGPU_PTE_READABLE | AMDGPU_PTE_WRITEABLE |
1213 			     AMDGPU_PTE_EXECUTABLE);
1214 
1215 	if (r) {
1216 		DRM_ERROR("failed to do bo_map on meta data, err=%d\n", r);
1217 		goto error_del_bo_va;
1218 	}
1219 
1220 	r = amdgpu_vm_bo_update(adev, bo_va, false);
1221 	if (r) {
1222 		DRM_ERROR("failed to do vm_bo_update on meta data\n");
1223 		goto error_del_bo_va;
1224 	}
1225 	amdgpu_sync_fence(&sync, bo_va->last_pt_update);
1226 
1227 	r = amdgpu_vm_update_pdes(adev, vm, false);
1228 	if (r) {
1229 		DRM_ERROR("failed to update pdes on meta data\n");
1230 		goto error_del_bo_va;
1231 	}
1232 	amdgpu_sync_fence(&sync, vm->last_update);
1233 
1234 	amdgpu_sync_wait(&sync, false);
1235 	drm_exec_fini(&exec);
1236 
1237 	amdgpu_sync_free(&sync);
1238 	ctx_data->meta_data_va = bo_va;
1239 	return 0;
1240 
1241 error_del_bo_va:
1242 	amdgpu_vm_bo_del(adev, bo_va);
1243 
1244 error_fini_exec:
1245 	drm_exec_fini(&exec);
1246 	amdgpu_sync_free(&sync);
1247 	return r;
1248 }
1249 
1250 int amdgpu_mes_ctx_unmap_meta_data(struct amdgpu_device *adev,
1251 				   struct amdgpu_mes_ctx_data *ctx_data)
1252 {
1253 	struct amdgpu_bo_va *bo_va = ctx_data->meta_data_va;
1254 	struct amdgpu_bo *bo = ctx_data->meta_data_obj;
1255 	struct amdgpu_vm *vm = bo_va->base.vm;
1256 	struct dma_fence *fence;
1257 	struct drm_exec exec;
1258 	long r;
1259 
1260 	drm_exec_init(&exec, 0, 0);
1261 	drm_exec_until_all_locked(&exec) {
1262 		r = drm_exec_lock_obj(&exec,
1263 				      &ctx_data->meta_data_obj->tbo.base);
1264 		drm_exec_retry_on_contention(&exec);
1265 		if (unlikely(r))
1266 			goto out_unlock;
1267 
1268 		r = amdgpu_vm_lock_pd(vm, &exec, 0);
1269 		drm_exec_retry_on_contention(&exec);
1270 		if (unlikely(r))
1271 			goto out_unlock;
1272 	}
1273 
1274 	amdgpu_vm_bo_del(adev, bo_va);
1275 	if (!amdgpu_vm_ready(vm))
1276 		goto out_unlock;
1277 
1278 	r = dma_resv_get_singleton(bo->tbo.base.resv, DMA_RESV_USAGE_BOOKKEEP,
1279 				   &fence);
1280 	if (r)
1281 		goto out_unlock;
1282 	if (fence) {
1283 		amdgpu_bo_fence(bo, fence, true);
1284 		fence = NULL;
1285 	}
1286 
1287 	r = amdgpu_vm_clear_freed(adev, vm, &fence);
1288 	if (r || !fence)
1289 		goto out_unlock;
1290 
1291 	dma_fence_wait(fence, false);
1292 	amdgpu_bo_fence(bo, fence, true);
1293 	dma_fence_put(fence);
1294 
1295 out_unlock:
1296 	if (unlikely(r < 0))
1297 		dev_err(adev->dev, "failed to clear page tables (%ld)\n", r);
1298 	drm_exec_fini(&exec);
1299 
1300 	return r;
1301 }
1302 
1303 static int amdgpu_mes_test_create_gang_and_queues(struct amdgpu_device *adev,
1304 					  int pasid, int *gang_id,
1305 					  int queue_type, int num_queue,
1306 					  struct amdgpu_ring **added_rings,
1307 					  struct amdgpu_mes_ctx_data *ctx_data)
1308 {
1309 	struct amdgpu_ring *ring;
1310 	struct amdgpu_mes_gang_properties gprops = {0};
1311 	int r, j;
1312 
1313 	/* create a gang for the process */
1314 	gprops.priority = AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
1315 	gprops.gang_quantum = adev->mes.default_gang_quantum;
1316 	gprops.inprocess_gang_priority = AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
1317 	gprops.priority_level = AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
1318 	gprops.global_priority_level = AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
1319 
1320 	r = amdgpu_mes_add_gang(adev, pasid, &gprops, gang_id);
1321 	if (r) {
1322 		DRM_ERROR("failed to add gang\n");
1323 		return r;
1324 	}
1325 
1326 	/* create queues for the gang */
1327 	for (j = 0; j < num_queue; j++) {
1328 		r = amdgpu_mes_add_ring(adev, *gang_id, queue_type, j,
1329 					ctx_data, &ring);
1330 		if (r) {
1331 			DRM_ERROR("failed to add ring\n");
1332 			break;
1333 		}
1334 
1335 		DRM_INFO("ring %s was added\n", ring->name);
1336 		added_rings[j] = ring;
1337 	}
1338 
1339 	return 0;
1340 }
1341 
1342 static int amdgpu_mes_test_queues(struct amdgpu_ring **added_rings)
1343 {
1344 	struct amdgpu_ring *ring;
1345 	int i, r;
1346 
1347 	for (i = 0; i < AMDGPU_MES_CTX_MAX_RINGS; i++) {
1348 		ring = added_rings[i];
1349 		if (!ring)
1350 			continue;
1351 
1352 		r = amdgpu_ring_test_helper(ring);
1353 		if (r)
1354 			return r;
1355 
1356 		r = amdgpu_ring_test_ib(ring, 1000 * 10);
1357 		if (r) {
1358 			DRM_DEV_ERROR(ring->adev->dev,
1359 				      "ring %s ib test failed (%d)\n",
1360 				      ring->name, r);
1361 			return r;
1362 		} else
1363 			DRM_INFO("ring %s ib test pass\n", ring->name);
1364 	}
1365 
1366 	return 0;
1367 }
1368 
1369 int amdgpu_mes_self_test(struct amdgpu_device *adev)
1370 {
1371 	struct amdgpu_vm *vm = NULL;
1372 	struct amdgpu_mes_ctx_data ctx_data = {0};
1373 	struct amdgpu_ring *added_rings[AMDGPU_MES_CTX_MAX_RINGS] = { NULL };
1374 	int gang_ids[3] = {0};
1375 	int queue_types[][2] = { { AMDGPU_RING_TYPE_GFX, 1 },
1376 				 { AMDGPU_RING_TYPE_COMPUTE, 1 },
1377 				 { AMDGPU_RING_TYPE_SDMA, 1} };
1378 	int i, r, pasid, k = 0;
1379 
1380 	pasid = amdgpu_pasid_alloc(16);
1381 	if (pasid < 0) {
1382 		dev_warn(adev->dev, "No more PASIDs available!");
1383 		pasid = 0;
1384 	}
1385 
1386 	vm = kzalloc(sizeof(*vm), GFP_KERNEL);
1387 	if (!vm) {
1388 		r = -ENOMEM;
1389 		goto error_pasid;
1390 	}
1391 
1392 	r = amdgpu_vm_init(adev, vm, -1);
1393 	if (r) {
1394 		DRM_ERROR("failed to initialize vm\n");
1395 		goto error_pasid;
1396 	}
1397 
1398 	r = amdgpu_mes_ctx_alloc_meta_data(adev, &ctx_data);
1399 	if (r) {
1400 		DRM_ERROR("failed to alloc ctx meta data\n");
1401 		goto error_fini;
1402 	}
1403 
1404 	ctx_data.meta_data_gpu_addr = AMDGPU_VA_RESERVED_BOTTOM;
1405 	r = amdgpu_mes_ctx_map_meta_data(adev, vm, &ctx_data);
1406 	if (r) {
1407 		DRM_ERROR("failed to map ctx meta data\n");
1408 		goto error_vm;
1409 	}
1410 
1411 	r = amdgpu_mes_create_process(adev, pasid, vm);
1412 	if (r) {
1413 		DRM_ERROR("failed to create MES process\n");
1414 		goto error_vm;
1415 	}
1416 
1417 	for (i = 0; i < ARRAY_SIZE(queue_types); i++) {
1418 		/* On GFX v10.3, fw hasn't supported to map sdma queue. */
1419 		if (amdgpu_ip_version(adev, GC_HWIP, 0) >=
1420 			    IP_VERSION(10, 3, 0) &&
1421 		    amdgpu_ip_version(adev, GC_HWIP, 0) <
1422 			    IP_VERSION(11, 0, 0) &&
1423 		    queue_types[i][0] == AMDGPU_RING_TYPE_SDMA)
1424 			continue;
1425 
1426 		r = amdgpu_mes_test_create_gang_and_queues(adev, pasid,
1427 							   &gang_ids[i],
1428 							   queue_types[i][0],
1429 							   queue_types[i][1],
1430 							   &added_rings[k],
1431 							   &ctx_data);
1432 		if (r)
1433 			goto error_queues;
1434 
1435 		k += queue_types[i][1];
1436 	}
1437 
1438 	/* start ring test and ib test for MES queues */
1439 	amdgpu_mes_test_queues(added_rings);
1440 
1441 error_queues:
1442 	/* remove all queues */
1443 	for (i = 0; i < ARRAY_SIZE(added_rings); i++) {
1444 		if (!added_rings[i])
1445 			continue;
1446 		amdgpu_mes_remove_ring(adev, added_rings[i]);
1447 	}
1448 
1449 	for (i = 0; i < ARRAY_SIZE(gang_ids); i++) {
1450 		if (!gang_ids[i])
1451 			continue;
1452 		amdgpu_mes_remove_gang(adev, gang_ids[i]);
1453 	}
1454 
1455 	amdgpu_mes_destroy_process(adev, pasid);
1456 
1457 error_vm:
1458 	amdgpu_mes_ctx_unmap_meta_data(adev, &ctx_data);
1459 
1460 error_fini:
1461 	amdgpu_vm_fini(adev, vm);
1462 
1463 error_pasid:
1464 	if (pasid)
1465 		amdgpu_pasid_free(pasid);
1466 
1467 	amdgpu_mes_ctx_free_meta_data(&ctx_data);
1468 	kfree(vm);
1469 	return 0;
1470 }
1471 
1472 int amdgpu_mes_init_microcode(struct amdgpu_device *adev, int pipe)
1473 {
1474 	const struct mes_firmware_header_v1_0 *mes_hdr;
1475 	struct amdgpu_firmware_info *info;
1476 	char ucode_prefix[30];
1477 	char fw_name[40];
1478 	bool need_retry = false;
1479 	int r;
1480 
1481 	amdgpu_ucode_ip_version_decode(adev, GC_HWIP, ucode_prefix,
1482 				       sizeof(ucode_prefix));
1483 	if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(11, 0, 0)) {
1484 		snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes%s.bin",
1485 			 ucode_prefix,
1486 			 pipe == AMDGPU_MES_SCHED_PIPE ? "_2" : "1");
1487 		need_retry = true;
1488 	} else {
1489 		snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes%s.bin",
1490 			 ucode_prefix,
1491 			 pipe == AMDGPU_MES_SCHED_PIPE ? "" : "1");
1492 	}
1493 
1494 	r = amdgpu_ucode_request(adev, &adev->mes.fw[pipe], fw_name);
1495 	if (r && need_retry && pipe == AMDGPU_MES_SCHED_PIPE) {
1496 		snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes.bin",
1497 			 ucode_prefix);
1498 		DRM_INFO("try to fall back to %s\n", fw_name);
1499 		r = amdgpu_ucode_request(adev, &adev->mes.fw[pipe],
1500 					 fw_name);
1501 	}
1502 
1503 	if (r)
1504 		goto out;
1505 
1506 	mes_hdr = (const struct mes_firmware_header_v1_0 *)
1507 		adev->mes.fw[pipe]->data;
1508 	adev->mes.uc_start_addr[pipe] =
1509 		le32_to_cpu(mes_hdr->mes_uc_start_addr_lo) |
1510 		((uint64_t)(le32_to_cpu(mes_hdr->mes_uc_start_addr_hi)) << 32);
1511 	adev->mes.data_start_addr[pipe] =
1512 		le32_to_cpu(mes_hdr->mes_data_start_addr_lo) |
1513 		((uint64_t)(le32_to_cpu(mes_hdr->mes_data_start_addr_hi)) << 32);
1514 
1515 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
1516 		int ucode, ucode_data;
1517 
1518 		if (pipe == AMDGPU_MES_SCHED_PIPE) {
1519 			ucode = AMDGPU_UCODE_ID_CP_MES;
1520 			ucode_data = AMDGPU_UCODE_ID_CP_MES_DATA;
1521 		} else {
1522 			ucode = AMDGPU_UCODE_ID_CP_MES1;
1523 			ucode_data = AMDGPU_UCODE_ID_CP_MES1_DATA;
1524 		}
1525 
1526 		info = &adev->firmware.ucode[ucode];
1527 		info->ucode_id = ucode;
1528 		info->fw = adev->mes.fw[pipe];
1529 		adev->firmware.fw_size +=
1530 			ALIGN(le32_to_cpu(mes_hdr->mes_ucode_size_bytes),
1531 			      PAGE_SIZE);
1532 
1533 		info = &adev->firmware.ucode[ucode_data];
1534 		info->ucode_id = ucode_data;
1535 		info->fw = adev->mes.fw[pipe];
1536 		adev->firmware.fw_size +=
1537 			ALIGN(le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes),
1538 			      PAGE_SIZE);
1539 	}
1540 
1541 	return 0;
1542 out:
1543 	amdgpu_ucode_release(&adev->mes.fw[pipe]);
1544 	return r;
1545 }
1546 
1547 #if defined(CONFIG_DEBUG_FS)
1548 
1549 static int amdgpu_debugfs_mes_event_log_show(struct seq_file *m, void *unused)
1550 {
1551 	struct amdgpu_device *adev = m->private;
1552 	uint32_t *mem = (uint32_t *)(adev->mes.event_log_cpu_addr);
1553 
1554 	seq_hex_dump(m, "", DUMP_PREFIX_OFFSET, 32, 4,
1555 		     mem, AMDGPU_MES_LOG_BUFFER_SIZE, false);
1556 
1557 	return 0;
1558 }
1559 
1560 DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_mes_event_log);
1561 
1562 #endif
1563 
1564 void amdgpu_debugfs_mes_event_log_init(struct amdgpu_device *adev)
1565 {
1566 
1567 #if defined(CONFIG_DEBUG_FS)
1568 	struct drm_minor *minor = adev_to_drm(adev)->primary;
1569 	struct dentry *root = minor->debugfs_root;
1570 	if (adev->enable_mes && amdgpu_mes_log_enable)
1571 		debugfs_create_file("amdgpu_mes_event_log", 0444, root,
1572 				    adev, &amdgpu_debugfs_mes_event_log_fops);
1573 
1574 #endif
1575 }
1576