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 
28 #include <linux/file.h>
29 #include <linux/pagemap.h>
30 #include <linux/sync_file.h>
31 #include <linux/dma-buf.h>
32 
33 #include <drm/amdgpu_drm.h>
34 #include <drm/drm_syncobj.h>
35 #include <drm/ttm/ttm_tt.h>
36 
37 #include "amdgpu_cs.h"
38 #include "amdgpu.h"
39 #include "amdgpu_trace.h"
40 #include "amdgpu_gmc.h"
41 #include "amdgpu_gem.h"
42 #include "amdgpu_ras.h"
43 
amdgpu_cs_parser_init(struct amdgpu_cs_parser * p,struct amdgpu_device * adev,struct drm_file * filp,union drm_amdgpu_cs * cs)44 static int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p,
45 				 struct amdgpu_device *adev,
46 				 struct drm_file *filp,
47 				 union drm_amdgpu_cs *cs)
48 {
49 	struct amdgpu_fpriv *fpriv = filp->driver_priv;
50 
51 	if (cs->in.num_chunks == 0)
52 		return -EINVAL;
53 
54 	memset(p, 0, sizeof(*p));
55 	p->adev = adev;
56 	p->filp = filp;
57 
58 	p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
59 	if (!p->ctx)
60 		return -EINVAL;
61 
62 	if (atomic_read(&p->ctx->guilty)) {
63 		amdgpu_ctx_put(p->ctx);
64 		return -ECANCELED;
65 	}
66 
67 	amdgpu_sync_create(&p->sync);
68 	return 0;
69 }
70 
amdgpu_cs_job_idx(struct amdgpu_cs_parser * p,struct drm_amdgpu_cs_chunk_ib * chunk_ib)71 static int amdgpu_cs_job_idx(struct amdgpu_cs_parser *p,
72 			     struct drm_amdgpu_cs_chunk_ib *chunk_ib)
73 {
74 	struct drm_sched_entity *entity;
75 	unsigned int i;
76 	int r;
77 
78 	r = amdgpu_ctx_get_entity(p->ctx, chunk_ib->ip_type,
79 				  chunk_ib->ip_instance,
80 				  chunk_ib->ring, &entity);
81 	if (r)
82 		return r;
83 
84 	/*
85 	 * Abort if there is no run queue associated with this entity.
86 	 * Possibly because of disabled HW IP.
87 	 */
88 	if (entity->rq == NULL)
89 		return -EINVAL;
90 
91 	/* Check if we can add this IB to some existing job */
92 	for (i = 0; i < p->gang_size; ++i)
93 		if (p->entities[i] == entity)
94 			return i;
95 
96 	/* If not increase the gang size if possible */
97 	if (i == AMDGPU_CS_GANG_SIZE)
98 		return -EINVAL;
99 
100 	p->entities[i] = entity;
101 	p->gang_size = i + 1;
102 	return i;
103 }
104 
amdgpu_cs_p1_ib(struct amdgpu_cs_parser * p,struct drm_amdgpu_cs_chunk_ib * chunk_ib,unsigned int * num_ibs)105 static int amdgpu_cs_p1_ib(struct amdgpu_cs_parser *p,
106 			   struct drm_amdgpu_cs_chunk_ib *chunk_ib,
107 			   unsigned int *num_ibs)
108 {
109 	int r;
110 
111 	r = amdgpu_cs_job_idx(p, chunk_ib);
112 	if (r < 0)
113 		return r;
114 
115 	++(num_ibs[r]);
116 	p->gang_leader_idx = r;
117 	return 0;
118 }
119 
amdgpu_cs_p1_user_fence(struct amdgpu_cs_parser * p,struct drm_amdgpu_cs_chunk_fence * data,uint32_t * offset)120 static int amdgpu_cs_p1_user_fence(struct amdgpu_cs_parser *p,
121 				   struct drm_amdgpu_cs_chunk_fence *data,
122 				   uint32_t *offset)
123 {
124 	struct drm_gem_object *gobj;
125 	struct amdgpu_bo *bo;
126 	unsigned long size;
127 	int r;
128 
129 	gobj = drm_gem_object_lookup(p->filp, data->handle);
130 	if (gobj == NULL)
131 		return -EINVAL;
132 
133 	bo = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
134 	p->uf_entry.priority = 0;
135 	p->uf_entry.tv.bo = &bo->tbo;
136 	/* One for TTM and two for the CS job */
137 	p->uf_entry.tv.num_shared = 3;
138 
139 	drm_gem_object_put(gobj);
140 
141 	size = amdgpu_bo_size(bo);
142 	if (size != PAGE_SIZE || (data->offset + 8) > size) {
143 		r = -EINVAL;
144 		goto error_unref;
145 	}
146 
147 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
148 		r = -EINVAL;
149 		goto error_unref;
150 	}
151 
152 	*offset = data->offset;
153 
154 	return 0;
155 
156 error_unref:
157 	amdgpu_bo_unref(&bo);
158 	return r;
159 }
160 
amdgpu_cs_p1_bo_handles(struct amdgpu_cs_parser * p,struct drm_amdgpu_bo_list_in * data)161 static int amdgpu_cs_p1_bo_handles(struct amdgpu_cs_parser *p,
162 				   struct drm_amdgpu_bo_list_in *data)
163 {
164 	struct drm_amdgpu_bo_list_entry *info;
165 	int r;
166 
167 	r = amdgpu_bo_create_list_entry_array(data, &info);
168 	if (r)
169 		return r;
170 
171 	r = amdgpu_bo_list_create(p->adev, p->filp, info, data->bo_number,
172 				  &p->bo_list);
173 	if (r)
174 		goto error_free;
175 
176 	kvfree(info);
177 	return 0;
178 
179 error_free:
180 	kvfree(info);
181 
182 	return r;
183 }
184 
185 /* Copy the data from userspace and go over it the first time */
amdgpu_cs_pass1(struct amdgpu_cs_parser * p,union drm_amdgpu_cs * cs)186 static int amdgpu_cs_pass1(struct amdgpu_cs_parser *p,
187 			   union drm_amdgpu_cs *cs)
188 {
189 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
190 	unsigned int num_ibs[AMDGPU_CS_GANG_SIZE] = { };
191 	struct amdgpu_vm *vm = &fpriv->vm;
192 	uint64_t *chunk_array_user;
193 	uint64_t *chunk_array;
194 	uint32_t uf_offset = 0;
195 	unsigned int size;
196 	int ret;
197 	int i;
198 
199 	chunk_array = kvmalloc_array(cs->in.num_chunks, sizeof(uint64_t),
200 				     GFP_KERNEL);
201 	if (!chunk_array)
202 		return -ENOMEM;
203 
204 	/* get chunks */
205 	chunk_array_user = u64_to_user_ptr(cs->in.chunks);
206 	if (copy_from_user(chunk_array, chunk_array_user,
207 			   sizeof(uint64_t)*cs->in.num_chunks)) {
208 		ret = -EFAULT;
209 		goto free_chunk;
210 	}
211 
212 	p->nchunks = cs->in.num_chunks;
213 	p->chunks = kvmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
214 			    GFP_KERNEL);
215 	if (!p->chunks) {
216 		ret = -ENOMEM;
217 		goto free_chunk;
218 	}
219 
220 	for (i = 0; i < p->nchunks; i++) {
221 		struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
222 		struct drm_amdgpu_cs_chunk user_chunk;
223 		uint32_t __user *cdata;
224 
225 		chunk_ptr = u64_to_user_ptr(chunk_array[i]);
226 		if (copy_from_user(&user_chunk, chunk_ptr,
227 				       sizeof(struct drm_amdgpu_cs_chunk))) {
228 			ret = -EFAULT;
229 			i--;
230 			goto free_partial_kdata;
231 		}
232 		p->chunks[i].chunk_id = user_chunk.chunk_id;
233 		p->chunks[i].length_dw = user_chunk.length_dw;
234 
235 		size = p->chunks[i].length_dw;
236 		cdata = u64_to_user_ptr(user_chunk.chunk_data);
237 
238 		p->chunks[i].kdata = kvmalloc_array(size, sizeof(uint32_t),
239 						    GFP_KERNEL);
240 		if (p->chunks[i].kdata == NULL) {
241 			ret = -ENOMEM;
242 			i--;
243 			goto free_partial_kdata;
244 		}
245 		size *= sizeof(uint32_t);
246 		if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
247 			ret = -EFAULT;
248 			goto free_partial_kdata;
249 		}
250 
251 		/* Assume the worst on the following checks */
252 		ret = -EINVAL;
253 		switch (p->chunks[i].chunk_id) {
254 		case AMDGPU_CHUNK_ID_IB:
255 			if (size < sizeof(struct drm_amdgpu_cs_chunk_ib))
256 				goto free_partial_kdata;
257 
258 			ret = amdgpu_cs_p1_ib(p, p->chunks[i].kdata, num_ibs);
259 			if (ret)
260 				goto free_partial_kdata;
261 			break;
262 
263 		case AMDGPU_CHUNK_ID_FENCE:
264 			if (size < sizeof(struct drm_amdgpu_cs_chunk_fence))
265 				goto free_partial_kdata;
266 
267 			ret = amdgpu_cs_p1_user_fence(p, p->chunks[i].kdata,
268 						      &uf_offset);
269 			if (ret)
270 				goto free_partial_kdata;
271 			break;
272 
273 		case AMDGPU_CHUNK_ID_BO_HANDLES:
274 			if (size < sizeof(struct drm_amdgpu_bo_list_in))
275 				goto free_partial_kdata;
276 
277 			ret = amdgpu_cs_p1_bo_handles(p, p->chunks[i].kdata);
278 			if (ret)
279 				goto free_partial_kdata;
280 			break;
281 
282 		case AMDGPU_CHUNK_ID_DEPENDENCIES:
283 		case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
284 		case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
285 		case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
286 		case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
287 		case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
288 			break;
289 
290 		default:
291 			goto free_partial_kdata;
292 		}
293 	}
294 
295 	if (!p->gang_size) {
296 		ret = -EINVAL;
297 		goto free_partial_kdata;
298 	}
299 
300 	for (i = 0; i < p->gang_size; ++i) {
301 		ret = amdgpu_job_alloc(p->adev, vm, p->entities[i], vm,
302 				       num_ibs[i], &p->jobs[i]);
303 		if (ret)
304 			goto free_all_kdata;
305 	}
306 	p->gang_leader = p->jobs[p->gang_leader_idx];
307 
308 	if (p->ctx->vram_lost_counter != p->gang_leader->vram_lost_counter) {
309 		ret = -ECANCELED;
310 		goto free_all_kdata;
311 	}
312 
313 	if (p->uf_entry.tv.bo)
314 		p->gang_leader->uf_addr = uf_offset;
315 	kvfree(chunk_array);
316 
317 	/* Use this opportunity to fill in task info for the vm */
318 	amdgpu_vm_set_task_info(vm);
319 
320 	return 0;
321 
322 free_all_kdata:
323 	i = p->nchunks - 1;
324 free_partial_kdata:
325 	for (; i >= 0; i--)
326 		kvfree(p->chunks[i].kdata);
327 	kvfree(p->chunks);
328 	p->chunks = NULL;
329 	p->nchunks = 0;
330 free_chunk:
331 	kvfree(chunk_array);
332 
333 	return ret;
334 }
335 
amdgpu_cs_p2_ib(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk,unsigned int * ce_preempt,unsigned int * de_preempt)336 static int amdgpu_cs_p2_ib(struct amdgpu_cs_parser *p,
337 			   struct amdgpu_cs_chunk *chunk,
338 			   unsigned int *ce_preempt,
339 			   unsigned int *de_preempt)
340 {
341 	struct drm_amdgpu_cs_chunk_ib *chunk_ib = chunk->kdata;
342 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
343 	struct amdgpu_vm *vm = &fpriv->vm;
344 	struct amdgpu_ring *ring;
345 	struct amdgpu_job *job;
346 	struct amdgpu_ib *ib;
347 	int r;
348 
349 	r = amdgpu_cs_job_idx(p, chunk_ib);
350 	if (r < 0)
351 		return r;
352 
353 	job = p->jobs[r];
354 	ring = amdgpu_job_ring(job);
355 	ib = &job->ibs[job->num_ibs++];
356 
357 	/* MM engine doesn't support user fences */
358 	if (p->uf_entry.tv.bo && ring->funcs->no_user_fence)
359 		return -EINVAL;
360 
361 	if (chunk_ib->ip_type == AMDGPU_HW_IP_GFX &&
362 	    chunk_ib->flags & AMDGPU_IB_FLAG_PREEMPT) {
363 		if (chunk_ib->flags & AMDGPU_IB_FLAG_CE)
364 			(*ce_preempt)++;
365 		else
366 			(*de_preempt)++;
367 
368 		/* Each GFX command submit allows only 1 IB max
369 		 * preemptible for CE & DE */
370 		if (*ce_preempt > 1 || *de_preempt > 1)
371 			return -EINVAL;
372 	}
373 
374 	if (chunk_ib->flags & AMDGPU_IB_FLAG_PREAMBLE)
375 		job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT;
376 
377 	r =  amdgpu_ib_get(p->adev, vm, ring->funcs->parse_cs ?
378 			   chunk_ib->ib_bytes : 0,
379 			   AMDGPU_IB_POOL_DELAYED, ib);
380 	if (r) {
381 		DRM_ERROR("Failed to get ib !\n");
382 		return r;
383 	}
384 
385 	ib->gpu_addr = chunk_ib->va_start;
386 	ib->length_dw = chunk_ib->ib_bytes / 4;
387 	ib->flags = chunk_ib->flags;
388 	return 0;
389 }
390 
amdgpu_cs_p2_dependencies(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)391 static int amdgpu_cs_p2_dependencies(struct amdgpu_cs_parser *p,
392 				     struct amdgpu_cs_chunk *chunk)
393 {
394 	struct drm_amdgpu_cs_chunk_dep *deps = chunk->kdata;
395 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
396 	unsigned num_deps;
397 	int i, r;
398 
399 	num_deps = chunk->length_dw * 4 /
400 		sizeof(struct drm_amdgpu_cs_chunk_dep);
401 
402 	for (i = 0; i < num_deps; ++i) {
403 		struct amdgpu_ctx *ctx;
404 		struct drm_sched_entity *entity;
405 		struct dma_fence *fence;
406 
407 		ctx = amdgpu_ctx_get(fpriv, deps[i].ctx_id);
408 		if (ctx == NULL)
409 			return -EINVAL;
410 
411 		r = amdgpu_ctx_get_entity(ctx, deps[i].ip_type,
412 					  deps[i].ip_instance,
413 					  deps[i].ring, &entity);
414 		if (r) {
415 			amdgpu_ctx_put(ctx);
416 			return r;
417 		}
418 
419 		fence = amdgpu_ctx_get_fence(ctx, entity, deps[i].handle);
420 		amdgpu_ctx_put(ctx);
421 
422 		if (IS_ERR(fence))
423 			return PTR_ERR(fence);
424 		else if (!fence)
425 			continue;
426 
427 		if (chunk->chunk_id == AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES) {
428 			struct drm_sched_fence *s_fence;
429 			struct dma_fence *old = fence;
430 
431 			s_fence = to_drm_sched_fence(fence);
432 			fence = dma_fence_get(&s_fence->scheduled);
433 			dma_fence_put(old);
434 		}
435 
436 		r = amdgpu_sync_fence(&p->sync, fence);
437 		dma_fence_put(fence);
438 		if (r)
439 			return r;
440 	}
441 	return 0;
442 }
443 
amdgpu_syncobj_lookup_and_add(struct amdgpu_cs_parser * p,uint32_t handle,u64 point,u64 flags)444 static int amdgpu_syncobj_lookup_and_add(struct amdgpu_cs_parser *p,
445 					 uint32_t handle, u64 point,
446 					 u64 flags)
447 {
448 	struct dma_fence *fence;
449 	int r;
450 
451 	r = drm_syncobj_find_fence(p->filp, handle, point, flags, &fence);
452 	if (r) {
453 		DRM_ERROR("syncobj %u failed to find fence @ %llu (%d)!\n",
454 			  handle, point, r);
455 		return r;
456 	}
457 
458 	r = amdgpu_sync_fence(&p->sync, fence);
459 	dma_fence_put(fence);
460 	return r;
461 }
462 
amdgpu_cs_p2_syncobj_in(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)463 static int amdgpu_cs_p2_syncobj_in(struct amdgpu_cs_parser *p,
464 				   struct amdgpu_cs_chunk *chunk)
465 {
466 	struct drm_amdgpu_cs_chunk_sem *deps = chunk->kdata;
467 	unsigned num_deps;
468 	int i, r;
469 
470 	num_deps = chunk->length_dw * 4 /
471 		sizeof(struct drm_amdgpu_cs_chunk_sem);
472 	for (i = 0; i < num_deps; ++i) {
473 		r = amdgpu_syncobj_lookup_and_add(p, deps[i].handle, 0, 0);
474 		if (r)
475 			return r;
476 	}
477 
478 	return 0;
479 }
480 
amdgpu_cs_p2_syncobj_timeline_wait(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)481 static int amdgpu_cs_p2_syncobj_timeline_wait(struct amdgpu_cs_parser *p,
482 					      struct amdgpu_cs_chunk *chunk)
483 {
484 	struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps = chunk->kdata;
485 	unsigned num_deps;
486 	int i, r;
487 
488 	num_deps = chunk->length_dw * 4 /
489 		sizeof(struct drm_amdgpu_cs_chunk_syncobj);
490 	for (i = 0; i < num_deps; ++i) {
491 		r = amdgpu_syncobj_lookup_and_add(p, syncobj_deps[i].handle,
492 						  syncobj_deps[i].point,
493 						  syncobj_deps[i].flags);
494 		if (r)
495 			return r;
496 	}
497 
498 	return 0;
499 }
500 
amdgpu_cs_p2_syncobj_out(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)501 static int amdgpu_cs_p2_syncobj_out(struct amdgpu_cs_parser *p,
502 				    struct amdgpu_cs_chunk *chunk)
503 {
504 	struct drm_amdgpu_cs_chunk_sem *deps = chunk->kdata;
505 	unsigned num_deps;
506 	int i;
507 
508 	num_deps = chunk->length_dw * 4 /
509 		sizeof(struct drm_amdgpu_cs_chunk_sem);
510 
511 	if (p->post_deps)
512 		return -EINVAL;
513 
514 	p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
515 				     GFP_KERNEL);
516 	p->num_post_deps = 0;
517 
518 	if (!p->post_deps)
519 		return -ENOMEM;
520 
521 
522 	for (i = 0; i < num_deps; ++i) {
523 		p->post_deps[i].syncobj =
524 			drm_syncobj_find(p->filp, deps[i].handle);
525 		if (!p->post_deps[i].syncobj)
526 			return -EINVAL;
527 		p->post_deps[i].chain = NULL;
528 		p->post_deps[i].point = 0;
529 		p->num_post_deps++;
530 	}
531 
532 	return 0;
533 }
534 
amdgpu_cs_p2_syncobj_timeline_signal(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)535 static int amdgpu_cs_p2_syncobj_timeline_signal(struct amdgpu_cs_parser *p,
536 						struct amdgpu_cs_chunk *chunk)
537 {
538 	struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps = chunk->kdata;
539 	unsigned num_deps;
540 	int i;
541 
542 	num_deps = chunk->length_dw * 4 /
543 		sizeof(struct drm_amdgpu_cs_chunk_syncobj);
544 
545 	if (p->post_deps)
546 		return -EINVAL;
547 
548 	p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
549 				     GFP_KERNEL);
550 	p->num_post_deps = 0;
551 
552 	if (!p->post_deps)
553 		return -ENOMEM;
554 
555 	for (i = 0; i < num_deps; ++i) {
556 		struct amdgpu_cs_post_dep *dep = &p->post_deps[i];
557 
558 		dep->chain = NULL;
559 		if (syncobj_deps[i].point) {
560 			dep->chain = dma_fence_chain_alloc();
561 			if (!dep->chain)
562 				return -ENOMEM;
563 		}
564 
565 		dep->syncobj = drm_syncobj_find(p->filp,
566 						syncobj_deps[i].handle);
567 		if (!dep->syncobj) {
568 			dma_fence_chain_free(dep->chain);
569 			return -EINVAL;
570 		}
571 		dep->point = syncobj_deps[i].point;
572 		p->num_post_deps++;
573 	}
574 
575 	return 0;
576 }
577 
amdgpu_cs_pass2(struct amdgpu_cs_parser * p)578 static int amdgpu_cs_pass2(struct amdgpu_cs_parser *p)
579 {
580 	unsigned int ce_preempt = 0, de_preempt = 0;
581 	int i, r;
582 
583 	for (i = 0; i < p->nchunks; ++i) {
584 		struct amdgpu_cs_chunk *chunk;
585 
586 		chunk = &p->chunks[i];
587 
588 		switch (chunk->chunk_id) {
589 		case AMDGPU_CHUNK_ID_IB:
590 			r = amdgpu_cs_p2_ib(p, chunk, &ce_preempt, &de_preempt);
591 			if (r)
592 				return r;
593 			break;
594 		case AMDGPU_CHUNK_ID_DEPENDENCIES:
595 		case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
596 			r = amdgpu_cs_p2_dependencies(p, chunk);
597 			if (r)
598 				return r;
599 			break;
600 		case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
601 			r = amdgpu_cs_p2_syncobj_in(p, chunk);
602 			if (r)
603 				return r;
604 			break;
605 		case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
606 			r = amdgpu_cs_p2_syncobj_out(p, chunk);
607 			if (r)
608 				return r;
609 			break;
610 		case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
611 			r = amdgpu_cs_p2_syncobj_timeline_wait(p, chunk);
612 			if (r)
613 				return r;
614 			break;
615 		case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
616 			r = amdgpu_cs_p2_syncobj_timeline_signal(p, chunk);
617 			if (r)
618 				return r;
619 			break;
620 		}
621 	}
622 
623 	return 0;
624 }
625 
626 /* Convert microseconds to bytes. */
us_to_bytes(struct amdgpu_device * adev,s64 us)627 static u64 us_to_bytes(struct amdgpu_device *adev, s64 us)
628 {
629 	if (us <= 0 || !adev->mm_stats.log2_max_MBps)
630 		return 0;
631 
632 	/* Since accum_us is incremented by a million per second, just
633 	 * multiply it by the number of MB/s to get the number of bytes.
634 	 */
635 	return us << adev->mm_stats.log2_max_MBps;
636 }
637 
bytes_to_us(struct amdgpu_device * adev,u64 bytes)638 static s64 bytes_to_us(struct amdgpu_device *adev, u64 bytes)
639 {
640 	if (!adev->mm_stats.log2_max_MBps)
641 		return 0;
642 
643 	return bytes >> adev->mm_stats.log2_max_MBps;
644 }
645 
646 /* Returns how many bytes TTM can move right now. If no bytes can be moved,
647  * it returns 0. If it returns non-zero, it's OK to move at least one buffer,
648  * which means it can go over the threshold once. If that happens, the driver
649  * will be in debt and no other buffer migrations can be done until that debt
650  * is repaid.
651  *
652  * This approach allows moving a buffer of any size (it's important to allow
653  * that).
654  *
655  * The currency is simply time in microseconds and it increases as the clock
656  * ticks. The accumulated microseconds (us) are converted to bytes and
657  * returned.
658  */
amdgpu_cs_get_threshold_for_moves(struct amdgpu_device * adev,u64 * max_bytes,u64 * max_vis_bytes)659 static void amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev,
660 					      u64 *max_bytes,
661 					      u64 *max_vis_bytes)
662 {
663 	s64 time_us, increment_us;
664 	u64 free_vram, total_vram, used_vram;
665 	/* Allow a maximum of 200 accumulated ms. This is basically per-IB
666 	 * throttling.
667 	 *
668 	 * It means that in order to get full max MBps, at least 5 IBs per
669 	 * second must be submitted and not more than 200ms apart from each
670 	 * other.
671 	 */
672 	const s64 us_upper_bound = 200000;
673 
674 	if (!adev->mm_stats.log2_max_MBps) {
675 		*max_bytes = 0;
676 		*max_vis_bytes = 0;
677 		return;
678 	}
679 
680 	total_vram = adev->gmc.real_vram_size - atomic64_read(&adev->vram_pin_size);
681 	used_vram = ttm_resource_manager_usage(&adev->mman.vram_mgr.manager);
682 	free_vram = used_vram >= total_vram ? 0 : total_vram - used_vram;
683 
684 	spin_lock(&adev->mm_stats.lock);
685 
686 	/* Increase the amount of accumulated us. */
687 	time_us = ktime_to_us(ktime_get());
688 	increment_us = time_us - adev->mm_stats.last_update_us;
689 	adev->mm_stats.last_update_us = time_us;
690 	adev->mm_stats.accum_us = min(adev->mm_stats.accum_us + increment_us,
691 				      us_upper_bound);
692 
693 	/* This prevents the short period of low performance when the VRAM
694 	 * usage is low and the driver is in debt or doesn't have enough
695 	 * accumulated us to fill VRAM quickly.
696 	 *
697 	 * The situation can occur in these cases:
698 	 * - a lot of VRAM is freed by userspace
699 	 * - the presence of a big buffer causes a lot of evictions
700 	 *   (solution: split buffers into smaller ones)
701 	 *
702 	 * If 128 MB or 1/8th of VRAM is free, start filling it now by setting
703 	 * accum_us to a positive number.
704 	 */
705 	if (free_vram >= 128 * 1024 * 1024 || free_vram >= total_vram / 8) {
706 		s64 min_us;
707 
708 		/* Be more aggressive on dGPUs. Try to fill a portion of free
709 		 * VRAM now.
710 		 */
711 		if (!(adev->flags & AMD_IS_APU))
712 			min_us = bytes_to_us(adev, free_vram / 4);
713 		else
714 			min_us = 0; /* Reset accum_us on APUs. */
715 
716 		adev->mm_stats.accum_us = max(min_us, adev->mm_stats.accum_us);
717 	}
718 
719 	/* This is set to 0 if the driver is in debt to disallow (optional)
720 	 * buffer moves.
721 	 */
722 	*max_bytes = us_to_bytes(adev, adev->mm_stats.accum_us);
723 
724 	/* Do the same for visible VRAM if half of it is free */
725 	if (!amdgpu_gmc_vram_full_visible(&adev->gmc)) {
726 		u64 total_vis_vram = adev->gmc.visible_vram_size;
727 		u64 used_vis_vram =
728 		  amdgpu_vram_mgr_vis_usage(&adev->mman.vram_mgr);
729 
730 		if (used_vis_vram < total_vis_vram) {
731 			u64 free_vis_vram = total_vis_vram - used_vis_vram;
732 			adev->mm_stats.accum_us_vis = min(adev->mm_stats.accum_us_vis +
733 							  increment_us, us_upper_bound);
734 
735 			if (free_vis_vram >= total_vis_vram / 2)
736 				adev->mm_stats.accum_us_vis =
737 					max(bytes_to_us(adev, free_vis_vram / 2),
738 					    adev->mm_stats.accum_us_vis);
739 		}
740 
741 		*max_vis_bytes = us_to_bytes(adev, adev->mm_stats.accum_us_vis);
742 	} else {
743 		*max_vis_bytes = 0;
744 	}
745 
746 	spin_unlock(&adev->mm_stats.lock);
747 }
748 
749 /* Report how many bytes have really been moved for the last command
750  * submission. This can result in a debt that can stop buffer migrations
751  * temporarily.
752  */
amdgpu_cs_report_moved_bytes(struct amdgpu_device * adev,u64 num_bytes,u64 num_vis_bytes)753 void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev, u64 num_bytes,
754 				  u64 num_vis_bytes)
755 {
756 	spin_lock(&adev->mm_stats.lock);
757 	adev->mm_stats.accum_us -= bytes_to_us(adev, num_bytes);
758 	adev->mm_stats.accum_us_vis -= bytes_to_us(adev, num_vis_bytes);
759 	spin_unlock(&adev->mm_stats.lock);
760 }
761 
amdgpu_cs_bo_validate(void * param,struct amdgpu_bo * bo)762 static int amdgpu_cs_bo_validate(void *param, struct amdgpu_bo *bo)
763 {
764 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
765 	struct amdgpu_cs_parser *p = param;
766 	struct ttm_operation_ctx ctx = {
767 		.interruptible = true,
768 		.no_wait_gpu = false,
769 		.resv = bo->tbo.base.resv
770 	};
771 	uint32_t domain;
772 	int r;
773 
774 	if (bo->tbo.pin_count)
775 		return 0;
776 
777 	/* Don't move this buffer if we have depleted our allowance
778 	 * to move it. Don't move anything if the threshold is zero.
779 	 */
780 	if (p->bytes_moved < p->bytes_moved_threshold &&
781 	    (!bo->tbo.base.dma_buf ||
782 	    list_empty(&bo->tbo.base.dma_buf->attachments))) {
783 		if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
784 		    (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)) {
785 			/* And don't move a CPU_ACCESS_REQUIRED BO to limited
786 			 * visible VRAM if we've depleted our allowance to do
787 			 * that.
788 			 */
789 			if (p->bytes_moved_vis < p->bytes_moved_vis_threshold)
790 				domain = bo->preferred_domains;
791 			else
792 				domain = bo->allowed_domains;
793 		} else {
794 			domain = bo->preferred_domains;
795 		}
796 	} else {
797 		domain = bo->allowed_domains;
798 	}
799 
800 retry:
801 	amdgpu_bo_placement_from_domain(bo, domain);
802 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
803 
804 	p->bytes_moved += ctx.bytes_moved;
805 	if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
806 	    amdgpu_bo_in_cpu_visible_vram(bo))
807 		p->bytes_moved_vis += ctx.bytes_moved;
808 
809 	if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
810 		domain = bo->allowed_domains;
811 		goto retry;
812 	}
813 
814 	return r;
815 }
816 
amdgpu_cs_list_validate(struct amdgpu_cs_parser * p,struct list_head * validated)817 static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
818 			    struct list_head *validated)
819 {
820 	struct ttm_operation_ctx ctx = { true, false };
821 	struct amdgpu_bo_list_entry *lobj;
822 	int r;
823 
824 	list_for_each_entry(lobj, validated, tv.head) {
825 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(lobj->tv.bo);
826 		struct mm_struct *usermm;
827 
828 		usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
829 		if (usermm && usermm != current->mm)
830 			return -EPERM;
831 
832 		if (amdgpu_ttm_tt_is_userptr(bo->tbo.ttm) &&
833 		    lobj->user_invalidated && lobj->user_pages) {
834 			amdgpu_bo_placement_from_domain(bo,
835 							AMDGPU_GEM_DOMAIN_CPU);
836 			r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
837 			if (r)
838 				return r;
839 
840 			amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm,
841 						     lobj->user_pages);
842 		}
843 
844 		r = amdgpu_cs_bo_validate(p, bo);
845 		if (r)
846 			return r;
847 
848 		kvfree(lobj->user_pages);
849 		lobj->user_pages = NULL;
850 	}
851 	return 0;
852 }
853 
amdgpu_cs_parser_bos(struct amdgpu_cs_parser * p,union drm_amdgpu_cs * cs)854 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
855 				union drm_amdgpu_cs *cs)
856 {
857 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
858 	struct amdgpu_vm *vm = &fpriv->vm;
859 	struct amdgpu_bo_list_entry *e;
860 	struct list_head duplicates;
861 	unsigned int i;
862 	int r;
863 
864 	INIT_LIST_HEAD(&p->validated);
865 
866 	/* p->bo_list could already be assigned if AMDGPU_CHUNK_ID_BO_HANDLES is present */
867 	if (cs->in.bo_list_handle) {
868 		if (p->bo_list)
869 			return -EINVAL;
870 
871 		r = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle,
872 				       &p->bo_list);
873 		if (r)
874 			return r;
875 	} else if (!p->bo_list) {
876 		/* Create a empty bo_list when no handle is provided */
877 		r = amdgpu_bo_list_create(p->adev, p->filp, NULL, 0,
878 					  &p->bo_list);
879 		if (r)
880 			return r;
881 	}
882 
883 	mutex_lock(&p->bo_list->bo_list_mutex);
884 
885 	/* One for TTM and one for the CS job */
886 	amdgpu_bo_list_for_each_entry(e, p->bo_list)
887 		e->tv.num_shared = 2;
888 
889 	amdgpu_bo_list_get_list(p->bo_list, &p->validated);
890 
891 	INIT_LIST_HEAD(&duplicates);
892 	amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
893 
894 	if (p->uf_entry.tv.bo && !ttm_to_amdgpu_bo(p->uf_entry.tv.bo)->parent)
895 		list_add(&p->uf_entry.tv.head, &p->validated);
896 
897 	/* Get userptr backing pages. If pages are updated after registered
898 	 * in amdgpu_gem_userptr_ioctl(), amdgpu_cs_list_validate() will do
899 	 * amdgpu_ttm_backend_bind() to flush and invalidate new pages
900 	 */
901 	amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
902 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
903 		bool userpage_invalidated = false;
904 		int i;
905 
906 		e->user_pages = kvmalloc_array(bo->tbo.ttm->num_pages,
907 					sizeof(struct page *),
908 					GFP_KERNEL | __GFP_ZERO);
909 		if (!e->user_pages) {
910 			DRM_ERROR("kvmalloc_array failure\n");
911 			r = -ENOMEM;
912 			goto out_free_user_pages;
913 		}
914 
915 		r = amdgpu_ttm_tt_get_user_pages(bo, e->user_pages, &e->range);
916 		if (r) {
917 			kvfree(e->user_pages);
918 			e->user_pages = NULL;
919 			goto out_free_user_pages;
920 		}
921 
922 		for (i = 0; i < bo->tbo.ttm->num_pages; i++) {
923 			if (bo->tbo.ttm->pages[i] != e->user_pages[i]) {
924 				userpage_invalidated = true;
925 				break;
926 			}
927 		}
928 		e->user_invalidated = userpage_invalidated;
929 	}
930 
931 	r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
932 				   &duplicates);
933 	if (unlikely(r != 0)) {
934 		if (r != -ERESTARTSYS)
935 			DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
936 		goto out_free_user_pages;
937 	}
938 
939 	amdgpu_bo_list_for_each_entry(e, p->bo_list) {
940 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
941 
942 		e->bo_va = amdgpu_vm_bo_find(vm, bo);
943 	}
944 
945 	amdgpu_cs_get_threshold_for_moves(p->adev, &p->bytes_moved_threshold,
946 					  &p->bytes_moved_vis_threshold);
947 	p->bytes_moved = 0;
948 	p->bytes_moved_vis = 0;
949 
950 	r = amdgpu_vm_validate_pt_bos(p->adev, &fpriv->vm,
951 				      amdgpu_cs_bo_validate, p);
952 	if (r) {
953 		DRM_ERROR("amdgpu_vm_validate_pt_bos() failed.\n");
954 		goto error_validate;
955 	}
956 
957 	r = amdgpu_cs_list_validate(p, &duplicates);
958 	if (r)
959 		goto error_validate;
960 
961 	r = amdgpu_cs_list_validate(p, &p->validated);
962 	if (r)
963 		goto error_validate;
964 
965 	if (p->uf_entry.tv.bo) {
966 		struct amdgpu_bo *uf = ttm_to_amdgpu_bo(p->uf_entry.tv.bo);
967 
968 		r = amdgpu_ttm_alloc_gart(&uf->tbo);
969 		if (r)
970 			goto error_validate;
971 
972 		p->gang_leader->uf_addr += amdgpu_bo_gpu_offset(uf);
973 	}
974 
975 	amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved,
976 				     p->bytes_moved_vis);
977 
978 	for (i = 0; i < p->gang_size; ++i)
979 		amdgpu_job_set_resources(p->jobs[i], p->bo_list->gds_obj,
980 					 p->bo_list->gws_obj,
981 					 p->bo_list->oa_obj);
982 	return 0;
983 
984 error_validate:
985 	ttm_eu_backoff_reservation(&p->ticket, &p->validated);
986 
987 out_free_user_pages:
988 	amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
989 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
990 
991 		if (!e->user_pages)
992 			continue;
993 		amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm, e->range);
994 		kvfree(e->user_pages);
995 		e->user_pages = NULL;
996 		e->range = NULL;
997 	}
998 	mutex_unlock(&p->bo_list->bo_list_mutex);
999 	return r;
1000 }
1001 
trace_amdgpu_cs_ibs(struct amdgpu_cs_parser * p)1002 static void trace_amdgpu_cs_ibs(struct amdgpu_cs_parser *p)
1003 {
1004 	int i, j;
1005 
1006 	if (!trace_amdgpu_cs_enabled())
1007 		return;
1008 
1009 	for (i = 0; i < p->gang_size; ++i) {
1010 		struct amdgpu_job *job = p->jobs[i];
1011 
1012 		for (j = 0; j < job->num_ibs; ++j)
1013 			trace_amdgpu_cs(p, job, &job->ibs[j]);
1014 	}
1015 }
1016 
amdgpu_cs_patch_ibs(struct amdgpu_cs_parser * p,struct amdgpu_job * job)1017 static int amdgpu_cs_patch_ibs(struct amdgpu_cs_parser *p,
1018 			       struct amdgpu_job *job)
1019 {
1020 	struct amdgpu_ring *ring = amdgpu_job_ring(job);
1021 	unsigned int i;
1022 	int r;
1023 
1024 	/* Only for UVD/VCE VM emulation */
1025 	if (!ring->funcs->parse_cs && !ring->funcs->patch_cs_in_place)
1026 		return 0;
1027 
1028 	for (i = 0; i < job->num_ibs; ++i) {
1029 		struct amdgpu_ib *ib = &job->ibs[i];
1030 		struct amdgpu_bo_va_mapping *m;
1031 		struct amdgpu_bo *aobj;
1032 		uint64_t va_start;
1033 		uint8_t *kptr;
1034 
1035 		va_start = ib->gpu_addr & AMDGPU_GMC_HOLE_MASK;
1036 		r = amdgpu_cs_find_mapping(p, va_start, &aobj, &m);
1037 		if (r) {
1038 			DRM_ERROR("IB va_start is invalid\n");
1039 			return r;
1040 		}
1041 
1042 		if ((va_start + ib->length_dw * 4) >
1043 		    (m->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
1044 			DRM_ERROR("IB va_start+ib_bytes is invalid\n");
1045 			return -EINVAL;
1046 		}
1047 
1048 		/* the IB should be reserved at this point */
1049 		r = amdgpu_bo_kmap(aobj, (void **)&kptr);
1050 		if (r) {
1051 			return r;
1052 		}
1053 
1054 		kptr += va_start - (m->start * AMDGPU_GPU_PAGE_SIZE);
1055 
1056 		if (ring->funcs->parse_cs) {
1057 			memcpy(ib->ptr, kptr, ib->length_dw * 4);
1058 			amdgpu_bo_kunmap(aobj);
1059 
1060 			r = amdgpu_ring_parse_cs(ring, p, job, ib);
1061 			if (r)
1062 				return r;
1063 		} else {
1064 			ib->ptr = (uint32_t *)kptr;
1065 			r = amdgpu_ring_patch_cs_in_place(ring, p, job, ib);
1066 			amdgpu_bo_kunmap(aobj);
1067 			if (r)
1068 				return r;
1069 		}
1070 	}
1071 
1072 	return 0;
1073 }
1074 
amdgpu_cs_patch_jobs(struct amdgpu_cs_parser * p)1075 static int amdgpu_cs_patch_jobs(struct amdgpu_cs_parser *p)
1076 {
1077 	unsigned int i;
1078 	int r;
1079 
1080 	for (i = 0; i < p->gang_size; ++i) {
1081 		r = amdgpu_cs_patch_ibs(p, p->jobs[i]);
1082 		if (r)
1083 			return r;
1084 	}
1085 	return 0;
1086 }
1087 
amdgpu_cs_vm_handling(struct amdgpu_cs_parser * p)1088 static int amdgpu_cs_vm_handling(struct amdgpu_cs_parser *p)
1089 {
1090 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1091 	struct amdgpu_job *job = p->gang_leader;
1092 	struct amdgpu_device *adev = p->adev;
1093 	struct amdgpu_vm *vm = &fpriv->vm;
1094 	struct amdgpu_bo_list_entry *e;
1095 	struct amdgpu_bo_va *bo_va;
1096 	struct amdgpu_bo *bo;
1097 	unsigned int i;
1098 	int r;
1099 
1100 	r = amdgpu_vm_clear_freed(adev, vm, NULL);
1101 	if (r)
1102 		return r;
1103 
1104 	r = amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
1105 	if (r)
1106 		return r;
1107 
1108 	r = amdgpu_sync_fence(&p->sync, fpriv->prt_va->last_pt_update);
1109 	if (r)
1110 		return r;
1111 
1112 	if (fpriv->csa_va) {
1113 		bo_va = fpriv->csa_va;
1114 		BUG_ON(!bo_va);
1115 		r = amdgpu_vm_bo_update(adev, bo_va, false);
1116 		if (r)
1117 			return r;
1118 
1119 		r = amdgpu_sync_fence(&p->sync, bo_va->last_pt_update);
1120 		if (r)
1121 			return r;
1122 	}
1123 
1124 	amdgpu_bo_list_for_each_entry(e, p->bo_list) {
1125 		/* ignore duplicates */
1126 		bo = ttm_to_amdgpu_bo(e->tv.bo);
1127 		if (!bo)
1128 			continue;
1129 
1130 		bo_va = e->bo_va;
1131 		if (bo_va == NULL)
1132 			continue;
1133 
1134 		r = amdgpu_vm_bo_update(adev, bo_va, false);
1135 		if (r)
1136 			return r;
1137 
1138 		r = amdgpu_sync_fence(&p->sync, bo_va->last_pt_update);
1139 		if (r)
1140 			return r;
1141 	}
1142 
1143 	r = amdgpu_vm_handle_moved(adev, vm);
1144 	if (r)
1145 		return r;
1146 
1147 	r = amdgpu_vm_update_pdes(adev, vm, false);
1148 	if (r)
1149 		return r;
1150 
1151 	r = amdgpu_sync_fence(&p->sync, vm->last_update);
1152 	if (r)
1153 		return r;
1154 
1155 	for (i = 0; i < p->gang_size; ++i) {
1156 		job = p->jobs[i];
1157 
1158 		if (!job->vm)
1159 			continue;
1160 
1161 		job->vm_pd_addr = amdgpu_gmc_pd_addr(vm->root.bo);
1162 	}
1163 
1164 	if (amdgpu_vm_debug) {
1165 		/* Invalidate all BOs to test for userspace bugs */
1166 		amdgpu_bo_list_for_each_entry(e, p->bo_list) {
1167 			struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
1168 
1169 			/* ignore duplicates */
1170 			if (!bo)
1171 				continue;
1172 
1173 			amdgpu_vm_bo_invalidate(adev, bo, false);
1174 		}
1175 	}
1176 
1177 	return 0;
1178 }
1179 
amdgpu_cs_sync_rings(struct amdgpu_cs_parser * p)1180 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
1181 {
1182 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1183 	struct drm_gpu_scheduler *sched;
1184 	struct amdgpu_bo_list_entry *e;
1185 	struct dma_fence *fence;
1186 	unsigned int i;
1187 	int r;
1188 
1189 	r = amdgpu_ctx_wait_prev_fence(p->ctx, p->entities[p->gang_leader_idx]);
1190 	if (r) {
1191 		if (r != -ERESTARTSYS)
1192 			DRM_ERROR("amdgpu_ctx_wait_prev_fence failed.\n");
1193 		return r;
1194 	}
1195 
1196 	list_for_each_entry(e, &p->validated, tv.head) {
1197 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
1198 		struct dma_resv *resv = bo->tbo.base.resv;
1199 		enum amdgpu_sync_mode sync_mode;
1200 
1201 		sync_mode = amdgpu_bo_explicit_sync(bo) ?
1202 			AMDGPU_SYNC_EXPLICIT : AMDGPU_SYNC_NE_OWNER;
1203 		r = amdgpu_sync_resv(p->adev, &p->sync, resv, sync_mode,
1204 				     &fpriv->vm);
1205 		if (r)
1206 			return r;
1207 	}
1208 
1209 	for (i = 0; i < p->gang_size; ++i) {
1210 		r = amdgpu_sync_push_to_job(&p->sync, p->jobs[i]);
1211 		if (r)
1212 			return r;
1213 	}
1214 
1215 	sched = p->gang_leader->base.entity->rq->sched;
1216 	while ((fence = amdgpu_sync_get_fence(&p->sync))) {
1217 		struct drm_sched_fence *s_fence = to_drm_sched_fence(fence);
1218 
1219 		/*
1220 		 * When we have an dependency it might be necessary to insert a
1221 		 * pipeline sync to make sure that all caches etc are flushed and the
1222 		 * next job actually sees the results from the previous one
1223 		 * before we start executing on the same scheduler ring.
1224 		 */
1225 		if (!s_fence || s_fence->sched != sched) {
1226 			dma_fence_put(fence);
1227 			continue;
1228 		}
1229 
1230 		r = amdgpu_sync_fence(&p->gang_leader->explicit_sync, fence);
1231 		dma_fence_put(fence);
1232 		if (r)
1233 			return r;
1234 	}
1235 	return 0;
1236 }
1237 
amdgpu_cs_post_dependencies(struct amdgpu_cs_parser * p)1238 static void amdgpu_cs_post_dependencies(struct amdgpu_cs_parser *p)
1239 {
1240 	int i;
1241 
1242 	for (i = 0; i < p->num_post_deps; ++i) {
1243 		if (p->post_deps[i].chain && p->post_deps[i].point) {
1244 			drm_syncobj_add_point(p->post_deps[i].syncobj,
1245 					      p->post_deps[i].chain,
1246 					      p->fence, p->post_deps[i].point);
1247 			p->post_deps[i].chain = NULL;
1248 		} else {
1249 			drm_syncobj_replace_fence(p->post_deps[i].syncobj,
1250 						  p->fence);
1251 		}
1252 	}
1253 }
1254 
amdgpu_cs_submit(struct amdgpu_cs_parser * p,union drm_amdgpu_cs * cs)1255 static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
1256 			    union drm_amdgpu_cs *cs)
1257 {
1258 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1259 	struct amdgpu_job *leader = p->gang_leader;
1260 	struct amdgpu_bo_list_entry *e;
1261 	unsigned int i;
1262 	uint64_t seq;
1263 	int r;
1264 
1265 	for (i = 0; i < p->gang_size; ++i)
1266 		drm_sched_job_arm(&p->jobs[i]->base);
1267 
1268 	for (i = 0; i < p->gang_size; ++i) {
1269 		struct dma_fence *fence;
1270 
1271 		if (p->jobs[i] == leader)
1272 			continue;
1273 
1274 		fence = &p->jobs[i]->base.s_fence->scheduled;
1275 		dma_fence_get(fence);
1276 		r = drm_sched_job_add_dependency(&leader->base, fence);
1277 		if (r) {
1278 			dma_fence_put(fence);
1279 			goto error_cleanup;
1280 		}
1281 	}
1282 
1283 	if (p->gang_size > 1) {
1284 		for (i = 0; i < p->gang_size; ++i)
1285 			amdgpu_job_set_gang_leader(p->jobs[i], leader);
1286 	}
1287 
1288 	/* No memory allocation is allowed while holding the notifier lock.
1289 	 * The lock is held until amdgpu_cs_submit is finished and fence is
1290 	 * added to BOs.
1291 	 */
1292 	mutex_lock(&p->adev->notifier_lock);
1293 
1294 	/* If userptr are invalidated after amdgpu_cs_parser_bos(), return
1295 	 * -EAGAIN, drmIoctl in libdrm will restart the amdgpu_cs_ioctl.
1296 	 */
1297 	r = 0;
1298 	amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
1299 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
1300 
1301 		r |= !amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm, e->range);
1302 		e->range = NULL;
1303 	}
1304 	if (r) {
1305 		r = -EAGAIN;
1306 		goto error_unlock;
1307 	}
1308 
1309 	p->fence = dma_fence_get(&leader->base.s_fence->finished);
1310 	list_for_each_entry(e, &p->validated, tv.head) {
1311 
1312 		/* Everybody except for the gang leader uses READ */
1313 		for (i = 0; i < p->gang_size; ++i) {
1314 			if (p->jobs[i] == leader)
1315 				continue;
1316 
1317 			dma_resv_add_fence(e->tv.bo->base.resv,
1318 					   &p->jobs[i]->base.s_fence->finished,
1319 					   DMA_RESV_USAGE_READ);
1320 		}
1321 
1322 		/* The gang leader is remembered as writer */
1323 		e->tv.num_shared = 0;
1324 	}
1325 
1326 	seq = amdgpu_ctx_add_fence(p->ctx, p->entities[p->gang_leader_idx],
1327 				   p->fence);
1328 	amdgpu_cs_post_dependencies(p);
1329 
1330 	if ((leader->preamble_status & AMDGPU_PREAMBLE_IB_PRESENT) &&
1331 	    !p->ctx->preamble_presented) {
1332 		leader->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
1333 		p->ctx->preamble_presented = true;
1334 	}
1335 
1336 	cs->out.handle = seq;
1337 	leader->uf_sequence = seq;
1338 
1339 	amdgpu_vm_bo_trace_cs(&fpriv->vm, &p->ticket);
1340 	for (i = 0; i < p->gang_size; ++i) {
1341 		amdgpu_job_free_resources(p->jobs[i]);
1342 		trace_amdgpu_cs_ioctl(p->jobs[i]);
1343 		drm_sched_entity_push_job(&p->jobs[i]->base);
1344 		p->jobs[i] = NULL;
1345 	}
1346 
1347 	amdgpu_vm_move_to_lru_tail(p->adev, &fpriv->vm);
1348 	ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);
1349 
1350 	mutex_unlock(&p->adev->notifier_lock);
1351 	mutex_unlock(&p->bo_list->bo_list_mutex);
1352 	return 0;
1353 
1354 error_unlock:
1355 	mutex_unlock(&p->adev->notifier_lock);
1356 
1357 error_cleanup:
1358 	for (i = 0; i < p->gang_size; ++i)
1359 		drm_sched_job_cleanup(&p->jobs[i]->base);
1360 	return r;
1361 }
1362 
1363 /* Cleanup the parser structure */
amdgpu_cs_parser_fini(struct amdgpu_cs_parser * parser)1364 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser)
1365 {
1366 	unsigned i;
1367 
1368 	amdgpu_sync_free(&parser->sync);
1369 	for (i = 0; i < parser->num_post_deps; i++) {
1370 		drm_syncobj_put(parser->post_deps[i].syncobj);
1371 		kfree(parser->post_deps[i].chain);
1372 	}
1373 	kfree(parser->post_deps);
1374 
1375 	dma_fence_put(parser->fence);
1376 
1377 	if (parser->ctx)
1378 		amdgpu_ctx_put(parser->ctx);
1379 	if (parser->bo_list)
1380 		amdgpu_bo_list_put(parser->bo_list);
1381 
1382 	for (i = 0; i < parser->nchunks; i++)
1383 		kvfree(parser->chunks[i].kdata);
1384 	kvfree(parser->chunks);
1385 	for (i = 0; i < parser->gang_size; ++i) {
1386 		if (parser->jobs[i])
1387 			amdgpu_job_free(parser->jobs[i]);
1388 	}
1389 	if (parser->uf_entry.tv.bo) {
1390 		struct amdgpu_bo *uf = ttm_to_amdgpu_bo(parser->uf_entry.tv.bo);
1391 
1392 		amdgpu_bo_unref(&uf);
1393 	}
1394 }
1395 
amdgpu_cs_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)1396 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
1397 {
1398 	struct amdgpu_device *adev = drm_to_adev(dev);
1399 	struct amdgpu_cs_parser parser;
1400 	int r;
1401 
1402 	if (amdgpu_ras_intr_triggered())
1403 		return -EHWPOISON;
1404 
1405 	if (!adev->accel_working)
1406 		return -EBUSY;
1407 
1408 	r = amdgpu_cs_parser_init(&parser, adev, filp, data);
1409 	if (r) {
1410 		if (printk_ratelimit())
1411 			DRM_ERROR("Failed to initialize parser %d!\n", r);
1412 		return r;
1413 	}
1414 
1415 	r = amdgpu_cs_pass1(&parser, data);
1416 	if (r)
1417 		goto error_fini;
1418 
1419 	r = amdgpu_cs_pass2(&parser);
1420 	if (r)
1421 		goto error_fini;
1422 
1423 	r = amdgpu_cs_parser_bos(&parser, data);
1424 	if (r) {
1425 		if (r == -ENOMEM)
1426 			DRM_ERROR("Not enough memory for command submission!\n");
1427 		else if (r != -ERESTARTSYS && r != -EAGAIN)
1428 			DRM_ERROR("Failed to process the buffer list %d!\n", r);
1429 		goto error_fini;
1430 	}
1431 
1432 	r = amdgpu_cs_patch_jobs(&parser);
1433 	if (r)
1434 		goto error_backoff;
1435 
1436 	r = amdgpu_cs_vm_handling(&parser);
1437 	if (r)
1438 		goto error_backoff;
1439 
1440 	r = amdgpu_cs_sync_rings(&parser);
1441 	if (r)
1442 		goto error_backoff;
1443 
1444 	trace_amdgpu_cs_ibs(&parser);
1445 
1446 	r = amdgpu_cs_submit(&parser, data);
1447 	if (r)
1448 		goto error_backoff;
1449 
1450 	amdgpu_cs_parser_fini(&parser);
1451 	return 0;
1452 
1453 error_backoff:
1454 	ttm_eu_backoff_reservation(&parser.ticket, &parser.validated);
1455 	mutex_unlock(&parser.bo_list->bo_list_mutex);
1456 
1457 error_fini:
1458 	amdgpu_cs_parser_fini(&parser);
1459 	return r;
1460 }
1461 
1462 /**
1463  * amdgpu_cs_wait_ioctl - wait for a command submission to finish
1464  *
1465  * @dev: drm device
1466  * @data: data from userspace
1467  * @filp: file private
1468  *
1469  * Wait for the command submission identified by handle to finish.
1470  */
amdgpu_cs_wait_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)1471 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
1472 			 struct drm_file *filp)
1473 {
1474 	union drm_amdgpu_wait_cs *wait = data;
1475 	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1476 	struct drm_sched_entity *entity;
1477 	struct amdgpu_ctx *ctx;
1478 	struct dma_fence *fence;
1479 	long r;
1480 
1481 	ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
1482 	if (ctx == NULL)
1483 		return -EINVAL;
1484 
1485 	r = amdgpu_ctx_get_entity(ctx, wait->in.ip_type, wait->in.ip_instance,
1486 				  wait->in.ring, &entity);
1487 	if (r) {
1488 		amdgpu_ctx_put(ctx);
1489 		return r;
1490 	}
1491 
1492 	fence = amdgpu_ctx_get_fence(ctx, entity, wait->in.handle);
1493 	if (IS_ERR(fence))
1494 		r = PTR_ERR(fence);
1495 	else if (fence) {
1496 		r = dma_fence_wait_timeout(fence, true, timeout);
1497 		if (r > 0 && fence->error)
1498 			r = fence->error;
1499 		dma_fence_put(fence);
1500 	} else
1501 		r = 1;
1502 
1503 	amdgpu_ctx_put(ctx);
1504 	if (r < 0)
1505 		return r;
1506 
1507 	memset(wait, 0, sizeof(*wait));
1508 	wait->out.status = (r == 0);
1509 
1510 	return 0;
1511 }
1512 
1513 /**
1514  * amdgpu_cs_get_fence - helper to get fence from drm_amdgpu_fence
1515  *
1516  * @adev: amdgpu device
1517  * @filp: file private
1518  * @user: drm_amdgpu_fence copied from user space
1519  */
amdgpu_cs_get_fence(struct amdgpu_device * adev,struct drm_file * filp,struct drm_amdgpu_fence * user)1520 static struct dma_fence *amdgpu_cs_get_fence(struct amdgpu_device *adev,
1521 					     struct drm_file *filp,
1522 					     struct drm_amdgpu_fence *user)
1523 {
1524 	struct drm_sched_entity *entity;
1525 	struct amdgpu_ctx *ctx;
1526 	struct dma_fence *fence;
1527 	int r;
1528 
1529 	ctx = amdgpu_ctx_get(filp->driver_priv, user->ctx_id);
1530 	if (ctx == NULL)
1531 		return ERR_PTR(-EINVAL);
1532 
1533 	r = amdgpu_ctx_get_entity(ctx, user->ip_type, user->ip_instance,
1534 				  user->ring, &entity);
1535 	if (r) {
1536 		amdgpu_ctx_put(ctx);
1537 		return ERR_PTR(r);
1538 	}
1539 
1540 	fence = amdgpu_ctx_get_fence(ctx, entity, user->seq_no);
1541 	amdgpu_ctx_put(ctx);
1542 
1543 	return fence;
1544 }
1545 
amdgpu_cs_fence_to_handle_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)1546 int amdgpu_cs_fence_to_handle_ioctl(struct drm_device *dev, void *data,
1547 				    struct drm_file *filp)
1548 {
1549 	struct amdgpu_device *adev = drm_to_adev(dev);
1550 	union drm_amdgpu_fence_to_handle *info = data;
1551 	struct dma_fence *fence;
1552 	struct drm_syncobj *syncobj;
1553 	struct sync_file *sync_file;
1554 	int fd, r;
1555 
1556 	fence = amdgpu_cs_get_fence(adev, filp, &info->in.fence);
1557 	if (IS_ERR(fence))
1558 		return PTR_ERR(fence);
1559 
1560 	if (!fence)
1561 		fence = dma_fence_get_stub();
1562 
1563 	switch (info->in.what) {
1564 	case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ:
1565 		r = drm_syncobj_create(&syncobj, 0, fence);
1566 		dma_fence_put(fence);
1567 		if (r)
1568 			return r;
1569 		r = drm_syncobj_get_handle(filp, syncobj, &info->out.handle);
1570 		drm_syncobj_put(syncobj);
1571 		return r;
1572 
1573 	case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ_FD:
1574 		r = drm_syncobj_create(&syncobj, 0, fence);
1575 		dma_fence_put(fence);
1576 		if (r)
1577 			return r;
1578 		r = drm_syncobj_get_fd(syncobj, (int *)&info->out.handle);
1579 		drm_syncobj_put(syncobj);
1580 		return r;
1581 
1582 	case AMDGPU_FENCE_TO_HANDLE_GET_SYNC_FILE_FD:
1583 		fd = get_unused_fd_flags(O_CLOEXEC);
1584 		if (fd < 0) {
1585 			dma_fence_put(fence);
1586 			return fd;
1587 		}
1588 
1589 		sync_file = sync_file_create(fence);
1590 		dma_fence_put(fence);
1591 		if (!sync_file) {
1592 			put_unused_fd(fd);
1593 			return -ENOMEM;
1594 		}
1595 
1596 		fd_install(fd, sync_file->file);
1597 		info->out.handle = fd;
1598 		return 0;
1599 
1600 	default:
1601 		dma_fence_put(fence);
1602 		return -EINVAL;
1603 	}
1604 }
1605 
1606 /**
1607  * amdgpu_cs_wait_all_fences - wait on all fences to signal
1608  *
1609  * @adev: amdgpu device
1610  * @filp: file private
1611  * @wait: wait parameters
1612  * @fences: array of drm_amdgpu_fence
1613  */
amdgpu_cs_wait_all_fences(struct amdgpu_device * adev,struct drm_file * filp,union drm_amdgpu_wait_fences * wait,struct drm_amdgpu_fence * fences)1614 static int amdgpu_cs_wait_all_fences(struct amdgpu_device *adev,
1615 				     struct drm_file *filp,
1616 				     union drm_amdgpu_wait_fences *wait,
1617 				     struct drm_amdgpu_fence *fences)
1618 {
1619 	uint32_t fence_count = wait->in.fence_count;
1620 	unsigned int i;
1621 	long r = 1;
1622 
1623 	for (i = 0; i < fence_count; i++) {
1624 		struct dma_fence *fence;
1625 		unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1626 
1627 		fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1628 		if (IS_ERR(fence))
1629 			return PTR_ERR(fence);
1630 		else if (!fence)
1631 			continue;
1632 
1633 		r = dma_fence_wait_timeout(fence, true, timeout);
1634 		dma_fence_put(fence);
1635 		if (r < 0)
1636 			return r;
1637 
1638 		if (r == 0)
1639 			break;
1640 
1641 		if (fence->error)
1642 			return fence->error;
1643 	}
1644 
1645 	memset(wait, 0, sizeof(*wait));
1646 	wait->out.status = (r > 0);
1647 
1648 	return 0;
1649 }
1650 
1651 /**
1652  * amdgpu_cs_wait_any_fence - wait on any fence to signal
1653  *
1654  * @adev: amdgpu device
1655  * @filp: file private
1656  * @wait: wait parameters
1657  * @fences: array of drm_amdgpu_fence
1658  */
amdgpu_cs_wait_any_fence(struct amdgpu_device * adev,struct drm_file * filp,union drm_amdgpu_wait_fences * wait,struct drm_amdgpu_fence * fences)1659 static int amdgpu_cs_wait_any_fence(struct amdgpu_device *adev,
1660 				    struct drm_file *filp,
1661 				    union drm_amdgpu_wait_fences *wait,
1662 				    struct drm_amdgpu_fence *fences)
1663 {
1664 	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1665 	uint32_t fence_count = wait->in.fence_count;
1666 	uint32_t first = ~0;
1667 	struct dma_fence **array;
1668 	unsigned int i;
1669 	long r;
1670 
1671 	/* Prepare the fence array */
1672 	array = kcalloc(fence_count, sizeof(struct dma_fence *), GFP_KERNEL);
1673 
1674 	if (array == NULL)
1675 		return -ENOMEM;
1676 
1677 	for (i = 0; i < fence_count; i++) {
1678 		struct dma_fence *fence;
1679 
1680 		fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1681 		if (IS_ERR(fence)) {
1682 			r = PTR_ERR(fence);
1683 			goto err_free_fence_array;
1684 		} else if (fence) {
1685 			array[i] = fence;
1686 		} else { /* NULL, the fence has been already signaled */
1687 			r = 1;
1688 			first = i;
1689 			goto out;
1690 		}
1691 	}
1692 
1693 	r = dma_fence_wait_any_timeout(array, fence_count, true, timeout,
1694 				       &first);
1695 	if (r < 0)
1696 		goto err_free_fence_array;
1697 
1698 out:
1699 	memset(wait, 0, sizeof(*wait));
1700 	wait->out.status = (r > 0);
1701 	wait->out.first_signaled = first;
1702 
1703 	if (first < fence_count && array[first])
1704 		r = array[first]->error;
1705 	else
1706 		r = 0;
1707 
1708 err_free_fence_array:
1709 	for (i = 0; i < fence_count; i++)
1710 		dma_fence_put(array[i]);
1711 	kfree(array);
1712 
1713 	return r;
1714 }
1715 
1716 /**
1717  * amdgpu_cs_wait_fences_ioctl - wait for multiple command submissions to finish
1718  *
1719  * @dev: drm device
1720  * @data: data from userspace
1721  * @filp: file private
1722  */
amdgpu_cs_wait_fences_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)1723 int amdgpu_cs_wait_fences_ioctl(struct drm_device *dev, void *data,
1724 				struct drm_file *filp)
1725 {
1726 	struct amdgpu_device *adev = drm_to_adev(dev);
1727 	union drm_amdgpu_wait_fences *wait = data;
1728 	uint32_t fence_count = wait->in.fence_count;
1729 	struct drm_amdgpu_fence *fences_user;
1730 	struct drm_amdgpu_fence *fences;
1731 	int r;
1732 
1733 	/* Get the fences from userspace */
1734 	fences = kmalloc_array(fence_count, sizeof(struct drm_amdgpu_fence),
1735 			GFP_KERNEL);
1736 	if (fences == NULL)
1737 		return -ENOMEM;
1738 
1739 	fences_user = u64_to_user_ptr(wait->in.fences);
1740 	if (copy_from_user(fences, fences_user,
1741 		sizeof(struct drm_amdgpu_fence) * fence_count)) {
1742 		r = -EFAULT;
1743 		goto err_free_fences;
1744 	}
1745 
1746 	if (wait->in.wait_all)
1747 		r = amdgpu_cs_wait_all_fences(adev, filp, wait, fences);
1748 	else
1749 		r = amdgpu_cs_wait_any_fence(adev, filp, wait, fences);
1750 
1751 err_free_fences:
1752 	kfree(fences);
1753 
1754 	return r;
1755 }
1756 
1757 /**
1758  * amdgpu_cs_find_mapping - find bo_va for VM address
1759  *
1760  * @parser: command submission parser context
1761  * @addr: VM address
1762  * @bo: resulting BO of the mapping found
1763  * @map: Placeholder to return found BO mapping
1764  *
1765  * Search the buffer objects in the command submission context for a certain
1766  * virtual memory address. Returns allocation structure when found, NULL
1767  * otherwise.
1768  */
amdgpu_cs_find_mapping(struct amdgpu_cs_parser * parser,uint64_t addr,struct amdgpu_bo ** bo,struct amdgpu_bo_va_mapping ** map)1769 int amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1770 			   uint64_t addr, struct amdgpu_bo **bo,
1771 			   struct amdgpu_bo_va_mapping **map)
1772 {
1773 	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
1774 	struct ttm_operation_ctx ctx = { false, false };
1775 	struct amdgpu_vm *vm = &fpriv->vm;
1776 	struct amdgpu_bo_va_mapping *mapping;
1777 	int r;
1778 
1779 	addr /= AMDGPU_GPU_PAGE_SIZE;
1780 
1781 	mapping = amdgpu_vm_bo_lookup_mapping(vm, addr);
1782 	if (!mapping || !mapping->bo_va || !mapping->bo_va->base.bo)
1783 		return -EINVAL;
1784 
1785 	*bo = mapping->bo_va->base.bo;
1786 	*map = mapping;
1787 
1788 	/* Double check that the BO is reserved by this CS */
1789 	if (dma_resv_locking_ctx((*bo)->tbo.base.resv) != &parser->ticket)
1790 		return -EINVAL;
1791 
1792 	if (!((*bo)->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)) {
1793 		(*bo)->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1794 		amdgpu_bo_placement_from_domain(*bo, (*bo)->allowed_domains);
1795 		r = ttm_bo_validate(&(*bo)->tbo, &(*bo)->placement, &ctx);
1796 		if (r)
1797 			return r;
1798 	}
1799 
1800 	return amdgpu_ttm_alloc_gart(&(*bo)->tbo);
1801 }
1802