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