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 static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
131 				      struct drm_amdgpu_cs_chunk_fence *fence_data)
132 {
133 	struct drm_gem_object *gobj;
134 	uint32_t handle;
135 
136 	handle = fence_data->handle;
137 	gobj = drm_gem_object_lookup(p->adev->ddev, p->filp,
138 				     fence_data->handle);
139 	if (gobj == NULL)
140 		return -EINVAL;
141 
142 	p->uf.bo = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
143 	p->uf.offset = fence_data->offset;
144 
145 	if (amdgpu_ttm_tt_has_userptr(p->uf.bo->tbo.ttm)) {
146 		drm_gem_object_unreference_unlocked(gobj);
147 		return -EINVAL;
148 	}
149 
150 	p->uf_entry.robj = amdgpu_bo_ref(p->uf.bo);
151 	p->uf_entry.prefered_domains = AMDGPU_GEM_DOMAIN_GTT;
152 	p->uf_entry.allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
153 	p->uf_entry.priority = 0;
154 	p->uf_entry.tv.bo = &p->uf_entry.robj->tbo;
155 	p->uf_entry.tv.shared = true;
156 
157 	drm_gem_object_unreference_unlocked(gobj);
158 	return 0;
159 }
160 
161 int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
162 {
163 	union drm_amdgpu_cs *cs = data;
164 	uint64_t *chunk_array_user;
165 	uint64_t *chunk_array;
166 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
167 	unsigned size;
168 	int i;
169 	int ret;
170 
171 	if (cs->in.num_chunks == 0)
172 		return 0;
173 
174 	chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
175 	if (!chunk_array)
176 		return -ENOMEM;
177 
178 	p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
179 	if (!p->ctx) {
180 		ret = -EINVAL;
181 		goto free_chunk;
182 	}
183 
184 	p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
185 
186 	/* get chunks */
187 	INIT_LIST_HEAD(&p->validated);
188 	chunk_array_user = (uint64_t __user *)(unsigned long)(cs->in.chunks);
189 	if (copy_from_user(chunk_array, chunk_array_user,
190 			   sizeof(uint64_t)*cs->in.num_chunks)) {
191 		ret = -EFAULT;
192 		goto put_bo_list;
193 	}
194 
195 	p->nchunks = cs->in.num_chunks;
196 	p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
197 			    GFP_KERNEL);
198 	if (!p->chunks) {
199 		ret = -ENOMEM;
200 		goto put_bo_list;
201 	}
202 
203 	for (i = 0; i < p->nchunks; i++) {
204 		struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
205 		struct drm_amdgpu_cs_chunk user_chunk;
206 		uint32_t __user *cdata;
207 
208 		chunk_ptr = (void __user *)(unsigned long)chunk_array[i];
209 		if (copy_from_user(&user_chunk, chunk_ptr,
210 				       sizeof(struct drm_amdgpu_cs_chunk))) {
211 			ret = -EFAULT;
212 			i--;
213 			goto free_partial_kdata;
214 		}
215 		p->chunks[i].chunk_id = user_chunk.chunk_id;
216 		p->chunks[i].length_dw = user_chunk.length_dw;
217 
218 		size = p->chunks[i].length_dw;
219 		cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
220 		p->chunks[i].user_ptr = cdata;
221 
222 		p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t));
223 		if (p->chunks[i].kdata == NULL) {
224 			ret = -ENOMEM;
225 			i--;
226 			goto free_partial_kdata;
227 		}
228 		size *= sizeof(uint32_t);
229 		if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
230 			ret = -EFAULT;
231 			goto free_partial_kdata;
232 		}
233 
234 		switch (p->chunks[i].chunk_id) {
235 		case AMDGPU_CHUNK_ID_IB:
236 			p->num_ibs++;
237 			break;
238 
239 		case AMDGPU_CHUNK_ID_FENCE:
240 			size = sizeof(struct drm_amdgpu_cs_chunk_fence);
241 			if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
242 				ret = -EINVAL;
243 				goto free_partial_kdata;
244 			}
245 
246 			ret = amdgpu_cs_user_fence_chunk(p, (void *)p->chunks[i].kdata);
247 			if (ret)
248 				goto free_partial_kdata;
249 
250 			break;
251 
252 		case AMDGPU_CHUNK_ID_DEPENDENCIES:
253 			break;
254 
255 		default:
256 			ret = -EINVAL;
257 			goto free_partial_kdata;
258 		}
259 	}
260 
261 
262 	p->ibs = kcalloc(p->num_ibs, sizeof(struct amdgpu_ib), GFP_KERNEL);
263 	if (!p->ibs) {
264 		ret = -ENOMEM;
265 		goto free_all_kdata;
266 	}
267 
268 	kfree(chunk_array);
269 	return 0;
270 
271 free_all_kdata:
272 	i = p->nchunks - 1;
273 free_partial_kdata:
274 	for (; i >= 0; i--)
275 		drm_free_large(p->chunks[i].kdata);
276 	kfree(p->chunks);
277 put_bo_list:
278 	if (p->bo_list)
279 		amdgpu_bo_list_put(p->bo_list);
280 	amdgpu_ctx_put(p->ctx);
281 free_chunk:
282 	kfree(chunk_array);
283 
284 	return ret;
285 }
286 
287 /* Returns how many bytes TTM can move per IB.
288  */
289 static u64 amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev)
290 {
291 	u64 real_vram_size = adev->mc.real_vram_size;
292 	u64 vram_usage = atomic64_read(&adev->vram_usage);
293 
294 	/* This function is based on the current VRAM usage.
295 	 *
296 	 * - If all of VRAM is free, allow relocating the number of bytes that
297 	 *   is equal to 1/4 of the size of VRAM for this IB.
298 
299 	 * - If more than one half of VRAM is occupied, only allow relocating
300 	 *   1 MB of data for this IB.
301 	 *
302 	 * - From 0 to one half of used VRAM, the threshold decreases
303 	 *   linearly.
304 	 *         __________________
305 	 * 1/4 of -|\               |
306 	 * VRAM    | \              |
307 	 *         |  \             |
308 	 *         |   \            |
309 	 *         |    \           |
310 	 *         |     \          |
311 	 *         |      \         |
312 	 *         |       \________|1 MB
313 	 *         |----------------|
314 	 *    VRAM 0 %             100 %
315 	 *         used            used
316 	 *
317 	 * Note: It's a threshold, not a limit. The threshold must be crossed
318 	 * for buffer relocations to stop, so any buffer of an arbitrary size
319 	 * can be moved as long as the threshold isn't crossed before
320 	 * the relocation takes place. We don't want to disable buffer
321 	 * relocations completely.
322 	 *
323 	 * The idea is that buffers should be placed in VRAM at creation time
324 	 * and TTM should only do a minimum number of relocations during
325 	 * command submission. In practice, you need to submit at least
326 	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
327 	 *
328 	 * Also, things can get pretty crazy under memory pressure and actual
329 	 * VRAM usage can change a lot, so playing safe even at 50% does
330 	 * consistently increase performance.
331 	 */
332 
333 	u64 half_vram = real_vram_size >> 1;
334 	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
335 	u64 bytes_moved_threshold = half_free_vram >> 1;
336 	return max(bytes_moved_threshold, 1024*1024ull);
337 }
338 
339 int amdgpu_cs_list_validate(struct amdgpu_device *adev,
340 			    struct amdgpu_vm *vm,
341 			    struct list_head *validated)
342 {
343 	struct amdgpu_bo_list_entry *lobj;
344 	struct amdgpu_bo *bo;
345 	u64 bytes_moved = 0, initial_bytes_moved;
346 	u64 bytes_moved_threshold = amdgpu_cs_get_threshold_for_moves(adev);
347 	int r;
348 
349 	list_for_each_entry(lobj, validated, tv.head) {
350 		bo = lobj->robj;
351 		if (!bo->pin_count) {
352 			u32 domain = lobj->prefered_domains;
353 			u32 current_domain =
354 				amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
355 
356 			/* Check if this buffer will be moved and don't move it
357 			 * if we have moved too many buffers for this IB already.
358 			 *
359 			 * Note that this allows moving at least one buffer of
360 			 * any size, because it doesn't take the current "bo"
361 			 * into account. We don't want to disallow buffer moves
362 			 * completely.
363 			 */
364 			if ((lobj->allowed_domains & current_domain) != 0 &&
365 			    (domain & current_domain) == 0 && /* will be moved */
366 			    bytes_moved > bytes_moved_threshold) {
367 				/* don't move it */
368 				domain = current_domain;
369 			}
370 
371 		retry:
372 			amdgpu_ttm_placement_from_domain(bo, domain);
373 			initial_bytes_moved = atomic64_read(&adev->num_bytes_moved);
374 			r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
375 			bytes_moved += atomic64_read(&adev->num_bytes_moved) -
376 				       initial_bytes_moved;
377 
378 			if (unlikely(r)) {
379 				if (r != -ERESTARTSYS && domain != lobj->allowed_domains) {
380 					domain = lobj->allowed_domains;
381 					goto retry;
382 				}
383 				return r;
384 			}
385 		}
386 		lobj->bo_va = amdgpu_vm_bo_find(vm, bo);
387 	}
388 	return 0;
389 }
390 
391 static int amdgpu_cs_parser_relocs(struct amdgpu_cs_parser *p)
392 {
393 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
394 	struct amdgpu_cs_buckets buckets;
395 	struct list_head duplicates;
396 	bool need_mmap_lock = false;
397 	int i, r;
398 
399 	if (p->bo_list) {
400 		need_mmap_lock = p->bo_list->has_userptr;
401 		amdgpu_cs_buckets_init(&buckets);
402 		for (i = 0; i < p->bo_list->num_entries; i++)
403 			amdgpu_cs_buckets_add(&buckets, &p->bo_list->array[i].tv.head,
404 								  p->bo_list->array[i].priority);
405 
406 		amdgpu_cs_buckets_get_list(&buckets, &p->validated);
407 	}
408 
409 	INIT_LIST_HEAD(&duplicates);
410 	amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
411 
412 	if (p->uf.bo)
413 		list_add(&p->uf_entry.tv.head, &p->validated);
414 
415 	if (need_mmap_lock)
416 		down_read(&current->mm->mmap_sem);
417 
418 	r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true, &duplicates);
419 	if (unlikely(r != 0))
420 		goto error_reserve;
421 
422 	amdgpu_vm_get_pt_bos(&fpriv->vm, &duplicates);
423 
424 	r = amdgpu_cs_list_validate(p->adev, &fpriv->vm, &duplicates);
425 	if (r)
426 		goto error_validate;
427 
428 	r = amdgpu_cs_list_validate(p->adev, &fpriv->vm, &p->validated);
429 
430 error_validate:
431 	if (r) {
432 		amdgpu_vm_move_pt_bos_in_lru(p->adev, &fpriv->vm);
433 		ttm_eu_backoff_reservation(&p->ticket, &p->validated);
434 	}
435 
436 error_reserve:
437 	if (need_mmap_lock)
438 		up_read(&current->mm->mmap_sem);
439 
440 	return r;
441 }
442 
443 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
444 {
445 	struct amdgpu_bo_list_entry *e;
446 	int r;
447 
448 	list_for_each_entry(e, &p->validated, tv.head) {
449 		struct reservation_object *resv = e->robj->tbo.resv;
450 		r = amdgpu_sync_resv(p->adev, &p->ibs[0].sync, resv, p->filp);
451 
452 		if (r)
453 			return r;
454 	}
455 	return 0;
456 }
457 
458 static int cmp_size_smaller_first(void *priv, struct list_head *a,
459 				  struct list_head *b)
460 {
461 	struct amdgpu_bo_list_entry *la = list_entry(a, struct amdgpu_bo_list_entry, tv.head);
462 	struct amdgpu_bo_list_entry *lb = list_entry(b, struct amdgpu_bo_list_entry, tv.head);
463 
464 	/* Sort A before B if A is smaller. */
465 	return (int)la->robj->tbo.num_pages - (int)lb->robj->tbo.num_pages;
466 }
467 
468 /**
469  * cs_parser_fini() - clean parser states
470  * @parser:	parser structure holding parsing context.
471  * @error:	error number
472  *
473  * If error is set than unvalidate buffer, otherwise just free memory
474  * used by parsing context.
475  **/
476 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bool backoff)
477 {
478 	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
479 	unsigned i;
480 
481 	amdgpu_vm_move_pt_bos_in_lru(parser->adev, &fpriv->vm);
482 
483 	if (!error) {
484 		/* Sort the buffer list from the smallest to largest buffer,
485 		 * which affects the order of buffers in the LRU list.
486 		 * This assures that the smallest buffers are added first
487 		 * to the LRU list, so they are likely to be later evicted
488 		 * first, instead of large buffers whose eviction is more
489 		 * expensive.
490 		 *
491 		 * This slightly lowers the number of bytes moved by TTM
492 		 * per frame under memory pressure.
493 		 */
494 		list_sort(NULL, &parser->validated, cmp_size_smaller_first);
495 
496 		ttm_eu_fence_buffer_objects(&parser->ticket,
497 					    &parser->validated,
498 					    parser->fence);
499 	} else if (backoff) {
500 		ttm_eu_backoff_reservation(&parser->ticket,
501 					   &parser->validated);
502 	}
503 	fence_put(parser->fence);
504 
505 	if (parser->ctx)
506 		amdgpu_ctx_put(parser->ctx);
507 	if (parser->bo_list)
508 		amdgpu_bo_list_put(parser->bo_list);
509 
510 	for (i = 0; i < parser->nchunks; i++)
511 		drm_free_large(parser->chunks[i].kdata);
512 	kfree(parser->chunks);
513 	if (parser->ibs)
514 		for (i = 0; i < parser->num_ibs; i++)
515 			amdgpu_ib_free(parser->adev, &parser->ibs[i]);
516 	kfree(parser->ibs);
517 	amdgpu_bo_unref(&parser->uf.bo);
518 	amdgpu_bo_unref(&parser->uf_entry.robj);
519 }
520 
521 static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p,
522 				   struct amdgpu_vm *vm)
523 {
524 	struct amdgpu_device *adev = p->adev;
525 	struct amdgpu_bo_va *bo_va;
526 	struct amdgpu_bo *bo;
527 	int i, r;
528 
529 	r = amdgpu_vm_update_page_directory(adev, vm);
530 	if (r)
531 		return r;
532 
533 	r = amdgpu_sync_fence(adev, &p->ibs[0].sync, vm->page_directory_fence);
534 	if (r)
535 		return r;
536 
537 	r = amdgpu_vm_clear_freed(adev, vm);
538 	if (r)
539 		return r;
540 
541 	if (p->bo_list) {
542 		for (i = 0; i < p->bo_list->num_entries; i++) {
543 			struct fence *f;
544 
545 			/* ignore duplicates */
546 			bo = p->bo_list->array[i].robj;
547 			if (!bo)
548 				continue;
549 
550 			bo_va = p->bo_list->array[i].bo_va;
551 			if (bo_va == NULL)
552 				continue;
553 
554 			r = amdgpu_vm_bo_update(adev, bo_va, &bo->tbo.mem);
555 			if (r)
556 				return r;
557 
558 			f = bo_va->last_pt_update;
559 			r = amdgpu_sync_fence(adev, &p->ibs[0].sync, f);
560 			if (r)
561 				return r;
562 		}
563 
564 	}
565 
566 	r = amdgpu_vm_clear_invalids(adev, vm, &p->ibs[0].sync);
567 
568 	if (amdgpu_vm_debug && p->bo_list) {
569 		/* Invalidate all BOs to test for userspace bugs */
570 		for (i = 0; i < p->bo_list->num_entries; i++) {
571 			/* ignore duplicates */
572 			bo = p->bo_list->array[i].robj;
573 			if (!bo)
574 				continue;
575 
576 			amdgpu_vm_bo_invalidate(adev, bo);
577 		}
578 	}
579 
580 	return r;
581 }
582 
583 static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
584 				 struct amdgpu_cs_parser *parser)
585 {
586 	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
587 	struct amdgpu_vm *vm = &fpriv->vm;
588 	struct amdgpu_ring *ring;
589 	int i, r;
590 
591 	if (parser->num_ibs == 0)
592 		return 0;
593 
594 	/* Only for UVD/VCE VM emulation */
595 	for (i = 0; i < parser->num_ibs; i++) {
596 		ring = parser->ibs[i].ring;
597 		if (ring->funcs->parse_cs) {
598 			r = amdgpu_ring_parse_cs(ring, parser, i);
599 			if (r)
600 				return r;
601 		}
602 	}
603 
604 	r = amdgpu_bo_vm_update_pte(parser, vm);
605 	if (!r)
606 		amdgpu_cs_sync_rings(parser);
607 
608 	return r;
609 }
610 
611 static int amdgpu_cs_handle_lockup(struct amdgpu_device *adev, int r)
612 {
613 	if (r == -EDEADLK) {
614 		r = amdgpu_gpu_reset(adev);
615 		if (!r)
616 			r = -EAGAIN;
617 	}
618 	return r;
619 }
620 
621 static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
622 			     struct amdgpu_cs_parser *parser)
623 {
624 	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
625 	struct amdgpu_vm *vm = &fpriv->vm;
626 	int i, j;
627 	int r;
628 
629 	for (i = 0, j = 0; i < parser->nchunks && j < parser->num_ibs; i++) {
630 		struct amdgpu_cs_chunk *chunk;
631 		struct amdgpu_ib *ib;
632 		struct drm_amdgpu_cs_chunk_ib *chunk_ib;
633 		struct amdgpu_ring *ring;
634 
635 		chunk = &parser->chunks[i];
636 		ib = &parser->ibs[j];
637 		chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
638 
639 		if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
640 			continue;
641 
642 		r = amdgpu_cs_get_ring(adev, chunk_ib->ip_type,
643 				       chunk_ib->ip_instance, chunk_ib->ring,
644 				       &ring);
645 		if (r)
646 			return r;
647 
648 		if (ring->funcs->parse_cs) {
649 			struct amdgpu_bo_va_mapping *m;
650 			struct amdgpu_bo *aobj = NULL;
651 			uint64_t offset;
652 			uint8_t *kptr;
653 
654 			m = amdgpu_cs_find_mapping(parser, chunk_ib->va_start,
655 						   &aobj);
656 			if (!aobj) {
657 				DRM_ERROR("IB va_start is invalid\n");
658 				return -EINVAL;
659 			}
660 
661 			if ((chunk_ib->va_start + chunk_ib->ib_bytes) >
662 			    (m->it.last + 1) * AMDGPU_GPU_PAGE_SIZE) {
663 				DRM_ERROR("IB va_start+ib_bytes is invalid\n");
664 				return -EINVAL;
665 			}
666 
667 			/* the IB should be reserved at this point */
668 			r = amdgpu_bo_kmap(aobj, (void **)&kptr);
669 			if (r) {
670 				return r;
671 			}
672 
673 			offset = ((uint64_t)m->it.start) * AMDGPU_GPU_PAGE_SIZE;
674 			kptr += chunk_ib->va_start - offset;
675 
676 			r =  amdgpu_ib_get(ring, NULL, chunk_ib->ib_bytes, ib);
677 			if (r) {
678 				DRM_ERROR("Failed to get ib !\n");
679 				return r;
680 			}
681 
682 			memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
683 			amdgpu_bo_kunmap(aobj);
684 		} else {
685 			r =  amdgpu_ib_get(ring, vm, 0, ib);
686 			if (r) {
687 				DRM_ERROR("Failed to get ib !\n");
688 				return r;
689 			}
690 
691 			ib->gpu_addr = chunk_ib->va_start;
692 		}
693 
694 		ib->length_dw = chunk_ib->ib_bytes / 4;
695 		ib->flags = chunk_ib->flags;
696 		ib->ctx = parser->ctx;
697 		j++;
698 	}
699 
700 	if (!parser->num_ibs)
701 		return 0;
702 
703 	/* add GDS resources to first IB */
704 	if (parser->bo_list) {
705 		struct amdgpu_bo *gds = parser->bo_list->gds_obj;
706 		struct amdgpu_bo *gws = parser->bo_list->gws_obj;
707 		struct amdgpu_bo *oa = parser->bo_list->oa_obj;
708 		struct amdgpu_ib *ib = &parser->ibs[0];
709 
710 		if (gds) {
711 			ib->gds_base = amdgpu_bo_gpu_offset(gds);
712 			ib->gds_size = amdgpu_bo_size(gds);
713 		}
714 		if (gws) {
715 			ib->gws_base = amdgpu_bo_gpu_offset(gws);
716 			ib->gws_size = amdgpu_bo_size(gws);
717 		}
718 		if (oa) {
719 			ib->oa_base = amdgpu_bo_gpu_offset(oa);
720 			ib->oa_size = amdgpu_bo_size(oa);
721 		}
722 	}
723 	/* wrap the last IB with user fence */
724 	if (parser->uf.bo) {
725 		struct amdgpu_ib *ib = &parser->ibs[parser->num_ibs - 1];
726 
727 		/* UVD & VCE fw doesn't support user fences */
728 		if (ib->ring->type == AMDGPU_RING_TYPE_UVD ||
729 		    ib->ring->type == AMDGPU_RING_TYPE_VCE)
730 			return -EINVAL;
731 
732 		ib->user = &parser->uf;
733 	}
734 
735 	return 0;
736 }
737 
738 static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
739 				  struct amdgpu_cs_parser *p)
740 {
741 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
742 	struct amdgpu_ib *ib;
743 	int i, j, r;
744 
745 	if (!p->num_ibs)
746 		return 0;
747 
748 	/* Add dependencies to first IB */
749 	ib = &p->ibs[0];
750 	for (i = 0; i < p->nchunks; ++i) {
751 		struct drm_amdgpu_cs_chunk_dep *deps;
752 		struct amdgpu_cs_chunk *chunk;
753 		unsigned num_deps;
754 
755 		chunk = &p->chunks[i];
756 
757 		if (chunk->chunk_id != AMDGPU_CHUNK_ID_DEPENDENCIES)
758 			continue;
759 
760 		deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
761 		num_deps = chunk->length_dw * 4 /
762 			sizeof(struct drm_amdgpu_cs_chunk_dep);
763 
764 		for (j = 0; j < num_deps; ++j) {
765 			struct amdgpu_ring *ring;
766 			struct amdgpu_ctx *ctx;
767 			struct fence *fence;
768 
769 			r = amdgpu_cs_get_ring(adev, deps[j].ip_type,
770 					       deps[j].ip_instance,
771 					       deps[j].ring, &ring);
772 			if (r)
773 				return r;
774 
775 			ctx = amdgpu_ctx_get(fpriv, deps[j].ctx_id);
776 			if (ctx == NULL)
777 				return -EINVAL;
778 
779 			fence = amdgpu_ctx_get_fence(ctx, ring,
780 						     deps[j].handle);
781 			if (IS_ERR(fence)) {
782 				r = PTR_ERR(fence);
783 				amdgpu_ctx_put(ctx);
784 				return r;
785 
786 			} else if (fence) {
787 				r = amdgpu_sync_fence(adev, &ib->sync, fence);
788 				fence_put(fence);
789 				amdgpu_ctx_put(ctx);
790 				if (r)
791 					return r;
792 			}
793 		}
794 	}
795 
796 	return 0;
797 }
798 
799 static int amdgpu_cs_free_job(struct amdgpu_job *job)
800 {
801 	int i;
802 	if (job->ibs)
803 		for (i = 0; i < job->num_ibs; i++)
804 			amdgpu_ib_free(job->adev, &job->ibs[i]);
805 	kfree(job->ibs);
806 	if (job->uf.bo)
807 		amdgpu_bo_unref(&job->uf.bo);
808 	return 0;
809 }
810 
811 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
812 {
813 	struct amdgpu_device *adev = dev->dev_private;
814 	union drm_amdgpu_cs *cs = data;
815 	struct amdgpu_cs_parser parser = {};
816 	bool reserved_buffers = false;
817 	int i, r;
818 
819 	if (!adev->accel_working)
820 		return -EBUSY;
821 
822 	parser.adev = adev;
823 	parser.filp = filp;
824 
825 	r = amdgpu_cs_parser_init(&parser, data);
826 	if (r) {
827 		DRM_ERROR("Failed to initialize parser !\n");
828 		amdgpu_cs_parser_fini(&parser, r, false);
829 		r = amdgpu_cs_handle_lockup(adev, r);
830 		return r;
831 	}
832 	r = amdgpu_cs_parser_relocs(&parser);
833 	if (r == -ENOMEM)
834 		DRM_ERROR("Not enough memory for command submission!\n");
835 	else if (r && r != -ERESTARTSYS)
836 		DRM_ERROR("Failed to process the buffer list %d!\n", r);
837 	else if (!r) {
838 		reserved_buffers = true;
839 		r = amdgpu_cs_ib_fill(adev, &parser);
840 	}
841 
842 	if (!r) {
843 		r = amdgpu_cs_dependencies(adev, &parser);
844 		if (r)
845 			DRM_ERROR("Failed in the dependencies handling %d!\n", r);
846 	}
847 
848 	if (r)
849 		goto out;
850 
851 	for (i = 0; i < parser.num_ibs; i++)
852 		trace_amdgpu_cs(&parser, i);
853 
854 	r = amdgpu_cs_ib_vm_chunk(adev, &parser);
855 	if (r)
856 		goto out;
857 
858 	if (amdgpu_enable_scheduler && parser.num_ibs) {
859 		struct amdgpu_ring * ring = parser.ibs->ring;
860 		struct amd_sched_fence *fence;
861 		struct amdgpu_job *job;
862 
863 		job = kzalloc(sizeof(struct amdgpu_job), GFP_KERNEL);
864 		if (!job) {
865 			r = -ENOMEM;
866 			goto out;
867 		}
868 
869 		job->base.sched = &ring->sched;
870 		job->base.s_entity = &parser.ctx->rings[ring->idx].entity;
871 		job->adev = parser.adev;
872 		job->owner = parser.filp;
873 		job->free_job = amdgpu_cs_free_job;
874 
875 		job->ibs = parser.ibs;
876 		job->num_ibs = parser.num_ibs;
877 		parser.ibs = NULL;
878 		parser.num_ibs = 0;
879 
880 		if (job->ibs[job->num_ibs - 1].user) {
881 			job->uf = parser.uf;
882 			job->ibs[job->num_ibs - 1].user = &job->uf;
883 			parser.uf.bo = NULL;
884 		}
885 
886 		fence = amd_sched_fence_create(job->base.s_entity,
887 					       parser.filp);
888 		if (!fence) {
889 			r = -ENOMEM;
890 			amdgpu_cs_free_job(job);
891 			kfree(job);
892 			goto out;
893 		}
894 		job->base.s_fence = fence;
895 		parser.fence = fence_get(&fence->base);
896 
897 		cs->out.handle = amdgpu_ctx_add_fence(parser.ctx, ring,
898 						      &fence->base);
899 		job->ibs[job->num_ibs - 1].sequence = cs->out.handle;
900 
901 		trace_amdgpu_cs_ioctl(job);
902 		amd_sched_entity_push_job(&job->base);
903 
904 	} else {
905 		struct amdgpu_fence *fence;
906 
907 		r = amdgpu_ib_schedule(adev, parser.num_ibs, parser.ibs,
908 				       parser.filp);
909 		fence = parser.ibs[parser.num_ibs - 1].fence;
910 		parser.fence = fence_get(&fence->base);
911 		cs->out.handle = parser.ibs[parser.num_ibs - 1].sequence;
912 	}
913 
914 out:
915 	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
916 	r = amdgpu_cs_handle_lockup(adev, r);
917 	return r;
918 }
919 
920 /**
921  * amdgpu_cs_wait_ioctl - wait for a command submission to finish
922  *
923  * @dev: drm device
924  * @data: data from userspace
925  * @filp: file private
926  *
927  * Wait for the command submission identified by handle to finish.
928  */
929 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
930 			 struct drm_file *filp)
931 {
932 	union drm_amdgpu_wait_cs *wait = data;
933 	struct amdgpu_device *adev = dev->dev_private;
934 	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
935 	struct amdgpu_ring *ring = NULL;
936 	struct amdgpu_ctx *ctx;
937 	struct fence *fence;
938 	long r;
939 
940 	r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance,
941 			       wait->in.ring, &ring);
942 	if (r)
943 		return r;
944 
945 	ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
946 	if (ctx == NULL)
947 		return -EINVAL;
948 
949 	fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
950 	if (IS_ERR(fence))
951 		r = PTR_ERR(fence);
952 	else if (fence) {
953 		r = fence_wait_timeout(fence, true, timeout);
954 		fence_put(fence);
955 	} else
956 		r = 1;
957 
958 	amdgpu_ctx_put(ctx);
959 	if (r < 0)
960 		return r;
961 
962 	memset(wait, 0, sizeof(*wait));
963 	wait->out.status = (r == 0);
964 
965 	return 0;
966 }
967 
968 /**
969  * amdgpu_cs_find_bo_va - find bo_va for VM address
970  *
971  * @parser: command submission parser context
972  * @addr: VM address
973  * @bo: resulting BO of the mapping found
974  *
975  * Search the buffer objects in the command submission context for a certain
976  * virtual memory address. Returns allocation structure when found, NULL
977  * otherwise.
978  */
979 struct amdgpu_bo_va_mapping *
980 amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
981 		       uint64_t addr, struct amdgpu_bo **bo)
982 {
983 	struct amdgpu_bo_list_entry *reloc;
984 	struct amdgpu_bo_va_mapping *mapping;
985 
986 	addr /= AMDGPU_GPU_PAGE_SIZE;
987 
988 	list_for_each_entry(reloc, &parser->validated, tv.head) {
989 		if (!reloc->bo_va)
990 			continue;
991 
992 		list_for_each_entry(mapping, &reloc->bo_va->valids, list) {
993 			if (mapping->it.start > addr ||
994 			    addr > mapping->it.last)
995 				continue;
996 
997 			*bo = reloc->bo_va->bo;
998 			return mapping;
999 		}
1000 
1001 		list_for_each_entry(mapping, &reloc->bo_va->invalids, list) {
1002 			if (mapping->it.start > addr ||
1003 			    addr > mapping->it.last)
1004 				continue;
1005 
1006 			*bo = reloc->bo_va->bo;
1007 			return mapping;
1008 		}
1009 	}
1010 
1011 	return NULL;
1012 }
1013