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 	/* in case amdgpu_device_recover_vram got NULL of 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 	amdgpu_bo_unref(&bo->parent);
98 
99 	kfree(bo->metadata);
100 	kfree(bo);
101 }
102 
103 /**
104  * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
105  * @bo: buffer object to be checked
106  *
107  * Uses destroy function associated with the object to determine if this is
108  * an &amdgpu_bo.
109  *
110  * Returns:
111  * true if the object belongs to &amdgpu_bo, false if not.
112  */
113 bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
114 {
115 	if (bo->destroy == &amdgpu_bo_destroy)
116 		return true;
117 	return false;
118 }
119 
120 /**
121  * amdgpu_bo_placement_from_domain - set buffer's placement
122  * @abo: &amdgpu_bo buffer object whose placement is to be set
123  * @domain: requested domain
124  *
125  * Sets buffer's placement according to requested domain and the buffer's
126  * flags.
127  */
128 void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
129 {
130 	struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
131 	struct ttm_placement *placement = &abo->placement;
132 	struct ttm_place *places = abo->placements;
133 	u64 flags = abo->flags;
134 	u32 c = 0;
135 
136 	if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
137 		unsigned visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
138 
139 		places[c].fpfn = 0;
140 		places[c].lpfn = 0;
141 		places[c].flags = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
142 			TTM_PL_FLAG_VRAM;
143 
144 		if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
145 			places[c].lpfn = visible_pfn;
146 		else
147 			places[c].flags |= TTM_PL_FLAG_TOPDOWN;
148 
149 		if (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
150 			places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
151 		c++;
152 	}
153 
154 	if (domain & AMDGPU_GEM_DOMAIN_GTT) {
155 		places[c].fpfn = 0;
156 		places[c].lpfn = 0;
157 		places[c].flags = TTM_PL_FLAG_TT;
158 		if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
159 			places[c].flags |= TTM_PL_FLAG_WC |
160 				TTM_PL_FLAG_UNCACHED;
161 		else
162 			places[c].flags |= TTM_PL_FLAG_CACHED;
163 		c++;
164 	}
165 
166 	if (domain & AMDGPU_GEM_DOMAIN_CPU) {
167 		places[c].fpfn = 0;
168 		places[c].lpfn = 0;
169 		places[c].flags = TTM_PL_FLAG_SYSTEM;
170 		if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
171 			places[c].flags |= TTM_PL_FLAG_WC |
172 				TTM_PL_FLAG_UNCACHED;
173 		else
174 			places[c].flags |= TTM_PL_FLAG_CACHED;
175 		c++;
176 	}
177 
178 	if (domain & AMDGPU_GEM_DOMAIN_GDS) {
179 		places[c].fpfn = 0;
180 		places[c].lpfn = 0;
181 		places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GDS;
182 		c++;
183 	}
184 
185 	if (domain & AMDGPU_GEM_DOMAIN_GWS) {
186 		places[c].fpfn = 0;
187 		places[c].lpfn = 0;
188 		places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GWS;
189 		c++;
190 	}
191 
192 	if (domain & AMDGPU_GEM_DOMAIN_OA) {
193 		places[c].fpfn = 0;
194 		places[c].lpfn = 0;
195 		places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_OA;
196 		c++;
197 	}
198 
199 	if (!c) {
200 		places[c].fpfn = 0;
201 		places[c].lpfn = 0;
202 		places[c].flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
203 		c++;
204 	}
205 
206 	BUG_ON(c >= AMDGPU_BO_MAX_PLACEMENTS);
207 
208 	placement->num_placement = c;
209 	placement->placement = places;
210 
211 	placement->num_busy_placement = c;
212 	placement->busy_placement = places;
213 }
214 
215 /**
216  * amdgpu_bo_create_reserved - create reserved BO for kernel use
217  *
218  * @adev: amdgpu device object
219  * @size: size for the new BO
220  * @align: alignment for the new BO
221  * @domain: where to place it
222  * @bo_ptr: used to initialize BOs in structures
223  * @gpu_addr: GPU addr of the pinned BO
224  * @cpu_addr: optional CPU address mapping
225  *
226  * Allocates and pins a BO for kernel internal use, and returns it still
227  * reserved.
228  *
229  * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
230  *
231  * Returns:
232  * 0 on success, negative error code otherwise.
233  */
234 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
235 			      unsigned long size, int align,
236 			      u32 domain, struct amdgpu_bo **bo_ptr,
237 			      u64 *gpu_addr, void **cpu_addr)
238 {
239 	struct amdgpu_bo_param bp;
240 	bool free = false;
241 	int r;
242 
243 	if (!size) {
244 		amdgpu_bo_unref(bo_ptr);
245 		return 0;
246 	}
247 
248 	memset(&bp, 0, sizeof(bp));
249 	bp.size = size;
250 	bp.byte_align = align;
251 	bp.domain = domain;
252 	bp.flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
253 		AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
254 	bp.type = ttm_bo_type_kernel;
255 	bp.resv = NULL;
256 
257 	if (!*bo_ptr) {
258 		r = amdgpu_bo_create(adev, &bp, bo_ptr);
259 		if (r) {
260 			dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
261 				r);
262 			return r;
263 		}
264 		free = true;
265 	}
266 
267 	r = amdgpu_bo_reserve(*bo_ptr, false);
268 	if (r) {
269 		dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
270 		goto error_free;
271 	}
272 
273 	r = amdgpu_bo_pin(*bo_ptr, domain);
274 	if (r) {
275 		dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
276 		goto error_unreserve;
277 	}
278 
279 	r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
280 	if (r) {
281 		dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
282 		goto error_unpin;
283 	}
284 
285 	if (gpu_addr)
286 		*gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
287 
288 	if (cpu_addr) {
289 		r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
290 		if (r) {
291 			dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
292 			goto error_unpin;
293 		}
294 	}
295 
296 	return 0;
297 
298 error_unpin:
299 	amdgpu_bo_unpin(*bo_ptr);
300 error_unreserve:
301 	amdgpu_bo_unreserve(*bo_ptr);
302 
303 error_free:
304 	if (free)
305 		amdgpu_bo_unref(bo_ptr);
306 
307 	return r;
308 }
309 
310 /**
311  * amdgpu_bo_create_kernel - create BO for kernel use
312  *
313  * @adev: amdgpu device object
314  * @size: size for the new BO
315  * @align: alignment for the new BO
316  * @domain: where to place it
317  * @bo_ptr:  used to initialize BOs in structures
318  * @gpu_addr: GPU addr of the pinned BO
319  * @cpu_addr: optional CPU address mapping
320  *
321  * Allocates and pins a BO for kernel internal use.
322  *
323  * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
324  *
325  * Returns:
326  * 0 on success, negative error code otherwise.
327  */
328 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
329 			    unsigned long size, int align,
330 			    u32 domain, struct amdgpu_bo **bo_ptr,
331 			    u64 *gpu_addr, void **cpu_addr)
332 {
333 	int r;
334 
335 	r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
336 				      gpu_addr, cpu_addr);
337 
338 	if (r)
339 		return r;
340 
341 	if (*bo_ptr)
342 		amdgpu_bo_unreserve(*bo_ptr);
343 
344 	return 0;
345 }
346 
347 /**
348  * amdgpu_bo_free_kernel - free BO for kernel use
349  *
350  * @bo: amdgpu BO to free
351  * @gpu_addr: pointer to where the BO's GPU memory space address was stored
352  * @cpu_addr: pointer to where the BO's CPU memory space address was stored
353  *
354  * unmaps and unpin a BO for kernel internal use.
355  */
356 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
357 			   void **cpu_addr)
358 {
359 	if (*bo == NULL)
360 		return;
361 
362 	if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
363 		if (cpu_addr)
364 			amdgpu_bo_kunmap(*bo);
365 
366 		amdgpu_bo_unpin(*bo);
367 		amdgpu_bo_unreserve(*bo);
368 	}
369 	amdgpu_bo_unref(bo);
370 
371 	if (gpu_addr)
372 		*gpu_addr = 0;
373 
374 	if (cpu_addr)
375 		*cpu_addr = NULL;
376 }
377 
378 /* Validate bo size is bit bigger then the request domain */
379 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
380 					  unsigned long size, u32 domain)
381 {
382 	struct ttm_mem_type_manager *man = NULL;
383 
384 	/*
385 	 * If GTT is part of requested domains the check must succeed to
386 	 * allow fall back to GTT
387 	 */
388 	if (domain & AMDGPU_GEM_DOMAIN_GTT) {
389 		man = &adev->mman.bdev.man[TTM_PL_TT];
390 
391 		if (size < (man->size << PAGE_SHIFT))
392 			return true;
393 		else
394 			goto fail;
395 	}
396 
397 	if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
398 		man = &adev->mman.bdev.man[TTM_PL_VRAM];
399 
400 		if (size < (man->size << PAGE_SHIFT))
401 			return true;
402 		else
403 			goto fail;
404 	}
405 
406 
407 	/* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */
408 	return true;
409 
410 fail:
411 	DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size,
412 		  man->size << PAGE_SHIFT);
413 	return false;
414 }
415 
416 static int amdgpu_bo_do_create(struct amdgpu_device *adev,
417 			       struct amdgpu_bo_param *bp,
418 			       struct amdgpu_bo **bo_ptr)
419 {
420 	struct ttm_operation_ctx ctx = {
421 		.interruptible = (bp->type != ttm_bo_type_kernel),
422 		.no_wait_gpu = false,
423 		.resv = bp->resv,
424 		.flags = TTM_OPT_FLAG_ALLOW_RES_EVICT
425 	};
426 	struct amdgpu_bo *bo;
427 	unsigned long page_align, size = bp->size;
428 	size_t acc_size;
429 	int r;
430 
431 	/* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
432 	if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
433 		/* GWS and OA don't need any alignment. */
434 		page_align = bp->byte_align;
435 		size <<= PAGE_SHIFT;
436 	} else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
437 		/* Both size and alignment must be a multiple of 4. */
438 		page_align = ALIGN(bp->byte_align, 4);
439 		size = ALIGN(size, 4) << PAGE_SHIFT;
440 	} else {
441 		/* Memory should be aligned at least to a page size. */
442 		page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
443 		size = ALIGN(size, PAGE_SIZE);
444 	}
445 
446 	if (!amdgpu_bo_validate_size(adev, size, bp->domain))
447 		return -ENOMEM;
448 
449 	*bo_ptr = NULL;
450 
451 	acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
452 				       sizeof(struct amdgpu_bo));
453 
454 	bo = kzalloc(sizeof(struct amdgpu_bo), GFP_KERNEL);
455 	if (bo == NULL)
456 		return -ENOMEM;
457 	drm_gem_private_object_init(adev->ddev, &bo->gem_base, size);
458 	INIT_LIST_HEAD(&bo->shadow_list);
459 	bo->vm_bo = NULL;
460 	bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
461 		bp->domain;
462 	bo->allowed_domains = bo->preferred_domains;
463 	if (bp->type != ttm_bo_type_kernel &&
464 	    bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
465 		bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
466 
467 	bo->flags = bp->flags;
468 
469 #ifdef CONFIG_X86_32
470 	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
471 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
472 	 */
473 	bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
474 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
475 	/* Don't try to enable write-combining when it can't work, or things
476 	 * may be slow
477 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
478 	 */
479 
480 #ifndef CONFIG_COMPILE_TEST
481 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
482 	 thanks to write-combining
483 #endif
484 
485 	if (bo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
486 		DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
487 			      "better performance thanks to write-combining\n");
488 	bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
489 #else
490 	/* For architectures that don't support WC memory,
491 	 * mask out the WC flag from the BO
492 	 */
493 	if (!drm_arch_can_wc_memory())
494 		bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
495 #endif
496 
497 	bo->tbo.bdev = &adev->mman.bdev;
498 	amdgpu_bo_placement_from_domain(bo, bp->domain);
499 	if (bp->type == ttm_bo_type_kernel)
500 		bo->tbo.priority = 1;
501 
502 	r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, bp->type,
503 				 &bo->placement, page_align, &ctx, acc_size,
504 				 NULL, bp->resv, &amdgpu_bo_destroy);
505 	if (unlikely(r != 0))
506 		return r;
507 
508 	if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
509 	    bo->tbo.mem.mem_type == TTM_PL_VRAM &&
510 	    bo->tbo.mem.start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
511 		amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
512 					     ctx.bytes_moved);
513 	else
514 		amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
515 
516 	if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
517 	    bo->tbo.mem.placement & TTM_PL_FLAG_VRAM) {
518 		struct dma_fence *fence;
519 
520 		r = amdgpu_fill_buffer(bo, 0, bo->tbo.resv, &fence);
521 		if (unlikely(r))
522 			goto fail_unreserve;
523 
524 		amdgpu_bo_fence(bo, fence, false);
525 		dma_fence_put(bo->tbo.moving);
526 		bo->tbo.moving = dma_fence_get(fence);
527 		dma_fence_put(fence);
528 	}
529 	if (!bp->resv)
530 		amdgpu_bo_unreserve(bo);
531 	*bo_ptr = bo;
532 
533 	trace_amdgpu_bo_create(bo);
534 
535 	/* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
536 	if (bp->type == ttm_bo_type_device)
537 		bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
538 
539 	return 0;
540 
541 fail_unreserve:
542 	if (!bp->resv)
543 		ww_mutex_unlock(&bo->tbo.resv->lock);
544 	amdgpu_bo_unref(&bo);
545 	return r;
546 }
547 
548 static int amdgpu_bo_create_shadow(struct amdgpu_device *adev,
549 				   unsigned long size,
550 				   struct amdgpu_bo *bo)
551 {
552 	struct amdgpu_bo_param bp;
553 	int r;
554 
555 	if (bo->shadow)
556 		return 0;
557 
558 	memset(&bp, 0, sizeof(bp));
559 	bp.size = size;
560 	bp.domain = AMDGPU_GEM_DOMAIN_GTT;
561 	bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC |
562 		AMDGPU_GEM_CREATE_SHADOW;
563 	bp.type = ttm_bo_type_kernel;
564 	bp.resv = bo->tbo.resv;
565 
566 	r = amdgpu_bo_do_create(adev, &bp, &bo->shadow);
567 	if (!r) {
568 		bo->shadow->parent = amdgpu_bo_ref(bo);
569 		mutex_lock(&adev->shadow_list_lock);
570 		list_add_tail(&bo->shadow->shadow_list, &adev->shadow_list);
571 		mutex_unlock(&adev->shadow_list_lock);
572 	}
573 
574 	return r;
575 }
576 
577 /**
578  * amdgpu_bo_create - create an &amdgpu_bo buffer object
579  * @adev: amdgpu device object
580  * @bp: parameters to be used for the buffer object
581  * @bo_ptr: pointer to the buffer object pointer
582  *
583  * Creates an &amdgpu_bo buffer object; and if requested, also creates a
584  * shadow object.
585  * Shadow object is used to backup the original buffer object, and is always
586  * in GTT.
587  *
588  * Returns:
589  * 0 for success or a negative error code on failure.
590  */
591 int amdgpu_bo_create(struct amdgpu_device *adev,
592 		     struct amdgpu_bo_param *bp,
593 		     struct amdgpu_bo **bo_ptr)
594 {
595 	u64 flags = bp->flags;
596 	int r;
597 
598 	bp->flags = bp->flags & ~AMDGPU_GEM_CREATE_SHADOW;
599 	r = amdgpu_bo_do_create(adev, bp, bo_ptr);
600 	if (r)
601 		return r;
602 
603 	if ((flags & AMDGPU_GEM_CREATE_SHADOW) && !(adev->flags & AMD_IS_APU)) {
604 		if (!bp->resv)
605 			WARN_ON(reservation_object_lock((*bo_ptr)->tbo.resv,
606 							NULL));
607 
608 		r = amdgpu_bo_create_shadow(adev, bp->size, *bo_ptr);
609 
610 		if (!bp->resv)
611 			reservation_object_unlock((*bo_ptr)->tbo.resv);
612 
613 		if (r)
614 			amdgpu_bo_unref(bo_ptr);
615 	}
616 
617 	return r;
618 }
619 
620 /**
621  * amdgpu_bo_validate - validate an &amdgpu_bo buffer object
622  * @bo: pointer to the buffer object
623  *
624  * Sets placement according to domain; and changes placement and caching
625  * policy of the buffer object according to the placement.
626  * This is used for validating shadow bos.  It calls ttm_bo_validate() to
627  * make sure the buffer is resident where it needs to be.
628  *
629  * Returns:
630  * 0 for success or a negative error code on failure.
631  */
632 int amdgpu_bo_validate(struct amdgpu_bo *bo)
633 {
634 	struct ttm_operation_ctx ctx = { false, false };
635 	uint32_t domain;
636 	int r;
637 
638 	if (bo->pin_count)
639 		return 0;
640 
641 	domain = bo->preferred_domains;
642 
643 retry:
644 	amdgpu_bo_placement_from_domain(bo, domain);
645 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
646 	if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
647 		domain = bo->allowed_domains;
648 		goto retry;
649 	}
650 
651 	return r;
652 }
653 
654 /**
655  * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow
656  *
657  * @shadow: &amdgpu_bo shadow to be restored
658  * @fence: dma_fence associated with the operation
659  *
660  * Copies a buffer object's shadow content back to the object.
661  * This is used for recovering a buffer from its shadow in case of a gpu
662  * reset where vram context may be lost.
663  *
664  * Returns:
665  * 0 for success or a negative error code on failure.
666  */
667 int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence)
668 
669 {
670 	struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev);
671 	struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
672 	uint64_t shadow_addr, parent_addr;
673 
674 	shadow_addr = amdgpu_bo_gpu_offset(shadow);
675 	parent_addr = amdgpu_bo_gpu_offset(shadow->parent);
676 
677 	return amdgpu_copy_buffer(ring, shadow_addr, parent_addr,
678 				  amdgpu_bo_size(shadow), NULL, fence,
679 				  true, false);
680 }
681 
682 /**
683  * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
684  * @bo: &amdgpu_bo buffer object to be mapped
685  * @ptr: kernel virtual address to be returned
686  *
687  * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
688  * amdgpu_bo_kptr() to get the kernel virtual address.
689  *
690  * Returns:
691  * 0 for success or a negative error code on failure.
692  */
693 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
694 {
695 	void *kptr;
696 	long r;
697 
698 	if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
699 		return -EPERM;
700 
701 	kptr = amdgpu_bo_kptr(bo);
702 	if (kptr) {
703 		if (ptr)
704 			*ptr = kptr;
705 		return 0;
706 	}
707 
708 	r = reservation_object_wait_timeout_rcu(bo->tbo.resv, false, false,
709 						MAX_SCHEDULE_TIMEOUT);
710 	if (r < 0)
711 		return r;
712 
713 	r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
714 	if (r)
715 		return r;
716 
717 	if (ptr)
718 		*ptr = amdgpu_bo_kptr(bo);
719 
720 	return 0;
721 }
722 
723 /**
724  * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
725  * @bo: &amdgpu_bo buffer object
726  *
727  * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
728  *
729  * Returns:
730  * the virtual address of a buffer object area.
731  */
732 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
733 {
734 	bool is_iomem;
735 
736 	return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
737 }
738 
739 /**
740  * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
741  * @bo: &amdgpu_bo buffer object to be unmapped
742  *
743  * Unmaps a kernel map set up by amdgpu_bo_kmap().
744  */
745 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
746 {
747 	if (bo->kmap.bo)
748 		ttm_bo_kunmap(&bo->kmap);
749 }
750 
751 /**
752  * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
753  * @bo: &amdgpu_bo buffer object
754  *
755  * References the contained &ttm_buffer_object.
756  *
757  * Returns:
758  * a refcounted pointer to the &amdgpu_bo buffer object.
759  */
760 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
761 {
762 	if (bo == NULL)
763 		return NULL;
764 
765 	ttm_bo_get(&bo->tbo);
766 	return bo;
767 }
768 
769 /**
770  * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
771  * @bo: &amdgpu_bo buffer object
772  *
773  * Unreferences the contained &ttm_buffer_object and clear the pointer
774  */
775 void amdgpu_bo_unref(struct amdgpu_bo **bo)
776 {
777 	struct ttm_buffer_object *tbo;
778 
779 	if ((*bo) == NULL)
780 		return;
781 
782 	tbo = &((*bo)->tbo);
783 	ttm_bo_put(tbo);
784 	*bo = NULL;
785 }
786 
787 /**
788  * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
789  * @bo: &amdgpu_bo buffer object to be pinned
790  * @domain: domain to be pinned to
791  * @min_offset: the start of requested address range
792  * @max_offset: the end of requested address range
793  *
794  * Pins the buffer object according to requested domain and address range. If
795  * the memory is unbound gart memory, binds the pages into gart table. Adjusts
796  * pin_count and pin_size accordingly.
797  *
798  * Pinning means to lock pages in memory along with keeping them at a fixed
799  * offset. It is required when a buffer can not be moved, for example, when
800  * a display buffer is being scanned out.
801  *
802  * Compared with amdgpu_bo_pin(), this function gives more flexibility on
803  * where to pin a buffer if there are specific restrictions on where a buffer
804  * must be located.
805  *
806  * Returns:
807  * 0 for success or a negative error code on failure.
808  */
809 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
810 			     u64 min_offset, u64 max_offset)
811 {
812 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
813 	struct ttm_operation_ctx ctx = { false, false };
814 	int r, i;
815 
816 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
817 		return -EPERM;
818 
819 	if (WARN_ON_ONCE(min_offset > max_offset))
820 		return -EINVAL;
821 
822 	/* A shared bo cannot be migrated to VRAM */
823 	if (bo->prime_shared_count) {
824 		if (domain & AMDGPU_GEM_DOMAIN_GTT)
825 			domain = AMDGPU_GEM_DOMAIN_GTT;
826 		else
827 			return -EINVAL;
828 	}
829 
830 	/* This assumes only APU display buffers are pinned with (VRAM|GTT).
831 	 * See function amdgpu_display_supported_domains()
832 	 */
833 	domain = amdgpu_bo_get_preferred_pin_domain(adev, domain);
834 
835 	if (bo->pin_count) {
836 		uint32_t mem_type = bo->tbo.mem.mem_type;
837 
838 		if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
839 			return -EINVAL;
840 
841 		bo->pin_count++;
842 
843 		if (max_offset != 0) {
844 			u64 domain_start = bo->tbo.bdev->man[mem_type].gpu_offset;
845 			WARN_ON_ONCE(max_offset <
846 				     (amdgpu_bo_gpu_offset(bo) - domain_start));
847 		}
848 
849 		return 0;
850 	}
851 
852 	bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
853 	/* force to pin into visible video ram */
854 	if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
855 		bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
856 	amdgpu_bo_placement_from_domain(bo, domain);
857 	for (i = 0; i < bo->placement.num_placement; i++) {
858 		unsigned fpfn, lpfn;
859 
860 		fpfn = min_offset >> PAGE_SHIFT;
861 		lpfn = max_offset >> PAGE_SHIFT;
862 
863 		if (fpfn > bo->placements[i].fpfn)
864 			bo->placements[i].fpfn = fpfn;
865 		if (!bo->placements[i].lpfn ||
866 		    (lpfn && lpfn < bo->placements[i].lpfn))
867 			bo->placements[i].lpfn = lpfn;
868 		bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
869 	}
870 
871 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
872 	if (unlikely(r)) {
873 		dev_err(adev->dev, "%p pin failed\n", bo);
874 		goto error;
875 	}
876 
877 	bo->pin_count = 1;
878 
879 	domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
880 	if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
881 		atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
882 		atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
883 			     &adev->visible_pin_size);
884 	} else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
885 		atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
886 	}
887 
888 error:
889 	return r;
890 }
891 
892 /**
893  * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
894  * @bo: &amdgpu_bo buffer object to be pinned
895  * @domain: domain to be pinned to
896  *
897  * A simple wrapper to amdgpu_bo_pin_restricted().
898  * Provides a simpler API for buffers that do not have any strict restrictions
899  * on where a buffer must be located.
900  *
901  * Returns:
902  * 0 for success or a negative error code on failure.
903  */
904 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
905 {
906 	return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
907 }
908 
909 /**
910  * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
911  * @bo: &amdgpu_bo buffer object to be unpinned
912  *
913  * Decreases the pin_count, and clears the flags if pin_count reaches 0.
914  * Changes placement and pin size accordingly.
915  *
916  * Returns:
917  * 0 for success or a negative error code on failure.
918  */
919 int amdgpu_bo_unpin(struct amdgpu_bo *bo)
920 {
921 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
922 	struct ttm_operation_ctx ctx = { false, false };
923 	int r, i;
924 
925 	if (WARN_ON_ONCE(!bo->pin_count)) {
926 		dev_warn(adev->dev, "%p unpin not necessary\n", bo);
927 		return 0;
928 	}
929 	bo->pin_count--;
930 	if (bo->pin_count)
931 		return 0;
932 
933 	amdgpu_bo_subtract_pin_size(bo);
934 
935 	for (i = 0; i < bo->placement.num_placement; i++) {
936 		bo->placements[i].lpfn = 0;
937 		bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
938 	}
939 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
940 	if (unlikely(r))
941 		dev_err(adev->dev, "%p validate failed for unpin\n", bo);
942 
943 	return r;
944 }
945 
946 /**
947  * amdgpu_bo_evict_vram - evict VRAM buffers
948  * @adev: amdgpu device object
949  *
950  * Evicts all VRAM buffers on the lru list of the memory type.
951  * Mainly used for evicting vram at suspend time.
952  *
953  * Returns:
954  * 0 for success or a negative error code on failure.
955  */
956 int amdgpu_bo_evict_vram(struct amdgpu_device *adev)
957 {
958 	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
959 #ifndef CONFIG_HIBERNATION
960 	if (adev->flags & AMD_IS_APU) {
961 		/* Useless to evict on IGP chips */
962 		return 0;
963 	}
964 #endif
965 	return ttm_bo_evict_mm(&adev->mman.bdev, TTM_PL_VRAM);
966 }
967 
968 static const char *amdgpu_vram_names[] = {
969 	"UNKNOWN",
970 	"GDDR1",
971 	"DDR2",
972 	"GDDR3",
973 	"GDDR4",
974 	"GDDR5",
975 	"HBM",
976 	"DDR3",
977 	"DDR4",
978 };
979 
980 /**
981  * amdgpu_bo_init - initialize memory manager
982  * @adev: amdgpu device object
983  *
984  * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
985  *
986  * Returns:
987  * 0 for success or a negative error code on failure.
988  */
989 int amdgpu_bo_init(struct amdgpu_device *adev)
990 {
991 	/* reserve PAT memory space to WC for VRAM */
992 	arch_io_reserve_memtype_wc(adev->gmc.aper_base,
993 				   adev->gmc.aper_size);
994 
995 	/* Add an MTRR for the VRAM */
996 	adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
997 					      adev->gmc.aper_size);
998 	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
999 		 adev->gmc.mc_vram_size >> 20,
1000 		 (unsigned long long)adev->gmc.aper_size >> 20);
1001 	DRM_INFO("RAM width %dbits %s\n",
1002 		 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1003 	return amdgpu_ttm_init(adev);
1004 }
1005 
1006 /**
1007  * amdgpu_bo_late_init - late init
1008  * @adev: amdgpu device object
1009  *
1010  * Calls amdgpu_ttm_late_init() to free resources used earlier during
1011  * initialization.
1012  *
1013  * Returns:
1014  * 0 for success or a negative error code on failure.
1015  */
1016 int amdgpu_bo_late_init(struct amdgpu_device *adev)
1017 {
1018 	amdgpu_ttm_late_init(adev);
1019 
1020 	return 0;
1021 }
1022 
1023 /**
1024  * amdgpu_bo_fini - tear down memory manager
1025  * @adev: amdgpu device object
1026  *
1027  * Reverses amdgpu_bo_init() to tear down memory manager.
1028  */
1029 void amdgpu_bo_fini(struct amdgpu_device *adev)
1030 {
1031 	amdgpu_ttm_fini(adev);
1032 	arch_phys_wc_del(adev->gmc.vram_mtrr);
1033 	arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1034 }
1035 
1036 /**
1037  * amdgpu_bo_fbdev_mmap - mmap fbdev memory
1038  * @bo: &amdgpu_bo buffer object
1039  * @vma: vma as input from the fbdev mmap method
1040  *
1041  * Calls ttm_fbdev_mmap() to mmap fbdev memory if it is backed by a bo.
1042  *
1043  * Returns:
1044  * 0 for success or a negative error code on failure.
1045  */
1046 int amdgpu_bo_fbdev_mmap(struct amdgpu_bo *bo,
1047 			     struct vm_area_struct *vma)
1048 {
1049 	return ttm_fbdev_mmap(vma, &bo->tbo);
1050 }
1051 
1052 /**
1053  * amdgpu_bo_set_tiling_flags - set tiling flags
1054  * @bo: &amdgpu_bo buffer object
1055  * @tiling_flags: new flags
1056  *
1057  * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1058  * kernel driver to set the tiling flags on a buffer.
1059  *
1060  * Returns:
1061  * 0 for success or a negative error code on failure.
1062  */
1063 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1064 {
1065 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1066 
1067 	if (adev->family <= AMDGPU_FAMILY_CZ &&
1068 	    AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1069 		return -EINVAL;
1070 
1071 	bo->tiling_flags = tiling_flags;
1072 	return 0;
1073 }
1074 
1075 /**
1076  * amdgpu_bo_get_tiling_flags - get tiling flags
1077  * @bo: &amdgpu_bo buffer object
1078  * @tiling_flags: returned flags
1079  *
1080  * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1081  * set the tiling flags on a buffer.
1082  */
1083 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1084 {
1085 	lockdep_assert_held(&bo->tbo.resv->lock.base);
1086 
1087 	if (tiling_flags)
1088 		*tiling_flags = bo->tiling_flags;
1089 }
1090 
1091 /**
1092  * amdgpu_bo_set_metadata - set metadata
1093  * @bo: &amdgpu_bo buffer object
1094  * @metadata: new metadata
1095  * @metadata_size: size of the new metadata
1096  * @flags: flags of the new metadata
1097  *
1098  * Sets buffer object's metadata, its size and flags.
1099  * Used via GEM ioctl.
1100  *
1101  * Returns:
1102  * 0 for success or a negative error code on failure.
1103  */
1104 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
1105 			    uint32_t metadata_size, uint64_t flags)
1106 {
1107 	void *buffer;
1108 
1109 	if (!metadata_size) {
1110 		if (bo->metadata_size) {
1111 			kfree(bo->metadata);
1112 			bo->metadata = NULL;
1113 			bo->metadata_size = 0;
1114 		}
1115 		return 0;
1116 	}
1117 
1118 	if (metadata == NULL)
1119 		return -EINVAL;
1120 
1121 	buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1122 	if (buffer == NULL)
1123 		return -ENOMEM;
1124 
1125 	kfree(bo->metadata);
1126 	bo->metadata_flags = flags;
1127 	bo->metadata = buffer;
1128 	bo->metadata_size = metadata_size;
1129 
1130 	return 0;
1131 }
1132 
1133 /**
1134  * amdgpu_bo_get_metadata - get metadata
1135  * @bo: &amdgpu_bo buffer object
1136  * @buffer: returned metadata
1137  * @buffer_size: size of the buffer
1138  * @metadata_size: size of the returned metadata
1139  * @flags: flags of the returned metadata
1140  *
1141  * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1142  * less than metadata_size.
1143  * Used via GEM ioctl.
1144  *
1145  * Returns:
1146  * 0 for success or a negative error code on failure.
1147  */
1148 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1149 			   size_t buffer_size, uint32_t *metadata_size,
1150 			   uint64_t *flags)
1151 {
1152 	if (!buffer && !metadata_size)
1153 		return -EINVAL;
1154 
1155 	if (buffer) {
1156 		if (buffer_size < bo->metadata_size)
1157 			return -EINVAL;
1158 
1159 		if (bo->metadata_size)
1160 			memcpy(buffer, bo->metadata, bo->metadata_size);
1161 	}
1162 
1163 	if (metadata_size)
1164 		*metadata_size = bo->metadata_size;
1165 	if (flags)
1166 		*flags = bo->metadata_flags;
1167 
1168 	return 0;
1169 }
1170 
1171 /**
1172  * amdgpu_bo_move_notify - notification about a memory move
1173  * @bo: pointer to a buffer object
1174  * @evict: if this move is evicting the buffer from the graphics address space
1175  * @new_mem: new information of the bufer object
1176  *
1177  * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1178  * bookkeeping.
1179  * TTM driver callback which is called when ttm moves a buffer.
1180  */
1181 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1182 			   bool evict,
1183 			   struct ttm_mem_reg *new_mem)
1184 {
1185 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1186 	struct amdgpu_bo *abo;
1187 	struct ttm_mem_reg *old_mem = &bo->mem;
1188 
1189 	if (!amdgpu_bo_is_amdgpu_bo(bo))
1190 		return;
1191 
1192 	abo = ttm_to_amdgpu_bo(bo);
1193 	amdgpu_vm_bo_invalidate(adev, abo, evict);
1194 
1195 	amdgpu_bo_kunmap(abo);
1196 
1197 	/* remember the eviction */
1198 	if (evict)
1199 		atomic64_inc(&adev->num_evictions);
1200 
1201 	/* update statistics */
1202 	if (!new_mem)
1203 		return;
1204 
1205 	/* move_notify is called before move happens */
1206 	trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
1207 }
1208 
1209 /**
1210  * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1211  * @bo: pointer to a buffer object
1212  *
1213  * Notifies the driver we are taking a fault on this BO and have reserved it,
1214  * also performs bookkeeping.
1215  * TTM driver callback for dealing with vm faults.
1216  *
1217  * Returns:
1218  * 0 for success or a negative error code on failure.
1219  */
1220 int amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1221 {
1222 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1223 	struct ttm_operation_ctx ctx = { false, false };
1224 	struct amdgpu_bo *abo;
1225 	unsigned long offset, size;
1226 	int r;
1227 
1228 	if (!amdgpu_bo_is_amdgpu_bo(bo))
1229 		return 0;
1230 
1231 	abo = ttm_to_amdgpu_bo(bo);
1232 
1233 	/* Remember that this BO was accessed by the CPU */
1234 	abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1235 
1236 	if (bo->mem.mem_type != TTM_PL_VRAM)
1237 		return 0;
1238 
1239 	size = bo->mem.num_pages << PAGE_SHIFT;
1240 	offset = bo->mem.start << PAGE_SHIFT;
1241 	if ((offset + size) <= adev->gmc.visible_vram_size)
1242 		return 0;
1243 
1244 	/* Can't move a pinned BO to visible VRAM */
1245 	if (abo->pin_count > 0)
1246 		return -EINVAL;
1247 
1248 	/* hurrah the memory is not visible ! */
1249 	atomic64_inc(&adev->num_vram_cpu_page_faults);
1250 	amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1251 					AMDGPU_GEM_DOMAIN_GTT);
1252 
1253 	/* Avoid costly evictions; only set GTT as a busy placement */
1254 	abo->placement.num_busy_placement = 1;
1255 	abo->placement.busy_placement = &abo->placements[1];
1256 
1257 	r = ttm_bo_validate(bo, &abo->placement, &ctx);
1258 	if (unlikely(r != 0))
1259 		return r;
1260 
1261 	offset = bo->mem.start << PAGE_SHIFT;
1262 	/* this should never happen */
1263 	if (bo->mem.mem_type == TTM_PL_VRAM &&
1264 	    (offset + size) > adev->gmc.visible_vram_size)
1265 		return -EINVAL;
1266 
1267 	return 0;
1268 }
1269 
1270 /**
1271  * amdgpu_bo_fence - add fence to buffer object
1272  *
1273  * @bo: buffer object in question
1274  * @fence: fence to add
1275  * @shared: true if fence should be added shared
1276  *
1277  */
1278 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1279 		     bool shared)
1280 {
1281 	struct reservation_object *resv = bo->tbo.resv;
1282 
1283 	if (shared)
1284 		reservation_object_add_shared_fence(resv, fence);
1285 	else
1286 		reservation_object_add_excl_fence(resv, fence);
1287 }
1288 
1289 /**
1290  * amdgpu_sync_wait_resv - Wait for BO reservation fences
1291  *
1292  * @bo: buffer object
1293  * @owner: fence owner
1294  * @intr: Whether the wait is interruptible
1295  *
1296  * Returns:
1297  * 0 on success, errno otherwise.
1298  */
1299 int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)
1300 {
1301 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1302 	struct amdgpu_sync sync;
1303 	int r;
1304 
1305 	amdgpu_sync_create(&sync);
1306 	amdgpu_sync_resv(adev, &sync, bo->tbo.resv, owner, false);
1307 	r = amdgpu_sync_wait(&sync, intr);
1308 	amdgpu_sync_free(&sync);
1309 
1310 	return r;
1311 }
1312 
1313 /**
1314  * amdgpu_bo_gpu_offset - return GPU offset of bo
1315  * @bo:	amdgpu object for which we query the offset
1316  *
1317  * Note: object should either be pinned or reserved when calling this
1318  * function, it might be useful to add check for this for debugging.
1319  *
1320  * Returns:
1321  * current GPU offset of the object.
1322  */
1323 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1324 {
1325 	WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_SYSTEM);
1326 	WARN_ON_ONCE(!ww_mutex_is_locked(&bo->tbo.resv->lock) &&
1327 		     !bo->pin_count && bo->tbo.type != ttm_bo_type_kernel);
1328 	WARN_ON_ONCE(bo->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET);
1329 	WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_VRAM &&
1330 		     !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1331 
1332 	return amdgpu_gmc_sign_extend(bo->tbo.offset);
1333 }
1334 
1335 /**
1336  * amdgpu_bo_get_preferred_pin_domain - get preferred domain for scanout
1337  * @adev: amdgpu device object
1338  * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1339  *
1340  * Returns:
1341  * Which of the allowed domains is preferred for pinning the BO for scanout.
1342  */
1343 uint32_t amdgpu_bo_get_preferred_pin_domain(struct amdgpu_device *adev,
1344 					    uint32_t domain)
1345 {
1346 	if (domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) {
1347 		domain = AMDGPU_GEM_DOMAIN_VRAM;
1348 		if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1349 			domain = AMDGPU_GEM_DOMAIN_GTT;
1350 	}
1351 	return domain;
1352 }
1353