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 "amdgpu_mes.h"
25 #include "amdgpu.h"
26 #include "soc15_common.h"
27 #include "amdgpu_mes_ctx.h"
28
29 #define AMDGPU_MES_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
30 #define AMDGPU_ONE_DOORBELL_SIZE 8
31
amdgpu_mes_doorbell_process_slice(struct amdgpu_device * adev)32 int amdgpu_mes_doorbell_process_slice(struct amdgpu_device *adev)
33 {
34 return roundup(AMDGPU_ONE_DOORBELL_SIZE *
35 AMDGPU_MES_MAX_NUM_OF_QUEUES_PER_PROCESS,
36 PAGE_SIZE);
37 }
38
amdgpu_mes_alloc_process_doorbells(struct amdgpu_device * adev,unsigned int * doorbell_index)39 int amdgpu_mes_alloc_process_doorbells(struct amdgpu_device *adev,
40 unsigned int *doorbell_index)
41 {
42 int r = ida_simple_get(&adev->mes.doorbell_ida, 2,
43 adev->mes.max_doorbell_slices,
44 GFP_KERNEL);
45 if (r > 0)
46 *doorbell_index = r;
47
48 return r;
49 }
50
amdgpu_mes_free_process_doorbells(struct amdgpu_device * adev,unsigned int doorbell_index)51 void amdgpu_mes_free_process_doorbells(struct amdgpu_device *adev,
52 unsigned int doorbell_index)
53 {
54 if (doorbell_index)
55 ida_simple_remove(&adev->mes.doorbell_ida, doorbell_index);
56 }
57
amdgpu_mes_get_doorbell_dw_offset_in_bar(struct amdgpu_device * adev,uint32_t doorbell_index,unsigned int doorbell_id)58 unsigned int amdgpu_mes_get_doorbell_dw_offset_in_bar(
59 struct amdgpu_device *adev,
60 uint32_t doorbell_index,
61 unsigned int doorbell_id)
62 {
63 return ((doorbell_index *
64 amdgpu_mes_doorbell_process_slice(adev)) / sizeof(u32) +
65 doorbell_id * 2);
66 }
67
amdgpu_mes_queue_doorbell_get(struct amdgpu_device * adev,struct amdgpu_mes_process * process,int ip_type,uint64_t * doorbell_index)68 static int amdgpu_mes_queue_doorbell_get(struct amdgpu_device *adev,
69 struct amdgpu_mes_process *process,
70 int ip_type, uint64_t *doorbell_index)
71 {
72 unsigned int offset, found;
73
74 if (ip_type == AMDGPU_RING_TYPE_SDMA) {
75 offset = adev->doorbell_index.sdma_engine[0];
76 found = find_next_zero_bit(process->doorbell_bitmap,
77 AMDGPU_MES_MAX_NUM_OF_QUEUES_PER_PROCESS,
78 offset);
79 } else {
80 found = find_first_zero_bit(process->doorbell_bitmap,
81 AMDGPU_MES_MAX_NUM_OF_QUEUES_PER_PROCESS);
82 }
83
84 if (found >= AMDGPU_MES_MAX_NUM_OF_QUEUES_PER_PROCESS) {
85 DRM_WARN("No doorbell available\n");
86 return -ENOSPC;
87 }
88
89 set_bit(found, process->doorbell_bitmap);
90
91 *doorbell_index = amdgpu_mes_get_doorbell_dw_offset_in_bar(adev,
92 process->doorbell_index, found);
93
94 return 0;
95 }
96
amdgpu_mes_queue_doorbell_free(struct amdgpu_device * adev,struct amdgpu_mes_process * process,uint32_t doorbell_index)97 static void amdgpu_mes_queue_doorbell_free(struct amdgpu_device *adev,
98 struct amdgpu_mes_process *process,
99 uint32_t doorbell_index)
100 {
101 unsigned int old, doorbell_id;
102
103 doorbell_id = doorbell_index -
104 (process->doorbell_index *
105 amdgpu_mes_doorbell_process_slice(adev)) / sizeof(u32);
106 doorbell_id /= 2;
107
108 old = test_and_clear_bit(doorbell_id, process->doorbell_bitmap);
109 WARN_ON(!old);
110 }
111
amdgpu_mes_doorbell_init(struct amdgpu_device * adev)112 static int amdgpu_mes_doorbell_init(struct amdgpu_device *adev)
113 {
114 size_t doorbell_start_offset;
115 size_t doorbell_aperture_size;
116 size_t doorbell_process_limit;
117 size_t aggregated_doorbell_start;
118 int i;
119
120 aggregated_doorbell_start = (adev->doorbell_index.max_assignment + 1) * sizeof(u32);
121 aggregated_doorbell_start =
122 roundup(aggregated_doorbell_start, PAGE_SIZE);
123
124 doorbell_start_offset = aggregated_doorbell_start + PAGE_SIZE;
125 doorbell_start_offset =
126 roundup(doorbell_start_offset,
127 amdgpu_mes_doorbell_process_slice(adev));
128
129 doorbell_aperture_size = adev->doorbell.size;
130 doorbell_aperture_size =
131 rounddown(doorbell_aperture_size,
132 amdgpu_mes_doorbell_process_slice(adev));
133
134 if (doorbell_aperture_size > doorbell_start_offset)
135 doorbell_process_limit =
136 (doorbell_aperture_size - doorbell_start_offset) /
137 amdgpu_mes_doorbell_process_slice(adev);
138 else
139 return -ENOSPC;
140
141 adev->mes.doorbell_id_offset = doorbell_start_offset / sizeof(u32);
142 adev->mes.max_doorbell_slices = doorbell_process_limit;
143
144 /* allocate Qword range for aggregated doorbell */
145 for (i = 0; i < AMDGPU_MES_PRIORITY_NUM_LEVELS; i++)
146 adev->mes.aggregated_doorbells[i] =
147 aggregated_doorbell_start / sizeof(u32) + i * 2;
148
149 DRM_INFO("max_doorbell_slices=%zu\n", doorbell_process_limit);
150 return 0;
151 }
152
amdgpu_mes_init(struct amdgpu_device * adev)153 int amdgpu_mes_init(struct amdgpu_device *adev)
154 {
155 int i, r;
156
157 adev->mes.adev = adev;
158
159 idr_init(&adev->mes.pasid_idr);
160 idr_init(&adev->mes.gang_id_idr);
161 idr_init(&adev->mes.queue_id_idr);
162 ida_init(&adev->mes.doorbell_ida);
163 spin_lock_init(&adev->mes.queue_id_lock);
164 spin_lock_init(&adev->mes.ring_lock);
165 mutex_init(&adev->mes.mutex_hidden);
166
167 adev->mes.total_max_queue = AMDGPU_FENCE_MES_QUEUE_ID_MASK;
168 adev->mes.vmid_mask_mmhub = 0xffffff00;
169 adev->mes.vmid_mask_gfxhub = 0xffffff00;
170
171 for (i = 0; i < AMDGPU_MES_MAX_COMPUTE_PIPES; i++) {
172 /* use only 1st MEC pipes */
173 if (i >= 4)
174 continue;
175 adev->mes.compute_hqd_mask[i] = 0xc;
176 }
177
178 for (i = 0; i < AMDGPU_MES_MAX_GFX_PIPES; i++)
179 adev->mes.gfx_hqd_mask[i] = i ? 0 : 0xfffffffe;
180
181 for (i = 0; i < AMDGPU_MES_MAX_SDMA_PIPES; i++) {
182 if (adev->ip_versions[SDMA0_HWIP][0] < IP_VERSION(6, 0, 0))
183 adev->mes.sdma_hqd_mask[i] = i ? 0 : 0x3fc;
184 /* zero sdma_hqd_mask for non-existent engine */
185 else if (adev->sdma.num_instances == 1)
186 adev->mes.sdma_hqd_mask[i] = i ? 0 : 0xfc;
187 else
188 adev->mes.sdma_hqd_mask[i] = 0xfc;
189 }
190
191 r = amdgpu_device_wb_get(adev, &adev->mes.sch_ctx_offs);
192 if (r) {
193 dev_err(adev->dev,
194 "(%d) ring trail_fence_offs wb alloc failed\n", r);
195 goto error_ids;
196 }
197 adev->mes.sch_ctx_gpu_addr =
198 adev->wb.gpu_addr + (adev->mes.sch_ctx_offs * 4);
199 adev->mes.sch_ctx_ptr =
200 (uint64_t *)&adev->wb.wb[adev->mes.sch_ctx_offs];
201
202 r = amdgpu_device_wb_get(adev, &adev->mes.query_status_fence_offs);
203 if (r) {
204 amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
205 dev_err(adev->dev,
206 "(%d) query_status_fence_offs wb alloc failed\n", r);
207 goto error_ids;
208 }
209 adev->mes.query_status_fence_gpu_addr =
210 adev->wb.gpu_addr + (adev->mes.query_status_fence_offs * 4);
211 adev->mes.query_status_fence_ptr =
212 (uint64_t *)&adev->wb.wb[adev->mes.query_status_fence_offs];
213
214 r = amdgpu_device_wb_get(adev, &adev->mes.read_val_offs);
215 if (r) {
216 amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
217 amdgpu_device_wb_free(adev, adev->mes.query_status_fence_offs);
218 dev_err(adev->dev,
219 "(%d) read_val_offs alloc failed\n", r);
220 goto error_ids;
221 }
222 adev->mes.read_val_gpu_addr =
223 adev->wb.gpu_addr + (adev->mes.read_val_offs * 4);
224 adev->mes.read_val_ptr =
225 (uint32_t *)&adev->wb.wb[adev->mes.read_val_offs];
226
227 r = amdgpu_mes_doorbell_init(adev);
228 if (r)
229 goto error;
230
231 return 0;
232
233 error:
234 amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
235 amdgpu_device_wb_free(adev, adev->mes.query_status_fence_offs);
236 amdgpu_device_wb_free(adev, adev->mes.read_val_offs);
237 error_ids:
238 idr_destroy(&adev->mes.pasid_idr);
239 idr_destroy(&adev->mes.gang_id_idr);
240 idr_destroy(&adev->mes.queue_id_idr);
241 ida_destroy(&adev->mes.doorbell_ida);
242 mutex_destroy(&adev->mes.mutex_hidden);
243 return r;
244 }
245
amdgpu_mes_fini(struct amdgpu_device * adev)246 void amdgpu_mes_fini(struct amdgpu_device *adev)
247 {
248 amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
249 amdgpu_device_wb_free(adev, adev->mes.query_status_fence_offs);
250 amdgpu_device_wb_free(adev, adev->mes.read_val_offs);
251
252 idr_destroy(&adev->mes.pasid_idr);
253 idr_destroy(&adev->mes.gang_id_idr);
254 idr_destroy(&adev->mes.queue_id_idr);
255 ida_destroy(&adev->mes.doorbell_ida);
256 mutex_destroy(&adev->mes.mutex_hidden);
257 }
258
amdgpu_mes_queue_free_mqd(struct amdgpu_mes_queue * q)259 static void amdgpu_mes_queue_free_mqd(struct amdgpu_mes_queue *q)
260 {
261 amdgpu_bo_free_kernel(&q->mqd_obj,
262 &q->mqd_gpu_addr,
263 &q->mqd_cpu_ptr);
264 }
265
amdgpu_mes_create_process(struct amdgpu_device * adev,int pasid,struct amdgpu_vm * vm)266 int amdgpu_mes_create_process(struct amdgpu_device *adev, int pasid,
267 struct amdgpu_vm *vm)
268 {
269 struct amdgpu_mes_process *process;
270 int r;
271
272 /* allocate the mes process buffer */
273 process = kzalloc(sizeof(struct amdgpu_mes_process), GFP_KERNEL);
274 if (!process) {
275 DRM_ERROR("no more memory to create mes process\n");
276 return -ENOMEM;
277 }
278
279 process->doorbell_bitmap =
280 kzalloc(DIV_ROUND_UP(AMDGPU_MES_MAX_NUM_OF_QUEUES_PER_PROCESS,
281 BITS_PER_BYTE), GFP_KERNEL);
282 if (!process->doorbell_bitmap) {
283 DRM_ERROR("failed to allocate doorbell bitmap\n");
284 kfree(process);
285 return -ENOMEM;
286 }
287
288 /* allocate the process context bo and map it */
289 r = amdgpu_bo_create_kernel(adev, AMDGPU_MES_PROC_CTX_SIZE, PAGE_SIZE,
290 AMDGPU_GEM_DOMAIN_GTT,
291 &process->proc_ctx_bo,
292 &process->proc_ctx_gpu_addr,
293 &process->proc_ctx_cpu_ptr);
294 if (r) {
295 DRM_ERROR("failed to allocate process context bo\n");
296 goto clean_up_memory;
297 }
298 memset(process->proc_ctx_cpu_ptr, 0, AMDGPU_MES_PROC_CTX_SIZE);
299
300 /*
301 * Avoid taking any other locks under MES lock to avoid circular
302 * lock dependencies.
303 */
304 amdgpu_mes_lock(&adev->mes);
305
306 /* add the mes process to idr list */
307 r = idr_alloc(&adev->mes.pasid_idr, process, pasid, pasid + 1,
308 GFP_KERNEL);
309 if (r < 0) {
310 DRM_ERROR("failed to lock pasid=%d\n", pasid);
311 goto clean_up_ctx;
312 }
313
314 /* allocate the starting doorbell index of the process */
315 r = amdgpu_mes_alloc_process_doorbells(adev, &process->doorbell_index);
316 if (r < 0) {
317 DRM_ERROR("failed to allocate doorbell for process\n");
318 goto clean_up_pasid;
319 }
320
321 DRM_DEBUG("process doorbell index = %d\n", process->doorbell_index);
322
323 INIT_LIST_HEAD(&process->gang_list);
324 process->vm = vm;
325 process->pasid = pasid;
326 process->process_quantum = adev->mes.default_process_quantum;
327 process->pd_gpu_addr = amdgpu_bo_gpu_offset(vm->root.bo);
328
329 amdgpu_mes_unlock(&adev->mes);
330 return 0;
331
332 clean_up_pasid:
333 idr_remove(&adev->mes.pasid_idr, pasid);
334 amdgpu_mes_unlock(&adev->mes);
335 clean_up_ctx:
336 amdgpu_bo_free_kernel(&process->proc_ctx_bo,
337 &process->proc_ctx_gpu_addr,
338 &process->proc_ctx_cpu_ptr);
339 clean_up_memory:
340 kfree(process->doorbell_bitmap);
341 kfree(process);
342 return r;
343 }
344
amdgpu_mes_destroy_process(struct amdgpu_device * adev,int pasid)345 void amdgpu_mes_destroy_process(struct amdgpu_device *adev, int pasid)
346 {
347 struct amdgpu_mes_process *process;
348 struct amdgpu_mes_gang *gang, *tmp1;
349 struct amdgpu_mes_queue *queue, *tmp2;
350 struct mes_remove_queue_input queue_input;
351 unsigned long flags;
352 int r;
353
354 /*
355 * Avoid taking any other locks under MES lock to avoid circular
356 * lock dependencies.
357 */
358 amdgpu_mes_lock(&adev->mes);
359
360 process = idr_find(&adev->mes.pasid_idr, pasid);
361 if (!process) {
362 DRM_WARN("pasid %d doesn't exist\n", pasid);
363 amdgpu_mes_unlock(&adev->mes);
364 return;
365 }
366
367 /* Remove all queues from hardware */
368 list_for_each_entry_safe(gang, tmp1, &process->gang_list, list) {
369 list_for_each_entry_safe(queue, tmp2, &gang->queue_list, list) {
370 spin_lock_irqsave(&adev->mes.queue_id_lock, flags);
371 idr_remove(&adev->mes.queue_id_idr, queue->queue_id);
372 spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
373
374 queue_input.doorbell_offset = queue->doorbell_off;
375 queue_input.gang_context_addr = gang->gang_ctx_gpu_addr;
376
377 r = adev->mes.funcs->remove_hw_queue(&adev->mes,
378 &queue_input);
379 if (r)
380 DRM_WARN("failed to remove hardware queue\n");
381 }
382
383 idr_remove(&adev->mes.gang_id_idr, gang->gang_id);
384 }
385
386 amdgpu_mes_free_process_doorbells(adev, process->doorbell_index);
387 idr_remove(&adev->mes.pasid_idr, pasid);
388 amdgpu_mes_unlock(&adev->mes);
389
390 /* free all memory allocated by the process */
391 list_for_each_entry_safe(gang, tmp1, &process->gang_list, list) {
392 /* free all queues in the gang */
393 list_for_each_entry_safe(queue, tmp2, &gang->queue_list, list) {
394 amdgpu_mes_queue_free_mqd(queue);
395 list_del(&queue->list);
396 kfree(queue);
397 }
398 amdgpu_bo_free_kernel(&gang->gang_ctx_bo,
399 &gang->gang_ctx_gpu_addr,
400 &gang->gang_ctx_cpu_ptr);
401 list_del(&gang->list);
402 kfree(gang);
403
404 }
405 amdgpu_bo_free_kernel(&process->proc_ctx_bo,
406 &process->proc_ctx_gpu_addr,
407 &process->proc_ctx_cpu_ptr);
408 kfree(process->doorbell_bitmap);
409 kfree(process);
410 }
411
amdgpu_mes_add_gang(struct amdgpu_device * adev,int pasid,struct amdgpu_mes_gang_properties * gprops,int * gang_id)412 int amdgpu_mes_add_gang(struct amdgpu_device *adev, int pasid,
413 struct amdgpu_mes_gang_properties *gprops,
414 int *gang_id)
415 {
416 struct amdgpu_mes_process *process;
417 struct amdgpu_mes_gang *gang;
418 int r;
419
420 /* allocate the mes gang buffer */
421 gang = kzalloc(sizeof(struct amdgpu_mes_gang), GFP_KERNEL);
422 if (!gang) {
423 return -ENOMEM;
424 }
425
426 /* allocate the gang context bo and map it to cpu space */
427 r = amdgpu_bo_create_kernel(adev, AMDGPU_MES_GANG_CTX_SIZE, PAGE_SIZE,
428 AMDGPU_GEM_DOMAIN_GTT,
429 &gang->gang_ctx_bo,
430 &gang->gang_ctx_gpu_addr,
431 &gang->gang_ctx_cpu_ptr);
432 if (r) {
433 DRM_ERROR("failed to allocate process context bo\n");
434 goto clean_up_mem;
435 }
436 memset(gang->gang_ctx_cpu_ptr, 0, AMDGPU_MES_GANG_CTX_SIZE);
437
438 /*
439 * Avoid taking any other locks under MES lock to avoid circular
440 * lock dependencies.
441 */
442 amdgpu_mes_lock(&adev->mes);
443
444 process = idr_find(&adev->mes.pasid_idr, pasid);
445 if (!process) {
446 DRM_ERROR("pasid %d doesn't exist\n", pasid);
447 r = -EINVAL;
448 goto clean_up_ctx;
449 }
450
451 /* add the mes gang to idr list */
452 r = idr_alloc(&adev->mes.gang_id_idr, gang, 1, 0,
453 GFP_KERNEL);
454 if (r < 0) {
455 DRM_ERROR("failed to allocate idr for gang\n");
456 goto clean_up_ctx;
457 }
458
459 gang->gang_id = r;
460 *gang_id = r;
461
462 INIT_LIST_HEAD(&gang->queue_list);
463 gang->process = process;
464 gang->priority = gprops->priority;
465 gang->gang_quantum = gprops->gang_quantum ?
466 gprops->gang_quantum : adev->mes.default_gang_quantum;
467 gang->global_priority_level = gprops->global_priority_level;
468 gang->inprocess_gang_priority = gprops->inprocess_gang_priority;
469 list_add_tail(&gang->list, &process->gang_list);
470
471 amdgpu_mes_unlock(&adev->mes);
472 return 0;
473
474 clean_up_ctx:
475 amdgpu_mes_unlock(&adev->mes);
476 amdgpu_bo_free_kernel(&gang->gang_ctx_bo,
477 &gang->gang_ctx_gpu_addr,
478 &gang->gang_ctx_cpu_ptr);
479 clean_up_mem:
480 kfree(gang);
481 return r;
482 }
483
amdgpu_mes_remove_gang(struct amdgpu_device * adev,int gang_id)484 int amdgpu_mes_remove_gang(struct amdgpu_device *adev, int gang_id)
485 {
486 struct amdgpu_mes_gang *gang;
487
488 /*
489 * Avoid taking any other locks under MES lock to avoid circular
490 * lock dependencies.
491 */
492 amdgpu_mes_lock(&adev->mes);
493
494 gang = idr_find(&adev->mes.gang_id_idr, gang_id);
495 if (!gang) {
496 DRM_ERROR("gang id %d doesn't exist\n", gang_id);
497 amdgpu_mes_unlock(&adev->mes);
498 return -EINVAL;
499 }
500
501 if (!list_empty(&gang->queue_list)) {
502 DRM_ERROR("queue list is not empty\n");
503 amdgpu_mes_unlock(&adev->mes);
504 return -EBUSY;
505 }
506
507 idr_remove(&adev->mes.gang_id_idr, gang->gang_id);
508 list_del(&gang->list);
509 amdgpu_mes_unlock(&adev->mes);
510
511 amdgpu_bo_free_kernel(&gang->gang_ctx_bo,
512 &gang->gang_ctx_gpu_addr,
513 &gang->gang_ctx_cpu_ptr);
514
515 kfree(gang);
516
517 return 0;
518 }
519
amdgpu_mes_suspend(struct amdgpu_device * adev)520 int amdgpu_mes_suspend(struct amdgpu_device *adev)
521 {
522 struct idr *idp;
523 struct amdgpu_mes_process *process;
524 struct amdgpu_mes_gang *gang;
525 struct mes_suspend_gang_input input;
526 int r, pasid;
527
528 /*
529 * Avoid taking any other locks under MES lock to avoid circular
530 * lock dependencies.
531 */
532 amdgpu_mes_lock(&adev->mes);
533
534 idp = &adev->mes.pasid_idr;
535
536 idr_for_each_entry(idp, process, pasid) {
537 list_for_each_entry(gang, &process->gang_list, list) {
538 r = adev->mes.funcs->suspend_gang(&adev->mes, &input);
539 if (r)
540 DRM_ERROR("failed to suspend pasid %d gangid %d",
541 pasid, gang->gang_id);
542 }
543 }
544
545 amdgpu_mes_unlock(&adev->mes);
546 return 0;
547 }
548
amdgpu_mes_resume(struct amdgpu_device * adev)549 int amdgpu_mes_resume(struct amdgpu_device *adev)
550 {
551 struct idr *idp;
552 struct amdgpu_mes_process *process;
553 struct amdgpu_mes_gang *gang;
554 struct mes_resume_gang_input input;
555 int r, pasid;
556
557 /*
558 * Avoid taking any other locks under MES lock to avoid circular
559 * lock dependencies.
560 */
561 amdgpu_mes_lock(&adev->mes);
562
563 idp = &adev->mes.pasid_idr;
564
565 idr_for_each_entry(idp, process, pasid) {
566 list_for_each_entry(gang, &process->gang_list, list) {
567 r = adev->mes.funcs->resume_gang(&adev->mes, &input);
568 if (r)
569 DRM_ERROR("failed to resume pasid %d gangid %d",
570 pasid, gang->gang_id);
571 }
572 }
573
574 amdgpu_mes_unlock(&adev->mes);
575 return 0;
576 }
577
amdgpu_mes_queue_alloc_mqd(struct amdgpu_device * adev,struct amdgpu_mes_queue * q,struct amdgpu_mes_queue_properties * p)578 static int amdgpu_mes_queue_alloc_mqd(struct amdgpu_device *adev,
579 struct amdgpu_mes_queue *q,
580 struct amdgpu_mes_queue_properties *p)
581 {
582 struct amdgpu_mqd *mqd_mgr = &adev->mqds[p->queue_type];
583 u32 mqd_size = mqd_mgr->mqd_size;
584 int r;
585
586 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
587 AMDGPU_GEM_DOMAIN_GTT,
588 &q->mqd_obj,
589 &q->mqd_gpu_addr, &q->mqd_cpu_ptr);
590 if (r) {
591 dev_warn(adev->dev, "failed to create queue mqd bo (%d)", r);
592 return r;
593 }
594 memset(q->mqd_cpu_ptr, 0, mqd_size);
595
596 r = amdgpu_bo_reserve(q->mqd_obj, false);
597 if (unlikely(r != 0))
598 goto clean_up;
599
600 return 0;
601
602 clean_up:
603 amdgpu_bo_free_kernel(&q->mqd_obj,
604 &q->mqd_gpu_addr,
605 &q->mqd_cpu_ptr);
606 return r;
607 }
608
amdgpu_mes_queue_init_mqd(struct amdgpu_device * adev,struct amdgpu_mes_queue * q,struct amdgpu_mes_queue_properties * p)609 static void amdgpu_mes_queue_init_mqd(struct amdgpu_device *adev,
610 struct amdgpu_mes_queue *q,
611 struct amdgpu_mes_queue_properties *p)
612 {
613 struct amdgpu_mqd *mqd_mgr = &adev->mqds[p->queue_type];
614 struct amdgpu_mqd_prop mqd_prop = {0};
615
616 mqd_prop.mqd_gpu_addr = q->mqd_gpu_addr;
617 mqd_prop.hqd_base_gpu_addr = p->hqd_base_gpu_addr;
618 mqd_prop.rptr_gpu_addr = p->rptr_gpu_addr;
619 mqd_prop.wptr_gpu_addr = p->wptr_gpu_addr;
620 mqd_prop.queue_size = p->queue_size;
621 mqd_prop.use_doorbell = true;
622 mqd_prop.doorbell_index = p->doorbell_off;
623 mqd_prop.eop_gpu_addr = p->eop_gpu_addr;
624 mqd_prop.hqd_pipe_priority = p->hqd_pipe_priority;
625 mqd_prop.hqd_queue_priority = p->hqd_queue_priority;
626 mqd_prop.hqd_active = false;
627
628 mqd_mgr->init_mqd(adev, q->mqd_cpu_ptr, &mqd_prop);
629
630 amdgpu_bo_unreserve(q->mqd_obj);
631 }
632
amdgpu_mes_add_hw_queue(struct amdgpu_device * adev,int gang_id,struct amdgpu_mes_queue_properties * qprops,int * queue_id)633 int amdgpu_mes_add_hw_queue(struct amdgpu_device *adev, int gang_id,
634 struct amdgpu_mes_queue_properties *qprops,
635 int *queue_id)
636 {
637 struct amdgpu_mes_queue *queue;
638 struct amdgpu_mes_gang *gang;
639 struct mes_add_queue_input queue_input;
640 unsigned long flags;
641 int r;
642
643 /* allocate the mes queue buffer */
644 queue = kzalloc(sizeof(struct amdgpu_mes_queue), GFP_KERNEL);
645 if (!queue) {
646 DRM_ERROR("Failed to allocate memory for queue\n");
647 return -ENOMEM;
648 }
649
650 /* Allocate the queue mqd */
651 r = amdgpu_mes_queue_alloc_mqd(adev, queue, qprops);
652 if (r)
653 goto clean_up_memory;
654
655 /*
656 * Avoid taking any other locks under MES lock to avoid circular
657 * lock dependencies.
658 */
659 amdgpu_mes_lock(&adev->mes);
660
661 gang = idr_find(&adev->mes.gang_id_idr, gang_id);
662 if (!gang) {
663 DRM_ERROR("gang id %d doesn't exist\n", gang_id);
664 r = -EINVAL;
665 goto clean_up_mqd;
666 }
667
668 /* add the mes gang to idr list */
669 spin_lock_irqsave(&adev->mes.queue_id_lock, flags);
670 r = idr_alloc(&adev->mes.queue_id_idr, queue, 1, 0,
671 GFP_ATOMIC);
672 if (r < 0) {
673 spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
674 goto clean_up_mqd;
675 }
676 spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
677 *queue_id = queue->queue_id = r;
678
679 /* allocate a doorbell index for the queue */
680 r = amdgpu_mes_queue_doorbell_get(adev, gang->process,
681 qprops->queue_type,
682 &qprops->doorbell_off);
683 if (r)
684 goto clean_up_queue_id;
685
686 /* initialize the queue mqd */
687 amdgpu_mes_queue_init_mqd(adev, queue, qprops);
688
689 /* add hw queue to mes */
690 queue_input.process_id = gang->process->pasid;
691
692 queue_input.page_table_base_addr =
693 adev->vm_manager.vram_base_offset + gang->process->pd_gpu_addr -
694 adev->gmc.vram_start;
695
696 queue_input.process_va_start = 0;
697 queue_input.process_va_end =
698 (adev->vm_manager.max_pfn - 1) << AMDGPU_GPU_PAGE_SHIFT;
699 queue_input.process_quantum = gang->process->process_quantum;
700 queue_input.process_context_addr = gang->process->proc_ctx_gpu_addr;
701 queue_input.gang_quantum = gang->gang_quantum;
702 queue_input.gang_context_addr = gang->gang_ctx_gpu_addr;
703 queue_input.inprocess_gang_priority = gang->inprocess_gang_priority;
704 queue_input.gang_global_priority_level = gang->global_priority_level;
705 queue_input.doorbell_offset = qprops->doorbell_off;
706 queue_input.mqd_addr = queue->mqd_gpu_addr;
707 queue_input.wptr_addr = qprops->wptr_gpu_addr;
708 queue_input.wptr_mc_addr = qprops->wptr_mc_addr;
709 queue_input.queue_type = qprops->queue_type;
710 queue_input.paging = qprops->paging;
711 queue_input.is_kfd_process = 0;
712
713 r = adev->mes.funcs->add_hw_queue(&adev->mes, &queue_input);
714 if (r) {
715 DRM_ERROR("failed to add hardware queue to MES, doorbell=0x%llx\n",
716 qprops->doorbell_off);
717 goto clean_up_doorbell;
718 }
719
720 DRM_DEBUG("MES hw queue was added, pasid=%d, gang id=%d, "
721 "queue type=%d, doorbell=0x%llx\n",
722 gang->process->pasid, gang_id, qprops->queue_type,
723 qprops->doorbell_off);
724
725 queue->ring = qprops->ring;
726 queue->doorbell_off = qprops->doorbell_off;
727 queue->wptr_gpu_addr = qprops->wptr_gpu_addr;
728 queue->queue_type = qprops->queue_type;
729 queue->paging = qprops->paging;
730 queue->gang = gang;
731 queue->ring->mqd_ptr = queue->mqd_cpu_ptr;
732 list_add_tail(&queue->list, &gang->queue_list);
733
734 amdgpu_mes_unlock(&adev->mes);
735 return 0;
736
737 clean_up_doorbell:
738 amdgpu_mes_queue_doorbell_free(adev, gang->process,
739 qprops->doorbell_off);
740 clean_up_queue_id:
741 spin_lock_irqsave(&adev->mes.queue_id_lock, flags);
742 idr_remove(&adev->mes.queue_id_idr, queue->queue_id);
743 spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
744 clean_up_mqd:
745 amdgpu_mes_unlock(&adev->mes);
746 amdgpu_mes_queue_free_mqd(queue);
747 clean_up_memory:
748 kfree(queue);
749 return r;
750 }
751
amdgpu_mes_remove_hw_queue(struct amdgpu_device * adev,int queue_id)752 int amdgpu_mes_remove_hw_queue(struct amdgpu_device *adev, int queue_id)
753 {
754 unsigned long flags;
755 struct amdgpu_mes_queue *queue;
756 struct amdgpu_mes_gang *gang;
757 struct mes_remove_queue_input queue_input;
758 int r;
759
760 /*
761 * Avoid taking any other locks under MES lock to avoid circular
762 * lock dependencies.
763 */
764 amdgpu_mes_lock(&adev->mes);
765
766 /* remove the mes gang from idr list */
767 spin_lock_irqsave(&adev->mes.queue_id_lock, flags);
768
769 queue = idr_find(&adev->mes.queue_id_idr, queue_id);
770 if (!queue) {
771 spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
772 amdgpu_mes_unlock(&adev->mes);
773 DRM_ERROR("queue id %d doesn't exist\n", queue_id);
774 return -EINVAL;
775 }
776
777 idr_remove(&adev->mes.queue_id_idr, queue_id);
778 spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
779
780 DRM_DEBUG("try to remove queue, doorbell off = 0x%llx\n",
781 queue->doorbell_off);
782
783 gang = queue->gang;
784 queue_input.doorbell_offset = queue->doorbell_off;
785 queue_input.gang_context_addr = gang->gang_ctx_gpu_addr;
786
787 r = adev->mes.funcs->remove_hw_queue(&adev->mes, &queue_input);
788 if (r)
789 DRM_ERROR("failed to remove hardware queue, queue id = %d\n",
790 queue_id);
791
792 list_del(&queue->list);
793 amdgpu_mes_queue_doorbell_free(adev, gang->process,
794 queue->doorbell_off);
795 amdgpu_mes_unlock(&adev->mes);
796
797 amdgpu_mes_queue_free_mqd(queue);
798 kfree(queue);
799 return 0;
800 }
801
amdgpu_mes_unmap_legacy_queue(struct amdgpu_device * adev,struct amdgpu_ring * ring,enum amdgpu_unmap_queues_action action,u64 gpu_addr,u64 seq)802 int amdgpu_mes_unmap_legacy_queue(struct amdgpu_device *adev,
803 struct amdgpu_ring *ring,
804 enum amdgpu_unmap_queues_action action,
805 u64 gpu_addr, u64 seq)
806 {
807 struct mes_unmap_legacy_queue_input queue_input;
808 int r;
809
810 queue_input.action = action;
811 queue_input.queue_type = ring->funcs->type;
812 queue_input.doorbell_offset = ring->doorbell_index;
813 queue_input.pipe_id = ring->pipe;
814 queue_input.queue_id = ring->queue;
815 queue_input.trail_fence_addr = gpu_addr;
816 queue_input.trail_fence_data = seq;
817
818 r = adev->mes.funcs->unmap_legacy_queue(&adev->mes, &queue_input);
819 if (r)
820 DRM_ERROR("failed to unmap legacy queue\n");
821
822 return r;
823 }
824
amdgpu_mes_rreg(struct amdgpu_device * adev,uint32_t reg)825 uint32_t amdgpu_mes_rreg(struct amdgpu_device *adev, uint32_t reg)
826 {
827 struct mes_misc_op_input op_input;
828 int r, val = 0;
829
830 op_input.op = MES_MISC_OP_READ_REG;
831 op_input.read_reg.reg_offset = reg;
832 op_input.read_reg.buffer_addr = adev->mes.read_val_gpu_addr;
833
834 if (!adev->mes.funcs->misc_op) {
835 DRM_ERROR("mes rreg is not supported!\n");
836 goto error;
837 }
838
839 r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
840 if (r)
841 DRM_ERROR("failed to read reg (0x%x)\n", reg);
842 else
843 val = *(adev->mes.read_val_ptr);
844
845 error:
846 return val;
847 }
848
amdgpu_mes_wreg(struct amdgpu_device * adev,uint32_t reg,uint32_t val)849 int amdgpu_mes_wreg(struct amdgpu_device *adev,
850 uint32_t reg, uint32_t val)
851 {
852 struct mes_misc_op_input op_input;
853 int r;
854
855 op_input.op = MES_MISC_OP_WRITE_REG;
856 op_input.write_reg.reg_offset = reg;
857 op_input.write_reg.reg_value = val;
858
859 if (!adev->mes.funcs->misc_op) {
860 DRM_ERROR("mes wreg is not supported!\n");
861 r = -EINVAL;
862 goto error;
863 }
864
865 r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
866 if (r)
867 DRM_ERROR("failed to write reg (0x%x)\n", reg);
868
869 error:
870 return r;
871 }
872
amdgpu_mes_reg_write_reg_wait(struct amdgpu_device * adev,uint32_t reg0,uint32_t reg1,uint32_t ref,uint32_t mask)873 int amdgpu_mes_reg_write_reg_wait(struct amdgpu_device *adev,
874 uint32_t reg0, uint32_t reg1,
875 uint32_t ref, uint32_t mask)
876 {
877 struct mes_misc_op_input op_input;
878 int r;
879
880 op_input.op = MES_MISC_OP_WRM_REG_WR_WAIT;
881 op_input.wrm_reg.reg0 = reg0;
882 op_input.wrm_reg.reg1 = reg1;
883 op_input.wrm_reg.ref = ref;
884 op_input.wrm_reg.mask = mask;
885
886 if (!adev->mes.funcs->misc_op) {
887 DRM_ERROR("mes reg_write_reg_wait is not supported!\n");
888 r = -EINVAL;
889 goto error;
890 }
891
892 r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
893 if (r)
894 DRM_ERROR("failed to reg_write_reg_wait\n");
895
896 error:
897 return r;
898 }
899
amdgpu_mes_reg_wait(struct amdgpu_device * adev,uint32_t reg,uint32_t val,uint32_t mask)900 int amdgpu_mes_reg_wait(struct amdgpu_device *adev, uint32_t reg,
901 uint32_t val, uint32_t mask)
902 {
903 struct mes_misc_op_input op_input;
904 int r;
905
906 op_input.op = MES_MISC_OP_WRM_REG_WAIT;
907 op_input.wrm_reg.reg0 = reg;
908 op_input.wrm_reg.ref = val;
909 op_input.wrm_reg.mask = mask;
910
911 if (!adev->mes.funcs->misc_op) {
912 DRM_ERROR("mes reg wait is not supported!\n");
913 r = -EINVAL;
914 goto error;
915 }
916
917 r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
918 if (r)
919 DRM_ERROR("failed to reg_write_reg_wait\n");
920
921 error:
922 return r;
923 }
924
925 static void
amdgpu_mes_ring_to_queue_props(struct amdgpu_device * adev,struct amdgpu_ring * ring,struct amdgpu_mes_queue_properties * props)926 amdgpu_mes_ring_to_queue_props(struct amdgpu_device *adev,
927 struct amdgpu_ring *ring,
928 struct amdgpu_mes_queue_properties *props)
929 {
930 props->queue_type = ring->funcs->type;
931 props->hqd_base_gpu_addr = ring->gpu_addr;
932 props->rptr_gpu_addr = ring->rptr_gpu_addr;
933 props->wptr_gpu_addr = ring->wptr_gpu_addr;
934 props->wptr_mc_addr =
935 ring->mes_ctx->meta_data_mc_addr + ring->wptr_offs;
936 props->queue_size = ring->ring_size;
937 props->eop_gpu_addr = ring->eop_gpu_addr;
938 props->hqd_pipe_priority = AMDGPU_GFX_PIPE_PRIO_NORMAL;
939 props->hqd_queue_priority = AMDGPU_GFX_QUEUE_PRIORITY_MINIMUM;
940 props->paging = false;
941 props->ring = ring;
942 }
943
944 #define DEFINE_AMDGPU_MES_CTX_GET_OFFS_ENG(_eng) \
945 do { \
946 if (id_offs < AMDGPU_MES_CTX_MAX_OFFS) \
947 return offsetof(struct amdgpu_mes_ctx_meta_data, \
948 _eng[ring->idx].slots[id_offs]); \
949 else if (id_offs == AMDGPU_MES_CTX_RING_OFFS) \
950 return offsetof(struct amdgpu_mes_ctx_meta_data, \
951 _eng[ring->idx].ring); \
952 else if (id_offs == AMDGPU_MES_CTX_IB_OFFS) \
953 return offsetof(struct amdgpu_mes_ctx_meta_data, \
954 _eng[ring->idx].ib); \
955 else if (id_offs == AMDGPU_MES_CTX_PADDING_OFFS) \
956 return offsetof(struct amdgpu_mes_ctx_meta_data, \
957 _eng[ring->idx].padding); \
958 } while(0)
959
amdgpu_mes_ctx_get_offs(struct amdgpu_ring * ring,unsigned int id_offs)960 int amdgpu_mes_ctx_get_offs(struct amdgpu_ring *ring, unsigned int id_offs)
961 {
962 switch (ring->funcs->type) {
963 case AMDGPU_RING_TYPE_GFX:
964 DEFINE_AMDGPU_MES_CTX_GET_OFFS_ENG(gfx);
965 break;
966 case AMDGPU_RING_TYPE_COMPUTE:
967 DEFINE_AMDGPU_MES_CTX_GET_OFFS_ENG(compute);
968 break;
969 case AMDGPU_RING_TYPE_SDMA:
970 DEFINE_AMDGPU_MES_CTX_GET_OFFS_ENG(sdma);
971 break;
972 default:
973 break;
974 }
975
976 WARN_ON(1);
977 return -EINVAL;
978 }
979
amdgpu_mes_add_ring(struct amdgpu_device * adev,int gang_id,int queue_type,int idx,struct amdgpu_mes_ctx_data * ctx_data,struct amdgpu_ring ** out)980 int amdgpu_mes_add_ring(struct amdgpu_device *adev, int gang_id,
981 int queue_type, int idx,
982 struct amdgpu_mes_ctx_data *ctx_data,
983 struct amdgpu_ring **out)
984 {
985 struct amdgpu_ring *ring;
986 struct amdgpu_mes_gang *gang;
987 struct amdgpu_mes_queue_properties qprops = {0};
988 int r, queue_id, pasid;
989
990 /*
991 * Avoid taking any other locks under MES lock to avoid circular
992 * lock dependencies.
993 */
994 amdgpu_mes_lock(&adev->mes);
995 gang = idr_find(&adev->mes.gang_id_idr, gang_id);
996 if (!gang) {
997 DRM_ERROR("gang id %d doesn't exist\n", gang_id);
998 amdgpu_mes_unlock(&adev->mes);
999 return -EINVAL;
1000 }
1001 pasid = gang->process->pasid;
1002
1003 ring = kzalloc(sizeof(struct amdgpu_ring), GFP_KERNEL);
1004 if (!ring) {
1005 amdgpu_mes_unlock(&adev->mes);
1006 return -ENOMEM;
1007 }
1008
1009 ring->ring_obj = NULL;
1010 ring->use_doorbell = true;
1011 ring->is_mes_queue = true;
1012 ring->mes_ctx = ctx_data;
1013 ring->idx = idx;
1014 ring->no_scheduler = true;
1015
1016 if (queue_type == AMDGPU_RING_TYPE_COMPUTE) {
1017 int offset = offsetof(struct amdgpu_mes_ctx_meta_data,
1018 compute[ring->idx].mec_hpd);
1019 ring->eop_gpu_addr =
1020 amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
1021 }
1022
1023 switch (queue_type) {
1024 case AMDGPU_RING_TYPE_GFX:
1025 ring->funcs = adev->gfx.gfx_ring[0].funcs;
1026 break;
1027 case AMDGPU_RING_TYPE_COMPUTE:
1028 ring->funcs = adev->gfx.compute_ring[0].funcs;
1029 break;
1030 case AMDGPU_RING_TYPE_SDMA:
1031 ring->funcs = adev->sdma.instance[0].ring.funcs;
1032 break;
1033 default:
1034 BUG();
1035 }
1036
1037 r = amdgpu_ring_init(adev, ring, 1024, NULL, 0,
1038 AMDGPU_RING_PRIO_DEFAULT, NULL);
1039 if (r)
1040 goto clean_up_memory;
1041
1042 amdgpu_mes_ring_to_queue_props(adev, ring, &qprops);
1043
1044 dma_fence_wait(gang->process->vm->last_update, false);
1045 dma_fence_wait(ctx_data->meta_data_va->last_pt_update, false);
1046 amdgpu_mes_unlock(&adev->mes);
1047
1048 r = amdgpu_mes_add_hw_queue(adev, gang_id, &qprops, &queue_id);
1049 if (r)
1050 goto clean_up_ring;
1051
1052 ring->hw_queue_id = queue_id;
1053 ring->doorbell_index = qprops.doorbell_off;
1054
1055 if (queue_type == AMDGPU_RING_TYPE_GFX)
1056 sprintf(ring->name, "gfx_%d.%d.%d", pasid, gang_id, queue_id);
1057 else if (queue_type == AMDGPU_RING_TYPE_COMPUTE)
1058 sprintf(ring->name, "compute_%d.%d.%d", pasid, gang_id,
1059 queue_id);
1060 else if (queue_type == AMDGPU_RING_TYPE_SDMA)
1061 sprintf(ring->name, "sdma_%d.%d.%d", pasid, gang_id,
1062 queue_id);
1063 else
1064 BUG();
1065
1066 *out = ring;
1067 return 0;
1068
1069 clean_up_ring:
1070 amdgpu_ring_fini(ring);
1071 clean_up_memory:
1072 kfree(ring);
1073 amdgpu_mes_unlock(&adev->mes);
1074 return r;
1075 }
1076
amdgpu_mes_remove_ring(struct amdgpu_device * adev,struct amdgpu_ring * ring)1077 void amdgpu_mes_remove_ring(struct amdgpu_device *adev,
1078 struct amdgpu_ring *ring)
1079 {
1080 if (!ring)
1081 return;
1082
1083 amdgpu_mes_remove_hw_queue(adev, ring->hw_queue_id);
1084 amdgpu_ring_fini(ring);
1085 kfree(ring);
1086 }
1087
amdgpu_mes_get_aggregated_doorbell_index(struct amdgpu_device * adev,enum amdgpu_mes_priority_level prio)1088 uint32_t amdgpu_mes_get_aggregated_doorbell_index(struct amdgpu_device *adev,
1089 enum amdgpu_mes_priority_level prio)
1090 {
1091 return adev->mes.aggregated_doorbells[prio];
1092 }
1093
amdgpu_mes_ctx_alloc_meta_data(struct amdgpu_device * adev,struct amdgpu_mes_ctx_data * ctx_data)1094 int amdgpu_mes_ctx_alloc_meta_data(struct amdgpu_device *adev,
1095 struct amdgpu_mes_ctx_data *ctx_data)
1096 {
1097 int r;
1098
1099 r = amdgpu_bo_create_kernel(adev,
1100 sizeof(struct amdgpu_mes_ctx_meta_data),
1101 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
1102 &ctx_data->meta_data_obj,
1103 &ctx_data->meta_data_mc_addr,
1104 &ctx_data->meta_data_ptr);
1105 if (!ctx_data->meta_data_obj)
1106 return -ENOMEM;
1107
1108 memset(ctx_data->meta_data_ptr, 0,
1109 sizeof(struct amdgpu_mes_ctx_meta_data));
1110
1111 return 0;
1112 }
1113
amdgpu_mes_ctx_free_meta_data(struct amdgpu_mes_ctx_data * ctx_data)1114 void amdgpu_mes_ctx_free_meta_data(struct amdgpu_mes_ctx_data *ctx_data)
1115 {
1116 if (ctx_data->meta_data_obj)
1117 amdgpu_bo_free_kernel(&ctx_data->meta_data_obj,
1118 &ctx_data->meta_data_mc_addr,
1119 &ctx_data->meta_data_ptr);
1120 }
1121
amdgpu_mes_ctx_map_meta_data(struct amdgpu_device * adev,struct amdgpu_vm * vm,struct amdgpu_mes_ctx_data * ctx_data)1122 int amdgpu_mes_ctx_map_meta_data(struct amdgpu_device *adev,
1123 struct amdgpu_vm *vm,
1124 struct amdgpu_mes_ctx_data *ctx_data)
1125 {
1126 struct amdgpu_bo_va *bo_va;
1127 struct ww_acquire_ctx ticket;
1128 struct list_head list;
1129 struct amdgpu_bo_list_entry pd;
1130 struct ttm_validate_buffer csa_tv;
1131 struct amdgpu_sync sync;
1132 int r;
1133
1134 amdgpu_sync_create(&sync);
1135 INIT_LIST_HEAD(&list);
1136 INIT_LIST_HEAD(&csa_tv.head);
1137
1138 csa_tv.bo = &ctx_data->meta_data_obj->tbo;
1139 csa_tv.num_shared = 1;
1140
1141 list_add(&csa_tv.head, &list);
1142 amdgpu_vm_get_pd_bo(vm, &list, &pd);
1143
1144 r = ttm_eu_reserve_buffers(&ticket, &list, true, NULL);
1145 if (r) {
1146 DRM_ERROR("failed to reserve meta data BO: err=%d\n", r);
1147 return r;
1148 }
1149
1150 bo_va = amdgpu_vm_bo_add(adev, vm, ctx_data->meta_data_obj);
1151 if (!bo_va) {
1152 ttm_eu_backoff_reservation(&ticket, &list);
1153 DRM_ERROR("failed to create bo_va for meta data BO\n");
1154 return -ENOMEM;
1155 }
1156
1157 r = amdgpu_vm_bo_map(adev, bo_va, ctx_data->meta_data_gpu_addr, 0,
1158 sizeof(struct amdgpu_mes_ctx_meta_data),
1159 AMDGPU_PTE_READABLE | AMDGPU_PTE_WRITEABLE |
1160 AMDGPU_PTE_EXECUTABLE);
1161
1162 if (r) {
1163 DRM_ERROR("failed to do bo_map on meta data, err=%d\n", r);
1164 goto error;
1165 }
1166
1167 r = amdgpu_vm_bo_update(adev, bo_va, false);
1168 if (r) {
1169 DRM_ERROR("failed to do vm_bo_update on meta data\n");
1170 goto error;
1171 }
1172 amdgpu_sync_fence(&sync, bo_va->last_pt_update);
1173
1174 r = amdgpu_vm_update_pdes(adev, vm, false);
1175 if (r) {
1176 DRM_ERROR("failed to update pdes on meta data\n");
1177 goto error;
1178 }
1179 amdgpu_sync_fence(&sync, vm->last_update);
1180
1181 amdgpu_sync_wait(&sync, false);
1182 ttm_eu_backoff_reservation(&ticket, &list);
1183
1184 amdgpu_sync_free(&sync);
1185 ctx_data->meta_data_va = bo_va;
1186 return 0;
1187
1188 error:
1189 amdgpu_vm_bo_del(adev, bo_va);
1190 ttm_eu_backoff_reservation(&ticket, &list);
1191 amdgpu_sync_free(&sync);
1192 return r;
1193 }
1194
amdgpu_mes_ctx_unmap_meta_data(struct amdgpu_device * adev,struct amdgpu_mes_ctx_data * ctx_data)1195 int amdgpu_mes_ctx_unmap_meta_data(struct amdgpu_device *adev,
1196 struct amdgpu_mes_ctx_data *ctx_data)
1197 {
1198 struct amdgpu_bo_va *bo_va = ctx_data->meta_data_va;
1199 struct amdgpu_bo *bo = ctx_data->meta_data_obj;
1200 struct amdgpu_vm *vm = bo_va->base.vm;
1201 struct amdgpu_bo_list_entry vm_pd;
1202 struct list_head list, duplicates;
1203 struct dma_fence *fence = NULL;
1204 struct ttm_validate_buffer tv;
1205 struct ww_acquire_ctx ticket;
1206 long r = 0;
1207
1208 INIT_LIST_HEAD(&list);
1209 INIT_LIST_HEAD(&duplicates);
1210
1211 tv.bo = &bo->tbo;
1212 tv.num_shared = 2;
1213 list_add(&tv.head, &list);
1214
1215 amdgpu_vm_get_pd_bo(vm, &list, &vm_pd);
1216
1217 r = ttm_eu_reserve_buffers(&ticket, &list, false, &duplicates);
1218 if (r) {
1219 dev_err(adev->dev, "leaking bo va because "
1220 "we fail to reserve bo (%ld)\n", r);
1221 return r;
1222 }
1223
1224 amdgpu_vm_bo_del(adev, bo_va);
1225 if (!amdgpu_vm_ready(vm))
1226 goto out_unlock;
1227
1228 r = dma_resv_get_singleton(bo->tbo.base.resv, DMA_RESV_USAGE_BOOKKEEP, &fence);
1229 if (r)
1230 goto out_unlock;
1231 if (fence) {
1232 amdgpu_bo_fence(bo, fence, true);
1233 fence = NULL;
1234 }
1235
1236 r = amdgpu_vm_clear_freed(adev, vm, &fence);
1237 if (r || !fence)
1238 goto out_unlock;
1239
1240 dma_fence_wait(fence, false);
1241 amdgpu_bo_fence(bo, fence, true);
1242 dma_fence_put(fence);
1243
1244 out_unlock:
1245 if (unlikely(r < 0))
1246 dev_err(adev->dev, "failed to clear page tables (%ld)\n", r);
1247 ttm_eu_backoff_reservation(&ticket, &list);
1248
1249 return r;
1250 }
1251
amdgpu_mes_test_create_gang_and_queues(struct amdgpu_device * adev,int pasid,int * gang_id,int queue_type,int num_queue,struct amdgpu_ring ** added_rings,struct amdgpu_mes_ctx_data * ctx_data)1252 static int amdgpu_mes_test_create_gang_and_queues(struct amdgpu_device *adev,
1253 int pasid, int *gang_id,
1254 int queue_type, int num_queue,
1255 struct amdgpu_ring **added_rings,
1256 struct amdgpu_mes_ctx_data *ctx_data)
1257 {
1258 struct amdgpu_ring *ring;
1259 struct amdgpu_mes_gang_properties gprops = {0};
1260 int r, j;
1261
1262 /* create a gang for the process */
1263 gprops.priority = AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
1264 gprops.gang_quantum = adev->mes.default_gang_quantum;
1265 gprops.inprocess_gang_priority = AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
1266 gprops.priority_level = AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
1267 gprops.global_priority_level = AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
1268
1269 r = amdgpu_mes_add_gang(adev, pasid, &gprops, gang_id);
1270 if (r) {
1271 DRM_ERROR("failed to add gang\n");
1272 return r;
1273 }
1274
1275 /* create queues for the gang */
1276 for (j = 0; j < num_queue; j++) {
1277 r = amdgpu_mes_add_ring(adev, *gang_id, queue_type, j,
1278 ctx_data, &ring);
1279 if (r) {
1280 DRM_ERROR("failed to add ring\n");
1281 break;
1282 }
1283
1284 DRM_INFO("ring %s was added\n", ring->name);
1285 added_rings[j] = ring;
1286 }
1287
1288 return 0;
1289 }
1290
amdgpu_mes_test_queues(struct amdgpu_ring ** added_rings)1291 static int amdgpu_mes_test_queues(struct amdgpu_ring **added_rings)
1292 {
1293 struct amdgpu_ring *ring;
1294 int i, r;
1295
1296 for (i = 0; i < AMDGPU_MES_CTX_MAX_RINGS; i++) {
1297 ring = added_rings[i];
1298 if (!ring)
1299 continue;
1300
1301 r = amdgpu_ring_test_ring(ring);
1302 if (r) {
1303 DRM_DEV_ERROR(ring->adev->dev,
1304 "ring %s test failed (%d)\n",
1305 ring->name, r);
1306 return r;
1307 } else
1308 DRM_INFO("ring %s test pass\n", ring->name);
1309
1310 r = amdgpu_ring_test_ib(ring, 1000 * 10);
1311 if (r) {
1312 DRM_DEV_ERROR(ring->adev->dev,
1313 "ring %s ib test failed (%d)\n",
1314 ring->name, r);
1315 return r;
1316 } else
1317 DRM_INFO("ring %s ib test pass\n", ring->name);
1318 }
1319
1320 return 0;
1321 }
1322
amdgpu_mes_self_test(struct amdgpu_device * adev)1323 int amdgpu_mes_self_test(struct amdgpu_device *adev)
1324 {
1325 struct amdgpu_vm *vm = NULL;
1326 struct amdgpu_mes_ctx_data ctx_data = {0};
1327 struct amdgpu_ring *added_rings[AMDGPU_MES_CTX_MAX_RINGS] = { NULL };
1328 int gang_ids[3] = {0};
1329 int queue_types[][2] = { { AMDGPU_RING_TYPE_GFX,
1330 AMDGPU_MES_CTX_MAX_GFX_RINGS},
1331 { AMDGPU_RING_TYPE_COMPUTE,
1332 AMDGPU_MES_CTX_MAX_COMPUTE_RINGS},
1333 { AMDGPU_RING_TYPE_SDMA,
1334 AMDGPU_MES_CTX_MAX_SDMA_RINGS } };
1335 int i, r, pasid, k = 0;
1336
1337 pasid = amdgpu_pasid_alloc(16);
1338 if (pasid < 0) {
1339 dev_warn(adev->dev, "No more PASIDs available!");
1340 pasid = 0;
1341 }
1342
1343 vm = kzalloc(sizeof(*vm), GFP_KERNEL);
1344 if (!vm) {
1345 r = -ENOMEM;
1346 goto error_pasid;
1347 }
1348
1349 r = amdgpu_vm_init(adev, vm);
1350 if (r) {
1351 DRM_ERROR("failed to initialize vm\n");
1352 goto error_pasid;
1353 }
1354
1355 r = amdgpu_mes_ctx_alloc_meta_data(adev, &ctx_data);
1356 if (r) {
1357 DRM_ERROR("failed to alloc ctx meta data\n");
1358 goto error_fini;
1359 }
1360
1361 ctx_data.meta_data_gpu_addr = AMDGPU_VA_RESERVED_SIZE;
1362 r = amdgpu_mes_ctx_map_meta_data(adev, vm, &ctx_data);
1363 if (r) {
1364 DRM_ERROR("failed to map ctx meta data\n");
1365 goto error_vm;
1366 }
1367
1368 r = amdgpu_mes_create_process(adev, pasid, vm);
1369 if (r) {
1370 DRM_ERROR("failed to create MES process\n");
1371 goto error_vm;
1372 }
1373
1374 for (i = 0; i < ARRAY_SIZE(queue_types); i++) {
1375 /* On GFX v10.3, fw hasn't supported to map sdma queue. */
1376 if (adev->ip_versions[GC_HWIP][0] >= IP_VERSION(10, 3, 0) &&
1377 adev->ip_versions[GC_HWIP][0] < IP_VERSION(11, 0, 0) &&
1378 queue_types[i][0] == AMDGPU_RING_TYPE_SDMA)
1379 continue;
1380
1381 r = amdgpu_mes_test_create_gang_and_queues(adev, pasid,
1382 &gang_ids[i],
1383 queue_types[i][0],
1384 queue_types[i][1],
1385 &added_rings[k],
1386 &ctx_data);
1387 if (r)
1388 goto error_queues;
1389
1390 k += queue_types[i][1];
1391 }
1392
1393 /* start ring test and ib test for MES queues */
1394 amdgpu_mes_test_queues(added_rings);
1395
1396 error_queues:
1397 /* remove all queues */
1398 for (i = 0; i < ARRAY_SIZE(added_rings); i++) {
1399 if (!added_rings[i])
1400 continue;
1401 amdgpu_mes_remove_ring(adev, added_rings[i]);
1402 }
1403
1404 for (i = 0; i < ARRAY_SIZE(gang_ids); i++) {
1405 if (!gang_ids[i])
1406 continue;
1407 amdgpu_mes_remove_gang(adev, gang_ids[i]);
1408 }
1409
1410 amdgpu_mes_destroy_process(adev, pasid);
1411
1412 error_vm:
1413 amdgpu_mes_ctx_unmap_meta_data(adev, &ctx_data);
1414
1415 error_fini:
1416 amdgpu_vm_fini(adev, vm);
1417
1418 error_pasid:
1419 if (pasid)
1420 amdgpu_pasid_free(pasid);
1421
1422 amdgpu_mes_ctx_free_meta_data(&ctx_data);
1423 kfree(vm);
1424 return 0;
1425 }
1426
