xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c (revision e1a0ebc8d82b64440d3ca7eac6a8489937ee2519)
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 < adev->sdma.num_instances) {
108 			*out_ring = &adev->sdma.instance[ring].ring;
109 		} else {
110 			DRM_ERROR("only %d SDMA rings are supported\n",
111 				  adev->sdma.num_instances);
112 			return -EINVAL;
113 		}
114 		break;
115 	case AMDGPU_HW_IP_UVD:
116 		*out_ring = &adev->uvd.ring;
117 		break;
118 	case AMDGPU_HW_IP_VCE:
119 		if (ring < 2){
120 			*out_ring = &adev->vce.ring[ring];
121 		} else {
122 			DRM_ERROR("only two VCE rings are supported\n");
123 			return -EINVAL;
124 		}
125 		break;
126 	}
127 	return 0;
128 }
129 
130 int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
131 {
132 	union drm_amdgpu_cs *cs = data;
133 	uint64_t *chunk_array_user;
134 	uint64_t *chunk_array;
135 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
136 	unsigned size;
137 	int i;
138 	int ret;
139 
140 	if (cs->in.num_chunks == 0)
141 		return 0;
142 
143 	chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
144 	if (!chunk_array)
145 		return -ENOMEM;
146 
147 	p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
148 	if (!p->ctx) {
149 		ret = -EINVAL;
150 		goto free_chunk;
151 	}
152 
153 	p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
154 
155 	/* get chunks */
156 	INIT_LIST_HEAD(&p->validated);
157 	chunk_array_user = (uint64_t __user *)(unsigned long)(cs->in.chunks);
158 	if (copy_from_user(chunk_array, chunk_array_user,
159 			   sizeof(uint64_t)*cs->in.num_chunks)) {
160 		ret = -EFAULT;
161 		goto put_bo_list;
162 	}
163 
164 	p->nchunks = cs->in.num_chunks;
165 	p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
166 			    GFP_KERNEL);
167 	if (!p->chunks) {
168 		ret = -ENOMEM;
169 		goto put_bo_list;
170 	}
171 
172 	for (i = 0; i < p->nchunks; i++) {
173 		struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
174 		struct drm_amdgpu_cs_chunk user_chunk;
175 		uint32_t __user *cdata;
176 
177 		chunk_ptr = (void __user *)(unsigned long)chunk_array[i];
178 		if (copy_from_user(&user_chunk, chunk_ptr,
179 				       sizeof(struct drm_amdgpu_cs_chunk))) {
180 			ret = -EFAULT;
181 			i--;
182 			goto free_partial_kdata;
183 		}
184 		p->chunks[i].chunk_id = user_chunk.chunk_id;
185 		p->chunks[i].length_dw = user_chunk.length_dw;
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 			ret = -ENOMEM;
194 			i--;
195 			goto free_partial_kdata;
196 		}
197 		size *= sizeof(uint32_t);
198 		if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
199 			ret = -EFAULT;
200 			goto free_partial_kdata;
201 		}
202 
203 		switch (p->chunks[i].chunk_id) {
204 		case AMDGPU_CHUNK_ID_IB:
205 			p->num_ibs++;
206 			break;
207 
208 		case AMDGPU_CHUNK_ID_FENCE:
209 			size = sizeof(struct drm_amdgpu_cs_chunk_fence);
210 			if (p->chunks[i].length_dw * sizeof(uint32_t) >= size) {
211 				uint32_t handle;
212 				struct drm_gem_object *gobj;
213 				struct drm_amdgpu_cs_chunk_fence *fence_data;
214 
215 				fence_data = (void *)p->chunks[i].kdata;
216 				handle = fence_data->handle;
217 				gobj = drm_gem_object_lookup(p->adev->ddev,
218 							     p->filp, handle);
219 				if (gobj == NULL) {
220 					ret = -EINVAL;
221 					goto free_partial_kdata;
222 				}
223 
224 				p->uf.bo = gem_to_amdgpu_bo(gobj);
225 				p->uf.offset = fence_data->offset;
226 			} else {
227 				ret = -EINVAL;
228 				goto free_partial_kdata;
229 			}
230 			break;
231 
232 		case AMDGPU_CHUNK_ID_DEPENDENCIES:
233 			break;
234 
235 		default:
236 			ret = -EINVAL;
237 			goto free_partial_kdata;
238 		}
239 	}
240 
241 
242 	p->ibs = kcalloc(p->num_ibs, sizeof(struct amdgpu_ib), GFP_KERNEL);
243 	if (!p->ibs) {
244 		ret = -ENOMEM;
245 		goto free_all_kdata;
246 	}
247 
248 	kfree(chunk_array);
249 	return 0;
250 
251 free_all_kdata:
252 	i = p->nchunks - 1;
253 free_partial_kdata:
254 	for (; i >= 0; i--)
255 		drm_free_large(p->chunks[i].kdata);
256 	kfree(p->chunks);
257 put_bo_list:
258 	if (p->bo_list)
259 		amdgpu_bo_list_put(p->bo_list);
260 	amdgpu_ctx_put(p->ctx);
261 free_chunk:
262 	kfree(chunk_array);
263 
264 	return ret;
265 }
266 
267 /* Returns how many bytes TTM can move per IB.
268  */
269 static u64 amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev)
270 {
271 	u64 real_vram_size = adev->mc.real_vram_size;
272 	u64 vram_usage = atomic64_read(&adev->vram_usage);
273 
274 	/* This function is based on the current VRAM usage.
275 	 *
276 	 * - If all of VRAM is free, allow relocating the number of bytes that
277 	 *   is equal to 1/4 of the size of VRAM for this IB.
278 
279 	 * - If more than one half of VRAM is occupied, only allow relocating
280 	 *   1 MB of data for this IB.
281 	 *
282 	 * - From 0 to one half of used VRAM, the threshold decreases
283 	 *   linearly.
284 	 *         __________________
285 	 * 1/4 of -|\               |
286 	 * VRAM    | \              |
287 	 *         |  \             |
288 	 *         |   \            |
289 	 *         |    \           |
290 	 *         |     \          |
291 	 *         |      \         |
292 	 *         |       \________|1 MB
293 	 *         |----------------|
294 	 *    VRAM 0 %             100 %
295 	 *         used            used
296 	 *
297 	 * Note: It's a threshold, not a limit. The threshold must be crossed
298 	 * for buffer relocations to stop, so any buffer of an arbitrary size
299 	 * can be moved as long as the threshold isn't crossed before
300 	 * the relocation takes place. We don't want to disable buffer
301 	 * relocations completely.
302 	 *
303 	 * The idea is that buffers should be placed in VRAM at creation time
304 	 * and TTM should only do a minimum number of relocations during
305 	 * command submission. In practice, you need to submit at least
306 	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
307 	 *
308 	 * Also, things can get pretty crazy under memory pressure and actual
309 	 * VRAM usage can change a lot, so playing safe even at 50% does
310 	 * consistently increase performance.
311 	 */
312 
313 	u64 half_vram = real_vram_size >> 1;
314 	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
315 	u64 bytes_moved_threshold = half_free_vram >> 1;
316 	return max(bytes_moved_threshold, 1024*1024ull);
317 }
318 
319 int amdgpu_cs_list_validate(struct amdgpu_device *adev,
320 			    struct amdgpu_vm *vm,
321 			    struct list_head *validated)
322 {
323 	struct amdgpu_bo_list_entry *lobj;
324 	struct amdgpu_bo *bo;
325 	u64 bytes_moved = 0, initial_bytes_moved;
326 	u64 bytes_moved_threshold = amdgpu_cs_get_threshold_for_moves(adev);
327 	int r;
328 
329 	list_for_each_entry(lobj, validated, tv.head) {
330 		bo = lobj->robj;
331 		if (!bo->pin_count) {
332 			u32 domain = lobj->prefered_domains;
333 			u32 current_domain =
334 				amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
335 
336 			/* Check if this buffer will be moved and don't move it
337 			 * if we have moved too many buffers for this IB already.
338 			 *
339 			 * Note that this allows moving at least one buffer of
340 			 * any size, because it doesn't take the current "bo"
341 			 * into account. We don't want to disallow buffer moves
342 			 * completely.
343 			 */
344 			if ((lobj->allowed_domains & current_domain) != 0 &&
345 			    (domain & current_domain) == 0 && /* will be moved */
346 			    bytes_moved > bytes_moved_threshold) {
347 				/* don't move it */
348 				domain = current_domain;
349 			}
350 
351 		retry:
352 			amdgpu_ttm_placement_from_domain(bo, domain);
353 			initial_bytes_moved = atomic64_read(&adev->num_bytes_moved);
354 			r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
355 			bytes_moved += atomic64_read(&adev->num_bytes_moved) -
356 				       initial_bytes_moved;
357 
358 			if (unlikely(r)) {
359 				if (r != -ERESTARTSYS && domain != lobj->allowed_domains) {
360 					domain = lobj->allowed_domains;
361 					goto retry;
362 				}
363 				return r;
364 			}
365 		}
366 		lobj->bo_va = amdgpu_vm_bo_find(vm, bo);
367 	}
368 	return 0;
369 }
370 
371 static int amdgpu_cs_parser_relocs(struct amdgpu_cs_parser *p)
372 {
373 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
374 	struct amdgpu_cs_buckets buckets;
375 	struct list_head duplicates;
376 	bool need_mmap_lock = false;
377 	int i, r;
378 
379 	if (p->bo_list) {
380 		need_mmap_lock = p->bo_list->has_userptr;
381 		amdgpu_cs_buckets_init(&buckets);
382 		for (i = 0; i < p->bo_list->num_entries; i++)
383 			amdgpu_cs_buckets_add(&buckets, &p->bo_list->array[i].tv.head,
384 								  p->bo_list->array[i].priority);
385 
386 		amdgpu_cs_buckets_get_list(&buckets, &p->validated);
387 	}
388 
389 	p->vm_bos = amdgpu_vm_get_bos(p->adev, &fpriv->vm,
390 				      &p->validated);
391 
392 	if (need_mmap_lock)
393 		down_read(&current->mm->mmap_sem);
394 
395 	INIT_LIST_HEAD(&duplicates);
396 	r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true, &duplicates);
397 	if (unlikely(r != 0))
398 		goto error_reserve;
399 
400 	r = amdgpu_cs_list_validate(p->adev, &fpriv->vm, &p->validated);
401 	if (r)
402 		goto error_validate;
403 
404 	r = amdgpu_cs_list_validate(p->adev, &fpriv->vm, &duplicates);
405 
406 error_validate:
407 	if (r)
408 		ttm_eu_backoff_reservation(&p->ticket, &p->validated);
409 
410 error_reserve:
411 	if (need_mmap_lock)
412 		up_read(&current->mm->mmap_sem);
413 
414 	return r;
415 }
416 
417 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
418 {
419 	struct amdgpu_bo_list_entry *e;
420 	int r;
421 
422 	list_for_each_entry(e, &p->validated, tv.head) {
423 		struct reservation_object *resv = e->robj->tbo.resv;
424 		r = amdgpu_sync_resv(p->adev, &p->ibs[0].sync, resv, p->filp);
425 
426 		if (r)
427 			return r;
428 	}
429 	return 0;
430 }
431 
432 static int cmp_size_smaller_first(void *priv, struct list_head *a,
433 				  struct list_head *b)
434 {
435 	struct amdgpu_bo_list_entry *la = list_entry(a, struct amdgpu_bo_list_entry, tv.head);
436 	struct amdgpu_bo_list_entry *lb = list_entry(b, struct amdgpu_bo_list_entry, tv.head);
437 
438 	/* Sort A before B if A is smaller. */
439 	return (int)la->robj->tbo.num_pages - (int)lb->robj->tbo.num_pages;
440 }
441 
442 /**
443  * cs_parser_fini() - clean parser states
444  * @parser:	parser structure holding parsing context.
445  * @error:	error number
446  *
447  * If error is set than unvalidate buffer, otherwise just free memory
448  * used by parsing context.
449  **/
450 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bool backoff)
451 {
452 	unsigned i;
453 
454 	if (!error) {
455 		/* Sort the buffer list from the smallest to largest buffer,
456 		 * which affects the order of buffers in the LRU list.
457 		 * This assures that the smallest buffers are added first
458 		 * to the LRU list, so they are likely to be later evicted
459 		 * first, instead of large buffers whose eviction is more
460 		 * expensive.
461 		 *
462 		 * This slightly lowers the number of bytes moved by TTM
463 		 * per frame under memory pressure.
464 		 */
465 		list_sort(NULL, &parser->validated, cmp_size_smaller_first);
466 
467 		ttm_eu_fence_buffer_objects(&parser->ticket,
468 					    &parser->validated,
469 					    parser->fence);
470 	} else if (backoff) {
471 		ttm_eu_backoff_reservation(&parser->ticket,
472 					   &parser->validated);
473 	}
474 	fence_put(parser->fence);
475 
476 	if (parser->ctx)
477 		amdgpu_ctx_put(parser->ctx);
478 	if (parser->bo_list)
479 		amdgpu_bo_list_put(parser->bo_list);
480 
481 	drm_free_large(parser->vm_bos);
482 	for (i = 0; i < parser->nchunks; i++)
483 		drm_free_large(parser->chunks[i].kdata);
484 	kfree(parser->chunks);
485 	if (parser->ibs)
486 		for (i = 0; i < parser->num_ibs; i++)
487 			amdgpu_ib_free(parser->adev, &parser->ibs[i]);
488 	kfree(parser->ibs);
489 	if (parser->uf.bo)
490 		drm_gem_object_unreference_unlocked(&parser->uf.bo->gem_base);
491 }
492 
493 static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p,
494 				   struct amdgpu_vm *vm)
495 {
496 	struct amdgpu_device *adev = p->adev;
497 	struct amdgpu_bo_va *bo_va;
498 	struct amdgpu_bo *bo;
499 	int i, r;
500 
501 	r = amdgpu_vm_update_page_directory(adev, vm);
502 	if (r)
503 		return r;
504 
505 	r = amdgpu_sync_fence(adev, &p->ibs[0].sync, vm->page_directory_fence);
506 	if (r)
507 		return r;
508 
509 	r = amdgpu_vm_clear_freed(adev, vm);
510 	if (r)
511 		return r;
512 
513 	if (p->bo_list) {
514 		for (i = 0; i < p->bo_list->num_entries; i++) {
515 			struct fence *f;
516 
517 			/* ignore duplicates */
518 			bo = p->bo_list->array[i].robj;
519 			if (!bo)
520 				continue;
521 
522 			bo_va = p->bo_list->array[i].bo_va;
523 			if (bo_va == NULL)
524 				continue;
525 
526 			r = amdgpu_vm_bo_update(adev, bo_va, &bo->tbo.mem);
527 			if (r)
528 				return r;
529 
530 			f = bo_va->last_pt_update;
531 			r = amdgpu_sync_fence(adev, &p->ibs[0].sync, f);
532 			if (r)
533 				return r;
534 		}
535 
536 	}
537 
538 	r = amdgpu_vm_clear_invalids(adev, vm, &p->ibs[0].sync);
539 
540 	if (amdgpu_vm_debug && p->bo_list) {
541 		/* Invalidate all BOs to test for userspace bugs */
542 		for (i = 0; i < p->bo_list->num_entries; i++) {
543 			/* ignore duplicates */
544 			bo = p->bo_list->array[i].robj;
545 			if (!bo)
546 				continue;
547 
548 			amdgpu_vm_bo_invalidate(adev, bo);
549 		}
550 	}
551 
552 	return r;
553 }
554 
555 static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
556 				 struct amdgpu_cs_parser *parser)
557 {
558 	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
559 	struct amdgpu_vm *vm = &fpriv->vm;
560 	struct amdgpu_ring *ring;
561 	int i, r;
562 
563 	if (parser->num_ibs == 0)
564 		return 0;
565 
566 	/* Only for UVD/VCE VM emulation */
567 	for (i = 0; i < parser->num_ibs; i++) {
568 		ring = parser->ibs[i].ring;
569 		if (ring->funcs->parse_cs) {
570 			r = amdgpu_ring_parse_cs(ring, parser, i);
571 			if (r)
572 				return r;
573 		}
574 	}
575 
576 	r = amdgpu_bo_vm_update_pte(parser, vm);
577 	if (!r)
578 		amdgpu_cs_sync_rings(parser);
579 
580 	return r;
581 }
582 
583 static int amdgpu_cs_handle_lockup(struct amdgpu_device *adev, int r)
584 {
585 	if (r == -EDEADLK) {
586 		r = amdgpu_gpu_reset(adev);
587 		if (!r)
588 			r = -EAGAIN;
589 	}
590 	return r;
591 }
592 
593 static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
594 			     struct amdgpu_cs_parser *parser)
595 {
596 	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
597 	struct amdgpu_vm *vm = &fpriv->vm;
598 	int i, j;
599 	int r;
600 
601 	for (i = 0, j = 0; i < parser->nchunks && j < parser->num_ibs; i++) {
602 		struct amdgpu_cs_chunk *chunk;
603 		struct amdgpu_ib *ib;
604 		struct drm_amdgpu_cs_chunk_ib *chunk_ib;
605 		struct amdgpu_ring *ring;
606 
607 		chunk = &parser->chunks[i];
608 		ib = &parser->ibs[j];
609 		chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
610 
611 		if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
612 			continue;
613 
614 		r = amdgpu_cs_get_ring(adev, chunk_ib->ip_type,
615 				       chunk_ib->ip_instance, chunk_ib->ring,
616 				       &ring);
617 		if (r)
618 			return r;
619 
620 		if (ring->funcs->parse_cs) {
621 			struct amdgpu_bo_va_mapping *m;
622 			struct amdgpu_bo *aobj = NULL;
623 			uint64_t offset;
624 			uint8_t *kptr;
625 
626 			m = amdgpu_cs_find_mapping(parser, chunk_ib->va_start,
627 						   &aobj);
628 			if (!aobj) {
629 				DRM_ERROR("IB va_start is invalid\n");
630 				return -EINVAL;
631 			}
632 
633 			if ((chunk_ib->va_start + chunk_ib->ib_bytes) >
634 			    (m->it.last + 1) * AMDGPU_GPU_PAGE_SIZE) {
635 				DRM_ERROR("IB va_start+ib_bytes is invalid\n");
636 				return -EINVAL;
637 			}
638 
639 			/* the IB should be reserved at this point */
640 			r = amdgpu_bo_kmap(aobj, (void **)&kptr);
641 			if (r) {
642 				return r;
643 			}
644 
645 			offset = ((uint64_t)m->it.start) * AMDGPU_GPU_PAGE_SIZE;
646 			kptr += chunk_ib->va_start - offset;
647 
648 			r =  amdgpu_ib_get(ring, NULL, chunk_ib->ib_bytes, ib);
649 			if (r) {
650 				DRM_ERROR("Failed to get ib !\n");
651 				return r;
652 			}
653 
654 			memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
655 			amdgpu_bo_kunmap(aobj);
656 		} else {
657 			r =  amdgpu_ib_get(ring, vm, 0, ib);
658 			if (r) {
659 				DRM_ERROR("Failed to get ib !\n");
660 				return r;
661 			}
662 
663 			ib->gpu_addr = chunk_ib->va_start;
664 		}
665 
666 		ib->length_dw = chunk_ib->ib_bytes / 4;
667 		ib->flags = chunk_ib->flags;
668 		ib->ctx = parser->ctx;
669 		j++;
670 	}
671 
672 	if (!parser->num_ibs)
673 		return 0;
674 
675 	/* add GDS resources to first IB */
676 	if (parser->bo_list) {
677 		struct amdgpu_bo *gds = parser->bo_list->gds_obj;
678 		struct amdgpu_bo *gws = parser->bo_list->gws_obj;
679 		struct amdgpu_bo *oa = parser->bo_list->oa_obj;
680 		struct amdgpu_ib *ib = &parser->ibs[0];
681 
682 		if (gds) {
683 			ib->gds_base = amdgpu_bo_gpu_offset(gds);
684 			ib->gds_size = amdgpu_bo_size(gds);
685 		}
686 		if (gws) {
687 			ib->gws_base = amdgpu_bo_gpu_offset(gws);
688 			ib->gws_size = amdgpu_bo_size(gws);
689 		}
690 		if (oa) {
691 			ib->oa_base = amdgpu_bo_gpu_offset(oa);
692 			ib->oa_size = amdgpu_bo_size(oa);
693 		}
694 	}
695 	/* wrap the last IB with user fence */
696 	if (parser->uf.bo) {
697 		struct amdgpu_ib *ib = &parser->ibs[parser->num_ibs - 1];
698 
699 		/* UVD & VCE fw doesn't support user fences */
700 		if (ib->ring->type == AMDGPU_RING_TYPE_UVD ||
701 		    ib->ring->type == AMDGPU_RING_TYPE_VCE)
702 			return -EINVAL;
703 
704 		ib->user = &parser->uf;
705 	}
706 
707 	return 0;
708 }
709 
710 static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
711 				  struct amdgpu_cs_parser *p)
712 {
713 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
714 	struct amdgpu_ib *ib;
715 	int i, j, r;
716 
717 	if (!p->num_ibs)
718 		return 0;
719 
720 	/* Add dependencies to first IB */
721 	ib = &p->ibs[0];
722 	for (i = 0; i < p->nchunks; ++i) {
723 		struct drm_amdgpu_cs_chunk_dep *deps;
724 		struct amdgpu_cs_chunk *chunk;
725 		unsigned num_deps;
726 
727 		chunk = &p->chunks[i];
728 
729 		if (chunk->chunk_id != AMDGPU_CHUNK_ID_DEPENDENCIES)
730 			continue;
731 
732 		deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
733 		num_deps = chunk->length_dw * 4 /
734 			sizeof(struct drm_amdgpu_cs_chunk_dep);
735 
736 		for (j = 0; j < num_deps; ++j) {
737 			struct amdgpu_ring *ring;
738 			struct amdgpu_ctx *ctx;
739 			struct fence *fence;
740 
741 			r = amdgpu_cs_get_ring(adev, deps[j].ip_type,
742 					       deps[j].ip_instance,
743 					       deps[j].ring, &ring);
744 			if (r)
745 				return r;
746 
747 			ctx = amdgpu_ctx_get(fpriv, deps[j].ctx_id);
748 			if (ctx == NULL)
749 				return -EINVAL;
750 
751 			fence = amdgpu_ctx_get_fence(ctx, ring,
752 						     deps[j].handle);
753 			if (IS_ERR(fence)) {
754 				r = PTR_ERR(fence);
755 				amdgpu_ctx_put(ctx);
756 				return r;
757 
758 			} else if (fence) {
759 				r = amdgpu_sync_fence(adev, &ib->sync, fence);
760 				fence_put(fence);
761 				amdgpu_ctx_put(ctx);
762 				if (r)
763 					return r;
764 			}
765 		}
766 	}
767 
768 	return 0;
769 }
770 
771 static int amdgpu_cs_free_job(struct amdgpu_job *job)
772 {
773 	int i;
774 	if (job->ibs)
775 		for (i = 0; i < job->num_ibs; i++)
776 			amdgpu_ib_free(job->adev, &job->ibs[i]);
777 	kfree(job->ibs);
778 	if (job->uf.bo)
779 		drm_gem_object_unreference_unlocked(&job->uf.bo->gem_base);
780 	return 0;
781 }
782 
783 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
784 {
785 	struct amdgpu_device *adev = dev->dev_private;
786 	union drm_amdgpu_cs *cs = data;
787 	struct amdgpu_cs_parser parser = {};
788 	bool reserved_buffers = false;
789 	int i, r;
790 
791 	if (!adev->accel_working)
792 		return -EBUSY;
793 
794 	parser.adev = adev;
795 	parser.filp = filp;
796 
797 	r = amdgpu_cs_parser_init(&parser, data);
798 	if (r) {
799 		DRM_ERROR("Failed to initialize parser !\n");
800 		amdgpu_cs_parser_fini(&parser, r, false);
801 		r = amdgpu_cs_handle_lockup(adev, r);
802 		return r;
803 	}
804 	r = amdgpu_cs_parser_relocs(&parser);
805 	if (r == -ENOMEM)
806 		DRM_ERROR("Not enough memory for command submission!\n");
807 	else if (r && r != -ERESTARTSYS)
808 		DRM_ERROR("Failed to process the buffer list %d!\n", r);
809 	else if (!r) {
810 		reserved_buffers = true;
811 		r = amdgpu_cs_ib_fill(adev, &parser);
812 	}
813 
814 	if (!r) {
815 		r = amdgpu_cs_dependencies(adev, &parser);
816 		if (r)
817 			DRM_ERROR("Failed in the dependencies handling %d!\n", r);
818 	}
819 
820 	if (r)
821 		goto out;
822 
823 	for (i = 0; i < parser.num_ibs; i++)
824 		trace_amdgpu_cs(&parser, i);
825 
826 	r = amdgpu_cs_ib_vm_chunk(adev, &parser);
827 	if (r)
828 		goto out;
829 
830 	if (amdgpu_enable_scheduler && parser.num_ibs) {
831 		struct amdgpu_ring * ring = parser.ibs->ring;
832 		struct amd_sched_fence *fence;
833 		struct amdgpu_job *job;
834 
835 		job = kzalloc(sizeof(struct amdgpu_job), GFP_KERNEL);
836 		if (!job) {
837 			r = -ENOMEM;
838 			goto out;
839 		}
840 
841 		job->base.sched = &ring->sched;
842 		job->base.s_entity = &parser.ctx->rings[ring->idx].entity;
843 		job->adev = parser.adev;
844 		job->owner = parser.filp;
845 		job->free_job = amdgpu_cs_free_job;
846 
847 		job->ibs = parser.ibs;
848 		job->num_ibs = parser.num_ibs;
849 		parser.ibs = NULL;
850 		parser.num_ibs = 0;
851 
852 		if (job->ibs[job->num_ibs - 1].user) {
853 			job->uf = parser.uf;
854 			job->ibs[job->num_ibs - 1].user = &job->uf;
855 			parser.uf.bo = NULL;
856 		}
857 
858 		fence = amd_sched_fence_create(job->base.s_entity,
859 					       parser.filp);
860 		if (!fence) {
861 			r = -ENOMEM;
862 			amdgpu_cs_free_job(job);
863 			kfree(job);
864 			goto out;
865 		}
866 		job->base.s_fence = fence;
867 		parser.fence = fence_get(&fence->base);
868 
869 		cs->out.handle = amdgpu_ctx_add_fence(parser.ctx, ring,
870 						      &fence->base);
871 		job->ibs[job->num_ibs - 1].sequence = cs->out.handle;
872 
873 		trace_amdgpu_cs_ioctl(job);
874 		amd_sched_entity_push_job(&job->base);
875 
876 	} else {
877 		struct amdgpu_fence *fence;
878 
879 		r = amdgpu_ib_schedule(adev, parser.num_ibs, parser.ibs,
880 				       parser.filp);
881 		fence = parser.ibs[parser.num_ibs - 1].fence;
882 		parser.fence = fence_get(&fence->base);
883 		cs->out.handle = parser.ibs[parser.num_ibs - 1].sequence;
884 	}
885 
886 out:
887 	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
888 	r = amdgpu_cs_handle_lockup(adev, r);
889 	return r;
890 }
891 
892 /**
893  * amdgpu_cs_wait_ioctl - wait for a command submission to finish
894  *
895  * @dev: drm device
896  * @data: data from userspace
897  * @filp: file private
898  *
899  * Wait for the command submission identified by handle to finish.
900  */
901 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
902 			 struct drm_file *filp)
903 {
904 	union drm_amdgpu_wait_cs *wait = data;
905 	struct amdgpu_device *adev = dev->dev_private;
906 	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
907 	struct amdgpu_ring *ring = NULL;
908 	struct amdgpu_ctx *ctx;
909 	struct fence *fence;
910 	long r;
911 
912 	r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance,
913 			       wait->in.ring, &ring);
914 	if (r)
915 		return r;
916 
917 	ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
918 	if (ctx == NULL)
919 		return -EINVAL;
920 
921 	fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
922 	if (IS_ERR(fence))
923 		r = PTR_ERR(fence);
924 	else if (fence) {
925 		r = fence_wait_timeout(fence, true, timeout);
926 		fence_put(fence);
927 	} else
928 		r = 1;
929 
930 	amdgpu_ctx_put(ctx);
931 	if (r < 0)
932 		return r;
933 
934 	memset(wait, 0, sizeof(*wait));
935 	wait->out.status = (r == 0);
936 
937 	return 0;
938 }
939 
940 /**
941  * amdgpu_cs_find_bo_va - find bo_va for VM address
942  *
943  * @parser: command submission parser context
944  * @addr: VM address
945  * @bo: resulting BO of the mapping found
946  *
947  * Search the buffer objects in the command submission context for a certain
948  * virtual memory address. Returns allocation structure when found, NULL
949  * otherwise.
950  */
951 struct amdgpu_bo_va_mapping *
952 amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
953 		       uint64_t addr, struct amdgpu_bo **bo)
954 {
955 	struct amdgpu_bo_list_entry *reloc;
956 	struct amdgpu_bo_va_mapping *mapping;
957 
958 	addr /= AMDGPU_GPU_PAGE_SIZE;
959 
960 	list_for_each_entry(reloc, &parser->validated, tv.head) {
961 		if (!reloc->bo_va)
962 			continue;
963 
964 		list_for_each_entry(mapping, &reloc->bo_va->valids, list) {
965 			if (mapping->it.start > addr ||
966 			    addr > mapping->it.last)
967 				continue;
968 
969 			*bo = reloc->bo_va->bo;
970 			return mapping;
971 		}
972 
973 		list_for_each_entry(mapping, &reloc->bo_va->invalids, list) {
974 			if (mapping->it.start > addr ||
975 			    addr > mapping->it.last)
976 				continue;
977 
978 			*bo = reloc->bo_va->bo;
979 			return mapping;
980 		}
981 	}
982 
983 	return NULL;
984 }
985