Loading...
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
35int 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
42static int amdgpu_mes_kernel_doorbell_get(struct amdgpu_device *adev,
43 struct amdgpu_mes_process *process,
44 int ip_type, uint64_t *doorbell_index)
45{
46 unsigned int offset, found;
47 struct amdgpu_mes *mes = &adev->mes;
48
49 if (ip_type == AMDGPU_RING_TYPE_SDMA)
50 offset = adev->doorbell_index.sdma_engine[0];
51 else
52 offset = 0;
53
54 found = find_next_zero_bit(mes->doorbell_bitmap, mes->num_mes_dbs, offset);
55 if (found >= mes->num_mes_dbs) {
56 DRM_WARN("No doorbell available\n");
57 return -ENOSPC;
58 }
59
60 set_bit(found, mes->doorbell_bitmap);
61
62 /* Get the absolute doorbell index on BAR */
63 *doorbell_index = mes->db_start_dw_offset + found * 2;
64 return 0;
65}
66
67static void amdgpu_mes_kernel_doorbell_free(struct amdgpu_device *adev,
68 struct amdgpu_mes_process *process,
69 uint32_t doorbell_index)
70{
71 unsigned int old, rel_index;
72 struct amdgpu_mes *mes = &adev->mes;
73
74 /* Find the relative index of the doorbell in this object */
75 rel_index = (doorbell_index - mes->db_start_dw_offset) / 2;
76 old = test_and_clear_bit(rel_index, mes->doorbell_bitmap);
77 WARN_ON(!old);
78}
79
80static int amdgpu_mes_doorbell_init(struct amdgpu_device *adev)
81{
82 int i;
83 struct amdgpu_mes *mes = &adev->mes;
84
85 /* Bitmap for dynamic allocation of kernel doorbells */
86 mes->doorbell_bitmap = bitmap_zalloc(PAGE_SIZE / sizeof(u32), GFP_KERNEL);
87 if (!mes->doorbell_bitmap) {
88 DRM_ERROR("Failed to allocate MES doorbell bitmap\n");
89 return -ENOMEM;
90 }
91
92 mes->num_mes_dbs = PAGE_SIZE / AMDGPU_ONE_DOORBELL_SIZE;
93 for (i = 0; i < AMDGPU_MES_PRIORITY_NUM_LEVELS; i++) {
94 adev->mes.aggregated_doorbells[i] = mes->db_start_dw_offset + i * 2;
95 set_bit(i, mes->doorbell_bitmap);
96 }
97
98 return 0;
99}
100
101static int amdgpu_mes_event_log_init(struct amdgpu_device *adev)
102{
103 int r;
104
105 if (!amdgpu_mes_log_enable)
106 return 0;
107
108 r = amdgpu_bo_create_kernel(adev, AMDGPU_MES_LOG_BUFFER_SIZE, PAGE_SIZE,
109 AMDGPU_GEM_DOMAIN_GTT,
110 &adev->mes.event_log_gpu_obj,
111 &adev->mes.event_log_gpu_addr,
112 &adev->mes.event_log_cpu_addr);
113 if (r) {
114 dev_warn(adev->dev, "failed to create MES event log buffer (%d)", r);
115 return r;
116 }
117
118 memset(adev->mes.event_log_cpu_addr, 0, PAGE_SIZE);
119
120 return 0;
121
122}
123
124static void amdgpu_mes_doorbell_free(struct amdgpu_device *adev)
125{
126 bitmap_free(adev->mes.doorbell_bitmap);
127}
128
129int amdgpu_mes_init(struct amdgpu_device *adev)
130{
131 int i, r;
132
133 adev->mes.adev = adev;
134
135 idr_init(&adev->mes.pasid_idr);
136 idr_init(&adev->mes.gang_id_idr);
137 idr_init(&adev->mes.queue_id_idr);
138 ida_init(&adev->mes.doorbell_ida);
139 spin_lock_init(&adev->mes.queue_id_lock);
140 spin_lock_init(&adev->mes.ring_lock);
141 mutex_init(&adev->mes.mutex_hidden);
142
143 adev->mes.total_max_queue = AMDGPU_FENCE_MES_QUEUE_ID_MASK;
144 adev->mes.vmid_mask_mmhub = 0xffffff00;
145 adev->mes.vmid_mask_gfxhub = 0xffffff00;
146
147 for (i = 0; i < AMDGPU_MES_MAX_COMPUTE_PIPES; i++) {
148 /* use only 1st MEC pipes */
149 if (i >= 4)
150 continue;
151 adev->mes.compute_hqd_mask[i] = 0xc;
152 }
153
154 for (i = 0; i < AMDGPU_MES_MAX_GFX_PIPES; i++)
155 adev->mes.gfx_hqd_mask[i] = i ? 0 : 0xfffffffe;
156
157 for (i = 0; i < AMDGPU_MES_MAX_SDMA_PIPES; i++) {
158 if (amdgpu_ip_version(adev, SDMA0_HWIP, 0) <
159 IP_VERSION(6, 0, 0))
160 adev->mes.sdma_hqd_mask[i] = i ? 0 : 0x3fc;
161 /* zero sdma_hqd_mask for non-existent engine */
162 else if (adev->sdma.num_instances == 1)
163 adev->mes.sdma_hqd_mask[i] = i ? 0 : 0xfc;
164 else
165 adev->mes.sdma_hqd_mask[i] = 0xfc;
166 }
167
168 r = amdgpu_device_wb_get(adev, &adev->mes.sch_ctx_offs);
169 if (r) {
170 dev_err(adev->dev,
171 "(%d) ring trail_fence_offs wb alloc failed\n", r);
172 goto error_ids;
173 }
174 adev->mes.sch_ctx_gpu_addr =
175 adev->wb.gpu_addr + (adev->mes.sch_ctx_offs * 4);
176 adev->mes.sch_ctx_ptr =
177 (uint64_t *)&adev->wb.wb[adev->mes.sch_ctx_offs];
178
179 r = amdgpu_device_wb_get(adev, &adev->mes.query_status_fence_offs);
180 if (r) {
181 amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
182 dev_err(adev->dev,
183 "(%d) query_status_fence_offs wb alloc failed\n", r);
184 goto error_ids;
185 }
186 adev->mes.query_status_fence_gpu_addr =
187 adev->wb.gpu_addr + (adev->mes.query_status_fence_offs * 4);
188 adev->mes.query_status_fence_ptr =
189 (uint64_t *)&adev->wb.wb[adev->mes.query_status_fence_offs];
190
191 r = amdgpu_device_wb_get(adev, &adev->mes.read_val_offs);
192 if (r) {
193 amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
194 amdgpu_device_wb_free(adev, adev->mes.query_status_fence_offs);
195 dev_err(adev->dev,
196 "(%d) read_val_offs alloc failed\n", r);
197 goto error_ids;
198 }
199 adev->mes.read_val_gpu_addr =
200 adev->wb.gpu_addr + (adev->mes.read_val_offs * 4);
201 adev->mes.read_val_ptr =
202 (uint32_t *)&adev->wb.wb[adev->mes.read_val_offs];
203
204 r = amdgpu_mes_doorbell_init(adev);
205 if (r)
206 goto error;
207
208 r = amdgpu_mes_event_log_init(adev);
209 if (r)
210 goto error_doorbell;
211
212 return 0;
213
214error_doorbell:
215 amdgpu_mes_doorbell_free(adev);
216error:
217 amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
218 amdgpu_device_wb_free(adev, adev->mes.query_status_fence_offs);
219 amdgpu_device_wb_free(adev, adev->mes.read_val_offs);
220error_ids:
221 idr_destroy(&adev->mes.pasid_idr);
222 idr_destroy(&adev->mes.gang_id_idr);
223 idr_destroy(&adev->mes.queue_id_idr);
224 ida_destroy(&adev->mes.doorbell_ida);
225 mutex_destroy(&adev->mes.mutex_hidden);
226 return r;
227}
228
229void amdgpu_mes_fini(struct amdgpu_device *adev)
230{
231 amdgpu_bo_free_kernel(&adev->mes.event_log_gpu_obj,
232 &adev->mes.event_log_gpu_addr,
233 &adev->mes.event_log_cpu_addr);
234
235 amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
236 amdgpu_device_wb_free(adev, adev->mes.query_status_fence_offs);
237 amdgpu_device_wb_free(adev, adev->mes.read_val_offs);
238 amdgpu_mes_doorbell_free(adev);
239
240 idr_destroy(&adev->mes.pasid_idr);
241 idr_destroy(&adev->mes.gang_id_idr);
242 idr_destroy(&adev->mes.queue_id_idr);
243 ida_destroy(&adev->mes.doorbell_ida);
244 mutex_destroy(&adev->mes.mutex_hidden);
245}
246
247static void amdgpu_mes_queue_free_mqd(struct amdgpu_mes_queue *q)
248{
249 amdgpu_bo_free_kernel(&q->mqd_obj,
250 &q->mqd_gpu_addr,
251 &q->mqd_cpu_ptr);
252}
253
254int amdgpu_mes_create_process(struct amdgpu_device *adev, int pasid,
255 struct amdgpu_vm *vm)
256{
257 struct amdgpu_mes_process *process;
258 int r;
259
260 /* allocate the mes process buffer */
261 process = kzalloc(sizeof(struct amdgpu_mes_process), GFP_KERNEL);
262 if (!process) {
263 DRM_ERROR("no more memory to create mes process\n");
264 return -ENOMEM;
265 }
266
267 /* allocate the process context bo and map it */
268 r = amdgpu_bo_create_kernel(adev, AMDGPU_MES_PROC_CTX_SIZE, PAGE_SIZE,
269 AMDGPU_GEM_DOMAIN_GTT,
270 &process->proc_ctx_bo,
271 &process->proc_ctx_gpu_addr,
272 &process->proc_ctx_cpu_ptr);
273 if (r) {
274 DRM_ERROR("failed to allocate process context bo\n");
275 goto clean_up_memory;
276 }
277 memset(process->proc_ctx_cpu_ptr, 0, AMDGPU_MES_PROC_CTX_SIZE);
278
279 /*
280 * Avoid taking any other locks under MES lock to avoid circular
281 * lock dependencies.
282 */
283 amdgpu_mes_lock(&adev->mes);
284
285 /* add the mes process to idr list */
286 r = idr_alloc(&adev->mes.pasid_idr, process, pasid, pasid + 1,
287 GFP_KERNEL);
288 if (r < 0) {
289 DRM_ERROR("failed to lock pasid=%d\n", pasid);
290 goto clean_up_ctx;
291 }
292
293 INIT_LIST_HEAD(&process->gang_list);
294 process->vm = vm;
295 process->pasid = pasid;
296 process->process_quantum = adev->mes.default_process_quantum;
297 process->pd_gpu_addr = amdgpu_bo_gpu_offset(vm->root.bo);
298
299 amdgpu_mes_unlock(&adev->mes);
300 return 0;
301
302clean_up_ctx:
303 amdgpu_mes_unlock(&adev->mes);
304 amdgpu_bo_free_kernel(&process->proc_ctx_bo,
305 &process->proc_ctx_gpu_addr,
306 &process->proc_ctx_cpu_ptr);
307clean_up_memory:
308 kfree(process);
309 return r;
310}
311
312void amdgpu_mes_destroy_process(struct amdgpu_device *adev, int pasid)
313{
314 struct amdgpu_mes_process *process;
315 struct amdgpu_mes_gang *gang, *tmp1;
316 struct amdgpu_mes_queue *queue, *tmp2;
317 struct mes_remove_queue_input queue_input;
318 unsigned long flags;
319 int r;
320
321 /*
322 * Avoid taking any other locks under MES lock to avoid circular
323 * lock dependencies.
324 */
325 amdgpu_mes_lock(&adev->mes);
326
327 process = idr_find(&adev->mes.pasid_idr, pasid);
328 if (!process) {
329 DRM_WARN("pasid %d doesn't exist\n", pasid);
330 amdgpu_mes_unlock(&adev->mes);
331 return;
332 }
333
334 /* Remove all queues from hardware */
335 list_for_each_entry_safe(gang, tmp1, &process->gang_list, list) {
336 list_for_each_entry_safe(queue, tmp2, &gang->queue_list, list) {
337 spin_lock_irqsave(&adev->mes.queue_id_lock, flags);
338 idr_remove(&adev->mes.queue_id_idr, queue->queue_id);
339 spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
340
341 queue_input.doorbell_offset = queue->doorbell_off;
342 queue_input.gang_context_addr = gang->gang_ctx_gpu_addr;
343
344 r = adev->mes.funcs->remove_hw_queue(&adev->mes,
345 &queue_input);
346 if (r)
347 DRM_WARN("failed to remove hardware queue\n");
348 }
349
350 idr_remove(&adev->mes.gang_id_idr, gang->gang_id);
351 }
352
353 idr_remove(&adev->mes.pasid_idr, pasid);
354 amdgpu_mes_unlock(&adev->mes);
355
356 /* free all memory allocated by the process */
357 list_for_each_entry_safe(gang, tmp1, &process->gang_list, list) {
358 /* free all queues in the gang */
359 list_for_each_entry_safe(queue, tmp2, &gang->queue_list, list) {
360 amdgpu_mes_queue_free_mqd(queue);
361 list_del(&queue->list);
362 kfree(queue);
363 }
364 amdgpu_bo_free_kernel(&gang->gang_ctx_bo,
365 &gang->gang_ctx_gpu_addr,
366 &gang->gang_ctx_cpu_ptr);
367 list_del(&gang->list);
368 kfree(gang);
369
370 }
371 amdgpu_bo_free_kernel(&process->proc_ctx_bo,
372 &process->proc_ctx_gpu_addr,
373 &process->proc_ctx_cpu_ptr);
374 kfree(process);
375}
376
377int amdgpu_mes_add_gang(struct amdgpu_device *adev, int pasid,
378 struct amdgpu_mes_gang_properties *gprops,
379 int *gang_id)
380{
381 struct amdgpu_mes_process *process;
382 struct amdgpu_mes_gang *gang;
383 int r;
384
385 /* allocate the mes gang buffer */
386 gang = kzalloc(sizeof(struct amdgpu_mes_gang), GFP_KERNEL);
387 if (!gang) {
388 return -ENOMEM;
389 }
390
391 /* allocate the gang context bo and map it to cpu space */
392 r = amdgpu_bo_create_kernel(adev, AMDGPU_MES_GANG_CTX_SIZE, PAGE_SIZE,
393 AMDGPU_GEM_DOMAIN_GTT,
394 &gang->gang_ctx_bo,
395 &gang->gang_ctx_gpu_addr,
396 &gang->gang_ctx_cpu_ptr);
397 if (r) {
398 DRM_ERROR("failed to allocate process context bo\n");
399 goto clean_up_mem;
400 }
401 memset(gang->gang_ctx_cpu_ptr, 0, AMDGPU_MES_GANG_CTX_SIZE);
402
403 /*
404 * Avoid taking any other locks under MES lock to avoid circular
405 * lock dependencies.
406 */
407 amdgpu_mes_lock(&adev->mes);
408
409 process = idr_find(&adev->mes.pasid_idr, pasid);
410 if (!process) {
411 DRM_ERROR("pasid %d doesn't exist\n", pasid);
412 r = -EINVAL;
413 goto clean_up_ctx;
414 }
415
416 /* add the mes gang to idr list */
417 r = idr_alloc(&adev->mes.gang_id_idr, gang, 1, 0,
418 GFP_KERNEL);
419 if (r < 0) {
420 DRM_ERROR("failed to allocate idr for gang\n");
421 goto clean_up_ctx;
422 }
423
424 gang->gang_id = r;
425 *gang_id = r;
426
427 INIT_LIST_HEAD(&gang->queue_list);
428 gang->process = process;
429 gang->priority = gprops->priority;
430 gang->gang_quantum = gprops->gang_quantum ?
431 gprops->gang_quantum : adev->mes.default_gang_quantum;
432 gang->global_priority_level = gprops->global_priority_level;
433 gang->inprocess_gang_priority = gprops->inprocess_gang_priority;
434 list_add_tail(&gang->list, &process->gang_list);
435
436 amdgpu_mes_unlock(&adev->mes);
437 return 0;
438
439clean_up_ctx:
440 amdgpu_mes_unlock(&adev->mes);
441 amdgpu_bo_free_kernel(&gang->gang_ctx_bo,
442 &gang->gang_ctx_gpu_addr,
443 &gang->gang_ctx_cpu_ptr);
444clean_up_mem:
445 kfree(gang);
446 return r;
447}
448
449int amdgpu_mes_remove_gang(struct amdgpu_device *adev, int gang_id)
450{
451 struct amdgpu_mes_gang *gang;
452
453 /*
454 * Avoid taking any other locks under MES lock to avoid circular
455 * lock dependencies.
456 */
457 amdgpu_mes_lock(&adev->mes);
458
459 gang = idr_find(&adev->mes.gang_id_idr, gang_id);
460 if (!gang) {
461 DRM_ERROR("gang id %d doesn't exist\n", gang_id);
462 amdgpu_mes_unlock(&adev->mes);
463 return -EINVAL;
464 }
465
466 if (!list_empty(&gang->queue_list)) {
467 DRM_ERROR("queue list is not empty\n");
468 amdgpu_mes_unlock(&adev->mes);
469 return -EBUSY;
470 }
471
472 idr_remove(&adev->mes.gang_id_idr, gang->gang_id);
473 list_del(&gang->list);
474 amdgpu_mes_unlock(&adev->mes);
475
476 amdgpu_bo_free_kernel(&gang->gang_ctx_bo,
477 &gang->gang_ctx_gpu_addr,
478 &gang->gang_ctx_cpu_ptr);
479
480 kfree(gang);
481
482 return 0;
483}
484
485int amdgpu_mes_suspend(struct amdgpu_device *adev)
486{
487 struct idr *idp;
488 struct amdgpu_mes_process *process;
489 struct amdgpu_mes_gang *gang;
490 struct mes_suspend_gang_input input;
491 int r, pasid;
492
493 /*
494 * Avoid taking any other locks under MES lock to avoid circular
495 * lock dependencies.
496 */
497 amdgpu_mes_lock(&adev->mes);
498
499 idp = &adev->mes.pasid_idr;
500
501 idr_for_each_entry(idp, process, pasid) {
502 list_for_each_entry(gang, &process->gang_list, list) {
503 r = adev->mes.funcs->suspend_gang(&adev->mes, &input);
504 if (r)
505 DRM_ERROR("failed to suspend pasid %d gangid %d",
506 pasid, gang->gang_id);
507 }
508 }
509
510 amdgpu_mes_unlock(&adev->mes);
511 return 0;
512}
513
514int amdgpu_mes_resume(struct amdgpu_device *adev)
515{
516 struct idr *idp;
517 struct amdgpu_mes_process *process;
518 struct amdgpu_mes_gang *gang;
519 struct mes_resume_gang_input input;
520 int r, pasid;
521
522 /*
523 * Avoid taking any other locks under MES lock to avoid circular
524 * lock dependencies.
525 */
526 amdgpu_mes_lock(&adev->mes);
527
528 idp = &adev->mes.pasid_idr;
529
530 idr_for_each_entry(idp, process, pasid) {
531 list_for_each_entry(gang, &process->gang_list, list) {
532 r = adev->mes.funcs->resume_gang(&adev->mes, &input);
533 if (r)
534 DRM_ERROR("failed to resume pasid %d gangid %d",
535 pasid, gang->gang_id);
536 }
537 }
538
539 amdgpu_mes_unlock(&adev->mes);
540 return 0;
541}
542
543static int amdgpu_mes_queue_alloc_mqd(struct amdgpu_device *adev,
544 struct amdgpu_mes_queue *q,
545 struct amdgpu_mes_queue_properties *p)
546{
547 struct amdgpu_mqd *mqd_mgr = &adev->mqds[p->queue_type];
548 u32 mqd_size = mqd_mgr->mqd_size;
549 int r;
550
551 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
552 AMDGPU_GEM_DOMAIN_GTT,
553 &q->mqd_obj,
554 &q->mqd_gpu_addr, &q->mqd_cpu_ptr);
555 if (r) {
556 dev_warn(adev->dev, "failed to create queue mqd bo (%d)", r);
557 return r;
558 }
559 memset(q->mqd_cpu_ptr, 0, mqd_size);
560
561 r = amdgpu_bo_reserve(q->mqd_obj, false);
562 if (unlikely(r != 0))
563 goto clean_up;
564
565 return 0;
566
567clean_up:
568 amdgpu_bo_free_kernel(&q->mqd_obj,
569 &q->mqd_gpu_addr,
570 &q->mqd_cpu_ptr);
571 return r;
572}
573
574static void amdgpu_mes_queue_init_mqd(struct amdgpu_device *adev,
575 struct amdgpu_mes_queue *q,
576 struct amdgpu_mes_queue_properties *p)
577{
578 struct amdgpu_mqd *mqd_mgr = &adev->mqds[p->queue_type];
579 struct amdgpu_mqd_prop mqd_prop = {0};
580
581 mqd_prop.mqd_gpu_addr = q->mqd_gpu_addr;
582 mqd_prop.hqd_base_gpu_addr = p->hqd_base_gpu_addr;
583 mqd_prop.rptr_gpu_addr = p->rptr_gpu_addr;
584 mqd_prop.wptr_gpu_addr = p->wptr_gpu_addr;
585 mqd_prop.queue_size = p->queue_size;
586 mqd_prop.use_doorbell = true;
587 mqd_prop.doorbell_index = p->doorbell_off;
588 mqd_prop.eop_gpu_addr = p->eop_gpu_addr;
589 mqd_prop.hqd_pipe_priority = p->hqd_pipe_priority;
590 mqd_prop.hqd_queue_priority = p->hqd_queue_priority;
591 mqd_prop.hqd_active = false;
592
593 if (p->queue_type == AMDGPU_RING_TYPE_GFX ||
594 p->queue_type == AMDGPU_RING_TYPE_COMPUTE) {
595 mutex_lock(&adev->srbm_mutex);
596 amdgpu_gfx_select_me_pipe_q(adev, p->ring->me, p->ring->pipe, 0, 0, 0);
597 }
598
599 mqd_mgr->init_mqd(adev, q->mqd_cpu_ptr, &mqd_prop);
600
601 if (p->queue_type == AMDGPU_RING_TYPE_GFX ||
602 p->queue_type == AMDGPU_RING_TYPE_COMPUTE) {
603 amdgpu_gfx_select_me_pipe_q(adev, 0, 0, 0, 0, 0);
604 mutex_unlock(&adev->srbm_mutex);
605 }
606
607 amdgpu_bo_unreserve(q->mqd_obj);
608}
609
610int amdgpu_mes_add_hw_queue(struct amdgpu_device *adev, int gang_id,
611 struct amdgpu_mes_queue_properties *qprops,
612 int *queue_id)
613{
614 struct amdgpu_mes_queue *queue;
615 struct amdgpu_mes_gang *gang;
616 struct mes_add_queue_input queue_input;
617 unsigned long flags;
618 int r;
619
620 memset(&queue_input, 0, sizeof(struct mes_add_queue_input));
621
622 /* allocate the mes queue buffer */
623 queue = kzalloc(sizeof(struct amdgpu_mes_queue), GFP_KERNEL);
624 if (!queue) {
625 DRM_ERROR("Failed to allocate memory for queue\n");
626 return -ENOMEM;
627 }
628
629 /* Allocate the queue mqd */
630 r = amdgpu_mes_queue_alloc_mqd(adev, queue, qprops);
631 if (r)
632 goto clean_up_memory;
633
634 /*
635 * Avoid taking any other locks under MES lock to avoid circular
636 * lock dependencies.
637 */
638 amdgpu_mes_lock(&adev->mes);
639
640 gang = idr_find(&adev->mes.gang_id_idr, gang_id);
641 if (!gang) {
642 DRM_ERROR("gang id %d doesn't exist\n", gang_id);
643 r = -EINVAL;
644 goto clean_up_mqd;
645 }
646
647 /* add the mes gang to idr list */
648 spin_lock_irqsave(&adev->mes.queue_id_lock, flags);
649 r = idr_alloc(&adev->mes.queue_id_idr, queue, 1, 0,
650 GFP_ATOMIC);
651 if (r < 0) {
652 spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
653 goto clean_up_mqd;
654 }
655 spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
656 *queue_id = queue->queue_id = r;
657
658 /* allocate a doorbell index for the queue */
659 r = amdgpu_mes_kernel_doorbell_get(adev, gang->process,
660 qprops->queue_type,
661 &qprops->doorbell_off);
662 if (r)
663 goto clean_up_queue_id;
664
665 /* initialize the queue mqd */
666 amdgpu_mes_queue_init_mqd(adev, queue, qprops);
667
668 /* add hw queue to mes */
669 queue_input.process_id = gang->process->pasid;
670
671 queue_input.page_table_base_addr =
672 adev->vm_manager.vram_base_offset + gang->process->pd_gpu_addr -
673 adev->gmc.vram_start;
674
675 queue_input.process_va_start = 0;
676 queue_input.process_va_end =
677 (adev->vm_manager.max_pfn - 1) << AMDGPU_GPU_PAGE_SHIFT;
678 queue_input.process_quantum = gang->process->process_quantum;
679 queue_input.process_context_addr = gang->process->proc_ctx_gpu_addr;
680 queue_input.gang_quantum = gang->gang_quantum;
681 queue_input.gang_context_addr = gang->gang_ctx_gpu_addr;
682 queue_input.inprocess_gang_priority = gang->inprocess_gang_priority;
683 queue_input.gang_global_priority_level = gang->global_priority_level;
684 queue_input.doorbell_offset = qprops->doorbell_off;
685 queue_input.mqd_addr = queue->mqd_gpu_addr;
686 queue_input.wptr_addr = qprops->wptr_gpu_addr;
687 queue_input.wptr_mc_addr = qprops->wptr_mc_addr;
688 queue_input.queue_type = qprops->queue_type;
689 queue_input.paging = qprops->paging;
690 queue_input.is_kfd_process = 0;
691
692 r = adev->mes.funcs->add_hw_queue(&adev->mes, &queue_input);
693 if (r) {
694 DRM_ERROR("failed to add hardware queue to MES, doorbell=0x%llx\n",
695 qprops->doorbell_off);
696 goto clean_up_doorbell;
697 }
698
699 DRM_DEBUG("MES hw queue was added, pasid=%d, gang id=%d, "
700 "queue type=%d, doorbell=0x%llx\n",
701 gang->process->pasid, gang_id, qprops->queue_type,
702 qprops->doorbell_off);
703
704 queue->ring = qprops->ring;
705 queue->doorbell_off = qprops->doorbell_off;
706 queue->wptr_gpu_addr = qprops->wptr_gpu_addr;
707 queue->queue_type = qprops->queue_type;
708 queue->paging = qprops->paging;
709 queue->gang = gang;
710 queue->ring->mqd_ptr = queue->mqd_cpu_ptr;
711 list_add_tail(&queue->list, &gang->queue_list);
712
713 amdgpu_mes_unlock(&adev->mes);
714 return 0;
715
716clean_up_doorbell:
717 amdgpu_mes_kernel_doorbell_free(adev, gang->process,
718 qprops->doorbell_off);
719clean_up_queue_id:
720 spin_lock_irqsave(&adev->mes.queue_id_lock, flags);
721 idr_remove(&adev->mes.queue_id_idr, queue->queue_id);
722 spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
723clean_up_mqd:
724 amdgpu_mes_unlock(&adev->mes);
725 amdgpu_mes_queue_free_mqd(queue);
726clean_up_memory:
727 kfree(queue);
728 return r;
729}
730
731int amdgpu_mes_remove_hw_queue(struct amdgpu_device *adev, int queue_id)
732{
733 unsigned long flags;
734 struct amdgpu_mes_queue *queue;
735 struct amdgpu_mes_gang *gang;
736 struct mes_remove_queue_input queue_input;
737 int r;
738
739 /*
740 * Avoid taking any other locks under MES lock to avoid circular
741 * lock dependencies.
742 */
743 amdgpu_mes_lock(&adev->mes);
744
745 /* remove the mes gang from idr list */
746 spin_lock_irqsave(&adev->mes.queue_id_lock, flags);
747
748 queue = idr_find(&adev->mes.queue_id_idr, queue_id);
749 if (!queue) {
750 spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
751 amdgpu_mes_unlock(&adev->mes);
752 DRM_ERROR("queue id %d doesn't exist\n", queue_id);
753 return -EINVAL;
754 }
755
756 idr_remove(&adev->mes.queue_id_idr, queue_id);
757 spin_unlock_irqrestore(&adev->mes.queue_id_lock, flags);
758
759 DRM_DEBUG("try to remove queue, doorbell off = 0x%llx\n",
760 queue->doorbell_off);
761
762 gang = queue->gang;
763 queue_input.doorbell_offset = queue->doorbell_off;
764 queue_input.gang_context_addr = gang->gang_ctx_gpu_addr;
765
766 r = adev->mes.funcs->remove_hw_queue(&adev->mes, &queue_input);
767 if (r)
768 DRM_ERROR("failed to remove hardware queue, queue id = %d\n",
769 queue_id);
770
771 list_del(&queue->list);
772 amdgpu_mes_kernel_doorbell_free(adev, gang->process,
773 queue->doorbell_off);
774 amdgpu_mes_unlock(&adev->mes);
775
776 amdgpu_mes_queue_free_mqd(queue);
777 kfree(queue);
778 return 0;
779}
780
781int amdgpu_mes_unmap_legacy_queue(struct amdgpu_device *adev,
782 struct amdgpu_ring *ring,
783 enum amdgpu_unmap_queues_action action,
784 u64 gpu_addr, u64 seq)
785{
786 struct mes_unmap_legacy_queue_input queue_input;
787 int r;
788
789 queue_input.action = action;
790 queue_input.queue_type = ring->funcs->type;
791 queue_input.doorbell_offset = ring->doorbell_index;
792 queue_input.pipe_id = ring->pipe;
793 queue_input.queue_id = ring->queue;
794 queue_input.trail_fence_addr = gpu_addr;
795 queue_input.trail_fence_data = seq;
796
797 r = adev->mes.funcs->unmap_legacy_queue(&adev->mes, &queue_input);
798 if (r)
799 DRM_ERROR("failed to unmap legacy queue\n");
800
801 return r;
802}
803
804uint32_t amdgpu_mes_rreg(struct amdgpu_device *adev, uint32_t reg)
805{
806 struct mes_misc_op_input op_input;
807 int r, val = 0;
808
809 op_input.op = MES_MISC_OP_READ_REG;
810 op_input.read_reg.reg_offset = reg;
811 op_input.read_reg.buffer_addr = adev->mes.read_val_gpu_addr;
812
813 if (!adev->mes.funcs->misc_op) {
814 DRM_ERROR("mes rreg is not supported!\n");
815 goto error;
816 }
817
818 r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
819 if (r)
820 DRM_ERROR("failed to read reg (0x%x)\n", reg);
821 else
822 val = *(adev->mes.read_val_ptr);
823
824error:
825 return val;
826}
827
828int amdgpu_mes_wreg(struct amdgpu_device *adev,
829 uint32_t reg, uint32_t val)
830{
831 struct mes_misc_op_input op_input;
832 int r;
833
834 op_input.op = MES_MISC_OP_WRITE_REG;
835 op_input.write_reg.reg_offset = reg;
836 op_input.write_reg.reg_value = val;
837
838 if (!adev->mes.funcs->misc_op) {
839 DRM_ERROR("mes wreg is not supported!\n");
840 r = -EINVAL;
841 goto error;
842 }
843
844 r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
845 if (r)
846 DRM_ERROR("failed to write reg (0x%x)\n", reg);
847
848error:
849 return r;
850}
851
852int amdgpu_mes_reg_write_reg_wait(struct amdgpu_device *adev,
853 uint32_t reg0, uint32_t reg1,
854 uint32_t ref, uint32_t mask)
855{
856 struct mes_misc_op_input op_input;
857 int r;
858
859 op_input.op = MES_MISC_OP_WRM_REG_WR_WAIT;
860 op_input.wrm_reg.reg0 = reg0;
861 op_input.wrm_reg.reg1 = reg1;
862 op_input.wrm_reg.ref = ref;
863 op_input.wrm_reg.mask = mask;
864
865 if (!adev->mes.funcs->misc_op) {
866 DRM_ERROR("mes reg_write_reg_wait is not supported!\n");
867 r = -EINVAL;
868 goto error;
869 }
870
871 r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
872 if (r)
873 DRM_ERROR("failed to reg_write_reg_wait\n");
874
875error:
876 return r;
877}
878
879int amdgpu_mes_reg_wait(struct amdgpu_device *adev, uint32_t reg,
880 uint32_t val, uint32_t mask)
881{
882 struct mes_misc_op_input op_input;
883 int r;
884
885 op_input.op = MES_MISC_OP_WRM_REG_WAIT;
886 op_input.wrm_reg.reg0 = reg;
887 op_input.wrm_reg.ref = val;
888 op_input.wrm_reg.mask = mask;
889
890 if (!adev->mes.funcs->misc_op) {
891 DRM_ERROR("mes reg wait is not supported!\n");
892 r = -EINVAL;
893 goto error;
894 }
895
896 r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
897 if (r)
898 DRM_ERROR("failed to reg_write_reg_wait\n");
899
900error:
901 return r;
902}
903
904int amdgpu_mes_set_shader_debugger(struct amdgpu_device *adev,
905 uint64_t process_context_addr,
906 uint32_t spi_gdbg_per_vmid_cntl,
907 const uint32_t *tcp_watch_cntl,
908 uint32_t flags,
909 bool trap_en)
910{
911 struct mes_misc_op_input op_input = {0};
912 int r;
913
914 if (!adev->mes.funcs->misc_op) {
915 DRM_ERROR("mes set shader debugger is not supported!\n");
916 return -EINVAL;
917 }
918
919 op_input.op = MES_MISC_OP_SET_SHADER_DEBUGGER;
920 op_input.set_shader_debugger.process_context_addr = process_context_addr;
921 op_input.set_shader_debugger.flags.u32all = flags;
922
923 /* use amdgpu mes_flush_shader_debugger instead */
924 if (op_input.set_shader_debugger.flags.process_ctx_flush)
925 return -EINVAL;
926
927 op_input.set_shader_debugger.spi_gdbg_per_vmid_cntl = spi_gdbg_per_vmid_cntl;
928 memcpy(op_input.set_shader_debugger.tcp_watch_cntl, tcp_watch_cntl,
929 sizeof(op_input.set_shader_debugger.tcp_watch_cntl));
930
931 if (((adev->mes.sched_version & AMDGPU_MES_API_VERSION_MASK) >>
932 AMDGPU_MES_API_VERSION_SHIFT) >= 14)
933 op_input.set_shader_debugger.trap_en = trap_en;
934
935 amdgpu_mes_lock(&adev->mes);
936
937 r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
938 if (r)
939 DRM_ERROR("failed to set_shader_debugger\n");
940
941 amdgpu_mes_unlock(&adev->mes);
942
943 return r;
944}
945
946int amdgpu_mes_flush_shader_debugger(struct amdgpu_device *adev,
947 uint64_t process_context_addr)
948{
949 struct mes_misc_op_input op_input = {0};
950 int r;
951
952 if (!adev->mes.funcs->misc_op) {
953 DRM_ERROR("mes flush shader debugger is not supported!\n");
954 return -EINVAL;
955 }
956
957 op_input.op = MES_MISC_OP_SET_SHADER_DEBUGGER;
958 op_input.set_shader_debugger.process_context_addr = process_context_addr;
959 op_input.set_shader_debugger.flags.process_ctx_flush = true;
960
961 amdgpu_mes_lock(&adev->mes);
962
963 r = adev->mes.funcs->misc_op(&adev->mes, &op_input);
964 if (r)
965 DRM_ERROR("failed to set_shader_debugger\n");
966
967 amdgpu_mes_unlock(&adev->mes);
968
969 return r;
970}
971
972static void
973amdgpu_mes_ring_to_queue_props(struct amdgpu_device *adev,
974 struct amdgpu_ring *ring,
975 struct amdgpu_mes_queue_properties *props)
976{
977 props->queue_type = ring->funcs->type;
978 props->hqd_base_gpu_addr = ring->gpu_addr;
979 props->rptr_gpu_addr = ring->rptr_gpu_addr;
980 props->wptr_gpu_addr = ring->wptr_gpu_addr;
981 props->wptr_mc_addr =
982 ring->mes_ctx->meta_data_mc_addr + ring->wptr_offs;
983 props->queue_size = ring->ring_size;
984 props->eop_gpu_addr = ring->eop_gpu_addr;
985 props->hqd_pipe_priority = AMDGPU_GFX_PIPE_PRIO_NORMAL;
986 props->hqd_queue_priority = AMDGPU_GFX_QUEUE_PRIORITY_MINIMUM;
987 props->paging = false;
988 props->ring = ring;
989}
990
991#define DEFINE_AMDGPU_MES_CTX_GET_OFFS_ENG(_eng) \
992do { \
993 if (id_offs < AMDGPU_MES_CTX_MAX_OFFS) \
994 return offsetof(struct amdgpu_mes_ctx_meta_data, \
995 _eng[ring->idx].slots[id_offs]); \
996 else if (id_offs == AMDGPU_MES_CTX_RING_OFFS) \
997 return offsetof(struct amdgpu_mes_ctx_meta_data, \
998 _eng[ring->idx].ring); \
999 else if (id_offs == AMDGPU_MES_CTX_IB_OFFS) \
1000 return offsetof(struct amdgpu_mes_ctx_meta_data, \
1001 _eng[ring->idx].ib); \
1002 else if (id_offs == AMDGPU_MES_CTX_PADDING_OFFS) \
1003 return offsetof(struct amdgpu_mes_ctx_meta_data, \
1004 _eng[ring->idx].padding); \
1005} while(0)
1006
1007int amdgpu_mes_ctx_get_offs(struct amdgpu_ring *ring, unsigned int id_offs)
1008{
1009 switch (ring->funcs->type) {
1010 case AMDGPU_RING_TYPE_GFX:
1011 DEFINE_AMDGPU_MES_CTX_GET_OFFS_ENG(gfx);
1012 break;
1013 case AMDGPU_RING_TYPE_COMPUTE:
1014 DEFINE_AMDGPU_MES_CTX_GET_OFFS_ENG(compute);
1015 break;
1016 case AMDGPU_RING_TYPE_SDMA:
1017 DEFINE_AMDGPU_MES_CTX_GET_OFFS_ENG(sdma);
1018 break;
1019 default:
1020 break;
1021 }
1022
1023 WARN_ON(1);
1024 return -EINVAL;
1025}
1026
1027int amdgpu_mes_add_ring(struct amdgpu_device *adev, int gang_id,
1028 int queue_type, int idx,
1029 struct amdgpu_mes_ctx_data *ctx_data,
1030 struct amdgpu_ring **out)
1031{
1032 struct amdgpu_ring *ring;
1033 struct amdgpu_mes_gang *gang;
1034 struct amdgpu_mes_queue_properties qprops = {0};
1035 int r, queue_id, pasid;
1036
1037 /*
1038 * Avoid taking any other locks under MES lock to avoid circular
1039 * lock dependencies.
1040 */
1041 amdgpu_mes_lock(&adev->mes);
1042 gang = idr_find(&adev->mes.gang_id_idr, gang_id);
1043 if (!gang) {
1044 DRM_ERROR("gang id %d doesn't exist\n", gang_id);
1045 amdgpu_mes_unlock(&adev->mes);
1046 return -EINVAL;
1047 }
1048 pasid = gang->process->pasid;
1049
1050 ring = kzalloc(sizeof(struct amdgpu_ring), GFP_KERNEL);
1051 if (!ring) {
1052 amdgpu_mes_unlock(&adev->mes);
1053 return -ENOMEM;
1054 }
1055
1056 ring->ring_obj = NULL;
1057 ring->use_doorbell = true;
1058 ring->is_mes_queue = true;
1059 ring->mes_ctx = ctx_data;
1060 ring->idx = idx;
1061 ring->no_scheduler = true;
1062
1063 if (queue_type == AMDGPU_RING_TYPE_COMPUTE) {
1064 int offset = offsetof(struct amdgpu_mes_ctx_meta_data,
1065 compute[ring->idx].mec_hpd);
1066 ring->eop_gpu_addr =
1067 amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
1068 }
1069
1070 switch (queue_type) {
1071 case AMDGPU_RING_TYPE_GFX:
1072 ring->funcs = adev->gfx.gfx_ring[0].funcs;
1073 ring->me = adev->gfx.gfx_ring[0].me;
1074 ring->pipe = adev->gfx.gfx_ring[0].pipe;
1075 break;
1076 case AMDGPU_RING_TYPE_COMPUTE:
1077 ring->funcs = adev->gfx.compute_ring[0].funcs;
1078 ring->me = adev->gfx.compute_ring[0].me;
1079 ring->pipe = adev->gfx.compute_ring[0].pipe;
1080 break;
1081 case AMDGPU_RING_TYPE_SDMA:
1082 ring->funcs = adev->sdma.instance[0].ring.funcs;
1083 break;
1084 default:
1085 BUG();
1086 }
1087
1088 r = amdgpu_ring_init(adev, ring, 1024, NULL, 0,
1089 AMDGPU_RING_PRIO_DEFAULT, NULL);
1090 if (r)
1091 goto clean_up_memory;
1092
1093 amdgpu_mes_ring_to_queue_props(adev, ring, &qprops);
1094
1095 dma_fence_wait(gang->process->vm->last_update, false);
1096 dma_fence_wait(ctx_data->meta_data_va->last_pt_update, false);
1097 amdgpu_mes_unlock(&adev->mes);
1098
1099 r = amdgpu_mes_add_hw_queue(adev, gang_id, &qprops, &queue_id);
1100 if (r)
1101 goto clean_up_ring;
1102
1103 ring->hw_queue_id = queue_id;
1104 ring->doorbell_index = qprops.doorbell_off;
1105
1106 if (queue_type == AMDGPU_RING_TYPE_GFX)
1107 sprintf(ring->name, "gfx_%d.%d.%d", pasid, gang_id, queue_id);
1108 else if (queue_type == AMDGPU_RING_TYPE_COMPUTE)
1109 sprintf(ring->name, "compute_%d.%d.%d", pasid, gang_id,
1110 queue_id);
1111 else if (queue_type == AMDGPU_RING_TYPE_SDMA)
1112 sprintf(ring->name, "sdma_%d.%d.%d", pasid, gang_id,
1113 queue_id);
1114 else
1115 BUG();
1116
1117 *out = ring;
1118 return 0;
1119
1120clean_up_ring:
1121 amdgpu_ring_fini(ring);
1122clean_up_memory:
1123 kfree(ring);
1124 amdgpu_mes_unlock(&adev->mes);
1125 return r;
1126}
1127
1128void amdgpu_mes_remove_ring(struct amdgpu_device *adev,
1129 struct amdgpu_ring *ring)
1130{
1131 if (!ring)
1132 return;
1133
1134 amdgpu_mes_remove_hw_queue(adev, ring->hw_queue_id);
1135 del_timer_sync(&ring->fence_drv.fallback_timer);
1136 amdgpu_ring_fini(ring);
1137 kfree(ring);
1138}
1139
1140uint32_t amdgpu_mes_get_aggregated_doorbell_index(struct amdgpu_device *adev,
1141 enum amdgpu_mes_priority_level prio)
1142{
1143 return adev->mes.aggregated_doorbells[prio];
1144}
1145
1146int amdgpu_mes_ctx_alloc_meta_data(struct amdgpu_device *adev,
1147 struct amdgpu_mes_ctx_data *ctx_data)
1148{
1149 int r;
1150
1151 r = amdgpu_bo_create_kernel(adev,
1152 sizeof(struct amdgpu_mes_ctx_meta_data),
1153 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
1154 &ctx_data->meta_data_obj,
1155 &ctx_data->meta_data_mc_addr,
1156 &ctx_data->meta_data_ptr);
1157 if (r) {
1158 dev_warn(adev->dev, "(%d) create CTX bo failed\n", r);
1159 return r;
1160 }
1161
1162 if (!ctx_data->meta_data_obj)
1163 return -ENOMEM;
1164
1165 memset(ctx_data->meta_data_ptr, 0,
1166 sizeof(struct amdgpu_mes_ctx_meta_data));
1167
1168 return 0;
1169}
1170
1171void amdgpu_mes_ctx_free_meta_data(struct amdgpu_mes_ctx_data *ctx_data)
1172{
1173 if (ctx_data->meta_data_obj)
1174 amdgpu_bo_free_kernel(&ctx_data->meta_data_obj,
1175 &ctx_data->meta_data_mc_addr,
1176 &ctx_data->meta_data_ptr);
1177}
1178
1179int amdgpu_mes_ctx_map_meta_data(struct amdgpu_device *adev,
1180 struct amdgpu_vm *vm,
1181 struct amdgpu_mes_ctx_data *ctx_data)
1182{
1183 struct amdgpu_bo_va *bo_va;
1184 struct amdgpu_sync sync;
1185 struct drm_exec exec;
1186 int r;
1187
1188 amdgpu_sync_create(&sync);
1189
1190 drm_exec_init(&exec, 0, 0);
1191 drm_exec_until_all_locked(&exec) {
1192 r = drm_exec_lock_obj(&exec,
1193 &ctx_data->meta_data_obj->tbo.base);
1194 drm_exec_retry_on_contention(&exec);
1195 if (unlikely(r))
1196 goto error_fini_exec;
1197
1198 r = amdgpu_vm_lock_pd(vm, &exec, 0);
1199 drm_exec_retry_on_contention(&exec);
1200 if (unlikely(r))
1201 goto error_fini_exec;
1202 }
1203
1204 bo_va = amdgpu_vm_bo_add(adev, vm, ctx_data->meta_data_obj);
1205 if (!bo_va) {
1206 DRM_ERROR("failed to create bo_va for meta data BO\n");
1207 r = -ENOMEM;
1208 goto error_fini_exec;
1209 }
1210
1211 r = amdgpu_vm_bo_map(adev, bo_va, ctx_data->meta_data_gpu_addr, 0,
1212 sizeof(struct amdgpu_mes_ctx_meta_data),
1213 AMDGPU_PTE_READABLE | AMDGPU_PTE_WRITEABLE |
1214 AMDGPU_PTE_EXECUTABLE);
1215
1216 if (r) {
1217 DRM_ERROR("failed to do bo_map on meta data, err=%d\n", r);
1218 goto error_del_bo_va;
1219 }
1220
1221 r = amdgpu_vm_bo_update(adev, bo_va, false);
1222 if (r) {
1223 DRM_ERROR("failed to do vm_bo_update on meta data\n");
1224 goto error_del_bo_va;
1225 }
1226 amdgpu_sync_fence(&sync, bo_va->last_pt_update);
1227
1228 r = amdgpu_vm_update_pdes(adev, vm, false);
1229 if (r) {
1230 DRM_ERROR("failed to update pdes on meta data\n");
1231 goto error_del_bo_va;
1232 }
1233 amdgpu_sync_fence(&sync, vm->last_update);
1234
1235 amdgpu_sync_wait(&sync, false);
1236 drm_exec_fini(&exec);
1237
1238 amdgpu_sync_free(&sync);
1239 ctx_data->meta_data_va = bo_va;
1240 return 0;
1241
1242error_del_bo_va:
1243 amdgpu_vm_bo_del(adev, bo_va);
1244
1245error_fini_exec:
1246 drm_exec_fini(&exec);
1247 amdgpu_sync_free(&sync);
1248 return r;
1249}
1250
1251int amdgpu_mes_ctx_unmap_meta_data(struct amdgpu_device *adev,
1252 struct amdgpu_mes_ctx_data *ctx_data)
1253{
1254 struct amdgpu_bo_va *bo_va = ctx_data->meta_data_va;
1255 struct amdgpu_bo *bo = ctx_data->meta_data_obj;
1256 struct amdgpu_vm *vm = bo_va->base.vm;
1257 struct dma_fence *fence;
1258 struct drm_exec exec;
1259 long r;
1260
1261 drm_exec_init(&exec, 0, 0);
1262 drm_exec_until_all_locked(&exec) {
1263 r = drm_exec_lock_obj(&exec,
1264 &ctx_data->meta_data_obj->tbo.base);
1265 drm_exec_retry_on_contention(&exec);
1266 if (unlikely(r))
1267 goto out_unlock;
1268
1269 r = amdgpu_vm_lock_pd(vm, &exec, 0);
1270 drm_exec_retry_on_contention(&exec);
1271 if (unlikely(r))
1272 goto out_unlock;
1273 }
1274
1275 amdgpu_vm_bo_del(adev, bo_va);
1276 if (!amdgpu_vm_ready(vm))
1277 goto out_unlock;
1278
1279 r = dma_resv_get_singleton(bo->tbo.base.resv, DMA_RESV_USAGE_BOOKKEEP,
1280 &fence);
1281 if (r)
1282 goto out_unlock;
1283 if (fence) {
1284 amdgpu_bo_fence(bo, fence, true);
1285 fence = NULL;
1286 }
1287
1288 r = amdgpu_vm_clear_freed(adev, vm, &fence);
1289 if (r || !fence)
1290 goto out_unlock;
1291
1292 dma_fence_wait(fence, false);
1293 amdgpu_bo_fence(bo, fence, true);
1294 dma_fence_put(fence);
1295
1296out_unlock:
1297 if (unlikely(r < 0))
1298 dev_err(adev->dev, "failed to clear page tables (%ld)\n", r);
1299 drm_exec_fini(&exec);
1300
1301 return r;
1302}
1303
1304static int amdgpu_mes_test_create_gang_and_queues(struct amdgpu_device *adev,
1305 int pasid, int *gang_id,
1306 int queue_type, int num_queue,
1307 struct amdgpu_ring **added_rings,
1308 struct amdgpu_mes_ctx_data *ctx_data)
1309{
1310 struct amdgpu_ring *ring;
1311 struct amdgpu_mes_gang_properties gprops = {0};
1312 int r, j;
1313
1314 /* create a gang for the process */
1315 gprops.priority = AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
1316 gprops.gang_quantum = adev->mes.default_gang_quantum;
1317 gprops.inprocess_gang_priority = AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
1318 gprops.priority_level = AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
1319 gprops.global_priority_level = AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
1320
1321 r = amdgpu_mes_add_gang(adev, pasid, &gprops, gang_id);
1322 if (r) {
1323 DRM_ERROR("failed to add gang\n");
1324 return r;
1325 }
1326
1327 /* create queues for the gang */
1328 for (j = 0; j < num_queue; j++) {
1329 r = amdgpu_mes_add_ring(adev, *gang_id, queue_type, j,
1330 ctx_data, &ring);
1331 if (r) {
1332 DRM_ERROR("failed to add ring\n");
1333 break;
1334 }
1335
1336 DRM_INFO("ring %s was added\n", ring->name);
1337 added_rings[j] = ring;
1338 }
1339
1340 return 0;
1341}
1342
1343static int amdgpu_mes_test_queues(struct amdgpu_ring **added_rings)
1344{
1345 struct amdgpu_ring *ring;
1346 int i, r;
1347
1348 for (i = 0; i < AMDGPU_MES_CTX_MAX_RINGS; i++) {
1349 ring = added_rings[i];
1350 if (!ring)
1351 continue;
1352
1353 r = amdgpu_ring_test_helper(ring);
1354 if (r)
1355 return r;
1356
1357 r = amdgpu_ring_test_ib(ring, 1000 * 10);
1358 if (r) {
1359 DRM_DEV_ERROR(ring->adev->dev,
1360 "ring %s ib test failed (%d)\n",
1361 ring->name, r);
1362 return r;
1363 } else
1364 DRM_INFO("ring %s ib test pass\n", ring->name);
1365 }
1366
1367 return 0;
1368}
1369
1370int amdgpu_mes_self_test(struct amdgpu_device *adev)
1371{
1372 struct amdgpu_vm *vm = NULL;
1373 struct amdgpu_mes_ctx_data ctx_data = {0};
1374 struct amdgpu_ring *added_rings[AMDGPU_MES_CTX_MAX_RINGS] = { NULL };
1375 int gang_ids[3] = {0};
1376 int queue_types[][2] = { { AMDGPU_RING_TYPE_GFX, 1 },
1377 { AMDGPU_RING_TYPE_COMPUTE, 1 },
1378 { AMDGPU_RING_TYPE_SDMA, 1} };
1379 int i, r, pasid, k = 0;
1380
1381 pasid = amdgpu_pasid_alloc(16);
1382 if (pasid < 0) {
1383 dev_warn(adev->dev, "No more PASIDs available!");
1384 pasid = 0;
1385 }
1386
1387 vm = kzalloc(sizeof(*vm), GFP_KERNEL);
1388 if (!vm) {
1389 r = -ENOMEM;
1390 goto error_pasid;
1391 }
1392
1393 r = amdgpu_vm_init(adev, vm, -1);
1394 if (r) {
1395 DRM_ERROR("failed to initialize vm\n");
1396 goto error_pasid;
1397 }
1398
1399 r = amdgpu_mes_ctx_alloc_meta_data(adev, &ctx_data);
1400 if (r) {
1401 DRM_ERROR("failed to alloc ctx meta data\n");
1402 goto error_fini;
1403 }
1404
1405 ctx_data.meta_data_gpu_addr = AMDGPU_VA_RESERVED_BOTTOM;
1406 r = amdgpu_mes_ctx_map_meta_data(adev, vm, &ctx_data);
1407 if (r) {
1408 DRM_ERROR("failed to map ctx meta data\n");
1409 goto error_vm;
1410 }
1411
1412 r = amdgpu_mes_create_process(adev, pasid, vm);
1413 if (r) {
1414 DRM_ERROR("failed to create MES process\n");
1415 goto error_vm;
1416 }
1417
1418 for (i = 0; i < ARRAY_SIZE(queue_types); i++) {
1419 /* On GFX v10.3, fw hasn't supported to map sdma queue. */
1420 if (amdgpu_ip_version(adev, GC_HWIP, 0) >=
1421 IP_VERSION(10, 3, 0) &&
1422 amdgpu_ip_version(adev, GC_HWIP, 0) <
1423 IP_VERSION(11, 0, 0) &&
1424 queue_types[i][0] == AMDGPU_RING_TYPE_SDMA)
1425 continue;
1426
1427 r = amdgpu_mes_test_create_gang_and_queues(adev, pasid,
1428 &gang_ids[i],
1429 queue_types[i][0],
1430 queue_types[i][1],
1431 &added_rings[k],
1432 &ctx_data);
1433 if (r)
1434 goto error_queues;
1435
1436 k += queue_types[i][1];
1437 }
1438
1439 /* start ring test and ib test for MES queues */
1440 amdgpu_mes_test_queues(added_rings);
1441
1442error_queues:
1443 /* remove all queues */
1444 for (i = 0; i < ARRAY_SIZE(added_rings); i++) {
1445 if (!added_rings[i])
1446 continue;
1447 amdgpu_mes_remove_ring(adev, added_rings[i]);
1448 }
1449
1450 for (i = 0; i < ARRAY_SIZE(gang_ids); i++) {
1451 if (!gang_ids[i])
1452 continue;
1453 amdgpu_mes_remove_gang(adev, gang_ids[i]);
1454 }
1455
1456 amdgpu_mes_destroy_process(adev, pasid);
1457
1458error_vm:
1459 amdgpu_mes_ctx_unmap_meta_data(adev, &ctx_data);
1460
1461error_fini:
1462 amdgpu_vm_fini(adev, vm);
1463
1464error_pasid:
1465 if (pasid)
1466 amdgpu_pasid_free(pasid);
1467
1468 amdgpu_mes_ctx_free_meta_data(&ctx_data);
1469 kfree(vm);
1470 return 0;
1471}
1472
1473int amdgpu_mes_init_microcode(struct amdgpu_device *adev, int pipe)
1474{
1475 const struct mes_firmware_header_v1_0 *mes_hdr;
1476 struct amdgpu_firmware_info *info;
1477 char ucode_prefix[30];
1478 char fw_name[40];
1479 bool need_retry = false;
1480 int r;
1481
1482 amdgpu_ucode_ip_version_decode(adev, GC_HWIP, ucode_prefix,
1483 sizeof(ucode_prefix));
1484 if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(11, 0, 0)) {
1485 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes%s.bin",
1486 ucode_prefix,
1487 pipe == AMDGPU_MES_SCHED_PIPE ? "_2" : "1");
1488 need_retry = true;
1489 } else {
1490 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes%s.bin",
1491 ucode_prefix,
1492 pipe == AMDGPU_MES_SCHED_PIPE ? "" : "1");
1493 }
1494
1495 r = amdgpu_ucode_request(adev, &adev->mes.fw[pipe], fw_name);
1496 if (r && need_retry && pipe == AMDGPU_MES_SCHED_PIPE) {
1497 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes.bin",
1498 ucode_prefix);
1499 DRM_INFO("try to fall back to %s\n", fw_name);
1500 r = amdgpu_ucode_request(adev, &adev->mes.fw[pipe],
1501 fw_name);
1502 }
1503
1504 if (r)
1505 goto out;
1506
1507 mes_hdr = (const struct mes_firmware_header_v1_0 *)
1508 adev->mes.fw[pipe]->data;
1509 adev->mes.uc_start_addr[pipe] =
1510 le32_to_cpu(mes_hdr->mes_uc_start_addr_lo) |
1511 ((uint64_t)(le32_to_cpu(mes_hdr->mes_uc_start_addr_hi)) << 32);
1512 adev->mes.data_start_addr[pipe] =
1513 le32_to_cpu(mes_hdr->mes_data_start_addr_lo) |
1514 ((uint64_t)(le32_to_cpu(mes_hdr->mes_data_start_addr_hi)) << 32);
1515
1516 if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
1517 int ucode, ucode_data;
1518
1519 if (pipe == AMDGPU_MES_SCHED_PIPE) {
1520 ucode = AMDGPU_UCODE_ID_CP_MES;
1521 ucode_data = AMDGPU_UCODE_ID_CP_MES_DATA;
1522 } else {
1523 ucode = AMDGPU_UCODE_ID_CP_MES1;
1524 ucode_data = AMDGPU_UCODE_ID_CP_MES1_DATA;
1525 }
1526
1527 info = &adev->firmware.ucode[ucode];
1528 info->ucode_id = ucode;
1529 info->fw = adev->mes.fw[pipe];
1530 adev->firmware.fw_size +=
1531 ALIGN(le32_to_cpu(mes_hdr->mes_ucode_size_bytes),
1532 PAGE_SIZE);
1533
1534 info = &adev->firmware.ucode[ucode_data];
1535 info->ucode_id = ucode_data;
1536 info->fw = adev->mes.fw[pipe];
1537 adev->firmware.fw_size +=
1538 ALIGN(le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes),
1539 PAGE_SIZE);
1540 }
1541
1542 return 0;
1543out:
1544 amdgpu_ucode_release(&adev->mes.fw[pipe]);
1545 return r;
1546}
1547
1548#if defined(CONFIG_DEBUG_FS)
1549
1550static int amdgpu_debugfs_mes_event_log_show(struct seq_file *m, void *unused)
1551{
1552 struct amdgpu_device *adev = m->private;
1553 uint32_t *mem = (uint32_t *)(adev->mes.event_log_cpu_addr);
1554
1555 seq_hex_dump(m, "", DUMP_PREFIX_OFFSET, 32, 4,
1556 mem, AMDGPU_MES_LOG_BUFFER_SIZE, false);
1557
1558 return 0;
1559}
1560
1561DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_mes_event_log);
1562
1563#endif
1564
1565void amdgpu_debugfs_mes_event_log_init(struct amdgpu_device *adev)
1566{
1567
1568#if defined(CONFIG_DEBUG_FS)
1569 struct drm_minor *minor = adev_to_drm(adev)->primary;
1570 struct dentry *root = minor->debugfs_root;
1571 if (adev->enable_mes && amdgpu_mes_log_enable)
1572 debugfs_create_file("amdgpu_mes_event_log", 0444, root,
1573 adev, &amdgpu_debugfs_mes_event_log_fops);
1574
1575#endif
1576}
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
32int 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
39int 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
51void 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
58unsigned 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
68static 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
97static 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
112static 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
153int 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
233error:
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);
237error_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
246void 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
259static 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
266int 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
332clean_up_pasid:
333 idr_remove(&adev->mes.pasid_idr, pasid);
334 amdgpu_mes_unlock(&adev->mes);
335clean_up_ctx:
336 amdgpu_bo_free_kernel(&process->proc_ctx_bo,
337 &process->proc_ctx_gpu_addr,
338 &process->proc_ctx_cpu_ptr);
339clean_up_memory:
340 kfree(process->doorbell_bitmap);
341 kfree(process);
342 return r;
343}
344
345void 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
412int 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
474clean_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);
479clean_up_mem:
480 kfree(gang);
481 return r;
482}
483
484int 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
520int 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
549int 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
578static 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
602clean_up:
603 amdgpu_bo_free_kernel(&q->mqd_obj,
604 &q->mqd_gpu_addr,
605 &q->mqd_cpu_ptr);
606 return r;
607}
608
609static 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
633int 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
737clean_up_doorbell:
738 amdgpu_mes_queue_doorbell_free(adev, gang->process,
739 qprops->doorbell_off);
740clean_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);
744clean_up_mqd:
745 amdgpu_mes_unlock(&adev->mes);
746 amdgpu_mes_queue_free_mqd(queue);
747clean_up_memory:
748 kfree(queue);
749 return r;
750}
751
752int 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
802int 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
825uint32_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
845error:
846 return val;
847}
848
849int 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
869error:
870 return r;
871}
872
873int 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
896error:
897 return r;
898}
899
900int 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
921error:
922 return r;
923}
924
925static void
926amdgpu_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) \
945do { \
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
960int 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
980int 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
1069clean_up_ring:
1070 amdgpu_ring_fini(ring);
1071clean_up_memory:
1072 kfree(ring);
1073 amdgpu_mes_unlock(&adev->mes);
1074 return r;
1075}
1076
1077void 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
1088uint32_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
1094int 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
1114void 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
1122int 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
1188error:
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
1195int 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
1244out_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
1252static 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
1291static 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
1323int 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
1396error_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
1412error_vm:
1413 amdgpu_mes_ctx_unmap_meta_data(adev, &ctx_data);
1414
1415error_fini:
1416 amdgpu_vm_fini(adev, vm);
1417
1418error_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}