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
30 #define AMDGPU_USERQ_PROC_CTX_SZ PAGE_SIZE
31 #define AMDGPU_USERQ_GANG_CTX_SZ PAGE_SIZE
32
33 static int
mes_userq_map_gtt_bo_to_gart(struct amdgpu_bo * bo)34 mes_userq_map_gtt_bo_to_gart(struct amdgpu_bo *bo)
35 {
36 int ret;
37
38 ret = amdgpu_bo_reserve(bo, true);
39 if (ret) {
40 DRM_ERROR("Failed to reserve bo. ret %d\n", ret);
41 goto err_reserve_bo_failed;
42 }
43
44 ret = amdgpu_ttm_alloc_gart(&bo->tbo);
45 if (ret) {
46 DRM_ERROR("Failed to bind bo to GART. ret %d\n", ret);
47 goto err_map_bo_gart_failed;
48 }
49
50 amdgpu_bo_unreserve(bo);
51 bo = amdgpu_bo_ref(bo);
52
53 return 0;
54
55 err_map_bo_gart_failed:
56 amdgpu_bo_unreserve(bo);
57 err_reserve_bo_failed:
58 return ret;
59 }
60
61 static int
mes_userq_create_wptr_mapping(struct amdgpu_userq_mgr * uq_mgr,struct amdgpu_usermode_queue * queue,uint64_t wptr)62 mes_userq_create_wptr_mapping(struct amdgpu_userq_mgr *uq_mgr,
63 struct amdgpu_usermode_queue *queue,
64 uint64_t wptr)
65 {
66 struct amdgpu_bo_va_mapping *wptr_mapping;
67 struct amdgpu_vm *wptr_vm;
68 struct amdgpu_userq_obj *wptr_obj = &queue->wptr_obj;
69 int ret;
70
71 wptr_vm = queue->vm;
72 ret = amdgpu_bo_reserve(wptr_vm->root.bo, false);
73 if (ret)
74 return ret;
75
76 wptr &= AMDGPU_GMC_HOLE_MASK;
77 wptr_mapping = amdgpu_vm_bo_lookup_mapping(wptr_vm, wptr >> PAGE_SHIFT);
78 amdgpu_bo_unreserve(wptr_vm->root.bo);
79 if (!wptr_mapping) {
80 DRM_ERROR("Failed to lookup wptr bo\n");
81 return -EINVAL;
82 }
83
84 wptr_obj->obj = wptr_mapping->bo_va->base.bo;
85 if (wptr_obj->obj->tbo.base.size > PAGE_SIZE) {
86 DRM_ERROR("Requested GART mapping for wptr bo larger than one page\n");
87 return -EINVAL;
88 }
89
90 ret = mes_userq_map_gtt_bo_to_gart(wptr_obj->obj);
91 if (ret) {
92 DRM_ERROR("Failed to map wptr bo to GART\n");
93 return ret;
94 }
95
96 queue->wptr_obj.gpu_addr = amdgpu_bo_gpu_offset_no_check(wptr_obj->obj);
97 return 0;
98 }
99
convert_to_mes_priority(int priority)100 static int convert_to_mes_priority(int priority)
101 {
102 switch (priority) {
103 case AMDGPU_USERQ_CREATE_FLAGS_QUEUE_PRIORITY_NORMAL_LOW:
104 default:
105 return AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
106 case AMDGPU_USERQ_CREATE_FLAGS_QUEUE_PRIORITY_LOW:
107 return AMDGPU_MES_PRIORITY_LEVEL_LOW;
108 case AMDGPU_USERQ_CREATE_FLAGS_QUEUE_PRIORITY_NORMAL_HIGH:
109 return AMDGPU_MES_PRIORITY_LEVEL_MEDIUM;
110 case AMDGPU_USERQ_CREATE_FLAGS_QUEUE_PRIORITY_HIGH:
111 return AMDGPU_MES_PRIORITY_LEVEL_HIGH;
112 }
113 }
114
mes_userq_map(struct amdgpu_userq_mgr * uq_mgr,struct amdgpu_usermode_queue * queue)115 static int mes_userq_map(struct amdgpu_userq_mgr *uq_mgr,
116 struct amdgpu_usermode_queue *queue)
117 {
118 struct amdgpu_device *adev = uq_mgr->adev;
119 struct amdgpu_userq_obj *ctx = &queue->fw_obj;
120 struct amdgpu_mqd_prop *userq_props = queue->userq_prop;
121 struct mes_add_queue_input queue_input;
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 = ctx->gpu_addr;
135 queue_input.gang_context_addr = ctx->gpu_addr + AMDGPU_USERQ_PROC_CTX_SZ;
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 amdgpu_mes_lock(&adev->mes);
149 r = adev->mes.funcs->add_hw_queue(&adev->mes, &queue_input);
150 amdgpu_mes_unlock(&adev->mes);
151 if (r) {
152 DRM_ERROR("Failed to map queue in HW, err (%d)\n", r);
153 return r;
154 }
155
156 DRM_DEBUG_DRIVER("Queue (doorbell:%d) mapped successfully\n", userq_props->doorbell_index);
157 return 0;
158 }
159
mes_userq_unmap(struct amdgpu_userq_mgr * uq_mgr,struct amdgpu_usermode_queue * queue)160 static int mes_userq_unmap(struct amdgpu_userq_mgr *uq_mgr,
161 struct amdgpu_usermode_queue *queue)
162 {
163 struct amdgpu_device *adev = uq_mgr->adev;
164 struct mes_remove_queue_input queue_input;
165 struct amdgpu_userq_obj *ctx = &queue->fw_obj;
166 int r;
167
168 memset(&queue_input, 0x0, sizeof(struct mes_remove_queue_input));
169 queue_input.doorbell_offset = queue->doorbell_index;
170 queue_input.gang_context_addr = ctx->gpu_addr + AMDGPU_USERQ_PROC_CTX_SZ;
171
172 amdgpu_mes_lock(&adev->mes);
173 r = adev->mes.funcs->remove_hw_queue(&adev->mes, &queue_input);
174 amdgpu_mes_unlock(&adev->mes);
175 if (r)
176 DRM_ERROR("Failed to unmap queue in HW, err (%d)\n", r);
177 return r;
178 }
179
mes_userq_create_ctx_space(struct amdgpu_userq_mgr * uq_mgr,struct amdgpu_usermode_queue * queue,struct drm_amdgpu_userq_in * mqd_user)180 static int mes_userq_create_ctx_space(struct amdgpu_userq_mgr *uq_mgr,
181 struct amdgpu_usermode_queue *queue,
182 struct drm_amdgpu_userq_in *mqd_user)
183 {
184 struct amdgpu_userq_obj *ctx = &queue->fw_obj;
185 int r, size;
186
187 /*
188 * The FW expects at least one page space allocated for
189 * process ctx and gang ctx each. Create an object
190 * for the same.
191 */
192 size = AMDGPU_USERQ_PROC_CTX_SZ + AMDGPU_USERQ_GANG_CTX_SZ;
193 r = amdgpu_userq_create_object(uq_mgr, ctx, size);
194 if (r) {
195 DRM_ERROR("Failed to allocate ctx space bo for userqueue, err:%d\n", r);
196 return r;
197 }
198
199 return 0;
200 }
201
mes_userq_detect_and_reset(struct amdgpu_device * adev,int queue_type)202 static int mes_userq_detect_and_reset(struct amdgpu_device *adev,
203 int queue_type)
204 {
205 int db_array_size = amdgpu_mes_get_hung_queue_db_array_size(adev);
206 struct mes_detect_and_reset_queue_input input;
207 struct amdgpu_usermode_queue *queue;
208 unsigned int hung_db_num = 0;
209 unsigned long queue_id;
210 u32 db_array[8];
211 bool found_hung_queue = false;
212 int r, i;
213
214 if (db_array_size > 8) {
215 dev_err(adev->dev, "DB array size (%d vs 8) too small\n",
216 db_array_size);
217 return -EINVAL;
218 }
219
220 memset(&input, 0x0, sizeof(struct mes_detect_and_reset_queue_input));
221
222 input.queue_type = queue_type;
223
224 amdgpu_mes_lock(&adev->mes);
225 r = amdgpu_mes_detect_and_reset_hung_queues(adev, queue_type, false,
226 &hung_db_num, db_array);
227 amdgpu_mes_unlock(&adev->mes);
228 if (r) {
229 dev_err(adev->dev, "Failed to detect and reset queues, err (%d)\n", r);
230 } else if (hung_db_num) {
231 xa_for_each(&adev->userq_doorbell_xa, queue_id, queue) {
232 if (queue->queue_type == queue_type) {
233 for (i = 0; i < hung_db_num; i++) {
234 if (queue->doorbell_index == db_array[i]) {
235 queue->state = AMDGPU_USERQ_STATE_HUNG;
236 found_hung_queue = true;
237 atomic_inc(&adev->gpu_reset_counter);
238 amdgpu_userq_fence_driver_force_completion(queue);
239 drm_dev_wedged_event(adev_to_drm(adev), DRM_WEDGE_RECOVERY_NONE, NULL);
240 }
241 }
242 }
243 }
244 }
245
246 if (found_hung_queue) {
247 /* Resume scheduling after hang recovery */
248 r = amdgpu_mes_resume(adev);
249 }
250
251 return r;
252 }
253
mes_userq_mqd_create(struct amdgpu_userq_mgr * uq_mgr,struct drm_amdgpu_userq_in * args_in,struct amdgpu_usermode_queue * queue)254 static int mes_userq_mqd_create(struct amdgpu_userq_mgr *uq_mgr,
255 struct drm_amdgpu_userq_in *args_in,
256 struct amdgpu_usermode_queue *queue)
257 {
258 struct amdgpu_device *adev = uq_mgr->adev;
259 struct amdgpu_mqd *mqd_hw_default = &adev->mqds[queue->queue_type];
260 struct drm_amdgpu_userq_in *mqd_user = args_in;
261 struct amdgpu_mqd_prop *userq_props;
262 int r;
263
264 /* Structure to initialize MQD for userqueue using generic MQD init function */
265 userq_props = kzalloc(sizeof(struct amdgpu_mqd_prop), GFP_KERNEL);
266 if (!userq_props) {
267 DRM_ERROR("Failed to allocate memory for userq_props\n");
268 return -ENOMEM;
269 }
270
271 r = amdgpu_userq_create_object(uq_mgr, &queue->mqd, mqd_hw_default->mqd_size);
272 if (r) {
273 DRM_ERROR("Failed to create MQD object for userqueue\n");
274 goto free_props;
275 }
276
277 /* Initialize the MQD BO with user given values */
278 userq_props->wptr_gpu_addr = mqd_user->wptr_va;
279 userq_props->rptr_gpu_addr = mqd_user->rptr_va;
280 userq_props->queue_size = mqd_user->queue_size;
281 userq_props->hqd_base_gpu_addr = mqd_user->queue_va;
282 userq_props->mqd_gpu_addr = queue->mqd.gpu_addr;
283 userq_props->use_doorbell = true;
284 userq_props->doorbell_index = queue->doorbell_index;
285 userq_props->fence_address = queue->fence_drv->gpu_addr;
286
287 if (queue->queue_type == AMDGPU_HW_IP_COMPUTE) {
288 struct drm_amdgpu_userq_mqd_compute_gfx11 *compute_mqd;
289
290 if (mqd_user->mqd_size != sizeof(*compute_mqd)) {
291 DRM_ERROR("Invalid compute IP MQD size\n");
292 r = -EINVAL;
293 goto free_mqd;
294 }
295
296 compute_mqd = memdup_user(u64_to_user_ptr(mqd_user->mqd), mqd_user->mqd_size);
297 if (IS_ERR(compute_mqd)) {
298 DRM_ERROR("Failed to read user MQD\n");
299 r = -ENOMEM;
300 goto free_mqd;
301 }
302
303 r = amdgpu_userq_input_va_validate(queue, compute_mqd->eop_va,
304 2048);
305 if (r)
306 goto free_mqd;
307
308 userq_props->eop_gpu_addr = compute_mqd->eop_va;
309 userq_props->hqd_pipe_priority = AMDGPU_GFX_PIPE_PRIO_NORMAL;
310 userq_props->hqd_queue_priority = AMDGPU_GFX_QUEUE_PRIORITY_MINIMUM;
311 userq_props->hqd_active = false;
312 userq_props->tmz_queue =
313 mqd_user->flags & AMDGPU_USERQ_CREATE_FLAGS_QUEUE_SECURE;
314 kfree(compute_mqd);
315 } else if (queue->queue_type == AMDGPU_HW_IP_GFX) {
316 struct drm_amdgpu_userq_mqd_gfx11 *mqd_gfx_v11;
317 struct amdgpu_gfx_shadow_info shadow_info;
318
319 if (adev->gfx.funcs->get_gfx_shadow_info) {
320 adev->gfx.funcs->get_gfx_shadow_info(adev, &shadow_info, true);
321 } else {
322 r = -EINVAL;
323 goto free_mqd;
324 }
325
326 if (mqd_user->mqd_size != sizeof(*mqd_gfx_v11) || !mqd_user->mqd) {
327 DRM_ERROR("Invalid GFX MQD\n");
328 r = -EINVAL;
329 goto free_mqd;
330 }
331
332 mqd_gfx_v11 = memdup_user(u64_to_user_ptr(mqd_user->mqd), mqd_user->mqd_size);
333 if (IS_ERR(mqd_gfx_v11)) {
334 DRM_ERROR("Failed to read user MQD\n");
335 r = -ENOMEM;
336 goto free_mqd;
337 }
338
339 userq_props->shadow_addr = mqd_gfx_v11->shadow_va;
340 userq_props->csa_addr = mqd_gfx_v11->csa_va;
341 userq_props->tmz_queue =
342 mqd_user->flags & AMDGPU_USERQ_CREATE_FLAGS_QUEUE_SECURE;
343
344 r = amdgpu_userq_input_va_validate(queue, mqd_gfx_v11->shadow_va,
345 shadow_info.shadow_size);
346 if (r)
347 goto free_mqd;
348 r = amdgpu_userq_input_va_validate(queue, mqd_gfx_v11->csa_va,
349 shadow_info.csa_size);
350 if (r)
351 goto free_mqd;
352
353 kfree(mqd_gfx_v11);
354 } else if (queue->queue_type == AMDGPU_HW_IP_DMA) {
355 struct drm_amdgpu_userq_mqd_sdma_gfx11 *mqd_sdma_v11;
356
357 if (mqd_user->mqd_size != sizeof(*mqd_sdma_v11) || !mqd_user->mqd) {
358 DRM_ERROR("Invalid SDMA MQD\n");
359 r = -EINVAL;
360 goto free_mqd;
361 }
362
363 mqd_sdma_v11 = memdup_user(u64_to_user_ptr(mqd_user->mqd), mqd_user->mqd_size);
364 if (IS_ERR(mqd_sdma_v11)) {
365 DRM_ERROR("Failed to read sdma user MQD\n");
366 r = -ENOMEM;
367 goto free_mqd;
368 }
369 r = amdgpu_userq_input_va_validate(queue, mqd_sdma_v11->csa_va,
370 32);
371 if (r)
372 goto free_mqd;
373
374 userq_props->csa_addr = mqd_sdma_v11->csa_va;
375 kfree(mqd_sdma_v11);
376 }
377
378 queue->userq_prop = userq_props;
379
380 r = mqd_hw_default->init_mqd(adev, (void *)queue->mqd.cpu_ptr, userq_props);
381 if (r) {
382 DRM_ERROR("Failed to initialize MQD for userqueue\n");
383 goto free_mqd;
384 }
385
386 /* Create BO for FW operations */
387 r = mes_userq_create_ctx_space(uq_mgr, queue, mqd_user);
388 if (r) {
389 DRM_ERROR("Failed to allocate BO for userqueue (%d)", r);
390 goto free_mqd;
391 }
392
393 /* FW expects WPTR BOs to be mapped into GART */
394 r = mes_userq_create_wptr_mapping(uq_mgr, queue, userq_props->wptr_gpu_addr);
395 if (r) {
396 DRM_ERROR("Failed to create WPTR mapping\n");
397 goto free_ctx;
398 }
399
400 return 0;
401
402 free_ctx:
403 amdgpu_userq_destroy_object(uq_mgr, &queue->fw_obj);
404
405 free_mqd:
406 amdgpu_userq_destroy_object(uq_mgr, &queue->mqd);
407
408 free_props:
409 kfree(userq_props);
410
411 return r;
412 }
413
414 static void
mes_userq_mqd_destroy(struct amdgpu_userq_mgr * uq_mgr,struct amdgpu_usermode_queue * queue)415 mes_userq_mqd_destroy(struct amdgpu_userq_mgr *uq_mgr,
416 struct amdgpu_usermode_queue *queue)
417 {
418 amdgpu_userq_destroy_object(uq_mgr, &queue->fw_obj);
419 kfree(queue->userq_prop);
420 amdgpu_userq_destroy_object(uq_mgr, &queue->mqd);
421 }
422
mes_userq_preempt(struct amdgpu_userq_mgr * uq_mgr,struct amdgpu_usermode_queue * queue)423 static int mes_userq_preempt(struct amdgpu_userq_mgr *uq_mgr,
424 struct amdgpu_usermode_queue *queue)
425 {
426 struct amdgpu_device *adev = uq_mgr->adev;
427 struct mes_suspend_gang_input queue_input;
428 struct amdgpu_userq_obj *ctx = &queue->fw_obj;
429 signed long timeout = 2100000; /* 2100 ms */
430 u64 fence_gpu_addr;
431 u32 fence_offset;
432 u64 *fence_ptr;
433 int i, r;
434
435 if (queue->state != AMDGPU_USERQ_STATE_MAPPED)
436 return 0;
437 r = amdgpu_device_wb_get(adev, &fence_offset);
438 if (r)
439 return r;
440
441 fence_gpu_addr = adev->wb.gpu_addr + (fence_offset * 4);
442 fence_ptr = (u64 *)&adev->wb.wb[fence_offset];
443 *fence_ptr = 0;
444
445 memset(&queue_input, 0x0, sizeof(struct mes_suspend_gang_input));
446 queue_input.gang_context_addr = ctx->gpu_addr + AMDGPU_USERQ_PROC_CTX_SZ;
447 queue_input.suspend_fence_addr = fence_gpu_addr;
448 queue_input.suspend_fence_value = 1;
449 amdgpu_mes_lock(&adev->mes);
450 r = adev->mes.funcs->suspend_gang(&adev->mes, &queue_input);
451 amdgpu_mes_unlock(&adev->mes);
452 if (r) {
453 DRM_ERROR("Failed to suspend gang: %d\n", r);
454 goto out;
455 }
456
457 for (i = 0; i < timeout; i++) {
458 if (*fence_ptr == 1)
459 goto out;
460 udelay(1);
461 }
462 r = -ETIMEDOUT;
463
464 out:
465 amdgpu_device_wb_free(adev, fence_offset);
466 return r;
467 }
468
mes_userq_restore(struct amdgpu_userq_mgr * uq_mgr,struct amdgpu_usermode_queue * queue)469 static int mes_userq_restore(struct amdgpu_userq_mgr *uq_mgr,
470 struct amdgpu_usermode_queue *queue)
471 {
472 struct amdgpu_device *adev = uq_mgr->adev;
473 struct mes_resume_gang_input queue_input;
474 struct amdgpu_userq_obj *ctx = &queue->fw_obj;
475 int r;
476
477 if (queue->state == AMDGPU_USERQ_STATE_HUNG)
478 return -EINVAL;
479 if (queue->state != AMDGPU_USERQ_STATE_PREEMPTED)
480 return 0;
481
482 memset(&queue_input, 0x0, sizeof(struct mes_resume_gang_input));
483 queue_input.gang_context_addr = ctx->gpu_addr + AMDGPU_USERQ_PROC_CTX_SZ;
484
485 amdgpu_mes_lock(&adev->mes);
486 r = adev->mes.funcs->resume_gang(&adev->mes, &queue_input);
487 amdgpu_mes_unlock(&adev->mes);
488 if (r)
489 dev_err(adev->dev, "Failed to resume queue, err (%d)\n", r);
490 return r;
491 }
492
493 const struct amdgpu_userq_funcs userq_mes_funcs = {
494 .mqd_create = mes_userq_mqd_create,
495 .mqd_destroy = mes_userq_mqd_destroy,
496 .unmap = mes_userq_unmap,
497 .map = mes_userq_map,
498 .detect_and_reset = mes_userq_detect_and_reset,
499 .preempt = mes_userq_preempt,
500 .restore = mes_userq_restore,
501 };
502