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 <linux/pagemap.h>
29 #include <drm/drmP.h>
30 #include <drm/amdgpu_drm.h>
31 #include "amdgpu.h"
32 #include "amdgpu_trace.h"
33 
34 int amdgpu_cs_get_ring(struct amdgpu_device *adev, u32 ip_type,
35 		       u32 ip_instance, u32 ring,
36 		       struct amdgpu_ring **out_ring)
37 {
38 	/* Right now all IPs have only one instance - multiple rings. */
39 	if (ip_instance != 0) {
40 		DRM_ERROR("invalid ip instance: %d\n", ip_instance);
41 		return -EINVAL;
42 	}
43 
44 	switch (ip_type) {
45 	default:
46 		DRM_ERROR("unknown ip type: %d\n", ip_type);
47 		return -EINVAL;
48 	case AMDGPU_HW_IP_GFX:
49 		if (ring < adev->gfx.num_gfx_rings) {
50 			*out_ring = &adev->gfx.gfx_ring[ring];
51 		} else {
52 			DRM_ERROR("only %d gfx rings are supported now\n",
53 				  adev->gfx.num_gfx_rings);
54 			return -EINVAL;
55 		}
56 		break;
57 	case AMDGPU_HW_IP_COMPUTE:
58 		if (ring < adev->gfx.num_compute_rings) {
59 			*out_ring = &adev->gfx.compute_ring[ring];
60 		} else {
61 			DRM_ERROR("only %d compute rings are supported now\n",
62 				  adev->gfx.num_compute_rings);
63 			return -EINVAL;
64 		}
65 		break;
66 	case AMDGPU_HW_IP_DMA:
67 		if (ring < adev->sdma.num_instances) {
68 			*out_ring = &adev->sdma.instance[ring].ring;
69 		} else {
70 			DRM_ERROR("only %d SDMA rings are supported\n",
71 				  adev->sdma.num_instances);
72 			return -EINVAL;
73 		}
74 		break;
75 	case AMDGPU_HW_IP_UVD:
76 		*out_ring = &adev->uvd.ring;
77 		break;
78 	case AMDGPU_HW_IP_VCE:
79 		if (ring < 2){
80 			*out_ring = &adev->vce.ring[ring];
81 		} else {
82 			DRM_ERROR("only two VCE rings are supported\n");
83 			return -EINVAL;
84 		}
85 		break;
86 	}
87 	return 0;
88 }
89 
90 static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
91 				      struct amdgpu_user_fence *uf,
92 				      struct drm_amdgpu_cs_chunk_fence *fence_data)
93 {
94 	struct drm_gem_object *gobj;
95 	uint32_t handle;
96 
97 	handle = fence_data->handle;
98 	gobj = drm_gem_object_lookup(p->adev->ddev, p->filp,
99 				     fence_data->handle);
100 	if (gobj == NULL)
101 		return -EINVAL;
102 
103 	uf->bo = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
104 	uf->offset = fence_data->offset;
105 
106 	if (amdgpu_ttm_tt_get_usermm(uf->bo->tbo.ttm)) {
107 		drm_gem_object_unreference_unlocked(gobj);
108 		return -EINVAL;
109 	}
110 
111 	p->uf_entry.robj = amdgpu_bo_ref(uf->bo);
112 	p->uf_entry.priority = 0;
113 	p->uf_entry.tv.bo = &p->uf_entry.robj->tbo;
114 	p->uf_entry.tv.shared = true;
115 	p->uf_entry.user_pages = NULL;
116 
117 	drm_gem_object_unreference_unlocked(gobj);
118 	return 0;
119 }
120 
121 int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
122 {
123 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
124 	union drm_amdgpu_cs *cs = data;
125 	uint64_t *chunk_array_user;
126 	uint64_t *chunk_array;
127 	struct amdgpu_user_fence uf = {};
128 	unsigned size, num_ibs = 0;
129 	int i;
130 	int ret;
131 
132 	if (cs->in.num_chunks == 0)
133 		return 0;
134 
135 	chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
136 	if (!chunk_array)
137 		return -ENOMEM;
138 
139 	p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
140 	if (!p->ctx) {
141 		ret = -EINVAL;
142 		goto free_chunk;
143 	}
144 
145 	/* get chunks */
146 	chunk_array_user = (uint64_t __user *)(unsigned long)(cs->in.chunks);
147 	if (copy_from_user(chunk_array, chunk_array_user,
148 			   sizeof(uint64_t)*cs->in.num_chunks)) {
149 		ret = -EFAULT;
150 		goto put_ctx;
151 	}
152 
153 	p->nchunks = cs->in.num_chunks;
154 	p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
155 			    GFP_KERNEL);
156 	if (!p->chunks) {
157 		ret = -ENOMEM;
158 		goto put_ctx;
159 	}
160 
161 	for (i = 0; i < p->nchunks; i++) {
162 		struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
163 		struct drm_amdgpu_cs_chunk user_chunk;
164 		uint32_t __user *cdata;
165 
166 		chunk_ptr = (void __user *)(unsigned long)chunk_array[i];
167 		if (copy_from_user(&user_chunk, chunk_ptr,
168 				       sizeof(struct drm_amdgpu_cs_chunk))) {
169 			ret = -EFAULT;
170 			i--;
171 			goto free_partial_kdata;
172 		}
173 		p->chunks[i].chunk_id = user_chunk.chunk_id;
174 		p->chunks[i].length_dw = user_chunk.length_dw;
175 
176 		size = p->chunks[i].length_dw;
177 		cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
178 
179 		p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t));
180 		if (p->chunks[i].kdata == NULL) {
181 			ret = -ENOMEM;
182 			i--;
183 			goto free_partial_kdata;
184 		}
185 		size *= sizeof(uint32_t);
186 		if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
187 			ret = -EFAULT;
188 			goto free_partial_kdata;
189 		}
190 
191 		switch (p->chunks[i].chunk_id) {
192 		case AMDGPU_CHUNK_ID_IB:
193 			++num_ibs;
194 			break;
195 
196 		case AMDGPU_CHUNK_ID_FENCE:
197 			size = sizeof(struct drm_amdgpu_cs_chunk_fence);
198 			if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
199 				ret = -EINVAL;
200 				goto free_partial_kdata;
201 			}
202 
203 			ret = amdgpu_cs_user_fence_chunk(p, &uf, (void *)p->chunks[i].kdata);
204 			if (ret)
205 				goto free_partial_kdata;
206 
207 			break;
208 
209 		case AMDGPU_CHUNK_ID_DEPENDENCIES:
210 			break;
211 
212 		default:
213 			ret = -EINVAL;
214 			goto free_partial_kdata;
215 		}
216 	}
217 
218 	ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job);
219 	if (ret)
220 		goto free_all_kdata;
221 
222 	p->job->uf = uf;
223 
224 	kfree(chunk_array);
225 	return 0;
226 
227 free_all_kdata:
228 	i = p->nchunks - 1;
229 free_partial_kdata:
230 	for (; i >= 0; i--)
231 		drm_free_large(p->chunks[i].kdata);
232 	kfree(p->chunks);
233 put_ctx:
234 	amdgpu_ctx_put(p->ctx);
235 free_chunk:
236 	kfree(chunk_array);
237 
238 	return ret;
239 }
240 
241 /* Returns how many bytes TTM can move per IB.
242  */
243 static u64 amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev)
244 {
245 	u64 real_vram_size = adev->mc.real_vram_size;
246 	u64 vram_usage = atomic64_read(&adev->vram_usage);
247 
248 	/* This function is based on the current VRAM usage.
249 	 *
250 	 * - If all of VRAM is free, allow relocating the number of bytes that
251 	 *   is equal to 1/4 of the size of VRAM for this IB.
252 
253 	 * - If more than one half of VRAM is occupied, only allow relocating
254 	 *   1 MB of data for this IB.
255 	 *
256 	 * - From 0 to one half of used VRAM, the threshold decreases
257 	 *   linearly.
258 	 *         __________________
259 	 * 1/4 of -|\               |
260 	 * VRAM    | \              |
261 	 *         |  \             |
262 	 *         |   \            |
263 	 *         |    \           |
264 	 *         |     \          |
265 	 *         |      \         |
266 	 *         |       \________|1 MB
267 	 *         |----------------|
268 	 *    VRAM 0 %             100 %
269 	 *         used            used
270 	 *
271 	 * Note: It's a threshold, not a limit. The threshold must be crossed
272 	 * for buffer relocations to stop, so any buffer of an arbitrary size
273 	 * can be moved as long as the threshold isn't crossed before
274 	 * the relocation takes place. We don't want to disable buffer
275 	 * relocations completely.
276 	 *
277 	 * The idea is that buffers should be placed in VRAM at creation time
278 	 * and TTM should only do a minimum number of relocations during
279 	 * command submission. In practice, you need to submit at least
280 	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
281 	 *
282 	 * Also, things can get pretty crazy under memory pressure and actual
283 	 * VRAM usage can change a lot, so playing safe even at 50% does
284 	 * consistently increase performance.
285 	 */
286 
287 	u64 half_vram = real_vram_size >> 1;
288 	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
289 	u64 bytes_moved_threshold = half_free_vram >> 1;
290 	return max(bytes_moved_threshold, 1024*1024ull);
291 }
292 
293 int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
294 			    struct list_head *validated)
295 {
296 	struct amdgpu_bo_list_entry *lobj;
297 	u64 initial_bytes_moved;
298 	int r;
299 
300 	list_for_each_entry(lobj, validated, tv.head) {
301 		struct amdgpu_bo *bo = lobj->robj;
302 		bool binding_userptr = false;
303 		struct mm_struct *usermm;
304 		uint32_t domain;
305 
306 		usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
307 		if (usermm && usermm != current->mm)
308 			return -EPERM;
309 
310 		/* Check if we have user pages and nobody bound the BO already */
311 		if (lobj->user_pages && bo->tbo.ttm->state != tt_bound) {
312 			size_t size = sizeof(struct page *);
313 
314 			size *= bo->tbo.ttm->num_pages;
315 			memcpy(bo->tbo.ttm->pages, lobj->user_pages, size);
316 			binding_userptr = true;
317 		}
318 
319 		if (bo->pin_count)
320 			continue;
321 
322 		/* Avoid moving this one if we have moved too many buffers
323 		 * for this IB already.
324 		 *
325 		 * Note that this allows moving at least one buffer of
326 		 * any size, because it doesn't take the current "bo"
327 		 * into account. We don't want to disallow buffer moves
328 		 * completely.
329 		 */
330 		if (p->bytes_moved <= p->bytes_moved_threshold)
331 			domain = bo->prefered_domains;
332 		else
333 			domain = bo->allowed_domains;
334 
335 	retry:
336 		amdgpu_ttm_placement_from_domain(bo, domain);
337 		initial_bytes_moved = atomic64_read(&bo->adev->num_bytes_moved);
338 		r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
339 		p->bytes_moved += atomic64_read(&bo->adev->num_bytes_moved) -
340 			       initial_bytes_moved;
341 
342 		if (unlikely(r)) {
343 			if (r != -ERESTARTSYS && domain != bo->allowed_domains) {
344 				domain = bo->allowed_domains;
345 				goto retry;
346 			}
347 			return r;
348 		}
349 
350 		if (binding_userptr) {
351 			drm_free_large(lobj->user_pages);
352 			lobj->user_pages = NULL;
353 		}
354 	}
355 	return 0;
356 }
357 
358 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
359 				union drm_amdgpu_cs *cs)
360 {
361 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
362 	struct amdgpu_bo_list_entry *e;
363 	struct list_head duplicates;
364 	bool need_mmap_lock = false;
365 	unsigned i, tries = 10;
366 	int r;
367 
368 	INIT_LIST_HEAD(&p->validated);
369 
370 	p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
371 	if (p->bo_list) {
372 		need_mmap_lock = p->bo_list->first_userptr !=
373 			p->bo_list->num_entries;
374 		amdgpu_bo_list_get_list(p->bo_list, &p->validated);
375 	}
376 
377 	INIT_LIST_HEAD(&duplicates);
378 	amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
379 
380 	if (p->job->uf.bo)
381 		list_add(&p->uf_entry.tv.head, &p->validated);
382 
383 	if (need_mmap_lock)
384 		down_read(&current->mm->mmap_sem);
385 
386 	while (1) {
387 		struct list_head need_pages;
388 		unsigned i;
389 
390 		r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
391 					   &duplicates);
392 		if (unlikely(r != 0))
393 			goto error_free_pages;
394 
395 		/* Without a BO list we don't have userptr BOs */
396 		if (!p->bo_list)
397 			break;
398 
399 		INIT_LIST_HEAD(&need_pages);
400 		for (i = p->bo_list->first_userptr;
401 		     i < p->bo_list->num_entries; ++i) {
402 
403 			e = &p->bo_list->array[i];
404 
405 			if (amdgpu_ttm_tt_userptr_invalidated(e->robj->tbo.ttm,
406 				 &e->user_invalidated) && e->user_pages) {
407 
408 				/* We acquired a page array, but somebody
409 				 * invalidated it. Free it an try again
410 				 */
411 				release_pages(e->user_pages,
412 					      e->robj->tbo.ttm->num_pages,
413 					      false);
414 				drm_free_large(e->user_pages);
415 				e->user_pages = NULL;
416 			}
417 
418 			if (e->robj->tbo.ttm->state != tt_bound &&
419 			    !e->user_pages) {
420 				list_del(&e->tv.head);
421 				list_add(&e->tv.head, &need_pages);
422 
423 				amdgpu_bo_unreserve(e->robj);
424 			}
425 		}
426 
427 		if (list_empty(&need_pages))
428 			break;
429 
430 		/* Unreserve everything again. */
431 		ttm_eu_backoff_reservation(&p->ticket, &p->validated);
432 
433 		/* We tried to often, just abort */
434 		if (!--tries) {
435 			r = -EDEADLK;
436 			goto error_free_pages;
437 		}
438 
439 		/* Fill the page arrays for all useptrs. */
440 		list_for_each_entry(e, &need_pages, tv.head) {
441 			struct ttm_tt *ttm = e->robj->tbo.ttm;
442 
443 			e->user_pages = drm_calloc_large(ttm->num_pages,
444 							 sizeof(struct page*));
445 			if (!e->user_pages) {
446 				r = -ENOMEM;
447 				goto error_free_pages;
448 			}
449 
450 			r = amdgpu_ttm_tt_get_user_pages(ttm, e->user_pages);
451 			if (r) {
452 				drm_free_large(e->user_pages);
453 				e->user_pages = NULL;
454 				goto error_free_pages;
455 			}
456 		}
457 
458 		/* And try again. */
459 		list_splice(&need_pages, &p->validated);
460 	}
461 
462 	amdgpu_vm_get_pt_bos(&fpriv->vm, &duplicates);
463 
464 	p->bytes_moved_threshold = amdgpu_cs_get_threshold_for_moves(p->adev);
465 	p->bytes_moved = 0;
466 
467 	r = amdgpu_cs_list_validate(p, &duplicates);
468 	if (r)
469 		goto error_validate;
470 
471 	r = amdgpu_cs_list_validate(p, &p->validated);
472 	if (r)
473 		goto error_validate;
474 
475 	if (p->bo_list) {
476 		struct amdgpu_vm *vm = &fpriv->vm;
477 		unsigned i;
478 
479 		for (i = 0; i < p->bo_list->num_entries; i++) {
480 			struct amdgpu_bo *bo = p->bo_list->array[i].robj;
481 
482 			p->bo_list->array[i].bo_va = amdgpu_vm_bo_find(vm, bo);
483 		}
484 	}
485 
486 error_validate:
487 	if (r) {
488 		amdgpu_vm_move_pt_bos_in_lru(p->adev, &fpriv->vm);
489 		ttm_eu_backoff_reservation(&p->ticket, &p->validated);
490 	}
491 
492 error_free_pages:
493 
494 	if (need_mmap_lock)
495 		up_read(&current->mm->mmap_sem);
496 
497 	if (p->bo_list) {
498 		for (i = p->bo_list->first_userptr;
499 		     i < p->bo_list->num_entries; ++i) {
500 			e = &p->bo_list->array[i];
501 
502 			if (!e->user_pages)
503 				continue;
504 
505 			release_pages(e->user_pages,
506 				      e->robj->tbo.ttm->num_pages,
507 				      false);
508 			drm_free_large(e->user_pages);
509 		}
510 	}
511 
512 	return r;
513 }
514 
515 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
516 {
517 	struct amdgpu_bo_list_entry *e;
518 	int r;
519 
520 	list_for_each_entry(e, &p->validated, tv.head) {
521 		struct reservation_object *resv = e->robj->tbo.resv;
522 		r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, p->filp);
523 
524 		if (r)
525 			return r;
526 	}
527 	return 0;
528 }
529 
530 static int cmp_size_smaller_first(void *priv, struct list_head *a,
531 				  struct list_head *b)
532 {
533 	struct amdgpu_bo_list_entry *la = list_entry(a, struct amdgpu_bo_list_entry, tv.head);
534 	struct amdgpu_bo_list_entry *lb = list_entry(b, struct amdgpu_bo_list_entry, tv.head);
535 
536 	/* Sort A before B if A is smaller. */
537 	return (int)la->robj->tbo.num_pages - (int)lb->robj->tbo.num_pages;
538 }
539 
540 /**
541  * cs_parser_fini() - clean parser states
542  * @parser:	parser structure holding parsing context.
543  * @error:	error number
544  *
545  * If error is set than unvalidate buffer, otherwise just free memory
546  * used by parsing context.
547  **/
548 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bool backoff)
549 {
550 	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
551 	unsigned i;
552 
553 	if (!error) {
554 		amdgpu_vm_move_pt_bos_in_lru(parser->adev, &fpriv->vm);
555 
556 		/* Sort the buffer list from the smallest to largest buffer,
557 		 * which affects the order of buffers in the LRU list.
558 		 * This assures that the smallest buffers are added first
559 		 * to the LRU list, so they are likely to be later evicted
560 		 * first, instead of large buffers whose eviction is more
561 		 * expensive.
562 		 *
563 		 * This slightly lowers the number of bytes moved by TTM
564 		 * per frame under memory pressure.
565 		 */
566 		list_sort(NULL, &parser->validated, cmp_size_smaller_first);
567 
568 		ttm_eu_fence_buffer_objects(&parser->ticket,
569 					    &parser->validated,
570 					    parser->fence);
571 	} else if (backoff) {
572 		ttm_eu_backoff_reservation(&parser->ticket,
573 					   &parser->validated);
574 	}
575 	fence_put(parser->fence);
576 
577 	if (parser->ctx)
578 		amdgpu_ctx_put(parser->ctx);
579 	if (parser->bo_list)
580 		amdgpu_bo_list_put(parser->bo_list);
581 
582 	for (i = 0; i < parser->nchunks; i++)
583 		drm_free_large(parser->chunks[i].kdata);
584 	kfree(parser->chunks);
585 	if (parser->job)
586 		amdgpu_job_free(parser->job);
587 	amdgpu_bo_unref(&parser->uf_entry.robj);
588 }
589 
590 static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p,
591 				   struct amdgpu_vm *vm)
592 {
593 	struct amdgpu_device *adev = p->adev;
594 	struct amdgpu_bo_va *bo_va;
595 	struct amdgpu_bo *bo;
596 	int i, r;
597 
598 	r = amdgpu_vm_update_page_directory(adev, vm);
599 	if (r)
600 		return r;
601 
602 	r = amdgpu_sync_fence(adev, &p->job->sync, vm->page_directory_fence);
603 	if (r)
604 		return r;
605 
606 	r = amdgpu_vm_clear_freed(adev, vm);
607 	if (r)
608 		return r;
609 
610 	if (p->bo_list) {
611 		for (i = 0; i < p->bo_list->num_entries; i++) {
612 			struct fence *f;
613 
614 			/* ignore duplicates */
615 			bo = p->bo_list->array[i].robj;
616 			if (!bo)
617 				continue;
618 
619 			bo_va = p->bo_list->array[i].bo_va;
620 			if (bo_va == NULL)
621 				continue;
622 
623 			r = amdgpu_vm_bo_update(adev, bo_va, &bo->tbo.mem);
624 			if (r)
625 				return r;
626 
627 			f = bo_va->last_pt_update;
628 			r = amdgpu_sync_fence(adev, &p->job->sync, f);
629 			if (r)
630 				return r;
631 		}
632 
633 	}
634 
635 	r = amdgpu_vm_clear_invalids(adev, vm, &p->job->sync);
636 
637 	if (amdgpu_vm_debug && p->bo_list) {
638 		/* Invalidate all BOs to test for userspace bugs */
639 		for (i = 0; i < p->bo_list->num_entries; i++) {
640 			/* ignore duplicates */
641 			bo = p->bo_list->array[i].robj;
642 			if (!bo)
643 				continue;
644 
645 			amdgpu_vm_bo_invalidate(adev, bo);
646 		}
647 	}
648 
649 	return r;
650 }
651 
652 static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
653 				 struct amdgpu_cs_parser *p)
654 {
655 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
656 	struct amdgpu_vm *vm = &fpriv->vm;
657 	struct amdgpu_ring *ring = p->job->ring;
658 	int i, r;
659 
660 	/* Only for UVD/VCE VM emulation */
661 	if (ring->funcs->parse_cs) {
662 		for (i = 0; i < p->job->num_ibs; i++) {
663 			r = amdgpu_ring_parse_cs(ring, p, i);
664 			if (r)
665 				return r;
666 		}
667 	}
668 
669 	r = amdgpu_bo_vm_update_pte(p, vm);
670 	if (!r)
671 		amdgpu_cs_sync_rings(p);
672 
673 	return r;
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 		ib->ctx = parser->ctx;
767 		j++;
768 	}
769 
770 	/* add GDS resources to first IB */
771 	if (parser->bo_list) {
772 		struct amdgpu_bo *gds = parser->bo_list->gds_obj;
773 		struct amdgpu_bo *gws = parser->bo_list->gws_obj;
774 		struct amdgpu_bo *oa = parser->bo_list->oa_obj;
775 		struct amdgpu_ib *ib = &parser->job->ibs[0];
776 
777 		if (gds) {
778 			ib->gds_base = amdgpu_bo_gpu_offset(gds);
779 			ib->gds_size = amdgpu_bo_size(gds);
780 		}
781 		if (gws) {
782 			ib->gws_base = amdgpu_bo_gpu_offset(gws);
783 			ib->gws_size = amdgpu_bo_size(gws);
784 		}
785 		if (oa) {
786 			ib->oa_base = amdgpu_bo_gpu_offset(oa);
787 			ib->oa_size = amdgpu_bo_size(oa);
788 		}
789 	}
790 	/* wrap the last IB with user fence */
791 	if (parser->job->uf.bo) {
792 		struct amdgpu_ib *ib = &parser->job->ibs[parser->job->num_ibs - 1];
793 
794 		/* UVD & VCE fw doesn't support user fences */
795 		if (parser->job->ring->type == AMDGPU_RING_TYPE_UVD ||
796 		    parser->job->ring->type == AMDGPU_RING_TYPE_VCE)
797 			return -EINVAL;
798 
799 		ib->user = &parser->job->uf;
800 	}
801 
802 	return 0;
803 }
804 
805 static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
806 				  struct amdgpu_cs_parser *p)
807 {
808 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
809 	int i, j, r;
810 
811 	for (i = 0; i < p->nchunks; ++i) {
812 		struct drm_amdgpu_cs_chunk_dep *deps;
813 		struct amdgpu_cs_chunk *chunk;
814 		unsigned num_deps;
815 
816 		chunk = &p->chunks[i];
817 
818 		if (chunk->chunk_id != AMDGPU_CHUNK_ID_DEPENDENCIES)
819 			continue;
820 
821 		deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
822 		num_deps = chunk->length_dw * 4 /
823 			sizeof(struct drm_amdgpu_cs_chunk_dep);
824 
825 		for (j = 0; j < num_deps; ++j) {
826 			struct amdgpu_ring *ring;
827 			struct amdgpu_ctx *ctx;
828 			struct fence *fence;
829 
830 			r = amdgpu_cs_get_ring(adev, deps[j].ip_type,
831 					       deps[j].ip_instance,
832 					       deps[j].ring, &ring);
833 			if (r)
834 				return r;
835 
836 			ctx = amdgpu_ctx_get(fpriv, deps[j].ctx_id);
837 			if (ctx == NULL)
838 				return -EINVAL;
839 
840 			fence = amdgpu_ctx_get_fence(ctx, ring,
841 						     deps[j].handle);
842 			if (IS_ERR(fence)) {
843 				r = PTR_ERR(fence);
844 				amdgpu_ctx_put(ctx);
845 				return r;
846 
847 			} else if (fence) {
848 				r = amdgpu_sync_fence(adev, &p->job->sync,
849 						      fence);
850 				fence_put(fence);
851 				amdgpu_ctx_put(ctx);
852 				if (r)
853 					return r;
854 			}
855 		}
856 	}
857 
858 	return 0;
859 }
860 
861 static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
862 			    union drm_amdgpu_cs *cs)
863 {
864 	struct amdgpu_ring *ring = p->job->ring;
865 	struct amd_sched_fence *fence;
866 	struct amdgpu_job *job;
867 
868 	job = p->job;
869 	p->job = NULL;
870 
871 	job->base.sched = &ring->sched;
872 	job->base.s_entity = &p->ctx->rings[ring->idx].entity;
873 	job->owner = p->filp;
874 
875 	fence = amd_sched_fence_create(job->base.s_entity, p->filp);
876 	if (!fence) {
877 		amdgpu_job_free(job);
878 		return -ENOMEM;
879 	}
880 
881 	job->base.s_fence = fence;
882 	p->fence = fence_get(&fence->base);
883 
884 	cs->out.handle = amdgpu_ctx_add_fence(p->ctx, ring,
885 					      &fence->base);
886 	job->ibs[job->num_ibs - 1].sequence = cs->out.handle;
887 
888 	trace_amdgpu_cs_ioctl(job);
889 	amd_sched_entity_push_job(&job->base);
890 
891 	return 0;
892 }
893 
894 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
895 {
896 	struct amdgpu_device *adev = dev->dev_private;
897 	union drm_amdgpu_cs *cs = data;
898 	struct amdgpu_cs_parser parser = {};
899 	bool reserved_buffers = false;
900 	int i, r;
901 
902 	if (!adev->accel_working)
903 		return -EBUSY;
904 
905 	parser.adev = adev;
906 	parser.filp = filp;
907 
908 	r = amdgpu_cs_parser_init(&parser, data);
909 	if (r) {
910 		DRM_ERROR("Failed to initialize parser !\n");
911 		amdgpu_cs_parser_fini(&parser, r, false);
912 		r = amdgpu_cs_handle_lockup(adev, r);
913 		return r;
914 	}
915 	r = amdgpu_cs_parser_bos(&parser, data);
916 	if (r == -ENOMEM)
917 		DRM_ERROR("Not enough memory for command submission!\n");
918 	else if (r && r != -ERESTARTSYS)
919 		DRM_ERROR("Failed to process the buffer list %d!\n", r);
920 	else if (!r) {
921 		reserved_buffers = true;
922 		r = amdgpu_cs_ib_fill(adev, &parser);
923 	}
924 
925 	if (!r) {
926 		r = amdgpu_cs_dependencies(adev, &parser);
927 		if (r)
928 			DRM_ERROR("Failed in the dependencies handling %d!\n", r);
929 	}
930 
931 	if (r)
932 		goto out;
933 
934 	for (i = 0; i < parser.job->num_ibs; i++)
935 		trace_amdgpu_cs(&parser, i);
936 
937 	r = amdgpu_cs_ib_vm_chunk(adev, &parser);
938 	if (r)
939 		goto out;
940 
941 	r = amdgpu_cs_submit(&parser, cs);
942 
943 out:
944 	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
945 	r = amdgpu_cs_handle_lockup(adev, r);
946 	return r;
947 }
948 
949 /**
950  * amdgpu_cs_wait_ioctl - wait for a command submission to finish
951  *
952  * @dev: drm device
953  * @data: data from userspace
954  * @filp: file private
955  *
956  * Wait for the command submission identified by handle to finish.
957  */
958 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
959 			 struct drm_file *filp)
960 {
961 	union drm_amdgpu_wait_cs *wait = data;
962 	struct amdgpu_device *adev = dev->dev_private;
963 	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
964 	struct amdgpu_ring *ring = NULL;
965 	struct amdgpu_ctx *ctx;
966 	struct fence *fence;
967 	long r;
968 
969 	r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance,
970 			       wait->in.ring, &ring);
971 	if (r)
972 		return r;
973 
974 	ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
975 	if (ctx == NULL)
976 		return -EINVAL;
977 
978 	fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
979 	if (IS_ERR(fence))
980 		r = PTR_ERR(fence);
981 	else if (fence) {
982 		r = fence_wait_timeout(fence, true, timeout);
983 		fence_put(fence);
984 	} else
985 		r = 1;
986 
987 	amdgpu_ctx_put(ctx);
988 	if (r < 0)
989 		return r;
990 
991 	memset(wait, 0, sizeof(*wait));
992 	wait->out.status = (r == 0);
993 
994 	return 0;
995 }
996 
997 /**
998  * amdgpu_cs_find_bo_va - find bo_va for VM address
999  *
1000  * @parser: command submission parser context
1001  * @addr: VM address
1002  * @bo: resulting BO of the mapping found
1003  *
1004  * Search the buffer objects in the command submission context for a certain
1005  * virtual memory address. Returns allocation structure when found, NULL
1006  * otherwise.
1007  */
1008 struct amdgpu_bo_va_mapping *
1009 amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1010 		       uint64_t addr, struct amdgpu_bo **bo)
1011 {
1012 	struct amdgpu_bo_va_mapping *mapping;
1013 	unsigned i;
1014 
1015 	if (!parser->bo_list)
1016 		return NULL;
1017 
1018 	addr /= AMDGPU_GPU_PAGE_SIZE;
1019 
1020 	for (i = 0; i < parser->bo_list->num_entries; i++) {
1021 		struct amdgpu_bo_list_entry *lobj;
1022 
1023 		lobj = &parser->bo_list->array[i];
1024 		if (!lobj->bo_va)
1025 			continue;
1026 
1027 		list_for_each_entry(mapping, &lobj->bo_va->valids, list) {
1028 			if (mapping->it.start > addr ||
1029 			    addr > mapping->it.last)
1030 				continue;
1031 
1032 			*bo = lobj->bo_va->bo;
1033 			return mapping;
1034 		}
1035 
1036 		list_for_each_entry(mapping, &lobj->bo_va->invalids, list) {
1037 			if (mapping->it.start > addr ||
1038 			    addr > mapping->it.last)
1039 				continue;
1040 
1041 			*bo = lobj->bo_va->bo;
1042 			return mapping;
1043 		}
1044 	}
1045 
1046 	return NULL;
1047 }
1048