xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c (revision d639d9fa162aadec1ae9980c4dcf6e50bd2f8290)
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 	if ((amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(12, 1, 0)) &&
316 		((adev->mes.sched_version & AMDGPU_MES_VERSION_MASK) >= 0x71))
317 		input.suspend_all_sdma_gangs = 1;
318 
319 	/*
320 	 * Avoid taking any other locks under MES lock to avoid circular
321 	 * lock dependencies.
322 	 */
323 	amdgpu_mes_lock(&adev->mes);
324 	r = adev->mes.funcs->suspend_gang(&adev->mes, &input);
325 	amdgpu_mes_unlock(&adev->mes);
326 	if (r)
327 		dev_err(adev->dev, "failed to suspend all gangs");
328 
329 	return r;
330 }
331 
332 int amdgpu_mes_resume(struct amdgpu_device *adev, u32 xcc_id)
333 {
334 	struct mes_resume_gang_input input;
335 	int r;
336 
337 	if (!amdgpu_mes_suspend_resume_all_supported(adev))
338 		return 0;
339 
340 	memset(&input, 0x0, sizeof(struct mes_resume_gang_input));
341 	input.resume_all_gangs = 1;
342 	input.xcc_id = xcc_id;
343 
344 	/*
345 	 * Avoid taking any other locks under MES lock to avoid circular
346 	 * lock dependencies.
347 	 */
348 	amdgpu_mes_lock(&adev->mes);
349 	r = adev->mes.funcs->resume_gang(&adev->mes, &input);
350 	amdgpu_mes_unlock(&adev->mes);
351 	if (r)
352 		dev_err(adev->dev, "failed to resume all gangs");
353 
354 	return r;
355 }
356 
357 int amdgpu_mes_map_legacy_queue(struct amdgpu_device *adev,
358 				struct amdgpu_ring *ring, uint32_t xcc_id)
359 {
360 	struct mes_map_legacy_queue_input queue_input;
361 	int r;
362 
363 	memset(&queue_input, 0, sizeof(queue_input));
364 
365 	queue_input.xcc_id = xcc_id;
366 	queue_input.queue_type = ring->funcs->type;
367 	queue_input.doorbell_offset = ring->doorbell_index;
368 	queue_input.pipe_id = ring->pipe;
369 	queue_input.queue_id = ring->queue;
370 	queue_input.mqd_addr = amdgpu_bo_gpu_offset(ring->mqd_obj);
371 	queue_input.wptr_addr = ring->wptr_gpu_addr;
372 
373 	amdgpu_mes_lock(&adev->mes);
374 	r = adev->mes.funcs->map_legacy_queue(&adev->mes, &queue_input);
375 	amdgpu_mes_unlock(&adev->mes);
376 	if (r)
377 		dev_err(adev->dev, "failed to map legacy queue\n");
378 
379 	return r;
380 }
381 
382 int amdgpu_mes_unmap_legacy_queue(struct amdgpu_device *adev,
383 				  struct amdgpu_ring *ring,
384 				  enum amdgpu_unmap_queues_action action,
385 				  u64 gpu_addr, u64 seq, uint32_t xcc_id)
386 {
387 	struct mes_unmap_legacy_queue_input queue_input;
388 	int r;
389 
390 	queue_input.xcc_id = xcc_id;
391 	queue_input.action = action;
392 	queue_input.queue_type = ring->funcs->type;
393 	queue_input.doorbell_offset = ring->doorbell_index;
394 	queue_input.pipe_id = ring->pipe;
395 	queue_input.queue_id = ring->queue;
396 	queue_input.trail_fence_addr = gpu_addr;
397 	queue_input.trail_fence_data = seq;
398 
399 	amdgpu_mes_lock(&adev->mes);
400 	r = adev->mes.funcs->unmap_legacy_queue(&adev->mes, &queue_input);
401 	amdgpu_mes_unlock(&adev->mes);
402 	if (r)
403 		dev_err(adev->dev, "failed to unmap legacy queue\n");
404 
405 	return r;
406 }
407 
408 int amdgpu_mes_reset_legacy_queue(struct amdgpu_device *adev,
409 				  struct amdgpu_ring *ring,
410 				  unsigned int vmid,
411 				  bool use_mmio,
412 				  uint32_t xcc_id)
413 {
414 	struct mes_reset_queue_input queue_input;
415 	int r;
416 
417 	memset(&queue_input, 0, sizeof(queue_input));
418 
419 	queue_input.xcc_id = xcc_id;
420 	queue_input.queue_type = ring->funcs->type;
421 	queue_input.doorbell_offset = ring->doorbell_index;
422 	queue_input.me_id = ring->me;
423 	queue_input.pipe_id = ring->pipe;
424 	queue_input.queue_id = ring->queue;
425 	queue_input.mqd_addr = ring->mqd_obj ? amdgpu_bo_gpu_offset(ring->mqd_obj) : 0;
426 	queue_input.wptr_addr = ring->wptr_gpu_addr;
427 	queue_input.vmid = vmid;
428 	queue_input.use_mmio = use_mmio;
429 	queue_input.is_kq = true;
430 	if (ring->funcs->type == AMDGPU_RING_TYPE_GFX)
431 		queue_input.legacy_gfx = true;
432 
433 	amdgpu_mes_lock(&adev->mes);
434 	r = adev->mes.funcs->reset_hw_queue(&adev->mes, &queue_input);
435 	amdgpu_mes_unlock(&adev->mes);
436 	if (r)
437 		dev_err(adev->dev, "failed to reset legacy queue\n");
438 
439 	return r;
440 }
441 
442 int amdgpu_mes_get_hung_queue_db_array_size(struct amdgpu_device *adev)
443 {
444 	return adev->mes.hung_queue_db_array_size;
445 }
446 
447 int amdgpu_mes_detect_and_reset_hung_queues(struct amdgpu_device *adev,
448 					    int queue_type,
449 					    bool detect_only,
450 					    unsigned int *hung_db_num,
451 					    u32 *hung_db_array,
452 					    uint32_t xcc_id)
453 {
454 	struct mes_detect_and_reset_queue_input input;
455 	u32 *db_array = adev->mes.hung_queue_db_array_cpu_addr[xcc_id];
456 	int hqd_info_offset = adev->mes.hung_queue_hqd_info_offset, r, i;
457 
458 	if (!hung_db_num || !hung_db_array)
459 		return -EINVAL;
460 
461 	if ((queue_type != AMDGPU_RING_TYPE_GFX) &&
462 	    (queue_type != AMDGPU_RING_TYPE_COMPUTE) &&
463 	    (queue_type != AMDGPU_RING_TYPE_SDMA))
464 		return -EINVAL;
465 
466 	/* Clear the doorbell array before detection */
467 	memset(adev->mes.hung_queue_db_array_cpu_addr[xcc_id], AMDGPU_MES_INVALID_DB_OFFSET,
468 		adev->mes.hung_queue_db_array_size * sizeof(u32));
469 	input.queue_type = queue_type;
470 	input.detect_only = detect_only;
471 	input.xcc_id = xcc_id;
472 
473 	r = adev->mes.funcs->detect_and_reset_hung_queues(&adev->mes,
474 							  &input);
475 
476 	if (r && detect_only) {
477 		dev_err(adev->dev, "Failed to detect hung queues\n");
478 		return r;
479 	}
480 
481 	*hung_db_num = 0;
482 	/* MES passes hung queues' doorbell to driver */
483 	for (i = 0; i < adev->mes.hung_queue_hqd_info_offset; i++) {
484 		/* Finding hung queues where db_array[i] is a valid doorbell */
485 		if (db_array[i] != AMDGPU_MES_INVALID_DB_OFFSET) {
486 			hung_db_array[i] = db_array[i];
487 			*hung_db_num += 1;
488 		}
489 	}
490 
491 	if (r && !(*hung_db_num)) {
492 		dev_err(adev->dev, "Failed to detect and reset hung queues\n");
493 		return r;
494 	}
495 
496 	for (i = hqd_info_offset; i < hqd_info_offset + *hung_db_num; i++)
497 		hung_db_array[i] = db_array[i];
498 
499 	return r;
500 }
501 
502 uint32_t amdgpu_mes_rreg(struct amdgpu_device *adev, uint32_t reg,
503 			 uint32_t xcc_id)
504 {
505 	struct mes_misc_op_input op_input;
506 	int r, val = 0;
507 	uint32_t addr_offset = 0;
508 	uint64_t read_val_gpu_addr;
509 	uint32_t *read_val_ptr;
510 
511 	if (amdgpu_device_wb_get(adev, &addr_offset)) {
512 		dev_err(adev->dev, "critical bug! too many mes readers\n");
513 		goto error;
514 	}
515 	read_val_gpu_addr = adev->wb.gpu_addr + (addr_offset * 4);
516 	read_val_ptr = (uint32_t *)&adev->wb.wb[addr_offset];
517 	op_input.xcc_id = xcc_id;
518 	op_input.op = MES_MISC_OP_READ_REG;
519 	op_input.read_reg.reg_offset = reg;
520 	op_input.read_reg.buffer_addr = read_val_gpu_addr;
521 
522 	if (!adev->mes.funcs->misc_op) {
523 		dev_err(adev->dev, "mes rreg is not supported!\n");
524 		goto error;
525 	}
526 
527 	amdgpu_mes_lock(&adev->mes);
528 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
529 	amdgpu_mes_unlock(&adev->mes);
530 	if (r)
531 		dev_err(adev->dev, "failed to read reg (0x%x)\n", reg);
532 	else
533 		val = *(read_val_ptr);
534 
535 error:
536 	if (addr_offset)
537 		amdgpu_device_wb_free(adev, addr_offset);
538 	return val;
539 }
540 
541 int amdgpu_mes_wreg(struct amdgpu_device *adev, uint32_t reg,
542 		    uint32_t val, uint32_t xcc_id)
543 {
544 	struct mes_misc_op_input op_input;
545 	int r;
546 
547 	op_input.xcc_id = xcc_id;
548 	op_input.op = MES_MISC_OP_WRITE_REG;
549 	op_input.write_reg.reg_offset = reg;
550 	op_input.write_reg.reg_value = val;
551 
552 	if (!adev->mes.funcs->misc_op) {
553 		dev_err(adev->dev, "mes wreg is not supported!\n");
554 		r = -EINVAL;
555 		goto error;
556 	}
557 
558 	amdgpu_mes_lock(&adev->mes);
559 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
560 	amdgpu_mes_unlock(&adev->mes);
561 	if (r)
562 		dev_err(adev->dev, "failed to write reg (0x%x)\n", reg);
563 
564 error:
565 	return r;
566 }
567 
568 int amdgpu_mes_reg_write_reg_wait(struct amdgpu_device *adev,
569 				  uint32_t reg0, uint32_t reg1,
570 				  uint32_t ref, uint32_t mask,
571 				  uint32_t xcc_id)
572 {
573 	struct mes_misc_op_input op_input;
574 	int r;
575 
576 	op_input.xcc_id = xcc_id;
577 	op_input.op = MES_MISC_OP_WRM_REG_WR_WAIT;
578 	op_input.wrm_reg.reg0 = reg0;
579 	op_input.wrm_reg.reg1 = reg1;
580 	op_input.wrm_reg.ref = ref;
581 	op_input.wrm_reg.mask = mask;
582 
583 	if (!adev->mes.funcs->misc_op) {
584 		dev_err(adev->dev, "mes reg_write_reg_wait is not supported!\n");
585 		r = -EINVAL;
586 		goto error;
587 	}
588 
589 	amdgpu_mes_lock(&adev->mes);
590 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
591 	amdgpu_mes_unlock(&adev->mes);
592 	if (r)
593 		dev_err(adev->dev, "failed to reg_write_reg_wait\n");
594 
595 error:
596 	return r;
597 }
598 
599 int amdgpu_mes_hdp_flush(struct amdgpu_device *adev)
600 {
601 	uint32_t hdp_flush_req_offset, hdp_flush_done_offset;
602 	struct amdgpu_ring *mes_ring;
603 	uint32_t ref_and_mask = 0, reg_mem_engine = 0;
604 
605 	if (!adev->gfx.funcs->get_hdp_flush_mask) {
606 		dev_err(adev->dev, "mes hdp flush is not supported.\n");
607 		return -EINVAL;
608 	}
609 
610 	mes_ring = &adev->mes.ring[0];
611 	hdp_flush_req_offset = adev->nbio.funcs->get_hdp_flush_req_offset(adev);
612 	hdp_flush_done_offset = adev->nbio.funcs->get_hdp_flush_done_offset(adev);
613 
614 	adev->gfx.funcs->get_hdp_flush_mask(mes_ring, &ref_and_mask, &reg_mem_engine);
615 
616 	return amdgpu_mes_reg_write_reg_wait(adev, hdp_flush_req_offset, hdp_flush_done_offset,
617 					     ref_and_mask, ref_and_mask, 0);
618 }
619 
620 int amdgpu_mes_set_shader_debugger(struct amdgpu_device *adev,
621 				uint64_t process_context_addr,
622 				uint32_t spi_gdbg_per_vmid_cntl,
623 				const uint32_t *tcp_watch_cntl,
624 				uint32_t flags,
625 				bool trap_en,
626 				uint32_t xcc_id)
627 {
628 	struct mes_misc_op_input op_input = {0};
629 	int r;
630 
631 	if (!adev->mes.funcs->misc_op) {
632 		dev_err(adev->dev,
633 			"mes set shader debugger is not supported!\n");
634 		return -EINVAL;
635 	}
636 
637 	op_input.xcc_id = xcc_id;
638 	op_input.op = MES_MISC_OP_SET_SHADER_DEBUGGER;
639 	op_input.set_shader_debugger.process_context_addr = process_context_addr;
640 	op_input.set_shader_debugger.flags.u32all = flags;
641 
642 	/* use amdgpu mes_flush_shader_debugger instead */
643 	if (op_input.set_shader_debugger.flags.process_ctx_flush)
644 		return -EINVAL;
645 
646 	op_input.set_shader_debugger.spi_gdbg_per_vmid_cntl = spi_gdbg_per_vmid_cntl;
647 	memcpy(op_input.set_shader_debugger.tcp_watch_cntl, tcp_watch_cntl,
648 			sizeof(op_input.set_shader_debugger.tcp_watch_cntl));
649 
650 	if (((adev->mes.sched_version & AMDGPU_MES_API_VERSION_MASK) >>
651 			AMDGPU_MES_API_VERSION_SHIFT) >= 14)
652 		op_input.set_shader_debugger.trap_en = trap_en;
653 
654 	amdgpu_mes_lock(&adev->mes);
655 
656 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
657 	if (r)
658 		dev_err(adev->dev, "failed to set_shader_debugger\n");
659 
660 	amdgpu_mes_unlock(&adev->mes);
661 
662 	return r;
663 }
664 
665 int amdgpu_mes_flush_shader_debugger(struct amdgpu_device *adev,
666 				     uint64_t process_context_addr,
667 				     uint32_t xcc_id)
668 {
669 	struct mes_misc_op_input op_input = {0};
670 	int r;
671 
672 	if (!adev->mes.funcs->misc_op) {
673 		dev_err(adev->dev,
674 			"mes flush shader debugger is not supported!\n");
675 		return -EINVAL;
676 	}
677 
678 	op_input.xcc_id = xcc_id;
679 	op_input.op = MES_MISC_OP_SET_SHADER_DEBUGGER;
680 	op_input.set_shader_debugger.process_context_addr = process_context_addr;
681 	op_input.set_shader_debugger.flags.process_ctx_flush = true;
682 
683 	amdgpu_mes_lock(&adev->mes);
684 
685 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
686 	if (r)
687 		dev_err(adev->dev, "failed to set_shader_debugger\n");
688 
689 	amdgpu_mes_unlock(&adev->mes);
690 
691 	return r;
692 }
693 
694 uint32_t amdgpu_mes_get_aggregated_doorbell_index(struct amdgpu_device *adev,
695 						   enum amdgpu_mes_priority_level prio)
696 {
697 	return adev->mes.aggregated_doorbells[prio];
698 }
699 
700 int amdgpu_mes_init_microcode(struct amdgpu_device *adev, int pipe)
701 {
702 	const struct mes_firmware_header_v1_0 *mes_hdr;
703 	struct amdgpu_firmware_info *info;
704 	char ucode_prefix[30];
705 	char fw_name[50];
706 	bool need_retry = false;
707 	u32 *ucode_ptr;
708 	int r;
709 
710 	amdgpu_ucode_ip_version_decode(adev, GC_HWIP, ucode_prefix,
711 				       sizeof(ucode_prefix));
712 	if (adev->enable_uni_mes) {
713 		snprintf(fw_name, sizeof(fw_name),
714 			 "amdgpu/%s_uni_mes.bin", ucode_prefix);
715 	} else if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(11, 0, 0) &&
716 	    amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(12, 0, 0)) {
717 		snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes%s.bin",
718 			 ucode_prefix,
719 			 pipe == AMDGPU_MES_SCHED_PIPE ? "_2" : "1");
720 		need_retry = true;
721 	} else {
722 		snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes%s.bin",
723 			 ucode_prefix,
724 			 pipe == AMDGPU_MES_SCHED_PIPE ? "" : "1");
725 	}
726 
727 	r = amdgpu_ucode_request(adev, &adev->mes.fw[pipe], AMDGPU_UCODE_REQUIRED,
728 				 "%s", fw_name);
729 	if (r && need_retry && pipe == AMDGPU_MES_SCHED_PIPE) {
730 		dev_info(adev->dev, "try to fall back to %s_mes.bin\n", ucode_prefix);
731 		r = amdgpu_ucode_request(adev, &adev->mes.fw[pipe],
732 					 AMDGPU_UCODE_REQUIRED,
733 					 "amdgpu/%s_mes.bin", ucode_prefix);
734 	}
735 
736 	if (r)
737 		goto out;
738 
739 	mes_hdr = (const struct mes_firmware_header_v1_0 *)
740 		adev->mes.fw[pipe]->data;
741 	adev->mes.uc_start_addr[pipe] =
742 		le32_to_cpu(mes_hdr->mes_uc_start_addr_lo) |
743 		((uint64_t)(le32_to_cpu(mes_hdr->mes_uc_start_addr_hi)) << 32);
744 	adev->mes.data_start_addr[pipe] =
745 		le32_to_cpu(mes_hdr->mes_data_start_addr_lo) |
746 		((uint64_t)(le32_to_cpu(mes_hdr->mes_data_start_addr_hi)) << 32);
747 	ucode_ptr = (u32 *)(adev->mes.fw[pipe]->data +
748 			  sizeof(union amdgpu_firmware_header));
749 	adev->mes.fw_version[pipe] =
750 		le32_to_cpu(ucode_ptr[24]) & AMDGPU_MES_VERSION_MASK;
751 
752 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
753 		int ucode, ucode_data;
754 
755 		if (pipe == AMDGPU_MES_SCHED_PIPE) {
756 			ucode = AMDGPU_UCODE_ID_CP_MES;
757 			ucode_data = AMDGPU_UCODE_ID_CP_MES_DATA;
758 		} else {
759 			ucode = AMDGPU_UCODE_ID_CP_MES1;
760 			ucode_data = AMDGPU_UCODE_ID_CP_MES1_DATA;
761 		}
762 
763 		info = &adev->firmware.ucode[ucode];
764 		info->ucode_id = ucode;
765 		info->fw = adev->mes.fw[pipe];
766 		adev->firmware.fw_size +=
767 			ALIGN(le32_to_cpu(mes_hdr->mes_ucode_size_bytes),
768 			      PAGE_SIZE);
769 
770 		info = &adev->firmware.ucode[ucode_data];
771 		info->ucode_id = ucode_data;
772 		info->fw = adev->mes.fw[pipe];
773 		adev->firmware.fw_size +=
774 			ALIGN(le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes),
775 			      PAGE_SIZE);
776 	}
777 
778 	return 0;
779 out:
780 	amdgpu_ucode_release(&adev->mes.fw[pipe]);
781 	return r;
782 }
783 
784 void amdgpu_mes_validate_fw_version(struct amdgpu_device *adev)
785 {
786 	u32 fw_from_ucode = adev->mes.fw_version[AMDGPU_MES_SCHED_PIPE];
787 	u32 fw_from_reg = adev->mes.sched_version & AMDGPU_MES_VERSION_MASK;
788 
789 	if (fw_from_ucode != fw_from_reg)
790 		dev_info(adev->dev,
791 			 "MES firmware reports incorrect version in ucode binary (0x%x vs 0x%x)\n",
792 			 fw_from_ucode, fw_from_reg);
793 }
794 
795 
796 bool amdgpu_mes_suspend_resume_all_supported(struct amdgpu_device *adev)
797 {
798 	uint32_t mes_rev = adev->mes.sched_version & AMDGPU_MES_VERSION_MASK;
799 
800 	return ((amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(11, 0, 0) &&
801 		 amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(12, 0, 0) &&
802 		 mes_rev >= 0x63) ||
803 		amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(12, 0, 0));
804 }
805 
806 bool amdgpu_mes_queue_reset_by_mes_supported(struct amdgpu_device *adev)
807 {
808 	return (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(12, 1, 0) &&
809 		(adev->mes.sched_version & AMDGPU_MES_VERSION_MASK) >= 0x73);
810 }
811 
812 /* Fix me -- node_id is used to identify the correct MES instances in the future */
813 static int amdgpu_mes_set_enforce_isolation(struct amdgpu_device *adev,
814 					    uint32_t node_id, bool enable)
815 {
816 	struct mes_misc_op_input op_input = {0};
817 	int r;
818 
819 	op_input.op = MES_MISC_OP_CHANGE_CONFIG;
820 	op_input.change_config.option.limit_single_process = enable ? 1 : 0;
821 
822 	if (!adev->mes.funcs->misc_op) {
823 		dev_err(adev->dev, "mes change config is not supported!\n");
824 		r = -EINVAL;
825 		goto error;
826 	}
827 
828 	amdgpu_mes_lock(&adev->mes);
829 	r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
830 	amdgpu_mes_unlock(&adev->mes);
831 	if (r)
832 		dev_err(adev->dev, "failed to change_config.\n");
833 
834 error:
835 	return r;
836 }
837 
838 int amdgpu_mes_update_enforce_isolation(struct amdgpu_device *adev)
839 {
840 	int i, r = 0;
841 
842 	if (adev->enable_mes && adev->gfx.enable_cleaner_shader) {
843 		mutex_lock(&adev->enforce_isolation_mutex);
844 		for (i = 0; i < (adev->xcp_mgr ? adev->xcp_mgr->num_xcps : 1); i++) {
845 			if (adev->enforce_isolation[i] == AMDGPU_ENFORCE_ISOLATION_ENABLE)
846 				r |= amdgpu_mes_set_enforce_isolation(adev, i, true);
847 			else
848 				r |= amdgpu_mes_set_enforce_isolation(adev, i, false);
849 		}
850 		mutex_unlock(&adev->enforce_isolation_mutex);
851 	}
852 	return r;
853 }
854 
855 #if defined(CONFIG_DEBUG_FS)
856 
857 static int amdgpu_debugfs_mes_event_log_show(struct seq_file *m, void *unused)
858 {
859 	struct amdgpu_device *adev = m->private;
860 	uint32_t *mem = (uint32_t *)(adev->mes.event_log_cpu_addr);
861 
862 	seq_hex_dump(m, "", DUMP_PREFIX_OFFSET, 32, 4,
863 		     mem, adev->mes.event_log_size, false);
864 
865 	return 0;
866 }
867 
868 DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_mes_event_log);
869 
870 #endif
871 
872 void amdgpu_debugfs_mes_event_log_init(struct amdgpu_device *adev)
873 {
874 
875 #if defined(CONFIG_DEBUG_FS)
876 	struct drm_minor *minor = adev_to_drm(adev)->primary;
877 	struct dentry *root = minor->debugfs_root;
878 	if (adev->enable_mes && amdgpu_mes_log_enable)
879 		debugfs_create_file("amdgpu_mes_event_log", 0444, root,
880 				    adev, &amdgpu_debugfs_mes_event_log_fops);
881 
882 #endif
883 }
884