xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c (revision 40004709a3d3b07041a473a163ca911ef04ab8bd)
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 					 int ip_type, uint64_t *doorbell_index)
44 {
45 	unsigned int offset, found;
46 	struct amdgpu_mes *mes = &adev->mes;
47 
48 	if (ip_type == AMDGPU_RING_TYPE_SDMA)
49 		offset = adev->doorbell_index.sdma_engine[0];
50 	else
51 		offset = 0;
52 
53 	found = find_next_zero_bit(mes->doorbell_bitmap, mes->num_mes_dbs, offset);
54 	if (found >= mes->num_mes_dbs) {
55 		DRM_WARN("No doorbell available\n");
56 		return -ENOSPC;
57 	}
58 
59 	set_bit(found, mes->doorbell_bitmap);
60 
61 	/* Get the absolute doorbell index on BAR */
62 	*doorbell_index = mes->db_start_dw_offset + found * 2;
63 	return 0;
64 }
65 
66 static void amdgpu_mes_kernel_doorbell_free(struct amdgpu_device *adev,
67 					   uint32_t doorbell_index)
68 {
69 	unsigned int old, rel_index;
70 	struct amdgpu_mes *mes = &adev->mes;
71 
72 	/* Find the relative index of the doorbell in this object */
73 	rel_index = (doorbell_index - mes->db_start_dw_offset) / 2;
74 	old = test_and_clear_bit(rel_index, mes->doorbell_bitmap);
75 	WARN_ON(!old);
76 }
77 
78 static int amdgpu_mes_doorbell_init(struct amdgpu_device *adev)
79 {
80 	int i;
81 	struct amdgpu_mes *mes = &adev->mes;
82 
83 	/* Bitmap for dynamic allocation of kernel doorbells */
84 	mes->doorbell_bitmap = bitmap_zalloc(PAGE_SIZE / sizeof(u32), GFP_KERNEL);
85 	if (!mes->doorbell_bitmap) {
86 		DRM_ERROR("Failed to allocate MES doorbell bitmap\n");
87 		return -ENOMEM;
88 	}
89 
90 	mes->num_mes_dbs = PAGE_SIZE / AMDGPU_ONE_DOORBELL_SIZE;
91 	for (i = 0; i < AMDGPU_MES_PRIORITY_NUM_LEVELS; i++) {
92 		adev->mes.aggregated_doorbells[i] = mes->db_start_dw_offset + i * 2;
93 		set_bit(i, mes->doorbell_bitmap);
94 	}
95 
96 	return 0;
97 }
98 
99 static int amdgpu_mes_event_log_init(struct amdgpu_device *adev)
100 {
101 	int r;
102 
103 	if (!amdgpu_mes_log_enable)
104 		return 0;
105 
106 	r = amdgpu_bo_create_kernel(adev, adev->mes.event_log_size, PAGE_SIZE,
107 				    AMDGPU_GEM_DOMAIN_GTT,
108 				    &adev->mes.event_log_gpu_obj,
109 				    &adev->mes.event_log_gpu_addr,
110 				    &adev->mes.event_log_cpu_addr);
111 	if (r) {
112 		dev_warn(adev->dev, "failed to create MES event log buffer (%d)", r);
113 		return r;
114 	}
115 
116 	memset(adev->mes.event_log_cpu_addr, 0, adev->mes.event_log_size);
117 
118 	return  0;
119 
120 }
121 
122 static void amdgpu_mes_doorbell_free(struct amdgpu_device *adev)
123 {
124 	bitmap_free(adev->mes.doorbell_bitmap);
125 }
126 
127 int amdgpu_mes_init(struct amdgpu_device *adev)
128 {
129 	int i, r;
130 
131 	adev->mes.adev = adev;
132 
133 	idr_init(&adev->mes.pasid_idr);
134 	idr_init(&adev->mes.gang_id_idr);
135 	idr_init(&adev->mes.queue_id_idr);
136 	ida_init(&adev->mes.doorbell_ida);
137 	spin_lock_init(&adev->mes.queue_id_lock);
138 	mutex_init(&adev->mes.mutex_hidden);
139 
140 	for (i = 0; i < AMDGPU_MAX_MES_PIPES; i++)
141 		spin_lock_init(&adev->mes.ring_lock[i]);
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 >= adev->gfx.mec.num_pipe_per_mec)
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 	for (i = 0; i < AMDGPU_MAX_MES_PIPES; i++) {
169 		r = amdgpu_device_wb_get(adev, &adev->mes.sch_ctx_offs[i]);
170 		if (r) {
171 			dev_err(adev->dev,
172 				"(%d) ring trail_fence_offs wb alloc failed\n",
173 				r);
174 			goto error;
175 		}
176 		adev->mes.sch_ctx_gpu_addr[i] =
177 			adev->wb.gpu_addr + (adev->mes.sch_ctx_offs[i] * 4);
178 		adev->mes.sch_ctx_ptr[i] =
179 			(uint64_t *)&adev->wb.wb[adev->mes.sch_ctx_offs[i]];
180 
181 		r = amdgpu_device_wb_get(adev,
182 				 &adev->mes.query_status_fence_offs[i]);
183 		if (r) {
184 			dev_err(adev->dev,
185 			      "(%d) query_status_fence_offs wb alloc failed\n",
186 			      r);
187 			goto error;
188 		}
189 		adev->mes.query_status_fence_gpu_addr[i] = adev->wb.gpu_addr +
190 			(adev->mes.query_status_fence_offs[i] * 4);
191 		adev->mes.query_status_fence_ptr[i] =
192 			(uint64_t *)&adev->wb.wb[adev->mes.query_status_fence_offs[i]];
193 	}
194 
195 	r = amdgpu_device_wb_get(adev, &adev->mes.read_val_offs);
196 	if (r) {
197 		dev_err(adev->dev,
198 			"(%d) read_val_offs alloc failed\n", r);
199 		goto error;
200 	}
201 	adev->mes.read_val_gpu_addr =
202 		adev->wb.gpu_addr + (adev->mes.read_val_offs * 4);
203 	adev->mes.read_val_ptr =
204 		(uint32_t *)&adev->wb.wb[adev->mes.read_val_offs];
205 
206 	r = amdgpu_mes_doorbell_init(adev);
207 	if (r)
208 		goto error;
209 
210 	r = amdgpu_mes_event_log_init(adev);
211 	if (r)
212 		goto error_doorbell;
213 
214 	return 0;
215 
216 error_doorbell:
217 	amdgpu_mes_doorbell_free(adev);
218 error:
219 	for (i = 0; i < AMDGPU_MAX_MES_PIPES; i++) {
220 		if (adev->mes.sch_ctx_ptr[i])
221 			amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs[i]);
222 		if (adev->mes.query_status_fence_ptr[i])
223 			amdgpu_device_wb_free(adev,
224 				      adev->mes.query_status_fence_offs[i]);
225 	}
226 	if (adev->mes.read_val_ptr)
227 		amdgpu_device_wb_free(adev, adev->mes.read_val_offs);
228 
229 	idr_destroy(&adev->mes.pasid_idr);
230 	idr_destroy(&adev->mes.gang_id_idr);
231 	idr_destroy(&adev->mes.queue_id_idr);
232 	ida_destroy(&adev->mes.doorbell_ida);
233 	mutex_destroy(&adev->mes.mutex_hidden);
234 	return r;
235 }
236 
237 void amdgpu_mes_fini(struct amdgpu_device *adev)
238 {
239 	int i;
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 	for (i = 0; i < AMDGPU_MAX_MES_PIPES; i++) {
246 		if (adev->mes.sch_ctx_ptr[i])
247 			amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs[i]);
248 		if (adev->mes.query_status_fence_ptr[i])
249 			amdgpu_device_wb_free(adev,
250 				      adev->mes.query_status_fence_offs[i]);
251 	}
252 	if (adev->mes.read_val_ptr)
253 		amdgpu_device_wb_free(adev, adev->mes.read_val_offs);
254 
255 	amdgpu_mes_doorbell_free(adev);
256 
257 	idr_destroy(&adev->mes.pasid_idr);
258 	idr_destroy(&adev->mes.gang_id_idr);
259 	idr_destroy(&adev->mes.queue_id_idr);
260 	ida_destroy(&adev->mes.doorbell_ida);
261 	mutex_destroy(&adev->mes.mutex_hidden);
262 }
263 
264 static void amdgpu_mes_queue_free_mqd(struct amdgpu_mes_queue *q)
265 {
266 	amdgpu_bo_free_kernel(&q->mqd_obj,
267 			      &q->mqd_gpu_addr,
268 			      &q->mqd_cpu_ptr);
269 }
270 
271 int amdgpu_mes_create_process(struct amdgpu_device *adev, int pasid,
272 			      struct amdgpu_vm *vm)
273 {
274 	struct amdgpu_mes_process *process;
275 	int r;
276 
277 	/* allocate the mes process buffer */
278 	process = kzalloc(sizeof(struct amdgpu_mes_process), GFP_KERNEL);
279 	if (!process) {
280 		DRM_ERROR("no more memory to create mes process\n");
281 		return -ENOMEM;
282 	}
283 
284 	/* allocate the process context bo and map it */
285 	r = amdgpu_bo_create_kernel(adev, AMDGPU_MES_PROC_CTX_SIZE, PAGE_SIZE,
286 				    AMDGPU_GEM_DOMAIN_GTT,
287 				    &process->proc_ctx_bo,
288 				    &process->proc_ctx_gpu_addr,
289 				    &process->proc_ctx_cpu_ptr);
290 	if (r) {
291 		DRM_ERROR("failed to allocate process context bo\n");
292 		goto clean_up_memory;
293 	}
294 	memset(process->proc_ctx_cpu_ptr, 0, AMDGPU_MES_PROC_CTX_SIZE);
295 
296 	/*
297 	 * Avoid taking any other locks under MES lock to avoid circular
298 	 * lock dependencies.
299 	 */
300 	amdgpu_mes_lock(&adev->mes);
301 
302 	/* add the mes process to idr list */
303 	r = idr_alloc(&adev->mes.pasid_idr, process, pasid, pasid + 1,
304 		      GFP_KERNEL);
305 	if (r < 0) {
306 		DRM_ERROR("failed to lock pasid=%d\n", pasid);
307 		goto clean_up_ctx;
308 	}
309 
310 	INIT_LIST_HEAD(&process->gang_list);
311 	process->vm = vm;
312 	process->pasid = pasid;
313 	process->process_quantum = adev->mes.default_process_quantum;
314 	process->pd_gpu_addr = amdgpu_bo_gpu_offset(vm->root.bo);
315 
316 	amdgpu_mes_unlock(&adev->mes);
317 	return 0;
318 
319 clean_up_ctx:
320 	amdgpu_mes_unlock(&adev->mes);
321 	amdgpu_bo_free_kernel(&process->proc_ctx_bo,
322 			      &process->proc_ctx_gpu_addr,
323 			      &process->proc_ctx_cpu_ptr);
324 clean_up_memory:
325 	kfree(process);
326 	return r;
327 }
328 
329 void amdgpu_mes_destroy_process(struct amdgpu_device *adev, int pasid)
330 {
331 	struct amdgpu_mes_process *process;
332 	struct amdgpu_mes_gang *gang, *tmp1;
333 	struct amdgpu_mes_queue *queue, *tmp2;
334 	struct mes_remove_queue_input queue_input;
335 	unsigned long flags;
336 	int r;
337 
338 	/*
339 	 * Avoid taking any other locks under MES lock to avoid circular
340 	 * lock dependencies.
341 	 */
342 	amdgpu_mes_lock(&adev->mes);
343 
344 	process = idr_find(&adev->mes.pasid_idr, pasid);
345 	if (!process) {
346 		DRM_WARN("pasid %d doesn't exist\n", pasid);
347 		amdgpu_mes_unlock(&adev->mes);
348 		return;
349 	}
350 
351 	/* Remove all queues from hardware */
352 	list_for_each_entry_safe(gang, tmp1, &process->gang_list, list) {
353 		list_for_each_entry_safe(queue, tmp2, &gang->queue_list, list) {
354 			spin_lock_irqsave(&adev->mes.queue_id_lock, flags);
355 			idr_remove(&adev->mes.queue_id_idr, queue->queue_id);
356 			spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
357 
358 			queue_input.doorbell_offset = queue->doorbell_off;
359 			queue_input.gang_context_addr = gang->gang_ctx_gpu_addr;
360 
361 			r = adev->mes.funcs->remove_hw_queue(&adev->mes,
362 							     &queue_input);
363 			if (r)
364 				DRM_WARN("failed to remove hardware queue\n");
365 		}
366 
367 		idr_remove(&adev->mes.gang_id_idr, gang->gang_id);
368 	}
369 
370 	idr_remove(&adev->mes.pasid_idr, pasid);
371 	amdgpu_mes_unlock(&adev->mes);
372 
373 	/* free all memory allocated by the process */
374 	list_for_each_entry_safe(gang, tmp1, &process->gang_list, list) {
375 		/* free all queues in the gang */
376 		list_for_each_entry_safe(queue, tmp2, &gang->queue_list, list) {
377 			amdgpu_mes_queue_free_mqd(queue);
378 			list_del(&queue->list);
379 			kfree(queue);
380 		}
381 		amdgpu_bo_free_kernel(&gang->gang_ctx_bo,
382 				      &gang->gang_ctx_gpu_addr,
383 				      &gang->gang_ctx_cpu_ptr);
384 		list_del(&gang->list);
385 		kfree(gang);
386 
387 	}
388 	amdgpu_bo_free_kernel(&process->proc_ctx_bo,
389 			      &process->proc_ctx_gpu_addr,
390 			      &process->proc_ctx_cpu_ptr);
391 	kfree(process);
392 }
393 
394 int amdgpu_mes_add_gang(struct amdgpu_device *adev, int pasid,
395 			struct amdgpu_mes_gang_properties *gprops,
396 			int *gang_id)
397 {
398 	struct amdgpu_mes_process *process;
399 	struct amdgpu_mes_gang *gang;
400 	int r;
401 
402 	/* allocate the mes gang buffer */
403 	gang = kzalloc(sizeof(struct amdgpu_mes_gang), GFP_KERNEL);
404 	if (!gang) {
405 		return -ENOMEM;
406 	}
407 
408 	/* allocate the gang context bo and map it to cpu space */
409 	r = amdgpu_bo_create_kernel(adev, AMDGPU_MES_GANG_CTX_SIZE, PAGE_SIZE,
410 				    AMDGPU_GEM_DOMAIN_GTT,
411 				    &gang->gang_ctx_bo,
412 				    &gang->gang_ctx_gpu_addr,
413 				    &gang->gang_ctx_cpu_ptr);
414 	if (r) {
415 		DRM_ERROR("failed to allocate process context bo\n");
416 		goto clean_up_mem;
417 	}
418 	memset(gang->gang_ctx_cpu_ptr, 0, AMDGPU_MES_GANG_CTX_SIZE);
419 
420 	/*
421 	 * Avoid taking any other locks under MES lock to avoid circular
422 	 * lock dependencies.
423 	 */
424 	amdgpu_mes_lock(&adev->mes);
425 
426 	process = idr_find(&adev->mes.pasid_idr, pasid);
427 	if (!process) {
428 		DRM_ERROR("pasid %d doesn't exist\n", pasid);
429 		r = -EINVAL;
430 		goto clean_up_ctx;
431 	}
432 
433 	/* add the mes gang to idr list */
434 	r = idr_alloc(&adev->mes.gang_id_idr, gang, 1, 0,
435 		      GFP_KERNEL);
436 	if (r < 0) {
437 		DRM_ERROR("failed to allocate idr for gang\n");
438 		goto clean_up_ctx;
439 	}
440 
441 	gang->gang_id = r;
442 	*gang_id = r;
443 
444 	INIT_LIST_HEAD(&gang->queue_list);
445 	gang->process = process;
446 	gang->priority = gprops->priority;
447 	gang->gang_quantum = gprops->gang_quantum ?
448 		gprops->gang_quantum : adev->mes.default_gang_quantum;
449 	gang->global_priority_level = gprops->global_priority_level;
450 	gang->inprocess_gang_priority = gprops->inprocess_gang_priority;
451 	list_add_tail(&gang->list, &process->gang_list);
452 
453 	amdgpu_mes_unlock(&adev->mes);
454 	return 0;
455 
456 clean_up_ctx:
457 	amdgpu_mes_unlock(&adev->mes);
458 	amdgpu_bo_free_kernel(&gang->gang_ctx_bo,
459 			      &gang->gang_ctx_gpu_addr,
460 			      &gang->gang_ctx_cpu_ptr);
461 clean_up_mem:
462 	kfree(gang);
463 	return r;
464 }
465 
466 int amdgpu_mes_remove_gang(struct amdgpu_device *adev, int gang_id)
467 {
468 	struct amdgpu_mes_gang *gang;
469 
470 	/*
471 	 * Avoid taking any other locks under MES lock to avoid circular
472 	 * lock dependencies.
473 	 */
474 	amdgpu_mes_lock(&adev->mes);
475 
476 	gang = idr_find(&adev->mes.gang_id_idr, gang_id);
477 	if (!gang) {
478 		DRM_ERROR("gang id %d doesn't exist\n", gang_id);
479 		amdgpu_mes_unlock(&adev->mes);
480 		return -EINVAL;
481 	}
482 
483 	if (!list_empty(&gang->queue_list)) {
484 		DRM_ERROR("queue list is not empty\n");
485 		amdgpu_mes_unlock(&adev->mes);
486 		return -EBUSY;
487 	}
488 
489 	idr_remove(&adev->mes.gang_id_idr, gang->gang_id);
490 	list_del(&gang->list);
491 	amdgpu_mes_unlock(&adev->mes);
492 
493 	amdgpu_bo_free_kernel(&gang->gang_ctx_bo,
494 			      &gang->gang_ctx_gpu_addr,
495 			      &gang->gang_ctx_cpu_ptr);
496 
497 	kfree(gang);
498 
499 	return 0;
500 }
501 
502 int amdgpu_mes_suspend(struct amdgpu_device *adev)
503 {
504 	struct mes_suspend_gang_input input;
505 	int r;
506 
507 	if (!amdgpu_mes_suspend_resume_all_supported(adev))
508 		return 0;
509 
510 	memset(&input, 0x0, sizeof(struct mes_suspend_gang_input));
511 	input.suspend_all_gangs = 1;
512 
513 	/*
514 	 * Avoid taking any other locks under MES lock to avoid circular
515 	 * lock dependencies.
516 	 */
517 	amdgpu_mes_lock(&adev->mes);
518 	r = adev->mes.funcs->suspend_gang(&adev->mes, &input);
519 	amdgpu_mes_unlock(&adev->mes);
520 	if (r)
521 		DRM_ERROR("failed to suspend all gangs");
522 
523 	return r;
524 }
525 
526 int amdgpu_mes_resume(struct amdgpu_device *adev)
527 {
528 	struct mes_resume_gang_input input;
529 	int r;
530 
531 	if (!amdgpu_mes_suspend_resume_all_supported(adev))
532 		return 0;
533 
534 	memset(&input, 0x0, sizeof(struct mes_resume_gang_input));
535 	input.resume_all_gangs = 1;
536 
537 	/*
538 	 * Avoid taking any other locks under MES lock to avoid circular
539 	 * lock dependencies.
540 	 */
541 	amdgpu_mes_lock(&adev->mes);
542 	r = adev->mes.funcs->resume_gang(&adev->mes, &input);
543 	amdgpu_mes_unlock(&adev->mes);
544 	if (r)
545 		DRM_ERROR("failed to resume all gangs");
546 
547 	return r;
548 }
549 
550 static int amdgpu_mes_queue_alloc_mqd(struct amdgpu_device *adev,
551 				     struct amdgpu_mes_queue *q,
552 				     struct amdgpu_mes_queue_properties *p)
553 {
554 	struct amdgpu_mqd *mqd_mgr = &adev->mqds[p->queue_type];
555 	u32 mqd_size = mqd_mgr->mqd_size;
556 	int r;
557 
558 	r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
559 				    AMDGPU_GEM_DOMAIN_GTT,
560 				    &q->mqd_obj,
561 				    &q->mqd_gpu_addr, &q->mqd_cpu_ptr);
562 	if (r) {
563 		dev_warn(adev->dev, "failed to create queue mqd bo (%d)", r);
564 		return r;
565 	}
566 	memset(q->mqd_cpu_ptr, 0, mqd_size);
567 
568 	r = amdgpu_bo_reserve(q->mqd_obj, false);
569 	if (unlikely(r != 0))
570 		goto clean_up;
571 
572 	return 0;
573 
574 clean_up:
575 	amdgpu_bo_free_kernel(&q->mqd_obj,
576 			      &q->mqd_gpu_addr,
577 			      &q->mqd_cpu_ptr);
578 	return r;
579 }
580 
581 static void amdgpu_mes_queue_init_mqd(struct amdgpu_device *adev,
582 				     struct amdgpu_mes_queue *q,
583 				     struct amdgpu_mes_queue_properties *p)
584 {
585 	struct amdgpu_mqd *mqd_mgr = &adev->mqds[p->queue_type];
586 	struct amdgpu_mqd_prop mqd_prop = {0};
587 
588 	mqd_prop.mqd_gpu_addr = q->mqd_gpu_addr;
589 	mqd_prop.hqd_base_gpu_addr = p->hqd_base_gpu_addr;
590 	mqd_prop.rptr_gpu_addr = p->rptr_gpu_addr;
591 	mqd_prop.wptr_gpu_addr = p->wptr_gpu_addr;
592 	mqd_prop.queue_size = p->queue_size;
593 	mqd_prop.use_doorbell = true;
594 	mqd_prop.doorbell_index = p->doorbell_off;
595 	mqd_prop.eop_gpu_addr = p->eop_gpu_addr;
596 	mqd_prop.hqd_pipe_priority = p->hqd_pipe_priority;
597 	mqd_prop.hqd_queue_priority = p->hqd_queue_priority;
598 	mqd_prop.hqd_active = false;
599 
600 	if (p->queue_type == AMDGPU_RING_TYPE_GFX ||
601 	    p->queue_type == AMDGPU_RING_TYPE_COMPUTE) {
602 		mutex_lock(&adev->srbm_mutex);
603 		amdgpu_gfx_select_me_pipe_q(adev, p->ring->me, p->ring->pipe, 0, 0, 0);
604 	}
605 
606 	mqd_mgr->init_mqd(adev, q->mqd_cpu_ptr, &mqd_prop);
607 
608 	if (p->queue_type == AMDGPU_RING_TYPE_GFX ||
609 	    p->queue_type == AMDGPU_RING_TYPE_COMPUTE) {
610 		amdgpu_gfx_select_me_pipe_q(adev, 0, 0, 0, 0, 0);
611 		mutex_unlock(&adev->srbm_mutex);
612 	}
613 
614 	amdgpu_bo_unreserve(q->mqd_obj);
615 }
616 
617 int amdgpu_mes_add_hw_queue(struct amdgpu_device *adev, int gang_id,
618 			    struct amdgpu_mes_queue_properties *qprops,
619 			    int *queue_id)
620 {
621 	struct amdgpu_mes_queue *queue;
622 	struct amdgpu_mes_gang *gang;
623 	struct mes_add_queue_input queue_input;
624 	unsigned long flags;
625 	int r;
626 
627 	memset(&queue_input, 0, sizeof(struct mes_add_queue_input));
628 
629 	/* allocate the mes queue buffer */
630 	queue = kzalloc(sizeof(struct amdgpu_mes_queue), GFP_KERNEL);
631 	if (!queue) {
632 		DRM_ERROR("Failed to allocate memory for queue\n");
633 		return -ENOMEM;
634 	}
635 
636 	/* Allocate the queue mqd */
637 	r = amdgpu_mes_queue_alloc_mqd(adev, queue, qprops);
638 	if (r)
639 		goto clean_up_memory;
640 
641 	/*
642 	 * Avoid taking any other locks under MES lock to avoid circular
643 	 * lock dependencies.
644 	 */
645 	amdgpu_mes_lock(&adev->mes);
646 
647 	gang = idr_find(&adev->mes.gang_id_idr, gang_id);
648 	if (!gang) {
649 		DRM_ERROR("gang id %d doesn't exist\n", gang_id);
650 		r = -EINVAL;
651 		goto clean_up_mqd;
652 	}
653 
654 	/* add the mes gang to idr list */
655 	spin_lock_irqsave(&adev->mes.queue_id_lock, flags);
656 	r = idr_alloc(&adev->mes.queue_id_idr, queue, 1, 0,
657 		      GFP_ATOMIC);
658 	if (r < 0) {
659 		spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
660 		goto clean_up_mqd;
661 	}
662 	spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
663 	*queue_id = queue->queue_id = r;
664 
665 	/* allocate a doorbell index for the queue */
666 	r = amdgpu_mes_kernel_doorbell_get(adev,
667 					  qprops->queue_type,
668 					  &qprops->doorbell_off);
669 	if (r)
670 		goto clean_up_queue_id;
671 
672 	/* initialize the queue mqd */
673 	amdgpu_mes_queue_init_mqd(adev, queue, qprops);
674 
675 	/* add hw queue to mes */
676 	queue_input.process_id = gang->process->pasid;
677 
678 	queue_input.page_table_base_addr =
679 		adev->vm_manager.vram_base_offset + gang->process->pd_gpu_addr -
680 		adev->gmc.vram_start;
681 
682 	queue_input.process_va_start = 0;
683 	queue_input.process_va_end =
684 		(adev->vm_manager.max_pfn - 1) << AMDGPU_GPU_PAGE_SHIFT;
685 	queue_input.process_quantum = gang->process->process_quantum;
686 	queue_input.process_context_addr = gang->process->proc_ctx_gpu_addr;
687 	queue_input.gang_quantum = gang->gang_quantum;
688 	queue_input.gang_context_addr = gang->gang_ctx_gpu_addr;
689 	queue_input.inprocess_gang_priority = gang->inprocess_gang_priority;
690 	queue_input.gang_global_priority_level = gang->global_priority_level;
691 	queue_input.doorbell_offset = qprops->doorbell_off;
692 	queue_input.mqd_addr = queue->mqd_gpu_addr;
693 	queue_input.wptr_addr = qprops->wptr_gpu_addr;
694 	queue_input.wptr_mc_addr = qprops->wptr_mc_addr;
695 	queue_input.queue_type = qprops->queue_type;
696 	queue_input.paging = qprops->paging;
697 	queue_input.is_kfd_process = 0;
698 
699 	r = adev->mes.funcs->add_hw_queue(&adev->mes, &queue_input);
700 	if (r) {
701 		DRM_ERROR("failed to add hardware queue to MES, doorbell=0x%llx\n",
702 			  qprops->doorbell_off);
703 		goto clean_up_doorbell;
704 	}
705 
706 	DRM_DEBUG("MES hw queue was added, pasid=%d, gang id=%d, "
707 		  "queue type=%d, doorbell=0x%llx\n",
708 		  gang->process->pasid, gang_id, qprops->queue_type,
709 		  qprops->doorbell_off);
710 
711 	queue->ring = qprops->ring;
712 	queue->doorbell_off = qprops->doorbell_off;
713 	queue->wptr_gpu_addr = qprops->wptr_gpu_addr;
714 	queue->queue_type = qprops->queue_type;
715 	queue->paging = qprops->paging;
716 	queue->gang = gang;
717 	queue->ring->mqd_ptr = queue->mqd_cpu_ptr;
718 	list_add_tail(&queue->list, &gang->queue_list);
719 
720 	amdgpu_mes_unlock(&adev->mes);
721 	return 0;
722 
723 clean_up_doorbell:
724 	amdgpu_mes_kernel_doorbell_free(adev, qprops->doorbell_off);
725 clean_up_queue_id:
726 	spin_lock_irqsave(&adev->mes.queue_id_lock, flags);
727 	idr_remove(&adev->mes.queue_id_idr, queue->queue_id);
728 	spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
729 clean_up_mqd:
730 	amdgpu_mes_unlock(&adev->mes);
731 	amdgpu_mes_queue_free_mqd(queue);
732 clean_up_memory:
733 	kfree(queue);
734 	return r;
735 }
736 
737 int amdgpu_mes_remove_hw_queue(struct amdgpu_device *adev, int queue_id)
738 {
739 	unsigned long flags;
740 	struct amdgpu_mes_queue *queue;
741 	struct amdgpu_mes_gang *gang;
742 	struct mes_remove_queue_input queue_input;
743 	int r;
744 
745 	/*
746 	 * Avoid taking any other locks under MES lock to avoid circular
747 	 * lock dependencies.
748 	 */
749 	amdgpu_mes_lock(&adev->mes);
750 
751 	/* remove the mes gang from idr list */
752 	spin_lock_irqsave(&adev->mes.queue_id_lock, flags);
753 
754 	queue = idr_find(&adev->mes.queue_id_idr, queue_id);
755 	if (!queue) {
756 		spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
757 		amdgpu_mes_unlock(&adev->mes);
758 		DRM_ERROR("queue id %d doesn't exist\n", queue_id);
759 		return -EINVAL;
760 	}
761 
762 	idr_remove(&adev->mes.queue_id_idr, queue_id);
763 	spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
764 
765 	DRM_DEBUG("try to remove queue, doorbell off = 0x%llx\n",
766 		  queue->doorbell_off);
767 
768 	gang = queue->gang;
769 	queue_input.doorbell_offset = queue->doorbell_off;
770 	queue_input.gang_context_addr = gang->gang_ctx_gpu_addr;
771 
772 	r = adev->mes.funcs->remove_hw_queue(&adev->mes, &queue_input);
773 	if (r)
774 		DRM_ERROR("failed to remove hardware queue, queue id = %d\n",
775 			  queue_id);
776 
777 	list_del(&queue->list);
778 	amdgpu_mes_kernel_doorbell_free(adev, queue->doorbell_off);
779 	amdgpu_mes_unlock(&adev->mes);
780 
781 	amdgpu_mes_queue_free_mqd(queue);
782 	kfree(queue);
783 	return 0;
784 }
785 
786 int amdgpu_mes_reset_hw_queue(struct amdgpu_device *adev, int queue_id)
787 {
788 	unsigned long flags;
789 	struct amdgpu_mes_queue *queue;
790 	struct amdgpu_mes_gang *gang;
791 	struct mes_reset_queue_input queue_input;
792 	int r;
793 
794 	/*
795 	 * Avoid taking any other locks under MES lock to avoid circular
796 	 * lock dependencies.
797 	 */
798 	amdgpu_mes_lock(&adev->mes);
799 
800 	/* remove the mes gang from idr list */
801 	spin_lock_irqsave(&adev->mes.queue_id_lock, flags);
802 
803 	queue = idr_find(&adev->mes.queue_id_idr, queue_id);
804 	if (!queue) {
805 		spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
806 		amdgpu_mes_unlock(&adev->mes);
807 		DRM_ERROR("queue id %d doesn't exist\n", queue_id);
808 		return -EINVAL;
809 	}
810 	spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
811 
812 	DRM_DEBUG("try to reset queue, doorbell off = 0x%llx\n",
813 		  queue->doorbell_off);
814 
815 	gang = queue->gang;
816 	queue_input.doorbell_offset = queue->doorbell_off;
817 	queue_input.gang_context_addr = gang->gang_ctx_gpu_addr;
818 
819 	r = adev->mes.funcs->reset_hw_queue(&adev->mes, &queue_input);
820 	if (r)
821 		DRM_ERROR("failed to reset hardware queue, queue id = %d\n",
822 			  queue_id);
823 
824 	amdgpu_mes_unlock(&adev->mes);
825 
826 	return 0;
827 }
828 
829 int amdgpu_mes_reset_hw_queue_mmio(struct amdgpu_device *adev, int queue_type,
830 				   int me_id, int pipe_id, int queue_id, int vmid)
831 {
832 	struct mes_reset_queue_input queue_input;
833 	int r;
834 
835 	queue_input.use_mmio = true;
836 	queue_input.me_id = me_id;
837 	queue_input.pipe_id = pipe_id;
838 	queue_input.queue_id = queue_id;
839 	queue_input.vmid = vmid;
840 	r = adev->mes.funcs->reset_hw_queue(&adev->mes, &queue_input);
841 	if (r)
842 		DRM_ERROR("failed to reset hardware queue by mmio, queue id = %d\n",
843 			  queue_id);
844 	return r;
845 }
846 
847 int amdgpu_mes_map_legacy_queue(struct amdgpu_device *adev,
848 				struct amdgpu_ring *ring)
849 {
850 	struct mes_map_legacy_queue_input queue_input;
851 	int r;
852 
853 	memset(&queue_input, 0, sizeof(queue_input));
854 
855 	queue_input.queue_type = ring->funcs->type;
856 	queue_input.doorbell_offset = ring->doorbell_index;
857 	queue_input.pipe_id = ring->pipe;
858 	queue_input.queue_id = ring->queue;
859 	queue_input.mqd_addr = amdgpu_bo_gpu_offset(ring->mqd_obj);
860 	queue_input.wptr_addr = ring->wptr_gpu_addr;
861 
862 	r = adev->mes.funcs->map_legacy_queue(&adev->mes, &queue_input);
863 	if (r)
864 		DRM_ERROR("failed to map legacy queue\n");
865 
866 	return r;
867 }
868 
869 int amdgpu_mes_unmap_legacy_queue(struct amdgpu_device *adev,
870 				  struct amdgpu_ring *ring,
871 				  enum amdgpu_unmap_queues_action action,
872 				  u64 gpu_addr, u64 seq)
873 {
874 	struct mes_unmap_legacy_queue_input queue_input;
875 	int r;
876 
877 	queue_input.action = action;
878 	queue_input.queue_type = ring->funcs->type;
879 	queue_input.doorbell_offset = ring->doorbell_index;
880 	queue_input.pipe_id = ring->pipe;
881 	queue_input.queue_id = ring->queue;
882 	queue_input.trail_fence_addr = gpu_addr;
883 	queue_input.trail_fence_data = seq;
884 
885 	r = adev->mes.funcs->unmap_legacy_queue(&adev->mes, &queue_input);
886 	if (r)
887 		DRM_ERROR("failed to unmap legacy queue\n");
888 
889 	return r;
890 }
891 
892 int amdgpu_mes_reset_legacy_queue(struct amdgpu_device *adev,
893 				  struct amdgpu_ring *ring,
894 				  unsigned int vmid,
895 				  bool use_mmio)
896 {
897 	struct mes_reset_legacy_queue_input queue_input;
898 	int r;
899 
900 	memset(&queue_input, 0, sizeof(queue_input));
901 
902 	queue_input.queue_type = ring->funcs->type;
903 	queue_input.doorbell_offset = ring->doorbell_index;
904 	queue_input.me_id = ring->me;
905 	queue_input.pipe_id = ring->pipe;
906 	queue_input.queue_id = ring->queue;
907 	queue_input.mqd_addr = amdgpu_bo_gpu_offset(ring->mqd_obj);
908 	queue_input.wptr_addr = ring->wptr_gpu_addr;
909 	queue_input.vmid = vmid;
910 	queue_input.use_mmio = use_mmio;
911 
912 	r = adev->mes.funcs->reset_legacy_queue(&adev->mes, &queue_input);
913 	if (r)
914 		DRM_ERROR("failed to reset legacy queue\n");
915 
916 	return r;
917 }
918 
919 uint32_t amdgpu_mes_rreg(struct amdgpu_device *adev, uint32_t reg)
920 {
921 	struct mes_misc_op_input op_input;
922 	int r, val = 0;
923 
924 	op_input.op = MES_MISC_OP_READ_REG;
925 	op_input.read_reg.reg_offset = reg;
926 	op_input.read_reg.buffer_addr = adev->mes.read_val_gpu_addr;
927 
928 	if (!adev->mes.funcs->misc_op) {
929 		DRM_ERROR("mes rreg is not supported!\n");
930 		goto error;
931 	}
932 
933 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
934 	if (r)
935 		DRM_ERROR("failed to read reg (0x%x)\n", reg);
936 	else
937 		val = *(adev->mes.read_val_ptr);
938 
939 error:
940 	return val;
941 }
942 
943 int amdgpu_mes_wreg(struct amdgpu_device *adev,
944 		    uint32_t reg, uint32_t val)
945 {
946 	struct mes_misc_op_input op_input;
947 	int r;
948 
949 	op_input.op = MES_MISC_OP_WRITE_REG;
950 	op_input.write_reg.reg_offset = reg;
951 	op_input.write_reg.reg_value = val;
952 
953 	if (!adev->mes.funcs->misc_op) {
954 		DRM_ERROR("mes wreg is not supported!\n");
955 		r = -EINVAL;
956 		goto error;
957 	}
958 
959 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
960 	if (r)
961 		DRM_ERROR("failed to write reg (0x%x)\n", reg);
962 
963 error:
964 	return r;
965 }
966 
967 int amdgpu_mes_reg_write_reg_wait(struct amdgpu_device *adev,
968 				  uint32_t reg0, uint32_t reg1,
969 				  uint32_t ref, uint32_t mask)
970 {
971 	struct mes_misc_op_input op_input;
972 	int r;
973 
974 	op_input.op = MES_MISC_OP_WRM_REG_WR_WAIT;
975 	op_input.wrm_reg.reg0 = reg0;
976 	op_input.wrm_reg.reg1 = reg1;
977 	op_input.wrm_reg.ref = ref;
978 	op_input.wrm_reg.mask = mask;
979 
980 	if (!adev->mes.funcs->misc_op) {
981 		DRM_ERROR("mes reg_write_reg_wait is not supported!\n");
982 		r = -EINVAL;
983 		goto error;
984 	}
985 
986 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
987 	if (r)
988 		DRM_ERROR("failed to reg_write_reg_wait\n");
989 
990 error:
991 	return r;
992 }
993 
994 int amdgpu_mes_reg_wait(struct amdgpu_device *adev, uint32_t reg,
995 			uint32_t val, uint32_t mask)
996 {
997 	struct mes_misc_op_input op_input;
998 	int r;
999 
1000 	op_input.op = MES_MISC_OP_WRM_REG_WAIT;
1001 	op_input.wrm_reg.reg0 = reg;
1002 	op_input.wrm_reg.ref = val;
1003 	op_input.wrm_reg.mask = mask;
1004 
1005 	if (!adev->mes.funcs->misc_op) {
1006 		DRM_ERROR("mes reg wait is not supported!\n");
1007 		r = -EINVAL;
1008 		goto error;
1009 	}
1010 
1011 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
1012 	if (r)
1013 		DRM_ERROR("failed to reg_write_reg_wait\n");
1014 
1015 error:
1016 	return r;
1017 }
1018 
1019 int amdgpu_mes_set_shader_debugger(struct amdgpu_device *adev,
1020 				uint64_t process_context_addr,
1021 				uint32_t spi_gdbg_per_vmid_cntl,
1022 				const uint32_t *tcp_watch_cntl,
1023 				uint32_t flags,
1024 				bool trap_en)
1025 {
1026 	struct mes_misc_op_input op_input = {0};
1027 	int r;
1028 
1029 	if (!adev->mes.funcs->misc_op) {
1030 		DRM_ERROR("mes set shader debugger is not supported!\n");
1031 		return -EINVAL;
1032 	}
1033 
1034 	op_input.op = MES_MISC_OP_SET_SHADER_DEBUGGER;
1035 	op_input.set_shader_debugger.process_context_addr = process_context_addr;
1036 	op_input.set_shader_debugger.flags.u32all = flags;
1037 
1038 	/* use amdgpu mes_flush_shader_debugger instead */
1039 	if (op_input.set_shader_debugger.flags.process_ctx_flush)
1040 		return -EINVAL;
1041 
1042 	op_input.set_shader_debugger.spi_gdbg_per_vmid_cntl = spi_gdbg_per_vmid_cntl;
1043 	memcpy(op_input.set_shader_debugger.tcp_watch_cntl, tcp_watch_cntl,
1044 			sizeof(op_input.set_shader_debugger.tcp_watch_cntl));
1045 
1046 	if (((adev->mes.sched_version & AMDGPU_MES_API_VERSION_MASK) >>
1047 			AMDGPU_MES_API_VERSION_SHIFT) >= 14)
1048 		op_input.set_shader_debugger.trap_en = trap_en;
1049 
1050 	amdgpu_mes_lock(&adev->mes);
1051 
1052 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
1053 	if (r)
1054 		DRM_ERROR("failed to set_shader_debugger\n");
1055 
1056 	amdgpu_mes_unlock(&adev->mes);
1057 
1058 	return r;
1059 }
1060 
1061 int amdgpu_mes_flush_shader_debugger(struct amdgpu_device *adev,
1062 				     uint64_t process_context_addr)
1063 {
1064 	struct mes_misc_op_input op_input = {0};
1065 	int r;
1066 
1067 	if (!adev->mes.funcs->misc_op) {
1068 		DRM_ERROR("mes flush shader debugger is not supported!\n");
1069 		return -EINVAL;
1070 	}
1071 
1072 	op_input.op = MES_MISC_OP_SET_SHADER_DEBUGGER;
1073 	op_input.set_shader_debugger.process_context_addr = process_context_addr;
1074 	op_input.set_shader_debugger.flags.process_ctx_flush = true;
1075 
1076 	amdgpu_mes_lock(&adev->mes);
1077 
1078 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
1079 	if (r)
1080 		DRM_ERROR("failed to set_shader_debugger\n");
1081 
1082 	amdgpu_mes_unlock(&adev->mes);
1083 
1084 	return r;
1085 }
1086 
1087 static void
1088 amdgpu_mes_ring_to_queue_props(struct amdgpu_device *adev,
1089 			       struct amdgpu_ring *ring,
1090 			       struct amdgpu_mes_queue_properties *props)
1091 {
1092 	props->queue_type = ring->funcs->type;
1093 	props->hqd_base_gpu_addr = ring->gpu_addr;
1094 	props->rptr_gpu_addr = ring->rptr_gpu_addr;
1095 	props->wptr_gpu_addr = ring->wptr_gpu_addr;
1096 	props->wptr_mc_addr =
1097 		ring->mes_ctx->meta_data_mc_addr + ring->wptr_offs;
1098 	props->queue_size = ring->ring_size;
1099 	props->eop_gpu_addr = ring->eop_gpu_addr;
1100 	props->hqd_pipe_priority = AMDGPU_GFX_PIPE_PRIO_NORMAL;
1101 	props->hqd_queue_priority = AMDGPU_GFX_QUEUE_PRIORITY_MINIMUM;
1102 	props->paging = false;
1103 	props->ring = ring;
1104 }
1105 
1106 #define DEFINE_AMDGPU_MES_CTX_GET_OFFS_ENG(_eng)			\
1107 do {									\
1108        if (id_offs < AMDGPU_MES_CTX_MAX_OFFS)				\
1109 		return offsetof(struct amdgpu_mes_ctx_meta_data,	\
1110 				_eng[ring->idx].slots[id_offs]);        \
1111        else if (id_offs == AMDGPU_MES_CTX_RING_OFFS)			\
1112 		return offsetof(struct amdgpu_mes_ctx_meta_data,        \
1113 				_eng[ring->idx].ring);                  \
1114        else if (id_offs == AMDGPU_MES_CTX_IB_OFFS)			\
1115 		return offsetof(struct amdgpu_mes_ctx_meta_data,        \
1116 				_eng[ring->idx].ib);                    \
1117        else if (id_offs == AMDGPU_MES_CTX_PADDING_OFFS)			\
1118 		return offsetof(struct amdgpu_mes_ctx_meta_data,        \
1119 				_eng[ring->idx].padding);               \
1120 } while(0)
1121 
1122 int amdgpu_mes_ctx_get_offs(struct amdgpu_ring *ring, unsigned int id_offs)
1123 {
1124 	switch (ring->funcs->type) {
1125 	case AMDGPU_RING_TYPE_GFX:
1126 		DEFINE_AMDGPU_MES_CTX_GET_OFFS_ENG(gfx);
1127 		break;
1128 	case AMDGPU_RING_TYPE_COMPUTE:
1129 		DEFINE_AMDGPU_MES_CTX_GET_OFFS_ENG(compute);
1130 		break;
1131 	case AMDGPU_RING_TYPE_SDMA:
1132 		DEFINE_AMDGPU_MES_CTX_GET_OFFS_ENG(sdma);
1133 		break;
1134 	default:
1135 		break;
1136 	}
1137 
1138 	WARN_ON(1);
1139 	return -EINVAL;
1140 }
1141 
1142 int amdgpu_mes_add_ring(struct amdgpu_device *adev, int gang_id,
1143 			int queue_type, int idx,
1144 			struct amdgpu_mes_ctx_data *ctx_data,
1145 			struct amdgpu_ring **out)
1146 {
1147 	struct amdgpu_ring *ring;
1148 	struct amdgpu_mes_gang *gang;
1149 	struct amdgpu_mes_queue_properties qprops = {0};
1150 	int r, queue_id, pasid;
1151 
1152 	/*
1153 	 * Avoid taking any other locks under MES lock to avoid circular
1154 	 * lock dependencies.
1155 	 */
1156 	amdgpu_mes_lock(&adev->mes);
1157 	gang = idr_find(&adev->mes.gang_id_idr, gang_id);
1158 	if (!gang) {
1159 		DRM_ERROR("gang id %d doesn't exist\n", gang_id);
1160 		amdgpu_mes_unlock(&adev->mes);
1161 		return -EINVAL;
1162 	}
1163 	pasid = gang->process->pasid;
1164 
1165 	ring = kzalloc(sizeof(struct amdgpu_ring), GFP_KERNEL);
1166 	if (!ring) {
1167 		amdgpu_mes_unlock(&adev->mes);
1168 		return -ENOMEM;
1169 	}
1170 
1171 	ring->ring_obj = NULL;
1172 	ring->use_doorbell = true;
1173 	ring->is_mes_queue = true;
1174 	ring->mes_ctx = ctx_data;
1175 	ring->idx = idx;
1176 	ring->no_scheduler = true;
1177 
1178 	if (queue_type == AMDGPU_RING_TYPE_COMPUTE) {
1179 		int offset = offsetof(struct amdgpu_mes_ctx_meta_data,
1180 				      compute[ring->idx].mec_hpd);
1181 		ring->eop_gpu_addr =
1182 			amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
1183 	}
1184 
1185 	switch (queue_type) {
1186 	case AMDGPU_RING_TYPE_GFX:
1187 		ring->funcs = adev->gfx.gfx_ring[0].funcs;
1188 		ring->me = adev->gfx.gfx_ring[0].me;
1189 		ring->pipe = adev->gfx.gfx_ring[0].pipe;
1190 		break;
1191 	case AMDGPU_RING_TYPE_COMPUTE:
1192 		ring->funcs = adev->gfx.compute_ring[0].funcs;
1193 		ring->me = adev->gfx.compute_ring[0].me;
1194 		ring->pipe = adev->gfx.compute_ring[0].pipe;
1195 		break;
1196 	case AMDGPU_RING_TYPE_SDMA:
1197 		ring->funcs = adev->sdma.instance[0].ring.funcs;
1198 		break;
1199 	default:
1200 		BUG();
1201 	}
1202 
1203 	r = amdgpu_ring_init(adev, ring, 1024, NULL, 0,
1204 			     AMDGPU_RING_PRIO_DEFAULT, NULL);
1205 	if (r)
1206 		goto clean_up_memory;
1207 
1208 	amdgpu_mes_ring_to_queue_props(adev, ring, &qprops);
1209 
1210 	dma_fence_wait(gang->process->vm->last_update, false);
1211 	dma_fence_wait(ctx_data->meta_data_va->last_pt_update, false);
1212 	amdgpu_mes_unlock(&adev->mes);
1213 
1214 	r = amdgpu_mes_add_hw_queue(adev, gang_id, &qprops, &queue_id);
1215 	if (r)
1216 		goto clean_up_ring;
1217 
1218 	ring->hw_queue_id = queue_id;
1219 	ring->doorbell_index = qprops.doorbell_off;
1220 
1221 	if (queue_type == AMDGPU_RING_TYPE_GFX)
1222 		sprintf(ring->name, "gfx_%d.%d.%d", pasid, gang_id, queue_id);
1223 	else if (queue_type == AMDGPU_RING_TYPE_COMPUTE)
1224 		sprintf(ring->name, "compute_%d.%d.%d", pasid, gang_id,
1225 			queue_id);
1226 	else if (queue_type == AMDGPU_RING_TYPE_SDMA)
1227 		sprintf(ring->name, "sdma_%d.%d.%d", pasid, gang_id,
1228 			queue_id);
1229 	else
1230 		BUG();
1231 
1232 	*out = ring;
1233 	return 0;
1234 
1235 clean_up_ring:
1236 	amdgpu_ring_fini(ring);
1237 clean_up_memory:
1238 	kfree(ring);
1239 	amdgpu_mes_unlock(&adev->mes);
1240 	return r;
1241 }
1242 
1243 void amdgpu_mes_remove_ring(struct amdgpu_device *adev,
1244 			    struct amdgpu_ring *ring)
1245 {
1246 	if (!ring)
1247 		return;
1248 
1249 	amdgpu_mes_remove_hw_queue(adev, ring->hw_queue_id);
1250 	del_timer_sync(&ring->fence_drv.fallback_timer);
1251 	amdgpu_ring_fini(ring);
1252 	kfree(ring);
1253 }
1254 
1255 uint32_t amdgpu_mes_get_aggregated_doorbell_index(struct amdgpu_device *adev,
1256 						   enum amdgpu_mes_priority_level prio)
1257 {
1258 	return adev->mes.aggregated_doorbells[prio];
1259 }
1260 
1261 int amdgpu_mes_ctx_alloc_meta_data(struct amdgpu_device *adev,
1262 				   struct amdgpu_mes_ctx_data *ctx_data)
1263 {
1264 	int r;
1265 
1266 	r = amdgpu_bo_create_kernel(adev,
1267 			    sizeof(struct amdgpu_mes_ctx_meta_data),
1268 			    PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
1269 			    &ctx_data->meta_data_obj,
1270 			    &ctx_data->meta_data_mc_addr,
1271 			    &ctx_data->meta_data_ptr);
1272 	if (r) {
1273 		dev_warn(adev->dev, "(%d) create CTX bo failed\n", r);
1274 		return r;
1275 	}
1276 
1277 	if (!ctx_data->meta_data_obj)
1278 		return -ENOMEM;
1279 
1280 	memset(ctx_data->meta_data_ptr, 0,
1281 	       sizeof(struct amdgpu_mes_ctx_meta_data));
1282 
1283 	return 0;
1284 }
1285 
1286 void amdgpu_mes_ctx_free_meta_data(struct amdgpu_mes_ctx_data *ctx_data)
1287 {
1288 	if (ctx_data->meta_data_obj)
1289 		amdgpu_bo_free_kernel(&ctx_data->meta_data_obj,
1290 				      &ctx_data->meta_data_mc_addr,
1291 				      &ctx_data->meta_data_ptr);
1292 }
1293 
1294 int amdgpu_mes_ctx_map_meta_data(struct amdgpu_device *adev,
1295 				 struct amdgpu_vm *vm,
1296 				 struct amdgpu_mes_ctx_data *ctx_data)
1297 {
1298 	struct amdgpu_bo_va *bo_va;
1299 	struct amdgpu_sync sync;
1300 	struct drm_exec exec;
1301 	int r;
1302 
1303 	amdgpu_sync_create(&sync);
1304 
1305 	drm_exec_init(&exec, 0, 0);
1306 	drm_exec_until_all_locked(&exec) {
1307 		r = drm_exec_lock_obj(&exec,
1308 				      &ctx_data->meta_data_obj->tbo.base);
1309 		drm_exec_retry_on_contention(&exec);
1310 		if (unlikely(r))
1311 			goto error_fini_exec;
1312 
1313 		r = amdgpu_vm_lock_pd(vm, &exec, 0);
1314 		drm_exec_retry_on_contention(&exec);
1315 		if (unlikely(r))
1316 			goto error_fini_exec;
1317 	}
1318 
1319 	bo_va = amdgpu_vm_bo_add(adev, vm, ctx_data->meta_data_obj);
1320 	if (!bo_va) {
1321 		DRM_ERROR("failed to create bo_va for meta data BO\n");
1322 		r = -ENOMEM;
1323 		goto error_fini_exec;
1324 	}
1325 
1326 	r = amdgpu_vm_bo_map(adev, bo_va, ctx_data->meta_data_gpu_addr, 0,
1327 			     sizeof(struct amdgpu_mes_ctx_meta_data),
1328 			     AMDGPU_PTE_READABLE | AMDGPU_PTE_WRITEABLE |
1329 			     AMDGPU_PTE_EXECUTABLE);
1330 
1331 	if (r) {
1332 		DRM_ERROR("failed to do bo_map on meta data, err=%d\n", r);
1333 		goto error_del_bo_va;
1334 	}
1335 
1336 	r = amdgpu_vm_bo_update(adev, bo_va, false);
1337 	if (r) {
1338 		DRM_ERROR("failed to do vm_bo_update on meta data\n");
1339 		goto error_del_bo_va;
1340 	}
1341 	amdgpu_sync_fence(&sync, bo_va->last_pt_update);
1342 
1343 	r = amdgpu_vm_update_pdes(adev, vm, false);
1344 	if (r) {
1345 		DRM_ERROR("failed to update pdes on meta data\n");
1346 		goto error_del_bo_va;
1347 	}
1348 	amdgpu_sync_fence(&sync, vm->last_update);
1349 
1350 	amdgpu_sync_wait(&sync, false);
1351 	drm_exec_fini(&exec);
1352 
1353 	amdgpu_sync_free(&sync);
1354 	ctx_data->meta_data_va = bo_va;
1355 	return 0;
1356 
1357 error_del_bo_va:
1358 	amdgpu_vm_bo_del(adev, bo_va);
1359 
1360 error_fini_exec:
1361 	drm_exec_fini(&exec);
1362 	amdgpu_sync_free(&sync);
1363 	return r;
1364 }
1365 
1366 int amdgpu_mes_ctx_unmap_meta_data(struct amdgpu_device *adev,
1367 				   struct amdgpu_mes_ctx_data *ctx_data)
1368 {
1369 	struct amdgpu_bo_va *bo_va = ctx_data->meta_data_va;
1370 	struct amdgpu_bo *bo = ctx_data->meta_data_obj;
1371 	struct amdgpu_vm *vm = bo_va->base.vm;
1372 	struct dma_fence *fence;
1373 	struct drm_exec exec;
1374 	long r;
1375 
1376 	drm_exec_init(&exec, 0, 0);
1377 	drm_exec_until_all_locked(&exec) {
1378 		r = drm_exec_lock_obj(&exec,
1379 				      &ctx_data->meta_data_obj->tbo.base);
1380 		drm_exec_retry_on_contention(&exec);
1381 		if (unlikely(r))
1382 			goto out_unlock;
1383 
1384 		r = amdgpu_vm_lock_pd(vm, &exec, 0);
1385 		drm_exec_retry_on_contention(&exec);
1386 		if (unlikely(r))
1387 			goto out_unlock;
1388 	}
1389 
1390 	amdgpu_vm_bo_del(adev, bo_va);
1391 	if (!amdgpu_vm_ready(vm))
1392 		goto out_unlock;
1393 
1394 	r = dma_resv_get_singleton(bo->tbo.base.resv, DMA_RESV_USAGE_BOOKKEEP,
1395 				   &fence);
1396 	if (r)
1397 		goto out_unlock;
1398 	if (fence) {
1399 		amdgpu_bo_fence(bo, fence, true);
1400 		fence = NULL;
1401 	}
1402 
1403 	r = amdgpu_vm_clear_freed(adev, vm, &fence);
1404 	if (r || !fence)
1405 		goto out_unlock;
1406 
1407 	dma_fence_wait(fence, false);
1408 	amdgpu_bo_fence(bo, fence, true);
1409 	dma_fence_put(fence);
1410 
1411 out_unlock:
1412 	if (unlikely(r < 0))
1413 		dev_err(adev->dev, "failed to clear page tables (%ld)\n", r);
1414 	drm_exec_fini(&exec);
1415 
1416 	return r;
1417 }
1418 
1419 static int amdgpu_mes_test_create_gang_and_queues(struct amdgpu_device *adev,
1420 					  int pasid, int *gang_id,
1421 					  int queue_type, int num_queue,
1422 					  struct amdgpu_ring **added_rings,
1423 					  struct amdgpu_mes_ctx_data *ctx_data)
1424 {
1425 	struct amdgpu_ring *ring;
1426 	struct amdgpu_mes_gang_properties gprops = {0};
1427 	int r, j;
1428 
1429 	/* create a gang for the process */
1430 	gprops.priority = AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
1431 	gprops.gang_quantum = adev->mes.default_gang_quantum;
1432 	gprops.inprocess_gang_priority = AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
1433 	gprops.priority_level = AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
1434 	gprops.global_priority_level = AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
1435 
1436 	r = amdgpu_mes_add_gang(adev, pasid, &gprops, gang_id);
1437 	if (r) {
1438 		DRM_ERROR("failed to add gang\n");
1439 		return r;
1440 	}
1441 
1442 	/* create queues for the gang */
1443 	for (j = 0; j < num_queue; j++) {
1444 		r = amdgpu_mes_add_ring(adev, *gang_id, queue_type, j,
1445 					ctx_data, &ring);
1446 		if (r) {
1447 			DRM_ERROR("failed to add ring\n");
1448 			break;
1449 		}
1450 
1451 		DRM_INFO("ring %s was added\n", ring->name);
1452 		added_rings[j] = ring;
1453 	}
1454 
1455 	return 0;
1456 }
1457 
1458 static int amdgpu_mes_test_queues(struct amdgpu_ring **added_rings)
1459 {
1460 	struct amdgpu_ring *ring;
1461 	int i, r;
1462 
1463 	for (i = 0; i < AMDGPU_MES_CTX_MAX_RINGS; i++) {
1464 		ring = added_rings[i];
1465 		if (!ring)
1466 			continue;
1467 
1468 		r = amdgpu_ring_test_helper(ring);
1469 		if (r)
1470 			return r;
1471 
1472 		r = amdgpu_ring_test_ib(ring, 1000 * 10);
1473 		if (r) {
1474 			DRM_DEV_ERROR(ring->adev->dev,
1475 				      "ring %s ib test failed (%d)\n",
1476 				      ring->name, r);
1477 			return r;
1478 		} else
1479 			DRM_INFO("ring %s ib test pass\n", ring->name);
1480 	}
1481 
1482 	return 0;
1483 }
1484 
1485 int amdgpu_mes_self_test(struct amdgpu_device *adev)
1486 {
1487 	struct amdgpu_vm *vm = NULL;
1488 	struct amdgpu_mes_ctx_data ctx_data = {0};
1489 	struct amdgpu_ring *added_rings[AMDGPU_MES_CTX_MAX_RINGS] = { NULL };
1490 	int gang_ids[3] = {0};
1491 	int queue_types[][2] = { { AMDGPU_RING_TYPE_GFX, 1 },
1492 				 { AMDGPU_RING_TYPE_COMPUTE, 1 },
1493 				 { AMDGPU_RING_TYPE_SDMA, 1} };
1494 	int i, r, pasid, k = 0;
1495 
1496 	pasid = amdgpu_pasid_alloc(16);
1497 	if (pasid < 0) {
1498 		dev_warn(adev->dev, "No more PASIDs available!");
1499 		pasid = 0;
1500 	}
1501 
1502 	vm = kzalloc(sizeof(*vm), GFP_KERNEL);
1503 	if (!vm) {
1504 		r = -ENOMEM;
1505 		goto error_pasid;
1506 	}
1507 
1508 	r = amdgpu_vm_init(adev, vm, -1);
1509 	if (r) {
1510 		DRM_ERROR("failed to initialize vm\n");
1511 		goto error_pasid;
1512 	}
1513 
1514 	r = amdgpu_mes_ctx_alloc_meta_data(adev, &ctx_data);
1515 	if (r) {
1516 		DRM_ERROR("failed to alloc ctx meta data\n");
1517 		goto error_fini;
1518 	}
1519 
1520 	ctx_data.meta_data_gpu_addr = AMDGPU_VA_RESERVED_BOTTOM;
1521 	r = amdgpu_mes_ctx_map_meta_data(adev, vm, &ctx_data);
1522 	if (r) {
1523 		DRM_ERROR("failed to map ctx meta data\n");
1524 		goto error_vm;
1525 	}
1526 
1527 	r = amdgpu_mes_create_process(adev, pasid, vm);
1528 	if (r) {
1529 		DRM_ERROR("failed to create MES process\n");
1530 		goto error_vm;
1531 	}
1532 
1533 	for (i = 0; i < ARRAY_SIZE(queue_types); i++) {
1534 		/* On GFX v10.3, fw hasn't supported to map sdma queue. */
1535 		if (amdgpu_ip_version(adev, GC_HWIP, 0) >=
1536 			    IP_VERSION(10, 3, 0) &&
1537 		    amdgpu_ip_version(adev, GC_HWIP, 0) <
1538 			    IP_VERSION(11, 0, 0) &&
1539 		    queue_types[i][0] == AMDGPU_RING_TYPE_SDMA)
1540 			continue;
1541 
1542 		r = amdgpu_mes_test_create_gang_and_queues(adev, pasid,
1543 							   &gang_ids[i],
1544 							   queue_types[i][0],
1545 							   queue_types[i][1],
1546 							   &added_rings[k],
1547 							   &ctx_data);
1548 		if (r)
1549 			goto error_queues;
1550 
1551 		k += queue_types[i][1];
1552 	}
1553 
1554 	/* start ring test and ib test for MES queues */
1555 	amdgpu_mes_test_queues(added_rings);
1556 
1557 error_queues:
1558 	/* remove all queues */
1559 	for (i = 0; i < ARRAY_SIZE(added_rings); i++) {
1560 		if (!added_rings[i])
1561 			continue;
1562 		amdgpu_mes_remove_ring(adev, added_rings[i]);
1563 	}
1564 
1565 	for (i = 0; i < ARRAY_SIZE(gang_ids); i++) {
1566 		if (!gang_ids[i])
1567 			continue;
1568 		amdgpu_mes_remove_gang(adev, gang_ids[i]);
1569 	}
1570 
1571 	amdgpu_mes_destroy_process(adev, pasid);
1572 
1573 error_vm:
1574 	amdgpu_mes_ctx_unmap_meta_data(adev, &ctx_data);
1575 
1576 error_fini:
1577 	amdgpu_vm_fini(adev, vm);
1578 
1579 error_pasid:
1580 	if (pasid)
1581 		amdgpu_pasid_free(pasid);
1582 
1583 	amdgpu_mes_ctx_free_meta_data(&ctx_data);
1584 	kfree(vm);
1585 	return 0;
1586 }
1587 
1588 int amdgpu_mes_init_microcode(struct amdgpu_device *adev, int pipe)
1589 {
1590 	const struct mes_firmware_header_v1_0 *mes_hdr;
1591 	struct amdgpu_firmware_info *info;
1592 	char ucode_prefix[30];
1593 	char fw_name[50];
1594 	bool need_retry = false;
1595 	int r;
1596 
1597 	amdgpu_ucode_ip_version_decode(adev, GC_HWIP, ucode_prefix,
1598 				       sizeof(ucode_prefix));
1599 	if (adev->enable_uni_mes) {
1600 		snprintf(fw_name, sizeof(fw_name),
1601 			 "amdgpu/%s_uni_mes.bin", ucode_prefix);
1602 	} else if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(11, 0, 0) &&
1603 	    amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(12, 0, 0)) {
1604 		snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes%s.bin",
1605 			 ucode_prefix,
1606 			 pipe == AMDGPU_MES_SCHED_PIPE ? "_2" : "1");
1607 		need_retry = true;
1608 	} else {
1609 		snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes%s.bin",
1610 			 ucode_prefix,
1611 			 pipe == AMDGPU_MES_SCHED_PIPE ? "" : "1");
1612 	}
1613 
1614 	r = amdgpu_ucode_request(adev, &adev->mes.fw[pipe], "%s", fw_name);
1615 	if (r && need_retry && pipe == AMDGPU_MES_SCHED_PIPE) {
1616 		dev_info(adev->dev, "try to fall back to %s_mes.bin\n", ucode_prefix);
1617 		r = amdgpu_ucode_request(adev, &adev->mes.fw[pipe],
1618 					 "amdgpu/%s_mes.bin", ucode_prefix);
1619 	}
1620 
1621 	if (r)
1622 		goto out;
1623 
1624 	mes_hdr = (const struct mes_firmware_header_v1_0 *)
1625 		adev->mes.fw[pipe]->data;
1626 	adev->mes.uc_start_addr[pipe] =
1627 		le32_to_cpu(mes_hdr->mes_uc_start_addr_lo) |
1628 		((uint64_t)(le32_to_cpu(mes_hdr->mes_uc_start_addr_hi)) << 32);
1629 	adev->mes.data_start_addr[pipe] =
1630 		le32_to_cpu(mes_hdr->mes_data_start_addr_lo) |
1631 		((uint64_t)(le32_to_cpu(mes_hdr->mes_data_start_addr_hi)) << 32);
1632 
1633 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
1634 		int ucode, ucode_data;
1635 
1636 		if (pipe == AMDGPU_MES_SCHED_PIPE) {
1637 			ucode = AMDGPU_UCODE_ID_CP_MES;
1638 			ucode_data = AMDGPU_UCODE_ID_CP_MES_DATA;
1639 		} else {
1640 			ucode = AMDGPU_UCODE_ID_CP_MES1;
1641 			ucode_data = AMDGPU_UCODE_ID_CP_MES1_DATA;
1642 		}
1643 
1644 		info = &adev->firmware.ucode[ucode];
1645 		info->ucode_id = ucode;
1646 		info->fw = adev->mes.fw[pipe];
1647 		adev->firmware.fw_size +=
1648 			ALIGN(le32_to_cpu(mes_hdr->mes_ucode_size_bytes),
1649 			      PAGE_SIZE);
1650 
1651 		info = &adev->firmware.ucode[ucode_data];
1652 		info->ucode_id = ucode_data;
1653 		info->fw = adev->mes.fw[pipe];
1654 		adev->firmware.fw_size +=
1655 			ALIGN(le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes),
1656 			      PAGE_SIZE);
1657 	}
1658 
1659 	return 0;
1660 out:
1661 	amdgpu_ucode_release(&adev->mes.fw[pipe]);
1662 	return r;
1663 }
1664 
1665 bool amdgpu_mes_suspend_resume_all_supported(struct amdgpu_device *adev)
1666 {
1667 	uint32_t mes_rev = adev->mes.sched_version & AMDGPU_MES_VERSION_MASK;
1668 	bool is_supported = false;
1669 
1670 	if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(11, 0, 0) &&
1671 	    amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(12, 0, 0) &&
1672 	    mes_rev >= 0x63)
1673 		is_supported = true;
1674 
1675 	return is_supported;
1676 }
1677 
1678 #if defined(CONFIG_DEBUG_FS)
1679 
1680 static int amdgpu_debugfs_mes_event_log_show(struct seq_file *m, void *unused)
1681 {
1682 	struct amdgpu_device *adev = m->private;
1683 	uint32_t *mem = (uint32_t *)(adev->mes.event_log_cpu_addr);
1684 
1685 	seq_hex_dump(m, "", DUMP_PREFIX_OFFSET, 32, 4,
1686 		     mem, adev->mes.event_log_size, false);
1687 
1688 	return 0;
1689 }
1690 
1691 DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_mes_event_log);
1692 
1693 #endif
1694 
1695 void amdgpu_debugfs_mes_event_log_init(struct amdgpu_device *adev)
1696 {
1697 
1698 #if defined(CONFIG_DEBUG_FS)
1699 	struct drm_minor *minor = adev_to_drm(adev)->primary;
1700 	struct dentry *root = minor->debugfs_root;
1701 	if (adev->enable_mes && amdgpu_mes_log_enable)
1702 		debugfs_create_file("amdgpu_mes_event_log", 0444, root,
1703 				    adev, &amdgpu_debugfs_mes_event_log_fops);
1704 
1705 #endif
1706 }
1707