xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c (revision 2eb0f624b709e78ec8e2f4c3412947703db99301)
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 static bool amdgpu_need_backup(struct amdgpu_device *adev)
42 {
43 	if (adev->flags & AMD_IS_APU)
44 		return false;
45 
46 	if (amdgpu_gpu_recovery == 0 ||
47 	    (amdgpu_gpu_recovery == -1  && !amdgpu_sriov_vf(adev)))
48 		return false;
49 
50 	return true;
51 }
52 
53 static void amdgpu_ttm_bo_destroy(struct ttm_buffer_object *tbo)
54 {
55 	struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
56 	struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
57 
58 	if (bo->kfd_bo)
59 		amdgpu_amdkfd_unreserve_system_memory_limit(bo);
60 
61 	amdgpu_bo_kunmap(bo);
62 
63 	if (bo->gem_base.import_attach)
64 		drm_prime_gem_destroy(&bo->gem_base, bo->tbo.sg);
65 	drm_gem_object_release(&bo->gem_base);
66 	amdgpu_bo_unref(&bo->parent);
67 	if (!list_empty(&bo->shadow_list)) {
68 		mutex_lock(&adev->shadow_list_lock);
69 		list_del_init(&bo->shadow_list);
70 		mutex_unlock(&adev->shadow_list_lock);
71 	}
72 	kfree(bo->metadata);
73 	kfree(bo);
74 }
75 
76 bool amdgpu_ttm_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
77 {
78 	if (bo->destroy == &amdgpu_ttm_bo_destroy)
79 		return true;
80 	return false;
81 }
82 
83 void amdgpu_ttm_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
84 {
85 	struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
86 	struct ttm_placement *placement = &abo->placement;
87 	struct ttm_place *places = abo->placements;
88 	u64 flags = abo->flags;
89 	u32 c = 0;
90 
91 	if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
92 		unsigned visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
93 
94 		places[c].fpfn = 0;
95 		places[c].lpfn = 0;
96 		places[c].flags = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
97 			TTM_PL_FLAG_VRAM;
98 
99 		if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
100 			places[c].lpfn = visible_pfn;
101 		else
102 			places[c].flags |= TTM_PL_FLAG_TOPDOWN;
103 
104 		if (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
105 			places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
106 		c++;
107 	}
108 
109 	if (domain & AMDGPU_GEM_DOMAIN_GTT) {
110 		places[c].fpfn = 0;
111 		if (flags & AMDGPU_GEM_CREATE_SHADOW)
112 			places[c].lpfn = adev->gmc.gart_size >> PAGE_SHIFT;
113 		else
114 			places[c].lpfn = 0;
115 		places[c].flags = TTM_PL_FLAG_TT;
116 		if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
117 			places[c].flags |= TTM_PL_FLAG_WC |
118 				TTM_PL_FLAG_UNCACHED;
119 		else
120 			places[c].flags |= TTM_PL_FLAG_CACHED;
121 		c++;
122 	}
123 
124 	if (domain & AMDGPU_GEM_DOMAIN_CPU) {
125 		places[c].fpfn = 0;
126 		places[c].lpfn = 0;
127 		places[c].flags = TTM_PL_FLAG_SYSTEM;
128 		if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
129 			places[c].flags |= TTM_PL_FLAG_WC |
130 				TTM_PL_FLAG_UNCACHED;
131 		else
132 			places[c].flags |= TTM_PL_FLAG_CACHED;
133 		c++;
134 	}
135 
136 	if (domain & AMDGPU_GEM_DOMAIN_GDS) {
137 		places[c].fpfn = 0;
138 		places[c].lpfn = 0;
139 		places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GDS;
140 		c++;
141 	}
142 
143 	if (domain & AMDGPU_GEM_DOMAIN_GWS) {
144 		places[c].fpfn = 0;
145 		places[c].lpfn = 0;
146 		places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GWS;
147 		c++;
148 	}
149 
150 	if (domain & AMDGPU_GEM_DOMAIN_OA) {
151 		places[c].fpfn = 0;
152 		places[c].lpfn = 0;
153 		places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_OA;
154 		c++;
155 	}
156 
157 	if (!c) {
158 		places[c].fpfn = 0;
159 		places[c].lpfn = 0;
160 		places[c].flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
161 		c++;
162 	}
163 
164 	placement->num_placement = c;
165 	placement->placement = places;
166 
167 	placement->num_busy_placement = c;
168 	placement->busy_placement = places;
169 }
170 
171 /**
172  * amdgpu_bo_create_reserved - create reserved BO for kernel use
173  *
174  * @adev: amdgpu device object
175  * @size: size for the new BO
176  * @align: alignment for the new BO
177  * @domain: where to place it
178  * @bo_ptr: used to initialize BOs in structures
179  * @gpu_addr: GPU addr of the pinned BO
180  * @cpu_addr: optional CPU address mapping
181  *
182  * Allocates and pins a BO for kernel internal use, and returns it still
183  * reserved.
184  *
185  * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
186  *
187  * Returns 0 on success, negative error code otherwise.
188  */
189 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
190 			      unsigned long size, int align,
191 			      u32 domain, struct amdgpu_bo **bo_ptr,
192 			      u64 *gpu_addr, void **cpu_addr)
193 {
194 	bool free = false;
195 	int r;
196 
197 	if (!*bo_ptr) {
198 		r = amdgpu_bo_create(adev, size, align, domain,
199 				     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
200 				     AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS,
201 				     ttm_bo_type_kernel, NULL, bo_ptr);
202 		if (r) {
203 			dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
204 				r);
205 			return r;
206 		}
207 		free = true;
208 	}
209 
210 	r = amdgpu_bo_reserve(*bo_ptr, false);
211 	if (r) {
212 		dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
213 		goto error_free;
214 	}
215 
216 	r = amdgpu_bo_pin(*bo_ptr, domain, gpu_addr);
217 	if (r) {
218 		dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
219 		goto error_unreserve;
220 	}
221 
222 	if (cpu_addr) {
223 		r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
224 		if (r) {
225 			dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
226 			goto error_unreserve;
227 		}
228 	}
229 
230 	return 0;
231 
232 error_unreserve:
233 	amdgpu_bo_unreserve(*bo_ptr);
234 
235 error_free:
236 	if (free)
237 		amdgpu_bo_unref(bo_ptr);
238 
239 	return r;
240 }
241 
242 /**
243  * amdgpu_bo_create_kernel - create BO for kernel use
244  *
245  * @adev: amdgpu device object
246  * @size: size for the new BO
247  * @align: alignment for the new BO
248  * @domain: where to place it
249  * @bo_ptr:  used to initialize BOs in structures
250  * @gpu_addr: GPU addr of the pinned BO
251  * @cpu_addr: optional CPU address mapping
252  *
253  * Allocates and pins a BO for kernel internal use.
254  *
255  * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
256  *
257  * Returns 0 on success, negative error code otherwise.
258  */
259 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
260 			    unsigned long size, int align,
261 			    u32 domain, struct amdgpu_bo **bo_ptr,
262 			    u64 *gpu_addr, void **cpu_addr)
263 {
264 	int r;
265 
266 	r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
267 				      gpu_addr, cpu_addr);
268 
269 	if (r)
270 		return r;
271 
272 	amdgpu_bo_unreserve(*bo_ptr);
273 
274 	return 0;
275 }
276 
277 /**
278  * amdgpu_bo_free_kernel - free BO for kernel use
279  *
280  * @bo: amdgpu BO to free
281  *
282  * unmaps and unpin a BO for kernel internal use.
283  */
284 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
285 			   void **cpu_addr)
286 {
287 	if (*bo == NULL)
288 		return;
289 
290 	if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
291 		if (cpu_addr)
292 			amdgpu_bo_kunmap(*bo);
293 
294 		amdgpu_bo_unpin(*bo);
295 		amdgpu_bo_unreserve(*bo);
296 	}
297 	amdgpu_bo_unref(bo);
298 
299 	if (gpu_addr)
300 		*gpu_addr = 0;
301 
302 	if (cpu_addr)
303 		*cpu_addr = NULL;
304 }
305 
306 /* Validate bo size is bit bigger then the request domain */
307 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
308 					  unsigned long size, u32 domain)
309 {
310 	struct ttm_mem_type_manager *man = NULL;
311 
312 	/*
313 	 * If GTT is part of requested domains the check must succeed to
314 	 * allow fall back to GTT
315 	 */
316 	if (domain & AMDGPU_GEM_DOMAIN_GTT) {
317 		man = &adev->mman.bdev.man[TTM_PL_TT];
318 
319 		if (size < (man->size << PAGE_SHIFT))
320 			return true;
321 		else
322 			goto fail;
323 	}
324 
325 	if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
326 		man = &adev->mman.bdev.man[TTM_PL_VRAM];
327 
328 		if (size < (man->size << PAGE_SHIFT))
329 			return true;
330 		else
331 			goto fail;
332 	}
333 
334 
335 	/* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */
336 	return true;
337 
338 fail:
339 	DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size,
340 		  man->size << PAGE_SHIFT);
341 	return false;
342 }
343 
344 static int amdgpu_bo_do_create(struct amdgpu_device *adev, unsigned long size,
345 			       int byte_align, u32 domain,
346 			       u64 flags, enum ttm_bo_type type,
347 			       struct reservation_object *resv,
348 			       struct amdgpu_bo **bo_ptr)
349 {
350 	struct ttm_operation_ctx ctx = {
351 		.interruptible = (type != ttm_bo_type_kernel),
352 		.no_wait_gpu = false,
353 		.resv = resv,
354 		.flags = TTM_OPT_FLAG_ALLOW_RES_EVICT
355 	};
356 	struct amdgpu_bo *bo;
357 	unsigned long page_align;
358 	size_t acc_size;
359 	u32 domains;
360 	int r;
361 
362 	page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
363 	size = ALIGN(size, PAGE_SIZE);
364 
365 	if (!amdgpu_bo_validate_size(adev, size, domain))
366 		return -ENOMEM;
367 
368 	*bo_ptr = NULL;
369 
370 	acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
371 				       sizeof(struct amdgpu_bo));
372 
373 	bo = kzalloc(sizeof(struct amdgpu_bo), GFP_KERNEL);
374 	if (bo == NULL)
375 		return -ENOMEM;
376 	drm_gem_private_object_init(adev->ddev, &bo->gem_base, size);
377 	INIT_LIST_HEAD(&bo->shadow_list);
378 	INIT_LIST_HEAD(&bo->va);
379 	bo->preferred_domains = domain & (AMDGPU_GEM_DOMAIN_VRAM |
380 					 AMDGPU_GEM_DOMAIN_GTT |
381 					 AMDGPU_GEM_DOMAIN_CPU |
382 					 AMDGPU_GEM_DOMAIN_GDS |
383 					 AMDGPU_GEM_DOMAIN_GWS |
384 					 AMDGPU_GEM_DOMAIN_OA);
385 	bo->allowed_domains = bo->preferred_domains;
386 	if (type != ttm_bo_type_kernel &&
387 	    bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
388 		bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
389 
390 	bo->flags = flags;
391 
392 #ifdef CONFIG_X86_32
393 	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
394 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
395 	 */
396 	bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
397 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
398 	/* Don't try to enable write-combining when it can't work, or things
399 	 * may be slow
400 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
401 	 */
402 
403 #ifndef CONFIG_COMPILE_TEST
404 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
405 	 thanks to write-combining
406 #endif
407 
408 	if (bo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
409 		DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
410 			      "better performance thanks to write-combining\n");
411 	bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
412 #else
413 	/* For architectures that don't support WC memory,
414 	 * mask out the WC flag from the BO
415 	 */
416 	if (!drm_arch_can_wc_memory())
417 		bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
418 #endif
419 
420 	bo->tbo.bdev = &adev->mman.bdev;
421 	domains = bo->preferred_domains;
422 retry:
423 	amdgpu_ttm_placement_from_domain(bo, domains);
424 	r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, type,
425 				 &bo->placement, page_align, &ctx, acc_size,
426 				 NULL, resv, &amdgpu_ttm_bo_destroy);
427 
428 	if (unlikely(r && r != -ERESTARTSYS)) {
429 		if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) {
430 			bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
431 			goto retry;
432 		} else if (domains != bo->preferred_domains) {
433 			domains = bo->allowed_domains;
434 			goto retry;
435 		}
436 	}
437 	if (unlikely(r))
438 		return r;
439 
440 	if (adev->gmc.visible_vram_size < adev->gmc.real_vram_size &&
441 	    bo->tbo.mem.mem_type == TTM_PL_VRAM &&
442 	    bo->tbo.mem.start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
443 		amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
444 					     ctx.bytes_moved);
445 	else
446 		amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
447 
448 	if (type == ttm_bo_type_kernel)
449 		bo->tbo.priority = 1;
450 
451 	if (flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
452 	    bo->tbo.mem.placement & TTM_PL_FLAG_VRAM) {
453 		struct dma_fence *fence;
454 
455 		r = amdgpu_fill_buffer(bo, 0, bo->tbo.resv, &fence);
456 		if (unlikely(r))
457 			goto fail_unreserve;
458 
459 		amdgpu_bo_fence(bo, fence, false);
460 		dma_fence_put(bo->tbo.moving);
461 		bo->tbo.moving = dma_fence_get(fence);
462 		dma_fence_put(fence);
463 	}
464 	if (!resv)
465 		amdgpu_bo_unreserve(bo);
466 	*bo_ptr = bo;
467 
468 	trace_amdgpu_bo_create(bo);
469 
470 	/* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
471 	if (type == ttm_bo_type_device)
472 		bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
473 
474 	return 0;
475 
476 fail_unreserve:
477 	if (!resv)
478 		ww_mutex_unlock(&bo->tbo.resv->lock);
479 	amdgpu_bo_unref(&bo);
480 	return r;
481 }
482 
483 static int amdgpu_bo_create_shadow(struct amdgpu_device *adev,
484 				   unsigned long size, int byte_align,
485 				   struct amdgpu_bo *bo)
486 {
487 	int r;
488 
489 	if (bo->shadow)
490 		return 0;
491 
492 	r = amdgpu_bo_do_create(adev, size, byte_align, AMDGPU_GEM_DOMAIN_GTT,
493 				AMDGPU_GEM_CREATE_CPU_GTT_USWC |
494 				AMDGPU_GEM_CREATE_SHADOW,
495 				ttm_bo_type_kernel,
496 				bo->tbo.resv, &bo->shadow);
497 	if (!r) {
498 		bo->shadow->parent = amdgpu_bo_ref(bo);
499 		mutex_lock(&adev->shadow_list_lock);
500 		list_add_tail(&bo->shadow_list, &adev->shadow_list);
501 		mutex_unlock(&adev->shadow_list_lock);
502 	}
503 
504 	return r;
505 }
506 
507 int amdgpu_bo_create(struct amdgpu_device *adev, unsigned long size,
508 		     int byte_align, u32 domain,
509 		     u64 flags, enum ttm_bo_type type,
510 		     struct reservation_object *resv,
511 		     struct amdgpu_bo **bo_ptr)
512 {
513 	uint64_t parent_flags = flags & ~AMDGPU_GEM_CREATE_SHADOW;
514 	int r;
515 
516 	r = amdgpu_bo_do_create(adev, size, byte_align, domain,
517 				parent_flags, type, resv, bo_ptr);
518 	if (r)
519 		return r;
520 
521 	if ((flags & AMDGPU_GEM_CREATE_SHADOW) && amdgpu_need_backup(adev)) {
522 		if (!resv)
523 			WARN_ON(reservation_object_lock((*bo_ptr)->tbo.resv,
524 							NULL));
525 
526 		r = amdgpu_bo_create_shadow(adev, size, byte_align, (*bo_ptr));
527 
528 		if (!resv)
529 			reservation_object_unlock((*bo_ptr)->tbo.resv);
530 
531 		if (r)
532 			amdgpu_bo_unref(bo_ptr);
533 	}
534 
535 	return r;
536 }
537 
538 int amdgpu_bo_backup_to_shadow(struct amdgpu_device *adev,
539 			       struct amdgpu_ring *ring,
540 			       struct amdgpu_bo *bo,
541 			       struct reservation_object *resv,
542 			       struct dma_fence **fence,
543 			       bool direct)
544 
545 {
546 	struct amdgpu_bo *shadow = bo->shadow;
547 	uint64_t bo_addr, shadow_addr;
548 	int r;
549 
550 	if (!shadow)
551 		return -EINVAL;
552 
553 	bo_addr = amdgpu_bo_gpu_offset(bo);
554 	shadow_addr = amdgpu_bo_gpu_offset(bo->shadow);
555 
556 	r = reservation_object_reserve_shared(bo->tbo.resv);
557 	if (r)
558 		goto err;
559 
560 	r = amdgpu_copy_buffer(ring, bo_addr, shadow_addr,
561 			       amdgpu_bo_size(bo), resv, fence,
562 			       direct, false);
563 	if (!r)
564 		amdgpu_bo_fence(bo, *fence, true);
565 
566 err:
567 	return r;
568 }
569 
570 int amdgpu_bo_validate(struct amdgpu_bo *bo)
571 {
572 	struct ttm_operation_ctx ctx = { false, false };
573 	uint32_t domain;
574 	int r;
575 
576 	if (bo->pin_count)
577 		return 0;
578 
579 	domain = bo->preferred_domains;
580 
581 retry:
582 	amdgpu_ttm_placement_from_domain(bo, domain);
583 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
584 	if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
585 		domain = bo->allowed_domains;
586 		goto retry;
587 	}
588 
589 	return r;
590 }
591 
592 int amdgpu_bo_restore_from_shadow(struct amdgpu_device *adev,
593 				  struct amdgpu_ring *ring,
594 				  struct amdgpu_bo *bo,
595 				  struct reservation_object *resv,
596 				  struct dma_fence **fence,
597 				  bool direct)
598 
599 {
600 	struct amdgpu_bo *shadow = bo->shadow;
601 	uint64_t bo_addr, shadow_addr;
602 	int r;
603 
604 	if (!shadow)
605 		return -EINVAL;
606 
607 	bo_addr = amdgpu_bo_gpu_offset(bo);
608 	shadow_addr = amdgpu_bo_gpu_offset(bo->shadow);
609 
610 	r = reservation_object_reserve_shared(bo->tbo.resv);
611 	if (r)
612 		goto err;
613 
614 	r = amdgpu_copy_buffer(ring, shadow_addr, bo_addr,
615 			       amdgpu_bo_size(bo), resv, fence,
616 			       direct, false);
617 	if (!r)
618 		amdgpu_bo_fence(bo, *fence, true);
619 
620 err:
621 	return r;
622 }
623 
624 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
625 {
626 	void *kptr;
627 	long r;
628 
629 	if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
630 		return -EPERM;
631 
632 	kptr = amdgpu_bo_kptr(bo);
633 	if (kptr) {
634 		if (ptr)
635 			*ptr = kptr;
636 		return 0;
637 	}
638 
639 	r = reservation_object_wait_timeout_rcu(bo->tbo.resv, false, false,
640 						MAX_SCHEDULE_TIMEOUT);
641 	if (r < 0)
642 		return r;
643 
644 	r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
645 	if (r)
646 		return r;
647 
648 	if (ptr)
649 		*ptr = amdgpu_bo_kptr(bo);
650 
651 	return 0;
652 }
653 
654 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
655 {
656 	bool is_iomem;
657 
658 	return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
659 }
660 
661 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
662 {
663 	if (bo->kmap.bo)
664 		ttm_bo_kunmap(&bo->kmap);
665 }
666 
667 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
668 {
669 	if (bo == NULL)
670 		return NULL;
671 
672 	ttm_bo_reference(&bo->tbo);
673 	return bo;
674 }
675 
676 void amdgpu_bo_unref(struct amdgpu_bo **bo)
677 {
678 	struct ttm_buffer_object *tbo;
679 
680 	if ((*bo) == NULL)
681 		return;
682 
683 	tbo = &((*bo)->tbo);
684 	ttm_bo_unref(&tbo);
685 	if (tbo == NULL)
686 		*bo = NULL;
687 }
688 
689 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
690 			     u64 min_offset, u64 max_offset,
691 			     u64 *gpu_addr)
692 {
693 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
694 	struct ttm_operation_ctx ctx = { false, false };
695 	int r, i;
696 
697 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
698 		return -EPERM;
699 
700 	if (WARN_ON_ONCE(min_offset > max_offset))
701 		return -EINVAL;
702 
703 	/* A shared bo cannot be migrated to VRAM */
704 	if (bo->prime_shared_count && (domain == AMDGPU_GEM_DOMAIN_VRAM))
705 		return -EINVAL;
706 
707 	if (bo->pin_count) {
708 		uint32_t mem_type = bo->tbo.mem.mem_type;
709 
710 		if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
711 			return -EINVAL;
712 
713 		bo->pin_count++;
714 		if (gpu_addr)
715 			*gpu_addr = amdgpu_bo_gpu_offset(bo);
716 
717 		if (max_offset != 0) {
718 			u64 domain_start = bo->tbo.bdev->man[mem_type].gpu_offset;
719 			WARN_ON_ONCE(max_offset <
720 				     (amdgpu_bo_gpu_offset(bo) - domain_start));
721 		}
722 
723 		return 0;
724 	}
725 
726 	bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
727 	/* force to pin into visible video ram */
728 	if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
729 		bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
730 	amdgpu_ttm_placement_from_domain(bo, domain);
731 	for (i = 0; i < bo->placement.num_placement; i++) {
732 		unsigned fpfn, lpfn;
733 
734 		fpfn = min_offset >> PAGE_SHIFT;
735 		lpfn = max_offset >> PAGE_SHIFT;
736 
737 		if (fpfn > bo->placements[i].fpfn)
738 			bo->placements[i].fpfn = fpfn;
739 		if (!bo->placements[i].lpfn ||
740 		    (lpfn && lpfn < bo->placements[i].lpfn))
741 			bo->placements[i].lpfn = lpfn;
742 		bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
743 	}
744 
745 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
746 	if (unlikely(r)) {
747 		dev_err(adev->dev, "%p pin failed\n", bo);
748 		goto error;
749 	}
750 
751 	r = amdgpu_ttm_alloc_gart(&bo->tbo);
752 	if (unlikely(r)) {
753 		dev_err(adev->dev, "%p bind failed\n", bo);
754 		goto error;
755 	}
756 
757 	bo->pin_count = 1;
758 	if (gpu_addr != NULL)
759 		*gpu_addr = amdgpu_bo_gpu_offset(bo);
760 
761 	domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
762 	if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
763 		adev->vram_pin_size += amdgpu_bo_size(bo);
764 		if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
765 			adev->invisible_pin_size += amdgpu_bo_size(bo);
766 	} else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
767 		adev->gart_pin_size += amdgpu_bo_size(bo);
768 	}
769 
770 error:
771 	return r;
772 }
773 
774 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain, u64 *gpu_addr)
775 {
776 	return amdgpu_bo_pin_restricted(bo, domain, 0, 0, gpu_addr);
777 }
778 
779 int amdgpu_bo_unpin(struct amdgpu_bo *bo)
780 {
781 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
782 	struct ttm_operation_ctx ctx = { false, false };
783 	int r, i;
784 
785 	if (!bo->pin_count) {
786 		dev_warn(adev->dev, "%p unpin not necessary\n", bo);
787 		return 0;
788 	}
789 	bo->pin_count--;
790 	if (bo->pin_count)
791 		return 0;
792 	for (i = 0; i < bo->placement.num_placement; i++) {
793 		bo->placements[i].lpfn = 0;
794 		bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
795 	}
796 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
797 	if (unlikely(r)) {
798 		dev_err(adev->dev, "%p validate failed for unpin\n", bo);
799 		goto error;
800 	}
801 
802 	if (bo->tbo.mem.mem_type == TTM_PL_VRAM) {
803 		adev->vram_pin_size -= amdgpu_bo_size(bo);
804 		if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
805 			adev->invisible_pin_size -= amdgpu_bo_size(bo);
806 	} else if (bo->tbo.mem.mem_type == TTM_PL_TT) {
807 		adev->gart_pin_size -= amdgpu_bo_size(bo);
808 	}
809 
810 error:
811 	return r;
812 }
813 
814 int amdgpu_bo_evict_vram(struct amdgpu_device *adev)
815 {
816 	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
817 	if (0 && (adev->flags & AMD_IS_APU)) {
818 		/* Useless to evict on IGP chips */
819 		return 0;
820 	}
821 	return ttm_bo_evict_mm(&adev->mman.bdev, TTM_PL_VRAM);
822 }
823 
824 static const char *amdgpu_vram_names[] = {
825 	"UNKNOWN",
826 	"GDDR1",
827 	"DDR2",
828 	"GDDR3",
829 	"GDDR4",
830 	"GDDR5",
831 	"HBM",
832 	"DDR3",
833 	"DDR4",
834 };
835 
836 int amdgpu_bo_init(struct amdgpu_device *adev)
837 {
838 	/* reserve PAT memory space to WC for VRAM */
839 	arch_io_reserve_memtype_wc(adev->gmc.aper_base,
840 				   adev->gmc.aper_size);
841 
842 	/* Add an MTRR for the VRAM */
843 	adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
844 					      adev->gmc.aper_size);
845 	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
846 		 adev->gmc.mc_vram_size >> 20,
847 		 (unsigned long long)adev->gmc.aper_size >> 20);
848 	DRM_INFO("RAM width %dbits %s\n",
849 		 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
850 	return amdgpu_ttm_init(adev);
851 }
852 
853 void amdgpu_bo_fini(struct amdgpu_device *adev)
854 {
855 	amdgpu_ttm_fini(adev);
856 	arch_phys_wc_del(adev->gmc.vram_mtrr);
857 	arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
858 }
859 
860 int amdgpu_bo_fbdev_mmap(struct amdgpu_bo *bo,
861 			     struct vm_area_struct *vma)
862 {
863 	return ttm_fbdev_mmap(vma, &bo->tbo);
864 }
865 
866 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
867 {
868 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
869 
870 	if (adev->family <= AMDGPU_FAMILY_CZ &&
871 	    AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
872 		return -EINVAL;
873 
874 	bo->tiling_flags = tiling_flags;
875 	return 0;
876 }
877 
878 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
879 {
880 	lockdep_assert_held(&bo->tbo.resv->lock.base);
881 
882 	if (tiling_flags)
883 		*tiling_flags = bo->tiling_flags;
884 }
885 
886 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
887 			    uint32_t metadata_size, uint64_t flags)
888 {
889 	void *buffer;
890 
891 	if (!metadata_size) {
892 		if (bo->metadata_size) {
893 			kfree(bo->metadata);
894 			bo->metadata = NULL;
895 			bo->metadata_size = 0;
896 		}
897 		return 0;
898 	}
899 
900 	if (metadata == NULL)
901 		return -EINVAL;
902 
903 	buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
904 	if (buffer == NULL)
905 		return -ENOMEM;
906 
907 	kfree(bo->metadata);
908 	bo->metadata_flags = flags;
909 	bo->metadata = buffer;
910 	bo->metadata_size = metadata_size;
911 
912 	return 0;
913 }
914 
915 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
916 			   size_t buffer_size, uint32_t *metadata_size,
917 			   uint64_t *flags)
918 {
919 	if (!buffer && !metadata_size)
920 		return -EINVAL;
921 
922 	if (buffer) {
923 		if (buffer_size < bo->metadata_size)
924 			return -EINVAL;
925 
926 		if (bo->metadata_size)
927 			memcpy(buffer, bo->metadata, bo->metadata_size);
928 	}
929 
930 	if (metadata_size)
931 		*metadata_size = bo->metadata_size;
932 	if (flags)
933 		*flags = bo->metadata_flags;
934 
935 	return 0;
936 }
937 
938 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
939 			   bool evict,
940 			   struct ttm_mem_reg *new_mem)
941 {
942 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
943 	struct amdgpu_bo *abo;
944 	struct ttm_mem_reg *old_mem = &bo->mem;
945 
946 	if (!amdgpu_ttm_bo_is_amdgpu_bo(bo))
947 		return;
948 
949 	abo = ttm_to_amdgpu_bo(bo);
950 	amdgpu_vm_bo_invalidate(adev, abo, evict);
951 
952 	amdgpu_bo_kunmap(abo);
953 
954 	/* remember the eviction */
955 	if (evict)
956 		atomic64_inc(&adev->num_evictions);
957 
958 	/* update statistics */
959 	if (!new_mem)
960 		return;
961 
962 	/* move_notify is called before move happens */
963 	trace_amdgpu_ttm_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
964 }
965 
966 int amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
967 {
968 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
969 	struct ttm_operation_ctx ctx = { false, false };
970 	struct amdgpu_bo *abo;
971 	unsigned long offset, size;
972 	int r;
973 
974 	if (!amdgpu_ttm_bo_is_amdgpu_bo(bo))
975 		return 0;
976 
977 	abo = ttm_to_amdgpu_bo(bo);
978 
979 	/* Remember that this BO was accessed by the CPU */
980 	abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
981 
982 	if (bo->mem.mem_type != TTM_PL_VRAM)
983 		return 0;
984 
985 	size = bo->mem.num_pages << PAGE_SHIFT;
986 	offset = bo->mem.start << PAGE_SHIFT;
987 	if ((offset + size) <= adev->gmc.visible_vram_size)
988 		return 0;
989 
990 	/* Can't move a pinned BO to visible VRAM */
991 	if (abo->pin_count > 0)
992 		return -EINVAL;
993 
994 	/* hurrah the memory is not visible ! */
995 	atomic64_inc(&adev->num_vram_cpu_page_faults);
996 	amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
997 					 AMDGPU_GEM_DOMAIN_GTT);
998 
999 	/* Avoid costly evictions; only set GTT as a busy placement */
1000 	abo->placement.num_busy_placement = 1;
1001 	abo->placement.busy_placement = &abo->placements[1];
1002 
1003 	r = ttm_bo_validate(bo, &abo->placement, &ctx);
1004 	if (unlikely(r != 0))
1005 		return r;
1006 
1007 	offset = bo->mem.start << PAGE_SHIFT;
1008 	/* this should never happen */
1009 	if (bo->mem.mem_type == TTM_PL_VRAM &&
1010 	    (offset + size) > adev->gmc.visible_vram_size)
1011 		return -EINVAL;
1012 
1013 	return 0;
1014 }
1015 
1016 /**
1017  * amdgpu_bo_fence - add fence to buffer object
1018  *
1019  * @bo: buffer object in question
1020  * @fence: fence to add
1021  * @shared: true if fence should be added shared
1022  *
1023  */
1024 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1025 		     bool shared)
1026 {
1027 	struct reservation_object *resv = bo->tbo.resv;
1028 
1029 	if (shared)
1030 		reservation_object_add_shared_fence(resv, fence);
1031 	else
1032 		reservation_object_add_excl_fence(resv, fence);
1033 }
1034 
1035 /**
1036  * amdgpu_bo_gpu_offset - return GPU offset of bo
1037  * @bo:	amdgpu object for which we query the offset
1038  *
1039  * Returns current GPU offset of the object.
1040  *
1041  * Note: object should either be pinned or reserved when calling this
1042  * function, it might be useful to add check for this for debugging.
1043  */
1044 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1045 {
1046 	WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_SYSTEM);
1047 	WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_TT &&
1048 		     !amdgpu_gtt_mgr_has_gart_addr(&bo->tbo.mem));
1049 	WARN_ON_ONCE(!ww_mutex_is_locked(&bo->tbo.resv->lock) &&
1050 		     !bo->pin_count);
1051 	WARN_ON_ONCE(bo->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET);
1052 	WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_VRAM &&
1053 		     !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1054 
1055 	return bo->tbo.offset;
1056 }
1057