xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c (revision d56f3af9e801970d21c57621de3b42bc17eac152)
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 
40 static bool amdgpu_need_backup(struct amdgpu_device *adev)
41 {
42 	if (adev->flags & AMD_IS_APU)
43 		return false;
44 
45 	if (amdgpu_gpu_recovery == 0 ||
46 	    (amdgpu_gpu_recovery == -1  && !amdgpu_sriov_vf(adev)))
47 		return false;
48 
49 	return true;
50 }
51 
52 static void amdgpu_ttm_bo_destroy(struct ttm_buffer_object *tbo)
53 {
54 	struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
55 	struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
56 
57 	amdgpu_bo_kunmap(bo);
58 
59 	if (bo->gem_base.import_attach)
60 		drm_prime_gem_destroy(&bo->gem_base, bo->tbo.sg);
61 	drm_gem_object_release(&bo->gem_base);
62 	amdgpu_bo_unref(&bo->parent);
63 	if (!list_empty(&bo->shadow_list)) {
64 		mutex_lock(&adev->shadow_list_lock);
65 		list_del_init(&bo->shadow_list);
66 		mutex_unlock(&adev->shadow_list_lock);
67 	}
68 	kfree(bo->metadata);
69 	kfree(bo);
70 }
71 
72 bool amdgpu_ttm_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
73 {
74 	if (bo->destroy == &amdgpu_ttm_bo_destroy)
75 		return true;
76 	return false;
77 }
78 
79 void amdgpu_ttm_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
80 {
81 	struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
82 	struct ttm_placement *placement = &abo->placement;
83 	struct ttm_place *places = abo->placements;
84 	u64 flags = abo->flags;
85 	u32 c = 0;
86 
87 	if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
88 		unsigned visible_pfn = adev->mc.visible_vram_size >> PAGE_SHIFT;
89 
90 		places[c].fpfn = 0;
91 		places[c].lpfn = 0;
92 		places[c].flags = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
93 			TTM_PL_FLAG_VRAM;
94 
95 		if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
96 			places[c].lpfn = visible_pfn;
97 		else
98 			places[c].flags |= TTM_PL_FLAG_TOPDOWN;
99 
100 		if (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
101 			places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
102 		c++;
103 	}
104 
105 	if (domain & AMDGPU_GEM_DOMAIN_GTT) {
106 		places[c].fpfn = 0;
107 		if (flags & AMDGPU_GEM_CREATE_SHADOW)
108 			places[c].lpfn = adev->mc.gart_size >> PAGE_SHIFT;
109 		else
110 			places[c].lpfn = 0;
111 		places[c].flags = TTM_PL_FLAG_TT;
112 		if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
113 			places[c].flags |= TTM_PL_FLAG_WC |
114 				TTM_PL_FLAG_UNCACHED;
115 		else
116 			places[c].flags |= TTM_PL_FLAG_CACHED;
117 		c++;
118 	}
119 
120 	if (domain & AMDGPU_GEM_DOMAIN_CPU) {
121 		places[c].fpfn = 0;
122 		places[c].lpfn = 0;
123 		places[c].flags = TTM_PL_FLAG_SYSTEM;
124 		if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
125 			places[c].flags |= TTM_PL_FLAG_WC |
126 				TTM_PL_FLAG_UNCACHED;
127 		else
128 			places[c].flags |= TTM_PL_FLAG_CACHED;
129 		c++;
130 	}
131 
132 	if (domain & AMDGPU_GEM_DOMAIN_GDS) {
133 		places[c].fpfn = 0;
134 		places[c].lpfn = 0;
135 		places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GDS;
136 		c++;
137 	}
138 
139 	if (domain & AMDGPU_GEM_DOMAIN_GWS) {
140 		places[c].fpfn = 0;
141 		places[c].lpfn = 0;
142 		places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GWS;
143 		c++;
144 	}
145 
146 	if (domain & AMDGPU_GEM_DOMAIN_OA) {
147 		places[c].fpfn = 0;
148 		places[c].lpfn = 0;
149 		places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_OA;
150 		c++;
151 	}
152 
153 	if (!c) {
154 		places[c].fpfn = 0;
155 		places[c].lpfn = 0;
156 		places[c].flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
157 		c++;
158 	}
159 
160 	placement->num_placement = c;
161 	placement->placement = places;
162 
163 	placement->num_busy_placement = c;
164 	placement->busy_placement = places;
165 }
166 
167 /**
168  * amdgpu_bo_create_reserved - create reserved BO for kernel use
169  *
170  * @adev: amdgpu device object
171  * @size: size for the new BO
172  * @align: alignment for the new BO
173  * @domain: where to place it
174  * @bo_ptr: resulting BO
175  * @gpu_addr: GPU addr of the pinned BO
176  * @cpu_addr: optional CPU address mapping
177  *
178  * Allocates and pins a BO for kernel internal use, and returns it still
179  * reserved.
180  *
181  * Returns 0 on success, negative error code otherwise.
182  */
183 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
184 			      unsigned long size, int align,
185 			      u32 domain, struct amdgpu_bo **bo_ptr,
186 			      u64 *gpu_addr, void **cpu_addr)
187 {
188 	bool free = false;
189 	int r;
190 
191 	if (!*bo_ptr) {
192 		r = amdgpu_bo_create(adev, size, align, true, domain,
193 				     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
194 				     AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS,
195 				     NULL, NULL, 0, bo_ptr);
196 		if (r) {
197 			dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
198 				r);
199 			return r;
200 		}
201 		free = true;
202 	}
203 
204 	r = amdgpu_bo_reserve(*bo_ptr, false);
205 	if (r) {
206 		dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
207 		goto error_free;
208 	}
209 
210 	r = amdgpu_bo_pin(*bo_ptr, domain, gpu_addr);
211 	if (r) {
212 		dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
213 		goto error_unreserve;
214 	}
215 
216 	if (cpu_addr) {
217 		r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
218 		if (r) {
219 			dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
220 			goto error_unreserve;
221 		}
222 	}
223 
224 	return 0;
225 
226 error_unreserve:
227 	amdgpu_bo_unreserve(*bo_ptr);
228 
229 error_free:
230 	if (free)
231 		amdgpu_bo_unref(bo_ptr);
232 
233 	return r;
234 }
235 
236 /**
237  * amdgpu_bo_create_kernel - create BO for kernel use
238  *
239  * @adev: amdgpu device object
240  * @size: size for the new BO
241  * @align: alignment for the new BO
242  * @domain: where to place it
243  * @bo_ptr: resulting BO
244  * @gpu_addr: GPU addr of the pinned BO
245  * @cpu_addr: optional CPU address mapping
246  *
247  * Allocates and pins a BO for kernel internal use.
248  *
249  * Returns 0 on success, negative error code otherwise.
250  */
251 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
252 			    unsigned long size, int align,
253 			    u32 domain, struct amdgpu_bo **bo_ptr,
254 			    u64 *gpu_addr, void **cpu_addr)
255 {
256 	int r;
257 
258 	r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
259 				      gpu_addr, cpu_addr);
260 
261 	if (r)
262 		return r;
263 
264 	amdgpu_bo_unreserve(*bo_ptr);
265 
266 	return 0;
267 }
268 
269 /**
270  * amdgpu_bo_free_kernel - free BO for kernel use
271  *
272  * @bo: amdgpu BO to free
273  *
274  * unmaps and unpin a BO for kernel internal use.
275  */
276 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
277 			   void **cpu_addr)
278 {
279 	if (*bo == NULL)
280 		return;
281 
282 	if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
283 		if (cpu_addr)
284 			amdgpu_bo_kunmap(*bo);
285 
286 		amdgpu_bo_unpin(*bo);
287 		amdgpu_bo_unreserve(*bo);
288 	}
289 	amdgpu_bo_unref(bo);
290 
291 	if (gpu_addr)
292 		*gpu_addr = 0;
293 
294 	if (cpu_addr)
295 		*cpu_addr = NULL;
296 }
297 
298 /* Validate bo size is bit bigger then the request domain */
299 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
300 					  unsigned long size, u32 domain)
301 {
302 	struct ttm_mem_type_manager *man = NULL;
303 
304 	/*
305 	 * If GTT is part of requested domains the check must succeed to
306 	 * allow fall back to GTT
307 	 */
308 	if (domain & AMDGPU_GEM_DOMAIN_GTT) {
309 		man = &adev->mman.bdev.man[TTM_PL_TT];
310 
311 		if (size < (man->size << PAGE_SHIFT))
312 			return true;
313 		else
314 			goto fail;
315 	}
316 
317 	if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
318 		man = &adev->mman.bdev.man[TTM_PL_VRAM];
319 
320 		if (size < (man->size << PAGE_SHIFT))
321 			return true;
322 		else
323 			goto fail;
324 	}
325 
326 
327 	/* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */
328 	return true;
329 
330 fail:
331 	DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size,
332 		  man->size << PAGE_SHIFT);
333 	return false;
334 }
335 
336 static int amdgpu_bo_do_create(struct amdgpu_device *adev,
337 			       unsigned long size, int byte_align,
338 			       bool kernel, u32 domain, u64 flags,
339 			       struct sg_table *sg,
340 			       struct reservation_object *resv,
341 			       uint64_t init_value,
342 			       struct amdgpu_bo **bo_ptr)
343 {
344 	struct ttm_operation_ctx ctx = {
345 		.interruptible = !kernel,
346 		.no_wait_gpu = false,
347 		.allow_reserved_eviction = true,
348 		.resv = resv
349 	};
350 	struct amdgpu_bo *bo;
351 	enum ttm_bo_type type;
352 	unsigned long page_align;
353 	size_t acc_size;
354 	int r;
355 
356 	page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
357 	size = ALIGN(size, PAGE_SIZE);
358 
359 	if (!amdgpu_bo_validate_size(adev, size, domain))
360 		return -ENOMEM;
361 
362 	if (kernel) {
363 		type = ttm_bo_type_kernel;
364 	} else if (sg) {
365 		type = ttm_bo_type_sg;
366 	} else {
367 		type = ttm_bo_type_device;
368 	}
369 	*bo_ptr = NULL;
370 
371 	acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
372 				       sizeof(struct amdgpu_bo));
373 
374 	bo = kzalloc(sizeof(struct amdgpu_bo), GFP_KERNEL);
375 	if (bo == NULL)
376 		return -ENOMEM;
377 	r = drm_gem_object_init(adev->ddev, &bo->gem_base, size);
378 	if (unlikely(r)) {
379 		kfree(bo);
380 		return r;
381 	}
382 	INIT_LIST_HEAD(&bo->shadow_list);
383 	INIT_LIST_HEAD(&bo->va);
384 	bo->preferred_domains = domain & (AMDGPU_GEM_DOMAIN_VRAM |
385 					 AMDGPU_GEM_DOMAIN_GTT |
386 					 AMDGPU_GEM_DOMAIN_CPU |
387 					 AMDGPU_GEM_DOMAIN_GDS |
388 					 AMDGPU_GEM_DOMAIN_GWS |
389 					 AMDGPU_GEM_DOMAIN_OA);
390 	bo->allowed_domains = bo->preferred_domains;
391 	if (!kernel && bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
392 		bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
393 
394 	bo->flags = flags;
395 
396 #ifdef CONFIG_X86_32
397 	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
398 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
399 	 */
400 	bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
401 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
402 	/* Don't try to enable write-combining when it can't work, or things
403 	 * may be slow
404 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
405 	 */
406 
407 #ifndef CONFIG_COMPILE_TEST
408 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
409 	 thanks to write-combining
410 #endif
411 
412 	if (bo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
413 		DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
414 			      "better performance thanks to write-combining\n");
415 	bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
416 #else
417 	/* For architectures that don't support WC memory,
418 	 * mask out the WC flag from the BO
419 	 */
420 	if (!drm_arch_can_wc_memory())
421 		bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
422 #endif
423 
424 	bo->tbo.bdev = &adev->mman.bdev;
425 	amdgpu_ttm_placement_from_domain(bo, domain);
426 
427 	r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, type,
428 				 &bo->placement, page_align, &ctx, NULL,
429 				 acc_size, sg, resv, &amdgpu_ttm_bo_destroy);
430 	if (unlikely(r != 0))
431 		return r;
432 
433 	if (adev->mc.visible_vram_size < adev->mc.real_vram_size &&
434 	    bo->tbo.mem.mem_type == TTM_PL_VRAM &&
435 	    bo->tbo.mem.start < adev->mc.visible_vram_size >> PAGE_SHIFT)
436 		amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
437 					     ctx.bytes_moved);
438 	else
439 		amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
440 
441 	if (kernel)
442 		bo->tbo.priority = 1;
443 
444 	if (flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
445 	    bo->tbo.mem.placement & TTM_PL_FLAG_VRAM) {
446 		struct dma_fence *fence;
447 
448 		r = amdgpu_fill_buffer(bo, init_value, bo->tbo.resv, &fence);
449 		if (unlikely(r))
450 			goto fail_unreserve;
451 
452 		amdgpu_bo_fence(bo, fence, false);
453 		dma_fence_put(bo->tbo.moving);
454 		bo->tbo.moving = dma_fence_get(fence);
455 		dma_fence_put(fence);
456 	}
457 	if (!resv)
458 		amdgpu_bo_unreserve(bo);
459 	*bo_ptr = bo;
460 
461 	trace_amdgpu_bo_create(bo);
462 
463 	/* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
464 	if (type == ttm_bo_type_device)
465 		bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
466 
467 	return 0;
468 
469 fail_unreserve:
470 	if (!resv)
471 		ww_mutex_unlock(&bo->tbo.resv->lock);
472 	amdgpu_bo_unref(&bo);
473 	return r;
474 }
475 
476 static int amdgpu_bo_create_shadow(struct amdgpu_device *adev,
477 				   unsigned long size, int byte_align,
478 				   struct amdgpu_bo *bo)
479 {
480 	int r;
481 
482 	if (bo->shadow)
483 		return 0;
484 
485 	r = amdgpu_bo_do_create(adev, size, byte_align, true,
486 				AMDGPU_GEM_DOMAIN_GTT,
487 				AMDGPU_GEM_CREATE_CPU_GTT_USWC |
488 				AMDGPU_GEM_CREATE_SHADOW,
489 				NULL, bo->tbo.resv, 0,
490 				&bo->shadow);
491 	if (!r) {
492 		bo->shadow->parent = amdgpu_bo_ref(bo);
493 		mutex_lock(&adev->shadow_list_lock);
494 		list_add_tail(&bo->shadow_list, &adev->shadow_list);
495 		mutex_unlock(&adev->shadow_list_lock);
496 	}
497 
498 	return r;
499 }
500 
501 /* init_value will only take effect when flags contains
502  * AMDGPU_GEM_CREATE_VRAM_CLEARED.
503  */
504 int amdgpu_bo_create(struct amdgpu_device *adev,
505 		     unsigned long size, int byte_align,
506 		     bool kernel, u32 domain, u64 flags,
507 		     struct sg_table *sg,
508 		     struct reservation_object *resv,
509 		     uint64_t init_value,
510 		     struct amdgpu_bo **bo_ptr)
511 {
512 	uint64_t parent_flags = flags & ~AMDGPU_GEM_CREATE_SHADOW;
513 	int r;
514 
515 	r = amdgpu_bo_do_create(adev, size, byte_align, kernel, domain,
516 				parent_flags, sg, resv, init_value, bo_ptr);
517 	if (r)
518 		return r;
519 
520 	if ((flags & AMDGPU_GEM_CREATE_SHADOW) && amdgpu_need_backup(adev)) {
521 		if (!resv)
522 			WARN_ON(reservation_object_lock((*bo_ptr)->tbo.resv,
523 							NULL));
524 
525 		r = amdgpu_bo_create_shadow(adev, size, byte_align, (*bo_ptr));
526 
527 		if (!resv)
528 			reservation_object_unlock((*bo_ptr)->tbo.resv);
529 
530 		if (r)
531 			amdgpu_bo_unref(bo_ptr);
532 	}
533 
534 	return r;
535 }
536 
537 int amdgpu_bo_backup_to_shadow(struct amdgpu_device *adev,
538 			       struct amdgpu_ring *ring,
539 			       struct amdgpu_bo *bo,
540 			       struct reservation_object *resv,
541 			       struct dma_fence **fence,
542 			       bool direct)
543 
544 {
545 	struct amdgpu_bo *shadow = bo->shadow;
546 	uint64_t bo_addr, shadow_addr;
547 	int r;
548 
549 	if (!shadow)
550 		return -EINVAL;
551 
552 	bo_addr = amdgpu_bo_gpu_offset(bo);
553 	shadow_addr = amdgpu_bo_gpu_offset(bo->shadow);
554 
555 	r = reservation_object_reserve_shared(bo->tbo.resv);
556 	if (r)
557 		goto err;
558 
559 	r = amdgpu_copy_buffer(ring, bo_addr, shadow_addr,
560 			       amdgpu_bo_size(bo), resv, fence,
561 			       direct, false);
562 	if (!r)
563 		amdgpu_bo_fence(bo, *fence, true);
564 
565 err:
566 	return r;
567 }
568 
569 int amdgpu_bo_validate(struct amdgpu_bo *bo)
570 {
571 	struct ttm_operation_ctx ctx = { false, false };
572 	uint32_t domain;
573 	int r;
574 
575 	if (bo->pin_count)
576 		return 0;
577 
578 	domain = bo->preferred_domains;
579 
580 retry:
581 	amdgpu_ttm_placement_from_domain(bo, domain);
582 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
583 	if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
584 		domain = bo->allowed_domains;
585 		goto retry;
586 	}
587 
588 	return r;
589 }
590 
591 int amdgpu_bo_restore_from_shadow(struct amdgpu_device *adev,
592 				  struct amdgpu_ring *ring,
593 				  struct amdgpu_bo *bo,
594 				  struct reservation_object *resv,
595 				  struct dma_fence **fence,
596 				  bool direct)
597 
598 {
599 	struct amdgpu_bo *shadow = bo->shadow;
600 	uint64_t bo_addr, shadow_addr;
601 	int r;
602 
603 	if (!shadow)
604 		return -EINVAL;
605 
606 	bo_addr = amdgpu_bo_gpu_offset(bo);
607 	shadow_addr = amdgpu_bo_gpu_offset(bo->shadow);
608 
609 	r = reservation_object_reserve_shared(bo->tbo.resv);
610 	if (r)
611 		goto err;
612 
613 	r = amdgpu_copy_buffer(ring, shadow_addr, bo_addr,
614 			       amdgpu_bo_size(bo), resv, fence,
615 			       direct, false);
616 	if (!r)
617 		amdgpu_bo_fence(bo, *fence, true);
618 
619 err:
620 	return r;
621 }
622 
623 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
624 {
625 	void *kptr;
626 	long r;
627 
628 	if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
629 		return -EPERM;
630 
631 	kptr = amdgpu_bo_kptr(bo);
632 	if (kptr) {
633 		if (ptr)
634 			*ptr = kptr;
635 		return 0;
636 	}
637 
638 	r = reservation_object_wait_timeout_rcu(bo->tbo.resv, false, false,
639 						MAX_SCHEDULE_TIMEOUT);
640 	if (r < 0)
641 		return r;
642 
643 	r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
644 	if (r)
645 		return r;
646 
647 	if (ptr)
648 		*ptr = amdgpu_bo_kptr(bo);
649 
650 	return 0;
651 }
652 
653 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
654 {
655 	bool is_iomem;
656 
657 	return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
658 }
659 
660 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
661 {
662 	if (bo->kmap.bo)
663 		ttm_bo_kunmap(&bo->kmap);
664 }
665 
666 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
667 {
668 	if (bo == NULL)
669 		return NULL;
670 
671 	ttm_bo_reference(&bo->tbo);
672 	return bo;
673 }
674 
675 void amdgpu_bo_unref(struct amdgpu_bo **bo)
676 {
677 	struct ttm_buffer_object *tbo;
678 
679 	if ((*bo) == NULL)
680 		return;
681 
682 	tbo = &((*bo)->tbo);
683 	ttm_bo_unref(&tbo);
684 	if (tbo == NULL)
685 		*bo = NULL;
686 }
687 
688 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
689 			     u64 min_offset, u64 max_offset,
690 			     u64 *gpu_addr)
691 {
692 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
693 	struct ttm_operation_ctx ctx = { false, false };
694 	int r, i;
695 
696 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
697 		return -EPERM;
698 
699 	if (WARN_ON_ONCE(min_offset > max_offset))
700 		return -EINVAL;
701 
702 	/* A shared bo cannot be migrated to VRAM */
703 	if (bo->prime_shared_count && (domain == AMDGPU_GEM_DOMAIN_VRAM))
704 		return -EINVAL;
705 
706 	if (bo->pin_count) {
707 		uint32_t mem_type = bo->tbo.mem.mem_type;
708 
709 		if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
710 			return -EINVAL;
711 
712 		bo->pin_count++;
713 		if (gpu_addr)
714 			*gpu_addr = amdgpu_bo_gpu_offset(bo);
715 
716 		if (max_offset != 0) {
717 			u64 domain_start = bo->tbo.bdev->man[mem_type].gpu_offset;
718 			WARN_ON_ONCE(max_offset <
719 				     (amdgpu_bo_gpu_offset(bo) - domain_start));
720 		}
721 
722 		return 0;
723 	}
724 
725 	bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
726 	/* force to pin into visible video ram */
727 	if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
728 		bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
729 	amdgpu_ttm_placement_from_domain(bo, domain);
730 	for (i = 0; i < bo->placement.num_placement; i++) {
731 		unsigned fpfn, lpfn;
732 
733 		fpfn = min_offset >> PAGE_SHIFT;
734 		lpfn = max_offset >> PAGE_SHIFT;
735 
736 		if (fpfn > bo->placements[i].fpfn)
737 			bo->placements[i].fpfn = fpfn;
738 		if (!bo->placements[i].lpfn ||
739 		    (lpfn && lpfn < bo->placements[i].lpfn))
740 			bo->placements[i].lpfn = lpfn;
741 		bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
742 	}
743 
744 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
745 	if (unlikely(r)) {
746 		dev_err(adev->dev, "%p pin failed\n", bo);
747 		goto error;
748 	}
749 
750 	r = amdgpu_ttm_alloc_gart(&bo->tbo);
751 	if (unlikely(r)) {
752 		dev_err(adev->dev, "%p bind failed\n", bo);
753 		goto error;
754 	}
755 
756 	bo->pin_count = 1;
757 	if (gpu_addr != NULL)
758 		*gpu_addr = amdgpu_bo_gpu_offset(bo);
759 
760 	domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
761 	if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
762 		adev->vram_pin_size += amdgpu_bo_size(bo);
763 		if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
764 			adev->invisible_pin_size += amdgpu_bo_size(bo);
765 	} else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
766 		adev->gart_pin_size += amdgpu_bo_size(bo);
767 	}
768 
769 error:
770 	return r;
771 }
772 
773 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain, u64 *gpu_addr)
774 {
775 	return amdgpu_bo_pin_restricted(bo, domain, 0, 0, gpu_addr);
776 }
777 
778 int amdgpu_bo_unpin(struct amdgpu_bo *bo)
779 {
780 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
781 	struct ttm_operation_ctx ctx = { false, false };
782 	int r, i;
783 
784 	if (!bo->pin_count) {
785 		dev_warn(adev->dev, "%p unpin not necessary\n", bo);
786 		return 0;
787 	}
788 	bo->pin_count--;
789 	if (bo->pin_count)
790 		return 0;
791 	for (i = 0; i < bo->placement.num_placement; i++) {
792 		bo->placements[i].lpfn = 0;
793 		bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
794 	}
795 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
796 	if (unlikely(r)) {
797 		dev_err(adev->dev, "%p validate failed for unpin\n", bo);
798 		goto error;
799 	}
800 
801 	if (bo->tbo.mem.mem_type == TTM_PL_VRAM) {
802 		adev->vram_pin_size -= amdgpu_bo_size(bo);
803 		if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
804 			adev->invisible_pin_size -= amdgpu_bo_size(bo);
805 	} else if (bo->tbo.mem.mem_type == TTM_PL_TT) {
806 		adev->gart_pin_size -= amdgpu_bo_size(bo);
807 	}
808 
809 error:
810 	return r;
811 }
812 
813 int amdgpu_bo_evict_vram(struct amdgpu_device *adev)
814 {
815 	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
816 	if (0 && (adev->flags & AMD_IS_APU)) {
817 		/* Useless to evict on IGP chips */
818 		return 0;
819 	}
820 	return ttm_bo_evict_mm(&adev->mman.bdev, TTM_PL_VRAM);
821 }
822 
823 static const char *amdgpu_vram_names[] = {
824 	"UNKNOWN",
825 	"GDDR1",
826 	"DDR2",
827 	"GDDR3",
828 	"GDDR4",
829 	"GDDR5",
830 	"HBM",
831 	"DDR3"
832 };
833 
834 int amdgpu_bo_init(struct amdgpu_device *adev)
835 {
836 	/* reserve PAT memory space to WC for VRAM */
837 	arch_io_reserve_memtype_wc(adev->mc.aper_base,
838 				   adev->mc.aper_size);
839 
840 	/* Add an MTRR for the VRAM */
841 	adev->mc.vram_mtrr = arch_phys_wc_add(adev->mc.aper_base,
842 					      adev->mc.aper_size);
843 	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
844 		 adev->mc.mc_vram_size >> 20,
845 		 (unsigned long long)adev->mc.aper_size >> 20);
846 	DRM_INFO("RAM width %dbits %s\n",
847 		 adev->mc.vram_width, amdgpu_vram_names[adev->mc.vram_type]);
848 	return amdgpu_ttm_init(adev);
849 }
850 
851 void amdgpu_bo_fini(struct amdgpu_device *adev)
852 {
853 	amdgpu_ttm_fini(adev);
854 	arch_phys_wc_del(adev->mc.vram_mtrr);
855 	arch_io_free_memtype_wc(adev->mc.aper_base, adev->mc.aper_size);
856 }
857 
858 int amdgpu_bo_fbdev_mmap(struct amdgpu_bo *bo,
859 			     struct vm_area_struct *vma)
860 {
861 	return ttm_fbdev_mmap(vma, &bo->tbo);
862 }
863 
864 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
865 {
866 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
867 
868 	if (adev->family <= AMDGPU_FAMILY_CZ &&
869 	    AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
870 		return -EINVAL;
871 
872 	bo->tiling_flags = tiling_flags;
873 	return 0;
874 }
875 
876 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
877 {
878 	lockdep_assert_held(&bo->tbo.resv->lock.base);
879 
880 	if (tiling_flags)
881 		*tiling_flags = bo->tiling_flags;
882 }
883 
884 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
885 			    uint32_t metadata_size, uint64_t flags)
886 {
887 	void *buffer;
888 
889 	if (!metadata_size) {
890 		if (bo->metadata_size) {
891 			kfree(bo->metadata);
892 			bo->metadata = NULL;
893 			bo->metadata_size = 0;
894 		}
895 		return 0;
896 	}
897 
898 	if (metadata == NULL)
899 		return -EINVAL;
900 
901 	buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
902 	if (buffer == NULL)
903 		return -ENOMEM;
904 
905 	kfree(bo->metadata);
906 	bo->metadata_flags = flags;
907 	bo->metadata = buffer;
908 	bo->metadata_size = metadata_size;
909 
910 	return 0;
911 }
912 
913 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
914 			   size_t buffer_size, uint32_t *metadata_size,
915 			   uint64_t *flags)
916 {
917 	if (!buffer && !metadata_size)
918 		return -EINVAL;
919 
920 	if (buffer) {
921 		if (buffer_size < bo->metadata_size)
922 			return -EINVAL;
923 
924 		if (bo->metadata_size)
925 			memcpy(buffer, bo->metadata, bo->metadata_size);
926 	}
927 
928 	if (metadata_size)
929 		*metadata_size = bo->metadata_size;
930 	if (flags)
931 		*flags = bo->metadata_flags;
932 
933 	return 0;
934 }
935 
936 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
937 			   bool evict,
938 			   struct ttm_mem_reg *new_mem)
939 {
940 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
941 	struct amdgpu_bo *abo;
942 	struct ttm_mem_reg *old_mem = &bo->mem;
943 
944 	if (!amdgpu_ttm_bo_is_amdgpu_bo(bo))
945 		return;
946 
947 	abo = ttm_to_amdgpu_bo(bo);
948 	amdgpu_vm_bo_invalidate(adev, abo, evict);
949 
950 	amdgpu_bo_kunmap(abo);
951 
952 	/* remember the eviction */
953 	if (evict)
954 		atomic64_inc(&adev->num_evictions);
955 
956 	/* update statistics */
957 	if (!new_mem)
958 		return;
959 
960 	/* move_notify is called before move happens */
961 	trace_amdgpu_ttm_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
962 }
963 
964 int amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
965 {
966 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
967 	struct ttm_operation_ctx ctx = { false, false };
968 	struct amdgpu_bo *abo;
969 	unsigned long offset, size;
970 	int r;
971 
972 	if (!amdgpu_ttm_bo_is_amdgpu_bo(bo))
973 		return 0;
974 
975 	abo = ttm_to_amdgpu_bo(bo);
976 
977 	/* Remember that this BO was accessed by the CPU */
978 	abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
979 
980 	if (bo->mem.mem_type != TTM_PL_VRAM)
981 		return 0;
982 
983 	size = bo->mem.num_pages << PAGE_SHIFT;
984 	offset = bo->mem.start << PAGE_SHIFT;
985 	if ((offset + size) <= adev->mc.visible_vram_size)
986 		return 0;
987 
988 	/* Can't move a pinned BO to visible VRAM */
989 	if (abo->pin_count > 0)
990 		return -EINVAL;
991 
992 	/* hurrah the memory is not visible ! */
993 	atomic64_inc(&adev->num_vram_cpu_page_faults);
994 	amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
995 					 AMDGPU_GEM_DOMAIN_GTT);
996 
997 	/* Avoid costly evictions; only set GTT as a busy placement */
998 	abo->placement.num_busy_placement = 1;
999 	abo->placement.busy_placement = &abo->placements[1];
1000 
1001 	r = ttm_bo_validate(bo, &abo->placement, &ctx);
1002 	if (unlikely(r != 0))
1003 		return r;
1004 
1005 	offset = bo->mem.start << PAGE_SHIFT;
1006 	/* this should never happen */
1007 	if (bo->mem.mem_type == TTM_PL_VRAM &&
1008 	    (offset + size) > adev->mc.visible_vram_size)
1009 		return -EINVAL;
1010 
1011 	return 0;
1012 }
1013 
1014 /**
1015  * amdgpu_bo_fence - add fence to buffer object
1016  *
1017  * @bo: buffer object in question
1018  * @fence: fence to add
1019  * @shared: true if fence should be added shared
1020  *
1021  */
1022 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1023 		     bool shared)
1024 {
1025 	struct reservation_object *resv = bo->tbo.resv;
1026 
1027 	if (shared)
1028 		reservation_object_add_shared_fence(resv, fence);
1029 	else
1030 		reservation_object_add_excl_fence(resv, fence);
1031 }
1032 
1033 /**
1034  * amdgpu_bo_gpu_offset - return GPU offset of bo
1035  * @bo:	amdgpu object for which we query the offset
1036  *
1037  * Returns current GPU offset of the object.
1038  *
1039  * Note: object should either be pinned or reserved when calling this
1040  * function, it might be useful to add check for this for debugging.
1041  */
1042 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1043 {
1044 	WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_SYSTEM);
1045 	WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_TT &&
1046 		     !amdgpu_gtt_mgr_has_gart_addr(&bo->tbo.mem));
1047 	WARN_ON_ONCE(!ww_mutex_is_locked(&bo->tbo.resv->lock) &&
1048 		     !bo->pin_count);
1049 	WARN_ON_ONCE(bo->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET);
1050 	WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_VRAM &&
1051 		     !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1052 
1053 	return bo->tbo.offset;
1054 }
1055