1 /*
2 * Copyright 2008 Jerome Glisse.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
22 * DEALINGS IN THE SOFTWARE.
23 *
24 * Authors:
25 * Jerome Glisse <glisse@freedesktop.org>
26 */
27
28 #include <linux/file.h>
29 #include <linux/pagemap.h>
30 #include <linux/sync_file.h>
31 #include <linux/dma-buf.h>
32
33 #include <drm/amdgpu_drm.h>
34 #include <drm/drm_syncobj.h>
35 #include <drm/ttm/ttm_tt.h>
36
37 #include "amdgpu_cs.h"
38 #include "amdgpu.h"
39 #include "amdgpu_trace.h"
40 #include "amdgpu_gmc.h"
41 #include "amdgpu_gem.h"
42 #include "amdgpu_ras.h"
43
amdgpu_cs_parser_init(struct amdgpu_cs_parser * p,struct amdgpu_device * adev,struct drm_file * filp,union drm_amdgpu_cs * cs)44 static int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p,
45 struct amdgpu_device *adev,
46 struct drm_file *filp,
47 union drm_amdgpu_cs *cs)
48 {
49 struct amdgpu_fpriv *fpriv = filp->driver_priv;
50
51 if (cs->in.num_chunks == 0)
52 return -EINVAL;
53
54 memset(p, 0, sizeof(*p));
55 p->adev = adev;
56 p->filp = filp;
57
58 p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
59 if (!p->ctx)
60 return -EINVAL;
61
62 if (atomic_read(&p->ctx->guilty)) {
63 amdgpu_ctx_put(p->ctx);
64 return -ECANCELED;
65 }
66
67 amdgpu_sync_create(&p->sync);
68 return 0;
69 }
70
amdgpu_cs_job_idx(struct amdgpu_cs_parser * p,struct drm_amdgpu_cs_chunk_ib * chunk_ib)71 static int amdgpu_cs_job_idx(struct amdgpu_cs_parser *p,
72 struct drm_amdgpu_cs_chunk_ib *chunk_ib)
73 {
74 struct drm_sched_entity *entity;
75 unsigned int i;
76 int r;
77
78 r = amdgpu_ctx_get_entity(p->ctx, chunk_ib->ip_type,
79 chunk_ib->ip_instance,
80 chunk_ib->ring, &entity);
81 if (r)
82 return r;
83
84 /*
85 * Abort if there is no run queue associated with this entity.
86 * Possibly because of disabled HW IP.
87 */
88 if (entity->rq == NULL)
89 return -EINVAL;
90
91 /* Check if we can add this IB to some existing job */
92 for (i = 0; i < p->gang_size; ++i)
93 if (p->entities[i] == entity)
94 return i;
95
96 /* If not increase the gang size if possible */
97 if (i == AMDGPU_CS_GANG_SIZE)
98 return -EINVAL;
99
100 p->entities[i] = entity;
101 p->gang_size = i + 1;
102 return i;
103 }
104
amdgpu_cs_p1_ib(struct amdgpu_cs_parser * p,struct drm_amdgpu_cs_chunk_ib * chunk_ib,unsigned int * num_ibs)105 static int amdgpu_cs_p1_ib(struct amdgpu_cs_parser *p,
106 struct drm_amdgpu_cs_chunk_ib *chunk_ib,
107 unsigned int *num_ibs)
108 {
109 int r;
110
111 r = amdgpu_cs_job_idx(p, chunk_ib);
112 if (r < 0)
113 return r;
114
115 ++(num_ibs[r]);
116 p->gang_leader_idx = r;
117 return 0;
118 }
119
amdgpu_cs_p1_user_fence(struct amdgpu_cs_parser * p,struct drm_amdgpu_cs_chunk_fence * data,uint32_t * offset)120 static int amdgpu_cs_p1_user_fence(struct amdgpu_cs_parser *p,
121 struct drm_amdgpu_cs_chunk_fence *data,
122 uint32_t *offset)
123 {
124 struct drm_gem_object *gobj;
125 struct amdgpu_bo *bo;
126 unsigned long size;
127 int r;
128
129 gobj = drm_gem_object_lookup(p->filp, data->handle);
130 if (gobj == NULL)
131 return -EINVAL;
132
133 bo = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
134 p->uf_entry.priority = 0;
135 p->uf_entry.tv.bo = &bo->tbo;
136 /* One for TTM and two for the CS job */
137 p->uf_entry.tv.num_shared = 3;
138
139 drm_gem_object_put(gobj);
140
141 size = amdgpu_bo_size(bo);
142 if (size != PAGE_SIZE || (data->offset + 8) > size) {
143 r = -EINVAL;
144 goto error_unref;
145 }
146
147 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
148 r = -EINVAL;
149 goto error_unref;
150 }
151
152 *offset = data->offset;
153
154 return 0;
155
156 error_unref:
157 amdgpu_bo_unref(&bo);
158 return r;
159 }
160
amdgpu_cs_p1_bo_handles(struct amdgpu_cs_parser * p,struct drm_amdgpu_bo_list_in * data)161 static int amdgpu_cs_p1_bo_handles(struct amdgpu_cs_parser *p,
162 struct drm_amdgpu_bo_list_in *data)
163 {
164 struct drm_amdgpu_bo_list_entry *info;
165 int r;
166
167 r = amdgpu_bo_create_list_entry_array(data, &info);
168 if (r)
169 return r;
170
171 r = amdgpu_bo_list_create(p->adev, p->filp, info, data->bo_number,
172 &p->bo_list);
173 if (r)
174 goto error_free;
175
176 kvfree(info);
177 return 0;
178
179 error_free:
180 kvfree(info);
181
182 return r;
183 }
184
185 /* Copy the data from userspace and go over it the first time */
amdgpu_cs_pass1(struct amdgpu_cs_parser * p,union drm_amdgpu_cs * cs)186 static int amdgpu_cs_pass1(struct amdgpu_cs_parser *p,
187 union drm_amdgpu_cs *cs)
188 {
189 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
190 unsigned int num_ibs[AMDGPU_CS_GANG_SIZE] = { };
191 struct amdgpu_vm *vm = &fpriv->vm;
192 uint64_t *chunk_array_user;
193 uint64_t *chunk_array;
194 uint32_t uf_offset = 0;
195 unsigned int size;
196 int ret;
197 int i;
198
199 chunk_array = kvmalloc_array(cs->in.num_chunks, sizeof(uint64_t),
200 GFP_KERNEL);
201 if (!chunk_array)
202 return -ENOMEM;
203
204 /* get chunks */
205 chunk_array_user = u64_to_user_ptr(cs->in.chunks);
206 if (copy_from_user(chunk_array, chunk_array_user,
207 sizeof(uint64_t)*cs->in.num_chunks)) {
208 ret = -EFAULT;
209 goto free_chunk;
210 }
211
212 p->nchunks = cs->in.num_chunks;
213 p->chunks = kvmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
214 GFP_KERNEL);
215 if (!p->chunks) {
216 ret = -ENOMEM;
217 goto free_chunk;
218 }
219
220 for (i = 0; i < p->nchunks; i++) {
221 struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
222 struct drm_amdgpu_cs_chunk user_chunk;
223 uint32_t __user *cdata;
224
225 chunk_ptr = u64_to_user_ptr(chunk_array[i]);
226 if (copy_from_user(&user_chunk, chunk_ptr,
227 sizeof(struct drm_amdgpu_cs_chunk))) {
228 ret = -EFAULT;
229 i--;
230 goto free_partial_kdata;
231 }
232 p->chunks[i].chunk_id = user_chunk.chunk_id;
233 p->chunks[i].length_dw = user_chunk.length_dw;
234
235 size = p->chunks[i].length_dw;
236 cdata = u64_to_user_ptr(user_chunk.chunk_data);
237
238 p->chunks[i].kdata = kvmalloc_array(size, sizeof(uint32_t),
239 GFP_KERNEL);
240 if (p->chunks[i].kdata == NULL) {
241 ret = -ENOMEM;
242 i--;
243 goto free_partial_kdata;
244 }
245 size *= sizeof(uint32_t);
246 if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
247 ret = -EFAULT;
248 goto free_partial_kdata;
249 }
250
251 /* Assume the worst on the following checks */
252 ret = -EINVAL;
253 switch (p->chunks[i].chunk_id) {
254 case AMDGPU_CHUNK_ID_IB:
255 if (size < sizeof(struct drm_amdgpu_cs_chunk_ib))
256 goto free_partial_kdata;
257
258 ret = amdgpu_cs_p1_ib(p, p->chunks[i].kdata, num_ibs);
259 if (ret)
260 goto free_partial_kdata;
261 break;
262
263 case AMDGPU_CHUNK_ID_FENCE:
264 if (size < sizeof(struct drm_amdgpu_cs_chunk_fence))
265 goto free_partial_kdata;
266
267 ret = amdgpu_cs_p1_user_fence(p, p->chunks[i].kdata,
268 &uf_offset);
269 if (ret)
270 goto free_partial_kdata;
271 break;
272
273 case AMDGPU_CHUNK_ID_BO_HANDLES:
274 if (size < sizeof(struct drm_amdgpu_bo_list_in))
275 goto free_partial_kdata;
276
277 ret = amdgpu_cs_p1_bo_handles(p, p->chunks[i].kdata);
278 if (ret)
279 goto free_partial_kdata;
280 break;
281
282 case AMDGPU_CHUNK_ID_DEPENDENCIES:
283 case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
284 case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
285 case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
286 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
287 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
288 break;
289
290 default:
291 goto free_partial_kdata;
292 }
293 }
294
295 if (!p->gang_size) {
296 ret = -EINVAL;
297 goto free_partial_kdata;
298 }
299
300 for (i = 0; i < p->gang_size; ++i) {
301 ret = amdgpu_job_alloc(p->adev, vm, p->entities[i], vm,
302 num_ibs[i], &p->jobs[i]);
303 if (ret)
304 goto free_all_kdata;
305 }
306 p->gang_leader = p->jobs[p->gang_leader_idx];
307
308 if (p->ctx->vram_lost_counter != p->gang_leader->vram_lost_counter) {
309 ret = -ECANCELED;
310 goto free_all_kdata;
311 }
312
313 if (p->uf_entry.tv.bo)
314 p->gang_leader->uf_addr = uf_offset;
315 kvfree(chunk_array);
316
317 /* Use this opportunity to fill in task info for the vm */
318 amdgpu_vm_set_task_info(vm);
319
320 return 0;
321
322 free_all_kdata:
323 i = p->nchunks - 1;
324 free_partial_kdata:
325 for (; i >= 0; i--)
326 kvfree(p->chunks[i].kdata);
327 kvfree(p->chunks);
328 p->chunks = NULL;
329 p->nchunks = 0;
330 free_chunk:
331 kvfree(chunk_array);
332
333 return ret;
334 }
335
amdgpu_cs_p2_ib(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk,unsigned int * ce_preempt,unsigned int * de_preempt)336 static int amdgpu_cs_p2_ib(struct amdgpu_cs_parser *p,
337 struct amdgpu_cs_chunk *chunk,
338 unsigned int *ce_preempt,
339 unsigned int *de_preempt)
340 {
341 struct drm_amdgpu_cs_chunk_ib *chunk_ib = chunk->kdata;
342 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
343 struct amdgpu_vm *vm = &fpriv->vm;
344 struct amdgpu_ring *ring;
345 struct amdgpu_job *job;
346 struct amdgpu_ib *ib;
347 int r;
348
349 r = amdgpu_cs_job_idx(p, chunk_ib);
350 if (r < 0)
351 return r;
352
353 job = p->jobs[r];
354 ring = amdgpu_job_ring(job);
355 ib = &job->ibs[job->num_ibs++];
356
357 /* MM engine doesn't support user fences */
358 if (p->uf_entry.tv.bo && ring->funcs->no_user_fence)
359 return -EINVAL;
360
361 if (chunk_ib->ip_type == AMDGPU_HW_IP_GFX &&
362 chunk_ib->flags & AMDGPU_IB_FLAG_PREEMPT) {
363 if (chunk_ib->flags & AMDGPU_IB_FLAG_CE)
364 (*ce_preempt)++;
365 else
366 (*de_preempt)++;
367
368 /* Each GFX command submit allows only 1 IB max
369 * preemptible for CE & DE */
370 if (*ce_preempt > 1 || *de_preempt > 1)
371 return -EINVAL;
372 }
373
374 if (chunk_ib->flags & AMDGPU_IB_FLAG_PREAMBLE)
375 job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT;
376
377 r = amdgpu_ib_get(p->adev, vm, ring->funcs->parse_cs ?
378 chunk_ib->ib_bytes : 0,
379 AMDGPU_IB_POOL_DELAYED, ib);
380 if (r) {
381 DRM_ERROR("Failed to get ib !\n");
382 return r;
383 }
384
385 ib->gpu_addr = chunk_ib->va_start;
386 ib->length_dw = chunk_ib->ib_bytes / 4;
387 ib->flags = chunk_ib->flags;
388 return 0;
389 }
390
amdgpu_cs_p2_dependencies(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)391 static int amdgpu_cs_p2_dependencies(struct amdgpu_cs_parser *p,
392 struct amdgpu_cs_chunk *chunk)
393 {
394 struct drm_amdgpu_cs_chunk_dep *deps = chunk->kdata;
395 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
396 unsigned num_deps;
397 int i, r;
398
399 num_deps = chunk->length_dw * 4 /
400 sizeof(struct drm_amdgpu_cs_chunk_dep);
401
402 for (i = 0; i < num_deps; ++i) {
403 struct amdgpu_ctx *ctx;
404 struct drm_sched_entity *entity;
405 struct dma_fence *fence;
406
407 ctx = amdgpu_ctx_get(fpriv, deps[i].ctx_id);
408 if (ctx == NULL)
409 return -EINVAL;
410
411 r = amdgpu_ctx_get_entity(ctx, deps[i].ip_type,
412 deps[i].ip_instance,
413 deps[i].ring, &entity);
414 if (r) {
415 amdgpu_ctx_put(ctx);
416 return r;
417 }
418
419 fence = amdgpu_ctx_get_fence(ctx, entity, deps[i].handle);
420 amdgpu_ctx_put(ctx);
421
422 if (IS_ERR(fence))
423 return PTR_ERR(fence);
424 else if (!fence)
425 continue;
426
427 if (chunk->chunk_id == AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES) {
428 struct drm_sched_fence *s_fence;
429 struct dma_fence *old = fence;
430
431 s_fence = to_drm_sched_fence(fence);
432 fence = dma_fence_get(&s_fence->scheduled);
433 dma_fence_put(old);
434 }
435
436 r = amdgpu_sync_fence(&p->sync, fence);
437 dma_fence_put(fence);
438 if (r)
439 return r;
440 }
441 return 0;
442 }
443
amdgpu_syncobj_lookup_and_add(struct amdgpu_cs_parser * p,uint32_t handle,u64 point,u64 flags)444 static int amdgpu_syncobj_lookup_and_add(struct amdgpu_cs_parser *p,
445 uint32_t handle, u64 point,
446 u64 flags)
447 {
448 struct dma_fence *fence;
449 int r;
450
451 r = drm_syncobj_find_fence(p->filp, handle, point, flags, &fence);
452 if (r) {
453 DRM_ERROR("syncobj %u failed to find fence @ %llu (%d)!\n",
454 handle, point, r);
455 return r;
456 }
457
458 r = amdgpu_sync_fence(&p->sync, fence);
459 dma_fence_put(fence);
460 return r;
461 }
462
amdgpu_cs_p2_syncobj_in(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)463 static int amdgpu_cs_p2_syncobj_in(struct amdgpu_cs_parser *p,
464 struct amdgpu_cs_chunk *chunk)
465 {
466 struct drm_amdgpu_cs_chunk_sem *deps = chunk->kdata;
467 unsigned num_deps;
468 int i, r;
469
470 num_deps = chunk->length_dw * 4 /
471 sizeof(struct drm_amdgpu_cs_chunk_sem);
472 for (i = 0; i < num_deps; ++i) {
473 r = amdgpu_syncobj_lookup_and_add(p, deps[i].handle, 0, 0);
474 if (r)
475 return r;
476 }
477
478 return 0;
479 }
480
amdgpu_cs_p2_syncobj_timeline_wait(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)481 static int amdgpu_cs_p2_syncobj_timeline_wait(struct amdgpu_cs_parser *p,
482 struct amdgpu_cs_chunk *chunk)
483 {
484 struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps = chunk->kdata;
485 unsigned num_deps;
486 int i, r;
487
488 num_deps = chunk->length_dw * 4 /
489 sizeof(struct drm_amdgpu_cs_chunk_syncobj);
490 for (i = 0; i < num_deps; ++i) {
491 r = amdgpu_syncobj_lookup_and_add(p, syncobj_deps[i].handle,
492 syncobj_deps[i].point,
493 syncobj_deps[i].flags);
494 if (r)
495 return r;
496 }
497
498 return 0;
499 }
500
amdgpu_cs_p2_syncobj_out(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)501 static int amdgpu_cs_p2_syncobj_out(struct amdgpu_cs_parser *p,
502 struct amdgpu_cs_chunk *chunk)
503 {
504 struct drm_amdgpu_cs_chunk_sem *deps = chunk->kdata;
505 unsigned num_deps;
506 int i;
507
508 num_deps = chunk->length_dw * 4 /
509 sizeof(struct drm_amdgpu_cs_chunk_sem);
510
511 if (p->post_deps)
512 return -EINVAL;
513
514 p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
515 GFP_KERNEL);
516 p->num_post_deps = 0;
517
518 if (!p->post_deps)
519 return -ENOMEM;
520
521
522 for (i = 0; i < num_deps; ++i) {
523 p->post_deps[i].syncobj =
524 drm_syncobj_find(p->filp, deps[i].handle);
525 if (!p->post_deps[i].syncobj)
526 return -EINVAL;
527 p->post_deps[i].chain = NULL;
528 p->post_deps[i].point = 0;
529 p->num_post_deps++;
530 }
531
532 return 0;
533 }
534
amdgpu_cs_p2_syncobj_timeline_signal(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)535 static int amdgpu_cs_p2_syncobj_timeline_signal(struct amdgpu_cs_parser *p,
536 struct amdgpu_cs_chunk *chunk)
537 {
538 struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps = chunk->kdata;
539 unsigned num_deps;
540 int i;
541
542 num_deps = chunk->length_dw * 4 /
543 sizeof(struct drm_amdgpu_cs_chunk_syncobj);
544
545 if (p->post_deps)
546 return -EINVAL;
547
548 p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
549 GFP_KERNEL);
550 p->num_post_deps = 0;
551
552 if (!p->post_deps)
553 return -ENOMEM;
554
555 for (i = 0; i < num_deps; ++i) {
556 struct amdgpu_cs_post_dep *dep = &p->post_deps[i];
557
558 dep->chain = NULL;
559 if (syncobj_deps[i].point) {
560 dep->chain = dma_fence_chain_alloc();
561 if (!dep->chain)
562 return -ENOMEM;
563 }
564
565 dep->syncobj = drm_syncobj_find(p->filp,
566 syncobj_deps[i].handle);
567 if (!dep->syncobj) {
568 dma_fence_chain_free(dep->chain);
569 return -EINVAL;
570 }
571 dep->point = syncobj_deps[i].point;
572 p->num_post_deps++;
573 }
574
575 return 0;
576 }
577
amdgpu_cs_pass2(struct amdgpu_cs_parser * p)578 static int amdgpu_cs_pass2(struct amdgpu_cs_parser *p)
579 {
580 unsigned int ce_preempt = 0, de_preempt = 0;
581 int i, r;
582
583 for (i = 0; i < p->nchunks; ++i) {
584 struct amdgpu_cs_chunk *chunk;
585
586 chunk = &p->chunks[i];
587
588 switch (chunk->chunk_id) {
589 case AMDGPU_CHUNK_ID_IB:
590 r = amdgpu_cs_p2_ib(p, chunk, &ce_preempt, &de_preempt);
591 if (r)
592 return r;
593 break;
594 case AMDGPU_CHUNK_ID_DEPENDENCIES:
595 case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
596 r = amdgpu_cs_p2_dependencies(p, chunk);
597 if (r)
598 return r;
599 break;
600 case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
601 r = amdgpu_cs_p2_syncobj_in(p, chunk);
602 if (r)
603 return r;
604 break;
605 case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
606 r = amdgpu_cs_p2_syncobj_out(p, chunk);
607 if (r)
608 return r;
609 break;
610 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
611 r = amdgpu_cs_p2_syncobj_timeline_wait(p, chunk);
612 if (r)
613 return r;
614 break;
615 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
616 r = amdgpu_cs_p2_syncobj_timeline_signal(p, chunk);
617 if (r)
618 return r;
619 break;
620 }
621 }
622
623 return 0;
624 }
625
626 /* Convert microseconds to bytes. */
us_to_bytes(struct amdgpu_device * adev,s64 us)627 static u64 us_to_bytes(struct amdgpu_device *adev, s64 us)
628 {
629 if (us <= 0 || !adev->mm_stats.log2_max_MBps)
630 return 0;
631
632 /* Since accum_us is incremented by a million per second, just
633 * multiply it by the number of MB/s to get the number of bytes.
634 */
635 return us << adev->mm_stats.log2_max_MBps;
636 }
637
bytes_to_us(struct amdgpu_device * adev,u64 bytes)638 static s64 bytes_to_us(struct amdgpu_device *adev, u64 bytes)
639 {
640 if (!adev->mm_stats.log2_max_MBps)
641 return 0;
642
643 return bytes >> adev->mm_stats.log2_max_MBps;
644 }
645
646 /* Returns how many bytes TTM can move right now. If no bytes can be moved,
647 * it returns 0. If it returns non-zero, it's OK to move at least one buffer,
648 * which means it can go over the threshold once. If that happens, the driver
649 * will be in debt and no other buffer migrations can be done until that debt
650 * is repaid.
651 *
652 * This approach allows moving a buffer of any size (it's important to allow
653 * that).
654 *
655 * The currency is simply time in microseconds and it increases as the clock
656 * ticks. The accumulated microseconds (us) are converted to bytes and
657 * returned.
658 */
amdgpu_cs_get_threshold_for_moves(struct amdgpu_device * adev,u64 * max_bytes,u64 * max_vis_bytes)659 static void amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev,
660 u64 *max_bytes,
661 u64 *max_vis_bytes)
662 {
663 s64 time_us, increment_us;
664 u64 free_vram, total_vram, used_vram;
665 /* Allow a maximum of 200 accumulated ms. This is basically per-IB
666 * throttling.
667 *
668 * It means that in order to get full max MBps, at least 5 IBs per
669 * second must be submitted and not more than 200ms apart from each
670 * other.
671 */
672 const s64 us_upper_bound = 200000;
673
674 if (!adev->mm_stats.log2_max_MBps) {
675 *max_bytes = 0;
676 *max_vis_bytes = 0;
677 return;
678 }
679
680 total_vram = adev->gmc.real_vram_size - atomic64_read(&adev->vram_pin_size);
681 used_vram = ttm_resource_manager_usage(&adev->mman.vram_mgr.manager);
682 free_vram = used_vram >= total_vram ? 0 : total_vram - used_vram;
683
684 spin_lock(&adev->mm_stats.lock);
685
686 /* Increase the amount of accumulated us. */
687 time_us = ktime_to_us(ktime_get());
688 increment_us = time_us - adev->mm_stats.last_update_us;
689 adev->mm_stats.last_update_us = time_us;
690 adev->mm_stats.accum_us = min(adev->mm_stats.accum_us + increment_us,
691 us_upper_bound);
692
693 /* This prevents the short period of low performance when the VRAM
694 * usage is low and the driver is in debt or doesn't have enough
695 * accumulated us to fill VRAM quickly.
696 *
697 * The situation can occur in these cases:
698 * - a lot of VRAM is freed by userspace
699 * - the presence of a big buffer causes a lot of evictions
700 * (solution: split buffers into smaller ones)
701 *
702 * If 128 MB or 1/8th of VRAM is free, start filling it now by setting
703 * accum_us to a positive number.
704 */
705 if (free_vram >= 128 * 1024 * 1024 || free_vram >= total_vram / 8) {
706 s64 min_us;
707
708 /* Be more aggressive on dGPUs. Try to fill a portion of free
709 * VRAM now.
710 */
711 if (!(adev->flags & AMD_IS_APU))
712 min_us = bytes_to_us(adev, free_vram / 4);
713 else
714 min_us = 0; /* Reset accum_us on APUs. */
715
716 adev->mm_stats.accum_us = max(min_us, adev->mm_stats.accum_us);
717 }
718
719 /* This is set to 0 if the driver is in debt to disallow (optional)
720 * buffer moves.
721 */
722 *max_bytes = us_to_bytes(adev, adev->mm_stats.accum_us);
723
724 /* Do the same for visible VRAM if half of it is free */
725 if (!amdgpu_gmc_vram_full_visible(&adev->gmc)) {
726 u64 total_vis_vram = adev->gmc.visible_vram_size;
727 u64 used_vis_vram =
728 amdgpu_vram_mgr_vis_usage(&adev->mman.vram_mgr);
729
730 if (used_vis_vram < total_vis_vram) {
731 u64 free_vis_vram = total_vis_vram - used_vis_vram;
732 adev->mm_stats.accum_us_vis = min(adev->mm_stats.accum_us_vis +
733 increment_us, us_upper_bound);
734
735 if (free_vis_vram >= total_vis_vram / 2)
736 adev->mm_stats.accum_us_vis =
737 max(bytes_to_us(adev, free_vis_vram / 2),
738 adev->mm_stats.accum_us_vis);
739 }
740
741 *max_vis_bytes = us_to_bytes(adev, adev->mm_stats.accum_us_vis);
742 } else {
743 *max_vis_bytes = 0;
744 }
745
746 spin_unlock(&adev->mm_stats.lock);
747 }
748
749 /* Report how many bytes have really been moved for the last command
750 * submission. This can result in a debt that can stop buffer migrations
751 * temporarily.
752 */
amdgpu_cs_report_moved_bytes(struct amdgpu_device * adev,u64 num_bytes,u64 num_vis_bytes)753 void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev, u64 num_bytes,
754 u64 num_vis_bytes)
755 {
756 spin_lock(&adev->mm_stats.lock);
757 adev->mm_stats.accum_us -= bytes_to_us(adev, num_bytes);
758 adev->mm_stats.accum_us_vis -= bytes_to_us(adev, num_vis_bytes);
759 spin_unlock(&adev->mm_stats.lock);
760 }
761
amdgpu_cs_bo_validate(void * param,struct amdgpu_bo * bo)762 static int amdgpu_cs_bo_validate(void *param, struct amdgpu_bo *bo)
763 {
764 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
765 struct amdgpu_cs_parser *p = param;
766 struct ttm_operation_ctx ctx = {
767 .interruptible = true,
768 .no_wait_gpu = false,
769 .resv = bo->tbo.base.resv
770 };
771 uint32_t domain;
772 int r;
773
774 if (bo->tbo.pin_count)
775 return 0;
776
777 /* Don't move this buffer if we have depleted our allowance
778 * to move it. Don't move anything if the threshold is zero.
779 */
780 if (p->bytes_moved < p->bytes_moved_threshold &&
781 (!bo->tbo.base.dma_buf ||
782 list_empty(&bo->tbo.base.dma_buf->attachments))) {
783 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
784 (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)) {
785 /* And don't move a CPU_ACCESS_REQUIRED BO to limited
786 * visible VRAM if we've depleted our allowance to do
787 * that.
788 */
789 if (p->bytes_moved_vis < p->bytes_moved_vis_threshold)
790 domain = bo->preferred_domains;
791 else
792 domain = bo->allowed_domains;
793 } else {
794 domain = bo->preferred_domains;
795 }
796 } else {
797 domain = bo->allowed_domains;
798 }
799
800 retry:
801 amdgpu_bo_placement_from_domain(bo, domain);
802 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
803
804 p->bytes_moved += ctx.bytes_moved;
805 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
806 amdgpu_bo_in_cpu_visible_vram(bo))
807 p->bytes_moved_vis += ctx.bytes_moved;
808
809 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
810 domain = bo->allowed_domains;
811 goto retry;
812 }
813
814 return r;
815 }
816
amdgpu_cs_list_validate(struct amdgpu_cs_parser * p,struct list_head * validated)817 static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
818 struct list_head *validated)
819 {
820 struct ttm_operation_ctx ctx = { true, false };
821 struct amdgpu_bo_list_entry *lobj;
822 int r;
823
824 list_for_each_entry(lobj, validated, tv.head) {
825 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(lobj->tv.bo);
826 struct mm_struct *usermm;
827
828 usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
829 if (usermm && usermm != current->mm)
830 return -EPERM;
831
832 if (amdgpu_ttm_tt_is_userptr(bo->tbo.ttm) &&
833 lobj->user_invalidated && lobj->user_pages) {
834 amdgpu_bo_placement_from_domain(bo,
835 AMDGPU_GEM_DOMAIN_CPU);
836 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
837 if (r)
838 return r;
839
840 amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm,
841 lobj->user_pages);
842 }
843
844 r = amdgpu_cs_bo_validate(p, bo);
845 if (r)
846 return r;
847
848 kvfree(lobj->user_pages);
849 lobj->user_pages = NULL;
850 }
851 return 0;
852 }
853
amdgpu_cs_parser_bos(struct amdgpu_cs_parser * p,union drm_amdgpu_cs * cs)854 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
855 union drm_amdgpu_cs *cs)
856 {
857 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
858 struct amdgpu_vm *vm = &fpriv->vm;
859 struct amdgpu_bo_list_entry *e;
860 struct list_head duplicates;
861 unsigned int i;
862 int r;
863
864 INIT_LIST_HEAD(&p->validated);
865
866 /* p->bo_list could already be assigned if AMDGPU_CHUNK_ID_BO_HANDLES is present */
867 if (cs->in.bo_list_handle) {
868 if (p->bo_list)
869 return -EINVAL;
870
871 r = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle,
872 &p->bo_list);
873 if (r)
874 return r;
875 } else if (!p->bo_list) {
876 /* Create a empty bo_list when no handle is provided */
877 r = amdgpu_bo_list_create(p->adev, p->filp, NULL, 0,
878 &p->bo_list);
879 if (r)
880 return r;
881 }
882
883 mutex_lock(&p->bo_list->bo_list_mutex);
884
885 /* One for TTM and one for the CS job */
886 amdgpu_bo_list_for_each_entry(e, p->bo_list)
887 e->tv.num_shared = 2;
888
889 amdgpu_bo_list_get_list(p->bo_list, &p->validated);
890
891 INIT_LIST_HEAD(&duplicates);
892 amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
893
894 if (p->uf_entry.tv.bo && !ttm_to_amdgpu_bo(p->uf_entry.tv.bo)->parent)
895 list_add(&p->uf_entry.tv.head, &p->validated);
896
897 /* Get userptr backing pages. If pages are updated after registered
898 * in amdgpu_gem_userptr_ioctl(), amdgpu_cs_list_validate() will do
899 * amdgpu_ttm_backend_bind() to flush and invalidate new pages
900 */
901 amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
902 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
903 bool userpage_invalidated = false;
904 int i;
905
906 e->user_pages = kvmalloc_array(bo->tbo.ttm->num_pages,
907 sizeof(struct page *),
908 GFP_KERNEL | __GFP_ZERO);
909 if (!e->user_pages) {
910 DRM_ERROR("kvmalloc_array failure\n");
911 r = -ENOMEM;
912 goto out_free_user_pages;
913 }
914
915 r = amdgpu_ttm_tt_get_user_pages(bo, e->user_pages, &e->range);
916 if (r) {
917 kvfree(e->user_pages);
918 e->user_pages = NULL;
919 goto out_free_user_pages;
920 }
921
922 for (i = 0; i < bo->tbo.ttm->num_pages; i++) {
923 if (bo->tbo.ttm->pages[i] != e->user_pages[i]) {
924 userpage_invalidated = true;
925 break;
926 }
927 }
928 e->user_invalidated = userpage_invalidated;
929 }
930
931 r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
932 &duplicates);
933 if (unlikely(r != 0)) {
934 if (r != -ERESTARTSYS)
935 DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
936 goto out_free_user_pages;
937 }
938
939 amdgpu_bo_list_for_each_entry(e, p->bo_list) {
940 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
941
942 e->bo_va = amdgpu_vm_bo_find(vm, bo);
943 }
944
945 amdgpu_cs_get_threshold_for_moves(p->adev, &p->bytes_moved_threshold,
946 &p->bytes_moved_vis_threshold);
947 p->bytes_moved = 0;
948 p->bytes_moved_vis = 0;
949
950 r = amdgpu_vm_validate_pt_bos(p->adev, &fpriv->vm,
951 amdgpu_cs_bo_validate, p);
952 if (r) {
953 DRM_ERROR("amdgpu_vm_validate_pt_bos() failed.\n");
954 goto error_validate;
955 }
956
957 r = amdgpu_cs_list_validate(p, &duplicates);
958 if (r)
959 goto error_validate;
960
961 r = amdgpu_cs_list_validate(p, &p->validated);
962 if (r)
963 goto error_validate;
964
965 if (p->uf_entry.tv.bo) {
966 struct amdgpu_bo *uf = ttm_to_amdgpu_bo(p->uf_entry.tv.bo);
967
968 r = amdgpu_ttm_alloc_gart(&uf->tbo);
969 if (r)
970 goto error_validate;
971
972 p->gang_leader->uf_addr += amdgpu_bo_gpu_offset(uf);
973 }
974
975 amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved,
976 p->bytes_moved_vis);
977
978 for (i = 0; i < p->gang_size; ++i)
979 amdgpu_job_set_resources(p->jobs[i], p->bo_list->gds_obj,
980 p->bo_list->gws_obj,
981 p->bo_list->oa_obj);
982 return 0;
983
984 error_validate:
985 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
986
987 out_free_user_pages:
988 amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
989 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
990
991 if (!e->user_pages)
992 continue;
993 amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm, e->range);
994 kvfree(e->user_pages);
995 e->user_pages = NULL;
996 e->range = NULL;
997 }
998 mutex_unlock(&p->bo_list->bo_list_mutex);
999 return r;
1000 }
1001
trace_amdgpu_cs_ibs(struct amdgpu_cs_parser * p)1002 static void trace_amdgpu_cs_ibs(struct amdgpu_cs_parser *p)
1003 {
1004 int i, j;
1005
1006 if (!trace_amdgpu_cs_enabled())
1007 return;
1008
1009 for (i = 0; i < p->gang_size; ++i) {
1010 struct amdgpu_job *job = p->jobs[i];
1011
1012 for (j = 0; j < job->num_ibs; ++j)
1013 trace_amdgpu_cs(p, job, &job->ibs[j]);
1014 }
1015 }
1016
amdgpu_cs_patch_ibs(struct amdgpu_cs_parser * p,struct amdgpu_job * job)1017 static int amdgpu_cs_patch_ibs(struct amdgpu_cs_parser *p,
1018 struct amdgpu_job *job)
1019 {
1020 struct amdgpu_ring *ring = amdgpu_job_ring(job);
1021 unsigned int i;
1022 int r;
1023
1024 /* Only for UVD/VCE VM emulation */
1025 if (!ring->funcs->parse_cs && !ring->funcs->patch_cs_in_place)
1026 return 0;
1027
1028 for (i = 0; i < job->num_ibs; ++i) {
1029 struct amdgpu_ib *ib = &job->ibs[i];
1030 struct amdgpu_bo_va_mapping *m;
1031 struct amdgpu_bo *aobj;
1032 uint64_t va_start;
1033 uint8_t *kptr;
1034
1035 va_start = ib->gpu_addr & AMDGPU_GMC_HOLE_MASK;
1036 r = amdgpu_cs_find_mapping(p, va_start, &aobj, &m);
1037 if (r) {
1038 DRM_ERROR("IB va_start is invalid\n");
1039 return r;
1040 }
1041
1042 if ((va_start + ib->length_dw * 4) >
1043 (m->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
1044 DRM_ERROR("IB va_start+ib_bytes is invalid\n");
1045 return -EINVAL;
1046 }
1047
1048 /* the IB should be reserved at this point */
1049 r = amdgpu_bo_kmap(aobj, (void **)&kptr);
1050 if (r) {
1051 return r;
1052 }
1053
1054 kptr += va_start - (m->start * AMDGPU_GPU_PAGE_SIZE);
1055
1056 if (ring->funcs->parse_cs) {
1057 memcpy(ib->ptr, kptr, ib->length_dw * 4);
1058 amdgpu_bo_kunmap(aobj);
1059
1060 r = amdgpu_ring_parse_cs(ring, p, job, ib);
1061 if (r)
1062 return r;
1063 } else {
1064 ib->ptr = (uint32_t *)kptr;
1065 r = amdgpu_ring_patch_cs_in_place(ring, p, job, ib);
1066 amdgpu_bo_kunmap(aobj);
1067 if (r)
1068 return r;
1069 }
1070 }
1071
1072 return 0;
1073 }
1074
amdgpu_cs_patch_jobs(struct amdgpu_cs_parser * p)1075 static int amdgpu_cs_patch_jobs(struct amdgpu_cs_parser *p)
1076 {
1077 unsigned int i;
1078 int r;
1079
1080 for (i = 0; i < p->gang_size; ++i) {
1081 r = amdgpu_cs_patch_ibs(p, p->jobs[i]);
1082 if (r)
1083 return r;
1084 }
1085 return 0;
1086 }
1087
amdgpu_cs_vm_handling(struct amdgpu_cs_parser * p)1088 static int amdgpu_cs_vm_handling(struct amdgpu_cs_parser *p)
1089 {
1090 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1091 struct amdgpu_job *job = p->gang_leader;
1092 struct amdgpu_device *adev = p->adev;
1093 struct amdgpu_vm *vm = &fpriv->vm;
1094 struct amdgpu_bo_list_entry *e;
1095 struct amdgpu_bo_va *bo_va;
1096 struct amdgpu_bo *bo;
1097 unsigned int i;
1098 int r;
1099
1100 r = amdgpu_vm_clear_freed(adev, vm, NULL);
1101 if (r)
1102 return r;
1103
1104 r = amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
1105 if (r)
1106 return r;
1107
1108 r = amdgpu_sync_fence(&p->sync, fpriv->prt_va->last_pt_update);
1109 if (r)
1110 return r;
1111
1112 if (fpriv->csa_va) {
1113 bo_va = fpriv->csa_va;
1114 BUG_ON(!bo_va);
1115 r = amdgpu_vm_bo_update(adev, bo_va, false);
1116 if (r)
1117 return r;
1118
1119 r = amdgpu_sync_fence(&p->sync, bo_va->last_pt_update);
1120 if (r)
1121 return r;
1122 }
1123
1124 amdgpu_bo_list_for_each_entry(e, p->bo_list) {
1125 /* ignore duplicates */
1126 bo = ttm_to_amdgpu_bo(e->tv.bo);
1127 if (!bo)
1128 continue;
1129
1130 bo_va = e->bo_va;
1131 if (bo_va == NULL)
1132 continue;
1133
1134 r = amdgpu_vm_bo_update(adev, bo_va, false);
1135 if (r)
1136 return r;
1137
1138 r = amdgpu_sync_fence(&p->sync, bo_va->last_pt_update);
1139 if (r)
1140 return r;
1141 }
1142
1143 r = amdgpu_vm_handle_moved(adev, vm);
1144 if (r)
1145 return r;
1146
1147 r = amdgpu_vm_update_pdes(adev, vm, false);
1148 if (r)
1149 return r;
1150
1151 r = amdgpu_sync_fence(&p->sync, vm->last_update);
1152 if (r)
1153 return r;
1154
1155 for (i = 0; i < p->gang_size; ++i) {
1156 job = p->jobs[i];
1157
1158 if (!job->vm)
1159 continue;
1160
1161 job->vm_pd_addr = amdgpu_gmc_pd_addr(vm->root.bo);
1162 }
1163
1164 if (amdgpu_vm_debug) {
1165 /* Invalidate all BOs to test for userspace bugs */
1166 amdgpu_bo_list_for_each_entry(e, p->bo_list) {
1167 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
1168
1169 /* ignore duplicates */
1170 if (!bo)
1171 continue;
1172
1173 amdgpu_vm_bo_invalidate(adev, bo, false);
1174 }
1175 }
1176
1177 return 0;
1178 }
1179
amdgpu_cs_sync_rings(struct amdgpu_cs_parser * p)1180 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
1181 {
1182 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1183 struct drm_gpu_scheduler *sched;
1184 struct amdgpu_bo_list_entry *e;
1185 struct dma_fence *fence;
1186 unsigned int i;
1187 int r;
1188
1189 r = amdgpu_ctx_wait_prev_fence(p->ctx, p->entities[p->gang_leader_idx]);
1190 if (r) {
1191 if (r != -ERESTARTSYS)
1192 DRM_ERROR("amdgpu_ctx_wait_prev_fence failed.\n");
1193 return r;
1194 }
1195
1196 list_for_each_entry(e, &p->validated, tv.head) {
1197 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
1198 struct dma_resv *resv = bo->tbo.base.resv;
1199 enum amdgpu_sync_mode sync_mode;
1200
1201 sync_mode = amdgpu_bo_explicit_sync(bo) ?
1202 AMDGPU_SYNC_EXPLICIT : AMDGPU_SYNC_NE_OWNER;
1203 r = amdgpu_sync_resv(p->adev, &p->sync, resv, sync_mode,
1204 &fpriv->vm);
1205 if (r)
1206 return r;
1207 }
1208
1209 for (i = 0; i < p->gang_size; ++i) {
1210 r = amdgpu_sync_push_to_job(&p->sync, p->jobs[i]);
1211 if (r)
1212 return r;
1213 }
1214
1215 sched = p->gang_leader->base.entity->rq->sched;
1216 while ((fence = amdgpu_sync_get_fence(&p->sync))) {
1217 struct drm_sched_fence *s_fence = to_drm_sched_fence(fence);
1218
1219 /*
1220 * When we have an dependency it might be necessary to insert a
1221 * pipeline sync to make sure that all caches etc are flushed and the
1222 * next job actually sees the results from the previous one
1223 * before we start executing on the same scheduler ring.
1224 */
1225 if (!s_fence || s_fence->sched != sched) {
1226 dma_fence_put(fence);
1227 continue;
1228 }
1229
1230 r = amdgpu_sync_fence(&p->gang_leader->explicit_sync, fence);
1231 dma_fence_put(fence);
1232 if (r)
1233 return r;
1234 }
1235 return 0;
1236 }
1237
amdgpu_cs_post_dependencies(struct amdgpu_cs_parser * p)1238 static void amdgpu_cs_post_dependencies(struct amdgpu_cs_parser *p)
1239 {
1240 int i;
1241
1242 for (i = 0; i < p->num_post_deps; ++i) {
1243 if (p->post_deps[i].chain && p->post_deps[i].point) {
1244 drm_syncobj_add_point(p->post_deps[i].syncobj,
1245 p->post_deps[i].chain,
1246 p->fence, p->post_deps[i].point);
1247 p->post_deps[i].chain = NULL;
1248 } else {
1249 drm_syncobj_replace_fence(p->post_deps[i].syncobj,
1250 p->fence);
1251 }
1252 }
1253 }
1254
amdgpu_cs_submit(struct amdgpu_cs_parser * p,union drm_amdgpu_cs * cs)1255 static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
1256 union drm_amdgpu_cs *cs)
1257 {
1258 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1259 struct amdgpu_job *leader = p->gang_leader;
1260 struct amdgpu_bo_list_entry *e;
1261 unsigned int i;
1262 uint64_t seq;
1263 int r;
1264
1265 for (i = 0; i < p->gang_size; ++i)
1266 drm_sched_job_arm(&p->jobs[i]->base);
1267
1268 for (i = 0; i < p->gang_size; ++i) {
1269 struct dma_fence *fence;
1270
1271 if (p->jobs[i] == leader)
1272 continue;
1273
1274 fence = &p->jobs[i]->base.s_fence->scheduled;
1275 dma_fence_get(fence);
1276 r = drm_sched_job_add_dependency(&leader->base, fence);
1277 if (r) {
1278 dma_fence_put(fence);
1279 goto error_cleanup;
1280 }
1281 }
1282
1283 if (p->gang_size > 1) {
1284 for (i = 0; i < p->gang_size; ++i)
1285 amdgpu_job_set_gang_leader(p->jobs[i], leader);
1286 }
1287
1288 /* No memory allocation is allowed while holding the notifier lock.
1289 * The lock is held until amdgpu_cs_submit is finished and fence is
1290 * added to BOs.
1291 */
1292 mutex_lock(&p->adev->notifier_lock);
1293
1294 /* If userptr are invalidated after amdgpu_cs_parser_bos(), return
1295 * -EAGAIN, drmIoctl in libdrm will restart the amdgpu_cs_ioctl.
1296 */
1297 r = 0;
1298 amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
1299 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
1300
1301 r |= !amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm, e->range);
1302 e->range = NULL;
1303 }
1304 if (r) {
1305 r = -EAGAIN;
1306 goto error_unlock;
1307 }
1308
1309 p->fence = dma_fence_get(&leader->base.s_fence->finished);
1310 list_for_each_entry(e, &p->validated, tv.head) {
1311
1312 /* Everybody except for the gang leader uses READ */
1313 for (i = 0; i < p->gang_size; ++i) {
1314 if (p->jobs[i] == leader)
1315 continue;
1316
1317 dma_resv_add_fence(e->tv.bo->base.resv,
1318 &p->jobs[i]->base.s_fence->finished,
1319 DMA_RESV_USAGE_READ);
1320 }
1321
1322 /* The gang leader is remembered as writer */
1323 e->tv.num_shared = 0;
1324 }
1325
1326 seq = amdgpu_ctx_add_fence(p->ctx, p->entities[p->gang_leader_idx],
1327 p->fence);
1328 amdgpu_cs_post_dependencies(p);
1329
1330 if ((leader->preamble_status & AMDGPU_PREAMBLE_IB_PRESENT) &&
1331 !p->ctx->preamble_presented) {
1332 leader->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
1333 p->ctx->preamble_presented = true;
1334 }
1335
1336 cs->out.handle = seq;
1337 leader->uf_sequence = seq;
1338
1339 amdgpu_vm_bo_trace_cs(&fpriv->vm, &p->ticket);
1340 for (i = 0; i < p->gang_size; ++i) {
1341 amdgpu_job_free_resources(p->jobs[i]);
1342 trace_amdgpu_cs_ioctl(p->jobs[i]);
1343 drm_sched_entity_push_job(&p->jobs[i]->base);
1344 p->jobs[i] = NULL;
1345 }
1346
1347 amdgpu_vm_move_to_lru_tail(p->adev, &fpriv->vm);
1348 ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);
1349
1350 mutex_unlock(&p->adev->notifier_lock);
1351 mutex_unlock(&p->bo_list->bo_list_mutex);
1352 return 0;
1353
1354 error_unlock:
1355 mutex_unlock(&p->adev->notifier_lock);
1356
1357 error_cleanup:
1358 for (i = 0; i < p->gang_size; ++i)
1359 drm_sched_job_cleanup(&p->jobs[i]->base);
1360 return r;
1361 }
1362
1363 /* Cleanup the parser structure */
amdgpu_cs_parser_fini(struct amdgpu_cs_parser * parser)1364 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser)
1365 {
1366 unsigned i;
1367
1368 amdgpu_sync_free(&parser->sync);
1369 for (i = 0; i < parser->num_post_deps; i++) {
1370 drm_syncobj_put(parser->post_deps[i].syncobj);
1371 kfree(parser->post_deps[i].chain);
1372 }
1373 kfree(parser->post_deps);
1374
1375 dma_fence_put(parser->fence);
1376
1377 if (parser->ctx)
1378 amdgpu_ctx_put(parser->ctx);
1379 if (parser->bo_list)
1380 amdgpu_bo_list_put(parser->bo_list);
1381
1382 for (i = 0; i < parser->nchunks; i++)
1383 kvfree(parser->chunks[i].kdata);
1384 kvfree(parser->chunks);
1385 for (i = 0; i < parser->gang_size; ++i) {
1386 if (parser->jobs[i])
1387 amdgpu_job_free(parser->jobs[i]);
1388 }
1389 if (parser->uf_entry.tv.bo) {
1390 struct amdgpu_bo *uf = ttm_to_amdgpu_bo(parser->uf_entry.tv.bo);
1391
1392 amdgpu_bo_unref(&uf);
1393 }
1394 }
1395
amdgpu_cs_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)1396 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
1397 {
1398 struct amdgpu_device *adev = drm_to_adev(dev);
1399 struct amdgpu_cs_parser parser;
1400 int r;
1401
1402 if (amdgpu_ras_intr_triggered())
1403 return -EHWPOISON;
1404
1405 if (!adev->accel_working)
1406 return -EBUSY;
1407
1408 r = amdgpu_cs_parser_init(&parser, adev, filp, data);
1409 if (r) {
1410 if (printk_ratelimit())
1411 DRM_ERROR("Failed to initialize parser %d!\n", r);
1412 return r;
1413 }
1414
1415 r = amdgpu_cs_pass1(&parser, data);
1416 if (r)
1417 goto error_fini;
1418
1419 r = amdgpu_cs_pass2(&parser);
1420 if (r)
1421 goto error_fini;
1422
1423 r = amdgpu_cs_parser_bos(&parser, data);
1424 if (r) {
1425 if (r == -ENOMEM)
1426 DRM_ERROR("Not enough memory for command submission!\n");
1427 else if (r != -ERESTARTSYS && r != -EAGAIN)
1428 DRM_ERROR("Failed to process the buffer list %d!\n", r);
1429 goto error_fini;
1430 }
1431
1432 r = amdgpu_cs_patch_jobs(&parser);
1433 if (r)
1434 goto error_backoff;
1435
1436 r = amdgpu_cs_vm_handling(&parser);
1437 if (r)
1438 goto error_backoff;
1439
1440 r = amdgpu_cs_sync_rings(&parser);
1441 if (r)
1442 goto error_backoff;
1443
1444 trace_amdgpu_cs_ibs(&parser);
1445
1446 r = amdgpu_cs_submit(&parser, data);
1447 if (r)
1448 goto error_backoff;
1449
1450 amdgpu_cs_parser_fini(&parser);
1451 return 0;
1452
1453 error_backoff:
1454 ttm_eu_backoff_reservation(&parser.ticket, &parser.validated);
1455 mutex_unlock(&parser.bo_list->bo_list_mutex);
1456
1457 error_fini:
1458 amdgpu_cs_parser_fini(&parser);
1459 return r;
1460 }
1461
1462 /**
1463 * amdgpu_cs_wait_ioctl - wait for a command submission to finish
1464 *
1465 * @dev: drm device
1466 * @data: data from userspace
1467 * @filp: file private
1468 *
1469 * Wait for the command submission identified by handle to finish.
1470 */
amdgpu_cs_wait_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)1471 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
1472 struct drm_file *filp)
1473 {
1474 union drm_amdgpu_wait_cs *wait = data;
1475 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1476 struct drm_sched_entity *entity;
1477 struct amdgpu_ctx *ctx;
1478 struct dma_fence *fence;
1479 long r;
1480
1481 ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
1482 if (ctx == NULL)
1483 return -EINVAL;
1484
1485 r = amdgpu_ctx_get_entity(ctx, wait->in.ip_type, wait->in.ip_instance,
1486 wait->in.ring, &entity);
1487 if (r) {
1488 amdgpu_ctx_put(ctx);
1489 return r;
1490 }
1491
1492 fence = amdgpu_ctx_get_fence(ctx, entity, wait->in.handle);
1493 if (IS_ERR(fence))
1494 r = PTR_ERR(fence);
1495 else if (fence) {
1496 r = dma_fence_wait_timeout(fence, true, timeout);
1497 if (r > 0 && fence->error)
1498 r = fence->error;
1499 dma_fence_put(fence);
1500 } else
1501 r = 1;
1502
1503 amdgpu_ctx_put(ctx);
1504 if (r < 0)
1505 return r;
1506
1507 memset(wait, 0, sizeof(*wait));
1508 wait->out.status = (r == 0);
1509
1510 return 0;
1511 }
1512
1513 /**
1514 * amdgpu_cs_get_fence - helper to get fence from drm_amdgpu_fence
1515 *
1516 * @adev: amdgpu device
1517 * @filp: file private
1518 * @user: drm_amdgpu_fence copied from user space
1519 */
amdgpu_cs_get_fence(struct amdgpu_device * adev,struct drm_file * filp,struct drm_amdgpu_fence * user)1520 static struct dma_fence *amdgpu_cs_get_fence(struct amdgpu_device *adev,
1521 struct drm_file *filp,
1522 struct drm_amdgpu_fence *user)
1523 {
1524 struct drm_sched_entity *entity;
1525 struct amdgpu_ctx *ctx;
1526 struct dma_fence *fence;
1527 int r;
1528
1529 ctx = amdgpu_ctx_get(filp->driver_priv, user->ctx_id);
1530 if (ctx == NULL)
1531 return ERR_PTR(-EINVAL);
1532
1533 r = amdgpu_ctx_get_entity(ctx, user->ip_type, user->ip_instance,
1534 user->ring, &entity);
1535 if (r) {
1536 amdgpu_ctx_put(ctx);
1537 return ERR_PTR(r);
1538 }
1539
1540 fence = amdgpu_ctx_get_fence(ctx, entity, user->seq_no);
1541 amdgpu_ctx_put(ctx);
1542
1543 return fence;
1544 }
1545
amdgpu_cs_fence_to_handle_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)1546 int amdgpu_cs_fence_to_handle_ioctl(struct drm_device *dev, void *data,
1547 struct drm_file *filp)
1548 {
1549 struct amdgpu_device *adev = drm_to_adev(dev);
1550 union drm_amdgpu_fence_to_handle *info = data;
1551 struct dma_fence *fence;
1552 struct drm_syncobj *syncobj;
1553 struct sync_file *sync_file;
1554 int fd, r;
1555
1556 fence = amdgpu_cs_get_fence(adev, filp, &info->in.fence);
1557 if (IS_ERR(fence))
1558 return PTR_ERR(fence);
1559
1560 if (!fence)
1561 fence = dma_fence_get_stub();
1562
1563 switch (info->in.what) {
1564 case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ:
1565 r = drm_syncobj_create(&syncobj, 0, fence);
1566 dma_fence_put(fence);
1567 if (r)
1568 return r;
1569 r = drm_syncobj_get_handle(filp, syncobj, &info->out.handle);
1570 drm_syncobj_put(syncobj);
1571 return r;
1572
1573 case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ_FD:
1574 r = drm_syncobj_create(&syncobj, 0, fence);
1575 dma_fence_put(fence);
1576 if (r)
1577 return r;
1578 r = drm_syncobj_get_fd(syncobj, (int *)&info->out.handle);
1579 drm_syncobj_put(syncobj);
1580 return r;
1581
1582 case AMDGPU_FENCE_TO_HANDLE_GET_SYNC_FILE_FD:
1583 fd = get_unused_fd_flags(O_CLOEXEC);
1584 if (fd < 0) {
1585 dma_fence_put(fence);
1586 return fd;
1587 }
1588
1589 sync_file = sync_file_create(fence);
1590 dma_fence_put(fence);
1591 if (!sync_file) {
1592 put_unused_fd(fd);
1593 return -ENOMEM;
1594 }
1595
1596 fd_install(fd, sync_file->file);
1597 info->out.handle = fd;
1598 return 0;
1599
1600 default:
1601 dma_fence_put(fence);
1602 return -EINVAL;
1603 }
1604 }
1605
1606 /**
1607 * amdgpu_cs_wait_all_fences - wait on all fences to signal
1608 *
1609 * @adev: amdgpu device
1610 * @filp: file private
1611 * @wait: wait parameters
1612 * @fences: array of drm_amdgpu_fence
1613 */
amdgpu_cs_wait_all_fences(struct amdgpu_device * adev,struct drm_file * filp,union drm_amdgpu_wait_fences * wait,struct drm_amdgpu_fence * fences)1614 static int amdgpu_cs_wait_all_fences(struct amdgpu_device *adev,
1615 struct drm_file *filp,
1616 union drm_amdgpu_wait_fences *wait,
1617 struct drm_amdgpu_fence *fences)
1618 {
1619 uint32_t fence_count = wait->in.fence_count;
1620 unsigned int i;
1621 long r = 1;
1622
1623 for (i = 0; i < fence_count; i++) {
1624 struct dma_fence *fence;
1625 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1626
1627 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1628 if (IS_ERR(fence))
1629 return PTR_ERR(fence);
1630 else if (!fence)
1631 continue;
1632
1633 r = dma_fence_wait_timeout(fence, true, timeout);
1634 dma_fence_put(fence);
1635 if (r < 0)
1636 return r;
1637
1638 if (r == 0)
1639 break;
1640
1641 if (fence->error)
1642 return fence->error;
1643 }
1644
1645 memset(wait, 0, sizeof(*wait));
1646 wait->out.status = (r > 0);
1647
1648 return 0;
1649 }
1650
1651 /**
1652 * amdgpu_cs_wait_any_fence - wait on any fence to signal
1653 *
1654 * @adev: amdgpu device
1655 * @filp: file private
1656 * @wait: wait parameters
1657 * @fences: array of drm_amdgpu_fence
1658 */
amdgpu_cs_wait_any_fence(struct amdgpu_device * adev,struct drm_file * filp,union drm_amdgpu_wait_fences * wait,struct drm_amdgpu_fence * fences)1659 static int amdgpu_cs_wait_any_fence(struct amdgpu_device *adev,
1660 struct drm_file *filp,
1661 union drm_amdgpu_wait_fences *wait,
1662 struct drm_amdgpu_fence *fences)
1663 {
1664 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1665 uint32_t fence_count = wait->in.fence_count;
1666 uint32_t first = ~0;
1667 struct dma_fence **array;
1668 unsigned int i;
1669 long r;
1670
1671 /* Prepare the fence array */
1672 array = kcalloc(fence_count, sizeof(struct dma_fence *), GFP_KERNEL);
1673
1674 if (array == NULL)
1675 return -ENOMEM;
1676
1677 for (i = 0; i < fence_count; i++) {
1678 struct dma_fence *fence;
1679
1680 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1681 if (IS_ERR(fence)) {
1682 r = PTR_ERR(fence);
1683 goto err_free_fence_array;
1684 } else if (fence) {
1685 array[i] = fence;
1686 } else { /* NULL, the fence has been already signaled */
1687 r = 1;
1688 first = i;
1689 goto out;
1690 }
1691 }
1692
1693 r = dma_fence_wait_any_timeout(array, fence_count, true, timeout,
1694 &first);
1695 if (r < 0)
1696 goto err_free_fence_array;
1697
1698 out:
1699 memset(wait, 0, sizeof(*wait));
1700 wait->out.status = (r > 0);
1701 wait->out.first_signaled = first;
1702
1703 if (first < fence_count && array[first])
1704 r = array[first]->error;
1705 else
1706 r = 0;
1707
1708 err_free_fence_array:
1709 for (i = 0; i < fence_count; i++)
1710 dma_fence_put(array[i]);
1711 kfree(array);
1712
1713 return r;
1714 }
1715
1716 /**
1717 * amdgpu_cs_wait_fences_ioctl - wait for multiple command submissions to finish
1718 *
1719 * @dev: drm device
1720 * @data: data from userspace
1721 * @filp: file private
1722 */
amdgpu_cs_wait_fences_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)1723 int amdgpu_cs_wait_fences_ioctl(struct drm_device *dev, void *data,
1724 struct drm_file *filp)
1725 {
1726 struct amdgpu_device *adev = drm_to_adev(dev);
1727 union drm_amdgpu_wait_fences *wait = data;
1728 uint32_t fence_count = wait->in.fence_count;
1729 struct drm_amdgpu_fence *fences_user;
1730 struct drm_amdgpu_fence *fences;
1731 int r;
1732
1733 /* Get the fences from userspace */
1734 fences = kmalloc_array(fence_count, sizeof(struct drm_amdgpu_fence),
1735 GFP_KERNEL);
1736 if (fences == NULL)
1737 return -ENOMEM;
1738
1739 fences_user = u64_to_user_ptr(wait->in.fences);
1740 if (copy_from_user(fences, fences_user,
1741 sizeof(struct drm_amdgpu_fence) * fence_count)) {
1742 r = -EFAULT;
1743 goto err_free_fences;
1744 }
1745
1746 if (wait->in.wait_all)
1747 r = amdgpu_cs_wait_all_fences(adev, filp, wait, fences);
1748 else
1749 r = amdgpu_cs_wait_any_fence(adev, filp, wait, fences);
1750
1751 err_free_fences:
1752 kfree(fences);
1753
1754 return r;
1755 }
1756
1757 /**
1758 * amdgpu_cs_find_mapping - find bo_va for VM address
1759 *
1760 * @parser: command submission parser context
1761 * @addr: VM address
1762 * @bo: resulting BO of the mapping found
1763 * @map: Placeholder to return found BO mapping
1764 *
1765 * Search the buffer objects in the command submission context for a certain
1766 * virtual memory address. Returns allocation structure when found, NULL
1767 * otherwise.
1768 */
amdgpu_cs_find_mapping(struct amdgpu_cs_parser * parser,uint64_t addr,struct amdgpu_bo ** bo,struct amdgpu_bo_va_mapping ** map)1769 int amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1770 uint64_t addr, struct amdgpu_bo **bo,
1771 struct amdgpu_bo_va_mapping **map)
1772 {
1773 struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
1774 struct ttm_operation_ctx ctx = { false, false };
1775 struct amdgpu_vm *vm = &fpriv->vm;
1776 struct amdgpu_bo_va_mapping *mapping;
1777 int r;
1778
1779 addr /= AMDGPU_GPU_PAGE_SIZE;
1780
1781 mapping = amdgpu_vm_bo_lookup_mapping(vm, addr);
1782 if (!mapping || !mapping->bo_va || !mapping->bo_va->base.bo)
1783 return -EINVAL;
1784
1785 *bo = mapping->bo_va->base.bo;
1786 *map = mapping;
1787
1788 /* Double check that the BO is reserved by this CS */
1789 if (dma_resv_locking_ctx((*bo)->tbo.base.resv) != &parser->ticket)
1790 return -EINVAL;
1791
1792 if (!((*bo)->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)) {
1793 (*bo)->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1794 amdgpu_bo_placement_from_domain(*bo, (*bo)->allowed_domains);
1795 r = ttm_bo_validate(&(*bo)->tbo, &(*bo)->placement, &ctx);
1796 if (r)
1797 return r;
1798 }
1799
1800 return amdgpu_ttm_alloc_gart(&(*bo)->tbo);
1801 }
1802