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 				amdgpu_bo_ref(p->uf.bo);
226 				drm_gem_object_unreference_unlocked(gobj);
227 				p->uf.offset = fence_data->offset;
228 			} else {
229 				ret = -EINVAL;
230 				goto free_partial_kdata;
231 			}
232 			break;
233 
234 		case AMDGPU_CHUNK_ID_DEPENDENCIES:
235 			break;
236 
237 		default:
238 			ret = -EINVAL;
239 			goto free_partial_kdata;
240 		}
241 	}
242 
243 
244 	p->ibs = kcalloc(p->num_ibs, sizeof(struct amdgpu_ib), GFP_KERNEL);
245 	if (!p->ibs) {
246 		ret = -ENOMEM;
247 		goto free_all_kdata;
248 	}
249 
250 	kfree(chunk_array);
251 	return 0;
252 
253 free_all_kdata:
254 	i = p->nchunks - 1;
255 free_partial_kdata:
256 	for (; i >= 0; i--)
257 		drm_free_large(p->chunks[i].kdata);
258 	kfree(p->chunks);
259 put_bo_list:
260 	if (p->bo_list)
261 		amdgpu_bo_list_put(p->bo_list);
262 	amdgpu_ctx_put(p->ctx);
263 free_chunk:
264 	kfree(chunk_array);
265 
266 	return ret;
267 }
268 
269 /* Returns how many bytes TTM can move per IB.
270  */
271 static u64 amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev)
272 {
273 	u64 real_vram_size = adev->mc.real_vram_size;
274 	u64 vram_usage = atomic64_read(&adev->vram_usage);
275 
276 	/* This function is based on the current VRAM usage.
277 	 *
278 	 * - If all of VRAM is free, allow relocating the number of bytes that
279 	 *   is equal to 1/4 of the size of VRAM for this IB.
280 
281 	 * - If more than one half of VRAM is occupied, only allow relocating
282 	 *   1 MB of data for this IB.
283 	 *
284 	 * - From 0 to one half of used VRAM, the threshold decreases
285 	 *   linearly.
286 	 *         __________________
287 	 * 1/4 of -|\               |
288 	 * VRAM    | \              |
289 	 *         |  \             |
290 	 *         |   \            |
291 	 *         |    \           |
292 	 *         |     \          |
293 	 *         |      \         |
294 	 *         |       \________|1 MB
295 	 *         |----------------|
296 	 *    VRAM 0 %             100 %
297 	 *         used            used
298 	 *
299 	 * Note: It's a threshold, not a limit. The threshold must be crossed
300 	 * for buffer relocations to stop, so any buffer of an arbitrary size
301 	 * can be moved as long as the threshold isn't crossed before
302 	 * the relocation takes place. We don't want to disable buffer
303 	 * relocations completely.
304 	 *
305 	 * The idea is that buffers should be placed in VRAM at creation time
306 	 * and TTM should only do a minimum number of relocations during
307 	 * command submission. In practice, you need to submit at least
308 	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
309 	 *
310 	 * Also, things can get pretty crazy under memory pressure and actual
311 	 * VRAM usage can change a lot, so playing safe even at 50% does
312 	 * consistently increase performance.
313 	 */
314 
315 	u64 half_vram = real_vram_size >> 1;
316 	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
317 	u64 bytes_moved_threshold = half_free_vram >> 1;
318 	return max(bytes_moved_threshold, 1024*1024ull);
319 }
320 
321 int amdgpu_cs_list_validate(struct amdgpu_device *adev,
322 			    struct amdgpu_vm *vm,
323 			    struct list_head *validated)
324 {
325 	struct amdgpu_bo_list_entry *lobj;
326 	struct amdgpu_bo *bo;
327 	u64 bytes_moved = 0, initial_bytes_moved;
328 	u64 bytes_moved_threshold = amdgpu_cs_get_threshold_for_moves(adev);
329 	int r;
330 
331 	list_for_each_entry(lobj, validated, tv.head) {
332 		bo = lobj->robj;
333 		if (!bo->pin_count) {
334 			u32 domain = lobj->prefered_domains;
335 			u32 current_domain =
336 				amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
337 
338 			/* Check if this buffer will be moved and don't move it
339 			 * if we have moved too many buffers for this IB already.
340 			 *
341 			 * Note that this allows moving at least one buffer of
342 			 * any size, because it doesn't take the current "bo"
343 			 * into account. We don't want to disallow buffer moves
344 			 * completely.
345 			 */
346 			if ((lobj->allowed_domains & current_domain) != 0 &&
347 			    (domain & current_domain) == 0 && /* will be moved */
348 			    bytes_moved > bytes_moved_threshold) {
349 				/* don't move it */
350 				domain = current_domain;
351 			}
352 
353 		retry:
354 			amdgpu_ttm_placement_from_domain(bo, domain);
355 			initial_bytes_moved = atomic64_read(&adev->num_bytes_moved);
356 			r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
357 			bytes_moved += atomic64_read(&adev->num_bytes_moved) -
358 				       initial_bytes_moved;
359 
360 			if (unlikely(r)) {
361 				if (r != -ERESTARTSYS && domain != lobj->allowed_domains) {
362 					domain = lobj->allowed_domains;
363 					goto retry;
364 				}
365 				return r;
366 			}
367 		}
368 		lobj->bo_va = amdgpu_vm_bo_find(vm, bo);
369 	}
370 	return 0;
371 }
372 
373 static int amdgpu_cs_parser_relocs(struct amdgpu_cs_parser *p)
374 {
375 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
376 	struct amdgpu_cs_buckets buckets;
377 	struct list_head duplicates;
378 	bool need_mmap_lock = false;
379 	int i, r;
380 
381 	if (p->bo_list) {
382 		need_mmap_lock = p->bo_list->has_userptr;
383 		amdgpu_cs_buckets_init(&buckets);
384 		for (i = 0; i < p->bo_list->num_entries; i++)
385 			amdgpu_cs_buckets_add(&buckets, &p->bo_list->array[i].tv.head,
386 								  p->bo_list->array[i].priority);
387 
388 		amdgpu_cs_buckets_get_list(&buckets, &p->validated);
389 	}
390 
391 	p->vm_bos = amdgpu_vm_get_bos(p->adev, &fpriv->vm,
392 				      &p->validated);
393 
394 	if (need_mmap_lock)
395 		down_read(&current->mm->mmap_sem);
396 
397 	INIT_LIST_HEAD(&duplicates);
398 	r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true, &duplicates);
399 	if (unlikely(r != 0))
400 		goto error_reserve;
401 
402 	r = amdgpu_cs_list_validate(p->adev, &fpriv->vm, &p->validated);
403 	if (r)
404 		goto error_validate;
405 
406 	r = amdgpu_cs_list_validate(p->adev, &fpriv->vm, &duplicates);
407 
408 error_validate:
409 	if (r)
410 		ttm_eu_backoff_reservation(&p->ticket, &p->validated);
411 
412 error_reserve:
413 	if (need_mmap_lock)
414 		up_read(&current->mm->mmap_sem);
415 
416 	return r;
417 }
418 
419 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
420 {
421 	struct amdgpu_bo_list_entry *e;
422 	int r;
423 
424 	list_for_each_entry(e, &p->validated, tv.head) {
425 		struct reservation_object *resv = e->robj->tbo.resv;
426 		r = amdgpu_sync_resv(p->adev, &p->ibs[0].sync, resv, p->filp);
427 
428 		if (r)
429 			return r;
430 	}
431 	return 0;
432 }
433 
434 static int cmp_size_smaller_first(void *priv, struct list_head *a,
435 				  struct list_head *b)
436 {
437 	struct amdgpu_bo_list_entry *la = list_entry(a, struct amdgpu_bo_list_entry, tv.head);
438 	struct amdgpu_bo_list_entry *lb = list_entry(b, struct amdgpu_bo_list_entry, tv.head);
439 
440 	/* Sort A before B if A is smaller. */
441 	return (int)la->robj->tbo.num_pages - (int)lb->robj->tbo.num_pages;
442 }
443 
444 /**
445  * cs_parser_fini() - clean parser states
446  * @parser:	parser structure holding parsing context.
447  * @error:	error number
448  *
449  * If error is set than unvalidate buffer, otherwise just free memory
450  * used by parsing context.
451  **/
452 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bool backoff)
453 {
454 	unsigned i;
455 
456 	if (!error) {
457 		/* Sort the buffer list from the smallest to largest buffer,
458 		 * which affects the order of buffers in the LRU list.
459 		 * This assures that the smallest buffers are added first
460 		 * to the LRU list, so they are likely to be later evicted
461 		 * first, instead of large buffers whose eviction is more
462 		 * expensive.
463 		 *
464 		 * This slightly lowers the number of bytes moved by TTM
465 		 * per frame under memory pressure.
466 		 */
467 		list_sort(NULL, &parser->validated, cmp_size_smaller_first);
468 
469 		ttm_eu_fence_buffer_objects(&parser->ticket,
470 					    &parser->validated,
471 					    parser->fence);
472 	} else if (backoff) {
473 		ttm_eu_backoff_reservation(&parser->ticket,
474 					   &parser->validated);
475 	}
476 	fence_put(parser->fence);
477 
478 	if (parser->ctx)
479 		amdgpu_ctx_put(parser->ctx);
480 	if (parser->bo_list)
481 		amdgpu_bo_list_put(parser->bo_list);
482 
483 	drm_free_large(parser->vm_bos);
484 	for (i = 0; i < parser->nchunks; i++)
485 		drm_free_large(parser->chunks[i].kdata);
486 	kfree(parser->chunks);
487 	if (parser->ibs)
488 		for (i = 0; i < parser->num_ibs; i++)
489 			amdgpu_ib_free(parser->adev, &parser->ibs[i]);
490 	kfree(parser->ibs);
491 	if (parser->uf.bo)
492 		amdgpu_bo_unref(&parser->uf.bo);
493 }
494 
495 static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p,
496 				   struct amdgpu_vm *vm)
497 {
498 	struct amdgpu_device *adev = p->adev;
499 	struct amdgpu_bo_va *bo_va;
500 	struct amdgpu_bo *bo;
501 	int i, r;
502 
503 	r = amdgpu_vm_update_page_directory(adev, vm);
504 	if (r)
505 		return r;
506 
507 	r = amdgpu_sync_fence(adev, &p->ibs[0].sync, vm->page_directory_fence);
508 	if (r)
509 		return r;
510 
511 	r = amdgpu_vm_clear_freed(adev, vm);
512 	if (r)
513 		return r;
514 
515 	if (p->bo_list) {
516 		for (i = 0; i < p->bo_list->num_entries; i++) {
517 			struct fence *f;
518 
519 			/* ignore duplicates */
520 			bo = p->bo_list->array[i].robj;
521 			if (!bo)
522 				continue;
523 
524 			bo_va = p->bo_list->array[i].bo_va;
525 			if (bo_va == NULL)
526 				continue;
527 
528 			r = amdgpu_vm_bo_update(adev, bo_va, &bo->tbo.mem);
529 			if (r)
530 				return r;
531 
532 			f = bo_va->last_pt_update;
533 			r = amdgpu_sync_fence(adev, &p->ibs[0].sync, f);
534 			if (r)
535 				return r;
536 		}
537 
538 	}
539 
540 	r = amdgpu_vm_clear_invalids(adev, vm, &p->ibs[0].sync);
541 
542 	if (amdgpu_vm_debug && p->bo_list) {
543 		/* Invalidate all BOs to test for userspace bugs */
544 		for (i = 0; i < p->bo_list->num_entries; i++) {
545 			/* ignore duplicates */
546 			bo = p->bo_list->array[i].robj;
547 			if (!bo)
548 				continue;
549 
550 			amdgpu_vm_bo_invalidate(adev, bo);
551 		}
552 	}
553 
554 	return r;
555 }
556 
557 static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
558 				 struct amdgpu_cs_parser *parser)
559 {
560 	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
561 	struct amdgpu_vm *vm = &fpriv->vm;
562 	struct amdgpu_ring *ring;
563 	int i, r;
564 
565 	if (parser->num_ibs == 0)
566 		return 0;
567 
568 	/* Only for UVD/VCE VM emulation */
569 	for (i = 0; i < parser->num_ibs; i++) {
570 		ring = parser->ibs[i].ring;
571 		if (ring->funcs->parse_cs) {
572 			r = amdgpu_ring_parse_cs(ring, parser, i);
573 			if (r)
574 				return r;
575 		}
576 	}
577 
578 	r = amdgpu_bo_vm_update_pte(parser, vm);
579 	if (!r)
580 		amdgpu_cs_sync_rings(parser);
581 
582 	return r;
583 }
584 
585 static int amdgpu_cs_handle_lockup(struct amdgpu_device *adev, int r)
586 {
587 	if (r == -EDEADLK) {
588 		r = amdgpu_gpu_reset(adev);
589 		if (!r)
590 			r = -EAGAIN;
591 	}
592 	return r;
593 }
594 
595 static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
596 			     struct amdgpu_cs_parser *parser)
597 {
598 	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
599 	struct amdgpu_vm *vm = &fpriv->vm;
600 	int i, j;
601 	int r;
602 
603 	for (i = 0, j = 0; i < parser->nchunks && j < parser->num_ibs; i++) {
604 		struct amdgpu_cs_chunk *chunk;
605 		struct amdgpu_ib *ib;
606 		struct drm_amdgpu_cs_chunk_ib *chunk_ib;
607 		struct amdgpu_ring *ring;
608 
609 		chunk = &parser->chunks[i];
610 		ib = &parser->ibs[j];
611 		chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
612 
613 		if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
614 			continue;
615 
616 		r = amdgpu_cs_get_ring(adev, chunk_ib->ip_type,
617 				       chunk_ib->ip_instance, chunk_ib->ring,
618 				       &ring);
619 		if (r)
620 			return r;
621 
622 		if (ring->funcs->parse_cs) {
623 			struct amdgpu_bo_va_mapping *m;
624 			struct amdgpu_bo *aobj = NULL;
625 			uint64_t offset;
626 			uint8_t *kptr;
627 
628 			m = amdgpu_cs_find_mapping(parser, chunk_ib->va_start,
629 						   &aobj);
630 			if (!aobj) {
631 				DRM_ERROR("IB va_start is invalid\n");
632 				return -EINVAL;
633 			}
634 
635 			if ((chunk_ib->va_start + chunk_ib->ib_bytes) >
636 			    (m->it.last + 1) * AMDGPU_GPU_PAGE_SIZE) {
637 				DRM_ERROR("IB va_start+ib_bytes is invalid\n");
638 				return -EINVAL;
639 			}
640 
641 			/* the IB should be reserved at this point */
642 			r = amdgpu_bo_kmap(aobj, (void **)&kptr);
643 			if (r) {
644 				return r;
645 			}
646 
647 			offset = ((uint64_t)m->it.start) * AMDGPU_GPU_PAGE_SIZE;
648 			kptr += chunk_ib->va_start - offset;
649 
650 			r =  amdgpu_ib_get(ring, NULL, chunk_ib->ib_bytes, ib);
651 			if (r) {
652 				DRM_ERROR("Failed to get ib !\n");
653 				return r;
654 			}
655 
656 			memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
657 			amdgpu_bo_kunmap(aobj);
658 		} else {
659 			r =  amdgpu_ib_get(ring, vm, 0, ib);
660 			if (r) {
661 				DRM_ERROR("Failed to get ib !\n");
662 				return r;
663 			}
664 
665 			ib->gpu_addr = chunk_ib->va_start;
666 		}
667 
668 		ib->length_dw = chunk_ib->ib_bytes / 4;
669 		ib->flags = chunk_ib->flags;
670 		ib->ctx = parser->ctx;
671 		j++;
672 	}
673 
674 	if (!parser->num_ibs)
675 		return 0;
676 
677 	/* add GDS resources to first IB */
678 	if (parser->bo_list) {
679 		struct amdgpu_bo *gds = parser->bo_list->gds_obj;
680 		struct amdgpu_bo *gws = parser->bo_list->gws_obj;
681 		struct amdgpu_bo *oa = parser->bo_list->oa_obj;
682 		struct amdgpu_ib *ib = &parser->ibs[0];
683 
684 		if (gds) {
685 			ib->gds_base = amdgpu_bo_gpu_offset(gds);
686 			ib->gds_size = amdgpu_bo_size(gds);
687 		}
688 		if (gws) {
689 			ib->gws_base = amdgpu_bo_gpu_offset(gws);
690 			ib->gws_size = amdgpu_bo_size(gws);
691 		}
692 		if (oa) {
693 			ib->oa_base = amdgpu_bo_gpu_offset(oa);
694 			ib->oa_size = amdgpu_bo_size(oa);
695 		}
696 	}
697 	/* wrap the last IB with user fence */
698 	if (parser->uf.bo) {
699 		struct amdgpu_ib *ib = &parser->ibs[parser->num_ibs - 1];
700 
701 		/* UVD & VCE fw doesn't support user fences */
702 		if (ib->ring->type == AMDGPU_RING_TYPE_UVD ||
703 		    ib->ring->type == AMDGPU_RING_TYPE_VCE)
704 			return -EINVAL;
705 
706 		ib->user = &parser->uf;
707 	}
708 
709 	return 0;
710 }
711 
712 static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
713 				  struct amdgpu_cs_parser *p)
714 {
715 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
716 	struct amdgpu_ib *ib;
717 	int i, j, r;
718 
719 	if (!p->num_ibs)
720 		return 0;
721 
722 	/* Add dependencies to first IB */
723 	ib = &p->ibs[0];
724 	for (i = 0; i < p->nchunks; ++i) {
725 		struct drm_amdgpu_cs_chunk_dep *deps;
726 		struct amdgpu_cs_chunk *chunk;
727 		unsigned num_deps;
728 
729 		chunk = &p->chunks[i];
730 
731 		if (chunk->chunk_id != AMDGPU_CHUNK_ID_DEPENDENCIES)
732 			continue;
733 
734 		deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
735 		num_deps = chunk->length_dw * 4 /
736 			sizeof(struct drm_amdgpu_cs_chunk_dep);
737 
738 		for (j = 0; j < num_deps; ++j) {
739 			struct amdgpu_ring *ring;
740 			struct amdgpu_ctx *ctx;
741 			struct fence *fence;
742 
743 			r = amdgpu_cs_get_ring(adev, deps[j].ip_type,
744 					       deps[j].ip_instance,
745 					       deps[j].ring, &ring);
746 			if (r)
747 				return r;
748 
749 			ctx = amdgpu_ctx_get(fpriv, deps[j].ctx_id);
750 			if (ctx == NULL)
751 				return -EINVAL;
752 
753 			fence = amdgpu_ctx_get_fence(ctx, ring,
754 						     deps[j].handle);
755 			if (IS_ERR(fence)) {
756 				r = PTR_ERR(fence);
757 				amdgpu_ctx_put(ctx);
758 				return r;
759 
760 			} else if (fence) {
761 				r = amdgpu_sync_fence(adev, &ib->sync, fence);
762 				fence_put(fence);
763 				amdgpu_ctx_put(ctx);
764 				if (r)
765 					return r;
766 			}
767 		}
768 	}
769 
770 	return 0;
771 }
772 
773 static int amdgpu_cs_free_job(struct amdgpu_job *job)
774 {
775 	int i;
776 	if (job->ibs)
777 		for (i = 0; i < job->num_ibs; i++)
778 			amdgpu_ib_free(job->adev, &job->ibs[i]);
779 	kfree(job->ibs);
780 	if (job->uf.bo)
781 		amdgpu_bo_unref(&job->uf.bo);
782 	return 0;
783 }
784 
785 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
786 {
787 	struct amdgpu_device *adev = dev->dev_private;
788 	union drm_amdgpu_cs *cs = data;
789 	struct amdgpu_cs_parser parser = {};
790 	bool reserved_buffers = false;
791 	int i, r;
792 
793 	if (!adev->accel_working)
794 		return -EBUSY;
795 
796 	parser.adev = adev;
797 	parser.filp = filp;
798 
799 	r = amdgpu_cs_parser_init(&parser, data);
800 	if (r) {
801 		DRM_ERROR("Failed to initialize parser !\n");
802 		amdgpu_cs_parser_fini(&parser, r, false);
803 		r = amdgpu_cs_handle_lockup(adev, r);
804 		return r;
805 	}
806 	r = amdgpu_cs_parser_relocs(&parser);
807 	if (r == -ENOMEM)
808 		DRM_ERROR("Not enough memory for command submission!\n");
809 	else if (r && r != -ERESTARTSYS)
810 		DRM_ERROR("Failed to process the buffer list %d!\n", r);
811 	else if (!r) {
812 		reserved_buffers = true;
813 		r = amdgpu_cs_ib_fill(adev, &parser);
814 	}
815 
816 	if (!r) {
817 		r = amdgpu_cs_dependencies(adev, &parser);
818 		if (r)
819 			DRM_ERROR("Failed in the dependencies handling %d!\n", r);
820 	}
821 
822 	if (r)
823 		goto out;
824 
825 	for (i = 0; i < parser.num_ibs; i++)
826 		trace_amdgpu_cs(&parser, i);
827 
828 	r = amdgpu_cs_ib_vm_chunk(adev, &parser);
829 	if (r)
830 		goto out;
831 
832 	if (amdgpu_enable_scheduler && parser.num_ibs) {
833 		struct amdgpu_ring * ring = parser.ibs->ring;
834 		struct amd_sched_fence *fence;
835 		struct amdgpu_job *job;
836 
837 		job = kzalloc(sizeof(struct amdgpu_job), GFP_KERNEL);
838 		if (!job) {
839 			r = -ENOMEM;
840 			goto out;
841 		}
842 
843 		job->base.sched = &ring->sched;
844 		job->base.s_entity = &parser.ctx->rings[ring->idx].entity;
845 		job->adev = parser.adev;
846 		job->owner = parser.filp;
847 		job->free_job = amdgpu_cs_free_job;
848 
849 		job->ibs = parser.ibs;
850 		job->num_ibs = parser.num_ibs;
851 		parser.ibs = NULL;
852 		parser.num_ibs = 0;
853 
854 		if (job->ibs[job->num_ibs - 1].user) {
855 			job->uf = parser.uf;
856 			job->ibs[job->num_ibs - 1].user = &job->uf;
857 			parser.uf.bo = NULL;
858 		}
859 
860 		fence = amd_sched_fence_create(job->base.s_entity,
861 					       parser.filp);
862 		if (!fence) {
863 			r = -ENOMEM;
864 			amdgpu_cs_free_job(job);
865 			kfree(job);
866 			goto out;
867 		}
868 		job->base.s_fence = fence;
869 		parser.fence = fence_get(&fence->base);
870 
871 		cs->out.handle = amdgpu_ctx_add_fence(parser.ctx, ring,
872 						      &fence->base);
873 		job->ibs[job->num_ibs - 1].sequence = cs->out.handle;
874 
875 		trace_amdgpu_cs_ioctl(job);
876 		amd_sched_entity_push_job(&job->base);
877 
878 	} else {
879 		struct amdgpu_fence *fence;
880 
881 		r = amdgpu_ib_schedule(adev, parser.num_ibs, parser.ibs,
882 				       parser.filp);
883 		fence = parser.ibs[parser.num_ibs - 1].fence;
884 		parser.fence = fence_get(&fence->base);
885 		cs->out.handle = parser.ibs[parser.num_ibs - 1].sequence;
886 	}
887 
888 out:
889 	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
890 	r = amdgpu_cs_handle_lockup(adev, r);
891 	return r;
892 }
893 
894 /**
895  * amdgpu_cs_wait_ioctl - wait for a command submission to finish
896  *
897  * @dev: drm device
898  * @data: data from userspace
899  * @filp: file private
900  *
901  * Wait for the command submission identified by handle to finish.
902  */
903 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
904 			 struct drm_file *filp)
905 {
906 	union drm_amdgpu_wait_cs *wait = data;
907 	struct amdgpu_device *adev = dev->dev_private;
908 	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
909 	struct amdgpu_ring *ring = NULL;
910 	struct amdgpu_ctx *ctx;
911 	struct fence *fence;
912 	long r;
913 
914 	r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance,
915 			       wait->in.ring, &ring);
916 	if (r)
917 		return r;
918 
919 	ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
920 	if (ctx == NULL)
921 		return -EINVAL;
922 
923 	fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
924 	if (IS_ERR(fence))
925 		r = PTR_ERR(fence);
926 	else if (fence) {
927 		r = fence_wait_timeout(fence, true, timeout);
928 		fence_put(fence);
929 	} else
930 		r = 1;
931 
932 	amdgpu_ctx_put(ctx);
933 	if (r < 0)
934 		return r;
935 
936 	memset(wait, 0, sizeof(*wait));
937 	wait->out.status = (r == 0);
938 
939 	return 0;
940 }
941 
942 /**
943  * amdgpu_cs_find_bo_va - find bo_va for VM address
944  *
945  * @parser: command submission parser context
946  * @addr: VM address
947  * @bo: resulting BO of the mapping found
948  *
949  * Search the buffer objects in the command submission context for a certain
950  * virtual memory address. Returns allocation structure when found, NULL
951  * otherwise.
952  */
953 struct amdgpu_bo_va_mapping *
954 amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
955 		       uint64_t addr, struct amdgpu_bo **bo)
956 {
957 	struct amdgpu_bo_list_entry *reloc;
958 	struct amdgpu_bo_va_mapping *mapping;
959 
960 	addr /= AMDGPU_GPU_PAGE_SIZE;
961 
962 	list_for_each_entry(reloc, &parser->validated, tv.head) {
963 		if (!reloc->bo_va)
964 			continue;
965 
966 		list_for_each_entry(mapping, &reloc->bo_va->valids, list) {
967 			if (mapping->it.start > addr ||
968 			    addr > mapping->it.last)
969 				continue;
970 
971 			*bo = reloc->bo_va->bo;
972 			return mapping;
973 		}
974 
975 		list_for_each_entry(mapping, &reloc->bo_va->invalids, list) {
976 			if (mapping->it.start > addr ||
977 			    addr > mapping->it.last)
978 				continue;
979 
980 			*bo = reloc->bo_va->bo;
981 			return mapping;
982 		}
983 	}
984 
985 	return NULL;
986 }
987