xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c (revision a8fe58cec351c25e09c393bf46117c0c47b5a17c)
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 int amdgpu_cs_get_ring(struct amdgpu_device *adev, u32 ip_type,
34 		       u32 ip_instance, u32 ring,
35 		       struct amdgpu_ring **out_ring)
36 {
37 	/* Right now all IPs have only one instance - multiple rings. */
38 	if (ip_instance != 0) {
39 		DRM_ERROR("invalid ip instance: %d\n", ip_instance);
40 		return -EINVAL;
41 	}
42 
43 	switch (ip_type) {
44 	default:
45 		DRM_ERROR("unknown ip type: %d\n", ip_type);
46 		return -EINVAL;
47 	case AMDGPU_HW_IP_GFX:
48 		if (ring < adev->gfx.num_gfx_rings) {
49 			*out_ring = &adev->gfx.gfx_ring[ring];
50 		} else {
51 			DRM_ERROR("only %d gfx rings are supported now\n",
52 				  adev->gfx.num_gfx_rings);
53 			return -EINVAL;
54 		}
55 		break;
56 	case AMDGPU_HW_IP_COMPUTE:
57 		if (ring < adev->gfx.num_compute_rings) {
58 			*out_ring = &adev->gfx.compute_ring[ring];
59 		} else {
60 			DRM_ERROR("only %d compute rings are supported now\n",
61 				  adev->gfx.num_compute_rings);
62 			return -EINVAL;
63 		}
64 		break;
65 	case AMDGPU_HW_IP_DMA:
66 		if (ring < adev->sdma.num_instances) {
67 			*out_ring = &adev->sdma.instance[ring].ring;
68 		} else {
69 			DRM_ERROR("only %d SDMA rings are supported\n",
70 				  adev->sdma.num_instances);
71 			return -EINVAL;
72 		}
73 		break;
74 	case AMDGPU_HW_IP_UVD:
75 		*out_ring = &adev->uvd.ring;
76 		break;
77 	case AMDGPU_HW_IP_VCE:
78 		if (ring < 2){
79 			*out_ring = &adev->vce.ring[ring];
80 		} else {
81 			DRM_ERROR("only two VCE rings are supported\n");
82 			return -EINVAL;
83 		}
84 		break;
85 	}
86 	return 0;
87 }
88 
89 static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
90 				      struct drm_amdgpu_cs_chunk_fence *fence_data)
91 {
92 	struct drm_gem_object *gobj;
93 	uint32_t handle;
94 
95 	handle = fence_data->handle;
96 	gobj = drm_gem_object_lookup(p->adev->ddev, p->filp,
97 				     fence_data->handle);
98 	if (gobj == NULL)
99 		return -EINVAL;
100 
101 	p->uf.bo = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
102 	p->uf.offset = fence_data->offset;
103 
104 	if (amdgpu_ttm_tt_has_userptr(p->uf.bo->tbo.ttm)) {
105 		drm_gem_object_unreference_unlocked(gobj);
106 		return -EINVAL;
107 	}
108 
109 	p->uf_entry.robj = amdgpu_bo_ref(p->uf.bo);
110 	p->uf_entry.priority = 0;
111 	p->uf_entry.tv.bo = &p->uf_entry.robj->tbo;
112 	p->uf_entry.tv.shared = true;
113 
114 	drm_gem_object_unreference_unlocked(gobj);
115 	return 0;
116 }
117 
118 int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
119 {
120 	union drm_amdgpu_cs *cs = data;
121 	uint64_t *chunk_array_user;
122 	uint64_t *chunk_array;
123 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
124 	unsigned size;
125 	int i;
126 	int ret;
127 
128 	if (cs->in.num_chunks == 0)
129 		return 0;
130 
131 	chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
132 	if (!chunk_array)
133 		return -ENOMEM;
134 
135 	p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
136 	if (!p->ctx) {
137 		ret = -EINVAL;
138 		goto free_chunk;
139 	}
140 
141 	/* get chunks */
142 	chunk_array_user = (uint64_t __user *)(unsigned long)(cs->in.chunks);
143 	if (copy_from_user(chunk_array, chunk_array_user,
144 			   sizeof(uint64_t)*cs->in.num_chunks)) {
145 		ret = -EFAULT;
146 		goto put_ctx;
147 	}
148 
149 	p->nchunks = cs->in.num_chunks;
150 	p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
151 			    GFP_KERNEL);
152 	if (!p->chunks) {
153 		ret = -ENOMEM;
154 		goto put_ctx;
155 	}
156 
157 	for (i = 0; i < p->nchunks; i++) {
158 		struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
159 		struct drm_amdgpu_cs_chunk user_chunk;
160 		uint32_t __user *cdata;
161 
162 		chunk_ptr = (void __user *)(unsigned long)chunk_array[i];
163 		if (copy_from_user(&user_chunk, chunk_ptr,
164 				       sizeof(struct drm_amdgpu_cs_chunk))) {
165 			ret = -EFAULT;
166 			i--;
167 			goto free_partial_kdata;
168 		}
169 		p->chunks[i].chunk_id = user_chunk.chunk_id;
170 		p->chunks[i].length_dw = user_chunk.length_dw;
171 
172 		size = p->chunks[i].length_dw;
173 		cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
174 
175 		p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t));
176 		if (p->chunks[i].kdata == NULL) {
177 			ret = -ENOMEM;
178 			i--;
179 			goto free_partial_kdata;
180 		}
181 		size *= sizeof(uint32_t);
182 		if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
183 			ret = -EFAULT;
184 			goto free_partial_kdata;
185 		}
186 
187 		switch (p->chunks[i].chunk_id) {
188 		case AMDGPU_CHUNK_ID_IB:
189 			p->num_ibs++;
190 			break;
191 
192 		case AMDGPU_CHUNK_ID_FENCE:
193 			size = sizeof(struct drm_amdgpu_cs_chunk_fence);
194 			if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
195 				ret = -EINVAL;
196 				goto free_partial_kdata;
197 			}
198 
199 			ret = amdgpu_cs_user_fence_chunk(p, (void *)p->chunks[i].kdata);
200 			if (ret)
201 				goto free_partial_kdata;
202 
203 			break;
204 
205 		case AMDGPU_CHUNK_ID_DEPENDENCIES:
206 			break;
207 
208 		default:
209 			ret = -EINVAL;
210 			goto free_partial_kdata;
211 		}
212 	}
213 
214 
215 	p->ibs = kcalloc(p->num_ibs, sizeof(struct amdgpu_ib), GFP_KERNEL);
216 	if (!p->ibs) {
217 		ret = -ENOMEM;
218 		goto free_all_kdata;
219 	}
220 
221 	kfree(chunk_array);
222 	return 0;
223 
224 free_all_kdata:
225 	i = p->nchunks - 1;
226 free_partial_kdata:
227 	for (; i >= 0; i--)
228 		drm_free_large(p->chunks[i].kdata);
229 	kfree(p->chunks);
230 put_ctx:
231 	amdgpu_ctx_put(p->ctx);
232 free_chunk:
233 	kfree(chunk_array);
234 
235 	return ret;
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 			    struct list_head *validated)
292 {
293 	struct amdgpu_bo_list_entry *lobj;
294 	u64 initial_bytes_moved;
295 	int r;
296 
297 	list_for_each_entry(lobj, validated, tv.head) {
298 		struct amdgpu_bo *bo = lobj->robj;
299 		uint32_t domain;
300 
301 		if (bo->pin_count)
302 			continue;
303 
304 		/* Avoid moving this one if we have moved too many buffers
305 		 * for this IB already.
306 		 *
307 		 * Note that this allows moving at least one buffer of
308 		 * any size, because it doesn't take the current "bo"
309 		 * into account. We don't want to disallow buffer moves
310 		 * completely.
311 		 */
312 		if (p->bytes_moved <= p->bytes_moved_threshold)
313 			domain = bo->prefered_domains;
314 		else
315 			domain = bo->allowed_domains;
316 
317 	retry:
318 		amdgpu_ttm_placement_from_domain(bo, domain);
319 		initial_bytes_moved = atomic64_read(&bo->adev->num_bytes_moved);
320 		r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
321 		p->bytes_moved += atomic64_read(&bo->adev->num_bytes_moved) -
322 			       initial_bytes_moved;
323 
324 		if (unlikely(r)) {
325 			if (r != -ERESTARTSYS && domain != bo->allowed_domains) {
326 				domain = bo->allowed_domains;
327 				goto retry;
328 			}
329 			return r;
330 		}
331 	}
332 	return 0;
333 }
334 
335 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
336 				union drm_amdgpu_cs *cs)
337 {
338 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
339 	struct list_head duplicates;
340 	bool need_mmap_lock = false;
341 	int r;
342 
343 	INIT_LIST_HEAD(&p->validated);
344 
345 	p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
346 	if (p->bo_list) {
347 		need_mmap_lock = p->bo_list->has_userptr;
348 		amdgpu_bo_list_get_list(p->bo_list, &p->validated);
349 	}
350 
351 	INIT_LIST_HEAD(&duplicates);
352 	amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
353 
354 	if (p->uf.bo)
355 		list_add(&p->uf_entry.tv.head, &p->validated);
356 
357 	if (need_mmap_lock)
358 		down_read(&current->mm->mmap_sem);
359 
360 	r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true, &duplicates);
361 	if (unlikely(r != 0))
362 		goto error_reserve;
363 
364 	amdgpu_vm_get_pt_bos(&fpriv->vm, &duplicates);
365 
366 	p->bytes_moved_threshold = amdgpu_cs_get_threshold_for_moves(p->adev);
367 	p->bytes_moved = 0;
368 
369 	r = amdgpu_cs_list_validate(p, &duplicates);
370 	if (r)
371 		goto error_validate;
372 
373 	r = amdgpu_cs_list_validate(p, &p->validated);
374 	if (r)
375 		goto error_validate;
376 
377 	if (p->bo_list) {
378 		struct amdgpu_vm *vm = &fpriv->vm;
379 		unsigned i;
380 
381 		for (i = 0; i < p->bo_list->num_entries; i++) {
382 			struct amdgpu_bo *bo = p->bo_list->array[i].robj;
383 
384 			p->bo_list->array[i].bo_va = amdgpu_vm_bo_find(vm, bo);
385 		}
386 	}
387 
388 error_validate:
389 	if (r) {
390 		amdgpu_vm_move_pt_bos_in_lru(p->adev, &fpriv->vm);
391 		ttm_eu_backoff_reservation(&p->ticket, &p->validated);
392 	}
393 
394 error_reserve:
395 	if (need_mmap_lock)
396 		up_read(&current->mm->mmap_sem);
397 
398 	return r;
399 }
400 
401 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
402 {
403 	struct amdgpu_bo_list_entry *e;
404 	int r;
405 
406 	list_for_each_entry(e, &p->validated, tv.head) {
407 		struct reservation_object *resv = e->robj->tbo.resv;
408 		r = amdgpu_sync_resv(p->adev, &p->ibs[0].sync, resv, p->filp);
409 
410 		if (r)
411 			return r;
412 	}
413 	return 0;
414 }
415 
416 static int cmp_size_smaller_first(void *priv, struct list_head *a,
417 				  struct list_head *b)
418 {
419 	struct amdgpu_bo_list_entry *la = list_entry(a, struct amdgpu_bo_list_entry, tv.head);
420 	struct amdgpu_bo_list_entry *lb = list_entry(b, struct amdgpu_bo_list_entry, tv.head);
421 
422 	/* Sort A before B if A is smaller. */
423 	return (int)la->robj->tbo.num_pages - (int)lb->robj->tbo.num_pages;
424 }
425 
426 /**
427  * cs_parser_fini() - clean parser states
428  * @parser:	parser structure holding parsing context.
429  * @error:	error number
430  *
431  * If error is set than unvalidate buffer, otherwise just free memory
432  * used by parsing context.
433  **/
434 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bool backoff)
435 {
436 	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
437 	unsigned i;
438 
439 	if (!error) {
440 		amdgpu_vm_move_pt_bos_in_lru(parser->adev, &fpriv->vm);
441 
442 		/* Sort the buffer list from the smallest to largest buffer,
443 		 * which affects the order of buffers in the LRU list.
444 		 * This assures that the smallest buffers are added first
445 		 * to the LRU list, so they are likely to be later evicted
446 		 * first, instead of large buffers whose eviction is more
447 		 * expensive.
448 		 *
449 		 * This slightly lowers the number of bytes moved by TTM
450 		 * per frame under memory pressure.
451 		 */
452 		list_sort(NULL, &parser->validated, cmp_size_smaller_first);
453 
454 		ttm_eu_fence_buffer_objects(&parser->ticket,
455 					    &parser->validated,
456 					    parser->fence);
457 	} else if (backoff) {
458 		ttm_eu_backoff_reservation(&parser->ticket,
459 					   &parser->validated);
460 	}
461 	fence_put(parser->fence);
462 
463 	if (parser->ctx)
464 		amdgpu_ctx_put(parser->ctx);
465 	if (parser->bo_list)
466 		amdgpu_bo_list_put(parser->bo_list);
467 
468 	for (i = 0; i < parser->nchunks; i++)
469 		drm_free_large(parser->chunks[i].kdata);
470 	kfree(parser->chunks);
471 	if (parser->ibs)
472 		for (i = 0; i < parser->num_ibs; i++)
473 			amdgpu_ib_free(parser->adev, &parser->ibs[i]);
474 	kfree(parser->ibs);
475 	amdgpu_bo_unref(&parser->uf.bo);
476 	amdgpu_bo_unref(&parser->uf_entry.robj);
477 }
478 
479 static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p,
480 				   struct amdgpu_vm *vm)
481 {
482 	struct amdgpu_device *adev = p->adev;
483 	struct amdgpu_bo_va *bo_va;
484 	struct amdgpu_bo *bo;
485 	int i, r;
486 
487 	r = amdgpu_vm_update_page_directory(adev, vm);
488 	if (r)
489 		return r;
490 
491 	r = amdgpu_sync_fence(adev, &p->ibs[0].sync, vm->page_directory_fence);
492 	if (r)
493 		return r;
494 
495 	r = amdgpu_vm_clear_freed(adev, vm);
496 	if (r)
497 		return r;
498 
499 	if (p->bo_list) {
500 		for (i = 0; i < p->bo_list->num_entries; i++) {
501 			struct fence *f;
502 
503 			/* ignore duplicates */
504 			bo = p->bo_list->array[i].robj;
505 			if (!bo)
506 				continue;
507 
508 			bo_va = p->bo_list->array[i].bo_va;
509 			if (bo_va == NULL)
510 				continue;
511 
512 			r = amdgpu_vm_bo_update(adev, bo_va, &bo->tbo.mem);
513 			if (r)
514 				return r;
515 
516 			f = bo_va->last_pt_update;
517 			r = amdgpu_sync_fence(adev, &p->ibs[0].sync, f);
518 			if (r)
519 				return r;
520 		}
521 
522 	}
523 
524 	r = amdgpu_vm_clear_invalids(adev, vm, &p->ibs[0].sync);
525 
526 	if (amdgpu_vm_debug && p->bo_list) {
527 		/* Invalidate all BOs to test for userspace bugs */
528 		for (i = 0; i < p->bo_list->num_entries; i++) {
529 			/* ignore duplicates */
530 			bo = p->bo_list->array[i].robj;
531 			if (!bo)
532 				continue;
533 
534 			amdgpu_vm_bo_invalidate(adev, bo);
535 		}
536 	}
537 
538 	return r;
539 }
540 
541 static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
542 				 struct amdgpu_cs_parser *parser)
543 {
544 	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
545 	struct amdgpu_vm *vm = &fpriv->vm;
546 	struct amdgpu_ring *ring;
547 	int i, r;
548 
549 	if (parser->num_ibs == 0)
550 		return 0;
551 
552 	/* Only for UVD/VCE VM emulation */
553 	for (i = 0; i < parser->num_ibs; i++) {
554 		ring = parser->ibs[i].ring;
555 		if (ring->funcs->parse_cs) {
556 			r = amdgpu_ring_parse_cs(ring, parser, i);
557 			if (r)
558 				return r;
559 		}
560 	}
561 
562 	r = amdgpu_bo_vm_update_pte(parser, vm);
563 	if (!r)
564 		amdgpu_cs_sync_rings(parser);
565 
566 	return r;
567 }
568 
569 static int amdgpu_cs_handle_lockup(struct amdgpu_device *adev, int r)
570 {
571 	if (r == -EDEADLK) {
572 		r = amdgpu_gpu_reset(adev);
573 		if (!r)
574 			r = -EAGAIN;
575 	}
576 	return r;
577 }
578 
579 static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
580 			     struct amdgpu_cs_parser *parser)
581 {
582 	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
583 	struct amdgpu_vm *vm = &fpriv->vm;
584 	int i, j;
585 	int r;
586 
587 	for (i = 0, j = 0; i < parser->nchunks && j < parser->num_ibs; i++) {
588 		struct amdgpu_cs_chunk *chunk;
589 		struct amdgpu_ib *ib;
590 		struct drm_amdgpu_cs_chunk_ib *chunk_ib;
591 		struct amdgpu_ring *ring;
592 
593 		chunk = &parser->chunks[i];
594 		ib = &parser->ibs[j];
595 		chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
596 
597 		if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
598 			continue;
599 
600 		r = amdgpu_cs_get_ring(adev, chunk_ib->ip_type,
601 				       chunk_ib->ip_instance, chunk_ib->ring,
602 				       &ring);
603 		if (r)
604 			return r;
605 
606 		if (ring->funcs->parse_cs) {
607 			struct amdgpu_bo_va_mapping *m;
608 			struct amdgpu_bo *aobj = NULL;
609 			uint64_t offset;
610 			uint8_t *kptr;
611 
612 			m = amdgpu_cs_find_mapping(parser, chunk_ib->va_start,
613 						   &aobj);
614 			if (!aobj) {
615 				DRM_ERROR("IB va_start is invalid\n");
616 				return -EINVAL;
617 			}
618 
619 			if ((chunk_ib->va_start + chunk_ib->ib_bytes) >
620 			    (m->it.last + 1) * AMDGPU_GPU_PAGE_SIZE) {
621 				DRM_ERROR("IB va_start+ib_bytes is invalid\n");
622 				return -EINVAL;
623 			}
624 
625 			/* the IB should be reserved at this point */
626 			r = amdgpu_bo_kmap(aobj, (void **)&kptr);
627 			if (r) {
628 				return r;
629 			}
630 
631 			offset = ((uint64_t)m->it.start) * AMDGPU_GPU_PAGE_SIZE;
632 			kptr += chunk_ib->va_start - offset;
633 
634 			r =  amdgpu_ib_get(ring, NULL, chunk_ib->ib_bytes, ib);
635 			if (r) {
636 				DRM_ERROR("Failed to get ib !\n");
637 				return r;
638 			}
639 
640 			memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
641 			amdgpu_bo_kunmap(aobj);
642 		} else {
643 			r =  amdgpu_ib_get(ring, vm, 0, ib);
644 			if (r) {
645 				DRM_ERROR("Failed to get ib !\n");
646 				return r;
647 			}
648 
649 			ib->gpu_addr = chunk_ib->va_start;
650 		}
651 
652 		ib->length_dw = chunk_ib->ib_bytes / 4;
653 		ib->flags = chunk_ib->flags;
654 		ib->ctx = parser->ctx;
655 		j++;
656 	}
657 
658 	if (!parser->num_ibs)
659 		return 0;
660 
661 	/* add GDS resources to first IB */
662 	if (parser->bo_list) {
663 		struct amdgpu_bo *gds = parser->bo_list->gds_obj;
664 		struct amdgpu_bo *gws = parser->bo_list->gws_obj;
665 		struct amdgpu_bo *oa = parser->bo_list->oa_obj;
666 		struct amdgpu_ib *ib = &parser->ibs[0];
667 
668 		if (gds) {
669 			ib->gds_base = amdgpu_bo_gpu_offset(gds);
670 			ib->gds_size = amdgpu_bo_size(gds);
671 		}
672 		if (gws) {
673 			ib->gws_base = amdgpu_bo_gpu_offset(gws);
674 			ib->gws_size = amdgpu_bo_size(gws);
675 		}
676 		if (oa) {
677 			ib->oa_base = amdgpu_bo_gpu_offset(oa);
678 			ib->oa_size = amdgpu_bo_size(oa);
679 		}
680 	}
681 	/* wrap the last IB with user fence */
682 	if (parser->uf.bo) {
683 		struct amdgpu_ib *ib = &parser->ibs[parser->num_ibs - 1];
684 
685 		/* UVD & VCE fw doesn't support user fences */
686 		if (ib->ring->type == AMDGPU_RING_TYPE_UVD ||
687 		    ib->ring->type == AMDGPU_RING_TYPE_VCE)
688 			return -EINVAL;
689 
690 		ib->user = &parser->uf;
691 	}
692 
693 	return 0;
694 }
695 
696 static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
697 				  struct amdgpu_cs_parser *p)
698 {
699 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
700 	struct amdgpu_ib *ib;
701 	int i, j, r;
702 
703 	if (!p->num_ibs)
704 		return 0;
705 
706 	/* Add dependencies to first IB */
707 	ib = &p->ibs[0];
708 	for (i = 0; i < p->nchunks; ++i) {
709 		struct drm_amdgpu_cs_chunk_dep *deps;
710 		struct amdgpu_cs_chunk *chunk;
711 		unsigned num_deps;
712 
713 		chunk = &p->chunks[i];
714 
715 		if (chunk->chunk_id != AMDGPU_CHUNK_ID_DEPENDENCIES)
716 			continue;
717 
718 		deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
719 		num_deps = chunk->length_dw * 4 /
720 			sizeof(struct drm_amdgpu_cs_chunk_dep);
721 
722 		for (j = 0; j < num_deps; ++j) {
723 			struct amdgpu_ring *ring;
724 			struct amdgpu_ctx *ctx;
725 			struct fence *fence;
726 
727 			r = amdgpu_cs_get_ring(adev, deps[j].ip_type,
728 					       deps[j].ip_instance,
729 					       deps[j].ring, &ring);
730 			if (r)
731 				return r;
732 
733 			ctx = amdgpu_ctx_get(fpriv, deps[j].ctx_id);
734 			if (ctx == NULL)
735 				return -EINVAL;
736 
737 			fence = amdgpu_ctx_get_fence(ctx, ring,
738 						     deps[j].handle);
739 			if (IS_ERR(fence)) {
740 				r = PTR_ERR(fence);
741 				amdgpu_ctx_put(ctx);
742 				return r;
743 
744 			} else if (fence) {
745 				r = amdgpu_sync_fence(adev, &ib->sync, fence);
746 				fence_put(fence);
747 				amdgpu_ctx_put(ctx);
748 				if (r)
749 					return r;
750 			}
751 		}
752 	}
753 
754 	return 0;
755 }
756 
757 static int amdgpu_cs_free_job(struct amdgpu_job *job)
758 {
759 	int i;
760 	if (job->ibs)
761 		for (i = 0; i < job->num_ibs; i++)
762 			amdgpu_ib_free(job->adev, &job->ibs[i]);
763 	kfree(job->ibs);
764 	if (job->uf.bo)
765 		amdgpu_bo_unref(&job->uf.bo);
766 	return 0;
767 }
768 
769 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
770 {
771 	struct amdgpu_device *adev = dev->dev_private;
772 	union drm_amdgpu_cs *cs = data;
773 	struct amdgpu_cs_parser parser = {};
774 	bool reserved_buffers = false;
775 	int i, r;
776 
777 	if (!adev->accel_working)
778 		return -EBUSY;
779 
780 	parser.adev = adev;
781 	parser.filp = filp;
782 
783 	r = amdgpu_cs_parser_init(&parser, data);
784 	if (r) {
785 		DRM_ERROR("Failed to initialize parser !\n");
786 		amdgpu_cs_parser_fini(&parser, r, false);
787 		r = amdgpu_cs_handle_lockup(adev, r);
788 		return r;
789 	}
790 	r = amdgpu_cs_parser_bos(&parser, data);
791 	if (r == -ENOMEM)
792 		DRM_ERROR("Not enough memory for command submission!\n");
793 	else if (r && r != -ERESTARTSYS)
794 		DRM_ERROR("Failed to process the buffer list %d!\n", r);
795 	else if (!r) {
796 		reserved_buffers = true;
797 		r = amdgpu_cs_ib_fill(adev, &parser);
798 	}
799 
800 	if (!r) {
801 		r = amdgpu_cs_dependencies(adev, &parser);
802 		if (r)
803 			DRM_ERROR("Failed in the dependencies handling %d!\n", r);
804 	}
805 
806 	if (r)
807 		goto out;
808 
809 	for (i = 0; i < parser.num_ibs; i++)
810 		trace_amdgpu_cs(&parser, i);
811 
812 	r = amdgpu_cs_ib_vm_chunk(adev, &parser);
813 	if (r)
814 		goto out;
815 
816 	if (parser.num_ibs) {
817 		struct amdgpu_ring * ring = parser.ibs->ring;
818 		struct amd_sched_fence *fence;
819 		struct amdgpu_job *job;
820 
821 		job = kzalloc(sizeof(struct amdgpu_job), GFP_KERNEL);
822 		if (!job) {
823 			r = -ENOMEM;
824 			goto out;
825 		}
826 
827 		job->base.sched = &ring->sched;
828 		job->base.s_entity = &parser.ctx->rings[ring->idx].entity;
829 		job->adev = parser.adev;
830 		job->owner = parser.filp;
831 		job->free_job = amdgpu_cs_free_job;
832 
833 		job->ibs = parser.ibs;
834 		job->num_ibs = parser.num_ibs;
835 		parser.ibs = NULL;
836 		parser.num_ibs = 0;
837 
838 		if (job->ibs[job->num_ibs - 1].user) {
839 			job->uf = parser.uf;
840 			job->ibs[job->num_ibs - 1].user = &job->uf;
841 			parser.uf.bo = NULL;
842 		}
843 
844 		fence = amd_sched_fence_create(job->base.s_entity,
845 					       parser.filp);
846 		if (!fence) {
847 			r = -ENOMEM;
848 			amdgpu_cs_free_job(job);
849 			kfree(job);
850 			goto out;
851 		}
852 		job->base.s_fence = fence;
853 		parser.fence = fence_get(&fence->base);
854 
855 		cs->out.handle = amdgpu_ctx_add_fence(parser.ctx, ring,
856 						      &fence->base);
857 		job->ibs[job->num_ibs - 1].sequence = cs->out.handle;
858 
859 		trace_amdgpu_cs_ioctl(job);
860 		amd_sched_entity_push_job(&job->base);
861 	}
862 
863 out:
864 	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
865 	r = amdgpu_cs_handle_lockup(adev, r);
866 	return r;
867 }
868 
869 /**
870  * amdgpu_cs_wait_ioctl - wait for a command submission to finish
871  *
872  * @dev: drm device
873  * @data: data from userspace
874  * @filp: file private
875  *
876  * Wait for the command submission identified by handle to finish.
877  */
878 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
879 			 struct drm_file *filp)
880 {
881 	union drm_amdgpu_wait_cs *wait = data;
882 	struct amdgpu_device *adev = dev->dev_private;
883 	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
884 	struct amdgpu_ring *ring = NULL;
885 	struct amdgpu_ctx *ctx;
886 	struct fence *fence;
887 	long r;
888 
889 	r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance,
890 			       wait->in.ring, &ring);
891 	if (r)
892 		return r;
893 
894 	ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
895 	if (ctx == NULL)
896 		return -EINVAL;
897 
898 	fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
899 	if (IS_ERR(fence))
900 		r = PTR_ERR(fence);
901 	else if (fence) {
902 		r = fence_wait_timeout(fence, true, timeout);
903 		fence_put(fence);
904 	} else
905 		r = 1;
906 
907 	amdgpu_ctx_put(ctx);
908 	if (r < 0)
909 		return r;
910 
911 	memset(wait, 0, sizeof(*wait));
912 	wait->out.status = (r == 0);
913 
914 	return 0;
915 }
916 
917 /**
918  * amdgpu_cs_find_bo_va - find bo_va for VM address
919  *
920  * @parser: command submission parser context
921  * @addr: VM address
922  * @bo: resulting BO of the mapping found
923  *
924  * Search the buffer objects in the command submission context for a certain
925  * virtual memory address. Returns allocation structure when found, NULL
926  * otherwise.
927  */
928 struct amdgpu_bo_va_mapping *
929 amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
930 		       uint64_t addr, struct amdgpu_bo **bo)
931 {
932 	struct amdgpu_bo_va_mapping *mapping;
933 	unsigned i;
934 
935 	if (!parser->bo_list)
936 		return NULL;
937 
938 	addr /= AMDGPU_GPU_PAGE_SIZE;
939 
940 	for (i = 0; i < parser->bo_list->num_entries; i++) {
941 		struct amdgpu_bo_list_entry *lobj;
942 
943 		lobj = &parser->bo_list->array[i];
944 		if (!lobj->bo_va)
945 			continue;
946 
947 		list_for_each_entry(mapping, &lobj->bo_va->valids, list) {
948 			if (mapping->it.start > addr ||
949 			    addr > mapping->it.last)
950 				continue;
951 
952 			*bo = lobj->bo_va->bo;
953 			return mapping;
954 		}
955 
956 		list_for_each_entry(mapping, &lobj->bo_va->invalids, list) {
957 			if (mapping->it.start > addr ||
958 			    addr > mapping->it.last)
959 				continue;
960 
961 			*bo = lobj->bo_va->bo;
962 			return mapping;
963 		}
964 	}
965 
966 	return NULL;
967 }
968