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/pagemap.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 *data,
91 				      uint32_t *offset)
92 {
93 	struct drm_gem_object *gobj;
94 
95 	gobj = drm_gem_object_lookup(p->filp, data->handle);
96 	if (gobj == NULL)
97 		return -EINVAL;
98 
99 	p->uf_entry.robj = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
100 	p->uf_entry.priority = 0;
101 	p->uf_entry.tv.bo = &p->uf_entry.robj->tbo;
102 	p->uf_entry.tv.shared = true;
103 	p->uf_entry.user_pages = NULL;
104 	*offset = data->offset;
105 
106 	drm_gem_object_unreference_unlocked(gobj);
107 
108 	if (amdgpu_ttm_tt_get_usermm(p->uf_entry.robj->tbo.ttm)) {
109 		amdgpu_bo_unref(&p->uf_entry.robj);
110 		return -EINVAL;
111 	}
112 
113 	return 0;
114 }
115 
116 int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
117 {
118 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
119 	struct amdgpu_vm *vm = &fpriv->vm;
120 	union drm_amdgpu_cs *cs = data;
121 	uint64_t *chunk_array_user;
122 	uint64_t *chunk_array;
123 	unsigned size, num_ibs = 0;
124 	uint32_t uf_offset = 0;
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 			++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, p->chunks[i].kdata,
200 							 &uf_offset);
201 			if (ret)
202 				goto free_partial_kdata;
203 
204 			break;
205 
206 		case AMDGPU_CHUNK_ID_DEPENDENCIES:
207 			break;
208 
209 		default:
210 			ret = -EINVAL;
211 			goto free_partial_kdata;
212 		}
213 	}
214 
215 	ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
216 	if (ret)
217 		goto free_all_kdata;
218 
219 	if (p->uf_entry.robj)
220 		p->job->uf_addr = uf_offset;
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 		bool binding_userptr = false;
300 		struct mm_struct *usermm;
301 		uint32_t domain;
302 
303 		usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
304 		if (usermm && usermm != current->mm)
305 			return -EPERM;
306 
307 		/* Check if we have user pages and nobody bound the BO already */
308 		if (lobj->user_pages && bo->tbo.ttm->state != tt_bound) {
309 			size_t size = sizeof(struct page *);
310 
311 			size *= bo->tbo.ttm->num_pages;
312 			memcpy(bo->tbo.ttm->pages, lobj->user_pages, size);
313 			binding_userptr = true;
314 		}
315 
316 		if (bo->pin_count)
317 			continue;
318 
319 		/* Avoid moving this one if we have moved too many buffers
320 		 * for this IB already.
321 		 *
322 		 * Note that this allows moving at least one buffer of
323 		 * any size, because it doesn't take the current "bo"
324 		 * into account. We don't want to disallow buffer moves
325 		 * completely.
326 		 */
327 		if (p->bytes_moved <= p->bytes_moved_threshold)
328 			domain = bo->prefered_domains;
329 		else
330 			domain = bo->allowed_domains;
331 
332 	retry:
333 		amdgpu_ttm_placement_from_domain(bo, domain);
334 		initial_bytes_moved = atomic64_read(&bo->adev->num_bytes_moved);
335 		r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
336 		p->bytes_moved += atomic64_read(&bo->adev->num_bytes_moved) -
337 			       initial_bytes_moved;
338 
339 		if (unlikely(r)) {
340 			if (r != -ERESTARTSYS && domain != bo->allowed_domains) {
341 				domain = bo->allowed_domains;
342 				goto retry;
343 			}
344 			return r;
345 		}
346 
347 		if (binding_userptr) {
348 			drm_free_large(lobj->user_pages);
349 			lobj->user_pages = NULL;
350 		}
351 	}
352 	return 0;
353 }
354 
355 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
356 				union drm_amdgpu_cs *cs)
357 {
358 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
359 	struct amdgpu_bo_list_entry *e;
360 	struct list_head duplicates;
361 	bool need_mmap_lock = false;
362 	unsigned i, tries = 10;
363 	int r;
364 
365 	INIT_LIST_HEAD(&p->validated);
366 
367 	p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
368 	if (p->bo_list) {
369 		need_mmap_lock = p->bo_list->first_userptr !=
370 			p->bo_list->num_entries;
371 		amdgpu_bo_list_get_list(p->bo_list, &p->validated);
372 	}
373 
374 	INIT_LIST_HEAD(&duplicates);
375 	amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
376 
377 	if (p->uf_entry.robj)
378 		list_add(&p->uf_entry.tv.head, &p->validated);
379 
380 	if (need_mmap_lock)
381 		down_read(&current->mm->mmap_sem);
382 
383 	while (1) {
384 		struct list_head need_pages;
385 		unsigned i;
386 
387 		r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
388 					   &duplicates);
389 		if (unlikely(r != 0))
390 			goto error_free_pages;
391 
392 		/* Without a BO list we don't have userptr BOs */
393 		if (!p->bo_list)
394 			break;
395 
396 		INIT_LIST_HEAD(&need_pages);
397 		for (i = p->bo_list->first_userptr;
398 		     i < p->bo_list->num_entries; ++i) {
399 
400 			e = &p->bo_list->array[i];
401 
402 			if (amdgpu_ttm_tt_userptr_invalidated(e->robj->tbo.ttm,
403 				 &e->user_invalidated) && e->user_pages) {
404 
405 				/* We acquired a page array, but somebody
406 				 * invalidated it. Free it an try again
407 				 */
408 				release_pages(e->user_pages,
409 					      e->robj->tbo.ttm->num_pages,
410 					      false);
411 				drm_free_large(e->user_pages);
412 				e->user_pages = NULL;
413 			}
414 
415 			if (e->robj->tbo.ttm->state != tt_bound &&
416 			    !e->user_pages) {
417 				list_del(&e->tv.head);
418 				list_add(&e->tv.head, &need_pages);
419 
420 				amdgpu_bo_unreserve(e->robj);
421 			}
422 		}
423 
424 		if (list_empty(&need_pages))
425 			break;
426 
427 		/* Unreserve everything again. */
428 		ttm_eu_backoff_reservation(&p->ticket, &p->validated);
429 
430 		/* We tried to often, just abort */
431 		if (!--tries) {
432 			r = -EDEADLK;
433 			goto error_free_pages;
434 		}
435 
436 		/* Fill the page arrays for all useptrs. */
437 		list_for_each_entry(e, &need_pages, tv.head) {
438 			struct ttm_tt *ttm = e->robj->tbo.ttm;
439 
440 			e->user_pages = drm_calloc_large(ttm->num_pages,
441 							 sizeof(struct page*));
442 			if (!e->user_pages) {
443 				r = -ENOMEM;
444 				goto error_free_pages;
445 			}
446 
447 			r = amdgpu_ttm_tt_get_user_pages(ttm, e->user_pages);
448 			if (r) {
449 				drm_free_large(e->user_pages);
450 				e->user_pages = NULL;
451 				goto error_free_pages;
452 			}
453 		}
454 
455 		/* And try again. */
456 		list_splice(&need_pages, &p->validated);
457 	}
458 
459 	amdgpu_vm_get_pt_bos(p->adev, &fpriv->vm, &duplicates);
460 
461 	p->bytes_moved_threshold = amdgpu_cs_get_threshold_for_moves(p->adev);
462 	p->bytes_moved = 0;
463 
464 	r = amdgpu_cs_list_validate(p, &duplicates);
465 	if (r)
466 		goto error_validate;
467 
468 	r = amdgpu_cs_list_validate(p, &p->validated);
469 	if (r)
470 		goto error_validate;
471 
472 	fpriv->vm.last_eviction_counter =
473 		atomic64_read(&p->adev->num_evictions);
474 
475 	if (p->bo_list) {
476 		struct amdgpu_bo *gds = p->bo_list->gds_obj;
477 		struct amdgpu_bo *gws = p->bo_list->gws_obj;
478 		struct amdgpu_bo *oa = p->bo_list->oa_obj;
479 		struct amdgpu_vm *vm = &fpriv->vm;
480 		unsigned i;
481 
482 		for (i = 0; i < p->bo_list->num_entries; i++) {
483 			struct amdgpu_bo *bo = p->bo_list->array[i].robj;
484 
485 			p->bo_list->array[i].bo_va = amdgpu_vm_bo_find(vm, bo);
486 		}
487 
488 		if (gds) {
489 			p->job->gds_base = amdgpu_bo_gpu_offset(gds);
490 			p->job->gds_size = amdgpu_bo_size(gds);
491 		}
492 		if (gws) {
493 			p->job->gws_base = amdgpu_bo_gpu_offset(gws);
494 			p->job->gws_size = amdgpu_bo_size(gws);
495 		}
496 		if (oa) {
497 			p->job->oa_base = amdgpu_bo_gpu_offset(oa);
498 			p->job->oa_size = amdgpu_bo_size(oa);
499 		}
500 	}
501 
502 	if (p->uf_entry.robj)
503 		p->job->uf_addr += amdgpu_bo_gpu_offset(p->uf_entry.robj);
504 
505 error_validate:
506 	if (r) {
507 		amdgpu_vm_move_pt_bos_in_lru(p->adev, &fpriv->vm);
508 		ttm_eu_backoff_reservation(&p->ticket, &p->validated);
509 	}
510 
511 error_free_pages:
512 
513 	if (need_mmap_lock)
514 		up_read(&current->mm->mmap_sem);
515 
516 	if (p->bo_list) {
517 		for (i = p->bo_list->first_userptr;
518 		     i < p->bo_list->num_entries; ++i) {
519 			e = &p->bo_list->array[i];
520 
521 			if (!e->user_pages)
522 				continue;
523 
524 			release_pages(e->user_pages,
525 				      e->robj->tbo.ttm->num_pages,
526 				      false);
527 			drm_free_large(e->user_pages);
528 		}
529 	}
530 
531 	return r;
532 }
533 
534 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
535 {
536 	struct amdgpu_bo_list_entry *e;
537 	int r;
538 
539 	list_for_each_entry(e, &p->validated, tv.head) {
540 		struct reservation_object *resv = e->robj->tbo.resv;
541 		r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, p->filp);
542 
543 		if (r)
544 			return r;
545 	}
546 	return 0;
547 }
548 
549 /**
550  * cs_parser_fini() - clean parser states
551  * @parser:	parser structure holding parsing context.
552  * @error:	error number
553  *
554  * If error is set than unvalidate buffer, otherwise just free memory
555  * used by parsing context.
556  **/
557 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bool backoff)
558 {
559 	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
560 	unsigned i;
561 
562 	if (!error) {
563 		amdgpu_vm_move_pt_bos_in_lru(parser->adev, &fpriv->vm);
564 
565 		ttm_eu_fence_buffer_objects(&parser->ticket,
566 					    &parser->validated,
567 					    parser->fence);
568 	} else if (backoff) {
569 		ttm_eu_backoff_reservation(&parser->ticket,
570 					   &parser->validated);
571 	}
572 	fence_put(parser->fence);
573 
574 	if (parser->ctx)
575 		amdgpu_ctx_put(parser->ctx);
576 	if (parser->bo_list)
577 		amdgpu_bo_list_put(parser->bo_list);
578 
579 	for (i = 0; i < parser->nchunks; i++)
580 		drm_free_large(parser->chunks[i].kdata);
581 	kfree(parser->chunks);
582 	if (parser->job)
583 		amdgpu_job_free(parser->job);
584 	amdgpu_bo_unref(&parser->uf_entry.robj);
585 }
586 
587 static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p,
588 				   struct amdgpu_vm *vm)
589 {
590 	struct amdgpu_device *adev = p->adev;
591 	struct amdgpu_bo_va *bo_va;
592 	struct amdgpu_bo *bo;
593 	int i, r;
594 
595 	r = amdgpu_vm_update_page_directory(adev, vm);
596 	if (r)
597 		return r;
598 
599 	r = amdgpu_sync_fence(adev, &p->job->sync, vm->page_directory_fence);
600 	if (r)
601 		return r;
602 
603 	r = amdgpu_vm_clear_freed(adev, vm);
604 	if (r)
605 		return r;
606 
607 	if (p->bo_list) {
608 		for (i = 0; i < p->bo_list->num_entries; i++) {
609 			struct fence *f;
610 
611 			/* ignore duplicates */
612 			bo = p->bo_list->array[i].robj;
613 			if (!bo)
614 				continue;
615 
616 			bo_va = p->bo_list->array[i].bo_va;
617 			if (bo_va == NULL)
618 				continue;
619 
620 			r = amdgpu_vm_bo_update(adev, bo_va, &bo->tbo.mem);
621 			if (r)
622 				return r;
623 
624 			f = bo_va->last_pt_update;
625 			r = amdgpu_sync_fence(adev, &p->job->sync, f);
626 			if (r)
627 				return r;
628 		}
629 
630 	}
631 
632 	r = amdgpu_vm_clear_invalids(adev, vm, &p->job->sync);
633 
634 	if (amdgpu_vm_debug && p->bo_list) {
635 		/* Invalidate all BOs to test for userspace bugs */
636 		for (i = 0; i < p->bo_list->num_entries; i++) {
637 			/* ignore duplicates */
638 			bo = p->bo_list->array[i].robj;
639 			if (!bo)
640 				continue;
641 
642 			amdgpu_vm_bo_invalidate(adev, bo);
643 		}
644 	}
645 
646 	return r;
647 }
648 
649 static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
650 				 struct amdgpu_cs_parser *p)
651 {
652 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
653 	struct amdgpu_vm *vm = &fpriv->vm;
654 	struct amdgpu_ring *ring = p->job->ring;
655 	int i, r;
656 
657 	/* Only for UVD/VCE VM emulation */
658 	if (ring->funcs->parse_cs) {
659 		p->job->vm = NULL;
660 		for (i = 0; i < p->job->num_ibs; i++) {
661 			r = amdgpu_ring_parse_cs(ring, p, i);
662 			if (r)
663 				return r;
664 		}
665 	} else {
666 		p->job->vm_pd_addr = amdgpu_bo_gpu_offset(vm->page_directory);
667 
668 		r = amdgpu_bo_vm_update_pte(p, vm);
669 		if (r)
670 			return r;
671 	}
672 
673 	return amdgpu_cs_sync_rings(p);
674 }
675 
676 static int amdgpu_cs_handle_lockup(struct amdgpu_device *adev, int r)
677 {
678 	if (r == -EDEADLK) {
679 		r = amdgpu_gpu_reset(adev);
680 		if (!r)
681 			r = -EAGAIN;
682 	}
683 	return r;
684 }
685 
686 static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
687 			     struct amdgpu_cs_parser *parser)
688 {
689 	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
690 	struct amdgpu_vm *vm = &fpriv->vm;
691 	int i, j;
692 	int r;
693 
694 	for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
695 		struct amdgpu_cs_chunk *chunk;
696 		struct amdgpu_ib *ib;
697 		struct drm_amdgpu_cs_chunk_ib *chunk_ib;
698 		struct amdgpu_ring *ring;
699 
700 		chunk = &parser->chunks[i];
701 		ib = &parser->job->ibs[j];
702 		chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
703 
704 		if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
705 			continue;
706 
707 		r = amdgpu_cs_get_ring(adev, chunk_ib->ip_type,
708 				       chunk_ib->ip_instance, chunk_ib->ring,
709 				       &ring);
710 		if (r)
711 			return r;
712 
713 		if (parser->job->ring && parser->job->ring != ring)
714 			return -EINVAL;
715 
716 		parser->job->ring = ring;
717 
718 		if (ring->funcs->parse_cs) {
719 			struct amdgpu_bo_va_mapping *m;
720 			struct amdgpu_bo *aobj = NULL;
721 			uint64_t offset;
722 			uint8_t *kptr;
723 
724 			m = amdgpu_cs_find_mapping(parser, chunk_ib->va_start,
725 						   &aobj);
726 			if (!aobj) {
727 				DRM_ERROR("IB va_start is invalid\n");
728 				return -EINVAL;
729 			}
730 
731 			if ((chunk_ib->va_start + chunk_ib->ib_bytes) >
732 			    (m->it.last + 1) * AMDGPU_GPU_PAGE_SIZE) {
733 				DRM_ERROR("IB va_start+ib_bytes is invalid\n");
734 				return -EINVAL;
735 			}
736 
737 			/* the IB should be reserved at this point */
738 			r = amdgpu_bo_kmap(aobj, (void **)&kptr);
739 			if (r) {
740 				return r;
741 			}
742 
743 			offset = ((uint64_t)m->it.start) * AMDGPU_GPU_PAGE_SIZE;
744 			kptr += chunk_ib->va_start - offset;
745 
746 			r =  amdgpu_ib_get(adev, NULL, chunk_ib->ib_bytes, ib);
747 			if (r) {
748 				DRM_ERROR("Failed to get ib !\n");
749 				return r;
750 			}
751 
752 			memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
753 			amdgpu_bo_kunmap(aobj);
754 		} else {
755 			r =  amdgpu_ib_get(adev, vm, 0, ib);
756 			if (r) {
757 				DRM_ERROR("Failed to get ib !\n");
758 				return r;
759 			}
760 
761 			ib->gpu_addr = chunk_ib->va_start;
762 		}
763 
764 		ib->length_dw = chunk_ib->ib_bytes / 4;
765 		ib->flags = chunk_ib->flags;
766 		j++;
767 	}
768 
769 	/* UVD & VCE fw doesn't support user fences */
770 	if (parser->job->uf_addr && (
771 	    parser->job->ring->type == AMDGPU_RING_TYPE_UVD ||
772 	    parser->job->ring->type == AMDGPU_RING_TYPE_VCE))
773 		return -EINVAL;
774 
775 	return 0;
776 }
777 
778 static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
779 				  struct amdgpu_cs_parser *p)
780 {
781 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
782 	int i, j, r;
783 
784 	for (i = 0; i < p->nchunks; ++i) {
785 		struct drm_amdgpu_cs_chunk_dep *deps;
786 		struct amdgpu_cs_chunk *chunk;
787 		unsigned num_deps;
788 
789 		chunk = &p->chunks[i];
790 
791 		if (chunk->chunk_id != AMDGPU_CHUNK_ID_DEPENDENCIES)
792 			continue;
793 
794 		deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
795 		num_deps = chunk->length_dw * 4 /
796 			sizeof(struct drm_amdgpu_cs_chunk_dep);
797 
798 		for (j = 0; j < num_deps; ++j) {
799 			struct amdgpu_ring *ring;
800 			struct amdgpu_ctx *ctx;
801 			struct fence *fence;
802 
803 			r = amdgpu_cs_get_ring(adev, deps[j].ip_type,
804 					       deps[j].ip_instance,
805 					       deps[j].ring, &ring);
806 			if (r)
807 				return r;
808 
809 			ctx = amdgpu_ctx_get(fpriv, deps[j].ctx_id);
810 			if (ctx == NULL)
811 				return -EINVAL;
812 
813 			fence = amdgpu_ctx_get_fence(ctx, ring,
814 						     deps[j].handle);
815 			if (IS_ERR(fence)) {
816 				r = PTR_ERR(fence);
817 				amdgpu_ctx_put(ctx);
818 				return r;
819 
820 			} else if (fence) {
821 				r = amdgpu_sync_fence(adev, &p->job->sync,
822 						      fence);
823 				fence_put(fence);
824 				amdgpu_ctx_put(ctx);
825 				if (r)
826 					return r;
827 			}
828 		}
829 	}
830 
831 	return 0;
832 }
833 
834 static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
835 			    union drm_amdgpu_cs *cs)
836 {
837 	struct amdgpu_ring *ring = p->job->ring;
838 	struct amd_sched_entity *entity = &p->ctx->rings[ring->idx].entity;
839 	struct amdgpu_job *job;
840 	int r;
841 
842 	job = p->job;
843 	p->job = NULL;
844 
845 	r = amd_sched_job_init(&job->base, &ring->sched, entity, p->filp);
846 	if (r) {
847 		amdgpu_job_free(job);
848 		return r;
849 	}
850 
851 	job->owner = p->filp;
852 	job->ctx = entity->fence_context;
853 	p->fence = fence_get(&job->base.s_fence->finished);
854 	cs->out.handle = amdgpu_ctx_add_fence(p->ctx, ring, p->fence);
855 	job->uf_sequence = cs->out.handle;
856 	amdgpu_job_free_resources(job);
857 
858 	trace_amdgpu_cs_ioctl(job);
859 	amd_sched_entity_push_job(&job->base);
860 
861 	return 0;
862 }
863 
864 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
865 {
866 	struct amdgpu_device *adev = dev->dev_private;
867 	union drm_amdgpu_cs *cs = data;
868 	struct amdgpu_cs_parser parser = {};
869 	bool reserved_buffers = false;
870 	int i, r;
871 
872 	if (!adev->accel_working)
873 		return -EBUSY;
874 
875 	parser.adev = adev;
876 	parser.filp = filp;
877 
878 	r = amdgpu_cs_parser_init(&parser, data);
879 	if (r) {
880 		DRM_ERROR("Failed to initialize parser !\n");
881 		amdgpu_cs_parser_fini(&parser, r, false);
882 		r = amdgpu_cs_handle_lockup(adev, r);
883 		return r;
884 	}
885 	r = amdgpu_cs_parser_bos(&parser, data);
886 	if (r == -ENOMEM)
887 		DRM_ERROR("Not enough memory for command submission!\n");
888 	else if (r && r != -ERESTARTSYS)
889 		DRM_ERROR("Failed to process the buffer list %d!\n", r);
890 	else if (!r) {
891 		reserved_buffers = true;
892 		r = amdgpu_cs_ib_fill(adev, &parser);
893 	}
894 
895 	if (!r) {
896 		r = amdgpu_cs_dependencies(adev, &parser);
897 		if (r)
898 			DRM_ERROR("Failed in the dependencies handling %d!\n", r);
899 	}
900 
901 	if (r)
902 		goto out;
903 
904 	for (i = 0; i < parser.job->num_ibs; i++)
905 		trace_amdgpu_cs(&parser, i);
906 
907 	r = amdgpu_cs_ib_vm_chunk(adev, &parser);
908 	if (r)
909 		goto out;
910 
911 	r = amdgpu_cs_submit(&parser, cs);
912 
913 out:
914 	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
915 	r = amdgpu_cs_handle_lockup(adev, r);
916 	return r;
917 }
918 
919 /**
920  * amdgpu_cs_wait_ioctl - wait for a command submission to finish
921  *
922  * @dev: drm device
923  * @data: data from userspace
924  * @filp: file private
925  *
926  * Wait for the command submission identified by handle to finish.
927  */
928 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
929 			 struct drm_file *filp)
930 {
931 	union drm_amdgpu_wait_cs *wait = data;
932 	struct amdgpu_device *adev = dev->dev_private;
933 	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
934 	struct amdgpu_ring *ring = NULL;
935 	struct amdgpu_ctx *ctx;
936 	struct fence *fence;
937 	long r;
938 
939 	r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance,
940 			       wait->in.ring, &ring);
941 	if (r)
942 		return r;
943 
944 	ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
945 	if (ctx == NULL)
946 		return -EINVAL;
947 
948 	fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
949 	if (IS_ERR(fence))
950 		r = PTR_ERR(fence);
951 	else if (fence) {
952 		r = fence_wait_timeout(fence, true, timeout);
953 		fence_put(fence);
954 	} else
955 		r = 1;
956 
957 	amdgpu_ctx_put(ctx);
958 	if (r < 0)
959 		return r;
960 
961 	memset(wait, 0, sizeof(*wait));
962 	wait->out.status = (r == 0);
963 
964 	return 0;
965 }
966 
967 /**
968  * amdgpu_cs_find_bo_va - find bo_va for VM address
969  *
970  * @parser: command submission parser context
971  * @addr: VM address
972  * @bo: resulting BO of the mapping found
973  *
974  * Search the buffer objects in the command submission context for a certain
975  * virtual memory address. Returns allocation structure when found, NULL
976  * otherwise.
977  */
978 struct amdgpu_bo_va_mapping *
979 amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
980 		       uint64_t addr, struct amdgpu_bo **bo)
981 {
982 	struct amdgpu_bo_va_mapping *mapping;
983 	unsigned i;
984 
985 	if (!parser->bo_list)
986 		return NULL;
987 
988 	addr /= AMDGPU_GPU_PAGE_SIZE;
989 
990 	for (i = 0; i < parser->bo_list->num_entries; i++) {
991 		struct amdgpu_bo_list_entry *lobj;
992 
993 		lobj = &parser->bo_list->array[i];
994 		if (!lobj->bo_va)
995 			continue;
996 
997 		list_for_each_entry(mapping, &lobj->bo_va->valids, list) {
998 			if (mapping->it.start > addr ||
999 			    addr > mapping->it.last)
1000 				continue;
1001 
1002 			*bo = lobj->bo_va->bo;
1003 			return mapping;
1004 		}
1005 
1006 		list_for_each_entry(mapping, &lobj->bo_va->invalids, list) {
1007 			if (mapping->it.start > addr ||
1008 			    addr > mapping->it.last)
1009 				continue;
1010 
1011 			*bo = lobj->bo_va->bo;
1012 			return mapping;
1013 		}
1014 	}
1015 
1016 	return NULL;
1017 }
1018