xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c (revision c06b6cde2a1c3bcbb561bd57bb6f34eae9030921)
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_device_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_device_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(12, 1, 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; 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 
255 	return 0;
256 
257 error_doorbell:
258 	amdgpu_mes_doorbell_free(adev);
259 error:
260 	for (i = 0; i < AMDGPU_MAX_MES_PIPES * num_xcc; i++) {
261 		if (adev->mes.sch_ctx_ptr[i])
262 			amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs[i]);
263 		if (adev->mes.query_status_fence_ptr[i])
264 			amdgpu_device_wb_free(adev,
265 				      adev->mes.query_status_fence_offs[i]);
266 		if (adev->mes.hung_queue_db_array_gpu_obj[i])
267 			amdgpu_bo_free_kernel(&adev->mes.hung_queue_db_array_gpu_obj[i],
268 					      &adev->mes.hung_queue_db_array_gpu_addr[i],
269 					      &adev->mes.hung_queue_db_array_cpu_addr[i]);
270 	}
271 
272 	ida_destroy(&adev->mes.doorbell_ida);
273 	mutex_destroy(&adev->mes.mutex_hidden);
274 	return r;
275 }
276 
277 void amdgpu_mes_fini(struct amdgpu_device *adev)
278 {
279 	int i;
280 	int num_xcc = adev->gfx.xcc_mask ? NUM_XCC(adev->gfx.xcc_mask) : 1;
281 
282 	amdgpu_bo_free_kernel(&adev->mes.event_log_gpu_obj,
283 			      &adev->mes.event_log_gpu_addr,
284 			      &adev->mes.event_log_cpu_addr);
285 
286 	for (i = 0; i < AMDGPU_MAX_MES_PIPES * num_xcc; i++) {
287 		if (adev->mes.hung_queue_db_array_gpu_obj[i])
288 			 amdgpu_bo_free_kernel(&adev->mes.hung_queue_db_array_gpu_obj[i],
289 					 &adev->mes.hung_queue_db_array_gpu_addr[i],
290 					 &adev->mes.hung_queue_db_array_cpu_addr[i]);
291 		if (adev->mes.sch_ctx_ptr[i])
292 			amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs[i]);
293 		if (adev->mes.query_status_fence_ptr[i])
294 			amdgpu_device_wb_free(adev,
295 				      adev->mes.query_status_fence_offs[i]);
296 	}
297 
298 	amdgpu_mes_doorbell_free(adev);
299 
300 	ida_destroy(&adev->mes.doorbell_ida);
301 	mutex_destroy(&adev->mes.mutex_hidden);
302 }
303 
304 int amdgpu_mes_suspend(struct amdgpu_device *adev, u32 xcc_id)
305 {
306 	struct mes_suspend_gang_input input;
307 	int r;
308 
309 	if (!amdgpu_mes_suspend_resume_all_supported(adev))
310 		return 0;
311 
312 	memset(&input, 0x0, sizeof(struct mes_suspend_gang_input));
313 	input.suspend_all_gangs = 1;
314 	input.xcc_id = xcc_id;
315 
316 	/*
317 	 * Avoid taking any other locks under MES lock to avoid circular
318 	 * lock dependencies.
319 	 */
320 	amdgpu_mes_lock(&adev->mes);
321 	r = adev->mes.funcs->suspend_gang(&adev->mes, &input);
322 	amdgpu_mes_unlock(&adev->mes);
323 	if (r)
324 		dev_err(adev->dev, "failed to suspend all gangs");
325 
326 	return r;
327 }
328 
329 int amdgpu_mes_resume(struct amdgpu_device *adev, u32 xcc_id)
330 {
331 	struct mes_resume_gang_input input;
332 	int r;
333 
334 	if (!amdgpu_mes_suspend_resume_all_supported(adev))
335 		return 0;
336 
337 	memset(&input, 0x0, sizeof(struct mes_resume_gang_input));
338 	input.resume_all_gangs = 1;
339 	input.xcc_id = xcc_id;
340 
341 	/*
342 	 * Avoid taking any other locks under MES lock to avoid circular
343 	 * lock dependencies.
344 	 */
345 	amdgpu_mes_lock(&adev->mes);
346 	r = adev->mes.funcs->resume_gang(&adev->mes, &input);
347 	amdgpu_mes_unlock(&adev->mes);
348 	if (r)
349 		dev_err(adev->dev, "failed to resume all gangs");
350 
351 	return r;
352 }
353 
354 int amdgpu_mes_map_legacy_queue(struct amdgpu_device *adev,
355 				struct amdgpu_ring *ring, uint32_t xcc_id)
356 {
357 	struct mes_map_legacy_queue_input queue_input;
358 	int r;
359 
360 	memset(&queue_input, 0, sizeof(queue_input));
361 
362 	queue_input.xcc_id = xcc_id;
363 	queue_input.queue_type = ring->funcs->type;
364 	queue_input.doorbell_offset = ring->doorbell_index;
365 	queue_input.pipe_id = ring->pipe;
366 	queue_input.queue_id = ring->queue;
367 	queue_input.mqd_addr = amdgpu_bo_gpu_offset(ring->mqd_obj);
368 	queue_input.wptr_addr = ring->wptr_gpu_addr;
369 
370 	amdgpu_mes_lock(&adev->mes);
371 	r = adev->mes.funcs->map_legacy_queue(&adev->mes, &queue_input);
372 	amdgpu_mes_unlock(&adev->mes);
373 	if (r)
374 		dev_err(adev->dev, "failed to map legacy queue\n");
375 
376 	return r;
377 }
378 
379 int amdgpu_mes_unmap_legacy_queue(struct amdgpu_device *adev,
380 				  struct amdgpu_ring *ring,
381 				  enum amdgpu_unmap_queues_action action,
382 				  u64 gpu_addr, u64 seq, uint32_t xcc_id)
383 {
384 	struct mes_unmap_legacy_queue_input queue_input;
385 	int r;
386 
387 	queue_input.xcc_id = xcc_id;
388 	queue_input.action = action;
389 	queue_input.queue_type = ring->funcs->type;
390 	queue_input.doorbell_offset = ring->doorbell_index;
391 	queue_input.pipe_id = ring->pipe;
392 	queue_input.queue_id = ring->queue;
393 	queue_input.trail_fence_addr = gpu_addr;
394 	queue_input.trail_fence_data = seq;
395 
396 	amdgpu_mes_lock(&adev->mes);
397 	r = adev->mes.funcs->unmap_legacy_queue(&adev->mes, &queue_input);
398 	amdgpu_mes_unlock(&adev->mes);
399 	if (r)
400 		dev_err(adev->dev, "failed to unmap legacy queue\n");
401 
402 	return r;
403 }
404 
405 int amdgpu_mes_reset_legacy_queue(struct amdgpu_device *adev,
406 				  struct amdgpu_ring *ring,
407 				  unsigned int vmid,
408 				  bool use_mmio,
409 				  uint32_t xcc_id)
410 {
411 	struct mes_reset_queue_input queue_input;
412 	int r;
413 
414 	memset(&queue_input, 0, sizeof(queue_input));
415 
416 	queue_input.xcc_id = xcc_id;
417 	queue_input.queue_type = ring->funcs->type;
418 	queue_input.doorbell_offset = ring->doorbell_index;
419 	queue_input.me_id = ring->me;
420 	queue_input.pipe_id = ring->pipe;
421 	queue_input.queue_id = ring->queue;
422 	queue_input.mqd_addr = ring->mqd_obj ? amdgpu_bo_gpu_offset(ring->mqd_obj) : 0;
423 	queue_input.wptr_addr = ring->wptr_gpu_addr;
424 	queue_input.vmid = vmid;
425 	queue_input.use_mmio = use_mmio;
426 	queue_input.is_kq = true;
427 	if (ring->funcs->type == AMDGPU_RING_TYPE_GFX)
428 		queue_input.legacy_gfx = true;
429 
430 	amdgpu_mes_lock(&adev->mes);
431 	r = adev->mes.funcs->reset_hw_queue(&adev->mes, &queue_input);
432 	amdgpu_mes_unlock(&adev->mes);
433 	if (r)
434 		dev_err(adev->dev, "failed to reset legacy queue\n");
435 
436 	return r;
437 }
438 
439 int amdgpu_mes_get_hung_queue_db_array_size(struct amdgpu_device *adev)
440 {
441 	return adev->mes.hung_queue_db_array_size;
442 }
443 
444 int amdgpu_mes_detect_and_reset_hung_queues(struct amdgpu_device *adev,
445 					    int queue_type,
446 					    bool detect_only,
447 					    unsigned int *hung_db_num,
448 					    u32 *hung_db_array,
449 					    uint32_t xcc_id)
450 {
451 	struct mes_detect_and_reset_queue_input input;
452 	u32 *db_array = adev->mes.hung_queue_db_array_cpu_addr[xcc_id];
453 	int hqd_info_offset = adev->mes.hung_queue_hqd_info_offset, r, i;
454 
455 	if (!hung_db_num || !hung_db_array)
456 		return -EINVAL;
457 
458 	if ((queue_type != AMDGPU_RING_TYPE_GFX) &&
459 	    (queue_type != AMDGPU_RING_TYPE_COMPUTE) &&
460 	    (queue_type != AMDGPU_RING_TYPE_SDMA))
461 		return -EINVAL;
462 
463 	/* Clear the doorbell array before detection */
464 	memset(adev->mes.hung_queue_db_array_cpu_addr[xcc_id], AMDGPU_MES_INVALID_DB_OFFSET,
465 		adev->mes.hung_queue_db_array_size * sizeof(u32));
466 	input.queue_type = queue_type;
467 	input.detect_only = detect_only;
468 	input.xcc_id = xcc_id;
469 
470 	r = adev->mes.funcs->detect_and_reset_hung_queues(&adev->mes,
471 							  &input);
472 
473 	if (r && detect_only) {
474 		dev_err(adev->dev, "Failed to detect hung queues\n");
475 		return r;
476 	}
477 
478 	*hung_db_num = 0;
479 	/* MES passes hung queues' doorbell to driver */
480 	for (i = 0; i < adev->mes.hung_queue_hqd_info_offset; i++) {
481 		/* Finding hung queues where db_array[i] is a valid doorbell */
482 		if (db_array[i] != AMDGPU_MES_INVALID_DB_OFFSET) {
483 			hung_db_array[i] = db_array[i];
484 			*hung_db_num += 1;
485 		}
486 	}
487 
488 	if (r && !(*hung_db_num)) {
489 		dev_err(adev->dev, "Failed to detect and reset hung queues\n");
490 		return r;
491 	}
492 
493 	for (i = hqd_info_offset; i < hqd_info_offset + *hung_db_num; i++)
494 		hung_db_array[i] = db_array[i];
495 
496 	return r;
497 }
498 
499 uint32_t amdgpu_mes_rreg(struct amdgpu_device *adev, uint32_t reg,
500 			 uint32_t xcc_id)
501 {
502 	struct mes_misc_op_input op_input;
503 	int r, val = 0;
504 	uint32_t addr_offset = 0;
505 	uint64_t read_val_gpu_addr;
506 	uint32_t *read_val_ptr;
507 
508 	if (amdgpu_device_wb_get(adev, &addr_offset)) {
509 		dev_err(adev->dev, "critical bug! too many mes readers\n");
510 		goto error;
511 	}
512 	read_val_gpu_addr = adev->wb.gpu_addr + (addr_offset * 4);
513 	read_val_ptr = (uint32_t *)&adev->wb.wb[addr_offset];
514 	op_input.xcc_id = xcc_id;
515 	op_input.op = MES_MISC_OP_READ_REG;
516 	op_input.read_reg.reg_offset = reg;
517 	op_input.read_reg.buffer_addr = read_val_gpu_addr;
518 
519 	if (!adev->mes.funcs->misc_op) {
520 		dev_err(adev->dev, "mes rreg is not supported!\n");
521 		goto error;
522 	}
523 
524 	amdgpu_mes_lock(&adev->mes);
525 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
526 	amdgpu_mes_unlock(&adev->mes);
527 	if (r)
528 		dev_err(adev->dev, "failed to read reg (0x%x)\n", reg);
529 	else
530 		val = *(read_val_ptr);
531 
532 error:
533 	if (addr_offset)
534 		amdgpu_device_wb_free(adev, addr_offset);
535 	return val;
536 }
537 
538 int amdgpu_mes_wreg(struct amdgpu_device *adev, uint32_t reg,
539 		    uint32_t val, uint32_t xcc_id)
540 {
541 	struct mes_misc_op_input op_input;
542 	int r;
543 
544 	op_input.xcc_id = xcc_id;
545 	op_input.op = MES_MISC_OP_WRITE_REG;
546 	op_input.write_reg.reg_offset = reg;
547 	op_input.write_reg.reg_value = val;
548 
549 	if (!adev->mes.funcs->misc_op) {
550 		dev_err(adev->dev, "mes wreg is not supported!\n");
551 		r = -EINVAL;
552 		goto error;
553 	}
554 
555 	amdgpu_mes_lock(&adev->mes);
556 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
557 	amdgpu_mes_unlock(&adev->mes);
558 	if (r)
559 		dev_err(adev->dev, "failed to write reg (0x%x)\n", reg);
560 
561 error:
562 	return r;
563 }
564 
565 int amdgpu_mes_reg_write_reg_wait(struct amdgpu_device *adev,
566 				  uint32_t reg0, uint32_t reg1,
567 				  uint32_t ref, uint32_t mask,
568 				  uint32_t xcc_id)
569 {
570 	struct mes_misc_op_input op_input;
571 	int r;
572 
573 	op_input.xcc_id = xcc_id;
574 	op_input.op = MES_MISC_OP_WRM_REG_WR_WAIT;
575 	op_input.wrm_reg.reg0 = reg0;
576 	op_input.wrm_reg.reg1 = reg1;
577 	op_input.wrm_reg.ref = ref;
578 	op_input.wrm_reg.mask = mask;
579 
580 	if (!adev->mes.funcs->misc_op) {
581 		dev_err(adev->dev, "mes reg_write_reg_wait is not supported!\n");
582 		r = -EINVAL;
583 		goto error;
584 	}
585 
586 	amdgpu_mes_lock(&adev->mes);
587 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
588 	amdgpu_mes_unlock(&adev->mes);
589 	if (r)
590 		dev_err(adev->dev, "failed to reg_write_reg_wait\n");
591 
592 error:
593 	return r;
594 }
595 
596 int amdgpu_mes_hdp_flush(struct amdgpu_device *adev)
597 {
598 	uint32_t hdp_flush_req_offset, hdp_flush_done_offset;
599 	struct amdgpu_ring *mes_ring;
600 	uint32_t ref_and_mask = 0, reg_mem_engine = 0;
601 
602 	if (!adev->gfx.funcs->get_hdp_flush_mask) {
603 		dev_err(adev->dev, "mes hdp flush is not supported.\n");
604 		return -EINVAL;
605 	}
606 
607 	mes_ring = &adev->mes.ring[0];
608 	hdp_flush_req_offset = adev->nbio.funcs->get_hdp_flush_req_offset(adev);
609 	hdp_flush_done_offset = adev->nbio.funcs->get_hdp_flush_done_offset(adev);
610 
611 	adev->gfx.funcs->get_hdp_flush_mask(mes_ring, &ref_and_mask, &reg_mem_engine);
612 
613 	return amdgpu_mes_reg_write_reg_wait(adev, hdp_flush_req_offset, hdp_flush_done_offset,
614 					     ref_and_mask, ref_and_mask, 0);
615 }
616 
617 int amdgpu_mes_set_shader_debugger(struct amdgpu_device *adev,
618 				uint64_t process_context_addr,
619 				uint32_t spi_gdbg_per_vmid_cntl,
620 				const uint32_t *tcp_watch_cntl,
621 				uint32_t flags,
622 				bool trap_en,
623 				uint32_t xcc_id)
624 {
625 	struct mes_misc_op_input op_input = {0};
626 	int r;
627 
628 	if (!adev->mes.funcs->misc_op) {
629 		dev_err(adev->dev,
630 			"mes set shader debugger is not supported!\n");
631 		return -EINVAL;
632 	}
633 
634 	op_input.xcc_id = xcc_id;
635 	op_input.op = MES_MISC_OP_SET_SHADER_DEBUGGER;
636 	op_input.set_shader_debugger.process_context_addr = process_context_addr;
637 	op_input.set_shader_debugger.flags.u32all = flags;
638 
639 	/* use amdgpu mes_flush_shader_debugger instead */
640 	if (op_input.set_shader_debugger.flags.process_ctx_flush)
641 		return -EINVAL;
642 
643 	op_input.set_shader_debugger.spi_gdbg_per_vmid_cntl = spi_gdbg_per_vmid_cntl;
644 	memcpy(op_input.set_shader_debugger.tcp_watch_cntl, tcp_watch_cntl,
645 			sizeof(op_input.set_shader_debugger.tcp_watch_cntl));
646 
647 	if (((adev->mes.sched_version & AMDGPU_MES_API_VERSION_MASK) >>
648 			AMDGPU_MES_API_VERSION_SHIFT) >= 14)
649 		op_input.set_shader_debugger.trap_en = trap_en;
650 
651 	amdgpu_mes_lock(&adev->mes);
652 
653 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
654 	if (r)
655 		dev_err(adev->dev, "failed to set_shader_debugger\n");
656 
657 	amdgpu_mes_unlock(&adev->mes);
658 
659 	return r;
660 }
661 
662 int amdgpu_mes_flush_shader_debugger(struct amdgpu_device *adev,
663 				     uint64_t process_context_addr,
664 				     uint32_t xcc_id)
665 {
666 	struct mes_misc_op_input op_input = {0};
667 	int r;
668 
669 	if (!adev->mes.funcs->misc_op) {
670 		dev_err(adev->dev,
671 			"mes flush shader debugger is not supported!\n");
672 		return -EINVAL;
673 	}
674 
675 	op_input.xcc_id = xcc_id;
676 	op_input.op = MES_MISC_OP_SET_SHADER_DEBUGGER;
677 	op_input.set_shader_debugger.process_context_addr = process_context_addr;
678 	op_input.set_shader_debugger.flags.process_ctx_flush = true;
679 
680 	amdgpu_mes_lock(&adev->mes);
681 
682 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
683 	if (r)
684 		dev_err(adev->dev, "failed to set_shader_debugger\n");
685 
686 	amdgpu_mes_unlock(&adev->mes);
687 
688 	return r;
689 }
690 
691 uint32_t amdgpu_mes_get_aggregated_doorbell_index(struct amdgpu_device *adev,
692 						   enum amdgpu_mes_priority_level prio)
693 {
694 	return adev->mes.aggregated_doorbells[prio];
695 }
696 
697 int amdgpu_mes_init_microcode(struct amdgpu_device *adev, int pipe)
698 {
699 	const struct mes_firmware_header_v1_0 *mes_hdr;
700 	struct amdgpu_firmware_info *info;
701 	char ucode_prefix[30];
702 	char fw_name[50];
703 	bool need_retry = false;
704 	u32 *ucode_ptr;
705 	int r;
706 
707 	amdgpu_ucode_ip_version_decode(adev, GC_HWIP, ucode_prefix,
708 				       sizeof(ucode_prefix));
709 	if (adev->enable_uni_mes) {
710 		snprintf(fw_name, sizeof(fw_name),
711 			 "amdgpu/%s_uni_mes.bin", ucode_prefix);
712 	} else if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(11, 0, 0) &&
713 	    amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(12, 0, 0)) {
714 		snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes%s.bin",
715 			 ucode_prefix,
716 			 pipe == AMDGPU_MES_SCHED_PIPE ? "_2" : "1");
717 		need_retry = true;
718 	} else {
719 		snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes%s.bin",
720 			 ucode_prefix,
721 			 pipe == AMDGPU_MES_SCHED_PIPE ? "" : "1");
722 	}
723 
724 	r = amdgpu_ucode_request(adev, &adev->mes.fw[pipe], AMDGPU_UCODE_REQUIRED,
725 				 "%s", fw_name);
726 	if (r && need_retry && pipe == AMDGPU_MES_SCHED_PIPE) {
727 		dev_info(adev->dev, "try to fall back to %s_mes.bin\n", ucode_prefix);
728 		r = amdgpu_ucode_request(adev, &adev->mes.fw[pipe],
729 					 AMDGPU_UCODE_REQUIRED,
730 					 "amdgpu/%s_mes.bin", ucode_prefix);
731 	}
732 
733 	if (r)
734 		goto out;
735 
736 	mes_hdr = (const struct mes_firmware_header_v1_0 *)
737 		adev->mes.fw[pipe]->data;
738 	adev->mes.uc_start_addr[pipe] =
739 		le32_to_cpu(mes_hdr->mes_uc_start_addr_lo) |
740 		((uint64_t)(le32_to_cpu(mes_hdr->mes_uc_start_addr_hi)) << 32);
741 	adev->mes.data_start_addr[pipe] =
742 		le32_to_cpu(mes_hdr->mes_data_start_addr_lo) |
743 		((uint64_t)(le32_to_cpu(mes_hdr->mes_data_start_addr_hi)) << 32);
744 	ucode_ptr = (u32 *)(adev->mes.fw[pipe]->data +
745 			  sizeof(union amdgpu_firmware_header));
746 	adev->mes.fw_version[pipe] =
747 		le32_to_cpu(ucode_ptr[24]) & AMDGPU_MES_VERSION_MASK;
748 
749 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
750 		int ucode, ucode_data;
751 
752 		if (pipe == AMDGPU_MES_SCHED_PIPE) {
753 			ucode = AMDGPU_UCODE_ID_CP_MES;
754 			ucode_data = AMDGPU_UCODE_ID_CP_MES_DATA;
755 		} else {
756 			ucode = AMDGPU_UCODE_ID_CP_MES1;
757 			ucode_data = AMDGPU_UCODE_ID_CP_MES1_DATA;
758 		}
759 
760 		info = &adev->firmware.ucode[ucode];
761 		info->ucode_id = ucode;
762 		info->fw = adev->mes.fw[pipe];
763 		adev->firmware.fw_size +=
764 			ALIGN(le32_to_cpu(mes_hdr->mes_ucode_size_bytes),
765 			      PAGE_SIZE);
766 
767 		info = &adev->firmware.ucode[ucode_data];
768 		info->ucode_id = ucode_data;
769 		info->fw = adev->mes.fw[pipe];
770 		adev->firmware.fw_size +=
771 			ALIGN(le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes),
772 			      PAGE_SIZE);
773 	}
774 
775 	return 0;
776 out:
777 	amdgpu_ucode_release(&adev->mes.fw[pipe]);
778 	return r;
779 }
780 
781 bool amdgpu_mes_suspend_resume_all_supported(struct amdgpu_device *adev)
782 {
783 	uint32_t mes_rev = adev->mes.sched_version & AMDGPU_MES_VERSION_MASK;
784 
785 	return ((amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(11, 0, 0) &&
786 		 amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(12, 0, 0) &&
787 		 mes_rev >= 0x63) ||
788 		amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(12, 0, 0));
789 }
790 
791 bool amdgpu_mes_queue_reset_by_mes_supported(struct amdgpu_device *adev)
792 {
793 	return (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(12, 1, 0) &&
794 		(adev->mes.sched_version & AMDGPU_MES_VERSION_MASK) >= 0x73);
795 }
796 
797 /* Fix me -- node_id is used to identify the correct MES instances in the future */
798 static int amdgpu_mes_set_enforce_isolation(struct amdgpu_device *adev,
799 					    uint32_t node_id, bool enable)
800 {
801 	struct mes_misc_op_input op_input = {0};
802 	int r;
803 
804 	op_input.op = MES_MISC_OP_CHANGE_CONFIG;
805 	op_input.change_config.option.limit_single_process = enable ? 1 : 0;
806 
807 	if (!adev->mes.funcs->misc_op) {
808 		dev_err(adev->dev, "mes change config is not supported!\n");
809 		r = -EINVAL;
810 		goto error;
811 	}
812 
813 	amdgpu_mes_lock(&adev->mes);
814 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
815 	amdgpu_mes_unlock(&adev->mes);
816 	if (r)
817 		dev_err(adev->dev, "failed to change_config.\n");
818 
819 error:
820 	return r;
821 }
822 
823 int amdgpu_mes_update_enforce_isolation(struct amdgpu_device *adev)
824 {
825 	int i, r = 0;
826 
827 	if (adev->enable_mes && adev->gfx.enable_cleaner_shader) {
828 		mutex_lock(&adev->enforce_isolation_mutex);
829 		for (i = 0; i < (adev->xcp_mgr ? adev->xcp_mgr->num_xcps : 1); i++) {
830 			if (adev->enforce_isolation[i] == AMDGPU_ENFORCE_ISOLATION_ENABLE)
831 				r |= amdgpu_mes_set_enforce_isolation(adev, i, true);
832 			else
833 				r |= amdgpu_mes_set_enforce_isolation(adev, i, false);
834 		}
835 		mutex_unlock(&adev->enforce_isolation_mutex);
836 	}
837 	return r;
838 }
839 
840 #if defined(CONFIG_DEBUG_FS)
841 
842 static int amdgpu_debugfs_mes_event_log_show(struct seq_file *m, void *unused)
843 {
844 	struct amdgpu_device *adev = m->private;
845 	uint32_t *mem = (uint32_t *)(adev->mes.event_log_cpu_addr);
846 
847 	seq_hex_dump(m, "", DUMP_PREFIX_OFFSET, 32, 4,
848 		     mem, adev->mes.event_log_size, false);
849 
850 	return 0;
851 }
852 
853 DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_mes_event_log);
854 
855 #endif
856 
857 void amdgpu_debugfs_mes_event_log_init(struct amdgpu_device *adev)
858 {
859 
860 #if defined(CONFIG_DEBUG_FS)
861 	struct drm_minor *minor = adev_to_drm(adev)->primary;
862 	struct dentry *root = minor->debugfs_root;
863 	if (adev->enable_mes && amdgpu_mes_log_enable)
864 		debugfs_create_file("amdgpu_mes_event_log", 0444, root,
865 				    adev, &amdgpu_debugfs_mes_event_log_fops);
866 
867 #endif
868 }
869