xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c (revision 994aeacbb3c039b4f3e02e76e6d39407920e76c6)
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 #include "amdgpu_vram_mgr.h"
43 
44 /**
45  * DOC: amdgpu_object
46  *
47  * This defines the interfaces to operate on an &amdgpu_bo buffer object which
48  * represents memory used by driver (VRAM, system memory, etc.). The driver
49  * provides DRM/GEM APIs to userspace. DRM/GEM APIs then use these interfaces
50  * to create/destroy/set buffer object which are then managed by the kernel TTM
51  * memory manager.
52  * The interfaces are also used internally by kernel clients, including gfx,
53  * uvd, etc. for kernel managed allocations used by the GPU.
54  *
55  */
56 
amdgpu_bo_destroy(struct ttm_buffer_object * tbo)57 static void amdgpu_bo_destroy(struct ttm_buffer_object *tbo)
58 {
59 	struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
60 
61 	amdgpu_bo_kunmap(bo);
62 
63 	if (bo->tbo.base.import_attach)
64 		drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
65 	drm_gem_object_release(&bo->tbo.base);
66 	amdgpu_bo_unref(&bo->parent);
67 	kvfree(bo);
68 }
69 
amdgpu_bo_user_destroy(struct ttm_buffer_object * tbo)70 static void amdgpu_bo_user_destroy(struct ttm_buffer_object *tbo)
71 {
72 	struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
73 	struct amdgpu_bo_user *ubo;
74 
75 	ubo = to_amdgpu_bo_user(bo);
76 	kfree(ubo->metadata);
77 	amdgpu_bo_destroy(tbo);
78 }
79 
80 /**
81  * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
82  * @bo: buffer object to be checked
83  *
84  * Uses destroy function associated with the object to determine if this is
85  * an &amdgpu_bo.
86  *
87  * Returns:
88  * true if the object belongs to &amdgpu_bo, false if not.
89  */
amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object * bo)90 bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
91 {
92 	if (bo->destroy == &amdgpu_bo_destroy ||
93 	    bo->destroy == &amdgpu_bo_user_destroy)
94 		return true;
95 
96 	return false;
97 }
98 
99 /**
100  * amdgpu_bo_placement_from_domain - set buffer's placement
101  * @abo: &amdgpu_bo buffer object whose placement is to be set
102  * @domain: requested domain
103  *
104  * Sets buffer's placement according to requested domain and the buffer's
105  * flags.
106  */
amdgpu_bo_placement_from_domain(struct amdgpu_bo * abo,u32 domain)107 void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
108 {
109 	struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
110 	struct ttm_placement *placement = &abo->placement;
111 	struct ttm_place *places = abo->placements;
112 	u64 flags = abo->flags;
113 	u32 c = 0;
114 
115 	if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
116 		unsigned int visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
117 		int8_t mem_id = KFD_XCP_MEM_ID(adev, abo->xcp_id);
118 
119 		if (adev->gmc.mem_partitions && mem_id >= 0) {
120 			places[c].fpfn = adev->gmc.mem_partitions[mem_id].range.fpfn;
121 			/*
122 			 * memory partition range lpfn is inclusive start + size - 1
123 			 * TTM place lpfn is exclusive start + size
124 			 */
125 			places[c].lpfn = adev->gmc.mem_partitions[mem_id].range.lpfn + 1;
126 		} else {
127 			places[c].fpfn = 0;
128 			places[c].lpfn = 0;
129 		}
130 		places[c].mem_type = TTM_PL_VRAM;
131 		places[c].flags = 0;
132 
133 		if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
134 			places[c].lpfn = min_not_zero(places[c].lpfn, visible_pfn);
135 		else
136 			places[c].flags |= TTM_PL_FLAG_TOPDOWN;
137 
138 		if (abo->tbo.type == ttm_bo_type_kernel &&
139 		    flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
140 			places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
141 
142 		c++;
143 	}
144 
145 	if (domain & AMDGPU_GEM_DOMAIN_DOORBELL) {
146 		places[c].fpfn = 0;
147 		places[c].lpfn = 0;
148 		places[c].mem_type = AMDGPU_PL_DOORBELL;
149 		places[c].flags = 0;
150 		c++;
151 	}
152 
153 	if (domain & AMDGPU_GEM_DOMAIN_GTT) {
154 		places[c].fpfn = 0;
155 		places[c].lpfn = 0;
156 		places[c].mem_type =
157 			abo->flags & AMDGPU_GEM_CREATE_PREEMPTIBLE ?
158 			AMDGPU_PL_PREEMPT : TTM_PL_TT;
159 		places[c].flags = 0;
160 		/*
161 		 * When GTT is just an alternative to VRAM make sure that we
162 		 * only use it as fallback and still try to fill up VRAM first.
163 		 */
164 		if (domain & abo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM)
165 			places[c].flags |= TTM_PL_FLAG_FALLBACK;
166 		c++;
167 	}
168 
169 	if (domain & AMDGPU_GEM_DOMAIN_CPU) {
170 		places[c].fpfn = 0;
171 		places[c].lpfn = 0;
172 		places[c].mem_type = TTM_PL_SYSTEM;
173 		places[c].flags = 0;
174 		c++;
175 	}
176 
177 	if (domain & AMDGPU_GEM_DOMAIN_GDS) {
178 		places[c].fpfn = 0;
179 		places[c].lpfn = 0;
180 		places[c].mem_type = AMDGPU_PL_GDS;
181 		places[c].flags = 0;
182 		c++;
183 	}
184 
185 	if (domain & AMDGPU_GEM_DOMAIN_GWS) {
186 		places[c].fpfn = 0;
187 		places[c].lpfn = 0;
188 		places[c].mem_type = AMDGPU_PL_GWS;
189 		places[c].flags = 0;
190 		c++;
191 	}
192 
193 	if (domain & AMDGPU_GEM_DOMAIN_OA) {
194 		places[c].fpfn = 0;
195 		places[c].lpfn = 0;
196 		places[c].mem_type = AMDGPU_PL_OA;
197 		places[c].flags = 0;
198 		c++;
199 	}
200 
201 	if (!c) {
202 		places[c].fpfn = 0;
203 		places[c].lpfn = 0;
204 		places[c].mem_type = TTM_PL_SYSTEM;
205 		places[c].flags = 0;
206 		c++;
207 	}
208 
209 	BUG_ON(c > AMDGPU_BO_MAX_PLACEMENTS);
210 
211 	placement->num_placement = c;
212 	placement->placement = places;
213 }
214 
215 /**
216  * amdgpu_bo_create_reserved - create reserved BO for kernel use
217  *
218  * @adev: amdgpu device object
219  * @size: size for the new BO
220  * @align: alignment for the new BO
221  * @domain: where to place it
222  * @bo_ptr: used to initialize BOs in structures
223  * @gpu_addr: GPU addr of the pinned BO
224  * @cpu_addr: optional CPU address mapping
225  *
226  * Allocates and pins a BO for kernel internal use, and returns it still
227  * reserved.
228  *
229  * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
230  *
231  * Returns:
232  * 0 on success, negative error code otherwise.
233  */
amdgpu_bo_create_reserved(struct amdgpu_device * adev,unsigned long size,int align,u32 domain,struct amdgpu_bo ** bo_ptr,u64 * gpu_addr,void ** cpu_addr)234 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
235 			      unsigned long size, int align,
236 			      u32 domain, struct amdgpu_bo **bo_ptr,
237 			      u64 *gpu_addr, void **cpu_addr)
238 {
239 	struct amdgpu_bo_param bp;
240 	bool free = false;
241 	int r;
242 
243 	if (!size) {
244 		amdgpu_bo_unref(bo_ptr);
245 		return 0;
246 	}
247 
248 	memset(&bp, 0, sizeof(bp));
249 	bp.size = size;
250 	bp.byte_align = align;
251 	bp.domain = domain;
252 	bp.flags = cpu_addr ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
253 		: AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
254 	bp.flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
255 	bp.type = ttm_bo_type_kernel;
256 	bp.resv = NULL;
257 	bp.bo_ptr_size = sizeof(struct amdgpu_bo);
258 
259 	if (!*bo_ptr) {
260 		r = amdgpu_bo_create(adev, &bp, bo_ptr);
261 		if (r) {
262 			dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
263 				r);
264 			return r;
265 		}
266 		free = true;
267 	}
268 
269 	r = amdgpu_bo_reserve(*bo_ptr, false);
270 	if (r) {
271 		dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
272 		goto error_free;
273 	}
274 
275 	r = amdgpu_bo_pin(*bo_ptr, domain);
276 	if (r) {
277 		dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
278 		goto error_unreserve;
279 	}
280 
281 	r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
282 	if (r) {
283 		dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
284 		goto error_unpin;
285 	}
286 
287 	if (gpu_addr)
288 		*gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
289 
290 	if (cpu_addr) {
291 		r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
292 		if (r) {
293 			dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
294 			goto error_unpin;
295 		}
296 	}
297 
298 	return 0;
299 
300 error_unpin:
301 	amdgpu_bo_unpin(*bo_ptr);
302 error_unreserve:
303 	amdgpu_bo_unreserve(*bo_ptr);
304 
305 error_free:
306 	if (free)
307 		amdgpu_bo_unref(bo_ptr);
308 
309 	return r;
310 }
311 
312 /**
313  * amdgpu_bo_create_kernel - create BO for kernel use
314  *
315  * @adev: amdgpu device object
316  * @size: size for the new BO
317  * @align: alignment for the new BO
318  * @domain: where to place it
319  * @bo_ptr:  used to initialize BOs in structures
320  * @gpu_addr: GPU addr of the pinned BO
321  * @cpu_addr: optional CPU address mapping
322  *
323  * Allocates and pins a BO for kernel internal use.
324  *
325  * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
326  *
327  * Returns:
328  * 0 on success, negative error code otherwise.
329  */
amdgpu_bo_create_kernel(struct amdgpu_device * adev,unsigned long size,int align,u32 domain,struct amdgpu_bo ** bo_ptr,u64 * gpu_addr,void ** cpu_addr)330 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
331 			    unsigned long size, int align,
332 			    u32 domain, struct amdgpu_bo **bo_ptr,
333 			    u64 *gpu_addr, void **cpu_addr)
334 {
335 	int r;
336 
337 	r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
338 				      gpu_addr, cpu_addr);
339 
340 	if (r)
341 		return r;
342 
343 	if (*bo_ptr)
344 		amdgpu_bo_unreserve(*bo_ptr);
345 
346 	return 0;
347 }
348 
349 /**
350  * amdgpu_bo_create_kernel_at - create BO for kernel use at specific location
351  *
352  * @adev: amdgpu device object
353  * @offset: offset of the BO
354  * @size: size of the BO
355  * @bo_ptr:  used to initialize BOs in structures
356  * @cpu_addr: optional CPU address mapping
357  *
358  * Creates a kernel BO at a specific offset in VRAM.
359  *
360  * Returns:
361  * 0 on success, negative error code otherwise.
362  */
amdgpu_bo_create_kernel_at(struct amdgpu_device * adev,uint64_t offset,uint64_t size,struct amdgpu_bo ** bo_ptr,void ** cpu_addr)363 int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev,
364 			       uint64_t offset, uint64_t size,
365 			       struct amdgpu_bo **bo_ptr, void **cpu_addr)
366 {
367 	struct ttm_operation_ctx ctx = { false, false };
368 	unsigned int i;
369 	int r;
370 
371 	offset &= PAGE_MASK;
372 	size = ALIGN(size, PAGE_SIZE);
373 
374 	r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE,
375 				      AMDGPU_GEM_DOMAIN_VRAM, bo_ptr, NULL,
376 				      cpu_addr);
377 	if (r)
378 		return r;
379 
380 	if ((*bo_ptr) == NULL)
381 		return 0;
382 
383 	/*
384 	 * Remove the original mem node and create a new one at the request
385 	 * position.
386 	 */
387 	if (cpu_addr)
388 		amdgpu_bo_kunmap(*bo_ptr);
389 
390 	ttm_resource_free(&(*bo_ptr)->tbo, &(*bo_ptr)->tbo.resource);
391 
392 	for (i = 0; i < (*bo_ptr)->placement.num_placement; ++i) {
393 		(*bo_ptr)->placements[i].fpfn = offset >> PAGE_SHIFT;
394 		(*bo_ptr)->placements[i].lpfn = (offset + size) >> PAGE_SHIFT;
395 	}
396 	r = ttm_bo_mem_space(&(*bo_ptr)->tbo, &(*bo_ptr)->placement,
397 			     &(*bo_ptr)->tbo.resource, &ctx);
398 	if (r)
399 		goto error;
400 
401 	if (cpu_addr) {
402 		r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
403 		if (r)
404 			goto error;
405 	}
406 
407 	amdgpu_bo_unreserve(*bo_ptr);
408 	return 0;
409 
410 error:
411 	amdgpu_bo_unreserve(*bo_ptr);
412 	amdgpu_bo_unref(bo_ptr);
413 	return r;
414 }
415 
416 /**
417  * amdgpu_bo_free_kernel - free BO for kernel use
418  *
419  * @bo: amdgpu BO to free
420  * @gpu_addr: pointer to where the BO's GPU memory space address was stored
421  * @cpu_addr: pointer to where the BO's CPU memory space address was stored
422  *
423  * unmaps and unpin a BO for kernel internal use.
424  */
amdgpu_bo_free_kernel(struct amdgpu_bo ** bo,u64 * gpu_addr,void ** cpu_addr)425 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
426 			   void **cpu_addr)
427 {
428 	if (*bo == NULL)
429 		return;
430 
431 	WARN_ON(amdgpu_ttm_adev((*bo)->tbo.bdev)->in_suspend);
432 
433 	if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
434 		if (cpu_addr)
435 			amdgpu_bo_kunmap(*bo);
436 
437 		amdgpu_bo_unpin(*bo);
438 		amdgpu_bo_unreserve(*bo);
439 	}
440 	amdgpu_bo_unref(bo);
441 
442 	if (gpu_addr)
443 		*gpu_addr = 0;
444 
445 	if (cpu_addr)
446 		*cpu_addr = NULL;
447 }
448 
449 /* Validate bo size is bit bigger than the request domain */
amdgpu_bo_validate_size(struct amdgpu_device * adev,unsigned long size,u32 domain)450 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
451 					  unsigned long size, u32 domain)
452 {
453 	struct ttm_resource_manager *man = NULL;
454 
455 	/*
456 	 * If GTT is part of requested domains the check must succeed to
457 	 * allow fall back to GTT.
458 	 */
459 	if (domain & AMDGPU_GEM_DOMAIN_GTT)
460 		man = ttm_manager_type(&adev->mman.bdev, TTM_PL_TT);
461 	else if (domain & AMDGPU_GEM_DOMAIN_VRAM)
462 		man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
463 	else
464 		return true;
465 
466 	if (!man) {
467 		if (domain & AMDGPU_GEM_DOMAIN_GTT)
468 			WARN_ON_ONCE("GTT domain requested but GTT mem manager uninitialized");
469 		return false;
470 	}
471 
472 	/* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU, _DOMAIN_DOORBELL */
473 	if (size < man->size)
474 		return true;
475 
476 	DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size, man->size);
477 	return false;
478 }
479 
amdgpu_bo_support_uswc(u64 bo_flags)480 bool amdgpu_bo_support_uswc(u64 bo_flags)
481 {
482 
483 #ifdef CONFIG_X86_32
484 	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
485 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
486 	 */
487 	return false;
488 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
489 	/* Don't try to enable write-combining when it can't work, or things
490 	 * may be slow
491 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
492 	 */
493 
494 #ifndef CONFIG_COMPILE_TEST
495 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
496 	 thanks to write-combining
497 #endif
498 
499 	if (bo_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
500 		DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
501 			      "better performance thanks to write-combining\n");
502 	return false;
503 #else
504 	/* For architectures that don't support WC memory,
505 	 * mask out the WC flag from the BO
506 	 */
507 	if (!drm_arch_can_wc_memory())
508 		return false;
509 
510 	return true;
511 #endif
512 }
513 
514 /**
515  * amdgpu_bo_create - create an &amdgpu_bo buffer object
516  * @adev: amdgpu device object
517  * @bp: parameters to be used for the buffer object
518  * @bo_ptr: pointer to the buffer object pointer
519  *
520  * Creates an &amdgpu_bo buffer object.
521  *
522  * Returns:
523  * 0 for success or a negative error code on failure.
524  */
amdgpu_bo_create(struct amdgpu_device * adev,struct amdgpu_bo_param * bp,struct amdgpu_bo ** bo_ptr)525 int amdgpu_bo_create(struct amdgpu_device *adev,
526 			       struct amdgpu_bo_param *bp,
527 			       struct amdgpu_bo **bo_ptr)
528 {
529 	struct ttm_operation_ctx ctx = {
530 		.interruptible = (bp->type != ttm_bo_type_kernel),
531 		.no_wait_gpu = bp->no_wait_gpu,
532 		/* We opt to avoid OOM on system pages allocations */
533 		.gfp_retry_mayfail = true,
534 		.allow_res_evict = bp->type != ttm_bo_type_kernel,
535 		.resv = bp->resv
536 	};
537 	struct amdgpu_bo *bo;
538 	unsigned long page_align, size = bp->size;
539 	int r;
540 
541 	/* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
542 	if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
543 		/* GWS and OA don't need any alignment. */
544 		page_align = bp->byte_align;
545 		size <<= PAGE_SHIFT;
546 
547 	} else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
548 		/* Both size and alignment must be a multiple of 4. */
549 		page_align = ALIGN(bp->byte_align, 4);
550 		size = ALIGN(size, 4) << PAGE_SHIFT;
551 	} else {
552 		/* Memory should be aligned at least to a page size. */
553 		page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
554 		size = ALIGN(size, PAGE_SIZE);
555 	}
556 
557 	if (!amdgpu_bo_validate_size(adev, size, bp->domain))
558 		return -ENOMEM;
559 
560 	BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo));
561 
562 	*bo_ptr = NULL;
563 	bo = kvzalloc(bp->bo_ptr_size, GFP_KERNEL);
564 	if (bo == NULL)
565 		return -ENOMEM;
566 	drm_gem_private_object_init(adev_to_drm(adev), &bo->tbo.base, size);
567 	bo->tbo.base.funcs = &amdgpu_gem_object_funcs;
568 	bo->vm_bo = NULL;
569 	bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
570 		bp->domain;
571 	bo->allowed_domains = bo->preferred_domains;
572 	if (bp->type != ttm_bo_type_kernel &&
573 	    !(bp->flags & AMDGPU_GEM_CREATE_DISCARDABLE) &&
574 	    bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
575 		bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
576 
577 	bo->flags = bp->flags;
578 
579 	if (adev->gmc.mem_partitions)
580 		/* For GPUs with spatial partitioning, bo->xcp_id=-1 means any partition */
581 		bo->xcp_id = bp->xcp_id_plus1 - 1;
582 	else
583 		/* For GPUs without spatial partitioning */
584 		bo->xcp_id = 0;
585 
586 	if (!amdgpu_bo_support_uswc(bo->flags))
587 		bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
588 
589 	bo->tbo.bdev = &adev->mman.bdev;
590 	if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
591 			  AMDGPU_GEM_DOMAIN_GDS))
592 		amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
593 	else
594 		amdgpu_bo_placement_from_domain(bo, bp->domain);
595 	if (bp->type == ttm_bo_type_kernel)
596 		bo->tbo.priority = 2;
597 	else if (!(bp->flags & AMDGPU_GEM_CREATE_DISCARDABLE))
598 		bo->tbo.priority = 1;
599 
600 	if (!bp->destroy)
601 		bp->destroy = &amdgpu_bo_destroy;
602 
603 	r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, bp->type,
604 				 &bo->placement, page_align, &ctx,  NULL,
605 				 bp->resv, bp->destroy);
606 	if (unlikely(r != 0))
607 		return r;
608 
609 	if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
610 	    amdgpu_res_cpu_visible(adev, bo->tbo.resource))
611 		amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
612 					     ctx.bytes_moved);
613 	else
614 		amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
615 
616 	if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
617 	    bo->tbo.resource->mem_type == TTM_PL_VRAM) {
618 		struct dma_fence *fence;
619 
620 		r = amdgpu_ttm_clear_buffer(bo, bo->tbo.base.resv, &fence);
621 		if (unlikely(r))
622 			goto fail_unreserve;
623 
624 		dma_resv_add_fence(bo->tbo.base.resv, fence,
625 				   DMA_RESV_USAGE_KERNEL);
626 		dma_fence_put(fence);
627 	}
628 	if (!bp->resv)
629 		amdgpu_bo_unreserve(bo);
630 	*bo_ptr = bo;
631 
632 	trace_amdgpu_bo_create(bo);
633 
634 	/* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
635 	if (bp->type == ttm_bo_type_device)
636 		bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
637 
638 	return 0;
639 
640 fail_unreserve:
641 	if (!bp->resv)
642 		dma_resv_unlock(bo->tbo.base.resv);
643 	amdgpu_bo_unref(&bo);
644 	return r;
645 }
646 
647 /**
648  * amdgpu_bo_create_user - create an &amdgpu_bo_user buffer object
649  * @adev: amdgpu device object
650  * @bp: parameters to be used for the buffer object
651  * @ubo_ptr: pointer to the buffer object pointer
652  *
653  * Create a BO to be used by user application;
654  *
655  * Returns:
656  * 0 for success or a negative error code on failure.
657  */
658 
amdgpu_bo_create_user(struct amdgpu_device * adev,struct amdgpu_bo_param * bp,struct amdgpu_bo_user ** ubo_ptr)659 int amdgpu_bo_create_user(struct amdgpu_device *adev,
660 			  struct amdgpu_bo_param *bp,
661 			  struct amdgpu_bo_user **ubo_ptr)
662 {
663 	struct amdgpu_bo *bo_ptr;
664 	int r;
665 
666 	bp->bo_ptr_size = sizeof(struct amdgpu_bo_user);
667 	bp->destroy = &amdgpu_bo_user_destroy;
668 	r = amdgpu_bo_create(adev, bp, &bo_ptr);
669 	if (r)
670 		return r;
671 
672 	*ubo_ptr = to_amdgpu_bo_user(bo_ptr);
673 	return r;
674 }
675 
676 /**
677  * amdgpu_bo_create_vm - create an &amdgpu_bo_vm buffer object
678  * @adev: amdgpu device object
679  * @bp: parameters to be used for the buffer object
680  * @vmbo_ptr: pointer to the buffer object pointer
681  *
682  * Create a BO to be for GPUVM.
683  *
684  * Returns:
685  * 0 for success or a negative error code on failure.
686  */
687 
amdgpu_bo_create_vm(struct amdgpu_device * adev,struct amdgpu_bo_param * bp,struct amdgpu_bo_vm ** vmbo_ptr)688 int amdgpu_bo_create_vm(struct amdgpu_device *adev,
689 			struct amdgpu_bo_param *bp,
690 			struct amdgpu_bo_vm **vmbo_ptr)
691 {
692 	struct amdgpu_bo *bo_ptr;
693 	int r;
694 
695 	/* bo_ptr_size will be determined by the caller and it depends on
696 	 * num of amdgpu_vm_pt entries.
697 	 */
698 	BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo_vm));
699 	r = amdgpu_bo_create(adev, bp, &bo_ptr);
700 	if (r)
701 		return r;
702 
703 	*vmbo_ptr = to_amdgpu_bo_vm(bo_ptr);
704 	return r;
705 }
706 
707 /**
708  * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
709  * @bo: &amdgpu_bo buffer object to be mapped
710  * @ptr: kernel virtual address to be returned
711  *
712  * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
713  * amdgpu_bo_kptr() to get the kernel virtual address.
714  *
715  * Returns:
716  * 0 for success or a negative error code on failure.
717  */
amdgpu_bo_kmap(struct amdgpu_bo * bo,void ** ptr)718 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
719 {
720 	void *kptr;
721 	long r;
722 
723 	if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
724 		return -EPERM;
725 
726 	r = dma_resv_wait_timeout(bo->tbo.base.resv, DMA_RESV_USAGE_KERNEL,
727 				  false, MAX_SCHEDULE_TIMEOUT);
728 	if (r < 0)
729 		return r;
730 
731 	kptr = amdgpu_bo_kptr(bo);
732 	if (kptr) {
733 		if (ptr)
734 			*ptr = kptr;
735 		return 0;
736 	}
737 
738 	r = ttm_bo_kmap(&bo->tbo, 0, PFN_UP(bo->tbo.base.size), &bo->kmap);
739 	if (r)
740 		return r;
741 
742 	if (ptr)
743 		*ptr = amdgpu_bo_kptr(bo);
744 
745 	return 0;
746 }
747 
748 /**
749  * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
750  * @bo: &amdgpu_bo buffer object
751  *
752  * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
753  *
754  * Returns:
755  * the virtual address of a buffer object area.
756  */
amdgpu_bo_kptr(struct amdgpu_bo * bo)757 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
758 {
759 	bool is_iomem;
760 
761 	return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
762 }
763 
764 /**
765  * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
766  * @bo: &amdgpu_bo buffer object to be unmapped
767  *
768  * Unmaps a kernel map set up by amdgpu_bo_kmap().
769  */
amdgpu_bo_kunmap(struct amdgpu_bo * bo)770 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
771 {
772 	if (bo->kmap.bo)
773 		ttm_bo_kunmap(&bo->kmap);
774 }
775 
776 /**
777  * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
778  * @bo: &amdgpu_bo buffer object
779  *
780  * References the contained &ttm_buffer_object.
781  *
782  * Returns:
783  * a refcounted pointer to the &amdgpu_bo buffer object.
784  */
amdgpu_bo_ref(struct amdgpu_bo * bo)785 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
786 {
787 	if (bo == NULL)
788 		return NULL;
789 
790 	drm_gem_object_get(&bo->tbo.base);
791 	return bo;
792 }
793 
794 /**
795  * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
796  * @bo: &amdgpu_bo buffer object
797  *
798  * Unreferences the contained &ttm_buffer_object and clear the pointer
799  */
amdgpu_bo_unref(struct amdgpu_bo ** bo)800 void amdgpu_bo_unref(struct amdgpu_bo **bo)
801 {
802 	if ((*bo) == NULL)
803 		return;
804 
805 	drm_gem_object_put(&(*bo)->tbo.base);
806 	*bo = NULL;
807 }
808 
809 /**
810  * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
811  * @bo: &amdgpu_bo buffer object to be pinned
812  * @domain: domain to be pinned to
813  *
814  * Pins the buffer object according to requested domain. If the memory is
815  * unbound gart memory, binds the pages into gart table. Adjusts pin_count and
816  * pin_size accordingly.
817  *
818  * Pinning means to lock pages in memory along with keeping them at a fixed
819  * offset. It is required when a buffer can not be moved, for example, when
820  * a display buffer is being scanned out.
821  *
822  * Returns:
823  * 0 for success or a negative error code on failure.
824  */
amdgpu_bo_pin(struct amdgpu_bo * bo,u32 domain)825 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
826 {
827 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
828 	struct ttm_operation_ctx ctx = { false, false };
829 	int r, i;
830 
831 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
832 		return -EPERM;
833 
834 	/* Check domain to be pinned to against preferred domains */
835 	if (bo->preferred_domains & domain)
836 		domain = bo->preferred_domains & domain;
837 
838 	/* A shared bo cannot be migrated to VRAM */
839 	if (bo->tbo.base.import_attach) {
840 		if (domain & AMDGPU_GEM_DOMAIN_GTT)
841 			domain = AMDGPU_GEM_DOMAIN_GTT;
842 		else
843 			return -EINVAL;
844 	}
845 
846 	if (bo->tbo.pin_count) {
847 		uint32_t mem_type = bo->tbo.resource->mem_type;
848 		uint32_t mem_flags = bo->tbo.resource->placement;
849 
850 		if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
851 			return -EINVAL;
852 
853 		if ((mem_type == TTM_PL_VRAM) &&
854 		    (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) &&
855 		    !(mem_flags & TTM_PL_FLAG_CONTIGUOUS))
856 			return -EINVAL;
857 
858 		ttm_bo_pin(&bo->tbo);
859 		return 0;
860 	}
861 
862 	/* This assumes only APU display buffers are pinned with (VRAM|GTT).
863 	 * See function amdgpu_display_supported_domains()
864 	 */
865 	domain = amdgpu_bo_get_preferred_domain(adev, domain);
866 
867 	if (bo->tbo.base.import_attach)
868 		dma_buf_pin(bo->tbo.base.import_attach);
869 
870 	/* force to pin into visible video ram */
871 	if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
872 		bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
873 	amdgpu_bo_placement_from_domain(bo, domain);
874 	for (i = 0; i < bo->placement.num_placement; i++) {
875 		if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS &&
876 		    bo->placements[i].mem_type == TTM_PL_VRAM)
877 			bo->placements[i].flags |= TTM_PL_FLAG_CONTIGUOUS;
878 	}
879 
880 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
881 	if (unlikely(r)) {
882 		dev_err(adev->dev, "%p pin failed\n", bo);
883 		goto error;
884 	}
885 
886 	ttm_bo_pin(&bo->tbo);
887 
888 	if (bo->tbo.resource->mem_type == TTM_PL_VRAM) {
889 		atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
890 		atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
891 			     &adev->visible_pin_size);
892 	} else if (bo->tbo.resource->mem_type == TTM_PL_TT) {
893 		atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
894 	}
895 
896 error:
897 	return r;
898 }
899 
900 /**
901  * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
902  * @bo: &amdgpu_bo buffer object to be unpinned
903  *
904  * Decreases the pin_count, and clears the flags if pin_count reaches 0.
905  * Changes placement and pin size accordingly.
906  *
907  * Returns:
908  * 0 for success or a negative error code on failure.
909  */
amdgpu_bo_unpin(struct amdgpu_bo * bo)910 void amdgpu_bo_unpin(struct amdgpu_bo *bo)
911 {
912 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
913 
914 	ttm_bo_unpin(&bo->tbo);
915 	if (bo->tbo.pin_count)
916 		return;
917 
918 	if (bo->tbo.base.import_attach)
919 		dma_buf_unpin(bo->tbo.base.import_attach);
920 
921 	if (bo->tbo.resource->mem_type == TTM_PL_VRAM) {
922 		atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
923 		atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
924 			     &adev->visible_pin_size);
925 	} else if (bo->tbo.resource->mem_type == TTM_PL_TT) {
926 		atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size);
927 	}
928 
929 }
930 
931 static const char * const amdgpu_vram_names[] = {
932 	"UNKNOWN",
933 	"GDDR1",
934 	"DDR2",
935 	"GDDR3",
936 	"GDDR4",
937 	"GDDR5",
938 	"HBM",
939 	"DDR3",
940 	"DDR4",
941 	"GDDR6",
942 	"DDR5",
943 	"LPDDR4",
944 	"LPDDR5"
945 };
946 
947 /**
948  * amdgpu_bo_init - initialize memory manager
949  * @adev: amdgpu device object
950  *
951  * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
952  *
953  * Returns:
954  * 0 for success or a negative error code on failure.
955  */
amdgpu_bo_init(struct amdgpu_device * adev)956 int amdgpu_bo_init(struct amdgpu_device *adev)
957 {
958 	/* On A+A platform, VRAM can be mapped as WB */
959 	if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) {
960 		/* reserve PAT memory space to WC for VRAM */
961 		int r = arch_io_reserve_memtype_wc(adev->gmc.aper_base,
962 				adev->gmc.aper_size);
963 
964 		if (r) {
965 			DRM_ERROR("Unable to set WC memtype for the aperture base\n");
966 			return r;
967 		}
968 
969 		/* Add an MTRR for the VRAM */
970 		adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
971 				adev->gmc.aper_size);
972 	}
973 
974 	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
975 		 adev->gmc.mc_vram_size >> 20,
976 		 (unsigned long long)adev->gmc.aper_size >> 20);
977 	DRM_INFO("RAM width %dbits %s\n",
978 		 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
979 	return amdgpu_ttm_init(adev);
980 }
981 
982 /**
983  * amdgpu_bo_fini - tear down memory manager
984  * @adev: amdgpu device object
985  *
986  * Reverses amdgpu_bo_init() to tear down memory manager.
987  */
amdgpu_bo_fini(struct amdgpu_device * adev)988 void amdgpu_bo_fini(struct amdgpu_device *adev)
989 {
990 	int idx;
991 
992 	amdgpu_ttm_fini(adev);
993 
994 	if (drm_dev_enter(adev_to_drm(adev), &idx)) {
995 		if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) {
996 			arch_phys_wc_del(adev->gmc.vram_mtrr);
997 			arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
998 		}
999 		drm_dev_exit(idx);
1000 	}
1001 }
1002 
1003 /**
1004  * amdgpu_bo_set_tiling_flags - set tiling flags
1005  * @bo: &amdgpu_bo buffer object
1006  * @tiling_flags: new flags
1007  *
1008  * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1009  * kernel driver to set the tiling flags on a buffer.
1010  *
1011  * Returns:
1012  * 0 for success or a negative error code on failure.
1013  */
amdgpu_bo_set_tiling_flags(struct amdgpu_bo * bo,u64 tiling_flags)1014 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1015 {
1016 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1017 	struct amdgpu_bo_user *ubo;
1018 
1019 	BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1020 	if (adev->family <= AMDGPU_FAMILY_CZ &&
1021 	    AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1022 		return -EINVAL;
1023 
1024 	ubo = to_amdgpu_bo_user(bo);
1025 	ubo->tiling_flags = tiling_flags;
1026 	return 0;
1027 }
1028 
1029 /**
1030  * amdgpu_bo_get_tiling_flags - get tiling flags
1031  * @bo: &amdgpu_bo buffer object
1032  * @tiling_flags: returned flags
1033  *
1034  * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1035  * set the tiling flags on a buffer.
1036  */
amdgpu_bo_get_tiling_flags(struct amdgpu_bo * bo,u64 * tiling_flags)1037 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1038 {
1039 	struct amdgpu_bo_user *ubo;
1040 
1041 	BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1042 	dma_resv_assert_held(bo->tbo.base.resv);
1043 	ubo = to_amdgpu_bo_user(bo);
1044 
1045 	if (tiling_flags)
1046 		*tiling_flags = ubo->tiling_flags;
1047 }
1048 
1049 /**
1050  * amdgpu_bo_set_metadata - set metadata
1051  * @bo: &amdgpu_bo buffer object
1052  * @metadata: new metadata
1053  * @metadata_size: size of the new metadata
1054  * @flags: flags of the new metadata
1055  *
1056  * Sets buffer object's metadata, its size and flags.
1057  * Used via GEM ioctl.
1058  *
1059  * Returns:
1060  * 0 for success or a negative error code on failure.
1061  */
amdgpu_bo_set_metadata(struct amdgpu_bo * bo,void * metadata,u32 metadata_size,uint64_t flags)1062 int amdgpu_bo_set_metadata(struct amdgpu_bo *bo, void *metadata,
1063 			   u32 metadata_size, uint64_t flags)
1064 {
1065 	struct amdgpu_bo_user *ubo;
1066 	void *buffer;
1067 
1068 	BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1069 	ubo = to_amdgpu_bo_user(bo);
1070 	if (!metadata_size) {
1071 		if (ubo->metadata_size) {
1072 			kfree(ubo->metadata);
1073 			ubo->metadata = NULL;
1074 			ubo->metadata_size = 0;
1075 		}
1076 		return 0;
1077 	}
1078 
1079 	if (metadata == NULL)
1080 		return -EINVAL;
1081 
1082 	buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1083 	if (buffer == NULL)
1084 		return -ENOMEM;
1085 
1086 	kfree(ubo->metadata);
1087 	ubo->metadata_flags = flags;
1088 	ubo->metadata = buffer;
1089 	ubo->metadata_size = metadata_size;
1090 
1091 	return 0;
1092 }
1093 
1094 /**
1095  * amdgpu_bo_get_metadata - get metadata
1096  * @bo: &amdgpu_bo buffer object
1097  * @buffer: returned metadata
1098  * @buffer_size: size of the buffer
1099  * @metadata_size: size of the returned metadata
1100  * @flags: flags of the returned metadata
1101  *
1102  * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1103  * less than metadata_size.
1104  * Used via GEM ioctl.
1105  *
1106  * Returns:
1107  * 0 for success or a negative error code on failure.
1108  */
amdgpu_bo_get_metadata(struct amdgpu_bo * bo,void * buffer,size_t buffer_size,uint32_t * metadata_size,uint64_t * flags)1109 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1110 			   size_t buffer_size, uint32_t *metadata_size,
1111 			   uint64_t *flags)
1112 {
1113 	struct amdgpu_bo_user *ubo;
1114 
1115 	if (!buffer && !metadata_size)
1116 		return -EINVAL;
1117 
1118 	BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1119 	ubo = to_amdgpu_bo_user(bo);
1120 	if (metadata_size)
1121 		*metadata_size = ubo->metadata_size;
1122 
1123 	if (buffer) {
1124 		if (buffer_size < ubo->metadata_size)
1125 			return -EINVAL;
1126 
1127 		if (ubo->metadata_size)
1128 			memcpy(buffer, ubo->metadata, ubo->metadata_size);
1129 	}
1130 
1131 	if (flags)
1132 		*flags = ubo->metadata_flags;
1133 
1134 	return 0;
1135 }
1136 
1137 /**
1138  * amdgpu_bo_move_notify - notification about a memory move
1139  * @bo: pointer to a buffer object
1140  * @evict: if this move is evicting the buffer from the graphics address space
1141  * @new_mem: new resource for backing the BO
1142  *
1143  * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1144  * bookkeeping.
1145  * TTM driver callback which is called when ttm moves a buffer.
1146  */
amdgpu_bo_move_notify(struct ttm_buffer_object * bo,bool evict,struct ttm_resource * new_mem)1147 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1148 			   bool evict,
1149 			   struct ttm_resource *new_mem)
1150 {
1151 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1152 	struct ttm_resource *old_mem = bo->resource;
1153 	struct amdgpu_bo *abo;
1154 
1155 	if (!amdgpu_bo_is_amdgpu_bo(bo))
1156 		return;
1157 
1158 	abo = ttm_to_amdgpu_bo(bo);
1159 	amdgpu_vm_bo_invalidate(adev, abo, evict);
1160 
1161 	amdgpu_bo_kunmap(abo);
1162 
1163 	if (abo->tbo.base.dma_buf && !abo->tbo.base.import_attach &&
1164 	    old_mem && old_mem->mem_type != TTM_PL_SYSTEM)
1165 		dma_buf_move_notify(abo->tbo.base.dma_buf);
1166 
1167 	/* move_notify is called before move happens */
1168 	trace_amdgpu_bo_move(abo, new_mem ? new_mem->mem_type : -1,
1169 			     old_mem ? old_mem->mem_type : -1);
1170 }
1171 
amdgpu_bo_get_memory(struct amdgpu_bo * bo,struct amdgpu_mem_stats * stats)1172 void amdgpu_bo_get_memory(struct amdgpu_bo *bo,
1173 			  struct amdgpu_mem_stats *stats)
1174 {
1175 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1176 	struct ttm_resource *res = bo->tbo.resource;
1177 	uint64_t size = amdgpu_bo_size(bo);
1178 	struct drm_gem_object *obj;
1179 	bool shared;
1180 
1181 	/* Abort if the BO doesn't currently have a backing store */
1182 	if (!res)
1183 		return;
1184 
1185 	obj = &bo->tbo.base;
1186 	shared = drm_gem_object_is_shared_for_memory_stats(obj);
1187 
1188 	switch (res->mem_type) {
1189 	case TTM_PL_VRAM:
1190 		stats->vram += size;
1191 		if (amdgpu_res_cpu_visible(adev, res))
1192 			stats->visible_vram += size;
1193 		if (shared)
1194 			stats->vram_shared += size;
1195 		break;
1196 	case TTM_PL_TT:
1197 		stats->gtt += size;
1198 		if (shared)
1199 			stats->gtt_shared += size;
1200 		break;
1201 	case TTM_PL_SYSTEM:
1202 	default:
1203 		stats->cpu += size;
1204 		if (shared)
1205 			stats->cpu_shared += size;
1206 		break;
1207 	}
1208 
1209 	if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) {
1210 		stats->requested_vram += size;
1211 		if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
1212 			stats->requested_visible_vram += size;
1213 
1214 		if (res->mem_type != TTM_PL_VRAM) {
1215 			stats->evicted_vram += size;
1216 			if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
1217 				stats->evicted_visible_vram += size;
1218 		}
1219 	} else if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_GTT) {
1220 		stats->requested_gtt += size;
1221 	}
1222 }
1223 
1224 /**
1225  * amdgpu_bo_release_notify - notification about a BO being released
1226  * @bo: pointer to a buffer object
1227  *
1228  * Wipes VRAM buffers whose contents should not be leaked before the
1229  * memory is released.
1230  */
amdgpu_bo_release_notify(struct ttm_buffer_object * bo)1231 void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)
1232 {
1233 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1234 	struct dma_fence *fence = NULL;
1235 	struct amdgpu_bo *abo;
1236 	int r;
1237 
1238 	if (!amdgpu_bo_is_amdgpu_bo(bo))
1239 		return;
1240 
1241 	abo = ttm_to_amdgpu_bo(bo);
1242 
1243 	WARN_ON(abo->vm_bo);
1244 
1245 	if (abo->kfd_bo)
1246 		amdgpu_amdkfd_release_notify(abo);
1247 
1248 	/* We only remove the fence if the resv has individualized. */
1249 	WARN_ON_ONCE(bo->type == ttm_bo_type_kernel
1250 			&& bo->base.resv != &bo->base._resv);
1251 	if (bo->base.resv == &bo->base._resv)
1252 		amdgpu_amdkfd_remove_fence_on_pt_pd_bos(abo);
1253 
1254 	if (!bo->resource || bo->resource->mem_type != TTM_PL_VRAM ||
1255 	    !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE) ||
1256 	    adev->in_suspend || drm_dev_is_unplugged(adev_to_drm(adev)))
1257 		return;
1258 
1259 	if (WARN_ON_ONCE(!dma_resv_trylock(bo->base.resv)))
1260 		return;
1261 
1262 	r = amdgpu_fill_buffer(abo, 0, bo->base.resv, &fence, true);
1263 	if (!WARN_ON(r)) {
1264 		amdgpu_vram_mgr_set_cleared(bo->resource);
1265 		amdgpu_bo_fence(abo, fence, false);
1266 		dma_fence_put(fence);
1267 	}
1268 
1269 	dma_resv_unlock(bo->base.resv);
1270 }
1271 
1272 /**
1273  * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1274  * @bo: pointer to a buffer object
1275  *
1276  * Notifies the driver we are taking a fault on this BO and have reserved it,
1277  * also performs bookkeeping.
1278  * TTM driver callback for dealing with vm faults.
1279  *
1280  * Returns:
1281  * 0 for success or a negative error code on failure.
1282  */
amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object * bo)1283 vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1284 {
1285 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1286 	struct ttm_operation_ctx ctx = { false, false };
1287 	struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1288 	int r;
1289 
1290 	/* Remember that this BO was accessed by the CPU */
1291 	abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1292 
1293 	if (amdgpu_res_cpu_visible(adev, bo->resource))
1294 		return 0;
1295 
1296 	/* Can't move a pinned BO to visible VRAM */
1297 	if (abo->tbo.pin_count > 0)
1298 		return VM_FAULT_SIGBUS;
1299 
1300 	/* hurrah the memory is not visible ! */
1301 	atomic64_inc(&adev->num_vram_cpu_page_faults);
1302 	amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1303 					AMDGPU_GEM_DOMAIN_GTT);
1304 
1305 	/* Avoid costly evictions; only set GTT as a busy placement */
1306 	abo->placements[0].flags |= TTM_PL_FLAG_DESIRED;
1307 
1308 	r = ttm_bo_validate(bo, &abo->placement, &ctx);
1309 	if (unlikely(r == -EBUSY || r == -ERESTARTSYS))
1310 		return VM_FAULT_NOPAGE;
1311 	else if (unlikely(r))
1312 		return VM_FAULT_SIGBUS;
1313 
1314 	/* this should never happen */
1315 	if (bo->resource->mem_type == TTM_PL_VRAM &&
1316 	    !amdgpu_res_cpu_visible(adev, bo->resource))
1317 		return VM_FAULT_SIGBUS;
1318 
1319 	ttm_bo_move_to_lru_tail_unlocked(bo);
1320 	return 0;
1321 }
1322 
1323 /**
1324  * amdgpu_bo_fence - add fence to buffer object
1325  *
1326  * @bo: buffer object in question
1327  * @fence: fence to add
1328  * @shared: true if fence should be added shared
1329  *
1330  */
amdgpu_bo_fence(struct amdgpu_bo * bo,struct dma_fence * fence,bool shared)1331 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1332 		     bool shared)
1333 {
1334 	struct dma_resv *resv = bo->tbo.base.resv;
1335 	int r;
1336 
1337 	r = dma_resv_reserve_fences(resv, 1);
1338 	if (r) {
1339 		/* As last resort on OOM we block for the fence */
1340 		dma_fence_wait(fence, false);
1341 		return;
1342 	}
1343 
1344 	dma_resv_add_fence(resv, fence, shared ? DMA_RESV_USAGE_READ :
1345 			   DMA_RESV_USAGE_WRITE);
1346 }
1347 
1348 /**
1349  * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences
1350  *
1351  * @adev: amdgpu device pointer
1352  * @resv: reservation object to sync to
1353  * @sync_mode: synchronization mode
1354  * @owner: fence owner
1355  * @intr: Whether the wait is interruptible
1356  *
1357  * Extract the fences from the reservation object and waits for them to finish.
1358  *
1359  * Returns:
1360  * 0 on success, errno otherwise.
1361  */
amdgpu_bo_sync_wait_resv(struct amdgpu_device * adev,struct dma_resv * resv,enum amdgpu_sync_mode sync_mode,void * owner,bool intr)1362 int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv,
1363 			     enum amdgpu_sync_mode sync_mode, void *owner,
1364 			     bool intr)
1365 {
1366 	struct amdgpu_sync sync;
1367 	int r;
1368 
1369 	amdgpu_sync_create(&sync);
1370 	amdgpu_sync_resv(adev, &sync, resv, sync_mode, owner);
1371 	r = amdgpu_sync_wait(&sync, intr);
1372 	amdgpu_sync_free(&sync);
1373 	return r;
1374 }
1375 
1376 /**
1377  * amdgpu_bo_sync_wait - Wrapper for amdgpu_bo_sync_wait_resv
1378  * @bo: buffer object to wait for
1379  * @owner: fence owner
1380  * @intr: Whether the wait is interruptible
1381  *
1382  * Wrapper to wait for fences in a BO.
1383  * Returns:
1384  * 0 on success, errno otherwise.
1385  */
amdgpu_bo_sync_wait(struct amdgpu_bo * bo,void * owner,bool intr)1386 int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)
1387 {
1388 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1389 
1390 	return amdgpu_bo_sync_wait_resv(adev, bo->tbo.base.resv,
1391 					AMDGPU_SYNC_NE_OWNER, owner, intr);
1392 }
1393 
1394 /**
1395  * amdgpu_bo_gpu_offset - return GPU offset of bo
1396  * @bo:	amdgpu object for which we query the offset
1397  *
1398  * Note: object should either be pinned or reserved when calling this
1399  * function, it might be useful to add check for this for debugging.
1400  *
1401  * Returns:
1402  * current GPU offset of the object.
1403  */
amdgpu_bo_gpu_offset(struct amdgpu_bo * bo)1404 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1405 {
1406 	WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_SYSTEM);
1407 	WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) &&
1408 		     !bo->tbo.pin_count && bo->tbo.type != ttm_bo_type_kernel);
1409 	WARN_ON_ONCE(bo->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET);
1410 	WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_VRAM &&
1411 		     !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1412 
1413 	return amdgpu_bo_gpu_offset_no_check(bo);
1414 }
1415 
1416 /**
1417  * amdgpu_bo_gpu_offset_no_check - return GPU offset of bo
1418  * @bo:	amdgpu object for which we query the offset
1419  *
1420  * Returns:
1421  * current GPU offset of the object without raising warnings.
1422  */
amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo * bo)1423 u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo)
1424 {
1425 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1426 	uint64_t offset = AMDGPU_BO_INVALID_OFFSET;
1427 
1428 	if (bo->tbo.resource->mem_type == TTM_PL_TT)
1429 		offset = amdgpu_gmc_agp_addr(&bo->tbo);
1430 
1431 	if (offset == AMDGPU_BO_INVALID_OFFSET)
1432 		offset = (bo->tbo.resource->start << PAGE_SHIFT) +
1433 			amdgpu_ttm_domain_start(adev, bo->tbo.resource->mem_type);
1434 
1435 	return amdgpu_gmc_sign_extend(offset);
1436 }
1437 
1438 /**
1439  * amdgpu_bo_get_preferred_domain - get preferred domain
1440  * @adev: amdgpu device object
1441  * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1442  *
1443  * Returns:
1444  * Which of the allowed domains is preferred for allocating the BO.
1445  */
amdgpu_bo_get_preferred_domain(struct amdgpu_device * adev,uint32_t domain)1446 uint32_t amdgpu_bo_get_preferred_domain(struct amdgpu_device *adev,
1447 					    uint32_t domain)
1448 {
1449 	if ((domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) &&
1450 	    ((adev->asic_type == CHIP_CARRIZO) || (adev->asic_type == CHIP_STONEY))) {
1451 		domain = AMDGPU_GEM_DOMAIN_VRAM;
1452 		if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1453 			domain = AMDGPU_GEM_DOMAIN_GTT;
1454 	}
1455 	return domain;
1456 }
1457 
1458 #if defined(CONFIG_DEBUG_FS)
1459 #define amdgpu_bo_print_flag(m, bo, flag)		        \
1460 	do {							\
1461 		if (bo->flags & (AMDGPU_GEM_CREATE_ ## flag)) {	\
1462 			seq_printf((m), " " #flag);		\
1463 		}						\
1464 	} while (0)
1465 
1466 /**
1467  * amdgpu_bo_print_info - print BO info in debugfs file
1468  *
1469  * @id: Index or Id of the BO
1470  * @bo: Requested BO for printing info
1471  * @m: debugfs file
1472  *
1473  * Print BO information in debugfs file
1474  *
1475  * Returns:
1476  * Size of the BO in bytes.
1477  */
amdgpu_bo_print_info(int id,struct amdgpu_bo * bo,struct seq_file * m)1478 u64 amdgpu_bo_print_info(int id, struct amdgpu_bo *bo, struct seq_file *m)
1479 {
1480 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1481 	struct dma_buf_attachment *attachment;
1482 	struct dma_buf *dma_buf;
1483 	const char *placement;
1484 	unsigned int pin_count;
1485 	u64 size;
1486 
1487 	if (dma_resv_trylock(bo->tbo.base.resv)) {
1488 		if (!bo->tbo.resource) {
1489 			placement = "NONE";
1490 		} else {
1491 			switch (bo->tbo.resource->mem_type) {
1492 			case TTM_PL_VRAM:
1493 				if (amdgpu_res_cpu_visible(adev, bo->tbo.resource))
1494 					placement = "VRAM VISIBLE";
1495 				else
1496 					placement = "VRAM";
1497 				break;
1498 			case TTM_PL_TT:
1499 				placement = "GTT";
1500 				break;
1501 			case AMDGPU_PL_GDS:
1502 				placement = "GDS";
1503 				break;
1504 			case AMDGPU_PL_GWS:
1505 				placement = "GWS";
1506 				break;
1507 			case AMDGPU_PL_OA:
1508 				placement = "OA";
1509 				break;
1510 			case AMDGPU_PL_PREEMPT:
1511 				placement = "PREEMPTIBLE";
1512 				break;
1513 			case AMDGPU_PL_DOORBELL:
1514 				placement = "DOORBELL";
1515 				break;
1516 			case TTM_PL_SYSTEM:
1517 			default:
1518 				placement = "CPU";
1519 				break;
1520 			}
1521 		}
1522 		dma_resv_unlock(bo->tbo.base.resv);
1523 	} else {
1524 		placement = "UNKNOWN";
1525 	}
1526 
1527 	size = amdgpu_bo_size(bo);
1528 	seq_printf(m, "\t\t0x%08x: %12lld byte %s",
1529 			id, size, placement);
1530 
1531 	pin_count = READ_ONCE(bo->tbo.pin_count);
1532 	if (pin_count)
1533 		seq_printf(m, " pin count %d", pin_count);
1534 
1535 	dma_buf = READ_ONCE(bo->tbo.base.dma_buf);
1536 	attachment = READ_ONCE(bo->tbo.base.import_attach);
1537 
1538 	if (attachment)
1539 		seq_printf(m, " imported from ino:%lu", file_inode(dma_buf->file)->i_ino);
1540 	else if (dma_buf)
1541 		seq_printf(m, " exported as ino:%lu", file_inode(dma_buf->file)->i_ino);
1542 
1543 	amdgpu_bo_print_flag(m, bo, CPU_ACCESS_REQUIRED);
1544 	amdgpu_bo_print_flag(m, bo, NO_CPU_ACCESS);
1545 	amdgpu_bo_print_flag(m, bo, CPU_GTT_USWC);
1546 	amdgpu_bo_print_flag(m, bo, VRAM_CLEARED);
1547 	amdgpu_bo_print_flag(m, bo, VRAM_CONTIGUOUS);
1548 	amdgpu_bo_print_flag(m, bo, VM_ALWAYS_VALID);
1549 	amdgpu_bo_print_flag(m, bo, EXPLICIT_SYNC);
1550 
1551 	seq_puts(m, "\n");
1552 
1553 	return size;
1554 }
1555 #endif
1556