]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blob - drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
8714d50c5b20280a70e6ca6418b05f99bb250397
[mirror_ubuntu-jammy-kernel.git] / drivers / gpu / drm / amd / amdgpu / amdgpu_object.c
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
55 static void amdgpu_bo_destroy(struct ttm_buffer_object *tbo)
56 {
57 struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
58 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
59 struct amdgpu_bo_user *ubo;
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 /* in case amdgpu_device_recover_vram got NULL of bo->parent */
67 if (!list_empty(&bo->shadow_list)) {
68 mutex_lock(&adev->shadow_list_lock);
69 list_del_init(&bo->shadow_list);
70 mutex_unlock(&adev->shadow_list_lock);
71 }
72 amdgpu_bo_unref(&bo->parent);
73
74 if (bo->tbo.type == ttm_bo_type_device) {
75 ubo = to_amdgpu_bo_user(bo);
76 kfree(ubo->metadata);
77 }
78
79 kfree(bo);
80 }
81
82 /**
83 * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
84 * @bo: buffer object to be checked
85 *
86 * Uses destroy function associated with the object to determine if this is
87 * an &amdgpu_bo.
88 *
89 * Returns:
90 * true if the object belongs to &amdgpu_bo, false if not.
91 */
92 bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
93 {
94 if (bo->destroy == &amdgpu_bo_destroy)
95 return true;
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 */
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 visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
117
118 places[c].fpfn = 0;
119 places[c].lpfn = 0;
120 places[c].mem_type = TTM_PL_VRAM;
121 places[c].flags = 0;
122
123 if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
124 places[c].lpfn = visible_pfn;
125 else
126 places[c].flags |= TTM_PL_FLAG_TOPDOWN;
127
128 if (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
129 places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
130 c++;
131 }
132
133 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
134 places[c].fpfn = 0;
135 places[c].lpfn = 0;
136 places[c].mem_type = TTM_PL_TT;
137 places[c].flags = 0;
138 c++;
139 }
140
141 if (domain & AMDGPU_GEM_DOMAIN_CPU) {
142 places[c].fpfn = 0;
143 places[c].lpfn = 0;
144 places[c].mem_type = TTM_PL_SYSTEM;
145 places[c].flags = 0;
146 c++;
147 }
148
149 if (domain & AMDGPU_GEM_DOMAIN_GDS) {
150 places[c].fpfn = 0;
151 places[c].lpfn = 0;
152 places[c].mem_type = AMDGPU_PL_GDS;
153 places[c].flags = 0;
154 c++;
155 }
156
157 if (domain & AMDGPU_GEM_DOMAIN_GWS) {
158 places[c].fpfn = 0;
159 places[c].lpfn = 0;
160 places[c].mem_type = AMDGPU_PL_GWS;
161 places[c].flags = 0;
162 c++;
163 }
164
165 if (domain & AMDGPU_GEM_DOMAIN_OA) {
166 places[c].fpfn = 0;
167 places[c].lpfn = 0;
168 places[c].mem_type = AMDGPU_PL_OA;
169 places[c].flags = 0;
170 c++;
171 }
172
173 if (!c) {
174 places[c].fpfn = 0;
175 places[c].lpfn = 0;
176 places[c].mem_type = TTM_PL_SYSTEM;
177 places[c].flags = 0;
178 c++;
179 }
180
181 BUG_ON(c >= AMDGPU_BO_MAX_PLACEMENTS);
182
183 placement->num_placement = c;
184 placement->placement = places;
185
186 placement->num_busy_placement = c;
187 placement->busy_placement = places;
188 }
189
190 /**
191 * amdgpu_bo_create_reserved - create reserved BO for kernel use
192 *
193 * @adev: amdgpu device object
194 * @size: size for the new BO
195 * @align: alignment for the new BO
196 * @domain: where to place it
197 * @bo_ptr: used to initialize BOs in structures
198 * @gpu_addr: GPU addr of the pinned BO
199 * @cpu_addr: optional CPU address mapping
200 *
201 * Allocates and pins a BO for kernel internal use, and returns it still
202 * reserved.
203 *
204 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
205 *
206 * Returns:
207 * 0 on success, negative error code otherwise.
208 */
209 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
210 unsigned long size, int align,
211 u32 domain, struct amdgpu_bo **bo_ptr,
212 u64 *gpu_addr, void **cpu_addr)
213 {
214 struct amdgpu_bo_param bp;
215 bool free = false;
216 int r;
217
218 if (!size) {
219 amdgpu_bo_unref(bo_ptr);
220 return 0;
221 }
222
223 memset(&bp, 0, sizeof(bp));
224 bp.size = size;
225 bp.byte_align = align;
226 bp.domain = domain;
227 bp.flags = cpu_addr ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
228 : AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
229 bp.flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
230 bp.type = ttm_bo_type_kernel;
231 bp.resv = NULL;
232 bp.bo_ptr_size = sizeof(struct amdgpu_bo);
233
234 if (!*bo_ptr) {
235 r = amdgpu_bo_create(adev, &bp, bo_ptr);
236 if (r) {
237 dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
238 r);
239 return r;
240 }
241 free = true;
242 }
243
244 r = amdgpu_bo_reserve(*bo_ptr, false);
245 if (r) {
246 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
247 goto error_free;
248 }
249
250 r = amdgpu_bo_pin(*bo_ptr, domain);
251 if (r) {
252 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
253 goto error_unreserve;
254 }
255
256 r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
257 if (r) {
258 dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
259 goto error_unpin;
260 }
261
262 if (gpu_addr)
263 *gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
264
265 if (cpu_addr) {
266 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
267 if (r) {
268 dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
269 goto error_unpin;
270 }
271 }
272
273 return 0;
274
275 error_unpin:
276 amdgpu_bo_unpin(*bo_ptr);
277 error_unreserve:
278 amdgpu_bo_unreserve(*bo_ptr);
279
280 error_free:
281 if (free)
282 amdgpu_bo_unref(bo_ptr);
283
284 return r;
285 }
286
287 /**
288 * amdgpu_bo_create_kernel - create BO for kernel use
289 *
290 * @adev: amdgpu device object
291 * @size: size for the new BO
292 * @align: alignment for the new BO
293 * @domain: where to place it
294 * @bo_ptr: used to initialize BOs in structures
295 * @gpu_addr: GPU addr of the pinned BO
296 * @cpu_addr: optional CPU address mapping
297 *
298 * Allocates and pins a BO for kernel internal use.
299 *
300 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
301 *
302 * Returns:
303 * 0 on success, negative error code otherwise.
304 */
305 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
306 unsigned long size, int align,
307 u32 domain, struct amdgpu_bo **bo_ptr,
308 u64 *gpu_addr, void **cpu_addr)
309 {
310 int r;
311
312 r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
313 gpu_addr, cpu_addr);
314
315 if (r)
316 return r;
317
318 if (*bo_ptr)
319 amdgpu_bo_unreserve(*bo_ptr);
320
321 return 0;
322 }
323
324 /**
325 * amdgpu_bo_create_kernel_at - create BO for kernel use at specific location
326 *
327 * @adev: amdgpu device object
328 * @offset: offset of the BO
329 * @size: size of the BO
330 * @domain: where to place it
331 * @bo_ptr: used to initialize BOs in structures
332 * @cpu_addr: optional CPU address mapping
333 *
334 * Creates a kernel BO at a specific offset in the address space of the domain.
335 *
336 * Returns:
337 * 0 on success, negative error code otherwise.
338 */
339 int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev,
340 uint64_t offset, uint64_t size, uint32_t domain,
341 struct amdgpu_bo **bo_ptr, void **cpu_addr)
342 {
343 struct ttm_operation_ctx ctx = { false, false };
344 unsigned int i;
345 int r;
346
347 offset &= PAGE_MASK;
348 size = ALIGN(size, PAGE_SIZE);
349
350 r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE, domain, bo_ptr,
351 NULL, cpu_addr);
352 if (r)
353 return r;
354
355 if ((*bo_ptr) == NULL)
356 return 0;
357
358 /*
359 * Remove the original mem node and create a new one at the request
360 * position.
361 */
362 if (cpu_addr)
363 amdgpu_bo_kunmap(*bo_ptr);
364
365 ttm_resource_free(&(*bo_ptr)->tbo, &(*bo_ptr)->tbo.mem);
366
367 for (i = 0; i < (*bo_ptr)->placement.num_placement; ++i) {
368 (*bo_ptr)->placements[i].fpfn = offset >> PAGE_SHIFT;
369 (*bo_ptr)->placements[i].lpfn = (offset + size) >> PAGE_SHIFT;
370 }
371 r = ttm_bo_mem_space(&(*bo_ptr)->tbo, &(*bo_ptr)->placement,
372 &(*bo_ptr)->tbo.mem, &ctx);
373 if (r)
374 goto error;
375
376 if (cpu_addr) {
377 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
378 if (r)
379 goto error;
380 }
381
382 amdgpu_bo_unreserve(*bo_ptr);
383 return 0;
384
385 error:
386 amdgpu_bo_unreserve(*bo_ptr);
387 amdgpu_bo_unref(bo_ptr);
388 return r;
389 }
390
391 /**
392 * amdgpu_bo_free_kernel - free BO for kernel use
393 *
394 * @bo: amdgpu BO to free
395 * @gpu_addr: pointer to where the BO's GPU memory space address was stored
396 * @cpu_addr: pointer to where the BO's CPU memory space address was stored
397 *
398 * unmaps and unpin a BO for kernel internal use.
399 */
400 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
401 void **cpu_addr)
402 {
403 if (*bo == NULL)
404 return;
405
406 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
407 if (cpu_addr)
408 amdgpu_bo_kunmap(*bo);
409
410 amdgpu_bo_unpin(*bo);
411 amdgpu_bo_unreserve(*bo);
412 }
413 amdgpu_bo_unref(bo);
414
415 if (gpu_addr)
416 *gpu_addr = 0;
417
418 if (cpu_addr)
419 *cpu_addr = NULL;
420 }
421
422 /* Validate bo size is bit bigger then the request domain */
423 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
424 unsigned long size, u32 domain)
425 {
426 struct ttm_resource_manager *man = NULL;
427
428 /*
429 * If GTT is part of requested domains the check must succeed to
430 * allow fall back to GTT
431 */
432 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
433 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_TT);
434
435 if (size < (man->size << PAGE_SHIFT))
436 return true;
437 else
438 goto fail;
439 }
440
441 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
442 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
443
444 if (size < (man->size << PAGE_SHIFT))
445 return true;
446 else
447 goto fail;
448 }
449
450
451 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */
452 return true;
453
454 fail:
455 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size,
456 man->size << PAGE_SHIFT);
457 return false;
458 }
459
460 bool amdgpu_bo_support_uswc(u64 bo_flags)
461 {
462
463 #ifdef CONFIG_X86_32
464 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
465 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
466 */
467 return false;
468 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
469 /* Don't try to enable write-combining when it can't work, or things
470 * may be slow
471 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
472 */
473
474 #ifndef CONFIG_COMPILE_TEST
475 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
476 thanks to write-combining
477 #endif
478
479 if (bo_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
480 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
481 "better performance thanks to write-combining\n");
482 return false;
483 #else
484 /* For architectures that don't support WC memory,
485 * mask out the WC flag from the BO
486 */
487 if (!drm_arch_can_wc_memory())
488 return false;
489
490 return true;
491 #endif
492 }
493
494 /**
495 * amdgpu_bo_create - create an &amdgpu_bo buffer object
496 * @adev: amdgpu device object
497 * @bp: parameters to be used for the buffer object
498 * @bo_ptr: pointer to the buffer object pointer
499 *
500 * Creates an &amdgpu_bo buffer object.
501 *
502 * Returns:
503 * 0 for success or a negative error code on failure.
504 */
505 int amdgpu_bo_create(struct amdgpu_device *adev,
506 struct amdgpu_bo_param *bp,
507 struct amdgpu_bo **bo_ptr)
508 {
509 struct ttm_operation_ctx ctx = {
510 .interruptible = (bp->type != ttm_bo_type_kernel),
511 .no_wait_gpu = bp->no_wait_gpu,
512 /* We opt to avoid OOM on system pages allocations */
513 .gfp_retry_mayfail = true,
514 .allow_res_evict = bp->type != ttm_bo_type_kernel,
515 .resv = bp->resv
516 };
517 struct amdgpu_bo *bo;
518 unsigned long page_align, size = bp->size;
519 int r;
520
521 /* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
522 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
523 /* GWS and OA don't need any alignment. */
524 page_align = bp->byte_align;
525 size <<= PAGE_SHIFT;
526 } else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
527 /* Both size and alignment must be a multiple of 4. */
528 page_align = ALIGN(bp->byte_align, 4);
529 size = ALIGN(size, 4) << PAGE_SHIFT;
530 } else {
531 /* Memory should be aligned at least to a page size. */
532 page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
533 size = ALIGN(size, PAGE_SIZE);
534 }
535
536 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
537 return -ENOMEM;
538
539 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo));
540
541 *bo_ptr = NULL;
542 bo = kzalloc(bp->bo_ptr_size, GFP_KERNEL);
543 if (bo == NULL)
544 return -ENOMEM;
545 drm_gem_private_object_init(adev_to_drm(adev), &bo->tbo.base, size);
546 INIT_LIST_HEAD(&bo->shadow_list);
547 bo->vm_bo = NULL;
548 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
549 bp->domain;
550 bo->allowed_domains = bo->preferred_domains;
551 if (bp->type != ttm_bo_type_kernel &&
552 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
553 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
554
555 bo->flags = bp->flags;
556
557 if (!amdgpu_bo_support_uswc(bo->flags))
558 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
559
560 bo->tbo.bdev = &adev->mman.bdev;
561 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
562 AMDGPU_GEM_DOMAIN_GDS))
563 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
564 else
565 amdgpu_bo_placement_from_domain(bo, bp->domain);
566 if (bp->type == ttm_bo_type_kernel)
567 bo->tbo.priority = 1;
568
569 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, bp->type,
570 &bo->placement, page_align, &ctx, NULL,
571 bp->resv, &amdgpu_bo_destroy);
572 if (unlikely(r != 0))
573 return r;
574
575 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
576 bo->tbo.mem.mem_type == TTM_PL_VRAM &&
577 bo->tbo.mem.start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
578 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
579 ctx.bytes_moved);
580 else
581 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
582
583 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
584 bo->tbo.mem.mem_type == TTM_PL_VRAM) {
585 struct dma_fence *fence;
586
587 r = amdgpu_fill_buffer(bo, 0, bo->tbo.base.resv, &fence);
588 if (unlikely(r))
589 goto fail_unreserve;
590
591 amdgpu_bo_fence(bo, fence, false);
592 dma_fence_put(bo->tbo.moving);
593 bo->tbo.moving = dma_fence_get(fence);
594 dma_fence_put(fence);
595 }
596 if (!bp->resv)
597 amdgpu_bo_unreserve(bo);
598 *bo_ptr = bo;
599
600 trace_amdgpu_bo_create(bo);
601
602 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
603 if (bp->type == ttm_bo_type_device)
604 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
605
606 return 0;
607
608 fail_unreserve:
609 if (!bp->resv)
610 dma_resv_unlock(bo->tbo.base.resv);
611 amdgpu_bo_unref(&bo);
612 return r;
613 }
614
615 int amdgpu_bo_create_shadow(struct amdgpu_device *adev,
616 unsigned long size,
617 struct amdgpu_bo *bo)
618 {
619 struct amdgpu_bo_param bp;
620 int r;
621
622 if (bo->shadow)
623 return 0;
624
625 memset(&bp, 0, sizeof(bp));
626 bp.size = size;
627 bp.domain = AMDGPU_GEM_DOMAIN_GTT;
628 bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC;
629 bp.type = ttm_bo_type_kernel;
630 bp.resv = bo->tbo.base.resv;
631 bp.bo_ptr_size = sizeof(struct amdgpu_bo);
632
633 r = amdgpu_bo_create(adev, &bp, &bo->shadow);
634 if (!r) {
635 bo->shadow->parent = amdgpu_bo_ref(bo);
636 mutex_lock(&adev->shadow_list_lock);
637 list_add_tail(&bo->shadow->shadow_list, &adev->shadow_list);
638 mutex_unlock(&adev->shadow_list_lock);
639 }
640
641 return r;
642 }
643
644 /**
645 * amdgpu_bo_create_user - create an &amdgpu_bo_user buffer object
646 * @adev: amdgpu device object
647 * @bp: parameters to be used for the buffer object
648 * @ubo_ptr: pointer to the buffer object pointer
649 *
650 * Create a BO to be used by user application;
651 *
652 * Returns:
653 * 0 for success or a negative error code on failure.
654 */
655
656 int amdgpu_bo_create_user(struct amdgpu_device *adev,
657 struct amdgpu_bo_param *bp,
658 struct amdgpu_bo_user **ubo_ptr)
659 {
660 struct amdgpu_bo *bo_ptr;
661 int r;
662
663 bp->bo_ptr_size = sizeof(struct amdgpu_bo_user);
664 r = amdgpu_bo_create(adev, bp, &bo_ptr);
665 if (r)
666 return r;
667
668 *ubo_ptr = to_amdgpu_bo_user(bo_ptr);
669 return r;
670 }
671 /**
672 * amdgpu_bo_validate - validate an &amdgpu_bo buffer object
673 * @bo: pointer to the buffer object
674 *
675 * Sets placement according to domain; and changes placement and caching
676 * policy of the buffer object according to the placement.
677 * This is used for validating shadow bos. It calls ttm_bo_validate() to
678 * make sure the buffer is resident where it needs to be.
679 *
680 * Returns:
681 * 0 for success or a negative error code on failure.
682 */
683 int amdgpu_bo_validate(struct amdgpu_bo *bo)
684 {
685 struct ttm_operation_ctx ctx = { false, false };
686 uint32_t domain;
687 int r;
688
689 if (bo->tbo.pin_count)
690 return 0;
691
692 domain = bo->preferred_domains;
693
694 retry:
695 amdgpu_bo_placement_from_domain(bo, domain);
696 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
697 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
698 domain = bo->allowed_domains;
699 goto retry;
700 }
701
702 return r;
703 }
704
705 /**
706 * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow
707 *
708 * @shadow: &amdgpu_bo shadow to be restored
709 * @fence: dma_fence associated with the operation
710 *
711 * Copies a buffer object's shadow content back to the object.
712 * This is used for recovering a buffer from its shadow in case of a gpu
713 * reset where vram context may be lost.
714 *
715 * Returns:
716 * 0 for success or a negative error code on failure.
717 */
718 int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence)
719
720 {
721 struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev);
722 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
723 uint64_t shadow_addr, parent_addr;
724
725 shadow_addr = amdgpu_bo_gpu_offset(shadow);
726 parent_addr = amdgpu_bo_gpu_offset(shadow->parent);
727
728 return amdgpu_copy_buffer(ring, shadow_addr, parent_addr,
729 amdgpu_bo_size(shadow), NULL, fence,
730 true, false, false);
731 }
732
733 /**
734 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
735 * @bo: &amdgpu_bo buffer object to be mapped
736 * @ptr: kernel virtual address to be returned
737 *
738 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
739 * amdgpu_bo_kptr() to get the kernel virtual address.
740 *
741 * Returns:
742 * 0 for success or a negative error code on failure.
743 */
744 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
745 {
746 void *kptr;
747 long r;
748
749 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
750 return -EPERM;
751
752 kptr = amdgpu_bo_kptr(bo);
753 if (kptr) {
754 if (ptr)
755 *ptr = kptr;
756 return 0;
757 }
758
759 r = dma_resv_wait_timeout_rcu(bo->tbo.base.resv, false, false,
760 MAX_SCHEDULE_TIMEOUT);
761 if (r < 0)
762 return r;
763
764 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.mem.num_pages, &bo->kmap);
765 if (r)
766 return r;
767
768 if (ptr)
769 *ptr = amdgpu_bo_kptr(bo);
770
771 return 0;
772 }
773
774 /**
775 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
776 * @bo: &amdgpu_bo buffer object
777 *
778 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
779 *
780 * Returns:
781 * the virtual address of a buffer object area.
782 */
783 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
784 {
785 bool is_iomem;
786
787 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
788 }
789
790 /**
791 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
792 * @bo: &amdgpu_bo buffer object to be unmapped
793 *
794 * Unmaps a kernel map set up by amdgpu_bo_kmap().
795 */
796 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
797 {
798 if (bo->kmap.bo)
799 ttm_bo_kunmap(&bo->kmap);
800 }
801
802 /**
803 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
804 * @bo: &amdgpu_bo buffer object
805 *
806 * References the contained &ttm_buffer_object.
807 *
808 * Returns:
809 * a refcounted pointer to the &amdgpu_bo buffer object.
810 */
811 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
812 {
813 if (bo == NULL)
814 return NULL;
815
816 ttm_bo_get(&bo->tbo);
817 return bo;
818 }
819
820 /**
821 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
822 * @bo: &amdgpu_bo buffer object
823 *
824 * Unreferences the contained &ttm_buffer_object and clear the pointer
825 */
826 void amdgpu_bo_unref(struct amdgpu_bo **bo)
827 {
828 struct ttm_buffer_object *tbo;
829
830 if ((*bo) == NULL)
831 return;
832
833 tbo = &((*bo)->tbo);
834 ttm_bo_put(tbo);
835 *bo = NULL;
836 }
837
838 /**
839 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
840 * @bo: &amdgpu_bo buffer object to be pinned
841 * @domain: domain to be pinned to
842 * @min_offset: the start of requested address range
843 * @max_offset: the end of requested address range
844 *
845 * Pins the buffer object according to requested domain and address range. If
846 * the memory is unbound gart memory, binds the pages into gart table. Adjusts
847 * pin_count and pin_size accordingly.
848 *
849 * Pinning means to lock pages in memory along with keeping them at a fixed
850 * offset. It is required when a buffer can not be moved, for example, when
851 * a display buffer is being scanned out.
852 *
853 * Compared with amdgpu_bo_pin(), this function gives more flexibility on
854 * where to pin a buffer if there are specific restrictions on where a buffer
855 * must be located.
856 *
857 * Returns:
858 * 0 for success or a negative error code on failure.
859 */
860 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
861 u64 min_offset, u64 max_offset)
862 {
863 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
864 struct ttm_operation_ctx ctx = { false, false };
865 int r, i;
866
867 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
868 return -EPERM;
869
870 if (WARN_ON_ONCE(min_offset > max_offset))
871 return -EINVAL;
872
873 /* A shared bo cannot be migrated to VRAM */
874 if (bo->prime_shared_count || bo->tbo.base.import_attach) {
875 if (domain & AMDGPU_GEM_DOMAIN_GTT)
876 domain = AMDGPU_GEM_DOMAIN_GTT;
877 else
878 return -EINVAL;
879 }
880
881 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
882 * See function amdgpu_display_supported_domains()
883 */
884 domain = amdgpu_bo_get_preferred_pin_domain(adev, domain);
885
886 if (bo->tbo.pin_count) {
887 uint32_t mem_type = bo->tbo.mem.mem_type;
888 uint32_t mem_flags = bo->tbo.mem.placement;
889
890 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
891 return -EINVAL;
892
893 if ((mem_type == TTM_PL_VRAM) &&
894 (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) &&
895 !(mem_flags & TTM_PL_FLAG_CONTIGUOUS))
896 return -EINVAL;
897
898 ttm_bo_pin(&bo->tbo);
899
900 if (max_offset != 0) {
901 u64 domain_start = amdgpu_ttm_domain_start(adev,
902 mem_type);
903 WARN_ON_ONCE(max_offset <
904 (amdgpu_bo_gpu_offset(bo) - domain_start));
905 }
906
907 return 0;
908 }
909
910 if (bo->tbo.base.import_attach)
911 dma_buf_pin(bo->tbo.base.import_attach);
912
913 /* force to pin into visible video ram */
914 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
915 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
916 amdgpu_bo_placement_from_domain(bo, domain);
917 for (i = 0; i < bo->placement.num_placement; i++) {
918 unsigned fpfn, lpfn;
919
920 fpfn = min_offset >> PAGE_SHIFT;
921 lpfn = max_offset >> PAGE_SHIFT;
922
923 if (fpfn > bo->placements[i].fpfn)
924 bo->placements[i].fpfn = fpfn;
925 if (!bo->placements[i].lpfn ||
926 (lpfn && lpfn < bo->placements[i].lpfn))
927 bo->placements[i].lpfn = lpfn;
928 }
929
930 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
931 if (unlikely(r)) {
932 dev_err(adev->dev, "%p pin failed\n", bo);
933 goto error;
934 }
935
936 ttm_bo_pin(&bo->tbo);
937
938 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
939 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
940 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
941 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
942 &adev->visible_pin_size);
943 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
944 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
945 }
946
947 error:
948 return r;
949 }
950
951 /**
952 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
953 * @bo: &amdgpu_bo buffer object to be pinned
954 * @domain: domain to be pinned to
955 *
956 * A simple wrapper to amdgpu_bo_pin_restricted().
957 * Provides a simpler API for buffers that do not have any strict restrictions
958 * on where a buffer must be located.
959 *
960 * Returns:
961 * 0 for success or a negative error code on failure.
962 */
963 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
964 {
965 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
966 return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
967 }
968
969 /**
970 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
971 * @bo: &amdgpu_bo buffer object to be unpinned
972 *
973 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
974 * Changes placement and pin size accordingly.
975 *
976 * Returns:
977 * 0 for success or a negative error code on failure.
978 */
979 void amdgpu_bo_unpin(struct amdgpu_bo *bo)
980 {
981 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
982
983 ttm_bo_unpin(&bo->tbo);
984 if (bo->tbo.pin_count)
985 return;
986
987 if (bo->tbo.base.import_attach)
988 dma_buf_unpin(bo->tbo.base.import_attach);
989
990 if (bo->tbo.mem.mem_type == TTM_PL_VRAM) {
991 atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
992 atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
993 &adev->visible_pin_size);
994 } else if (bo->tbo.mem.mem_type == TTM_PL_TT) {
995 atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size);
996 }
997 }
998
999 /**
1000 * amdgpu_bo_evict_vram - evict VRAM buffers
1001 * @adev: amdgpu device object
1002 *
1003 * Evicts all VRAM buffers on the lru list of the memory type.
1004 * Mainly used for evicting vram at suspend time.
1005 *
1006 * Returns:
1007 * 0 for success or a negative error code on failure.
1008 */
1009 int amdgpu_bo_evict_vram(struct amdgpu_device *adev)
1010 {
1011 struct ttm_resource_manager *man;
1012
1013 if (adev->in_s3 && (adev->flags & AMD_IS_APU)) {
1014 /* No need to evict vram on APUs for suspend to ram */
1015 return 0;
1016 }
1017
1018 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
1019 return ttm_resource_manager_evict_all(&adev->mman.bdev, man);
1020 }
1021
1022 static const char *amdgpu_vram_names[] = {
1023 "UNKNOWN",
1024 "GDDR1",
1025 "DDR2",
1026 "GDDR3",
1027 "GDDR4",
1028 "GDDR5",
1029 "HBM",
1030 "DDR3",
1031 "DDR4",
1032 "GDDR6",
1033 "DDR5"
1034 };
1035
1036 /**
1037 * amdgpu_bo_init - initialize memory manager
1038 * @adev: amdgpu device object
1039 *
1040 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
1041 *
1042 * Returns:
1043 * 0 for success or a negative error code on failure.
1044 */
1045 int amdgpu_bo_init(struct amdgpu_device *adev)
1046 {
1047 /* On A+A platform, VRAM can be mapped as WB */
1048 if (!adev->gmc.xgmi.connected_to_cpu) {
1049 /* reserve PAT memory space to WC for VRAM */
1050 arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1051 adev->gmc.aper_size);
1052
1053 /* Add an MTRR for the VRAM */
1054 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1055 adev->gmc.aper_size);
1056 }
1057
1058 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1059 adev->gmc.mc_vram_size >> 20,
1060 (unsigned long long)adev->gmc.aper_size >> 20);
1061 DRM_INFO("RAM width %dbits %s\n",
1062 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1063 return amdgpu_ttm_init(adev);
1064 }
1065
1066 /**
1067 * amdgpu_bo_fini - tear down memory manager
1068 * @adev: amdgpu device object
1069 *
1070 * Reverses amdgpu_bo_init() to tear down memory manager.
1071 */
1072 void amdgpu_bo_fini(struct amdgpu_device *adev)
1073 {
1074 amdgpu_ttm_fini(adev);
1075 if (!adev->gmc.xgmi.connected_to_cpu) {
1076 arch_phys_wc_del(adev->gmc.vram_mtrr);
1077 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1078 }
1079 }
1080
1081 /**
1082 * amdgpu_bo_set_tiling_flags - set tiling flags
1083 * @bo: &amdgpu_bo buffer object
1084 * @tiling_flags: new flags
1085 *
1086 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1087 * kernel driver to set the tiling flags on a buffer.
1088 *
1089 * Returns:
1090 * 0 for success or a negative error code on failure.
1091 */
1092 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1093 {
1094 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1095 struct amdgpu_bo_user *ubo;
1096
1097 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1098 if (adev->family <= AMDGPU_FAMILY_CZ &&
1099 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1100 return -EINVAL;
1101
1102 ubo = to_amdgpu_bo_user(bo);
1103 ubo->tiling_flags = tiling_flags;
1104 return 0;
1105 }
1106
1107 /**
1108 * amdgpu_bo_get_tiling_flags - get tiling flags
1109 * @bo: &amdgpu_bo buffer object
1110 * @tiling_flags: returned flags
1111 *
1112 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1113 * set the tiling flags on a buffer.
1114 */
1115 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1116 {
1117 struct amdgpu_bo_user *ubo;
1118
1119 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1120 dma_resv_assert_held(bo->tbo.base.resv);
1121 ubo = to_amdgpu_bo_user(bo);
1122
1123 if (tiling_flags)
1124 *tiling_flags = ubo->tiling_flags;
1125 }
1126
1127 /**
1128 * amdgpu_bo_set_metadata - set metadata
1129 * @bo: &amdgpu_bo buffer object
1130 * @metadata: new metadata
1131 * @metadata_size: size of the new metadata
1132 * @flags: flags of the new metadata
1133 *
1134 * Sets buffer object's metadata, its size and flags.
1135 * Used via GEM ioctl.
1136 *
1137 * Returns:
1138 * 0 for success or a negative error code on failure.
1139 */
1140 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
1141 uint32_t metadata_size, uint64_t flags)
1142 {
1143 struct amdgpu_bo_user *ubo;
1144 void *buffer;
1145
1146 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1147 ubo = to_amdgpu_bo_user(bo);
1148 if (!metadata_size) {
1149 if (ubo->metadata_size) {
1150 kfree(ubo->metadata);
1151 ubo->metadata = NULL;
1152 ubo->metadata_size = 0;
1153 }
1154 return 0;
1155 }
1156
1157 if (metadata == NULL)
1158 return -EINVAL;
1159
1160 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1161 if (buffer == NULL)
1162 return -ENOMEM;
1163
1164 kfree(ubo->metadata);
1165 ubo->metadata_flags = flags;
1166 ubo->metadata = buffer;
1167 ubo->metadata_size = metadata_size;
1168
1169 return 0;
1170 }
1171
1172 /**
1173 * amdgpu_bo_get_metadata - get metadata
1174 * @bo: &amdgpu_bo buffer object
1175 * @buffer: returned metadata
1176 * @buffer_size: size of the buffer
1177 * @metadata_size: size of the returned metadata
1178 * @flags: flags of the returned metadata
1179 *
1180 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1181 * less than metadata_size.
1182 * Used via GEM ioctl.
1183 *
1184 * Returns:
1185 * 0 for success or a negative error code on failure.
1186 */
1187 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1188 size_t buffer_size, uint32_t *metadata_size,
1189 uint64_t *flags)
1190 {
1191 struct amdgpu_bo_user *ubo;
1192
1193 if (!buffer && !metadata_size)
1194 return -EINVAL;
1195
1196 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1197 ubo = to_amdgpu_bo_user(bo);
1198 if (buffer) {
1199 if (buffer_size < ubo->metadata_size)
1200 return -EINVAL;
1201
1202 if (ubo->metadata_size)
1203 memcpy(buffer, ubo->metadata, ubo->metadata_size);
1204 }
1205
1206 if (metadata_size)
1207 *metadata_size = ubo->metadata_size;
1208 if (flags)
1209 *flags = ubo->metadata_flags;
1210
1211 return 0;
1212 }
1213
1214 /**
1215 * amdgpu_bo_move_notify - notification about a memory move
1216 * @bo: pointer to a buffer object
1217 * @evict: if this move is evicting the buffer from the graphics address space
1218 * @new_mem: new information of the bufer object
1219 *
1220 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1221 * bookkeeping.
1222 * TTM driver callback which is called when ttm moves a buffer.
1223 */
1224 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1225 bool evict,
1226 struct ttm_resource *new_mem)
1227 {
1228 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1229 struct amdgpu_bo *abo;
1230 struct ttm_resource *old_mem = &bo->mem;
1231
1232 if (!amdgpu_bo_is_amdgpu_bo(bo))
1233 return;
1234
1235 abo = ttm_to_amdgpu_bo(bo);
1236 amdgpu_vm_bo_invalidate(adev, abo, evict);
1237
1238 amdgpu_bo_kunmap(abo);
1239
1240 if (abo->tbo.base.dma_buf && !abo->tbo.base.import_attach &&
1241 bo->mem.mem_type != TTM_PL_SYSTEM)
1242 dma_buf_move_notify(abo->tbo.base.dma_buf);
1243
1244 /* remember the eviction */
1245 if (evict)
1246 atomic64_inc(&adev->num_evictions);
1247
1248 /* update statistics */
1249 if (!new_mem)
1250 return;
1251
1252 /* move_notify is called before move happens */
1253 trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
1254 }
1255
1256 void amdgpu_bo_get_memory(struct amdgpu_bo *bo, uint64_t *vram_mem,
1257 uint64_t *gtt_mem, uint64_t *cpu_mem)
1258 {
1259 unsigned int domain;
1260
1261 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
1262 switch (domain) {
1263 case AMDGPU_GEM_DOMAIN_VRAM:
1264 *vram_mem += amdgpu_bo_size(bo);
1265 break;
1266 case AMDGPU_GEM_DOMAIN_GTT:
1267 *gtt_mem += amdgpu_bo_size(bo);
1268 break;
1269 case AMDGPU_GEM_DOMAIN_CPU:
1270 default:
1271 *cpu_mem += amdgpu_bo_size(bo);
1272 break;
1273 }
1274 }
1275
1276 /**
1277 * amdgpu_bo_release_notify - notification about a BO being released
1278 * @bo: pointer to a buffer object
1279 *
1280 * Wipes VRAM buffers whose contents should not be leaked before the
1281 * memory is released.
1282 */
1283 void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)
1284 {
1285 struct dma_fence *fence = NULL;
1286 struct amdgpu_bo *abo;
1287 int r;
1288
1289 if (!amdgpu_bo_is_amdgpu_bo(bo))
1290 return;
1291
1292 abo = ttm_to_amdgpu_bo(bo);
1293
1294 if (abo->kfd_bo)
1295 amdgpu_amdkfd_unreserve_memory_limit(abo);
1296
1297 /* We only remove the fence if the resv has individualized. */
1298 WARN_ON_ONCE(bo->type == ttm_bo_type_kernel
1299 && bo->base.resv != &bo->base._resv);
1300 if (bo->base.resv == &bo->base._resv)
1301 amdgpu_amdkfd_remove_fence_on_pt_pd_bos(abo);
1302
1303 if (bo->mem.mem_type != TTM_PL_VRAM || !bo->mem.mm_node ||
1304 !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE))
1305 return;
1306
1307 dma_resv_lock(bo->base.resv, NULL);
1308
1309 r = amdgpu_fill_buffer(abo, AMDGPU_POISON, bo->base.resv, &fence);
1310 if (!WARN_ON(r)) {
1311 amdgpu_bo_fence(abo, fence, false);
1312 dma_fence_put(fence);
1313 }
1314
1315 dma_resv_unlock(bo->base.resv);
1316 }
1317
1318 /**
1319 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1320 * @bo: pointer to a buffer object
1321 *
1322 * Notifies the driver we are taking a fault on this BO and have reserved it,
1323 * also performs bookkeeping.
1324 * TTM driver callback for dealing with vm faults.
1325 *
1326 * Returns:
1327 * 0 for success or a negative error code on failure.
1328 */
1329 vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1330 {
1331 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1332 struct ttm_operation_ctx ctx = { false, false };
1333 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1334 unsigned long offset;
1335 int r;
1336
1337 /* Remember that this BO was accessed by the CPU */
1338 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1339
1340 if (bo->mem.mem_type != TTM_PL_VRAM)
1341 return 0;
1342
1343 offset = bo->mem.start << PAGE_SHIFT;
1344 if ((offset + bo->base.size) <= adev->gmc.visible_vram_size)
1345 return 0;
1346
1347 /* Can't move a pinned BO to visible VRAM */
1348 if (abo->tbo.pin_count > 0)
1349 return VM_FAULT_SIGBUS;
1350
1351 /* hurrah the memory is not visible ! */
1352 atomic64_inc(&adev->num_vram_cpu_page_faults);
1353 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1354 AMDGPU_GEM_DOMAIN_GTT);
1355
1356 /* Avoid costly evictions; only set GTT as a busy placement */
1357 abo->placement.num_busy_placement = 1;
1358 abo->placement.busy_placement = &abo->placements[1];
1359
1360 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1361 if (unlikely(r == -EBUSY || r == -ERESTARTSYS))
1362 return VM_FAULT_NOPAGE;
1363 else if (unlikely(r))
1364 return VM_FAULT_SIGBUS;
1365
1366 offset = bo->mem.start << PAGE_SHIFT;
1367 /* this should never happen */
1368 if (bo->mem.mem_type == TTM_PL_VRAM &&
1369 (offset + bo->base.size) > adev->gmc.visible_vram_size)
1370 return VM_FAULT_SIGBUS;
1371
1372 ttm_bo_move_to_lru_tail_unlocked(bo);
1373 return 0;
1374 }
1375
1376 /**
1377 * amdgpu_bo_fence - add fence to buffer object
1378 *
1379 * @bo: buffer object in question
1380 * @fence: fence to add
1381 * @shared: true if fence should be added shared
1382 *
1383 */
1384 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1385 bool shared)
1386 {
1387 struct dma_resv *resv = bo->tbo.base.resv;
1388
1389 if (shared)
1390 dma_resv_add_shared_fence(resv, fence);
1391 else
1392 dma_resv_add_excl_fence(resv, fence);
1393 }
1394
1395 /**
1396 * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences
1397 *
1398 * @adev: amdgpu device pointer
1399 * @resv: reservation object to sync to
1400 * @sync_mode: synchronization mode
1401 * @owner: fence owner
1402 * @intr: Whether the wait is interruptible
1403 *
1404 * Extract the fences from the reservation object and waits for them to finish.
1405 *
1406 * Returns:
1407 * 0 on success, errno otherwise.
1408 */
1409 int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv,
1410 enum amdgpu_sync_mode sync_mode, void *owner,
1411 bool intr)
1412 {
1413 struct amdgpu_sync sync;
1414 int r;
1415
1416 amdgpu_sync_create(&sync);
1417 amdgpu_sync_resv(adev, &sync, resv, sync_mode, owner);
1418 r = amdgpu_sync_wait(&sync, intr);
1419 amdgpu_sync_free(&sync);
1420 return r;
1421 }
1422
1423 /**
1424 * amdgpu_bo_sync_wait - Wrapper for amdgpu_bo_sync_wait_resv
1425 * @bo: buffer object to wait for
1426 * @owner: fence owner
1427 * @intr: Whether the wait is interruptible
1428 *
1429 * Wrapper to wait for fences in a BO.
1430 * Returns:
1431 * 0 on success, errno otherwise.
1432 */
1433 int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)
1434 {
1435 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1436
1437 return amdgpu_bo_sync_wait_resv(adev, bo->tbo.base.resv,
1438 AMDGPU_SYNC_NE_OWNER, owner, intr);
1439 }
1440
1441 /**
1442 * amdgpu_bo_gpu_offset - return GPU offset of bo
1443 * @bo: amdgpu object for which we query the offset
1444 *
1445 * Note: object should either be pinned or reserved when calling this
1446 * function, it might be useful to add check for this for debugging.
1447 *
1448 * Returns:
1449 * current GPU offset of the object.
1450 */
1451 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1452 {
1453 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_SYSTEM);
1454 WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) &&
1455 !bo->tbo.pin_count && bo->tbo.type != ttm_bo_type_kernel);
1456 WARN_ON_ONCE(bo->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET);
1457 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_VRAM &&
1458 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1459
1460 return amdgpu_bo_gpu_offset_no_check(bo);
1461 }
1462
1463 /**
1464 * amdgpu_bo_gpu_offset_no_check - return GPU offset of bo
1465 * @bo: amdgpu object for which we query the offset
1466 *
1467 * Returns:
1468 * current GPU offset of the object without raising warnings.
1469 */
1470 u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo)
1471 {
1472 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1473 uint64_t offset;
1474
1475 offset = (bo->tbo.mem.start << PAGE_SHIFT) +
1476 amdgpu_ttm_domain_start(adev, bo->tbo.mem.mem_type);
1477
1478 return amdgpu_gmc_sign_extend(offset);
1479 }
1480
1481 /**
1482 * amdgpu_bo_get_preferred_pin_domain - get preferred domain for scanout
1483 * @adev: amdgpu device object
1484 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1485 *
1486 * Returns:
1487 * Which of the allowed domains is preferred for pinning the BO for scanout.
1488 */
1489 uint32_t amdgpu_bo_get_preferred_pin_domain(struct amdgpu_device *adev,
1490 uint32_t domain)
1491 {
1492 if (domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) {
1493 domain = AMDGPU_GEM_DOMAIN_VRAM;
1494 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1495 domain = AMDGPU_GEM_DOMAIN_GTT;
1496 }
1497 return domain;
1498 }
1499
1500 #if defined(CONFIG_DEBUG_FS)
1501 #define amdgpu_bo_print_flag(m, bo, flag) \
1502 do { \
1503 if (bo->flags & (AMDGPU_GEM_CREATE_ ## flag)) { \
1504 seq_printf((m), " " #flag); \
1505 } \
1506 } while (0)
1507
1508 /**
1509 * amdgpu_bo_print_info - print BO info in debugfs file
1510 *
1511 * @id: Index or Id of the BO
1512 * @bo: Requested BO for printing info
1513 * @m: debugfs file
1514 *
1515 * Print BO information in debugfs file
1516 *
1517 * Returns:
1518 * Size of the BO in bytes.
1519 */
1520 u64 amdgpu_bo_print_info(int id, struct amdgpu_bo *bo, struct seq_file *m)
1521 {
1522 struct dma_buf_attachment *attachment;
1523 struct dma_buf *dma_buf;
1524 unsigned int domain;
1525 const char *placement;
1526 unsigned int pin_count;
1527 u64 size;
1528
1529 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
1530 switch (domain) {
1531 case AMDGPU_GEM_DOMAIN_VRAM:
1532 placement = "VRAM";
1533 break;
1534 case AMDGPU_GEM_DOMAIN_GTT:
1535 placement = " GTT";
1536 break;
1537 case AMDGPU_GEM_DOMAIN_CPU:
1538 default:
1539 placement = " CPU";
1540 break;
1541 }
1542
1543 size = amdgpu_bo_size(bo);
1544 seq_printf(m, "\t\t0x%08x: %12lld byte %s",
1545 id, size, placement);
1546
1547 pin_count = READ_ONCE(bo->tbo.pin_count);
1548 if (pin_count)
1549 seq_printf(m, " pin count %d", pin_count);
1550
1551 dma_buf = READ_ONCE(bo->tbo.base.dma_buf);
1552 attachment = READ_ONCE(bo->tbo.base.import_attach);
1553
1554 if (attachment)
1555 seq_printf(m, " imported from %p", dma_buf);
1556 else if (dma_buf)
1557 seq_printf(m, " exported as %p", dma_buf);
1558
1559 amdgpu_bo_print_flag(m, bo, CPU_ACCESS_REQUIRED);
1560 amdgpu_bo_print_flag(m, bo, NO_CPU_ACCESS);
1561 amdgpu_bo_print_flag(m, bo, CPU_GTT_USWC);
1562 amdgpu_bo_print_flag(m, bo, VRAM_CLEARED);
1563 amdgpu_bo_print_flag(m, bo, VRAM_CONTIGUOUS);
1564 amdgpu_bo_print_flag(m, bo, VM_ALWAYS_VALID);
1565 amdgpu_bo_print_flag(m, bo, EXPLICIT_SYNC);
1566
1567 seq_puts(m, "\n");
1568
1569 return size;
1570 }
1571 #endif