xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c (revision 570f7e331f5febb30f1384817463c7e42b65ca7d)
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_doorbell_init(struct amdgpu_device *adev)
43 {
44 	int i;
45 	struct amdgpu_mes *mes = &adev->mes;
46 
47 	/* Bitmap for dynamic allocation of kernel doorbells */
48 	mes->doorbell_bitmap = bitmap_zalloc(PAGE_SIZE / sizeof(u32), GFP_KERNEL);
49 	if (!mes->doorbell_bitmap) {
50 		dev_err(adev->dev, "Failed to allocate MES doorbell bitmap\n");
51 		return -ENOMEM;
52 	}
53 
54 	mes->num_mes_dbs = PAGE_SIZE / AMDGPU_ONE_DOORBELL_SIZE;
55 	for (i = 0; i < AMDGPU_MES_PRIORITY_NUM_LEVELS; i++) {
56 		adev->mes.aggregated_doorbells[i] = mes->db_start_dw_offset + i * 2;
57 		set_bit(i, mes->doorbell_bitmap);
58 	}
59 
60 	return 0;
61 }
62 
63 static int amdgpu_mes_event_log_init(struct amdgpu_device *adev)
64 {
65 	int r;
66 
67 	if (!amdgpu_mes_log_enable)
68 		return 0;
69 
70 	r = amdgpu_bo_create_kernel(adev, adev->mes.event_log_size, PAGE_SIZE,
71 				    AMDGPU_GEM_DOMAIN_VRAM,
72 				    &adev->mes.event_log_gpu_obj,
73 				    &adev->mes.event_log_gpu_addr,
74 				    &adev->mes.event_log_cpu_addr);
75 	if (r) {
76 		dev_warn(adev->dev, "failed to create MES event log buffer (%d)", r);
77 		return r;
78 	}
79 
80 	memset(adev->mes.event_log_cpu_addr, 0, adev->mes.event_log_size);
81 
82 	return  0;
83 
84 }
85 
86 static void amdgpu_mes_doorbell_free(struct amdgpu_device *adev)
87 {
88 	bitmap_free(adev->mes.doorbell_bitmap);
89 }
90 
91 static inline u32 amdgpu_mes_get_hqd_mask(u32 num_pipe,
92 					  u32 num_hqd_per_pipe,
93 					  u32 num_reserved_hqd)
94 {
95 	if (num_pipe == 0)
96 		return 0;
97 
98 	u32 total_hqd_mask = (u32)((1ULL << num_hqd_per_pipe) - 1);
99 	u32 reserved_hqd_mask = (u32)((1ULL << DIV_ROUND_UP(num_reserved_hqd, num_pipe)) - 1);
100 
101 	return (total_hqd_mask & ~reserved_hqd_mask);
102 }
103 
104 int amdgpu_mes_init(struct amdgpu_device *adev)
105 {
106 	int i, r, num_pipes, num_queues = 0;
107 	u32 total_vmid_mask, reserved_vmid_mask;
108 	int num_xcc = adev->gfx.xcc_mask ? NUM_XCC(adev->gfx.xcc_mask) : 1;
109 	u32 gfx_hqd_mask = amdgpu_mes_get_hqd_mask(adev->gfx.me.num_pipe_per_me,
110 				adev->gfx.me.num_queue_per_pipe,
111 				adev->gfx.disable_kq ? 0 : adev->gfx.num_gfx_rings);
112 	u32 compute_hqd_mask = amdgpu_mes_get_hqd_mask(adev->gfx.mec.num_pipe_per_mec,
113 				adev->gfx.mec.num_queue_per_pipe,
114 				adev->gfx.disable_kq ? 0 : adev->gfx.num_compute_rings);
115 
116 	adev->mes.adev = adev;
117 
118 	ida_init(&adev->mes.doorbell_ida);
119 	spin_lock_init(&adev->mes.queue_id_lock);
120 	mutex_init(&adev->mes.mutex_hidden);
121 
122 	for (i = 0; i < AMDGPU_MAX_MES_PIPES * num_xcc; i++)
123 		spin_lock_init(&adev->mes.ring_lock[i]);
124 
125 	adev->mes.total_max_queue = AMDGPU_FENCE_MES_QUEUE_ID_MASK;
126 	total_vmid_mask = (u32)((1UL << 16) - 1);
127 	reserved_vmid_mask = (u32)((1UL << adev->vm_manager.first_kfd_vmid) - 1);
128 
129 	adev->mes.vmid_mask_mmhub = 0xFF00;
130 	adev->mes.vmid_mask_gfxhub = total_vmid_mask & ~reserved_vmid_mask;
131 
132 	num_pipes = adev->gfx.me.num_pipe_per_me * adev->gfx.me.num_me;
133 	if (num_pipes > AMDGPU_MES_MAX_GFX_PIPES)
134 		dev_warn(adev->dev, "more gfx pipes than supported by MES! (%d vs %d)\n",
135 			 num_pipes, AMDGPU_MES_MAX_GFX_PIPES);
136 
137 	for (i = 0; i < AMDGPU_MES_MAX_GFX_PIPES; i++) {
138 		if (i >= num_pipes)
139 			break;
140 
141 		adev->mes.gfx_hqd_mask[i] = gfx_hqd_mask;
142 	}
143 
144 	num_pipes = adev->gfx.mec.num_pipe_per_mec * adev->gfx.mec.num_mec;
145 	if (num_pipes > AMDGPU_MES_MAX_COMPUTE_PIPES)
146 		dev_warn(adev->dev, "more compute pipes than supported by MES! (%d vs %d)\n",
147 			 num_pipes, AMDGPU_MES_MAX_COMPUTE_PIPES);
148 
149 	for (i = 0; i < AMDGPU_MES_MAX_COMPUTE_PIPES; i++) {
150 		/*
151 		 * Currently, only MEC1 is used for both kernel and user compute queue.
152 		 * To enable other MEC, we need to redistribute queues per pipe and
153 		 * adjust queue resource shared with kfd that needs a separate patch.
154 		 * Skip other MEC for now to avoid potential issues.
155 		 */
156 		if (i >= adev->gfx.mec.num_pipe_per_mec)
157 			break;
158 
159 		adev->mes.compute_hqd_mask[i] = compute_hqd_mask;
160 	}
161 
162 	num_pipes = adev->sdma.num_inst_per_xcc ?
163 		adev->sdma.num_inst_per_xcc : adev->sdma.num_instances;
164 	if (num_pipes > AMDGPU_MES_MAX_SDMA_PIPES)
165 		dev_warn(adev->dev, "more SDMA pipes than supported by MES! (%d vs %d)\n",
166 			 num_pipes, AMDGPU_MES_MAX_SDMA_PIPES);
167 
168 	for (i = 0; i < AMDGPU_MES_MAX_SDMA_PIPES; i++) {
169 		if (i >= num_pipes)
170 			break;
171 		adev->mes.sdma_hqd_mask[i] = 0xfc;
172 	}
173 
174 	dev_info(adev->dev,
175 			 "MES: vmid_mask_mmhub 0x%08x, vmid_mask_gfxhub 0x%08x\n",
176 			 adev->mes.vmid_mask_mmhub,
177 			 adev->mes.vmid_mask_gfxhub);
178 
179 	dev_info(adev->dev,
180 			 "MES: gfx_hqd_mask 0x%08x, compute_hqd_mask 0x%08x, sdma_hqd_mask 0x%08x\n",
181 			 adev->mes.gfx_hqd_mask[0],
182 			 adev->mes.compute_hqd_mask[0],
183 			 adev->mes.sdma_hqd_mask[0]);
184 
185 	for (i = 0; i < AMDGPU_MAX_MES_PIPES * num_xcc; i++) {
186 		r = amdgpu_wb_get(adev, &adev->mes.sch_ctx_offs[i]);
187 		if (r) {
188 			dev_err(adev->dev,
189 				"(%d) ring trail_fence_offs wb alloc failed\n",
190 				r);
191 			goto error;
192 		}
193 		adev->mes.sch_ctx_gpu_addr[i] =
194 			adev->wb.gpu_addr + (adev->mes.sch_ctx_offs[i] * 4);
195 		adev->mes.sch_ctx_ptr[i] =
196 			(uint64_t *)&adev->wb.wb[adev->mes.sch_ctx_offs[i]];
197 
198 		r = amdgpu_wb_get(adev,
199 				 &adev->mes.query_status_fence_offs[i]);
200 		if (r) {
201 			dev_err(adev->dev,
202 			      "(%d) query_status_fence_offs wb alloc failed\n",
203 			      r);
204 			goto error;
205 		}
206 		adev->mes.query_status_fence_gpu_addr[i] = adev->wb.gpu_addr +
207 			(adev->mes.query_status_fence_offs[i] * 4);
208 		adev->mes.query_status_fence_ptr[i] =
209 			(uint64_t *)&adev->wb.wb[adev->mes.query_status_fence_offs[i]];
210 	}
211 
212 	r = amdgpu_mes_doorbell_init(adev);
213 	if (r)
214 		goto error;
215 
216 	r = amdgpu_mes_event_log_init(adev);
217 	if (r)
218 		goto error_doorbell;
219 
220 	if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(11, 0, 0)) {
221 		/* When queue/pipe reset is done in MES instead of in the
222 		 * driver, MES passes hung queues information to the driver in
223 		 * hung_queue_hqd_info. Calculate required space to store this
224 		 * information.
225 		 */
226 		for (i = 0; i < AMDGPU_MES_MAX_GFX_PIPES; i++)
227 			num_queues += hweight32(adev->mes.gfx_hqd_mask[i]);
228 
229 		for (i = 0; i < AMDGPU_MES_MAX_COMPUTE_PIPES; i++)
230 			num_queues += hweight32(adev->mes.compute_hqd_mask[i]);
231 
232 		for (i = 0; i < AMDGPU_MES_MAX_SDMA_PIPES; i++)
233 			num_queues += hweight32(adev->mes.sdma_hqd_mask[i]) * num_xcc;
234 
235 		adev->mes.hung_queue_hqd_info_offset = num_queues;
236 		adev->mes.hung_queue_db_array_size = num_queues * 2;
237 	}
238 
239 	if (adev->mes.hung_queue_db_array_size) {
240 		for (i = 0; i < AMDGPU_MAX_MES_PIPES * num_xcc; i++) {
241 			r = amdgpu_bo_create_kernel(adev,
242 						    adev->mes.hung_queue_db_array_size * sizeof(u32),
243 						    PAGE_SIZE,
244 						    AMDGPU_GEM_DOMAIN_GTT,
245 						    &adev->mes.hung_queue_db_array_gpu_obj[i],
246 						    &adev->mes.hung_queue_db_array_gpu_addr[i],
247 						    &adev->mes.hung_queue_db_array_cpu_addr[i]);
248 			if (r) {
249 				dev_warn(adev->dev, "failed to create MES hung db array buffer (%d)", r);
250 				goto error_doorbell;
251 			}
252 		}
253 
254 		adev->gfx.mec.mes_hung_db_array =
255 			kcalloc(amdgpu_mes_get_hung_queue_db_array_size(adev),
256 				sizeof(u32), GFP_KERNEL);
257 
258 		if (!adev->gfx.mec.mes_hung_db_array) {
259 			r = -ENOMEM;
260 			goto error_doorbell;
261 		}
262 	}
263 
264 	return 0;
265 
266 error_doorbell:
267 	amdgpu_mes_doorbell_free(adev);
268 error:
269 	for (i = 0; i < AMDGPU_MAX_MES_PIPES * num_xcc; i++) {
270 		if (adev->mes.sch_ctx_ptr[i])
271 			amdgpu_wb_free(adev, adev->mes.sch_ctx_offs[i]);
272 		if (adev->mes.query_status_fence_ptr[i])
273 			amdgpu_wb_free(adev,
274 				      adev->mes.query_status_fence_offs[i]);
275 		if (adev->mes.hung_queue_db_array_gpu_obj[i])
276 			amdgpu_bo_free_kernel(&adev->mes.hung_queue_db_array_gpu_obj[i],
277 					      &adev->mes.hung_queue_db_array_gpu_addr[i],
278 					      &adev->mes.hung_queue_db_array_cpu_addr[i]);
279 	}
280 
281 	ida_destroy(&adev->mes.doorbell_ida);
282 	mutex_destroy(&adev->mes.mutex_hidden);
283 	return r;
284 }
285 
286 void amdgpu_mes_fini(struct amdgpu_device *adev)
287 {
288 	int i;
289 	int num_xcc = adev->gfx.xcc_mask ? NUM_XCC(adev->gfx.xcc_mask) : 1;
290 
291 	kfree(adev->gfx.mec.mes_hung_db_array);
292 
293 	amdgpu_bo_free_kernel(&adev->mes.event_log_gpu_obj,
294 			      &adev->mes.event_log_gpu_addr,
295 			      &adev->mes.event_log_cpu_addr);
296 
297 	for (i = 0; i < AMDGPU_MAX_MES_PIPES * num_xcc; i++) {
298 		if (adev->mes.hung_queue_db_array_gpu_obj[i])
299 			 amdgpu_bo_free_kernel(&adev->mes.hung_queue_db_array_gpu_obj[i],
300 					 &adev->mes.hung_queue_db_array_gpu_addr[i],
301 					 &adev->mes.hung_queue_db_array_cpu_addr[i]);
302 		if (adev->mes.sch_ctx_ptr[i])
303 			amdgpu_wb_free(adev, adev->mes.sch_ctx_offs[i]);
304 		if (adev->mes.query_status_fence_ptr[i])
305 			amdgpu_wb_free(adev,
306 				      adev->mes.query_status_fence_offs[i]);
307 	}
308 
309 	amdgpu_mes_doorbell_free(adev);
310 
311 	if (adev->mes.use_rs64mem)
312 		amdgpu_mes_rs64mem_fini(&adev->mes);
313 
314 	ida_destroy(&adev->mes.doorbell_ida);
315 	mutex_destroy(&adev->mes.mutex_hidden);
316 
317 }
318 
319 int amdgpu_mes_suspend(struct amdgpu_device *adev, u32 xcc_id)
320 {
321 	struct mes_suspend_gang_input input;
322 	int r;
323 
324 	if (!amdgpu_mes_suspend_resume_all_supported(adev))
325 		return 0;
326 
327 	memset(&input, 0x0, sizeof(struct mes_suspend_gang_input));
328 	input.suspend_all_gangs = 1;
329 	input.xcc_id = xcc_id;
330 	if ((amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(12, 1, 0)) &&
331 		((adev->mes.sched_version & AMDGPU_MES_VERSION_MASK) >= 0x71))
332 		input.suspend_all_sdma_gangs = 1;
333 
334 	/*
335 	 * Avoid taking any other locks under MES lock to avoid circular
336 	 * lock dependencies.
337 	 */
338 	amdgpu_mes_lock(&adev->mes);
339 	r = adev->mes.funcs->suspend_gang(&adev->mes, &input);
340 	amdgpu_mes_unlock(&adev->mes);
341 	if (r)
342 		dev_err(adev->dev, "failed to suspend all gangs");
343 
344 	return r;
345 }
346 
347 int amdgpu_mes_resume(struct amdgpu_device *adev, u32 xcc_id)
348 {
349 	struct mes_resume_gang_input input;
350 	int r;
351 
352 	if (!amdgpu_mes_suspend_resume_all_supported(adev))
353 		return 0;
354 
355 	memset(&input, 0x0, sizeof(struct mes_resume_gang_input));
356 	input.resume_all_gangs = 1;
357 	input.xcc_id = xcc_id;
358 
359 	/*
360 	 * Avoid taking any other locks under MES lock to avoid circular
361 	 * lock dependencies.
362 	 */
363 	amdgpu_mes_lock(&adev->mes);
364 	r = adev->mes.funcs->resume_gang(&adev->mes, &input);
365 	amdgpu_mes_unlock(&adev->mes);
366 	if (r)
367 		dev_err(adev->dev, "failed to resume all gangs");
368 
369 	return r;
370 }
371 
372 int amdgpu_mes_map_legacy_queue(struct amdgpu_device *adev,
373 				struct amdgpu_ring *ring, uint32_t xcc_id)
374 {
375 	struct mes_map_legacy_queue_input queue_input;
376 	int r;
377 
378 	memset(&queue_input, 0, sizeof(queue_input));
379 
380 	queue_input.xcc_id = xcc_id;
381 	queue_input.queue_type = ring->funcs->type;
382 	queue_input.doorbell_offset = ring->doorbell_index;
383 	queue_input.pipe_id = ring->pipe;
384 	queue_input.queue_id = ring->queue;
385 	queue_input.mqd_addr = amdgpu_bo_gpu_offset(ring->mqd_obj);
386 	queue_input.wptr_addr = ring->wptr_gpu_addr;
387 
388 	amdgpu_mes_lock(&adev->mes);
389 	r = adev->mes.funcs->map_legacy_queue(&adev->mes, &queue_input);
390 	amdgpu_mes_unlock(&adev->mes);
391 	if (r)
392 		dev_err(adev->dev, "failed to map legacy queue\n");
393 
394 	return r;
395 }
396 
397 int amdgpu_mes_unmap_legacy_queue(struct amdgpu_device *adev,
398 				  struct amdgpu_ring *ring,
399 				  enum amdgpu_unmap_queues_action action,
400 				  u64 gpu_addr, u64 seq, uint32_t xcc_id)
401 {
402 	struct mes_unmap_legacy_queue_input queue_input;
403 	int r;
404 
405 	queue_input.xcc_id = xcc_id;
406 	queue_input.action = action;
407 	queue_input.queue_type = ring->funcs->type;
408 	queue_input.doorbell_offset = ring->doorbell_index;
409 	queue_input.pipe_id = ring->pipe;
410 	queue_input.queue_id = ring->queue;
411 	queue_input.trail_fence_addr = gpu_addr;
412 	queue_input.trail_fence_data = seq;
413 
414 	amdgpu_mes_lock(&adev->mes);
415 	r = adev->mes.funcs->unmap_legacy_queue(&adev->mes, &queue_input);
416 	amdgpu_mes_unlock(&adev->mes);
417 	if (r)
418 		dev_err(adev->dev, "failed to unmap legacy queue\n");
419 
420 	return r;
421 }
422 
423 int amdgpu_mes_reset_legacy_queue(struct amdgpu_device *adev,
424 				  struct amdgpu_ring *ring,
425 				  unsigned int vmid,
426 				  bool use_mmio,
427 				  uint32_t xcc_id)
428 {
429 	struct mes_reset_queue_input queue_input;
430 	int r;
431 
432 	memset(&queue_input, 0, sizeof(queue_input));
433 
434 	queue_input.xcc_id = xcc_id;
435 	queue_input.queue_type = ring->funcs->type;
436 	queue_input.doorbell_offset = ring->doorbell_index;
437 	queue_input.me_id = ring->me;
438 	queue_input.pipe_id = ring->pipe;
439 	queue_input.queue_id = ring->queue;
440 	queue_input.mqd_addr = ring->mqd_obj ? amdgpu_bo_gpu_offset(ring->mqd_obj) : 0;
441 	queue_input.wptr_addr = ring->wptr_gpu_addr;
442 	queue_input.vmid = vmid;
443 	queue_input.use_mmio = use_mmio;
444 	queue_input.is_kq = true;
445 	if (ring->funcs->type == AMDGPU_RING_TYPE_GFX)
446 		queue_input.legacy_gfx = true;
447 
448 	amdgpu_mes_lock(&adev->mes);
449 	r = adev->mes.funcs->reset_hw_queue(&adev->mes, &queue_input);
450 	amdgpu_mes_unlock(&adev->mes);
451 	if (r)
452 		dev_err(adev->dev, "failed to reset legacy queue\n");
453 
454 	return r;
455 }
456 
457 int amdgpu_mes_reset_queue_mmio(struct amdgpu_device *adev,
458 				int queue_type,
459 				unsigned int vmid,
460 				unsigned int me,
461 				unsigned int pipe,
462 				unsigned int queue,
463 				uint32_t xcc_id)
464 {
465 	struct mes_reset_queue_input queue_input;
466 	int r;
467 
468 	memset(&queue_input, 0, sizeof(queue_input));
469 
470 	queue_input.xcc_id = xcc_id;
471 	queue_input.me_id = me;
472 	queue_input.pipe_id = pipe;
473 	queue_input.queue_id = queue;
474 	queue_input.vmid = vmid;
475 	queue_input.queue_type = queue_type;
476 	queue_input.use_mmio = true;
477 
478 	amdgpu_mes_lock(&adev->mes);
479 	r = adev->mes.funcs->reset_hw_queue(&adev->mes, &queue_input);
480 	amdgpu_mes_unlock(&adev->mes);
481 	if (r)
482 		dev_err(adev->dev, "failed to reset legacy queue\n");
483 
484 	return r;
485 }
486 
487 int amdgpu_mes_reset_user_queue(struct amdgpu_device *adev,
488 				int queue_type,
489 				unsigned int doorbell_index,
490 				unsigned int xcc_id)
491 {
492 	struct mes_reset_queue_input queue_input;
493 	int r;
494 
495 	memset(&queue_input, 0, sizeof(queue_input));
496 
497 	queue_input.xcc_id = xcc_id;
498 	queue_input.queue_type = queue_type;
499 	queue_input.doorbell_offset = doorbell_index;
500 
501 	amdgpu_mes_lock(&adev->mes);
502 	r = adev->mes.funcs->reset_hw_queue(&adev->mes, &queue_input);
503 	amdgpu_mes_unlock(&adev->mes);
504 	if (r)
505 		dev_err(adev->dev, "failed to reset user queue\n");
506 
507 	return r;
508 }
509 
510 int amdgpu_mes_get_hung_queue_db_array_size(struct amdgpu_device *adev)
511 {
512 	return adev->mes.hung_queue_db_array_size;
513 }
514 
515 int amdgpu_mes_detect_and_reset_hung_queues(struct amdgpu_device *adev,
516 					    int queue_type,
517 					    bool detect_only,
518 					    unsigned int *hung_db_num,
519 					    u32 *hung_db_array,
520 					    uint32_t xcc_id)
521 {
522 	struct mes_detect_and_reset_queue_input input;
523 	u32 *db_array = adev->mes.hung_queue_db_array_cpu_addr[xcc_id];
524 	int hqd_info_offset = adev->mes.hung_queue_hqd_info_offset, r, i;
525 
526 	if (!hung_db_num || !hung_db_array)
527 		return -EINVAL;
528 
529 	if ((queue_type != AMDGPU_RING_TYPE_GFX) &&
530 	    (queue_type != AMDGPU_RING_TYPE_COMPUTE) &&
531 	    (queue_type != AMDGPU_RING_TYPE_SDMA))
532 		return -EINVAL;
533 
534 	/* Clear the doorbell array before detection */
535 	memset(adev->mes.hung_queue_db_array_cpu_addr[xcc_id], AMDGPU_MES_INVALID_DB_OFFSET,
536 		adev->mes.hung_queue_db_array_size * sizeof(u32));
537 	input.queue_type = queue_type;
538 	input.detect_only = detect_only;
539 	input.xcc_id = xcc_id;
540 
541 	r = adev->mes.funcs->detect_and_reset_hung_queues(&adev->mes,
542 							  &input);
543 
544 	if (r && detect_only) {
545 		dev_err(adev->dev, "Failed to detect hung queues\n");
546 		return r;
547 	}
548 
549 	*hung_db_num = 0;
550 	/* MES passes hung queues' doorbell to driver */
551 	for (i = 0; i < adev->mes.hung_queue_hqd_info_offset; i++) {
552 		/* Finding hung queues where db_array[i] is a valid doorbell */
553 		if (db_array[i] != AMDGPU_MES_INVALID_DB_OFFSET) {
554 			hung_db_array[i] = db_array[i];
555 			*hung_db_num += 1;
556 		}
557 	}
558 
559 	if (r && !(*hung_db_num)) {
560 		dev_err(adev->dev, "Failed to detect and reset hung queues\n");
561 		return r;
562 	}
563 
564 	for (i = hqd_info_offset; i < hqd_info_offset + *hung_db_num; i++)
565 		hung_db_array[i] = db_array[i];
566 
567 	return r;
568 }
569 
570 uint32_t amdgpu_mes_rreg(struct amdgpu_device *adev, uint32_t reg,
571 			 uint32_t xcc_id)
572 {
573 	struct mes_misc_op_input op_input;
574 	int r, val = 0;
575 	uint32_t addr_offset = 0;
576 	uint64_t read_val_gpu_addr;
577 	uint32_t *read_val_ptr;
578 
579 	if (amdgpu_wb_get(adev, &addr_offset)) {
580 		dev_err(adev->dev, "critical bug! too many mes readers\n");
581 		goto error;
582 	}
583 	read_val_gpu_addr = adev->wb.gpu_addr + (addr_offset * 4);
584 	read_val_ptr = (uint32_t *)&adev->wb.wb[addr_offset];
585 	op_input.xcc_id = xcc_id;
586 	op_input.op = MES_MISC_OP_READ_REG;
587 	op_input.read_reg.reg_offset = reg;
588 	op_input.read_reg.buffer_addr = read_val_gpu_addr;
589 
590 	if (!adev->mes.funcs->misc_op) {
591 		dev_err(adev->dev, "mes rreg is not supported!\n");
592 		goto error;
593 	}
594 
595 	amdgpu_mes_lock(&adev->mes);
596 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
597 	amdgpu_mes_unlock(&adev->mes);
598 	if (r)
599 		dev_err(adev->dev, "failed to read reg (0x%x)\n", reg);
600 	else
601 		val = *(read_val_ptr);
602 
603 error:
604 	if (addr_offset)
605 		amdgpu_wb_free(adev, addr_offset);
606 	return val;
607 }
608 
609 int amdgpu_mes_wreg(struct amdgpu_device *adev, uint32_t reg,
610 		    uint32_t val, uint32_t xcc_id)
611 {
612 	struct mes_misc_op_input op_input;
613 	int r;
614 
615 	op_input.xcc_id = xcc_id;
616 	op_input.op = MES_MISC_OP_WRITE_REG;
617 	op_input.write_reg.reg_offset = reg;
618 	op_input.write_reg.reg_value = val;
619 
620 	if (!adev->mes.funcs->misc_op) {
621 		dev_err(adev->dev, "mes wreg is not supported!\n");
622 		r = -EINVAL;
623 		goto error;
624 	}
625 
626 	amdgpu_mes_lock(&adev->mes);
627 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
628 	amdgpu_mes_unlock(&adev->mes);
629 	if (r)
630 		dev_err(adev->dev, "failed to write reg (0x%x)\n", reg);
631 
632 error:
633 	return r;
634 }
635 
636 int amdgpu_mes_reg_write_reg_wait(struct amdgpu_device *adev,
637 				  uint32_t reg0, uint32_t reg1,
638 				  uint32_t ref, uint32_t mask,
639 				  uint32_t xcc_id)
640 {
641 	struct mes_misc_op_input op_input;
642 	int r;
643 
644 	op_input.xcc_id = xcc_id;
645 	op_input.op = MES_MISC_OP_WRM_REG_WR_WAIT;
646 	op_input.wrm_reg.reg0 = reg0;
647 	op_input.wrm_reg.reg1 = reg1;
648 	op_input.wrm_reg.ref = ref;
649 	op_input.wrm_reg.mask = mask;
650 
651 	if (!adev->mes.funcs->misc_op) {
652 		dev_err(adev->dev, "mes reg_write_reg_wait is not supported!\n");
653 		r = -EINVAL;
654 		goto error;
655 	}
656 
657 	amdgpu_mes_lock(&adev->mes);
658 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
659 	amdgpu_mes_unlock(&adev->mes);
660 	if (r)
661 		dev_err(adev->dev, "failed to reg_write_reg_wait\n");
662 
663 error:
664 	return r;
665 }
666 
667 int amdgpu_mes_hdp_flush(struct amdgpu_device *adev)
668 {
669 	uint32_t hdp_flush_req_offset, hdp_flush_done_offset;
670 	struct amdgpu_ring *mes_ring;
671 	uint32_t ref_and_mask = 0, reg_mem_engine = 0;
672 
673 	if (!adev->gfx.funcs->get_hdp_flush_mask) {
674 		dev_err(adev->dev, "mes hdp flush is not supported.\n");
675 		return -EINVAL;
676 	}
677 
678 	mes_ring = &adev->mes.ring[0];
679 	hdp_flush_req_offset = adev->nbio.funcs->get_hdp_flush_req_offset(adev);
680 	hdp_flush_done_offset = adev->nbio.funcs->get_hdp_flush_done_offset(adev);
681 
682 	adev->gfx.funcs->get_hdp_flush_mask(mes_ring, &ref_and_mask, &reg_mem_engine);
683 
684 	return amdgpu_mes_reg_write_reg_wait(adev, hdp_flush_req_offset, hdp_flush_done_offset,
685 					     ref_and_mask, ref_and_mask, 0);
686 }
687 
688 int amdgpu_mes_set_shader_debugger(struct amdgpu_device *adev,
689 				uint64_t process_context_addr,
690 				uint32_t spi_gdbg_per_vmid_cntl,
691 				const uint32_t *tcp_watch_cntl,
692 				uint32_t flags,
693 				bool trap_en,
694 				uint32_t xcc_id)
695 {
696 	struct mes_misc_op_input op_input = {0};
697 	int r;
698 
699 	if (!adev->mes.funcs->misc_op) {
700 		dev_err(adev->dev,
701 			"mes set shader debugger is not supported!\n");
702 		return -EINVAL;
703 	}
704 
705 	op_input.xcc_id = xcc_id;
706 	op_input.op = MES_MISC_OP_SET_SHADER_DEBUGGER;
707 	op_input.set_shader_debugger.process_context_addr = process_context_addr;
708 	op_input.set_shader_debugger.flags.u32all = flags;
709 
710 	/* use amdgpu mes_flush_shader_debugger instead */
711 	if (op_input.set_shader_debugger.flags.process_ctx_flush)
712 		return -EINVAL;
713 
714 	op_input.set_shader_debugger.spi_gdbg_per_vmid_cntl = spi_gdbg_per_vmid_cntl;
715 	memcpy(op_input.set_shader_debugger.tcp_watch_cntl, tcp_watch_cntl,
716 			sizeof(op_input.set_shader_debugger.tcp_watch_cntl));
717 
718 	if (((adev->mes.sched_version & AMDGPU_MES_API_VERSION_MASK) >>
719 			AMDGPU_MES_API_VERSION_SHIFT) >= 14)
720 		op_input.set_shader_debugger.trap_en = trap_en;
721 
722 	amdgpu_mes_lock(&adev->mes);
723 
724 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
725 	if (r)
726 		dev_err(adev->dev, "failed to set_shader_debugger\n");
727 
728 	amdgpu_mes_unlock(&adev->mes);
729 
730 	return r;
731 }
732 
733 int amdgpu_mes_flush_shader_debugger(struct amdgpu_device *adev,
734 				     uint64_t process_context_addr,
735 				     uint32_t xcc_id)
736 {
737 	struct mes_misc_op_input op_input = {0};
738 	int r;
739 
740 	if (!adev->mes.funcs->misc_op) {
741 		dev_err(adev->dev,
742 			"mes flush shader debugger is not supported!\n");
743 		return -EINVAL;
744 	}
745 
746 	op_input.xcc_id = xcc_id;
747 	op_input.op = MES_MISC_OP_SET_SHADER_DEBUGGER;
748 	op_input.set_shader_debugger.process_context_addr = process_context_addr;
749 	op_input.set_shader_debugger.flags.process_ctx_flush = true;
750 
751 	amdgpu_mes_lock(&adev->mes);
752 
753 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
754 	if (r)
755 		dev_err(adev->dev, "failed to set_shader_debugger\n");
756 
757 	amdgpu_mes_unlock(&adev->mes);
758 
759 	return r;
760 }
761 
762 uint32_t amdgpu_mes_get_aggregated_doorbell_index(struct amdgpu_device *adev,
763 						   enum amdgpu_mes_priority_level prio)
764 {
765 	return adev->mes.aggregated_doorbells[prio];
766 }
767 
768 int amdgpu_mes_init_microcode(struct amdgpu_device *adev, int pipe)
769 {
770 	const struct mes_firmware_header_v1_0 *mes_hdr;
771 	struct amdgpu_firmware_info *info;
772 	char ucode_prefix[30];
773 	char fw_name[50];
774 	bool need_retry = false;
775 	u32 *ucode_ptr;
776 	int r;
777 
778 	amdgpu_ucode_ip_version_decode(adev, GC_HWIP, ucode_prefix,
779 				       sizeof(ucode_prefix));
780 	if (adev->enable_uni_mes) {
781 		snprintf(fw_name, sizeof(fw_name),
782 			 "amdgpu/%s_uni_mes.bin", ucode_prefix);
783 	} else if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(11, 0, 0) &&
784 	    amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(12, 0, 0)) {
785 		snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes%s.bin",
786 			 ucode_prefix,
787 			 pipe == AMDGPU_MES_SCHED_PIPE ? "_2" : "1");
788 		need_retry = true;
789 	} else {
790 		snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes%s.bin",
791 			 ucode_prefix,
792 			 pipe == AMDGPU_MES_SCHED_PIPE ? "" : "1");
793 	}
794 
795 	r = amdgpu_ucode_request(adev, &adev->mes.fw[pipe], AMDGPU_UCODE_REQUIRED,
796 				 "%s", fw_name);
797 	if (r && need_retry && pipe == AMDGPU_MES_SCHED_PIPE) {
798 		dev_info(adev->dev, "try to fall back to %s_mes.bin\n", ucode_prefix);
799 		r = amdgpu_ucode_request(adev, &adev->mes.fw[pipe],
800 					 AMDGPU_UCODE_REQUIRED,
801 					 "amdgpu/%s_mes.bin", ucode_prefix);
802 	}
803 
804 	if (r)
805 		goto out;
806 
807 	mes_hdr = (const struct mes_firmware_header_v1_0 *)
808 		adev->mes.fw[pipe]->data;
809 	adev->mes.uc_start_addr[pipe] =
810 		le32_to_cpu(mes_hdr->mes_uc_start_addr_lo) |
811 		((uint64_t)(le32_to_cpu(mes_hdr->mes_uc_start_addr_hi)) << 32);
812 	adev->mes.data_start_addr[pipe] =
813 		le32_to_cpu(mes_hdr->mes_data_start_addr_lo) |
814 		((uint64_t)(le32_to_cpu(mes_hdr->mes_data_start_addr_hi)) << 32);
815 	ucode_ptr = (u32 *)(adev->mes.fw[pipe]->data +
816 			  sizeof(union amdgpu_firmware_header));
817 	adev->mes.fw_version[pipe] =
818 		le32_to_cpu(ucode_ptr[24]) & AMDGPU_MES_VERSION_MASK;
819 
820 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
821 		int ucode, ucode_data;
822 
823 		if (pipe == AMDGPU_MES_SCHED_PIPE) {
824 			ucode = AMDGPU_UCODE_ID_CP_MES;
825 			ucode_data = AMDGPU_UCODE_ID_CP_MES_DATA;
826 		} else {
827 			ucode = AMDGPU_UCODE_ID_CP_MES1;
828 			ucode_data = AMDGPU_UCODE_ID_CP_MES1_DATA;
829 		}
830 
831 		info = &adev->firmware.ucode[ucode];
832 		info->ucode_id = ucode;
833 		info->fw = adev->mes.fw[pipe];
834 		adev->firmware.fw_size +=
835 			ALIGN(le32_to_cpu(mes_hdr->mes_ucode_size_bytes),
836 			      PAGE_SIZE);
837 
838 		info = &adev->firmware.ucode[ucode_data];
839 		info->ucode_id = ucode_data;
840 		info->fw = adev->mes.fw[pipe];
841 		adev->firmware.fw_size +=
842 			ALIGN(le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes),
843 			      PAGE_SIZE);
844 	}
845 
846 	return 0;
847 out:
848 	amdgpu_ucode_release(&adev->mes.fw[pipe]);
849 	return r;
850 }
851 
852 void amdgpu_mes_validate_fw_version(struct amdgpu_device *adev)
853 {
854 	u32 fw_from_ucode = adev->mes.fw_version[AMDGPU_MES_SCHED_PIPE];
855 	u32 fw_from_reg = adev->mes.sched_version & AMDGPU_MES_VERSION_MASK;
856 
857 	if (fw_from_ucode != fw_from_reg)
858 		dev_info(adev->dev,
859 			 "MES firmware reports incorrect version in ucode binary (0x%x vs 0x%x)\n",
860 			 fw_from_ucode, fw_from_reg);
861 }
862 
863 
864 bool amdgpu_mes_suspend_resume_all_supported(struct amdgpu_device *adev)
865 {
866 	uint32_t mes_rev = adev->mes.sched_version & AMDGPU_MES_VERSION_MASK;
867 
868 	return ((amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(11, 0, 0) &&
869 		 amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(12, 0, 0) &&
870 		 mes_rev >= 0x63) ||
871 		amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(12, 0, 0));
872 }
873 
874 bool amdgpu_mes_queue_reset_by_mes_supported(struct amdgpu_device *adev)
875 {
876 	u32 ip_maj = IP_VERSION_MAJ(amdgpu_ip_version(adev, GC_HWIP, 0));
877 	u32 ip_min = IP_VERSION_MIN(amdgpu_ip_version(adev, GC_HWIP, 0));
878 	u32 mes_sched = adev->mes.sched_version & AMDGPU_MES_VERSION_MASK;
879 
880 	return (ip_maj == 11 && mes_sched >= 0x8c) ||
881 		((ip_maj == 12 && ip_min == 0) && mes_sched >= 0x8d) ||
882 		((ip_maj == 12 && ip_min == 1) && mes_sched >= 0x73);
883 }
884 
885 /* Fix me -- node_id is used to identify the correct MES instances in the future */
886 static int amdgpu_mes_set_enforce_isolation(struct amdgpu_device *adev,
887 					    uint32_t node_id, bool enable)
888 {
889 	struct mes_misc_op_input op_input = {0};
890 	int r;
891 
892 	op_input.op = MES_MISC_OP_CHANGE_CONFIG;
893 	op_input.change_config.option.limit_single_process = enable ? 1 : 0;
894 
895 	if (!adev->mes.funcs->misc_op) {
896 		dev_err(adev->dev, "mes change config is not supported!\n");
897 		r = -EINVAL;
898 		goto error;
899 	}
900 
901 	amdgpu_mes_lock(&adev->mes);
902 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
903 	amdgpu_mes_unlock(&adev->mes);
904 	if (r)
905 		dev_err(adev->dev, "failed to change_config.\n");
906 
907 error:
908 	return r;
909 }
910 
911 int amdgpu_mes_update_enforce_isolation(struct amdgpu_device *adev)
912 {
913 	int i, r = 0;
914 
915 	if (adev->enable_mes && adev->gfx.enable_cleaner_shader) {
916 		mutex_lock(&adev->enforce_isolation_mutex);
917 		for (i = 0; i < (adev->xcp_mgr ? adev->xcp_mgr->num_xcps : 1); i++) {
918 			if (adev->enforce_isolation[i] == AMDGPU_ENFORCE_ISOLATION_ENABLE)
919 				r |= amdgpu_mes_set_enforce_isolation(adev, i, true);
920 			else
921 				r |= amdgpu_mes_set_enforce_isolation(adev, i, false);
922 		}
923 		mutex_unlock(&adev->enforce_isolation_mutex);
924 	}
925 	return r;
926 }
927 
928 /**
929  * amdgpu_mes_rs64mem_init - initialize RS64 local memory context arrays
930  *
931  * @mes: MES instance
932  *
933  * Returns 0 on success, negative errno on failure.
934  */
935 int amdgpu_mes_rs64mem_init(struct amdgpu_mes *mes)
936 {
937 	struct amdgpu_device *adev = container_of(mes, struct amdgpu_device, mes);
938 	int r;
939 
940 	if (!mes->use_rs64mem)
941 		return 0;
942 
943 	r = amdgpu_bo_create_kernel(adev, PAGE_SIZE, PAGE_SIZE,
944 				    AMDGPU_GEM_DOMAIN_GTT,
945 				    &mes->ctx_array_size_bo,
946 				    &mes->ctx_array_size_gpu_addr,
947 				    (void **)&mes->ctx_array_size_cpu_ptr);
948 	if (r) {
949 		dev_err(adev->dev,
950 			"Failed to allocate ctx array size BO, r=%d\n", r);
951 		return r;
952 	}
953 
954 	memset(mes->ctx_array_size_cpu_ptr, 0, PAGE_SIZE);
955 
956 	return 0;
957 }
958 
959  /**
960   * amdgpu_mes_rs64mem_fini - tear down RS64 local memory management
961   *
962   * @mes: MES instance
963   */
964 void amdgpu_mes_rs64mem_fini(struct amdgpu_mes *mes)
965 {
966 	if (mes->ctx_array_size_bo) {
967 		amdgpu_bo_free_kernel(&mes->ctx_array_size_bo,
968 				      &mes->ctx_array_size_gpu_addr,
969 				      (void **)&mes->ctx_array_size_cpu_ptr);
970 	}
971 	bitmap_free(mes->proc_ctx_bitmap);
972 	bitmap_free(mes->gang_ctx_bitmap);
973 	mes->use_rs64mem = false;
974 }
975 
976 /**
977  * amdgpu_mes_rs64mem_setup_bitmaps - allocate bitmaps after querying MES
978  *
979  * Called after QUERY_SCHEDULER_STATUS returns and MES has written
980  * the array sizes to the GPU buffer. Reads the sizes and allocates
981  * the tracking bitmaps.
982  *
983  * @mes: MES instance
984  *
985  * Returns 0 on success, negative errno on failure.
986  */
987 int amdgpu_mes_rs64mem_setup_bitmaps(struct amdgpu_mes *mes)
988 {
989 	struct amdgpu_device *adev = container_of(mes, struct amdgpu_device, mes);
990 
991 	if (!mes->use_rs64mem || !mes->ctx_array_size_cpu_ptr)
992 		return 0;
993 
994 	/*
995 	 * MES FW wrote the sizes to the GPU buffer:
996 	 *   ctx_array_size_cpu_ptr[0] = proc_ctx_array_size (N)
997 	 *   ctx_array_size_cpu_ptr[1] = gang_ctx_array_size (M)
998 	 */
999 	mes->proc_ctx_array_size = mes->ctx_array_size_cpu_ptr[0];
1000 	mes->gang_ctx_array_size = mes->ctx_array_size_cpu_ptr[1];
1001 
1002 	/* Sanity check - MES FW typically returns N=50, M=300 */
1003 	if (mes->proc_ctx_array_size == 0 || mes->gang_ctx_array_size == 0) {
1004 		dev_warn(adev->dev,
1005 			 "MES returned zero ctx array sizes (proc=%u, gang=%u), "
1006 			 "disabling RS64 local memory optimization\n",
1007 			 mes->proc_ctx_array_size, mes->gang_ctx_array_size);
1008 		mes->use_rs64mem = false;
1009 		return 0;
1010 	}
1011 
1012 	/* Cap to safety limits */
1013 	if (mes->proc_ctx_array_size > AMDGPU_MES_PROC_CTX_ARRAY_MAX)
1014 		mes->proc_ctx_array_size = AMDGPU_MES_PROC_CTX_ARRAY_MAX;
1015 	if (mes->gang_ctx_array_size > AMDGPU_MES_GANG_CTX_ARRAY_MAX)
1016 		mes->gang_ctx_array_size = AMDGPU_MES_GANG_CTX_ARRAY_MAX;
1017 
1018 	dev_info(adev->dev,
1019 		 "MES RS64 local memory: proc_ctx_array_size:%u, "
1020 		 "gang_ctx_array_size:%u\n",
1021 		 mes->proc_ctx_array_size, mes->gang_ctx_array_size);
1022 
1023 	/* Allocate bitmaps */
1024 	mes->proc_ctx_bitmap = bitmap_zalloc(mes->proc_ctx_array_size,
1025 					     GFP_KERNEL);
1026 	if (!mes->proc_ctx_bitmap) {
1027 		mes->use_rs64mem = false;
1028 		return -ENOMEM;
1029 	}
1030 
1031 	mes->gang_ctx_bitmap = bitmap_zalloc(mes->gang_ctx_array_size,
1032 					     GFP_KERNEL);
1033 	if (!mes->gang_ctx_bitmap) {
1034 		bitmap_free(mes->proc_ctx_bitmap);
1035 		mes->proc_ctx_bitmap = NULL;
1036 		mes->use_rs64mem = false;
1037 		return -ENOMEM;
1038 	}
1039 
1040 	return 0;
1041 }
1042 
1043 /**
1044  * amdgpu_mes_alloc_proc_ctx_index - allocate a process context slot
1045  *
1046  * @mes: MES instance
1047  * @index: the allocated process context index
1048  *
1049  * Returns 0 on success, -ENOSPC if all slots are used, or
1050  * -EOPNOTSUPP if RS64 local memory is unavailable.
1051  */
1052 int amdgpu_mes_alloc_proc_ctx_index(struct amdgpu_mes *mes,
1053 				    uint32_t *index)
1054 {
1055 	unsigned long bit;
1056 
1057 	if (!mes->use_rs64mem || !mes->proc_ctx_bitmap)
1058 		return -EOPNOTSUPP;
1059 
1060 	amdgpu_mes_lock(mes);
1061 	bit = find_first_zero_bit(mes->proc_ctx_bitmap,
1062 				  mes->proc_ctx_array_size);
1063 	if (bit >= mes->proc_ctx_array_size) {
1064 		amdgpu_mes_unlock(mes);
1065 		return -ENOSPC;
1066 	}
1067 	set_bit(bit, mes->proc_ctx_bitmap);
1068 	*index = (uint32_t)bit;
1069 	amdgpu_mes_unlock(mes);
1070 
1071 	return 0;
1072 }
1073 
1074  /**
1075   * amdgpu_mes_free_proc_ctx_index - free a process context slot
1076   *
1077   * @mes: MES instance
1078   * @index: process context index is released
1079   */
1080 void amdgpu_mes_free_proc_ctx_index(struct amdgpu_mes *mes,
1081 				    uint32_t index)
1082 {
1083 	if (!mes->use_rs64mem || !mes->proc_ctx_bitmap)
1084 		return;
1085 	if (index >= mes->proc_ctx_array_size)
1086 		return;
1087 
1088 	amdgpu_mes_lock(mes);
1089 	clear_bit(index, mes->proc_ctx_bitmap);
1090 	amdgpu_mes_unlock(mes);
1091 }
1092 
1093  /**
1094   * amdgpu_mes_alloc_gang_ctx_index - allocate a gang context slot
1095   *
1096   * @mes: MES instance
1097   * @index: the allocated gang context index
1098   *
1099   * Returns 0 on success, -ENOSPC if all slots are used, or
1100   * -EOPNOTSUPP if RS64 local memory is unavailable.
1101   */
1102 int amdgpu_mes_alloc_gang_ctx_index(struct amdgpu_mes *mes,
1103 				    uint32_t *index)
1104 {
1105 	unsigned long bit;
1106 
1107 	if (!mes->use_rs64mem || !mes->gang_ctx_bitmap)
1108 		return -EOPNOTSUPP;
1109 
1110 	amdgpu_mes_lock(mes);
1111 	bit = find_first_zero_bit(mes->gang_ctx_bitmap,
1112 				  mes->gang_ctx_array_size);
1113 	if (bit >= mes->gang_ctx_array_size) {
1114 		amdgpu_mes_unlock(mes);
1115 		return -ENOSPC;
1116 	}
1117 	set_bit(bit, mes->gang_ctx_bitmap);
1118 	*index = bit;
1119 	amdgpu_mes_unlock(mes);
1120 
1121 	return 0;
1122 }
1123 
1124  /**
1125   * amdgpu_mes_free_gang_ctx_index - free a gang context slot
1126   *
1127   * @mes: MES instance
1128   * @index: gang context index is released
1129   */
1130 void amdgpu_mes_free_gang_ctx_index(struct amdgpu_mes *mes,
1131 				    uint32_t index)
1132 {
1133 	if (!mes->use_rs64mem || !mes->gang_ctx_bitmap)
1134 		return;
1135 	if (index >= mes->gang_ctx_array_size)
1136 		return;
1137 
1138 	amdgpu_mes_lock(mes);
1139 	clear_bit(index, mes->gang_ctx_bitmap);
1140 	amdgpu_mes_unlock(mes);
1141 }
1142 
1143 #if defined(CONFIG_DEBUG_FS)
1144 
1145 static int amdgpu_debugfs_mes_event_log_show(struct seq_file *m, void *unused)
1146 {
1147 	struct amdgpu_device *adev = m->private;
1148 	uint32_t *mem = (uint32_t *)(adev->mes.event_log_cpu_addr);
1149 
1150 	seq_hex_dump(m, "", DUMP_PREFIX_OFFSET, 32, 4,
1151 		     mem, adev->mes.event_log_size, false);
1152 
1153 	return 0;
1154 }
1155 
1156 DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_mes_event_log);
1157 
1158 #endif
1159 
1160 void amdgpu_debugfs_mes_event_log_init(struct amdgpu_device *adev)
1161 {
1162 
1163 #if defined(CONFIG_DEBUG_FS)
1164 	struct drm_minor *minor = adev_to_drm(adev)->primary;
1165 	struct dentry *root = minor->debugfs_root;
1166 	if (adev->enable_mes && amdgpu_mes_log_enable)
1167 		debugfs_create_file("amdgpu_mes_event_log", 0444, root,
1168 				    adev, &amdgpu_debugfs_mes_event_log_fops);
1169 
1170 #endif
1171 }
1172