1 /*
2  * Copyright 2009 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
7  * "Software"), to deal in the Software without restriction, including
8  * without limitation the rights to use, copy, modify, merge, publish,
9  * distribute, sub license, and/or sell copies of the Software, and to
10  * permit persons to whom the Software is furnished to do so, subject to
11  * the following conditions:
12  *
13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
20  *
21  * The above copyright notice and this permission notice (including the
22  * next paragraph) shall be included in all copies or substantial portions
23  * of the Software.
24  *
25  */
26 /*
27  * Authors:
28  *    Jerome Glisse <glisse@freedesktop.org>
29  *    Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
30  *    Dave Airlie
31  */
32 #include <linux/list.h>
33 #include <linux/slab.h>
34 #include <drm/drmP.h>
35 #include <drm/amdgpu_drm.h>
36 #include <drm/drm_cache.h>
37 #include "amdgpu.h"
38 #include "amdgpu_trace.h"
39 #include "amdgpu_amdkfd.h"
40 
41 /**
42  * DOC: amdgpu_object
43  *
44  * This defines the interfaces to operate on an &amdgpu_bo buffer object which
45  * represents memory used by driver (VRAM, system memory, etc.). The driver
46  * provides DRM/GEM APIs to userspace. DRM/GEM APIs then use these interfaces
47  * to create/destroy/set buffer object which are then managed by the kernel TTM
48  * memory manager.
49  * The interfaces are also used internally by kernel clients, including gfx,
50  * uvd, etc. for kernel managed allocations used by the GPU.
51  *
52  */
53 
54 /**
55  * amdgpu_bo_subtract_pin_size - Remove BO from pin_size accounting
56  *
57  * @bo: &amdgpu_bo buffer object
58  *
59  * This function is called when a BO stops being pinned, and updates the
60  * &amdgpu_device pin_size values accordingly.
61  */
62 static void amdgpu_bo_subtract_pin_size(struct amdgpu_bo *bo)
63 {
64 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
65 
66 	if (bo->tbo.mem.mem_type == TTM_PL_VRAM) {
67 		atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
68 		atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
69 			     &adev->visible_pin_size);
70 	} else if (bo->tbo.mem.mem_type == TTM_PL_TT) {
71 		atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size);
72 	}
73 }
74 
75 static void amdgpu_bo_destroy(struct ttm_buffer_object *tbo)
76 {
77 	struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
78 	struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
79 
80 	if (bo->pin_count > 0)
81 		amdgpu_bo_subtract_pin_size(bo);
82 
83 	if (bo->kfd_bo)
84 		amdgpu_amdkfd_unreserve_memory_limit(bo);
85 
86 	amdgpu_bo_kunmap(bo);
87 
88 	if (bo->gem_base.import_attach)
89 		drm_prime_gem_destroy(&bo->gem_base, bo->tbo.sg);
90 	drm_gem_object_release(&bo->gem_base);
91 	amdgpu_bo_unref(&bo->parent);
92 	if (!list_empty(&bo->shadow_list)) {
93 		mutex_lock(&adev->shadow_list_lock);
94 		list_del_init(&bo->shadow_list);
95 		mutex_unlock(&adev->shadow_list_lock);
96 	}
97 	kfree(bo->metadata);
98 	kfree(bo);
99 }
100 
101 /**
102  * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
103  * @bo: buffer object to be checked
104  *
105  * Uses destroy function associated with the object to determine if this is
106  * an &amdgpu_bo.
107  *
108  * Returns:
109  * true if the object belongs to &amdgpu_bo, false if not.
110  */
111 bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
112 {
113 	if (bo->destroy == &amdgpu_bo_destroy)
114 		return true;
115 	return false;
116 }
117 
118 /**
119  * amdgpu_bo_placement_from_domain - set buffer's placement
120  * @abo: &amdgpu_bo buffer object whose placement is to be set
121  * @domain: requested domain
122  *
123  * Sets buffer's placement according to requested domain and the buffer's
124  * flags.
125  */
126 void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
127 {
128 	struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
129 	struct ttm_placement *placement = &abo->placement;
130 	struct ttm_place *places = abo->placements;
131 	u64 flags = abo->flags;
132 	u32 c = 0;
133 
134 	if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
135 		unsigned visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
136 
137 		places[c].fpfn = 0;
138 		places[c].lpfn = 0;
139 		places[c].flags = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
140 			TTM_PL_FLAG_VRAM;
141 
142 		if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
143 			places[c].lpfn = visible_pfn;
144 		else
145 			places[c].flags |= TTM_PL_FLAG_TOPDOWN;
146 
147 		if (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
148 			places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
149 		c++;
150 	}
151 
152 	if (domain & AMDGPU_GEM_DOMAIN_GTT) {
153 		places[c].fpfn = 0;
154 		places[c].lpfn = 0;
155 		places[c].flags = TTM_PL_FLAG_TT;
156 		if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
157 			places[c].flags |= TTM_PL_FLAG_WC |
158 				TTM_PL_FLAG_UNCACHED;
159 		else
160 			places[c].flags |= TTM_PL_FLAG_CACHED;
161 		c++;
162 	}
163 
164 	if (domain & AMDGPU_GEM_DOMAIN_CPU) {
165 		places[c].fpfn = 0;
166 		places[c].lpfn = 0;
167 		places[c].flags = TTM_PL_FLAG_SYSTEM;
168 		if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
169 			places[c].flags |= TTM_PL_FLAG_WC |
170 				TTM_PL_FLAG_UNCACHED;
171 		else
172 			places[c].flags |= TTM_PL_FLAG_CACHED;
173 		c++;
174 	}
175 
176 	if (domain & AMDGPU_GEM_DOMAIN_GDS) {
177 		places[c].fpfn = 0;
178 		places[c].lpfn = 0;
179 		places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GDS;
180 		c++;
181 	}
182 
183 	if (domain & AMDGPU_GEM_DOMAIN_GWS) {
184 		places[c].fpfn = 0;
185 		places[c].lpfn = 0;
186 		places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GWS;
187 		c++;
188 	}
189 
190 	if (domain & AMDGPU_GEM_DOMAIN_OA) {
191 		places[c].fpfn = 0;
192 		places[c].lpfn = 0;
193 		places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_OA;
194 		c++;
195 	}
196 
197 	if (!c) {
198 		places[c].fpfn = 0;
199 		places[c].lpfn = 0;
200 		places[c].flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
201 		c++;
202 	}
203 
204 	BUG_ON(c >= AMDGPU_BO_MAX_PLACEMENTS);
205 
206 	placement->num_placement = c;
207 	placement->placement = places;
208 
209 	placement->num_busy_placement = c;
210 	placement->busy_placement = places;
211 }
212 
213 /**
214  * amdgpu_bo_create_reserved - create reserved BO for kernel use
215  *
216  * @adev: amdgpu device object
217  * @size: size for the new BO
218  * @align: alignment for the new BO
219  * @domain: where to place it
220  * @bo_ptr: used to initialize BOs in structures
221  * @gpu_addr: GPU addr of the pinned BO
222  * @cpu_addr: optional CPU address mapping
223  *
224  * Allocates and pins a BO for kernel internal use, and returns it still
225  * reserved.
226  *
227  * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
228  *
229  * Returns:
230  * 0 on success, negative error code otherwise.
231  */
232 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
233 			      unsigned long size, int align,
234 			      u32 domain, struct amdgpu_bo **bo_ptr,
235 			      u64 *gpu_addr, void **cpu_addr)
236 {
237 	struct amdgpu_bo_param bp;
238 	bool free = false;
239 	int r;
240 
241 	if (!size) {
242 		amdgpu_bo_unref(bo_ptr);
243 		return 0;
244 	}
245 
246 	memset(&bp, 0, sizeof(bp));
247 	bp.size = size;
248 	bp.byte_align = align;
249 	bp.domain = domain;
250 	bp.flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
251 		AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
252 	bp.type = ttm_bo_type_kernel;
253 	bp.resv = NULL;
254 
255 	if (!*bo_ptr) {
256 		r = amdgpu_bo_create(adev, &bp, bo_ptr);
257 		if (r) {
258 			dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
259 				r);
260 			return r;
261 		}
262 		free = true;
263 	}
264 
265 	r = amdgpu_bo_reserve(*bo_ptr, false);
266 	if (r) {
267 		dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
268 		goto error_free;
269 	}
270 
271 	r = amdgpu_bo_pin(*bo_ptr, domain);
272 	if (r) {
273 		dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
274 		goto error_unreserve;
275 	}
276 
277 	r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
278 	if (r) {
279 		dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
280 		goto error_unpin;
281 	}
282 
283 	if (gpu_addr)
284 		*gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
285 
286 	if (cpu_addr) {
287 		r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
288 		if (r) {
289 			dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
290 			goto error_unpin;
291 		}
292 	}
293 
294 	return 0;
295 
296 error_unpin:
297 	amdgpu_bo_unpin(*bo_ptr);
298 error_unreserve:
299 	amdgpu_bo_unreserve(*bo_ptr);
300 
301 error_free:
302 	if (free)
303 		amdgpu_bo_unref(bo_ptr);
304 
305 	return r;
306 }
307 
308 /**
309  * amdgpu_bo_create_kernel - create BO for kernel use
310  *
311  * @adev: amdgpu device object
312  * @size: size for the new BO
313  * @align: alignment for the new BO
314  * @domain: where to place it
315  * @bo_ptr:  used to initialize BOs in structures
316  * @gpu_addr: GPU addr of the pinned BO
317  * @cpu_addr: optional CPU address mapping
318  *
319  * Allocates and pins a BO for kernel internal use.
320  *
321  * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
322  *
323  * Returns:
324  * 0 on success, negative error code otherwise.
325  */
326 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
327 			    unsigned long size, int align,
328 			    u32 domain, struct amdgpu_bo **bo_ptr,
329 			    u64 *gpu_addr, void **cpu_addr)
330 {
331 	int r;
332 
333 	r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
334 				      gpu_addr, cpu_addr);
335 
336 	if (r)
337 		return r;
338 
339 	if (*bo_ptr)
340 		amdgpu_bo_unreserve(*bo_ptr);
341 
342 	return 0;
343 }
344 
345 /**
346  * amdgpu_bo_free_kernel - free BO for kernel use
347  *
348  * @bo: amdgpu BO to free
349  * @gpu_addr: pointer to where the BO's GPU memory space address was stored
350  * @cpu_addr: pointer to where the BO's CPU memory space address was stored
351  *
352  * unmaps and unpin a BO for kernel internal use.
353  */
354 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
355 			   void **cpu_addr)
356 {
357 	if (*bo == NULL)
358 		return;
359 
360 	if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
361 		if (cpu_addr)
362 			amdgpu_bo_kunmap(*bo);
363 
364 		amdgpu_bo_unpin(*bo);
365 		amdgpu_bo_unreserve(*bo);
366 	}
367 	amdgpu_bo_unref(bo);
368 
369 	if (gpu_addr)
370 		*gpu_addr = 0;
371 
372 	if (cpu_addr)
373 		*cpu_addr = NULL;
374 }
375 
376 /* Validate bo size is bit bigger then the request domain */
377 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
378 					  unsigned long size, u32 domain)
379 {
380 	struct ttm_mem_type_manager *man = NULL;
381 
382 	/*
383 	 * If GTT is part of requested domains the check must succeed to
384 	 * allow fall back to GTT
385 	 */
386 	if (domain & AMDGPU_GEM_DOMAIN_GTT) {
387 		man = &adev->mman.bdev.man[TTM_PL_TT];
388 
389 		if (size < (man->size << PAGE_SHIFT))
390 			return true;
391 		else
392 			goto fail;
393 	}
394 
395 	if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
396 		man = &adev->mman.bdev.man[TTM_PL_VRAM];
397 
398 		if (size < (man->size << PAGE_SHIFT))
399 			return true;
400 		else
401 			goto fail;
402 	}
403 
404 
405 	/* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */
406 	return true;
407 
408 fail:
409 	DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size,
410 		  man->size << PAGE_SHIFT);
411 	return false;
412 }
413 
414 static int amdgpu_bo_do_create(struct amdgpu_device *adev,
415 			       struct amdgpu_bo_param *bp,
416 			       struct amdgpu_bo **bo_ptr)
417 {
418 	struct ttm_operation_ctx ctx = {
419 		.interruptible = (bp->type != ttm_bo_type_kernel),
420 		.no_wait_gpu = false,
421 		.resv = bp->resv,
422 		.flags = TTM_OPT_FLAG_ALLOW_RES_EVICT
423 	};
424 	struct amdgpu_bo *bo;
425 	unsigned long page_align, size = bp->size;
426 	size_t acc_size;
427 	int r;
428 
429 	page_align = roundup(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
430 	if (bp->domain & (AMDGPU_GEM_DOMAIN_GDS | AMDGPU_GEM_DOMAIN_GWS |
431 			  AMDGPU_GEM_DOMAIN_OA))
432 		size <<= PAGE_SHIFT;
433 	else
434 		size = ALIGN(size, PAGE_SIZE);
435 
436 	if (!amdgpu_bo_validate_size(adev, size, bp->domain))
437 		return -ENOMEM;
438 
439 	*bo_ptr = NULL;
440 
441 	acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
442 				       sizeof(struct amdgpu_bo));
443 
444 	bo = kzalloc(sizeof(struct amdgpu_bo), GFP_KERNEL);
445 	if (bo == NULL)
446 		return -ENOMEM;
447 	drm_gem_private_object_init(adev->ddev, &bo->gem_base, size);
448 	INIT_LIST_HEAD(&bo->shadow_list);
449 	bo->vm_bo = NULL;
450 	bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
451 		bp->domain;
452 	bo->allowed_domains = bo->preferred_domains;
453 	if (bp->type != ttm_bo_type_kernel &&
454 	    bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
455 		bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
456 
457 	bo->flags = bp->flags;
458 
459 #ifdef CONFIG_X86_32
460 	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
461 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
462 	 */
463 	bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
464 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
465 	/* Don't try to enable write-combining when it can't work, or things
466 	 * may be slow
467 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
468 	 */
469 
470 #ifndef CONFIG_COMPILE_TEST
471 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
472 	 thanks to write-combining
473 #endif
474 
475 	if (bo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
476 		DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
477 			      "better performance thanks to write-combining\n");
478 	bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
479 #else
480 	/* For architectures that don't support WC memory,
481 	 * mask out the WC flag from the BO
482 	 */
483 	if (!drm_arch_can_wc_memory())
484 		bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
485 #endif
486 
487 	bo->tbo.bdev = &adev->mman.bdev;
488 	amdgpu_bo_placement_from_domain(bo, bp->domain);
489 	if (bp->type == ttm_bo_type_kernel)
490 		bo->tbo.priority = 1;
491 
492 	r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, bp->type,
493 				 &bo->placement, page_align, &ctx, acc_size,
494 				 NULL, bp->resv, &amdgpu_bo_destroy);
495 	if (unlikely(r != 0))
496 		return r;
497 
498 	if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
499 	    bo->tbo.mem.mem_type == TTM_PL_VRAM &&
500 	    bo->tbo.mem.start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
501 		amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
502 					     ctx.bytes_moved);
503 	else
504 		amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
505 
506 	if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
507 	    bo->tbo.mem.placement & TTM_PL_FLAG_VRAM) {
508 		struct dma_fence *fence;
509 
510 		r = amdgpu_fill_buffer(bo, 0, bo->tbo.resv, &fence);
511 		if (unlikely(r))
512 			goto fail_unreserve;
513 
514 		amdgpu_bo_fence(bo, fence, false);
515 		dma_fence_put(bo->tbo.moving);
516 		bo->tbo.moving = dma_fence_get(fence);
517 		dma_fence_put(fence);
518 	}
519 	if (!bp->resv)
520 		amdgpu_bo_unreserve(bo);
521 	*bo_ptr = bo;
522 
523 	trace_amdgpu_bo_create(bo);
524 
525 	/* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
526 	if (bp->type == ttm_bo_type_device)
527 		bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
528 
529 	return 0;
530 
531 fail_unreserve:
532 	if (!bp->resv)
533 		ww_mutex_unlock(&bo->tbo.resv->lock);
534 	amdgpu_bo_unref(&bo);
535 	return r;
536 }
537 
538 static int amdgpu_bo_create_shadow(struct amdgpu_device *adev,
539 				   unsigned long size,
540 				   struct amdgpu_bo *bo)
541 {
542 	struct amdgpu_bo_param bp;
543 	int r;
544 
545 	if (bo->shadow)
546 		return 0;
547 
548 	memset(&bp, 0, sizeof(bp));
549 	bp.size = size;
550 	bp.domain = AMDGPU_GEM_DOMAIN_GTT;
551 	bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC |
552 		AMDGPU_GEM_CREATE_SHADOW;
553 	bp.type = ttm_bo_type_kernel;
554 	bp.resv = bo->tbo.resv;
555 
556 	r = amdgpu_bo_do_create(adev, &bp, &bo->shadow);
557 	if (!r) {
558 		bo->shadow->parent = amdgpu_bo_ref(bo);
559 		mutex_lock(&adev->shadow_list_lock);
560 		list_add_tail(&bo->shadow->shadow_list, &adev->shadow_list);
561 		mutex_unlock(&adev->shadow_list_lock);
562 	}
563 
564 	return r;
565 }
566 
567 /**
568  * amdgpu_bo_create - create an &amdgpu_bo buffer object
569  * @adev: amdgpu device object
570  * @bp: parameters to be used for the buffer object
571  * @bo_ptr: pointer to the buffer object pointer
572  *
573  * Creates an &amdgpu_bo buffer object; and if requested, also creates a
574  * shadow object.
575  * Shadow object is used to backup the original buffer object, and is always
576  * in GTT.
577  *
578  * Returns:
579  * 0 for success or a negative error code on failure.
580  */
581 int amdgpu_bo_create(struct amdgpu_device *adev,
582 		     struct amdgpu_bo_param *bp,
583 		     struct amdgpu_bo **bo_ptr)
584 {
585 	u64 flags = bp->flags;
586 	int r;
587 
588 	bp->flags = bp->flags & ~AMDGPU_GEM_CREATE_SHADOW;
589 	r = amdgpu_bo_do_create(adev, bp, bo_ptr);
590 	if (r)
591 		return r;
592 
593 	if ((flags & AMDGPU_GEM_CREATE_SHADOW) && !(adev->flags & AMD_IS_APU)) {
594 		if (!bp->resv)
595 			WARN_ON(reservation_object_lock((*bo_ptr)->tbo.resv,
596 							NULL));
597 
598 		r = amdgpu_bo_create_shadow(adev, bp->size, *bo_ptr);
599 
600 		if (!bp->resv)
601 			reservation_object_unlock((*bo_ptr)->tbo.resv);
602 
603 		if (r)
604 			amdgpu_bo_unref(bo_ptr);
605 	}
606 
607 	return r;
608 }
609 
610 /**
611  * amdgpu_bo_validate - validate an &amdgpu_bo buffer object
612  * @bo: pointer to the buffer object
613  *
614  * Sets placement according to domain; and changes placement and caching
615  * policy of the buffer object according to the placement.
616  * This is used for validating shadow bos.  It calls ttm_bo_validate() to
617  * make sure the buffer is resident where it needs to be.
618  *
619  * Returns:
620  * 0 for success or a negative error code on failure.
621  */
622 int amdgpu_bo_validate(struct amdgpu_bo *bo)
623 {
624 	struct ttm_operation_ctx ctx = { false, false };
625 	uint32_t domain;
626 	int r;
627 
628 	if (bo->pin_count)
629 		return 0;
630 
631 	domain = bo->preferred_domains;
632 
633 retry:
634 	amdgpu_bo_placement_from_domain(bo, domain);
635 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
636 	if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
637 		domain = bo->allowed_domains;
638 		goto retry;
639 	}
640 
641 	return r;
642 }
643 
644 /**
645  * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow
646  *
647  * @shadow: &amdgpu_bo shadow to be restored
648  * @fence: dma_fence associated with the operation
649  *
650  * Copies a buffer object's shadow content back to the object.
651  * This is used for recovering a buffer from its shadow in case of a gpu
652  * reset where vram context may be lost.
653  *
654  * Returns:
655  * 0 for success or a negative error code on failure.
656  */
657 int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence)
658 
659 {
660 	struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev);
661 	struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
662 	uint64_t shadow_addr, parent_addr;
663 
664 	shadow_addr = amdgpu_bo_gpu_offset(shadow);
665 	parent_addr = amdgpu_bo_gpu_offset(shadow->parent);
666 
667 	return amdgpu_copy_buffer(ring, shadow_addr, parent_addr,
668 				  amdgpu_bo_size(shadow), NULL, fence,
669 				  true, false);
670 }
671 
672 /**
673  * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
674  * @bo: &amdgpu_bo buffer object to be mapped
675  * @ptr: kernel virtual address to be returned
676  *
677  * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
678  * amdgpu_bo_kptr() to get the kernel virtual address.
679  *
680  * Returns:
681  * 0 for success or a negative error code on failure.
682  */
683 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
684 {
685 	void *kptr;
686 	long r;
687 
688 	if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
689 		return -EPERM;
690 
691 	kptr = amdgpu_bo_kptr(bo);
692 	if (kptr) {
693 		if (ptr)
694 			*ptr = kptr;
695 		return 0;
696 	}
697 
698 	r = reservation_object_wait_timeout_rcu(bo->tbo.resv, false, false,
699 						MAX_SCHEDULE_TIMEOUT);
700 	if (r < 0)
701 		return r;
702 
703 	r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
704 	if (r)
705 		return r;
706 
707 	if (ptr)
708 		*ptr = amdgpu_bo_kptr(bo);
709 
710 	return 0;
711 }
712 
713 /**
714  * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
715  * @bo: &amdgpu_bo buffer object
716  *
717  * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
718  *
719  * Returns:
720  * the virtual address of a buffer object area.
721  */
722 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
723 {
724 	bool is_iomem;
725 
726 	return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
727 }
728 
729 /**
730  * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
731  * @bo: &amdgpu_bo buffer object to be unmapped
732  *
733  * Unmaps a kernel map set up by amdgpu_bo_kmap().
734  */
735 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
736 {
737 	if (bo->kmap.bo)
738 		ttm_bo_kunmap(&bo->kmap);
739 }
740 
741 /**
742  * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
743  * @bo: &amdgpu_bo buffer object
744  *
745  * References the contained &ttm_buffer_object.
746  *
747  * Returns:
748  * a refcounted pointer to the &amdgpu_bo buffer object.
749  */
750 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
751 {
752 	if (bo == NULL)
753 		return NULL;
754 
755 	ttm_bo_get(&bo->tbo);
756 	return bo;
757 }
758 
759 /**
760  * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
761  * @bo: &amdgpu_bo buffer object
762  *
763  * Unreferences the contained &ttm_buffer_object and clear the pointer
764  */
765 void amdgpu_bo_unref(struct amdgpu_bo **bo)
766 {
767 	struct ttm_buffer_object *tbo;
768 
769 	if ((*bo) == NULL)
770 		return;
771 
772 	tbo = &((*bo)->tbo);
773 	ttm_bo_put(tbo);
774 	*bo = NULL;
775 }
776 
777 /**
778  * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
779  * @bo: &amdgpu_bo buffer object to be pinned
780  * @domain: domain to be pinned to
781  * @min_offset: the start of requested address range
782  * @max_offset: the end of requested address range
783  *
784  * Pins the buffer object according to requested domain and address range. If
785  * the memory is unbound gart memory, binds the pages into gart table. Adjusts
786  * pin_count and pin_size accordingly.
787  *
788  * Pinning means to lock pages in memory along with keeping them at a fixed
789  * offset. It is required when a buffer can not be moved, for example, when
790  * a display buffer is being scanned out.
791  *
792  * Compared with amdgpu_bo_pin(), this function gives more flexibility on
793  * where to pin a buffer if there are specific restrictions on where a buffer
794  * must be located.
795  *
796  * Returns:
797  * 0 for success or a negative error code on failure.
798  */
799 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
800 			     u64 min_offset, u64 max_offset)
801 {
802 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
803 	struct ttm_operation_ctx ctx = { false, false };
804 	int r, i;
805 
806 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
807 		return -EPERM;
808 
809 	if (WARN_ON_ONCE(min_offset > max_offset))
810 		return -EINVAL;
811 
812 	/* A shared bo cannot be migrated to VRAM */
813 	if (bo->prime_shared_count) {
814 		if (domain & AMDGPU_GEM_DOMAIN_GTT)
815 			domain = AMDGPU_GEM_DOMAIN_GTT;
816 		else
817 			return -EINVAL;
818 	}
819 
820 	/* This assumes only APU display buffers are pinned with (VRAM|GTT).
821 	 * See function amdgpu_display_supported_domains()
822 	 */
823 	domain = amdgpu_bo_get_preferred_pin_domain(adev, domain);
824 
825 	if (bo->pin_count) {
826 		uint32_t mem_type = bo->tbo.mem.mem_type;
827 
828 		if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
829 			return -EINVAL;
830 
831 		bo->pin_count++;
832 
833 		if (max_offset != 0) {
834 			u64 domain_start = bo->tbo.bdev->man[mem_type].gpu_offset;
835 			WARN_ON_ONCE(max_offset <
836 				     (amdgpu_bo_gpu_offset(bo) - domain_start));
837 		}
838 
839 		return 0;
840 	}
841 
842 	bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
843 	/* force to pin into visible video ram */
844 	if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
845 		bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
846 	amdgpu_bo_placement_from_domain(bo, domain);
847 	for (i = 0; i < bo->placement.num_placement; i++) {
848 		unsigned fpfn, lpfn;
849 
850 		fpfn = min_offset >> PAGE_SHIFT;
851 		lpfn = max_offset >> PAGE_SHIFT;
852 
853 		if (fpfn > bo->placements[i].fpfn)
854 			bo->placements[i].fpfn = fpfn;
855 		if (!bo->placements[i].lpfn ||
856 		    (lpfn && lpfn < bo->placements[i].lpfn))
857 			bo->placements[i].lpfn = lpfn;
858 		bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
859 	}
860 
861 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
862 	if (unlikely(r)) {
863 		dev_err(adev->dev, "%p pin failed\n", bo);
864 		goto error;
865 	}
866 
867 	bo->pin_count = 1;
868 
869 	domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
870 	if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
871 		atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
872 		atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
873 			     &adev->visible_pin_size);
874 	} else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
875 		atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
876 	}
877 
878 error:
879 	return r;
880 }
881 
882 /**
883  * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
884  * @bo: &amdgpu_bo buffer object to be pinned
885  * @domain: domain to be pinned to
886  *
887  * A simple wrapper to amdgpu_bo_pin_restricted().
888  * Provides a simpler API for buffers that do not have any strict restrictions
889  * on where a buffer must be located.
890  *
891  * Returns:
892  * 0 for success or a negative error code on failure.
893  */
894 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
895 {
896 	return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
897 }
898 
899 /**
900  * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
901  * @bo: &amdgpu_bo buffer object to be unpinned
902  *
903  * Decreases the pin_count, and clears the flags if pin_count reaches 0.
904  * Changes placement and pin size accordingly.
905  *
906  * Returns:
907  * 0 for success or a negative error code on failure.
908  */
909 int amdgpu_bo_unpin(struct amdgpu_bo *bo)
910 {
911 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
912 	struct ttm_operation_ctx ctx = { false, false };
913 	int r, i;
914 
915 	if (WARN_ON_ONCE(!bo->pin_count)) {
916 		dev_warn(adev->dev, "%p unpin not necessary\n", bo);
917 		return 0;
918 	}
919 	bo->pin_count--;
920 	if (bo->pin_count)
921 		return 0;
922 
923 	amdgpu_bo_subtract_pin_size(bo);
924 
925 	for (i = 0; i < bo->placement.num_placement; i++) {
926 		bo->placements[i].lpfn = 0;
927 		bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
928 	}
929 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
930 	if (unlikely(r))
931 		dev_err(adev->dev, "%p validate failed for unpin\n", bo);
932 
933 	return r;
934 }
935 
936 /**
937  * amdgpu_bo_evict_vram - evict VRAM buffers
938  * @adev: amdgpu device object
939  *
940  * Evicts all VRAM buffers on the lru list of the memory type.
941  * Mainly used for evicting vram at suspend time.
942  *
943  * Returns:
944  * 0 for success or a negative error code on failure.
945  */
946 int amdgpu_bo_evict_vram(struct amdgpu_device *adev)
947 {
948 	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
949 #ifndef CONFIG_HIBERNATION
950 	if (adev->flags & AMD_IS_APU) {
951 		/* Useless to evict on IGP chips */
952 		return 0;
953 	}
954 #endif
955 	return ttm_bo_evict_mm(&adev->mman.bdev, TTM_PL_VRAM);
956 }
957 
958 static const char *amdgpu_vram_names[] = {
959 	"UNKNOWN",
960 	"GDDR1",
961 	"DDR2",
962 	"GDDR3",
963 	"GDDR4",
964 	"GDDR5",
965 	"HBM",
966 	"DDR3",
967 	"DDR4",
968 };
969 
970 /**
971  * amdgpu_bo_init - initialize memory manager
972  * @adev: amdgpu device object
973  *
974  * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
975  *
976  * Returns:
977  * 0 for success or a negative error code on failure.
978  */
979 int amdgpu_bo_init(struct amdgpu_device *adev)
980 {
981 	/* reserve PAT memory space to WC for VRAM */
982 	arch_io_reserve_memtype_wc(adev->gmc.aper_base,
983 				   adev->gmc.aper_size);
984 
985 	/* Add an MTRR for the VRAM */
986 	adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
987 					      adev->gmc.aper_size);
988 	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
989 		 adev->gmc.mc_vram_size >> 20,
990 		 (unsigned long long)adev->gmc.aper_size >> 20);
991 	DRM_INFO("RAM width %dbits %s\n",
992 		 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
993 	return amdgpu_ttm_init(adev);
994 }
995 
996 /**
997  * amdgpu_bo_late_init - late init
998  * @adev: amdgpu device object
999  *
1000  * Calls amdgpu_ttm_late_init() to free resources used earlier during
1001  * initialization.
1002  *
1003  * Returns:
1004  * 0 for success or a negative error code on failure.
1005  */
1006 int amdgpu_bo_late_init(struct amdgpu_device *adev)
1007 {
1008 	amdgpu_ttm_late_init(adev);
1009 
1010 	return 0;
1011 }
1012 
1013 /**
1014  * amdgpu_bo_fini - tear down memory manager
1015  * @adev: amdgpu device object
1016  *
1017  * Reverses amdgpu_bo_init() to tear down memory manager.
1018  */
1019 void amdgpu_bo_fini(struct amdgpu_device *adev)
1020 {
1021 	amdgpu_ttm_fini(adev);
1022 	arch_phys_wc_del(adev->gmc.vram_mtrr);
1023 	arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1024 }
1025 
1026 /**
1027  * amdgpu_bo_fbdev_mmap - mmap fbdev memory
1028  * @bo: &amdgpu_bo buffer object
1029  * @vma: vma as input from the fbdev mmap method
1030  *
1031  * Calls ttm_fbdev_mmap() to mmap fbdev memory if it is backed by a bo.
1032  *
1033  * Returns:
1034  * 0 for success or a negative error code on failure.
1035  */
1036 int amdgpu_bo_fbdev_mmap(struct amdgpu_bo *bo,
1037 			     struct vm_area_struct *vma)
1038 {
1039 	return ttm_fbdev_mmap(vma, &bo->tbo);
1040 }
1041 
1042 /**
1043  * amdgpu_bo_set_tiling_flags - set tiling flags
1044  * @bo: &amdgpu_bo buffer object
1045  * @tiling_flags: new flags
1046  *
1047  * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1048  * kernel driver to set the tiling flags on a buffer.
1049  *
1050  * Returns:
1051  * 0 for success or a negative error code on failure.
1052  */
1053 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1054 {
1055 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1056 
1057 	if (adev->family <= AMDGPU_FAMILY_CZ &&
1058 	    AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1059 		return -EINVAL;
1060 
1061 	bo->tiling_flags = tiling_flags;
1062 	return 0;
1063 }
1064 
1065 /**
1066  * amdgpu_bo_get_tiling_flags - get tiling flags
1067  * @bo: &amdgpu_bo buffer object
1068  * @tiling_flags: returned flags
1069  *
1070  * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1071  * set the tiling flags on a buffer.
1072  */
1073 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1074 {
1075 	lockdep_assert_held(&bo->tbo.resv->lock.base);
1076 
1077 	if (tiling_flags)
1078 		*tiling_flags = bo->tiling_flags;
1079 }
1080 
1081 /**
1082  * amdgpu_bo_set_metadata - set metadata
1083  * @bo: &amdgpu_bo buffer object
1084  * @metadata: new metadata
1085  * @metadata_size: size of the new metadata
1086  * @flags: flags of the new metadata
1087  *
1088  * Sets buffer object's metadata, its size and flags.
1089  * Used via GEM ioctl.
1090  *
1091  * Returns:
1092  * 0 for success or a negative error code on failure.
1093  */
1094 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
1095 			    uint32_t metadata_size, uint64_t flags)
1096 {
1097 	void *buffer;
1098 
1099 	if (!metadata_size) {
1100 		if (bo->metadata_size) {
1101 			kfree(bo->metadata);
1102 			bo->metadata = NULL;
1103 			bo->metadata_size = 0;
1104 		}
1105 		return 0;
1106 	}
1107 
1108 	if (metadata == NULL)
1109 		return -EINVAL;
1110 
1111 	buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1112 	if (buffer == NULL)
1113 		return -ENOMEM;
1114 
1115 	kfree(bo->metadata);
1116 	bo->metadata_flags = flags;
1117 	bo->metadata = buffer;
1118 	bo->metadata_size = metadata_size;
1119 
1120 	return 0;
1121 }
1122 
1123 /**
1124  * amdgpu_bo_get_metadata - get metadata
1125  * @bo: &amdgpu_bo buffer object
1126  * @buffer: returned metadata
1127  * @buffer_size: size of the buffer
1128  * @metadata_size: size of the returned metadata
1129  * @flags: flags of the returned metadata
1130  *
1131  * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1132  * less than metadata_size.
1133  * Used via GEM ioctl.
1134  *
1135  * Returns:
1136  * 0 for success or a negative error code on failure.
1137  */
1138 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1139 			   size_t buffer_size, uint32_t *metadata_size,
1140 			   uint64_t *flags)
1141 {
1142 	if (!buffer && !metadata_size)
1143 		return -EINVAL;
1144 
1145 	if (buffer) {
1146 		if (buffer_size < bo->metadata_size)
1147 			return -EINVAL;
1148 
1149 		if (bo->metadata_size)
1150 			memcpy(buffer, bo->metadata, bo->metadata_size);
1151 	}
1152 
1153 	if (metadata_size)
1154 		*metadata_size = bo->metadata_size;
1155 	if (flags)
1156 		*flags = bo->metadata_flags;
1157 
1158 	return 0;
1159 }
1160 
1161 /**
1162  * amdgpu_bo_move_notify - notification about a memory move
1163  * @bo: pointer to a buffer object
1164  * @evict: if this move is evicting the buffer from the graphics address space
1165  * @new_mem: new information of the bufer object
1166  *
1167  * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1168  * bookkeeping.
1169  * TTM driver callback which is called when ttm moves a buffer.
1170  */
1171 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1172 			   bool evict,
1173 			   struct ttm_mem_reg *new_mem)
1174 {
1175 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1176 	struct amdgpu_bo *abo;
1177 	struct ttm_mem_reg *old_mem = &bo->mem;
1178 
1179 	if (!amdgpu_bo_is_amdgpu_bo(bo))
1180 		return;
1181 
1182 	abo = ttm_to_amdgpu_bo(bo);
1183 	amdgpu_vm_bo_invalidate(adev, abo, evict);
1184 
1185 	amdgpu_bo_kunmap(abo);
1186 
1187 	/* remember the eviction */
1188 	if (evict)
1189 		atomic64_inc(&adev->num_evictions);
1190 
1191 	/* update statistics */
1192 	if (!new_mem)
1193 		return;
1194 
1195 	/* move_notify is called before move happens */
1196 	trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
1197 }
1198 
1199 /**
1200  * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1201  * @bo: pointer to a buffer object
1202  *
1203  * Notifies the driver we are taking a fault on this BO and have reserved it,
1204  * also performs bookkeeping.
1205  * TTM driver callback for dealing with vm faults.
1206  *
1207  * Returns:
1208  * 0 for success or a negative error code on failure.
1209  */
1210 int amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1211 {
1212 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1213 	struct ttm_operation_ctx ctx = { false, false };
1214 	struct amdgpu_bo *abo;
1215 	unsigned long offset, size;
1216 	int r;
1217 
1218 	if (!amdgpu_bo_is_amdgpu_bo(bo))
1219 		return 0;
1220 
1221 	abo = ttm_to_amdgpu_bo(bo);
1222 
1223 	/* Remember that this BO was accessed by the CPU */
1224 	abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1225 
1226 	if (bo->mem.mem_type != TTM_PL_VRAM)
1227 		return 0;
1228 
1229 	size = bo->mem.num_pages << PAGE_SHIFT;
1230 	offset = bo->mem.start << PAGE_SHIFT;
1231 	if ((offset + size) <= adev->gmc.visible_vram_size)
1232 		return 0;
1233 
1234 	/* Can't move a pinned BO to visible VRAM */
1235 	if (abo->pin_count > 0)
1236 		return -EINVAL;
1237 
1238 	/* hurrah the memory is not visible ! */
1239 	atomic64_inc(&adev->num_vram_cpu_page_faults);
1240 	amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1241 					AMDGPU_GEM_DOMAIN_GTT);
1242 
1243 	/* Avoid costly evictions; only set GTT as a busy placement */
1244 	abo->placement.num_busy_placement = 1;
1245 	abo->placement.busy_placement = &abo->placements[1];
1246 
1247 	r = ttm_bo_validate(bo, &abo->placement, &ctx);
1248 	if (unlikely(r != 0))
1249 		return r;
1250 
1251 	offset = bo->mem.start << PAGE_SHIFT;
1252 	/* this should never happen */
1253 	if (bo->mem.mem_type == TTM_PL_VRAM &&
1254 	    (offset + size) > adev->gmc.visible_vram_size)
1255 		return -EINVAL;
1256 
1257 	return 0;
1258 }
1259 
1260 /**
1261  * amdgpu_bo_fence - add fence to buffer object
1262  *
1263  * @bo: buffer object in question
1264  * @fence: fence to add
1265  * @shared: true if fence should be added shared
1266  *
1267  */
1268 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1269 		     bool shared)
1270 {
1271 	struct reservation_object *resv = bo->tbo.resv;
1272 
1273 	if (shared)
1274 		reservation_object_add_shared_fence(resv, fence);
1275 	else
1276 		reservation_object_add_excl_fence(resv, fence);
1277 }
1278 
1279 /**
1280  * amdgpu_bo_gpu_offset - return GPU offset of bo
1281  * @bo:	amdgpu object for which we query the offset
1282  *
1283  * Note: object should either be pinned or reserved when calling this
1284  * function, it might be useful to add check for this for debugging.
1285  *
1286  * Returns:
1287  * current GPU offset of the object.
1288  */
1289 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1290 {
1291 	WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_SYSTEM);
1292 	WARN_ON_ONCE(!ww_mutex_is_locked(&bo->tbo.resv->lock) &&
1293 		     !bo->pin_count && bo->tbo.type != ttm_bo_type_kernel);
1294 	WARN_ON_ONCE(bo->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET);
1295 	WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_VRAM &&
1296 		     !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1297 
1298 	return amdgpu_gmc_sign_extend(bo->tbo.offset);
1299 }
1300 
1301 /**
1302  * amdgpu_bo_get_preferred_pin_domain - get preferred domain for scanout
1303  * @adev: amdgpu device object
1304  * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1305  *
1306  * Returns:
1307  * Which of the allowed domains is preferred for pinning the BO for scanout.
1308  */
1309 uint32_t amdgpu_bo_get_preferred_pin_domain(struct amdgpu_device *adev,
1310 					    uint32_t domain)
1311 {
1312 	if (domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) {
1313 		domain = AMDGPU_GEM_DOMAIN_VRAM;
1314 		if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1315 			domain = AMDGPU_GEM_DOMAIN_GTT;
1316 	}
1317 	return domain;
1318 }
1319