xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c (revision 3e44c471a2dab210f7e9b1e5f7d4d54d52df59eb)
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 #include <linux/list_sort.h>
28 #include <drm/drmP.h>
29 #include <drm/amdgpu_drm.h>
30 #include "amdgpu.h"
31 #include "amdgpu_trace.h"
32 
33 #define AMDGPU_CS_MAX_PRIORITY		32u
34 #define AMDGPU_CS_NUM_BUCKETS		(AMDGPU_CS_MAX_PRIORITY + 1)
35 
36 /* This is based on the bucket sort with O(n) time complexity.
37  * An item with priority "i" is added to bucket[i]. The lists are then
38  * concatenated in descending order.
39  */
40 struct amdgpu_cs_buckets {
41 	struct list_head bucket[AMDGPU_CS_NUM_BUCKETS];
42 };
43 
44 static void amdgpu_cs_buckets_init(struct amdgpu_cs_buckets *b)
45 {
46 	unsigned i;
47 
48 	for (i = 0; i < AMDGPU_CS_NUM_BUCKETS; i++)
49 		INIT_LIST_HEAD(&b->bucket[i]);
50 }
51 
52 static void amdgpu_cs_buckets_add(struct amdgpu_cs_buckets *b,
53 				  struct list_head *item, unsigned priority)
54 {
55 	/* Since buffers which appear sooner in the relocation list are
56 	 * likely to be used more often than buffers which appear later
57 	 * in the list, the sort mustn't change the ordering of buffers
58 	 * with the same priority, i.e. it must be stable.
59 	 */
60 	list_add_tail(item, &b->bucket[min(priority, AMDGPU_CS_MAX_PRIORITY)]);
61 }
62 
63 static void amdgpu_cs_buckets_get_list(struct amdgpu_cs_buckets *b,
64 				       struct list_head *out_list)
65 {
66 	unsigned i;
67 
68 	/* Connect the sorted buckets in the output list. */
69 	for (i = 0; i < AMDGPU_CS_NUM_BUCKETS; i++) {
70 		list_splice(&b->bucket[i], out_list);
71 	}
72 }
73 
74 int amdgpu_cs_get_ring(struct amdgpu_device *adev, u32 ip_type,
75 		       u32 ip_instance, u32 ring,
76 		       struct amdgpu_ring **out_ring)
77 {
78 	/* Right now all IPs have only one instance - multiple rings. */
79 	if (ip_instance != 0) {
80 		DRM_ERROR("invalid ip instance: %d\n", ip_instance);
81 		return -EINVAL;
82 	}
83 
84 	switch (ip_type) {
85 	default:
86 		DRM_ERROR("unknown ip type: %d\n", ip_type);
87 		return -EINVAL;
88 	case AMDGPU_HW_IP_GFX:
89 		if (ring < adev->gfx.num_gfx_rings) {
90 			*out_ring = &adev->gfx.gfx_ring[ring];
91 		} else {
92 			DRM_ERROR("only %d gfx rings are supported now\n",
93 				  adev->gfx.num_gfx_rings);
94 			return -EINVAL;
95 		}
96 		break;
97 	case AMDGPU_HW_IP_COMPUTE:
98 		if (ring < adev->gfx.num_compute_rings) {
99 			*out_ring = &adev->gfx.compute_ring[ring];
100 		} else {
101 			DRM_ERROR("only %d compute rings are supported now\n",
102 				  adev->gfx.num_compute_rings);
103 			return -EINVAL;
104 		}
105 		break;
106 	case AMDGPU_HW_IP_DMA:
107 		if (ring < 2) {
108 			*out_ring = &adev->sdma[ring].ring;
109 		} else {
110 			DRM_ERROR("only two SDMA rings are supported\n");
111 			return -EINVAL;
112 		}
113 		break;
114 	case AMDGPU_HW_IP_UVD:
115 		*out_ring = &adev->uvd.ring;
116 		break;
117 	case AMDGPU_HW_IP_VCE:
118 		if (ring < 2){
119 			*out_ring = &adev->vce.ring[ring];
120 		} else {
121 			DRM_ERROR("only two VCE rings are supported\n");
122 			return -EINVAL;
123 		}
124 		break;
125 	}
126 	return 0;
127 }
128 
129 int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
130 {
131 	union drm_amdgpu_cs *cs = data;
132 	uint64_t *chunk_array_user;
133 	uint64_t *chunk_array = NULL;
134 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
135 	unsigned size, i;
136 	int r = 0;
137 
138 	if (!cs->in.num_chunks)
139 		goto out;
140 
141 	p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
142 	if (!p->ctx) {
143 		r = -EINVAL;
144 		goto out;
145 	}
146 	p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
147 
148 	/* get chunks */
149 	INIT_LIST_HEAD(&p->validated);
150 	chunk_array = kcalloc(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
151 	if (chunk_array == NULL) {
152 		r = -ENOMEM;
153 		goto out;
154 	}
155 
156 	chunk_array_user = (uint64_t *)(unsigned long)(cs->in.chunks);
157 	if (copy_from_user(chunk_array, chunk_array_user,
158 			   sizeof(uint64_t)*cs->in.num_chunks)) {
159 		r = -EFAULT;
160 		goto out;
161 	}
162 
163 	p->nchunks = cs->in.num_chunks;
164 	p->chunks = kcalloc(p->nchunks, sizeof(struct amdgpu_cs_chunk),
165 			    GFP_KERNEL);
166 	if (p->chunks == NULL) {
167 		r = -ENOMEM;
168 		goto out;
169 	}
170 
171 	for (i = 0; i < p->nchunks; i++) {
172 		struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
173 		struct drm_amdgpu_cs_chunk user_chunk;
174 		uint32_t __user *cdata;
175 
176 		chunk_ptr = (void __user *)(unsigned long)chunk_array[i];
177 		if (copy_from_user(&user_chunk, chunk_ptr,
178 				       sizeof(struct drm_amdgpu_cs_chunk))) {
179 			r = -EFAULT;
180 			goto out;
181 		}
182 		p->chunks[i].chunk_id = user_chunk.chunk_id;
183 		p->chunks[i].length_dw = user_chunk.length_dw;
184 		if (p->chunks[i].chunk_id == AMDGPU_CHUNK_ID_IB)
185 			p->num_ibs++;
186 
187 		size = p->chunks[i].length_dw;
188 		cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
189 		p->chunks[i].user_ptr = cdata;
190 
191 		p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t));
192 		if (p->chunks[i].kdata == NULL) {
193 			r = -ENOMEM;
194 			goto out;
195 		}
196 		size *= sizeof(uint32_t);
197 		if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
198 			r = -EFAULT;
199 			goto out;
200 		}
201 
202 		if (p->chunks[i].chunk_id == AMDGPU_CHUNK_ID_FENCE) {
203 			size = sizeof(struct drm_amdgpu_cs_chunk_fence);
204 			if (p->chunks[i].length_dw * sizeof(uint32_t) >= size) {
205 				uint32_t handle;
206 				struct drm_gem_object *gobj;
207 				struct drm_amdgpu_cs_chunk_fence *fence_data;
208 
209 				fence_data = (void *)p->chunks[i].kdata;
210 				handle = fence_data->handle;
211 				gobj = drm_gem_object_lookup(p->adev->ddev,
212 							     p->filp, handle);
213 				if (gobj == NULL) {
214 					r = -EINVAL;
215 					goto out;
216 				}
217 
218 				p->uf.bo = gem_to_amdgpu_bo(gobj);
219 				p->uf.offset = fence_data->offset;
220 			} else {
221 				r = -EINVAL;
222 				goto out;
223 			}
224 		}
225 	}
226 
227 	p->ibs = kcalloc(p->num_ibs, sizeof(struct amdgpu_ib), GFP_KERNEL);
228 	if (!p->ibs) {
229 		r = -ENOMEM;
230 		goto out;
231 	}
232 
233 out:
234 	kfree(chunk_array);
235 	return r;
236 }
237 
238 /* Returns how many bytes TTM can move per IB.
239  */
240 static u64 amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev)
241 {
242 	u64 real_vram_size = adev->mc.real_vram_size;
243 	u64 vram_usage = atomic64_read(&adev->vram_usage);
244 
245 	/* This function is based on the current VRAM usage.
246 	 *
247 	 * - If all of VRAM is free, allow relocating the number of bytes that
248 	 *   is equal to 1/4 of the size of VRAM for this IB.
249 
250 	 * - If more than one half of VRAM is occupied, only allow relocating
251 	 *   1 MB of data for this IB.
252 	 *
253 	 * - From 0 to one half of used VRAM, the threshold decreases
254 	 *   linearly.
255 	 *         __________________
256 	 * 1/4 of -|\               |
257 	 * VRAM    | \              |
258 	 *         |  \             |
259 	 *         |   \            |
260 	 *         |    \           |
261 	 *         |     \          |
262 	 *         |      \         |
263 	 *         |       \________|1 MB
264 	 *         |----------------|
265 	 *    VRAM 0 %             100 %
266 	 *         used            used
267 	 *
268 	 * Note: It's a threshold, not a limit. The threshold must be crossed
269 	 * for buffer relocations to stop, so any buffer of an arbitrary size
270 	 * can be moved as long as the threshold isn't crossed before
271 	 * the relocation takes place. We don't want to disable buffer
272 	 * relocations completely.
273 	 *
274 	 * The idea is that buffers should be placed in VRAM at creation time
275 	 * and TTM should only do a minimum number of relocations during
276 	 * command submission. In practice, you need to submit at least
277 	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
278 	 *
279 	 * Also, things can get pretty crazy under memory pressure and actual
280 	 * VRAM usage can change a lot, so playing safe even at 50% does
281 	 * consistently increase performance.
282 	 */
283 
284 	u64 half_vram = real_vram_size >> 1;
285 	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
286 	u64 bytes_moved_threshold = half_free_vram >> 1;
287 	return max(bytes_moved_threshold, 1024*1024ull);
288 }
289 
290 int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p)
291 {
292 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
293 	struct amdgpu_vm *vm = &fpriv->vm;
294 	struct amdgpu_device *adev = p->adev;
295 	struct amdgpu_bo_list_entry *lobj;
296 	struct list_head duplicates;
297 	struct amdgpu_bo *bo;
298 	u64 bytes_moved = 0, initial_bytes_moved;
299 	u64 bytes_moved_threshold = amdgpu_cs_get_threshold_for_moves(adev);
300 	int r;
301 
302 	INIT_LIST_HEAD(&duplicates);
303 	r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true, &duplicates);
304 	if (unlikely(r != 0)) {
305 		return r;
306 	}
307 
308 	list_for_each_entry(lobj, &p->validated, tv.head) {
309 		bo = lobj->robj;
310 		if (!bo->pin_count) {
311 			u32 domain = lobj->prefered_domains;
312 			u32 current_domain =
313 				amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
314 
315 			/* Check if this buffer will be moved and don't move it
316 			 * if we have moved too many buffers for this IB already.
317 			 *
318 			 * Note that this allows moving at least one buffer of
319 			 * any size, because it doesn't take the current "bo"
320 			 * into account. We don't want to disallow buffer moves
321 			 * completely.
322 			 */
323 			if (current_domain != AMDGPU_GEM_DOMAIN_CPU &&
324 			    (domain & current_domain) == 0 && /* will be moved */
325 			    bytes_moved > bytes_moved_threshold) {
326 				/* don't move it */
327 				domain = current_domain;
328 			}
329 
330 		retry:
331 			amdgpu_ttm_placement_from_domain(bo, domain);
332 			initial_bytes_moved = atomic64_read(&adev->num_bytes_moved);
333 			r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
334 			bytes_moved += atomic64_read(&adev->num_bytes_moved) -
335 				       initial_bytes_moved;
336 
337 			if (unlikely(r)) {
338 				if (r != -ERESTARTSYS && domain != lobj->allowed_domains) {
339 					domain = lobj->allowed_domains;
340 					goto retry;
341 				}
342 				ttm_eu_backoff_reservation(&p->ticket, &p->validated);
343 				return r;
344 			}
345 		}
346 		lobj->bo_va = amdgpu_vm_bo_find(vm, bo);
347 	}
348 	return 0;
349 }
350 
351 static int amdgpu_cs_parser_relocs(struct amdgpu_cs_parser *p)
352 {
353 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
354 	struct amdgpu_cs_buckets buckets;
355 	bool need_mmap_lock = false;
356 	int i, r;
357 
358 	if (p->bo_list) {
359 		need_mmap_lock = p->bo_list->has_userptr;
360 		amdgpu_cs_buckets_init(&buckets);
361 		for (i = 0; i < p->bo_list->num_entries; i++)
362 			amdgpu_cs_buckets_add(&buckets, &p->bo_list->array[i].tv.head,
363 								  p->bo_list->array[i].priority);
364 
365 		amdgpu_cs_buckets_get_list(&buckets, &p->validated);
366 	}
367 
368 	p->vm_bos = amdgpu_vm_get_bos(p->adev, &fpriv->vm,
369 				      &p->validated);
370 
371 	if (need_mmap_lock)
372 		down_read(&current->mm->mmap_sem);
373 
374 	r = amdgpu_cs_list_validate(p);
375 
376 	if (need_mmap_lock)
377 		up_read(&current->mm->mmap_sem);
378 
379 	return r;
380 }
381 
382 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
383 {
384 	struct amdgpu_bo_list_entry *e;
385 	int r;
386 
387 	list_for_each_entry(e, &p->validated, tv.head) {
388 		struct reservation_object *resv = e->robj->tbo.resv;
389 		r = amdgpu_sync_resv(p->adev, &p->ibs[0].sync, resv, p->filp);
390 
391 		if (r)
392 			return r;
393 	}
394 	return 0;
395 }
396 
397 static int cmp_size_smaller_first(void *priv, struct list_head *a,
398 				  struct list_head *b)
399 {
400 	struct amdgpu_bo_list_entry *la = list_entry(a, struct amdgpu_bo_list_entry, tv.head);
401 	struct amdgpu_bo_list_entry *lb = list_entry(b, struct amdgpu_bo_list_entry, tv.head);
402 
403 	/* Sort A before B if A is smaller. */
404 	return (int)la->robj->tbo.num_pages - (int)lb->robj->tbo.num_pages;
405 }
406 
407 /**
408  * cs_parser_fini() - clean parser states
409  * @parser:	parser structure holding parsing context.
410  * @error:	error number
411  *
412  * If error is set than unvalidate buffer, otherwise just free memory
413  * used by parsing context.
414  **/
415 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bool backoff)
416 {
417 	unsigned i;
418 
419 	if (!error) {
420 		/* Sort the buffer list from the smallest to largest buffer,
421 		 * which affects the order of buffers in the LRU list.
422 		 * This assures that the smallest buffers are added first
423 		 * to the LRU list, so they are likely to be later evicted
424 		 * first, instead of large buffers whose eviction is more
425 		 * expensive.
426 		 *
427 		 * This slightly lowers the number of bytes moved by TTM
428 		 * per frame under memory pressure.
429 		 */
430 		list_sort(NULL, &parser->validated, cmp_size_smaller_first);
431 
432 		ttm_eu_fence_buffer_objects(&parser->ticket,
433 				&parser->validated,
434 				&parser->ibs[parser->num_ibs-1].fence->base);
435 	} else if (backoff) {
436 		ttm_eu_backoff_reservation(&parser->ticket,
437 					   &parser->validated);
438 	}
439 
440 	if (parser->ctx)
441 		amdgpu_ctx_put(parser->ctx);
442 	if (parser->bo_list)
443 		amdgpu_bo_list_put(parser->bo_list);
444 	drm_free_large(parser->vm_bos);
445 	for (i = 0; i < parser->nchunks; i++)
446 		drm_free_large(parser->chunks[i].kdata);
447 	kfree(parser->chunks);
448 	for (i = 0; i < parser->num_ibs; i++)
449 		amdgpu_ib_free(parser->adev, &parser->ibs[i]);
450 	kfree(parser->ibs);
451 	if (parser->uf.bo)
452 		drm_gem_object_unreference_unlocked(&parser->uf.bo->gem_base);
453 }
454 
455 static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p,
456 				   struct amdgpu_vm *vm)
457 {
458 	struct amdgpu_device *adev = p->adev;
459 	struct amdgpu_bo_va *bo_va;
460 	struct amdgpu_bo *bo;
461 	int i, r;
462 
463 	r = amdgpu_vm_update_page_directory(adev, vm);
464 	if (r)
465 		return r;
466 
467 	r = amdgpu_vm_clear_freed(adev, vm);
468 	if (r)
469 		return r;
470 
471 	if (p->bo_list) {
472 		for (i = 0; i < p->bo_list->num_entries; i++) {
473 			/* ignore duplicates */
474 			bo = p->bo_list->array[i].robj;
475 			if (!bo)
476 				continue;
477 
478 			bo_va = p->bo_list->array[i].bo_va;
479 			if (bo_va == NULL)
480 				continue;
481 
482 			r = amdgpu_vm_bo_update(adev, bo_va, &bo->tbo.mem);
483 			if (r)
484 				return r;
485 
486 			amdgpu_sync_fence(&p->ibs[0].sync, bo_va->last_pt_update);
487 		}
488 	}
489 
490 	return amdgpu_vm_clear_invalids(adev, vm, &p->ibs[0].sync);
491 }
492 
493 static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
494 				 struct amdgpu_cs_parser *parser)
495 {
496 	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
497 	struct amdgpu_vm *vm = &fpriv->vm;
498 	struct amdgpu_ring *ring;
499 	int i, r;
500 
501 	if (parser->num_ibs == 0)
502 		return 0;
503 
504 	/* Only for UVD/VCE VM emulation */
505 	for (i = 0; i < parser->num_ibs; i++) {
506 		ring = parser->ibs[i].ring;
507 		if (ring->funcs->parse_cs) {
508 			r = amdgpu_ring_parse_cs(ring, parser, i);
509 			if (r)
510 				return r;
511 		}
512 	}
513 
514 	mutex_lock(&vm->mutex);
515 	r = amdgpu_bo_vm_update_pte(parser, vm);
516 	if (r) {
517 		goto out;
518 	}
519 	amdgpu_cs_sync_rings(parser);
520 
521 	r = amdgpu_ib_schedule(adev, parser->num_ibs, parser->ibs,
522 			       parser->filp);
523 
524 out:
525 	mutex_unlock(&vm->mutex);
526 	return r;
527 }
528 
529 static int amdgpu_cs_handle_lockup(struct amdgpu_device *adev, int r)
530 {
531 	if (r == -EDEADLK) {
532 		r = amdgpu_gpu_reset(adev);
533 		if (!r)
534 			r = -EAGAIN;
535 	}
536 	return r;
537 }
538 
539 static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
540 			     struct amdgpu_cs_parser *parser)
541 {
542 	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
543 	struct amdgpu_vm *vm = &fpriv->vm;
544 	int i, j;
545 	int r;
546 
547 	for (i = 0, j = 0; i < parser->nchunks && j < parser->num_ibs; i++) {
548 		struct amdgpu_cs_chunk *chunk;
549 		struct amdgpu_ib *ib;
550 		struct drm_amdgpu_cs_chunk_ib *chunk_ib;
551 		struct amdgpu_ring *ring;
552 
553 		chunk = &parser->chunks[i];
554 		ib = &parser->ibs[j];
555 		chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
556 
557 		if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
558 			continue;
559 
560 		r = amdgpu_cs_get_ring(adev, chunk_ib->ip_type,
561 				       chunk_ib->ip_instance, chunk_ib->ring,
562 				       &ring);
563 		if (r)
564 			return r;
565 
566 		if (ring->funcs->parse_cs) {
567 			struct amdgpu_bo_va_mapping *m;
568 			struct amdgpu_bo *aobj = NULL;
569 			uint64_t offset;
570 			uint8_t *kptr;
571 
572 			m = amdgpu_cs_find_mapping(parser, chunk_ib->va_start,
573 						   &aobj);
574 			if (!aobj) {
575 				DRM_ERROR("IB va_start is invalid\n");
576 				return -EINVAL;
577 			}
578 
579 			if ((chunk_ib->va_start + chunk_ib->ib_bytes) >
580 			    (m->it.last + 1) * AMDGPU_GPU_PAGE_SIZE) {
581 				DRM_ERROR("IB va_start+ib_bytes is invalid\n");
582 				return -EINVAL;
583 			}
584 
585 			/* the IB should be reserved at this point */
586 			r = amdgpu_bo_kmap(aobj, (void **)&kptr);
587 			if (r) {
588 				return r;
589 			}
590 
591 			offset = ((uint64_t)m->it.start) * AMDGPU_GPU_PAGE_SIZE;
592 			kptr += chunk_ib->va_start - offset;
593 
594 			r =  amdgpu_ib_get(ring, NULL, chunk_ib->ib_bytes, ib);
595 			if (r) {
596 				DRM_ERROR("Failed to get ib !\n");
597 				return r;
598 			}
599 
600 			memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
601 			amdgpu_bo_kunmap(aobj);
602 		} else {
603 			r =  amdgpu_ib_get(ring, vm, 0, ib);
604 			if (r) {
605 				DRM_ERROR("Failed to get ib !\n");
606 				return r;
607 			}
608 
609 			ib->gpu_addr = chunk_ib->va_start;
610 		}
611 
612 		ib->length_dw = chunk_ib->ib_bytes / 4;
613 		ib->flags = chunk_ib->flags;
614 		ib->ctx = parser->ctx;
615 		j++;
616 	}
617 
618 	if (!parser->num_ibs)
619 		return 0;
620 
621 	/* add GDS resources to first IB */
622 	if (parser->bo_list) {
623 		struct amdgpu_bo *gds = parser->bo_list->gds_obj;
624 		struct amdgpu_bo *gws = parser->bo_list->gws_obj;
625 		struct amdgpu_bo *oa = parser->bo_list->oa_obj;
626 		struct amdgpu_ib *ib = &parser->ibs[0];
627 
628 		if (gds) {
629 			ib->gds_base = amdgpu_bo_gpu_offset(gds);
630 			ib->gds_size = amdgpu_bo_size(gds);
631 		}
632 		if (gws) {
633 			ib->gws_base = amdgpu_bo_gpu_offset(gws);
634 			ib->gws_size = amdgpu_bo_size(gws);
635 		}
636 		if (oa) {
637 			ib->oa_base = amdgpu_bo_gpu_offset(oa);
638 			ib->oa_size = amdgpu_bo_size(oa);
639 		}
640 	}
641 
642 	/* wrap the last IB with user fence */
643 	if (parser->uf.bo) {
644 		struct amdgpu_ib *ib = &parser->ibs[parser->num_ibs - 1];
645 
646 		/* UVD & VCE fw doesn't support user fences */
647 		if (ib->ring->type == AMDGPU_RING_TYPE_UVD ||
648 		    ib->ring->type == AMDGPU_RING_TYPE_VCE)
649 			return -EINVAL;
650 
651 		ib->user = &parser->uf;
652 	}
653 
654 	return 0;
655 }
656 
657 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
658 {
659 	struct amdgpu_device *adev = dev->dev_private;
660 	union drm_amdgpu_cs *cs = data;
661 	struct amdgpu_cs_parser parser;
662 	int r, i;
663 	bool reserved_buffers = false;
664 
665 	down_read(&adev->exclusive_lock);
666 	if (!adev->accel_working) {
667 		up_read(&adev->exclusive_lock);
668 		return -EBUSY;
669 	}
670 	/* initialize parser */
671 	memset(&parser, 0, sizeof(struct amdgpu_cs_parser));
672 	parser.filp = filp;
673 	parser.adev = adev;
674 	r = amdgpu_cs_parser_init(&parser, data);
675 	if (r) {
676 		DRM_ERROR("Failed to initialize parser !\n");
677 		amdgpu_cs_parser_fini(&parser, r, false);
678 		up_read(&adev->exclusive_lock);
679 		r = amdgpu_cs_handle_lockup(adev, r);
680 		return r;
681 	}
682 
683 	r = amdgpu_cs_parser_relocs(&parser);
684 	if (r) {
685 		if (r != -ERESTARTSYS) {
686 			if (r == -ENOMEM)
687 				DRM_ERROR("Not enough memory for command submission!\n");
688 			else
689 				DRM_ERROR("Failed to process the buffer list %d!\n", r);
690 		}
691 	} else {
692 		reserved_buffers = true;
693 		r = amdgpu_cs_ib_fill(adev, &parser);
694 	}
695 
696 	if (r) {
697 		amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
698 		up_read(&adev->exclusive_lock);
699 		r = amdgpu_cs_handle_lockup(adev, r);
700 		return r;
701 	}
702 
703 	for (i = 0; i < parser.num_ibs; i++)
704 		trace_amdgpu_cs(&parser, i);
705 
706 	r = amdgpu_cs_ib_vm_chunk(adev, &parser);
707 	if (r) {
708 		goto out;
709 	}
710 
711 	cs->out.handle = parser.ibs[parser.num_ibs - 1].fence->seq;
712 out:
713 	amdgpu_cs_parser_fini(&parser, r, true);
714 	up_read(&adev->exclusive_lock);
715 	r = amdgpu_cs_handle_lockup(adev, r);
716 	return r;
717 }
718 
719 /**
720  * amdgpu_cs_wait_ioctl - wait for a command submission to finish
721  *
722  * @dev: drm device
723  * @data: data from userspace
724  * @filp: file private
725  *
726  * Wait for the command submission identified by handle to finish.
727  */
728 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
729 			 struct drm_file *filp)
730 {
731 	union drm_amdgpu_wait_cs *wait = data;
732 	struct amdgpu_device *adev = dev->dev_private;
733 	uint64_t seq[AMDGPU_MAX_RINGS] = {0};
734 	struct amdgpu_ring *ring = NULL;
735 	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
736 	struct amdgpu_ctx *ctx;
737 	long r;
738 
739 	ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
740 	if (ctx == NULL)
741 		return -EINVAL;
742 
743 	r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance,
744 			       wait->in.ring, &ring);
745 	if (r)
746 		return r;
747 
748 	seq[ring->idx] = wait->in.handle;
749 
750 	r = amdgpu_fence_wait_seq_timeout(adev, seq, true, timeout);
751 	amdgpu_ctx_put(ctx);
752 	if (r < 0)
753 		return r;
754 
755 	memset(wait, 0, sizeof(*wait));
756 	wait->out.status = (r == 0);
757 
758 	return 0;
759 }
760 
761 /**
762  * amdgpu_cs_find_bo_va - find bo_va for VM address
763  *
764  * @parser: command submission parser context
765  * @addr: VM address
766  * @bo: resulting BO of the mapping found
767  *
768  * Search the buffer objects in the command submission context for a certain
769  * virtual memory address. Returns allocation structure when found, NULL
770  * otherwise.
771  */
772 struct amdgpu_bo_va_mapping *
773 amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
774 		       uint64_t addr, struct amdgpu_bo **bo)
775 {
776 	struct amdgpu_bo_list_entry *reloc;
777 	struct amdgpu_bo_va_mapping *mapping;
778 
779 	addr /= AMDGPU_GPU_PAGE_SIZE;
780 
781 	list_for_each_entry(reloc, &parser->validated, tv.head) {
782 		if (!reloc->bo_va)
783 			continue;
784 
785 		list_for_each_entry(mapping, &reloc->bo_va->mappings, list) {
786 			if (mapping->it.start > addr ||
787 			    addr > mapping->it.last)
788 				continue;
789 
790 			*bo = reloc->bo_va->bo;
791 			return mapping;
792 		}
793 	}
794 
795 	return NULL;
796 }
797