xref: /linux/drivers/gpu/drm/amd/amdgpu/mes_userqueue.c (revision 570f7e331f5febb30f1384817463c7e42b65ca7d)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright 2024 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  *
23  */
24 #include <drm/drm_drv.h>
25 #include "amdgpu.h"
26 #include "amdgpu_gfx.h"
27 #include "mes_userqueue.h"
28 #include "amdgpu_userq_fence.h"
29 #include "amdgpu_trace.h"
30 
31 #define AMDGPU_USERQ_PROC_CTX_SZ PAGE_SIZE
32 #define AMDGPU_USERQ_GANG_CTX_SZ PAGE_SIZE
33 
34 static int
35 mes_userq_create_wptr_mapping(struct amdgpu_device *adev,
36 			      struct amdgpu_userq_mgr *uq_mgr,
37 			      struct amdgpu_usermode_queue *queue,
38 			      uint64_t wptr)
39 {
40 	struct amdgpu_bo_va_mapping *wptr_mapping;
41 	struct amdgpu_userq_obj *wptr_obj = &queue->wptr_obj;
42 	struct amdgpu_bo *obj;
43 	struct amdgpu_vm *vm = queue->vm;
44 	struct drm_exec exec;
45 	int ret;
46 
47 	wptr &= AMDGPU_GMC_HOLE_MASK;
48 
49 	drm_exec_init(&exec, DRM_EXEC_IGNORE_DUPLICATES, 2);
50 	drm_exec_until_all_locked(&exec) {
51 		ret = amdgpu_vm_lock_pd(vm, &exec, 1);
52 		drm_exec_retry_on_contention(&exec);
53 		if (unlikely(ret))
54 			goto fail_lock;
55 
56 		wptr_mapping = amdgpu_vm_bo_lookup_mapping(vm, wptr >> PAGE_SHIFT);
57 		if (!wptr_mapping) {
58 			ret = -EINVAL;
59 			goto fail_lock;
60 		}
61 
62 		obj = wptr_mapping->bo_va->base.bo;
63 		ret = drm_exec_lock_obj(&exec, &obj->tbo.base);
64 		drm_exec_retry_on_contention(&exec);
65 		if (unlikely(ret))
66 			goto fail_lock;
67 	}
68 
69 	wptr_obj->obj = amdgpu_bo_ref(wptr_mapping->bo_va->base.bo);
70 	if (wptr_obj->obj->tbo.base.size > PAGE_SIZE) {
71 		ret = -EINVAL;
72 		goto fail_map;
73 	}
74 	/* Keep WPTR BO under eviction-fence control instead of pinning. */
75 	ret = amdgpu_evf_mgr_attach_fence(&uq_mgr_to_fpriv(uq_mgr)->evf_mgr, wptr_obj->obj);
76 	if (ret) {
77 		DRM_ERROR("Failed to attach eviction fence to wptr bo. ret %d\n", ret);
78 		goto fail_map;
79 	}
80 
81 	ret = amdgpu_ttm_alloc_gart(&wptr_obj->obj->tbo);
82 	if (ret) {
83 		DRM_ERROR("Failed to bind wptr bo to GART. ret %d\n", ret);
84 		goto fail_map;
85 	}
86 
87 	queue->wptr_obj.gpu_addr = amdgpu_bo_gpu_offset(wptr_obj->obj);
88 
89 	drm_exec_fini(&exec);
90 	return 0;
91 fail_map:
92 	amdgpu_bo_unref(&wptr_obj->obj);
93 fail_lock:
94 	drm_exec_fini(&exec);
95 	return ret;
96 
97 }
98 
99 static int convert_to_mes_priority(int priority)
100 {
101 	switch (priority) {
102 	case AMDGPU_USERQ_CREATE_FLAGS_QUEUE_PRIORITY_NORMAL_LOW:
103 	default:
104 		return AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
105 	case AMDGPU_USERQ_CREATE_FLAGS_QUEUE_PRIORITY_LOW:
106 		return AMDGPU_MES_PRIORITY_LEVEL_LOW;
107 	case AMDGPU_USERQ_CREATE_FLAGS_QUEUE_PRIORITY_NORMAL_HIGH:
108 		return AMDGPU_MES_PRIORITY_LEVEL_MEDIUM;
109 	case AMDGPU_USERQ_CREATE_FLAGS_QUEUE_PRIORITY_HIGH:
110 		return AMDGPU_MES_PRIORITY_LEVEL_HIGH;
111 	}
112 }
113 
114 static int mes_userq_map(struct amdgpu_usermode_queue *queue)
115 {
116 	struct amdgpu_userq_mgr *uq_mgr = queue->userq_mgr;
117 	struct amdgpu_device *adev = uq_mgr->adev;
118 	struct amdgpu_userq_obj *ctx = &queue->fw_obj;
119 	struct amdgpu_mqd_prop *userq_props = queue->userq_prop;
120 	struct mes_add_queue_input queue_input;
121 	struct amdgpu_mes *mes = &adev->mes;
122 	int r;
123 
124 	memset(&queue_input, 0x0, sizeof(struct mes_add_queue_input));
125 
126 	queue_input.process_va_start = 0;
127 	queue_input.process_va_end = adev->vm_manager.max_pfn - 1;
128 
129 	/* set process quantum to 10 ms and gang quantum to 1 ms as default */
130 	queue_input.process_quantum = 100000;
131 	queue_input.gang_quantum = 10000;
132 	queue_input.paging = false;
133 
134 	queue_input.process_context_addr = uq_mgr->proc_ctx_obj.gpu_addr;
135 	queue_input.gang_context_addr = ctx->gpu_addr;
136 	queue_input.inprocess_gang_priority = AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
137 	queue_input.gang_global_priority_level = convert_to_mes_priority(queue->priority);
138 
139 	queue_input.process_id = queue->vm->pasid;
140 	queue_input.queue_type = queue->queue_type;
141 	queue_input.mqd_addr = queue->mqd.gpu_addr;
142 	queue_input.wptr_addr = userq_props->wptr_gpu_addr;
143 	queue_input.queue_size = userq_props->queue_size >> 2;
144 	queue_input.doorbell_offset = userq_props->doorbell_index;
145 	queue_input.page_table_base_addr = amdgpu_gmc_pd_addr(queue->vm->root.bo);
146 	queue_input.wptr_mc_addr = queue->wptr_obj.gpu_addr;
147 
148 	if (mes->use_rs64mem) {
149 		 if (!uq_mgr->proc_ctx_allocated) {
150 			r = amdgpu_mes_alloc_proc_ctx_index(mes, &uq_mgr->proc_ctx_array_index);
151 			if (r) {
152 				DRM_ERROR("Failed to allocate userq process index err:%d\n", r);
153 				return r;
154 			}
155 			uq_mgr->proc_ctx_allocated = true;
156 		}
157 
158 		r = amdgpu_mes_alloc_gang_ctx_index(mes, &queue->gang_ctx_array_index);
159 		if (r) {
160 			DRM_ERROR("Failed to allocate userq gang index err:%d\n", r);
161 			return r;
162 		}
163 		queue_input.process_context_array_index = uq_mgr->proc_ctx_array_index;
164 		queue_input.gang_context_array_index = queue->gang_ctx_array_index;
165 	}
166 	amdgpu_mes_lock(&adev->mes);
167 	r = adev->mes.funcs->add_hw_queue(&adev->mes, &queue_input);
168 	amdgpu_mes_unlock(&adev->mes);
169 	if (r) {
170 		DRM_ERROR("Failed to map queue in HW, err (%d)\n", r);
171 		return r;
172 	}
173 
174 	DRM_DEBUG_DRIVER("Queue (doorbell:%d) mapped successfully\n", userq_props->doorbell_index);
175 	return 0;
176 }
177 
178 static int mes_userq_unmap(struct amdgpu_usermode_queue *queue)
179 {
180 	struct amdgpu_userq_mgr *uq_mgr = queue->userq_mgr;
181 	struct amdgpu_device *adev = uq_mgr->adev;
182 	struct mes_remove_queue_input queue_input;
183 	struct amdgpu_userq_obj *ctx = &queue->fw_obj;
184 	struct amdgpu_mes *mes = &adev->mes;
185 	int r;
186 
187 	memset(&queue_input, 0x0, sizeof(struct mes_remove_queue_input));
188 	queue_input.gang_context_array_index = queue->gang_ctx_array_index;
189 	queue_input.doorbell_offset = queue->doorbell_index;
190 	queue_input.gang_context_addr = ctx->gpu_addr;
191 	queue_input.queue_type = queue->queue_type;
192 
193 	amdgpu_mes_lock(&adev->mes);
194 	r = adev->mes.funcs->remove_hw_queue(&adev->mes, &queue_input);
195 	amdgpu_mes_unlock(&adev->mes);
196 	if (mes->use_rs64mem)
197 		amdgpu_mes_free_gang_ctx_index(mes, queue->gang_ctx_array_index);
198 	if (r)
199 		DRM_ERROR("Failed to unmap queue in HW, err (%d)\n", r);
200 	return r;
201 }
202 
203 int mes_userq_reset(struct amdgpu_usermode_queue *queue)
204 {
205 	struct amdgpu_userq_mgr *uq_mgr = queue->userq_mgr;
206 	struct amdgpu_device *adev = uq_mgr->adev;
207 	struct mes_reset_queue_input queue_input;
208 	int r;
209 
210 	/* XXX: add a FW version check for SDMA per queue reset */
211 	memset(&queue_input, 0x0, sizeof(struct mes_reset_queue_input));
212 	queue_input.doorbell_offset = queue->doorbell_index;
213 	queue_input.queue_type = queue->queue_type;
214 
215 	amdgpu_mes_lock(&adev->mes);
216 	r = adev->mes.funcs->reset_hw_queue(&adev->mes, &queue_input);
217 	amdgpu_mes_unlock(&adev->mes);
218 	if (r)
219 		return r;
220 
221 	/* mes_userq_unmap() does not update queue->state; mark it UNMAPPED so the
222 	 * destroy path does not issue a second REMOVE_QUEUE for the removed queue.
223 	 */
224 	r = mes_userq_unmap(queue);
225 	if (!r) {
226 		trace_amdgpu_userq_state_changed(queue, AMDGPU_USERQ_STATE_UNMAPPED);
227 		queue->state = AMDGPU_USERQ_STATE_UNMAPPED;
228 	}
229 	return r;
230 }
231 
232 int mes_userq_reset_queue(struct amdgpu_device *adev,
233 			  struct amdgpu_usermode_queue *guilty_uq,
234 			  int queue_type,
235 			  unsigned int pipe,
236 			  unsigned int queue,
237 			  unsigned int db)
238 {
239 	struct amdgpu_usermode_queue *uq;
240 	bool use_mmio = adev->gfx.mec.use_mmio_for_reset;
241 	unsigned long uq_id;
242 	int r;
243 
244 	xa_for_each(&adev->userq_doorbell_xa, uq_id, uq) {
245 		if (uq->queue_type == queue_type) {
246 			if (uq == guilty_uq)
247 				continue;
248 			if (uq->doorbell_index == db) {
249 				uq->state = AMDGPU_USERQ_STATE_HUNG;
250 				if (use_mmio)
251 					r = amdgpu_mes_reset_queue_mmio(adev, queue_type, 0, 1, pipe, queue, 0);
252 				else
253 					r = amdgpu_mes_reset_user_queue(adev, queue_type, db, 0);
254 				if (r)
255 					return r;
256 				r = mes_userq_unmap(uq);
257 				if (r)
258 					return r;
259 				amdgpu_userq_fence_driver_force_completion(uq);
260 				break;
261 			}
262 		}
263 	}
264 	return 0;
265 }
266 
267 static int mes_userq_create_ctx_space(struct amdgpu_userq_mgr *uq_mgr,
268 				      struct amdgpu_usermode_queue *queue,
269 				      struct drm_amdgpu_userq_in *mqd_user)
270 {
271 	struct amdgpu_userq_obj *ctx = &queue->fw_obj;
272 	int r, size;
273 
274 	/* The FW expects at least one page space allocated for gang ctx. */
275 	size = AMDGPU_USERQ_GANG_CTX_SZ;
276 	r = amdgpu_bo_create_kernel(uq_mgr->adev, size, 0,
277 				    AMDGPU_GEM_DOMAIN_GTT,
278 				    &ctx->obj, &ctx->gpu_addr,
279 				    &ctx->cpu_ptr);
280 	if (r) {
281 		DRM_ERROR("Failed to allocate ctx space bo for userqueue, err:%d\n", r);
282 		return r;
283 	}
284 
285 	memset(ctx->cpu_ptr, 0, size);
286 	return 0;
287 }
288 
289 static int mes_userq_create_proc_ctx_space(struct amdgpu_userq_mgr *uq_mgr)
290 {
291 	int r = 0;
292 
293 	mutex_lock(&uq_mgr->proc_ctx_lock);
294 	/* This check is a necessary because amdgpu_bo_create_kernel()
295 	 * calls helpers like amdgpu_bo_pin() and memset() unconditionally
296 	 */
297 	if (!uq_mgr->proc_ctx_obj.obj) {
298 		r = amdgpu_bo_create_kernel(uq_mgr->adev, AMDGPU_USERQ_PROC_CTX_SZ,
299 					    0, AMDGPU_GEM_DOMAIN_GTT,
300 					    &uq_mgr->proc_ctx_obj.obj,
301 					    &uq_mgr->proc_ctx_obj.gpu_addr,
302 					    &uq_mgr->proc_ctx_obj.cpu_ptr);
303 
304 		if (!r)
305 			memset(uq_mgr->proc_ctx_obj.cpu_ptr, 0, AMDGPU_USERQ_PROC_CTX_SZ);
306 	}
307 
308 	mutex_unlock(&uq_mgr->proc_ctx_lock);
309 
310 	return r;
311 }
312 
313 static int mes_userq_mqd_create(struct amdgpu_usermode_queue *queue,
314 				struct drm_amdgpu_userq_in *args_in)
315 {
316 	struct amdgpu_userq_mgr *uq_mgr = queue->userq_mgr;
317 	struct amdgpu_device *adev = uq_mgr->adev;
318 	struct amdgpu_mqd *mqd_hw_default = &adev->mqds[queue->queue_type];
319 	struct drm_amdgpu_userq_in *mqd_user = args_in;
320 	struct amdgpu_mqd_prop *userq_props;
321 	int r;
322 
323 	/* Structure to initialize MQD for userqueue using generic MQD init function */
324 	userq_props = kzalloc_obj(struct amdgpu_mqd_prop);
325 	if (!userq_props) {
326 		DRM_ERROR("Failed to allocate memory for userq_props\n");
327 		return -ENOMEM;
328 	}
329 
330 	r = amdgpu_bo_create_kernel(adev,
331 				    AMDGPU_MQD_SIZE_ALIGN(mqd_hw_default->mqd_size),
332 				    0, AMDGPU_GEM_DOMAIN_GTT,
333 				    &queue->mqd.obj, &queue->mqd.gpu_addr,
334 				    &queue->mqd.cpu_ptr);
335 	if (r) {
336 		DRM_ERROR("Failed to create MQD object for userqueue\n");
337 		goto free_props;
338 	}
339 
340 	memset(queue->mqd.cpu_ptr, 0,
341 	       AMDGPU_MQD_SIZE_ALIGN(mqd_hw_default->mqd_size));
342 
343 	/* Initialize the MQD BO with user given values */
344 	userq_props->wptr_gpu_addr = mqd_user->wptr_va;
345 	userq_props->rptr_gpu_addr = mqd_user->rptr_va;
346 	userq_props->queue_size = mqd_user->queue_size;
347 	userq_props->hqd_base_gpu_addr = mqd_user->queue_va;
348 	userq_props->mqd_gpu_addr = queue->mqd.gpu_addr;
349 	userq_props->use_doorbell = true;
350 	userq_props->doorbell_index = queue->doorbell_index;
351 	userq_props->fence_address = queue->fence_drv->gpu_addr;
352 
353 	if (queue->queue_type == AMDGPU_HW_IP_COMPUTE) {
354 		struct drm_amdgpu_userq_mqd_compute_gfx11 *compute_mqd;
355 
356 		if (mqd_user->mqd_size != sizeof(*compute_mqd)) {
357 			DRM_ERROR("Invalid compute IP MQD size\n");
358 			r = -EINVAL;
359 			goto free_mqd;
360 		}
361 
362 		compute_mqd = memdup_user(u64_to_user_ptr(mqd_user->mqd), mqd_user->mqd_size);
363 		if (IS_ERR(compute_mqd)) {
364 			DRM_ERROR("Failed to read user MQD\n");
365 			r = -ENOMEM;
366 			goto free_mqd;
367 		}
368 
369 		r = amdgpu_bo_reserve(queue->vm->root.bo, false);
370 		if (r) {
371 			kfree(compute_mqd);
372 			goto free_mqd;
373 		}
374 		r = amdgpu_userq_input_va_validate(adev, queue,
375 						   compute_mqd->eop_va, 2048,
376 						   &queue->userq_vas.va.eop);
377 		amdgpu_bo_unreserve(queue->vm->root.bo);
378 		if (r) {
379 			kfree(compute_mqd);
380 			goto free_mqd;
381 		}
382 
383 		userq_props->eop_gpu_addr = compute_mqd->eop_va;
384 		userq_props->hqd_pipe_priority = AMDGPU_GFX_PIPE_PRIO_NORMAL;
385 		userq_props->hqd_queue_priority = AMDGPU_GFX_QUEUE_PRIORITY_MINIMUM;
386 		userq_props->hqd_active = false;
387 		userq_props->tmz_queue =
388 			mqd_user->flags & AMDGPU_USERQ_CREATE_FLAGS_QUEUE_SECURE;
389 		kfree(compute_mqd);
390 	} else if (queue->queue_type == AMDGPU_HW_IP_GFX) {
391 		struct drm_amdgpu_userq_mqd_gfx11 *mqd_gfx_v11;
392 		struct amdgpu_gfx_shadow_info shadow_info;
393 
394 		if (adev->gfx.funcs->get_gfx_shadow_info) {
395 			adev->gfx.funcs->get_gfx_shadow_info(adev, &shadow_info, true);
396 		} else {
397 			r = -EINVAL;
398 			goto free_mqd;
399 		}
400 
401 		if (mqd_user->mqd_size != sizeof(*mqd_gfx_v11) || !mqd_user->mqd) {
402 			DRM_ERROR("Invalid GFX MQD\n");
403 			r = -EINVAL;
404 			goto free_mqd;
405 		}
406 
407 		mqd_gfx_v11 = memdup_user(u64_to_user_ptr(mqd_user->mqd), mqd_user->mqd_size);
408 		if (IS_ERR(mqd_gfx_v11)) {
409 			DRM_ERROR("Failed to read user MQD\n");
410 			r = -ENOMEM;
411 			goto free_mqd;
412 		}
413 
414 		userq_props->shadow_addr = mqd_gfx_v11->shadow_va;
415 		userq_props->csa_addr = mqd_gfx_v11->csa_va;
416 		userq_props->tmz_queue =
417 			mqd_user->flags & AMDGPU_USERQ_CREATE_FLAGS_QUEUE_SECURE;
418 
419 		r = amdgpu_bo_reserve(queue->vm->root.bo, false);
420 		if (r) {
421 			kfree(mqd_gfx_v11);
422 			goto free_mqd;
423 		}
424 		r = amdgpu_userq_input_va_validate(adev, queue, mqd_gfx_v11->shadow_va,
425 						   shadow_info.shadow_size,
426 						   &queue->userq_vas.va.shadow);
427 		if (r) {
428 			amdgpu_bo_unreserve(queue->vm->root.bo);
429 			kfree(mqd_gfx_v11);
430 			goto free_mqd;
431 		}
432 
433 		r = amdgpu_userq_input_va_validate(adev, queue, mqd_gfx_v11->csa_va,
434 						   shadow_info.csa_size,
435 						   &queue->userq_vas.va.csa);
436 		amdgpu_bo_unreserve(queue->vm->root.bo);
437 		if (r) {
438 			kfree(mqd_gfx_v11);
439 			goto free_mqd;
440 		}
441 
442 		kfree(mqd_gfx_v11);
443 	} else if (queue->queue_type == AMDGPU_HW_IP_DMA) {
444 		struct drm_amdgpu_userq_mqd_sdma_gfx11 *mqd_sdma_v11;
445 
446 		if (mqd_user->mqd_size != sizeof(*mqd_sdma_v11) || !mqd_user->mqd) {
447 			DRM_ERROR("Invalid SDMA MQD\n");
448 			r = -EINVAL;
449 			goto free_mqd;
450 		}
451 
452 		mqd_sdma_v11 = memdup_user(u64_to_user_ptr(mqd_user->mqd), mqd_user->mqd_size);
453 		if (IS_ERR(mqd_sdma_v11)) {
454 			DRM_ERROR("Failed to read sdma user MQD\n");
455 			r = -ENOMEM;
456 			goto free_mqd;
457 		}
458 
459 		r = amdgpu_bo_reserve(queue->vm->root.bo, false);
460 		if (r) {
461 			kfree(mqd_sdma_v11);
462 			goto free_mqd;
463 		}
464 		r = amdgpu_userq_input_va_validate(adev, queue, mqd_sdma_v11->csa_va,
465 						   32,
466 						   &queue->userq_vas.va.csa);
467 		amdgpu_bo_unreserve(queue->vm->root.bo);
468 		if (r) {
469 			kfree(mqd_sdma_v11);
470 			goto free_mqd;
471 		}
472 
473 		userq_props->csa_addr = mqd_sdma_v11->csa_va;
474 		kfree(mqd_sdma_v11);
475 	}
476 
477 	queue->userq_prop = userq_props;
478 
479 	r = mqd_hw_default->init_mqd(adev, (void *)queue->mqd.cpu_ptr, userq_props);
480 	if (r) {
481 		DRM_ERROR("Failed to initialize MQD for userqueue\n");
482 		goto free_mqd;
483 	}
484 
485 	/* Create per-process MES process context BO */
486 	r = mes_userq_create_proc_ctx_space(uq_mgr);
487 	if (r) {
488 		DRM_ERROR("Failed to allocate MES process context space bo, error: %d\n", r);
489 		goto free_mqd;
490 	}
491 
492 	/* Create BO of a gang for FW operations */
493 	r = mes_userq_create_ctx_space(uq_mgr, queue, mqd_user);
494 	if (r) {
495 		DRM_ERROR("Failed to allocate BO for userqueue (%d)", r);
496 		goto free_mqd;
497 	}
498 
499 	/* FW expects WPTR BOs to be mapped into GART */
500 	r = mes_userq_create_wptr_mapping(adev, uq_mgr, queue, userq_props->wptr_gpu_addr);
501 	if (r) {
502 		DRM_ERROR("Failed to create WPTR mapping\n");
503 		goto free_ctx;
504 	}
505 
506 	return 0;
507 
508 free_ctx:
509 	amdgpu_bo_free_kernel(&queue->fw_obj.obj, &queue->fw_obj.gpu_addr,
510 			      &queue->fw_obj.cpu_ptr);
511 
512 free_mqd:
513 	amdgpu_bo_free_kernel(&queue->mqd.obj, &queue->mqd.gpu_addr,
514 			      &queue->mqd.cpu_ptr);
515 
516 free_props:
517 	kfree(userq_props);
518 
519 	return r;
520 }
521 
522 static void mes_userq_mqd_destroy(struct amdgpu_usermode_queue *queue)
523 {
524 
525 	amdgpu_bo_free_kernel(&queue->fw_obj.obj, &queue->fw_obj.gpu_addr,
526 			      &queue->fw_obj.cpu_ptr);
527 	kfree(queue->userq_prop);
528 	amdgpu_bo_free_kernel(&queue->mqd.obj, &queue->mqd.gpu_addr,
529 			      &queue->mqd.cpu_ptr);
530 
531 	amdgpu_bo_unref(&queue->wptr_obj.obj);
532 }
533 
534 static int mes_userq_preempt(struct amdgpu_usermode_queue *queue)
535 {
536 	struct amdgpu_userq_mgr *uq_mgr = queue->userq_mgr;
537 	struct amdgpu_device *adev = uq_mgr->adev;
538 	struct mes_suspend_gang_input queue_input;
539 	struct amdgpu_userq_obj *ctx = &queue->fw_obj;
540 	signed long timeout = 2100000; /* 2100 ms */
541 	u64 fence_gpu_addr;
542 	u32 fence_offset;
543 	u64 *fence_ptr;
544 	int i, r;
545 
546 	if (queue->state != AMDGPU_USERQ_STATE_MAPPED)
547 		return 0;
548 	r = amdgpu_wb_get(adev, &fence_offset);
549 	if (r)
550 		return r;
551 
552 	fence_gpu_addr = adev->wb.gpu_addr + (fence_offset * 4);
553 	fence_ptr = (u64 *)&adev->wb.wb[fence_offset];
554 	*fence_ptr = 0;
555 
556 	memset(&queue_input, 0x0, sizeof(struct mes_suspend_gang_input));
557 	queue_input.gang_context_addr = ctx->gpu_addr;
558 	queue_input.suspend_fence_addr = fence_gpu_addr;
559 	queue_input.suspend_fence_value = 1;
560 	amdgpu_mes_lock(&adev->mes);
561 	r = adev->mes.funcs->suspend_gang(&adev->mes, &queue_input);
562 	amdgpu_mes_unlock(&adev->mes);
563 	if (r) {
564 		DRM_ERROR("Failed to suspend gang: %d\n", r);
565 		goto out;
566 	}
567 
568 	for (i = 0; i < timeout; i++) {
569 		if (*fence_ptr == 1)
570 			goto out;
571 		udelay(1);
572 	}
573 	r = -ETIMEDOUT;
574 
575 out:
576 	amdgpu_wb_free(adev, fence_offset);
577 	return r;
578 }
579 
580 static int mes_userq_restore(struct amdgpu_usermode_queue *queue)
581 {
582 	struct amdgpu_userq_mgr *uq_mgr = queue->userq_mgr;
583 	struct amdgpu_device *adev = uq_mgr->adev;
584 	struct mes_resume_gang_input queue_input;
585 	struct amdgpu_userq_obj *ctx = &queue->fw_obj;
586 	int r;
587 
588 	if (queue->state == AMDGPU_USERQ_STATE_HUNG)
589 		return -EINVAL;
590 	if (queue->state != AMDGPU_USERQ_STATE_PREEMPTED)
591 		return 0;
592 
593 	memset(&queue_input, 0x0, sizeof(struct mes_resume_gang_input));
594 	queue_input.gang_context_addr = ctx->gpu_addr;
595 
596 	amdgpu_mes_lock(&adev->mes);
597 	r = adev->mes.funcs->resume_gang(&adev->mes, &queue_input);
598 	amdgpu_mes_unlock(&adev->mes);
599 	if (r)
600 		dev_err(adev->dev, "Failed to resume queue, err (%d)\n", r);
601 	return r;
602 }
603 
604 const struct amdgpu_userq_funcs userq_mes_funcs = {
605 	.mqd_create = mes_userq_mqd_create,
606 	.mqd_destroy = mes_userq_mqd_destroy,
607 	.unmap = mes_userq_unmap,
608 	.map = mes_userq_map,
609 	.preempt = mes_userq_preempt,
610 	.restore = mes_userq_restore,
611 	.reset = mes_userq_reset,
612 };
613