xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c (revision f1bf3bc6cb932b2094c71d5b45cf4e56b8450852)
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 <linux/dma-buf.h>
35 
36 #include <drm/drm_drv.h>
37 #include <drm/amdgpu_drm.h>
38 #include <drm/drm_cache.h>
39 #include "amdgpu.h"
40 #include "amdgpu_trace.h"
41 #include "amdgpu_amdkfd.h"
42 
43 /**
44  * DOC: amdgpu_object
45  *
46  * This defines the interfaces to operate on an &amdgpu_bo buffer object which
47  * represents memory used by driver (VRAM, system memory, etc.). The driver
48  * provides DRM/GEM APIs to userspace. DRM/GEM APIs then use these interfaces
49  * to create/destroy/set buffer object which are then managed by the kernel TTM
50  * memory manager.
51  * The interfaces are also used internally by kernel clients, including gfx,
52  * uvd, etc. for kernel managed allocations used by the GPU.
53  *
54  */
55 
56 static void amdgpu_bo_destroy(struct ttm_buffer_object *tbo)
57 {
58 	struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
59 
60 	amdgpu_bo_kunmap(bo);
61 
62 	if (bo->tbo.base.import_attach)
63 		drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
64 	drm_gem_object_release(&bo->tbo.base);
65 	amdgpu_bo_unref(&bo->parent);
66 	kvfree(bo);
67 }
68 
69 static void amdgpu_bo_user_destroy(struct ttm_buffer_object *tbo)
70 {
71 	struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
72 	struct amdgpu_bo_user *ubo;
73 
74 	ubo = to_amdgpu_bo_user(bo);
75 	kfree(ubo->metadata);
76 	amdgpu_bo_destroy(tbo);
77 }
78 
79 static void amdgpu_bo_vm_destroy(struct ttm_buffer_object *tbo)
80 {
81 	struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
82 	struct amdgpu_bo *shadow_bo = ttm_to_amdgpu_bo(tbo), *bo;
83 	struct amdgpu_bo_vm *vmbo;
84 
85 	bo = shadow_bo->parent;
86 	vmbo = to_amdgpu_bo_vm(bo);
87 	/* in case amdgpu_device_recover_vram got NULL of bo->parent */
88 	if (!list_empty(&vmbo->shadow_list)) {
89 		mutex_lock(&adev->shadow_list_lock);
90 		list_del_init(&vmbo->shadow_list);
91 		mutex_unlock(&adev->shadow_list_lock);
92 	}
93 
94 	amdgpu_bo_destroy(tbo);
95 }
96 
97 /**
98  * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
99  * @bo: buffer object to be checked
100  *
101  * Uses destroy function associated with the object to determine if this is
102  * an &amdgpu_bo.
103  *
104  * Returns:
105  * true if the object belongs to &amdgpu_bo, false if not.
106  */
107 bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
108 {
109 	if (bo->destroy == &amdgpu_bo_destroy ||
110 	    bo->destroy == &amdgpu_bo_user_destroy ||
111 	    bo->destroy == &amdgpu_bo_vm_destroy)
112 		return true;
113 
114 	return false;
115 }
116 
117 /**
118  * amdgpu_bo_placement_from_domain - set buffer's placement
119  * @abo: &amdgpu_bo buffer object whose placement is to be set
120  * @domain: requested domain
121  *
122  * Sets buffer's placement according to requested domain and the buffer's
123  * flags.
124  */
125 void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
126 {
127 	struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
128 	struct ttm_placement *placement = &abo->placement;
129 	struct ttm_place *places = abo->placements;
130 	u64 flags = abo->flags;
131 	u32 c = 0;
132 
133 	if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
134 		unsigned int visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
135 		int8_t mem_id = KFD_XCP_MEM_ID(adev, abo->xcp_id);
136 
137 		if (adev->gmc.mem_partitions && mem_id >= 0) {
138 			places[c].fpfn = adev->gmc.mem_partitions[mem_id].range.fpfn;
139 			/*
140 			 * memory partition range lpfn is inclusive start + size - 1
141 			 * TTM place lpfn is exclusive start + size
142 			 */
143 			places[c].lpfn = adev->gmc.mem_partitions[mem_id].range.lpfn + 1;
144 		} else {
145 			places[c].fpfn = 0;
146 			places[c].lpfn = 0;
147 		}
148 		places[c].mem_type = TTM_PL_VRAM;
149 		places[c].flags = 0;
150 
151 		if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
152 			places[c].lpfn = min_not_zero(places[c].lpfn, visible_pfn);
153 		else
154 			places[c].flags |= TTM_PL_FLAG_TOPDOWN;
155 
156 		if (abo->tbo.type == ttm_bo_type_kernel &&
157 		    flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
158 			places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
159 
160 		c++;
161 	}
162 
163 	if (domain & AMDGPU_GEM_DOMAIN_DOORBELL) {
164 		places[c].fpfn = 0;
165 		places[c].lpfn = 0;
166 		places[c].mem_type = AMDGPU_PL_DOORBELL;
167 		places[c].flags = 0;
168 		c++;
169 	}
170 
171 	if (domain & AMDGPU_GEM_DOMAIN_GTT) {
172 		places[c].fpfn = 0;
173 		places[c].lpfn = 0;
174 		places[c].mem_type =
175 			abo->flags & AMDGPU_GEM_CREATE_PREEMPTIBLE ?
176 			AMDGPU_PL_PREEMPT : TTM_PL_TT;
177 		places[c].flags = 0;
178 		/*
179 		 * When GTT is just an alternative to VRAM make sure that we
180 		 * only use it as fallback and still try to fill up VRAM first.
181 		 */
182 		if (domain & abo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM)
183 			places[c].flags |= TTM_PL_FLAG_FALLBACK;
184 		c++;
185 	}
186 
187 	if (domain & AMDGPU_GEM_DOMAIN_CPU) {
188 		places[c].fpfn = 0;
189 		places[c].lpfn = 0;
190 		places[c].mem_type = TTM_PL_SYSTEM;
191 		places[c].flags = 0;
192 		c++;
193 	}
194 
195 	if (domain & AMDGPU_GEM_DOMAIN_GDS) {
196 		places[c].fpfn = 0;
197 		places[c].lpfn = 0;
198 		places[c].mem_type = AMDGPU_PL_GDS;
199 		places[c].flags = 0;
200 		c++;
201 	}
202 
203 	if (domain & AMDGPU_GEM_DOMAIN_GWS) {
204 		places[c].fpfn = 0;
205 		places[c].lpfn = 0;
206 		places[c].mem_type = AMDGPU_PL_GWS;
207 		places[c].flags = 0;
208 		c++;
209 	}
210 
211 	if (domain & AMDGPU_GEM_DOMAIN_OA) {
212 		places[c].fpfn = 0;
213 		places[c].lpfn = 0;
214 		places[c].mem_type = AMDGPU_PL_OA;
215 		places[c].flags = 0;
216 		c++;
217 	}
218 
219 	if (!c) {
220 		places[c].fpfn = 0;
221 		places[c].lpfn = 0;
222 		places[c].mem_type = TTM_PL_SYSTEM;
223 		places[c].flags = 0;
224 		c++;
225 	}
226 
227 	BUG_ON(c > AMDGPU_BO_MAX_PLACEMENTS);
228 
229 	placement->num_placement = c;
230 	placement->placement = places;
231 }
232 
233 /**
234  * amdgpu_bo_create_reserved - create reserved BO for kernel use
235  *
236  * @adev: amdgpu device object
237  * @size: size for the new BO
238  * @align: alignment for the new BO
239  * @domain: where to place it
240  * @bo_ptr: used to initialize BOs in structures
241  * @gpu_addr: GPU addr of the pinned BO
242  * @cpu_addr: optional CPU address mapping
243  *
244  * Allocates and pins a BO for kernel internal use, and returns it still
245  * reserved.
246  *
247  * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
248  *
249  * Returns:
250  * 0 on success, negative error code otherwise.
251  */
252 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
253 			      unsigned long size, int align,
254 			      u32 domain, struct amdgpu_bo **bo_ptr,
255 			      u64 *gpu_addr, void **cpu_addr)
256 {
257 	struct amdgpu_bo_param bp;
258 	bool free = false;
259 	int r;
260 
261 	if (!size) {
262 		amdgpu_bo_unref(bo_ptr);
263 		return 0;
264 	}
265 
266 	memset(&bp, 0, sizeof(bp));
267 	bp.size = size;
268 	bp.byte_align = align;
269 	bp.domain = domain;
270 	bp.flags = cpu_addr ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
271 		: AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
272 	bp.flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
273 	bp.type = ttm_bo_type_kernel;
274 	bp.resv = NULL;
275 	bp.bo_ptr_size = sizeof(struct amdgpu_bo);
276 
277 	if (!*bo_ptr) {
278 		r = amdgpu_bo_create(adev, &bp, bo_ptr);
279 		if (r) {
280 			dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
281 				r);
282 			return r;
283 		}
284 		free = true;
285 	}
286 
287 	r = amdgpu_bo_reserve(*bo_ptr, false);
288 	if (r) {
289 		dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
290 		goto error_free;
291 	}
292 
293 	r = amdgpu_bo_pin(*bo_ptr, domain);
294 	if (r) {
295 		dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
296 		goto error_unreserve;
297 	}
298 
299 	r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
300 	if (r) {
301 		dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
302 		goto error_unpin;
303 	}
304 
305 	if (gpu_addr)
306 		*gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
307 
308 	if (cpu_addr) {
309 		r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
310 		if (r) {
311 			dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
312 			goto error_unpin;
313 		}
314 	}
315 
316 	return 0;
317 
318 error_unpin:
319 	amdgpu_bo_unpin(*bo_ptr);
320 error_unreserve:
321 	amdgpu_bo_unreserve(*bo_ptr);
322 
323 error_free:
324 	if (free)
325 		amdgpu_bo_unref(bo_ptr);
326 
327 	return r;
328 }
329 
330 /**
331  * amdgpu_bo_create_kernel - create BO for kernel use
332  *
333  * @adev: amdgpu device object
334  * @size: size for the new BO
335  * @align: alignment for the new BO
336  * @domain: where to place it
337  * @bo_ptr:  used to initialize BOs in structures
338  * @gpu_addr: GPU addr of the pinned BO
339  * @cpu_addr: optional CPU address mapping
340  *
341  * Allocates and pins a BO for kernel internal use.
342  *
343  * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
344  *
345  * Returns:
346  * 0 on success, negative error code otherwise.
347  */
348 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
349 			    unsigned long size, int align,
350 			    u32 domain, struct amdgpu_bo **bo_ptr,
351 			    u64 *gpu_addr, void **cpu_addr)
352 {
353 	int r;
354 
355 	r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
356 				      gpu_addr, cpu_addr);
357 
358 	if (r)
359 		return r;
360 
361 	if (*bo_ptr)
362 		amdgpu_bo_unreserve(*bo_ptr);
363 
364 	return 0;
365 }
366 
367 /**
368  * amdgpu_bo_create_kernel_at - create BO for kernel use at specific location
369  *
370  * @adev: amdgpu device object
371  * @offset: offset of the BO
372  * @size: size of the BO
373  * @bo_ptr:  used to initialize BOs in structures
374  * @cpu_addr: optional CPU address mapping
375  *
376  * Creates a kernel BO at a specific offset in VRAM.
377  *
378  * Returns:
379  * 0 on success, negative error code otherwise.
380  */
381 int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev,
382 			       uint64_t offset, uint64_t size,
383 			       struct amdgpu_bo **bo_ptr, void **cpu_addr)
384 {
385 	struct ttm_operation_ctx ctx = { false, false };
386 	unsigned int i;
387 	int r;
388 
389 	offset &= PAGE_MASK;
390 	size = ALIGN(size, PAGE_SIZE);
391 
392 	r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE,
393 				      AMDGPU_GEM_DOMAIN_VRAM, bo_ptr, NULL,
394 				      cpu_addr);
395 	if (r)
396 		return r;
397 
398 	if ((*bo_ptr) == NULL)
399 		return 0;
400 
401 	/*
402 	 * Remove the original mem node and create a new one at the request
403 	 * position.
404 	 */
405 	if (cpu_addr)
406 		amdgpu_bo_kunmap(*bo_ptr);
407 
408 	ttm_resource_free(&(*bo_ptr)->tbo, &(*bo_ptr)->tbo.resource);
409 
410 	for (i = 0; i < (*bo_ptr)->placement.num_placement; ++i) {
411 		(*bo_ptr)->placements[i].fpfn = offset >> PAGE_SHIFT;
412 		(*bo_ptr)->placements[i].lpfn = (offset + size) >> PAGE_SHIFT;
413 	}
414 	r = ttm_bo_mem_space(&(*bo_ptr)->tbo, &(*bo_ptr)->placement,
415 			     &(*bo_ptr)->tbo.resource, &ctx);
416 	if (r)
417 		goto error;
418 
419 	if (cpu_addr) {
420 		r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
421 		if (r)
422 			goto error;
423 	}
424 
425 	amdgpu_bo_unreserve(*bo_ptr);
426 	return 0;
427 
428 error:
429 	amdgpu_bo_unreserve(*bo_ptr);
430 	amdgpu_bo_unref(bo_ptr);
431 	return r;
432 }
433 
434 /**
435  * amdgpu_bo_free_kernel - free BO for kernel use
436  *
437  * @bo: amdgpu BO to free
438  * @gpu_addr: pointer to where the BO's GPU memory space address was stored
439  * @cpu_addr: pointer to where the BO's CPU memory space address was stored
440  *
441  * unmaps and unpin a BO for kernel internal use.
442  */
443 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
444 			   void **cpu_addr)
445 {
446 	if (*bo == NULL)
447 		return;
448 
449 	WARN_ON(amdgpu_ttm_adev((*bo)->tbo.bdev)->in_suspend);
450 
451 	if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
452 		if (cpu_addr)
453 			amdgpu_bo_kunmap(*bo);
454 
455 		amdgpu_bo_unpin(*bo);
456 		amdgpu_bo_unreserve(*bo);
457 	}
458 	amdgpu_bo_unref(bo);
459 
460 	if (gpu_addr)
461 		*gpu_addr = 0;
462 
463 	if (cpu_addr)
464 		*cpu_addr = NULL;
465 }
466 
467 /* Validate bo size is bit bigger than the request domain */
468 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
469 					  unsigned long size, u32 domain)
470 {
471 	struct ttm_resource_manager *man = NULL;
472 
473 	/*
474 	 * If GTT is part of requested domains the check must succeed to
475 	 * allow fall back to GTT.
476 	 */
477 	if (domain & AMDGPU_GEM_DOMAIN_GTT)
478 		man = ttm_manager_type(&adev->mman.bdev, TTM_PL_TT);
479 	else if (domain & AMDGPU_GEM_DOMAIN_VRAM)
480 		man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
481 	else
482 		return true;
483 
484 	if (!man) {
485 		if (domain & AMDGPU_GEM_DOMAIN_GTT)
486 			WARN_ON_ONCE("GTT domain requested but GTT mem manager uninitialized");
487 		return false;
488 	}
489 
490 	/* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU, _DOMAIN_DOORBELL */
491 	if (size < man->size)
492 		return true;
493 
494 	DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size, man->size);
495 	return false;
496 }
497 
498 bool amdgpu_bo_support_uswc(u64 bo_flags)
499 {
500 
501 #ifdef CONFIG_X86_32
502 	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
503 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
504 	 */
505 	return false;
506 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
507 	/* Don't try to enable write-combining when it can't work, or things
508 	 * may be slow
509 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
510 	 */
511 
512 #ifndef CONFIG_COMPILE_TEST
513 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
514 	 thanks to write-combining
515 #endif
516 
517 	if (bo_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
518 		DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
519 			      "better performance thanks to write-combining\n");
520 	return false;
521 #else
522 	/* For architectures that don't support WC memory,
523 	 * mask out the WC flag from the BO
524 	 */
525 	if (!drm_arch_can_wc_memory())
526 		return false;
527 
528 	return true;
529 #endif
530 }
531 
532 /**
533  * amdgpu_bo_create - create an &amdgpu_bo buffer object
534  * @adev: amdgpu device object
535  * @bp: parameters to be used for the buffer object
536  * @bo_ptr: pointer to the buffer object pointer
537  *
538  * Creates an &amdgpu_bo buffer object.
539  *
540  * Returns:
541  * 0 for success or a negative error code on failure.
542  */
543 int amdgpu_bo_create(struct amdgpu_device *adev,
544 			       struct amdgpu_bo_param *bp,
545 			       struct amdgpu_bo **bo_ptr)
546 {
547 	struct ttm_operation_ctx ctx = {
548 		.interruptible = (bp->type != ttm_bo_type_kernel),
549 		.no_wait_gpu = bp->no_wait_gpu,
550 		/* We opt to avoid OOM on system pages allocations */
551 		.gfp_retry_mayfail = true,
552 		.allow_res_evict = bp->type != ttm_bo_type_kernel,
553 		.resv = bp->resv
554 	};
555 	struct amdgpu_bo *bo;
556 	unsigned long page_align, size = bp->size;
557 	int r;
558 
559 	/* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
560 	if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
561 		/* GWS and OA don't need any alignment. */
562 		page_align = bp->byte_align;
563 		size <<= PAGE_SHIFT;
564 
565 	} else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
566 		/* Both size and alignment must be a multiple of 4. */
567 		page_align = ALIGN(bp->byte_align, 4);
568 		size = ALIGN(size, 4) << PAGE_SHIFT;
569 	} else {
570 		/* Memory should be aligned at least to a page size. */
571 		page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
572 		size = ALIGN(size, PAGE_SIZE);
573 	}
574 
575 	if (!amdgpu_bo_validate_size(adev, size, bp->domain))
576 		return -ENOMEM;
577 
578 	BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo));
579 
580 	*bo_ptr = NULL;
581 	bo = kvzalloc(bp->bo_ptr_size, GFP_KERNEL);
582 	if (bo == NULL)
583 		return -ENOMEM;
584 	drm_gem_private_object_init(adev_to_drm(adev), &bo->tbo.base, size);
585 	bo->vm_bo = NULL;
586 	bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
587 		bp->domain;
588 	bo->allowed_domains = bo->preferred_domains;
589 	if (bp->type != ttm_bo_type_kernel &&
590 	    !(bp->flags & AMDGPU_GEM_CREATE_DISCARDABLE) &&
591 	    bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
592 		bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
593 
594 	bo->flags = bp->flags;
595 
596 	if (adev->gmc.mem_partitions)
597 		/* For GPUs with spatial partitioning, bo->xcp_id=-1 means any partition */
598 		bo->xcp_id = bp->xcp_id_plus1 - 1;
599 	else
600 		/* For GPUs without spatial partitioning */
601 		bo->xcp_id = 0;
602 
603 	if (!amdgpu_bo_support_uswc(bo->flags))
604 		bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
605 
606 	if (adev->ras_enabled)
607 		bo->flags |= AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE;
608 
609 	bo->tbo.bdev = &adev->mman.bdev;
610 	if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
611 			  AMDGPU_GEM_DOMAIN_GDS))
612 		amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
613 	else
614 		amdgpu_bo_placement_from_domain(bo, bp->domain);
615 	if (bp->type == ttm_bo_type_kernel)
616 		bo->tbo.priority = 2;
617 	else if (!(bp->flags & AMDGPU_GEM_CREATE_DISCARDABLE))
618 		bo->tbo.priority = 1;
619 
620 	if (!bp->destroy)
621 		bp->destroy = &amdgpu_bo_destroy;
622 
623 	r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, bp->type,
624 				 &bo->placement, page_align, &ctx,  NULL,
625 				 bp->resv, bp->destroy);
626 	if (unlikely(r != 0))
627 		return r;
628 
629 	if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
630 	    amdgpu_res_cpu_visible(adev, bo->tbo.resource))
631 		amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
632 					     ctx.bytes_moved);
633 	else
634 		amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
635 
636 	if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
637 	    bo->tbo.resource->mem_type == TTM_PL_VRAM) {
638 		struct dma_fence *fence;
639 
640 		r = amdgpu_fill_buffer(bo, 0, bo->tbo.base.resv, &fence, true);
641 		if (unlikely(r))
642 			goto fail_unreserve;
643 
644 		dma_resv_add_fence(bo->tbo.base.resv, fence,
645 				   DMA_RESV_USAGE_KERNEL);
646 		dma_fence_put(fence);
647 	}
648 	if (!bp->resv)
649 		amdgpu_bo_unreserve(bo);
650 	*bo_ptr = bo;
651 
652 	trace_amdgpu_bo_create(bo);
653 
654 	/* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
655 	if (bp->type == ttm_bo_type_device)
656 		bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
657 
658 	return 0;
659 
660 fail_unreserve:
661 	if (!bp->resv)
662 		dma_resv_unlock(bo->tbo.base.resv);
663 	amdgpu_bo_unref(&bo);
664 	return r;
665 }
666 
667 /**
668  * amdgpu_bo_create_user - create an &amdgpu_bo_user buffer object
669  * @adev: amdgpu device object
670  * @bp: parameters to be used for the buffer object
671  * @ubo_ptr: pointer to the buffer object pointer
672  *
673  * Create a BO to be used by user application;
674  *
675  * Returns:
676  * 0 for success or a negative error code on failure.
677  */
678 
679 int amdgpu_bo_create_user(struct amdgpu_device *adev,
680 			  struct amdgpu_bo_param *bp,
681 			  struct amdgpu_bo_user **ubo_ptr)
682 {
683 	struct amdgpu_bo *bo_ptr;
684 	int r;
685 
686 	bp->bo_ptr_size = sizeof(struct amdgpu_bo_user);
687 	bp->destroy = &amdgpu_bo_user_destroy;
688 	r = amdgpu_bo_create(adev, bp, &bo_ptr);
689 	if (r)
690 		return r;
691 
692 	*ubo_ptr = to_amdgpu_bo_user(bo_ptr);
693 	return r;
694 }
695 
696 /**
697  * amdgpu_bo_create_vm - create an &amdgpu_bo_vm buffer object
698  * @adev: amdgpu device object
699  * @bp: parameters to be used for the buffer object
700  * @vmbo_ptr: pointer to the buffer object pointer
701  *
702  * Create a BO to be for GPUVM.
703  *
704  * Returns:
705  * 0 for success or a negative error code on failure.
706  */
707 
708 int amdgpu_bo_create_vm(struct amdgpu_device *adev,
709 			struct amdgpu_bo_param *bp,
710 			struct amdgpu_bo_vm **vmbo_ptr)
711 {
712 	struct amdgpu_bo *bo_ptr;
713 	int r;
714 
715 	/* bo_ptr_size will be determined by the caller and it depends on
716 	 * num of amdgpu_vm_pt entries.
717 	 */
718 	BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo_vm));
719 	r = amdgpu_bo_create(adev, bp, &bo_ptr);
720 	if (r)
721 		return r;
722 
723 	*vmbo_ptr = to_amdgpu_bo_vm(bo_ptr);
724 	return r;
725 }
726 
727 /**
728  * amdgpu_bo_add_to_shadow_list - add a BO to the shadow list
729  *
730  * @vmbo: BO that will be inserted into the shadow list
731  *
732  * Insert a BO to the shadow list.
733  */
734 void amdgpu_bo_add_to_shadow_list(struct amdgpu_bo_vm *vmbo)
735 {
736 	struct amdgpu_device *adev = amdgpu_ttm_adev(vmbo->bo.tbo.bdev);
737 
738 	mutex_lock(&adev->shadow_list_lock);
739 	list_add_tail(&vmbo->shadow_list, &adev->shadow_list);
740 	vmbo->shadow->parent = amdgpu_bo_ref(&vmbo->bo);
741 	vmbo->shadow->tbo.destroy = &amdgpu_bo_vm_destroy;
742 	mutex_unlock(&adev->shadow_list_lock);
743 }
744 
745 /**
746  * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow
747  *
748  * @shadow: &amdgpu_bo shadow to be restored
749  * @fence: dma_fence associated with the operation
750  *
751  * Copies a buffer object's shadow content back to the object.
752  * This is used for recovering a buffer from its shadow in case of a gpu
753  * reset where vram context may be lost.
754  *
755  * Returns:
756  * 0 for success or a negative error code on failure.
757  */
758 int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence)
759 
760 {
761 	struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev);
762 	struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
763 	uint64_t shadow_addr, parent_addr;
764 
765 	shadow_addr = amdgpu_bo_gpu_offset(shadow);
766 	parent_addr = amdgpu_bo_gpu_offset(shadow->parent);
767 
768 	return amdgpu_copy_buffer(ring, shadow_addr, parent_addr,
769 				  amdgpu_bo_size(shadow), NULL, fence,
770 				  true, false, 0);
771 }
772 
773 /**
774  * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
775  * @bo: &amdgpu_bo buffer object to be mapped
776  * @ptr: kernel virtual address to be returned
777  *
778  * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
779  * amdgpu_bo_kptr() to get the kernel virtual address.
780  *
781  * Returns:
782  * 0 for success or a negative error code on failure.
783  */
784 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
785 {
786 	void *kptr;
787 	long r;
788 
789 	if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
790 		return -EPERM;
791 
792 	r = dma_resv_wait_timeout(bo->tbo.base.resv, DMA_RESV_USAGE_KERNEL,
793 				  false, MAX_SCHEDULE_TIMEOUT);
794 	if (r < 0)
795 		return r;
796 
797 	kptr = amdgpu_bo_kptr(bo);
798 	if (kptr) {
799 		if (ptr)
800 			*ptr = kptr;
801 		return 0;
802 	}
803 
804 	r = ttm_bo_kmap(&bo->tbo, 0, PFN_UP(bo->tbo.base.size), &bo->kmap);
805 	if (r)
806 		return r;
807 
808 	if (ptr)
809 		*ptr = amdgpu_bo_kptr(bo);
810 
811 	return 0;
812 }
813 
814 /**
815  * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
816  * @bo: &amdgpu_bo buffer object
817  *
818  * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
819  *
820  * Returns:
821  * the virtual address of a buffer object area.
822  */
823 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
824 {
825 	bool is_iomem;
826 
827 	return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
828 }
829 
830 /**
831  * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
832  * @bo: &amdgpu_bo buffer object to be unmapped
833  *
834  * Unmaps a kernel map set up by amdgpu_bo_kmap().
835  */
836 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
837 {
838 	if (bo->kmap.bo)
839 		ttm_bo_kunmap(&bo->kmap);
840 }
841 
842 /**
843  * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
844  * @bo: &amdgpu_bo buffer object
845  *
846  * References the contained &ttm_buffer_object.
847  *
848  * Returns:
849  * a refcounted pointer to the &amdgpu_bo buffer object.
850  */
851 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
852 {
853 	if (bo == NULL)
854 		return NULL;
855 
856 	ttm_bo_get(&bo->tbo);
857 	return bo;
858 }
859 
860 /**
861  * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
862  * @bo: &amdgpu_bo buffer object
863  *
864  * Unreferences the contained &ttm_buffer_object and clear the pointer
865  */
866 void amdgpu_bo_unref(struct amdgpu_bo **bo)
867 {
868 	struct ttm_buffer_object *tbo;
869 
870 	if ((*bo) == NULL)
871 		return;
872 
873 	tbo = &((*bo)->tbo);
874 	ttm_bo_put(tbo);
875 	*bo = NULL;
876 }
877 
878 /**
879  * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
880  * @bo: &amdgpu_bo buffer object to be pinned
881  * @domain: domain to be pinned to
882  * @min_offset: the start of requested address range
883  * @max_offset: the end of requested address range
884  *
885  * Pins the buffer object according to requested domain and address range. If
886  * the memory is unbound gart memory, binds the pages into gart table. Adjusts
887  * pin_count and pin_size accordingly.
888  *
889  * Pinning means to lock pages in memory along with keeping them at a fixed
890  * offset. It is required when a buffer can not be moved, for example, when
891  * a display buffer is being scanned out.
892  *
893  * Compared with amdgpu_bo_pin(), this function gives more flexibility on
894  * where to pin a buffer if there are specific restrictions on where a buffer
895  * must be located.
896  *
897  * Returns:
898  * 0 for success or a negative error code on failure.
899  */
900 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
901 			     u64 min_offset, u64 max_offset)
902 {
903 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
904 	struct ttm_operation_ctx ctx = { false, false };
905 	int r, i;
906 
907 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
908 		return -EPERM;
909 
910 	if (WARN_ON_ONCE(min_offset > max_offset))
911 		return -EINVAL;
912 
913 	/* Check domain to be pinned to against preferred domains */
914 	if (bo->preferred_domains & domain)
915 		domain = bo->preferred_domains & domain;
916 
917 	/* A shared bo cannot be migrated to VRAM */
918 	if (bo->tbo.base.import_attach) {
919 		if (domain & AMDGPU_GEM_DOMAIN_GTT)
920 			domain = AMDGPU_GEM_DOMAIN_GTT;
921 		else
922 			return -EINVAL;
923 	}
924 
925 	if (bo->tbo.pin_count) {
926 		uint32_t mem_type = bo->tbo.resource->mem_type;
927 		uint32_t mem_flags = bo->tbo.resource->placement;
928 
929 		if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
930 			return -EINVAL;
931 
932 		if ((mem_type == TTM_PL_VRAM) &&
933 		    (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) &&
934 		    !(mem_flags & TTM_PL_FLAG_CONTIGUOUS))
935 			return -EINVAL;
936 
937 		ttm_bo_pin(&bo->tbo);
938 
939 		if (max_offset != 0) {
940 			u64 domain_start = amdgpu_ttm_domain_start(adev,
941 								   mem_type);
942 			WARN_ON_ONCE(max_offset <
943 				     (amdgpu_bo_gpu_offset(bo) - domain_start));
944 		}
945 
946 		return 0;
947 	}
948 
949 	/* This assumes only APU display buffers are pinned with (VRAM|GTT).
950 	 * See function amdgpu_display_supported_domains()
951 	 */
952 	domain = amdgpu_bo_get_preferred_domain(adev, domain);
953 
954 	if (bo->tbo.base.import_attach)
955 		dma_buf_pin(bo->tbo.base.import_attach);
956 
957 	/* force to pin into visible video ram */
958 	if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
959 		bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
960 	amdgpu_bo_placement_from_domain(bo, domain);
961 	for (i = 0; i < bo->placement.num_placement; i++) {
962 		unsigned int fpfn, lpfn;
963 
964 		fpfn = min_offset >> PAGE_SHIFT;
965 		lpfn = max_offset >> PAGE_SHIFT;
966 
967 		if (fpfn > bo->placements[i].fpfn)
968 			bo->placements[i].fpfn = fpfn;
969 		if (!bo->placements[i].lpfn ||
970 		    (lpfn && lpfn < bo->placements[i].lpfn))
971 			bo->placements[i].lpfn = lpfn;
972 
973 		if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS &&
974 		    bo->placements[i].mem_type == TTM_PL_VRAM)
975 			bo->placements[i].flags |= TTM_PL_FLAG_CONTIGUOUS;
976 	}
977 
978 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
979 	if (unlikely(r)) {
980 		dev_err(adev->dev, "%p pin failed\n", bo);
981 		goto error;
982 	}
983 
984 	ttm_bo_pin(&bo->tbo);
985 
986 	domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type);
987 	if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
988 		atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
989 		atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
990 			     &adev->visible_pin_size);
991 	} else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
992 		atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
993 	}
994 
995 error:
996 	return r;
997 }
998 
999 /**
1000  * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
1001  * @bo: &amdgpu_bo buffer object to be pinned
1002  * @domain: domain to be pinned to
1003  *
1004  * A simple wrapper to amdgpu_bo_pin_restricted().
1005  * Provides a simpler API for buffers that do not have any strict restrictions
1006  * on where a buffer must be located.
1007  *
1008  * Returns:
1009  * 0 for success or a negative error code on failure.
1010  */
1011 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
1012 {
1013 	bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1014 	return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
1015 }
1016 
1017 /**
1018  * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
1019  * @bo: &amdgpu_bo buffer object to be unpinned
1020  *
1021  * Decreases the pin_count, and clears the flags if pin_count reaches 0.
1022  * Changes placement and pin size accordingly.
1023  *
1024  * Returns:
1025  * 0 for success or a negative error code on failure.
1026  */
1027 void amdgpu_bo_unpin(struct amdgpu_bo *bo)
1028 {
1029 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1030 
1031 	ttm_bo_unpin(&bo->tbo);
1032 	if (bo->tbo.pin_count)
1033 		return;
1034 
1035 	if (bo->tbo.base.import_attach)
1036 		dma_buf_unpin(bo->tbo.base.import_attach);
1037 
1038 	if (bo->tbo.resource->mem_type == TTM_PL_VRAM) {
1039 		atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
1040 		atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
1041 			     &adev->visible_pin_size);
1042 	} else if (bo->tbo.resource->mem_type == TTM_PL_TT) {
1043 		atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size);
1044 	}
1045 
1046 }
1047 
1048 static const char * const amdgpu_vram_names[] = {
1049 	"UNKNOWN",
1050 	"GDDR1",
1051 	"DDR2",
1052 	"GDDR3",
1053 	"GDDR4",
1054 	"GDDR5",
1055 	"HBM",
1056 	"DDR3",
1057 	"DDR4",
1058 	"GDDR6",
1059 	"DDR5",
1060 	"LPDDR4",
1061 	"LPDDR5"
1062 };
1063 
1064 /**
1065  * amdgpu_bo_init - initialize memory manager
1066  * @adev: amdgpu device object
1067  *
1068  * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
1069  *
1070  * Returns:
1071  * 0 for success or a negative error code on failure.
1072  */
1073 int amdgpu_bo_init(struct amdgpu_device *adev)
1074 {
1075 	/* On A+A platform, VRAM can be mapped as WB */
1076 	if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) {
1077 		/* reserve PAT memory space to WC for VRAM */
1078 		int r = arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1079 				adev->gmc.aper_size);
1080 
1081 		if (r) {
1082 			DRM_ERROR("Unable to set WC memtype for the aperture base\n");
1083 			return r;
1084 		}
1085 
1086 		/* Add an MTRR for the VRAM */
1087 		adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1088 				adev->gmc.aper_size);
1089 	}
1090 
1091 	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1092 		 adev->gmc.mc_vram_size >> 20,
1093 		 (unsigned long long)adev->gmc.aper_size >> 20);
1094 	DRM_INFO("RAM width %dbits %s\n",
1095 		 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1096 	return amdgpu_ttm_init(adev);
1097 }
1098 
1099 /**
1100  * amdgpu_bo_fini - tear down memory manager
1101  * @adev: amdgpu device object
1102  *
1103  * Reverses amdgpu_bo_init() to tear down memory manager.
1104  */
1105 void amdgpu_bo_fini(struct amdgpu_device *adev)
1106 {
1107 	int idx;
1108 
1109 	amdgpu_ttm_fini(adev);
1110 
1111 	if (drm_dev_enter(adev_to_drm(adev), &idx)) {
1112 		if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) {
1113 			arch_phys_wc_del(adev->gmc.vram_mtrr);
1114 			arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1115 		}
1116 		drm_dev_exit(idx);
1117 	}
1118 }
1119 
1120 /**
1121  * amdgpu_bo_set_tiling_flags - set tiling flags
1122  * @bo: &amdgpu_bo buffer object
1123  * @tiling_flags: new flags
1124  *
1125  * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1126  * kernel driver to set the tiling flags on a buffer.
1127  *
1128  * Returns:
1129  * 0 for success or a negative error code on failure.
1130  */
1131 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1132 {
1133 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1134 	struct amdgpu_bo_user *ubo;
1135 
1136 	BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1137 	if (adev->family <= AMDGPU_FAMILY_CZ &&
1138 	    AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1139 		return -EINVAL;
1140 
1141 	ubo = to_amdgpu_bo_user(bo);
1142 	ubo->tiling_flags = tiling_flags;
1143 	return 0;
1144 }
1145 
1146 /**
1147  * amdgpu_bo_get_tiling_flags - get tiling flags
1148  * @bo: &amdgpu_bo buffer object
1149  * @tiling_flags: returned flags
1150  *
1151  * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1152  * set the tiling flags on a buffer.
1153  */
1154 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1155 {
1156 	struct amdgpu_bo_user *ubo;
1157 
1158 	BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1159 	dma_resv_assert_held(bo->tbo.base.resv);
1160 	ubo = to_amdgpu_bo_user(bo);
1161 
1162 	if (tiling_flags)
1163 		*tiling_flags = ubo->tiling_flags;
1164 }
1165 
1166 /**
1167  * amdgpu_bo_set_metadata - set metadata
1168  * @bo: &amdgpu_bo buffer object
1169  * @metadata: new metadata
1170  * @metadata_size: size of the new metadata
1171  * @flags: flags of the new metadata
1172  *
1173  * Sets buffer object's metadata, its size and flags.
1174  * Used via GEM ioctl.
1175  *
1176  * Returns:
1177  * 0 for success or a negative error code on failure.
1178  */
1179 int amdgpu_bo_set_metadata(struct amdgpu_bo *bo, void *metadata,
1180 			   u32 metadata_size, uint64_t flags)
1181 {
1182 	struct amdgpu_bo_user *ubo;
1183 	void *buffer;
1184 
1185 	BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1186 	ubo = to_amdgpu_bo_user(bo);
1187 	if (!metadata_size) {
1188 		if (ubo->metadata_size) {
1189 			kfree(ubo->metadata);
1190 			ubo->metadata = NULL;
1191 			ubo->metadata_size = 0;
1192 		}
1193 		return 0;
1194 	}
1195 
1196 	if (metadata == NULL)
1197 		return -EINVAL;
1198 
1199 	buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1200 	if (buffer == NULL)
1201 		return -ENOMEM;
1202 
1203 	kfree(ubo->metadata);
1204 	ubo->metadata_flags = flags;
1205 	ubo->metadata = buffer;
1206 	ubo->metadata_size = metadata_size;
1207 
1208 	return 0;
1209 }
1210 
1211 /**
1212  * amdgpu_bo_get_metadata - get metadata
1213  * @bo: &amdgpu_bo buffer object
1214  * @buffer: returned metadata
1215  * @buffer_size: size of the buffer
1216  * @metadata_size: size of the returned metadata
1217  * @flags: flags of the returned metadata
1218  *
1219  * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1220  * less than metadata_size.
1221  * Used via GEM ioctl.
1222  *
1223  * Returns:
1224  * 0 for success or a negative error code on failure.
1225  */
1226 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1227 			   size_t buffer_size, uint32_t *metadata_size,
1228 			   uint64_t *flags)
1229 {
1230 	struct amdgpu_bo_user *ubo;
1231 
1232 	if (!buffer && !metadata_size)
1233 		return -EINVAL;
1234 
1235 	BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1236 	ubo = to_amdgpu_bo_user(bo);
1237 	if (metadata_size)
1238 		*metadata_size = ubo->metadata_size;
1239 
1240 	if (buffer) {
1241 		if (buffer_size < ubo->metadata_size)
1242 			return -EINVAL;
1243 
1244 		if (ubo->metadata_size)
1245 			memcpy(buffer, ubo->metadata, ubo->metadata_size);
1246 	}
1247 
1248 	if (flags)
1249 		*flags = ubo->metadata_flags;
1250 
1251 	return 0;
1252 }
1253 
1254 /**
1255  * amdgpu_bo_move_notify - notification about a memory move
1256  * @bo: pointer to a buffer object
1257  * @evict: if this move is evicting the buffer from the graphics address space
1258  * @new_mem: new resource for backing the BO
1259  *
1260  * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1261  * bookkeeping.
1262  * TTM driver callback which is called when ttm moves a buffer.
1263  */
1264 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1265 			   bool evict,
1266 			   struct ttm_resource *new_mem)
1267 {
1268 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1269 	struct ttm_resource *old_mem = bo->resource;
1270 	struct amdgpu_bo *abo;
1271 
1272 	if (!amdgpu_bo_is_amdgpu_bo(bo))
1273 		return;
1274 
1275 	abo = ttm_to_amdgpu_bo(bo);
1276 	amdgpu_vm_bo_invalidate(adev, abo, evict);
1277 
1278 	amdgpu_bo_kunmap(abo);
1279 
1280 	if (abo->tbo.base.dma_buf && !abo->tbo.base.import_attach &&
1281 	    old_mem && old_mem->mem_type != TTM_PL_SYSTEM)
1282 		dma_buf_move_notify(abo->tbo.base.dma_buf);
1283 
1284 	/* move_notify is called before move happens */
1285 	trace_amdgpu_bo_move(abo, new_mem ? new_mem->mem_type : -1,
1286 			     old_mem ? old_mem->mem_type : -1);
1287 }
1288 
1289 void amdgpu_bo_get_memory(struct amdgpu_bo *bo,
1290 			  struct amdgpu_mem_stats *stats)
1291 {
1292 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1293 	struct ttm_resource *res = bo->tbo.resource;
1294 	uint64_t size = amdgpu_bo_size(bo);
1295 	struct drm_gem_object *obj;
1296 	unsigned int domain;
1297 	bool shared;
1298 
1299 	/* Abort if the BO doesn't currently have a backing store */
1300 	if (!res)
1301 		return;
1302 
1303 	obj = &bo->tbo.base;
1304 	shared = drm_gem_object_is_shared_for_memory_stats(obj);
1305 
1306 	domain = amdgpu_mem_type_to_domain(res->mem_type);
1307 	switch (domain) {
1308 	case AMDGPU_GEM_DOMAIN_VRAM:
1309 		stats->vram += size;
1310 		if (amdgpu_res_cpu_visible(adev, bo->tbo.resource))
1311 			stats->visible_vram += size;
1312 		if (shared)
1313 			stats->vram_shared += size;
1314 		break;
1315 	case AMDGPU_GEM_DOMAIN_GTT:
1316 		stats->gtt += size;
1317 		if (shared)
1318 			stats->gtt_shared += size;
1319 		break;
1320 	case AMDGPU_GEM_DOMAIN_CPU:
1321 	default:
1322 		stats->cpu += size;
1323 		if (shared)
1324 			stats->cpu_shared += size;
1325 		break;
1326 	}
1327 
1328 	if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) {
1329 		stats->requested_vram += size;
1330 		if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
1331 			stats->requested_visible_vram += size;
1332 
1333 		if (domain != AMDGPU_GEM_DOMAIN_VRAM) {
1334 			stats->evicted_vram += size;
1335 			if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
1336 				stats->evicted_visible_vram += size;
1337 		}
1338 	} else if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_GTT) {
1339 		stats->requested_gtt += size;
1340 	}
1341 }
1342 
1343 /**
1344  * amdgpu_bo_release_notify - notification about a BO being released
1345  * @bo: pointer to a buffer object
1346  *
1347  * Wipes VRAM buffers whose contents should not be leaked before the
1348  * memory is released.
1349  */
1350 void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)
1351 {
1352 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1353 	struct dma_fence *fence = NULL;
1354 	struct amdgpu_bo *abo;
1355 	int r;
1356 
1357 	if (!amdgpu_bo_is_amdgpu_bo(bo))
1358 		return;
1359 
1360 	abo = ttm_to_amdgpu_bo(bo);
1361 
1362 	WARN_ON(abo->vm_bo);
1363 
1364 	if (abo->kfd_bo)
1365 		amdgpu_amdkfd_release_notify(abo);
1366 
1367 	/* We only remove the fence if the resv has individualized. */
1368 	WARN_ON_ONCE(bo->type == ttm_bo_type_kernel
1369 			&& bo->base.resv != &bo->base._resv);
1370 	if (bo->base.resv == &bo->base._resv)
1371 		amdgpu_amdkfd_remove_fence_on_pt_pd_bos(abo);
1372 
1373 	if (!bo->resource || bo->resource->mem_type != TTM_PL_VRAM ||
1374 	    !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE) ||
1375 	    adev->in_suspend || drm_dev_is_unplugged(adev_to_drm(adev)))
1376 		return;
1377 
1378 	if (WARN_ON_ONCE(!dma_resv_trylock(bo->base.resv)))
1379 		return;
1380 
1381 	r = amdgpu_fill_buffer(abo, AMDGPU_POISON, bo->base.resv, &fence, true);
1382 	if (!WARN_ON(r)) {
1383 		amdgpu_bo_fence(abo, fence, false);
1384 		dma_fence_put(fence);
1385 	}
1386 
1387 	dma_resv_unlock(bo->base.resv);
1388 }
1389 
1390 /**
1391  * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1392  * @bo: pointer to a buffer object
1393  *
1394  * Notifies the driver we are taking a fault on this BO and have reserved it,
1395  * also performs bookkeeping.
1396  * TTM driver callback for dealing with vm faults.
1397  *
1398  * Returns:
1399  * 0 for success or a negative error code on failure.
1400  */
1401 vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1402 {
1403 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1404 	struct ttm_operation_ctx ctx = { false, false };
1405 	struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1406 	int r;
1407 
1408 	/* Remember that this BO was accessed by the CPU */
1409 	abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1410 
1411 	if (amdgpu_res_cpu_visible(adev, bo->resource))
1412 		return 0;
1413 
1414 	/* Can't move a pinned BO to visible VRAM */
1415 	if (abo->tbo.pin_count > 0)
1416 		return VM_FAULT_SIGBUS;
1417 
1418 	/* hurrah the memory is not visible ! */
1419 	atomic64_inc(&adev->num_vram_cpu_page_faults);
1420 	amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1421 					AMDGPU_GEM_DOMAIN_GTT);
1422 
1423 	/* Avoid costly evictions; only set GTT as a busy placement */
1424 	abo->placements[0].flags |= TTM_PL_FLAG_DESIRED;
1425 
1426 	r = ttm_bo_validate(bo, &abo->placement, &ctx);
1427 	if (unlikely(r == -EBUSY || r == -ERESTARTSYS))
1428 		return VM_FAULT_NOPAGE;
1429 	else if (unlikely(r))
1430 		return VM_FAULT_SIGBUS;
1431 
1432 	/* this should never happen */
1433 	if (bo->resource->mem_type == TTM_PL_VRAM &&
1434 	    !amdgpu_res_cpu_visible(adev, bo->resource))
1435 		return VM_FAULT_SIGBUS;
1436 
1437 	ttm_bo_move_to_lru_tail_unlocked(bo);
1438 	return 0;
1439 }
1440 
1441 /**
1442  * amdgpu_bo_fence - add fence to buffer object
1443  *
1444  * @bo: buffer object in question
1445  * @fence: fence to add
1446  * @shared: true if fence should be added shared
1447  *
1448  */
1449 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1450 		     bool shared)
1451 {
1452 	struct dma_resv *resv = bo->tbo.base.resv;
1453 	int r;
1454 
1455 	r = dma_resv_reserve_fences(resv, 1);
1456 	if (r) {
1457 		/* As last resort on OOM we block for the fence */
1458 		dma_fence_wait(fence, false);
1459 		return;
1460 	}
1461 
1462 	dma_resv_add_fence(resv, fence, shared ? DMA_RESV_USAGE_READ :
1463 			   DMA_RESV_USAGE_WRITE);
1464 }
1465 
1466 /**
1467  * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences
1468  *
1469  * @adev: amdgpu device pointer
1470  * @resv: reservation object to sync to
1471  * @sync_mode: synchronization mode
1472  * @owner: fence owner
1473  * @intr: Whether the wait is interruptible
1474  *
1475  * Extract the fences from the reservation object and waits for them to finish.
1476  *
1477  * Returns:
1478  * 0 on success, errno otherwise.
1479  */
1480 int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv,
1481 			     enum amdgpu_sync_mode sync_mode, void *owner,
1482 			     bool intr)
1483 {
1484 	struct amdgpu_sync sync;
1485 	int r;
1486 
1487 	amdgpu_sync_create(&sync);
1488 	amdgpu_sync_resv(adev, &sync, resv, sync_mode, owner);
1489 	r = amdgpu_sync_wait(&sync, intr);
1490 	amdgpu_sync_free(&sync);
1491 	return r;
1492 }
1493 
1494 /**
1495  * amdgpu_bo_sync_wait - Wrapper for amdgpu_bo_sync_wait_resv
1496  * @bo: buffer object to wait for
1497  * @owner: fence owner
1498  * @intr: Whether the wait is interruptible
1499  *
1500  * Wrapper to wait for fences in a BO.
1501  * Returns:
1502  * 0 on success, errno otherwise.
1503  */
1504 int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)
1505 {
1506 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1507 
1508 	return amdgpu_bo_sync_wait_resv(adev, bo->tbo.base.resv,
1509 					AMDGPU_SYNC_NE_OWNER, owner, intr);
1510 }
1511 
1512 /**
1513  * amdgpu_bo_gpu_offset - return GPU offset of bo
1514  * @bo:	amdgpu object for which we query the offset
1515  *
1516  * Note: object should either be pinned or reserved when calling this
1517  * function, it might be useful to add check for this for debugging.
1518  *
1519  * Returns:
1520  * current GPU offset of the object.
1521  */
1522 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1523 {
1524 	WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_SYSTEM);
1525 	WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) &&
1526 		     !bo->tbo.pin_count && bo->tbo.type != ttm_bo_type_kernel);
1527 	WARN_ON_ONCE(bo->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET);
1528 	WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_VRAM &&
1529 		     !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1530 
1531 	return amdgpu_bo_gpu_offset_no_check(bo);
1532 }
1533 
1534 /**
1535  * amdgpu_bo_gpu_offset_no_check - return GPU offset of bo
1536  * @bo:	amdgpu object for which we query the offset
1537  *
1538  * Returns:
1539  * current GPU offset of the object without raising warnings.
1540  */
1541 u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo)
1542 {
1543 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1544 	uint64_t offset = AMDGPU_BO_INVALID_OFFSET;
1545 
1546 	if (bo->tbo.resource->mem_type == TTM_PL_TT)
1547 		offset = amdgpu_gmc_agp_addr(&bo->tbo);
1548 
1549 	if (offset == AMDGPU_BO_INVALID_OFFSET)
1550 		offset = (bo->tbo.resource->start << PAGE_SHIFT) +
1551 			amdgpu_ttm_domain_start(adev, bo->tbo.resource->mem_type);
1552 
1553 	return amdgpu_gmc_sign_extend(offset);
1554 }
1555 
1556 /**
1557  * amdgpu_bo_get_preferred_domain - get preferred domain
1558  * @adev: amdgpu device object
1559  * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1560  *
1561  * Returns:
1562  * Which of the allowed domains is preferred for allocating the BO.
1563  */
1564 uint32_t amdgpu_bo_get_preferred_domain(struct amdgpu_device *adev,
1565 					    uint32_t domain)
1566 {
1567 	if ((domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) &&
1568 	    ((adev->asic_type == CHIP_CARRIZO) || (adev->asic_type == CHIP_STONEY))) {
1569 		domain = AMDGPU_GEM_DOMAIN_VRAM;
1570 		if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1571 			domain = AMDGPU_GEM_DOMAIN_GTT;
1572 	}
1573 	return domain;
1574 }
1575 
1576 #if defined(CONFIG_DEBUG_FS)
1577 #define amdgpu_bo_print_flag(m, bo, flag)		        \
1578 	do {							\
1579 		if (bo->flags & (AMDGPU_GEM_CREATE_ ## flag)) {	\
1580 			seq_printf((m), " " #flag);		\
1581 		}						\
1582 	} while (0)
1583 
1584 /**
1585  * amdgpu_bo_print_info - print BO info in debugfs file
1586  *
1587  * @id: Index or Id of the BO
1588  * @bo: Requested BO for printing info
1589  * @m: debugfs file
1590  *
1591  * Print BO information in debugfs file
1592  *
1593  * Returns:
1594  * Size of the BO in bytes.
1595  */
1596 u64 amdgpu_bo_print_info(int id, struct amdgpu_bo *bo, struct seq_file *m)
1597 {
1598 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1599 	struct dma_buf_attachment *attachment;
1600 	struct dma_buf *dma_buf;
1601 	const char *placement;
1602 	unsigned int pin_count;
1603 	u64 size;
1604 
1605 	if (dma_resv_trylock(bo->tbo.base.resv)) {
1606 		unsigned int domain;
1607 
1608 		domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type);
1609 		switch (domain) {
1610 		case AMDGPU_GEM_DOMAIN_VRAM:
1611 			if (amdgpu_res_cpu_visible(adev, bo->tbo.resource))
1612 				placement = "VRAM VISIBLE";
1613 			else
1614 				placement = "VRAM";
1615 			break;
1616 		case AMDGPU_GEM_DOMAIN_GTT:
1617 			placement = "GTT";
1618 			break;
1619 		case AMDGPU_GEM_DOMAIN_CPU:
1620 		default:
1621 			placement = "CPU";
1622 			break;
1623 		}
1624 		dma_resv_unlock(bo->tbo.base.resv);
1625 	} else {
1626 		placement = "UNKNOWN";
1627 	}
1628 
1629 	size = amdgpu_bo_size(bo);
1630 	seq_printf(m, "\t\t0x%08x: %12lld byte %s",
1631 			id, size, placement);
1632 
1633 	pin_count = READ_ONCE(bo->tbo.pin_count);
1634 	if (pin_count)
1635 		seq_printf(m, " pin count %d", pin_count);
1636 
1637 	dma_buf = READ_ONCE(bo->tbo.base.dma_buf);
1638 	attachment = READ_ONCE(bo->tbo.base.import_attach);
1639 
1640 	if (attachment)
1641 		seq_printf(m, " imported from ino:%lu", file_inode(dma_buf->file)->i_ino);
1642 	else if (dma_buf)
1643 		seq_printf(m, " exported as ino:%lu", file_inode(dma_buf->file)->i_ino);
1644 
1645 	amdgpu_bo_print_flag(m, bo, CPU_ACCESS_REQUIRED);
1646 	amdgpu_bo_print_flag(m, bo, NO_CPU_ACCESS);
1647 	amdgpu_bo_print_flag(m, bo, CPU_GTT_USWC);
1648 	amdgpu_bo_print_flag(m, bo, VRAM_CLEARED);
1649 	amdgpu_bo_print_flag(m, bo, VRAM_CONTIGUOUS);
1650 	amdgpu_bo_print_flag(m, bo, VM_ALWAYS_VALID);
1651 	amdgpu_bo_print_flag(m, bo, EXPLICIT_SYNC);
1652 
1653 	seq_puts(m, "\n");
1654 
1655 	return size;
1656 }
1657 #endif
1658