Loading...
1/*
2 * Copyright 2009 Jerome Glisse.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sub license, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
12 *
13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19 * USE OR OTHER DEALINGS IN THE SOFTWARE.
20 *
21 * The above copyright notice and this permission notice (including the
22 * next paragraph) shall be included in all copies or substantial portions
23 * of the Software.
24 *
25 */
26/*
27 * Authors:
28 * Jerome Glisse <glisse@freedesktop.org>
29 * Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
30 * Dave Airlie
31 */
32#include <linux/list.h>
33#include <linux/slab.h>
34#include <drm/drmP.h>
35#include <drm/amdgpu_drm.h>
36#include <drm/drm_cache.h>
37#include "amdgpu.h"
38#include "amdgpu_trace.h"
39#include "amdgpu_amdkfd.h"
40
41static bool amdgpu_need_backup(struct amdgpu_device *adev)
42{
43 if (adev->flags & AMD_IS_APU)
44 return false;
45
46 if (amdgpu_gpu_recovery == 0 ||
47 (amdgpu_gpu_recovery == -1 && !amdgpu_sriov_vf(adev)))
48 return false;
49
50 return true;
51}
52
53static void amdgpu_ttm_bo_destroy(struct ttm_buffer_object *tbo)
54{
55 struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
56 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
57
58 if (bo->kfd_bo)
59 amdgpu_amdkfd_unreserve_system_memory_limit(bo);
60
61 amdgpu_bo_kunmap(bo);
62
63 if (bo->gem_base.import_attach)
64 drm_prime_gem_destroy(&bo->gem_base, bo->tbo.sg);
65 drm_gem_object_release(&bo->gem_base);
66 amdgpu_bo_unref(&bo->parent);
67 if (!list_empty(&bo->shadow_list)) {
68 mutex_lock(&adev->shadow_list_lock);
69 list_del_init(&bo->shadow_list);
70 mutex_unlock(&adev->shadow_list_lock);
71 }
72 kfree(bo->metadata);
73 kfree(bo);
74}
75
76bool amdgpu_ttm_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
77{
78 if (bo->destroy == &amdgpu_ttm_bo_destroy)
79 return true;
80 return false;
81}
82
83void amdgpu_ttm_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
84{
85 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
86 struct ttm_placement *placement = &abo->placement;
87 struct ttm_place *places = abo->placements;
88 u64 flags = abo->flags;
89 u32 c = 0;
90
91 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
92 unsigned visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
93
94 places[c].fpfn = 0;
95 places[c].lpfn = 0;
96 places[c].flags = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
97 TTM_PL_FLAG_VRAM;
98
99 if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
100 places[c].lpfn = visible_pfn;
101 else
102 places[c].flags |= TTM_PL_FLAG_TOPDOWN;
103
104 if (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
105 places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
106 c++;
107 }
108
109 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
110 places[c].fpfn = 0;
111 if (flags & AMDGPU_GEM_CREATE_SHADOW)
112 places[c].lpfn = adev->gmc.gart_size >> PAGE_SHIFT;
113 else
114 places[c].lpfn = 0;
115 places[c].flags = TTM_PL_FLAG_TT;
116 if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
117 places[c].flags |= TTM_PL_FLAG_WC |
118 TTM_PL_FLAG_UNCACHED;
119 else
120 places[c].flags |= TTM_PL_FLAG_CACHED;
121 c++;
122 }
123
124 if (domain & AMDGPU_GEM_DOMAIN_CPU) {
125 places[c].fpfn = 0;
126 places[c].lpfn = 0;
127 places[c].flags = TTM_PL_FLAG_SYSTEM;
128 if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
129 places[c].flags |= TTM_PL_FLAG_WC |
130 TTM_PL_FLAG_UNCACHED;
131 else
132 places[c].flags |= TTM_PL_FLAG_CACHED;
133 c++;
134 }
135
136 if (domain & AMDGPU_GEM_DOMAIN_GDS) {
137 places[c].fpfn = 0;
138 places[c].lpfn = 0;
139 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GDS;
140 c++;
141 }
142
143 if (domain & AMDGPU_GEM_DOMAIN_GWS) {
144 places[c].fpfn = 0;
145 places[c].lpfn = 0;
146 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GWS;
147 c++;
148 }
149
150 if (domain & AMDGPU_GEM_DOMAIN_OA) {
151 places[c].fpfn = 0;
152 places[c].lpfn = 0;
153 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_OA;
154 c++;
155 }
156
157 if (!c) {
158 places[c].fpfn = 0;
159 places[c].lpfn = 0;
160 places[c].flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
161 c++;
162 }
163
164 placement->num_placement = c;
165 placement->placement = places;
166
167 placement->num_busy_placement = c;
168 placement->busy_placement = places;
169}
170
171/**
172 * amdgpu_bo_create_reserved - create reserved BO for kernel use
173 *
174 * @adev: amdgpu device object
175 * @size: size for the new BO
176 * @align: alignment for the new BO
177 * @domain: where to place it
178 * @bo_ptr: used to initialize BOs in structures
179 * @gpu_addr: GPU addr of the pinned BO
180 * @cpu_addr: optional CPU address mapping
181 *
182 * Allocates and pins a BO for kernel internal use, and returns it still
183 * reserved.
184 *
185 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
186 *
187 * Returns 0 on success, negative error code otherwise.
188 */
189int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
190 unsigned long size, int align,
191 u32 domain, struct amdgpu_bo **bo_ptr,
192 u64 *gpu_addr, void **cpu_addr)
193{
194 bool free = false;
195 int r;
196
197 if (!*bo_ptr) {
198 r = amdgpu_bo_create(adev, size, align, domain,
199 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
200 AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS,
201 ttm_bo_type_kernel, NULL, bo_ptr);
202 if (r) {
203 dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
204 r);
205 return r;
206 }
207 free = true;
208 }
209
210 r = amdgpu_bo_reserve(*bo_ptr, false);
211 if (r) {
212 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
213 goto error_free;
214 }
215
216 r = amdgpu_bo_pin(*bo_ptr, domain, gpu_addr);
217 if (r) {
218 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
219 goto error_unreserve;
220 }
221
222 if (cpu_addr) {
223 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
224 if (r) {
225 dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
226 goto error_unreserve;
227 }
228 }
229
230 return 0;
231
232error_unreserve:
233 amdgpu_bo_unreserve(*bo_ptr);
234
235error_free:
236 if (free)
237 amdgpu_bo_unref(bo_ptr);
238
239 return r;
240}
241
242/**
243 * amdgpu_bo_create_kernel - create BO for kernel use
244 *
245 * @adev: amdgpu device object
246 * @size: size for the new BO
247 * @align: alignment for the new BO
248 * @domain: where to place it
249 * @bo_ptr: used to initialize BOs in structures
250 * @gpu_addr: GPU addr of the pinned BO
251 * @cpu_addr: optional CPU address mapping
252 *
253 * Allocates and pins a BO for kernel internal use.
254 *
255 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
256 *
257 * Returns 0 on success, negative error code otherwise.
258 */
259int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
260 unsigned long size, int align,
261 u32 domain, struct amdgpu_bo **bo_ptr,
262 u64 *gpu_addr, void **cpu_addr)
263{
264 int r;
265
266 r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
267 gpu_addr, cpu_addr);
268
269 if (r)
270 return r;
271
272 amdgpu_bo_unreserve(*bo_ptr);
273
274 return 0;
275}
276
277/**
278 * amdgpu_bo_free_kernel - free BO for kernel use
279 *
280 * @bo: amdgpu BO to free
281 *
282 * unmaps and unpin a BO for kernel internal use.
283 */
284void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
285 void **cpu_addr)
286{
287 if (*bo == NULL)
288 return;
289
290 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
291 if (cpu_addr)
292 amdgpu_bo_kunmap(*bo);
293
294 amdgpu_bo_unpin(*bo);
295 amdgpu_bo_unreserve(*bo);
296 }
297 amdgpu_bo_unref(bo);
298
299 if (gpu_addr)
300 *gpu_addr = 0;
301
302 if (cpu_addr)
303 *cpu_addr = NULL;
304}
305
306/* Validate bo size is bit bigger then the request domain */
307static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
308 unsigned long size, u32 domain)
309{
310 struct ttm_mem_type_manager *man = NULL;
311
312 /*
313 * If GTT is part of requested domains the check must succeed to
314 * allow fall back to GTT
315 */
316 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
317 man = &adev->mman.bdev.man[TTM_PL_TT];
318
319 if (size < (man->size << PAGE_SHIFT))
320 return true;
321 else
322 goto fail;
323 }
324
325 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
326 man = &adev->mman.bdev.man[TTM_PL_VRAM];
327
328 if (size < (man->size << PAGE_SHIFT))
329 return true;
330 else
331 goto fail;
332 }
333
334
335 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */
336 return true;
337
338fail:
339 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size,
340 man->size << PAGE_SHIFT);
341 return false;
342}
343
344static int amdgpu_bo_do_create(struct amdgpu_device *adev, unsigned long size,
345 int byte_align, u32 domain,
346 u64 flags, enum ttm_bo_type type,
347 struct reservation_object *resv,
348 struct amdgpu_bo **bo_ptr)
349{
350 struct ttm_operation_ctx ctx = {
351 .interruptible = (type != ttm_bo_type_kernel),
352 .no_wait_gpu = false,
353 .resv = resv,
354 .flags = TTM_OPT_FLAG_ALLOW_RES_EVICT
355 };
356 struct amdgpu_bo *bo;
357 unsigned long page_align;
358 size_t acc_size;
359 int r;
360
361 page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
362 size = ALIGN(size, PAGE_SIZE);
363
364 if (!amdgpu_bo_validate_size(adev, size, domain))
365 return -ENOMEM;
366
367 *bo_ptr = NULL;
368
369 acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
370 sizeof(struct amdgpu_bo));
371
372 bo = kzalloc(sizeof(struct amdgpu_bo), GFP_KERNEL);
373 if (bo == NULL)
374 return -ENOMEM;
375 drm_gem_private_object_init(adev->ddev, &bo->gem_base, size);
376 INIT_LIST_HEAD(&bo->shadow_list);
377 INIT_LIST_HEAD(&bo->va);
378 bo->preferred_domains = domain & (AMDGPU_GEM_DOMAIN_VRAM |
379 AMDGPU_GEM_DOMAIN_GTT |
380 AMDGPU_GEM_DOMAIN_CPU |
381 AMDGPU_GEM_DOMAIN_GDS |
382 AMDGPU_GEM_DOMAIN_GWS |
383 AMDGPU_GEM_DOMAIN_OA);
384 bo->allowed_domains = bo->preferred_domains;
385 if (type != ttm_bo_type_kernel &&
386 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
387 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
388
389 bo->flags = flags;
390
391#ifdef CONFIG_X86_32
392 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
393 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
394 */
395 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
396#elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
397 /* Don't try to enable write-combining when it can't work, or things
398 * may be slow
399 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
400 */
401
402#ifndef CONFIG_COMPILE_TEST
403#warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
404 thanks to write-combining
405#endif
406
407 if (bo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
408 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
409 "better performance thanks to write-combining\n");
410 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
411#else
412 /* For architectures that don't support WC memory,
413 * mask out the WC flag from the BO
414 */
415 if (!drm_arch_can_wc_memory())
416 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
417#endif
418
419 bo->tbo.bdev = &adev->mman.bdev;
420 amdgpu_ttm_placement_from_domain(bo, domain);
421
422 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, type,
423 &bo->placement, page_align, &ctx, acc_size,
424 NULL, resv, &amdgpu_ttm_bo_destroy);
425 if (unlikely(r != 0))
426 return r;
427
428 if (adev->gmc.visible_vram_size < adev->gmc.real_vram_size &&
429 bo->tbo.mem.mem_type == TTM_PL_VRAM &&
430 bo->tbo.mem.start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
431 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
432 ctx.bytes_moved);
433 else
434 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
435
436 if (type == ttm_bo_type_kernel)
437 bo->tbo.priority = 1;
438
439 if (flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
440 bo->tbo.mem.placement & TTM_PL_FLAG_VRAM) {
441 struct dma_fence *fence;
442
443 r = amdgpu_fill_buffer(bo, 0, bo->tbo.resv, &fence);
444 if (unlikely(r))
445 goto fail_unreserve;
446
447 amdgpu_bo_fence(bo, fence, false);
448 dma_fence_put(bo->tbo.moving);
449 bo->tbo.moving = dma_fence_get(fence);
450 dma_fence_put(fence);
451 }
452 if (!resv)
453 amdgpu_bo_unreserve(bo);
454 *bo_ptr = bo;
455
456 trace_amdgpu_bo_create(bo);
457
458 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
459 if (type == ttm_bo_type_device)
460 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
461
462 return 0;
463
464fail_unreserve:
465 if (!resv)
466 ww_mutex_unlock(&bo->tbo.resv->lock);
467 amdgpu_bo_unref(&bo);
468 return r;
469}
470
471static int amdgpu_bo_create_shadow(struct amdgpu_device *adev,
472 unsigned long size, int byte_align,
473 struct amdgpu_bo *bo)
474{
475 int r;
476
477 if (bo->shadow)
478 return 0;
479
480 r = amdgpu_bo_do_create(adev, size, byte_align, AMDGPU_GEM_DOMAIN_GTT,
481 AMDGPU_GEM_CREATE_CPU_GTT_USWC |
482 AMDGPU_GEM_CREATE_SHADOW,
483 ttm_bo_type_kernel,
484 bo->tbo.resv, &bo->shadow);
485 if (!r) {
486 bo->shadow->parent = amdgpu_bo_ref(bo);
487 mutex_lock(&adev->shadow_list_lock);
488 list_add_tail(&bo->shadow_list, &adev->shadow_list);
489 mutex_unlock(&adev->shadow_list_lock);
490 }
491
492 return r;
493}
494
495int amdgpu_bo_create(struct amdgpu_device *adev, unsigned long size,
496 int byte_align, u32 domain,
497 u64 flags, enum ttm_bo_type type,
498 struct reservation_object *resv,
499 struct amdgpu_bo **bo_ptr)
500{
501 uint64_t parent_flags = flags & ~AMDGPU_GEM_CREATE_SHADOW;
502 int r;
503
504 r = amdgpu_bo_do_create(adev, size, byte_align, domain,
505 parent_flags, type, resv, bo_ptr);
506 if (r)
507 return r;
508
509 if ((flags & AMDGPU_GEM_CREATE_SHADOW) && amdgpu_need_backup(adev)) {
510 if (!resv)
511 WARN_ON(reservation_object_lock((*bo_ptr)->tbo.resv,
512 NULL));
513
514 r = amdgpu_bo_create_shadow(adev, size, byte_align, (*bo_ptr));
515
516 if (!resv)
517 reservation_object_unlock((*bo_ptr)->tbo.resv);
518
519 if (r)
520 amdgpu_bo_unref(bo_ptr);
521 }
522
523 return r;
524}
525
526int amdgpu_bo_backup_to_shadow(struct amdgpu_device *adev,
527 struct amdgpu_ring *ring,
528 struct amdgpu_bo *bo,
529 struct reservation_object *resv,
530 struct dma_fence **fence,
531 bool direct)
532
533{
534 struct amdgpu_bo *shadow = bo->shadow;
535 uint64_t bo_addr, shadow_addr;
536 int r;
537
538 if (!shadow)
539 return -EINVAL;
540
541 bo_addr = amdgpu_bo_gpu_offset(bo);
542 shadow_addr = amdgpu_bo_gpu_offset(bo->shadow);
543
544 r = reservation_object_reserve_shared(bo->tbo.resv);
545 if (r)
546 goto err;
547
548 r = amdgpu_copy_buffer(ring, bo_addr, shadow_addr,
549 amdgpu_bo_size(bo), resv, fence,
550 direct, false);
551 if (!r)
552 amdgpu_bo_fence(bo, *fence, true);
553
554err:
555 return r;
556}
557
558int amdgpu_bo_validate(struct amdgpu_bo *bo)
559{
560 struct ttm_operation_ctx ctx = { false, false };
561 uint32_t domain;
562 int r;
563
564 if (bo->pin_count)
565 return 0;
566
567 domain = bo->preferred_domains;
568
569retry:
570 amdgpu_ttm_placement_from_domain(bo, domain);
571 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
572 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
573 domain = bo->allowed_domains;
574 goto retry;
575 }
576
577 return r;
578}
579
580int amdgpu_bo_restore_from_shadow(struct amdgpu_device *adev,
581 struct amdgpu_ring *ring,
582 struct amdgpu_bo *bo,
583 struct reservation_object *resv,
584 struct dma_fence **fence,
585 bool direct)
586
587{
588 struct amdgpu_bo *shadow = bo->shadow;
589 uint64_t bo_addr, shadow_addr;
590 int r;
591
592 if (!shadow)
593 return -EINVAL;
594
595 bo_addr = amdgpu_bo_gpu_offset(bo);
596 shadow_addr = amdgpu_bo_gpu_offset(bo->shadow);
597
598 r = reservation_object_reserve_shared(bo->tbo.resv);
599 if (r)
600 goto err;
601
602 r = amdgpu_copy_buffer(ring, shadow_addr, bo_addr,
603 amdgpu_bo_size(bo), resv, fence,
604 direct, false);
605 if (!r)
606 amdgpu_bo_fence(bo, *fence, true);
607
608err:
609 return r;
610}
611
612int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
613{
614 void *kptr;
615 long r;
616
617 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
618 return -EPERM;
619
620 kptr = amdgpu_bo_kptr(bo);
621 if (kptr) {
622 if (ptr)
623 *ptr = kptr;
624 return 0;
625 }
626
627 r = reservation_object_wait_timeout_rcu(bo->tbo.resv, false, false,
628 MAX_SCHEDULE_TIMEOUT);
629 if (r < 0)
630 return r;
631
632 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
633 if (r)
634 return r;
635
636 if (ptr)
637 *ptr = amdgpu_bo_kptr(bo);
638
639 return 0;
640}
641
642void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
643{
644 bool is_iomem;
645
646 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
647}
648
649void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
650{
651 if (bo->kmap.bo)
652 ttm_bo_kunmap(&bo->kmap);
653}
654
655struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
656{
657 if (bo == NULL)
658 return NULL;
659
660 ttm_bo_reference(&bo->tbo);
661 return bo;
662}
663
664void amdgpu_bo_unref(struct amdgpu_bo **bo)
665{
666 struct ttm_buffer_object *tbo;
667
668 if ((*bo) == NULL)
669 return;
670
671 tbo = &((*bo)->tbo);
672 ttm_bo_unref(&tbo);
673 if (tbo == NULL)
674 *bo = NULL;
675}
676
677int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
678 u64 min_offset, u64 max_offset,
679 u64 *gpu_addr)
680{
681 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
682 struct ttm_operation_ctx ctx = { false, false };
683 int r, i;
684
685 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
686 return -EPERM;
687
688 if (WARN_ON_ONCE(min_offset > max_offset))
689 return -EINVAL;
690
691 /* A shared bo cannot be migrated to VRAM */
692 if (bo->prime_shared_count && (domain == AMDGPU_GEM_DOMAIN_VRAM))
693 return -EINVAL;
694
695 if (bo->pin_count) {
696 uint32_t mem_type = bo->tbo.mem.mem_type;
697
698 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
699 return -EINVAL;
700
701 bo->pin_count++;
702 if (gpu_addr)
703 *gpu_addr = amdgpu_bo_gpu_offset(bo);
704
705 if (max_offset != 0) {
706 u64 domain_start = bo->tbo.bdev->man[mem_type].gpu_offset;
707 WARN_ON_ONCE(max_offset <
708 (amdgpu_bo_gpu_offset(bo) - domain_start));
709 }
710
711 return 0;
712 }
713
714 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
715 /* force to pin into visible video ram */
716 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
717 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
718 amdgpu_ttm_placement_from_domain(bo, domain);
719 for (i = 0; i < bo->placement.num_placement; i++) {
720 unsigned fpfn, lpfn;
721
722 fpfn = min_offset >> PAGE_SHIFT;
723 lpfn = max_offset >> PAGE_SHIFT;
724
725 if (fpfn > bo->placements[i].fpfn)
726 bo->placements[i].fpfn = fpfn;
727 if (!bo->placements[i].lpfn ||
728 (lpfn && lpfn < bo->placements[i].lpfn))
729 bo->placements[i].lpfn = lpfn;
730 bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
731 }
732
733 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
734 if (unlikely(r)) {
735 dev_err(adev->dev, "%p pin failed\n", bo);
736 goto error;
737 }
738
739 r = amdgpu_ttm_alloc_gart(&bo->tbo);
740 if (unlikely(r)) {
741 dev_err(adev->dev, "%p bind failed\n", bo);
742 goto error;
743 }
744
745 bo->pin_count = 1;
746 if (gpu_addr != NULL)
747 *gpu_addr = amdgpu_bo_gpu_offset(bo);
748
749 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
750 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
751 adev->vram_pin_size += amdgpu_bo_size(bo);
752 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
753 adev->invisible_pin_size += amdgpu_bo_size(bo);
754 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
755 adev->gart_pin_size += amdgpu_bo_size(bo);
756 }
757
758error:
759 return r;
760}
761
762int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain, u64 *gpu_addr)
763{
764 return amdgpu_bo_pin_restricted(bo, domain, 0, 0, gpu_addr);
765}
766
767int amdgpu_bo_unpin(struct amdgpu_bo *bo)
768{
769 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
770 struct ttm_operation_ctx ctx = { false, false };
771 int r, i;
772
773 if (!bo->pin_count) {
774 dev_warn(adev->dev, "%p unpin not necessary\n", bo);
775 return 0;
776 }
777 bo->pin_count--;
778 if (bo->pin_count)
779 return 0;
780 for (i = 0; i < bo->placement.num_placement; i++) {
781 bo->placements[i].lpfn = 0;
782 bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
783 }
784 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
785 if (unlikely(r)) {
786 dev_err(adev->dev, "%p validate failed for unpin\n", bo);
787 goto error;
788 }
789
790 if (bo->tbo.mem.mem_type == TTM_PL_VRAM) {
791 adev->vram_pin_size -= amdgpu_bo_size(bo);
792 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
793 adev->invisible_pin_size -= amdgpu_bo_size(bo);
794 } else if (bo->tbo.mem.mem_type == TTM_PL_TT) {
795 adev->gart_pin_size -= amdgpu_bo_size(bo);
796 }
797
798error:
799 return r;
800}
801
802int amdgpu_bo_evict_vram(struct amdgpu_device *adev)
803{
804 /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
805 if (0 && (adev->flags & AMD_IS_APU)) {
806 /* Useless to evict on IGP chips */
807 return 0;
808 }
809 return ttm_bo_evict_mm(&adev->mman.bdev, TTM_PL_VRAM);
810}
811
812static const char *amdgpu_vram_names[] = {
813 "UNKNOWN",
814 "GDDR1",
815 "DDR2",
816 "GDDR3",
817 "GDDR4",
818 "GDDR5",
819 "HBM",
820 "DDR3",
821 "DDR4",
822};
823
824int amdgpu_bo_init(struct amdgpu_device *adev)
825{
826 /* reserve PAT memory space to WC for VRAM */
827 arch_io_reserve_memtype_wc(adev->gmc.aper_base,
828 adev->gmc.aper_size);
829
830 /* Add an MTRR for the VRAM */
831 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
832 adev->gmc.aper_size);
833 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
834 adev->gmc.mc_vram_size >> 20,
835 (unsigned long long)adev->gmc.aper_size >> 20);
836 DRM_INFO("RAM width %dbits %s\n",
837 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
838 return amdgpu_ttm_init(adev);
839}
840
841void amdgpu_bo_fini(struct amdgpu_device *adev)
842{
843 amdgpu_ttm_fini(adev);
844 arch_phys_wc_del(adev->gmc.vram_mtrr);
845 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
846}
847
848int amdgpu_bo_fbdev_mmap(struct amdgpu_bo *bo,
849 struct vm_area_struct *vma)
850{
851 return ttm_fbdev_mmap(vma, &bo->tbo);
852}
853
854int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
855{
856 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
857
858 if (adev->family <= AMDGPU_FAMILY_CZ &&
859 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
860 return -EINVAL;
861
862 bo->tiling_flags = tiling_flags;
863 return 0;
864}
865
866void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
867{
868 lockdep_assert_held(&bo->tbo.resv->lock.base);
869
870 if (tiling_flags)
871 *tiling_flags = bo->tiling_flags;
872}
873
874int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
875 uint32_t metadata_size, uint64_t flags)
876{
877 void *buffer;
878
879 if (!metadata_size) {
880 if (bo->metadata_size) {
881 kfree(bo->metadata);
882 bo->metadata = NULL;
883 bo->metadata_size = 0;
884 }
885 return 0;
886 }
887
888 if (metadata == NULL)
889 return -EINVAL;
890
891 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
892 if (buffer == NULL)
893 return -ENOMEM;
894
895 kfree(bo->metadata);
896 bo->metadata_flags = flags;
897 bo->metadata = buffer;
898 bo->metadata_size = metadata_size;
899
900 return 0;
901}
902
903int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
904 size_t buffer_size, uint32_t *metadata_size,
905 uint64_t *flags)
906{
907 if (!buffer && !metadata_size)
908 return -EINVAL;
909
910 if (buffer) {
911 if (buffer_size < bo->metadata_size)
912 return -EINVAL;
913
914 if (bo->metadata_size)
915 memcpy(buffer, bo->metadata, bo->metadata_size);
916 }
917
918 if (metadata_size)
919 *metadata_size = bo->metadata_size;
920 if (flags)
921 *flags = bo->metadata_flags;
922
923 return 0;
924}
925
926void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
927 bool evict,
928 struct ttm_mem_reg *new_mem)
929{
930 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
931 struct amdgpu_bo *abo;
932 struct ttm_mem_reg *old_mem = &bo->mem;
933
934 if (!amdgpu_ttm_bo_is_amdgpu_bo(bo))
935 return;
936
937 abo = ttm_to_amdgpu_bo(bo);
938 amdgpu_vm_bo_invalidate(adev, abo, evict);
939
940 amdgpu_bo_kunmap(abo);
941
942 /* remember the eviction */
943 if (evict)
944 atomic64_inc(&adev->num_evictions);
945
946 /* update statistics */
947 if (!new_mem)
948 return;
949
950 /* move_notify is called before move happens */
951 trace_amdgpu_ttm_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
952}
953
954int amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
955{
956 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
957 struct ttm_operation_ctx ctx = { false, false };
958 struct amdgpu_bo *abo;
959 unsigned long offset, size;
960 int r;
961
962 if (!amdgpu_ttm_bo_is_amdgpu_bo(bo))
963 return 0;
964
965 abo = ttm_to_amdgpu_bo(bo);
966
967 /* Remember that this BO was accessed by the CPU */
968 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
969
970 if (bo->mem.mem_type != TTM_PL_VRAM)
971 return 0;
972
973 size = bo->mem.num_pages << PAGE_SHIFT;
974 offset = bo->mem.start << PAGE_SHIFT;
975 if ((offset + size) <= adev->gmc.visible_vram_size)
976 return 0;
977
978 /* Can't move a pinned BO to visible VRAM */
979 if (abo->pin_count > 0)
980 return -EINVAL;
981
982 /* hurrah the memory is not visible ! */
983 atomic64_inc(&adev->num_vram_cpu_page_faults);
984 amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
985 AMDGPU_GEM_DOMAIN_GTT);
986
987 /* Avoid costly evictions; only set GTT as a busy placement */
988 abo->placement.num_busy_placement = 1;
989 abo->placement.busy_placement = &abo->placements[1];
990
991 r = ttm_bo_validate(bo, &abo->placement, &ctx);
992 if (unlikely(r != 0))
993 return r;
994
995 offset = bo->mem.start << PAGE_SHIFT;
996 /* this should never happen */
997 if (bo->mem.mem_type == TTM_PL_VRAM &&
998 (offset + size) > adev->gmc.visible_vram_size)
999 return -EINVAL;
1000
1001 return 0;
1002}
1003
1004/**
1005 * amdgpu_bo_fence - add fence to buffer object
1006 *
1007 * @bo: buffer object in question
1008 * @fence: fence to add
1009 * @shared: true if fence should be added shared
1010 *
1011 */
1012void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1013 bool shared)
1014{
1015 struct reservation_object *resv = bo->tbo.resv;
1016
1017 if (shared)
1018 reservation_object_add_shared_fence(resv, fence);
1019 else
1020 reservation_object_add_excl_fence(resv, fence);
1021}
1022
1023/**
1024 * amdgpu_bo_gpu_offset - return GPU offset of bo
1025 * @bo: amdgpu object for which we query the offset
1026 *
1027 * Returns current GPU offset of the object.
1028 *
1029 * Note: object should either be pinned or reserved when calling this
1030 * function, it might be useful to add check for this for debugging.
1031 */
1032u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1033{
1034 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_SYSTEM);
1035 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_TT &&
1036 !amdgpu_gtt_mgr_has_gart_addr(&bo->tbo.mem));
1037 WARN_ON_ONCE(!ww_mutex_is_locked(&bo->tbo.resv->lock) &&
1038 !bo->pin_count);
1039 WARN_ON_ONCE(bo->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET);
1040 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_VRAM &&
1041 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1042
1043 return bo->tbo.offset;
1044}
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
35#include <drm/amdgpu_drm.h>
36#include <drm/drm_cache.h>
37#include "amdgpu.h"
38#include "amdgpu_trace.h"
39#include "amdgpu_amdkfd.h"
40
41/**
42 * DOC: amdgpu_object
43 *
44 * This defines the interfaces to operate on an &amdgpu_bo buffer object which
45 * represents memory used by driver (VRAM, system memory, etc.). The driver
46 * provides DRM/GEM APIs to userspace. DRM/GEM APIs then use these interfaces
47 * to create/destroy/set buffer object which are then managed by the kernel TTM
48 * memory manager.
49 * The interfaces are also used internally by kernel clients, including gfx,
50 * uvd, etc. for kernel managed allocations used by the GPU.
51 *
52 */
53
54/**
55 * amdgpu_bo_subtract_pin_size - Remove BO from pin_size accounting
56 *
57 * @bo: &amdgpu_bo buffer object
58 *
59 * This function is called when a BO stops being pinned, and updates the
60 * &amdgpu_device pin_size values accordingly.
61 */
62static void amdgpu_bo_subtract_pin_size(struct amdgpu_bo *bo)
63{
64 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
65
66 if (bo->tbo.mem.mem_type == TTM_PL_VRAM) {
67 atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
68 atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
69 &adev->visible_pin_size);
70 } else if (bo->tbo.mem.mem_type == TTM_PL_TT) {
71 atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size);
72 }
73}
74
75static void amdgpu_bo_destroy(struct ttm_buffer_object *tbo)
76{
77 struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
78 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
79
80 if (bo->pin_count > 0)
81 amdgpu_bo_subtract_pin_size(bo);
82
83 amdgpu_bo_kunmap(bo);
84
85 if (bo->tbo.base.import_attach)
86 drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
87 drm_gem_object_release(&bo->tbo.base);
88 /* in case amdgpu_device_recover_vram got NULL of bo->parent */
89 if (!list_empty(&bo->shadow_list)) {
90 mutex_lock(&adev->shadow_list_lock);
91 list_del_init(&bo->shadow_list);
92 mutex_unlock(&adev->shadow_list_lock);
93 }
94 amdgpu_bo_unref(&bo->parent);
95
96 kfree(bo->metadata);
97 kfree(bo);
98}
99
100/**
101 * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
102 * @bo: buffer object to be checked
103 *
104 * Uses destroy function associated with the object to determine if this is
105 * an &amdgpu_bo.
106 *
107 * Returns:
108 * true if the object belongs to &amdgpu_bo, false if not.
109 */
110bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
111{
112 if (bo->destroy == &amdgpu_bo_destroy)
113 return true;
114 return false;
115}
116
117/**
118 * amdgpu_bo_placement_from_domain - set buffer's placement
119 * @abo: &amdgpu_bo buffer object whose placement is to be set
120 * @domain: requested domain
121 *
122 * Sets buffer's placement according to requested domain and the buffer's
123 * flags.
124 */
125void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
126{
127 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
128 struct ttm_placement *placement = &abo->placement;
129 struct ttm_place *places = abo->placements;
130 u64 flags = abo->flags;
131 u32 c = 0;
132
133 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
134 unsigned visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
135
136 places[c].fpfn = 0;
137 places[c].lpfn = 0;
138 places[c].flags = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
139 TTM_PL_FLAG_VRAM;
140
141 if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
142 places[c].lpfn = visible_pfn;
143 else
144 places[c].flags |= TTM_PL_FLAG_TOPDOWN;
145
146 if (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
147 places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
148 c++;
149 }
150
151 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
152 places[c].fpfn = 0;
153 places[c].lpfn = 0;
154 places[c].flags = TTM_PL_FLAG_TT;
155 if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
156 places[c].flags |= TTM_PL_FLAG_WC |
157 TTM_PL_FLAG_UNCACHED;
158 else
159 places[c].flags |= TTM_PL_FLAG_CACHED;
160 c++;
161 }
162
163 if (domain & AMDGPU_GEM_DOMAIN_CPU) {
164 places[c].fpfn = 0;
165 places[c].lpfn = 0;
166 places[c].flags = TTM_PL_FLAG_SYSTEM;
167 if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
168 places[c].flags |= TTM_PL_FLAG_WC |
169 TTM_PL_FLAG_UNCACHED;
170 else
171 places[c].flags |= TTM_PL_FLAG_CACHED;
172 c++;
173 }
174
175 if (domain & AMDGPU_GEM_DOMAIN_GDS) {
176 places[c].fpfn = 0;
177 places[c].lpfn = 0;
178 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GDS;
179 c++;
180 }
181
182 if (domain & AMDGPU_GEM_DOMAIN_GWS) {
183 places[c].fpfn = 0;
184 places[c].lpfn = 0;
185 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GWS;
186 c++;
187 }
188
189 if (domain & AMDGPU_GEM_DOMAIN_OA) {
190 places[c].fpfn = 0;
191 places[c].lpfn = 0;
192 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_OA;
193 c++;
194 }
195
196 if (!c) {
197 places[c].fpfn = 0;
198 places[c].lpfn = 0;
199 places[c].flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
200 c++;
201 }
202
203 BUG_ON(c >= AMDGPU_BO_MAX_PLACEMENTS);
204
205 placement->num_placement = c;
206 placement->placement = places;
207
208 placement->num_busy_placement = c;
209 placement->busy_placement = places;
210}
211
212/**
213 * amdgpu_bo_create_reserved - create reserved BO for kernel use
214 *
215 * @adev: amdgpu device object
216 * @size: size for the new BO
217 * @align: alignment for the new BO
218 * @domain: where to place it
219 * @bo_ptr: used to initialize BOs in structures
220 * @gpu_addr: GPU addr of the pinned BO
221 * @cpu_addr: optional CPU address mapping
222 *
223 * Allocates and pins a BO for kernel internal use, and returns it still
224 * reserved.
225 *
226 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
227 *
228 * Returns:
229 * 0 on success, negative error code otherwise.
230 */
231int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
232 unsigned long size, int align,
233 u32 domain, struct amdgpu_bo **bo_ptr,
234 u64 *gpu_addr, void **cpu_addr)
235{
236 struct amdgpu_bo_param bp;
237 bool free = false;
238 int r;
239
240 if (!size) {
241 amdgpu_bo_unref(bo_ptr);
242 return 0;
243 }
244
245 memset(&bp, 0, sizeof(bp));
246 bp.size = size;
247 bp.byte_align = align;
248 bp.domain = domain;
249 bp.flags = cpu_addr ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
250 : AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
251 bp.flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
252 bp.type = ttm_bo_type_kernel;
253 bp.resv = NULL;
254
255 if (!*bo_ptr) {
256 r = amdgpu_bo_create(adev, &bp, bo_ptr);
257 if (r) {
258 dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
259 r);
260 return r;
261 }
262 free = true;
263 }
264
265 r = amdgpu_bo_reserve(*bo_ptr, false);
266 if (r) {
267 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
268 goto error_free;
269 }
270
271 r = amdgpu_bo_pin(*bo_ptr, domain);
272 if (r) {
273 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
274 goto error_unreserve;
275 }
276
277 r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
278 if (r) {
279 dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
280 goto error_unpin;
281 }
282
283 if (gpu_addr)
284 *gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
285
286 if (cpu_addr) {
287 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
288 if (r) {
289 dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
290 goto error_unpin;
291 }
292 }
293
294 return 0;
295
296error_unpin:
297 amdgpu_bo_unpin(*bo_ptr);
298error_unreserve:
299 amdgpu_bo_unreserve(*bo_ptr);
300
301error_free:
302 if (free)
303 amdgpu_bo_unref(bo_ptr);
304
305 return r;
306}
307
308/**
309 * amdgpu_bo_create_kernel - create BO for kernel use
310 *
311 * @adev: amdgpu device object
312 * @size: size for the new BO
313 * @align: alignment for the new BO
314 * @domain: where to place it
315 * @bo_ptr: used to initialize BOs in structures
316 * @gpu_addr: GPU addr of the pinned BO
317 * @cpu_addr: optional CPU address mapping
318 *
319 * Allocates and pins a BO for kernel internal use.
320 *
321 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
322 *
323 * Returns:
324 * 0 on success, negative error code otherwise.
325 */
326int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
327 unsigned long size, int align,
328 u32 domain, struct amdgpu_bo **bo_ptr,
329 u64 *gpu_addr, void **cpu_addr)
330{
331 int r;
332
333 r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
334 gpu_addr, cpu_addr);
335
336 if (r)
337 return r;
338
339 if (*bo_ptr)
340 amdgpu_bo_unreserve(*bo_ptr);
341
342 return 0;
343}
344
345/**
346 * amdgpu_bo_free_kernel - free BO for kernel use
347 *
348 * @bo: amdgpu BO to free
349 * @gpu_addr: pointer to where the BO's GPU memory space address was stored
350 * @cpu_addr: pointer to where the BO's CPU memory space address was stored
351 *
352 * unmaps and unpin a BO for kernel internal use.
353 */
354void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
355 void **cpu_addr)
356{
357 if (*bo == NULL)
358 return;
359
360 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
361 if (cpu_addr)
362 amdgpu_bo_kunmap(*bo);
363
364 amdgpu_bo_unpin(*bo);
365 amdgpu_bo_unreserve(*bo);
366 }
367 amdgpu_bo_unref(bo);
368
369 if (gpu_addr)
370 *gpu_addr = 0;
371
372 if (cpu_addr)
373 *cpu_addr = NULL;
374}
375
376/* Validate bo size is bit bigger then the request domain */
377static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
378 unsigned long size, u32 domain)
379{
380 struct ttm_mem_type_manager *man = NULL;
381
382 /*
383 * If GTT is part of requested domains the check must succeed to
384 * allow fall back to GTT
385 */
386 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
387 man = &adev->mman.bdev.man[TTM_PL_TT];
388
389 if (size < (man->size << PAGE_SHIFT))
390 return true;
391 else
392 goto fail;
393 }
394
395 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
396 man = &adev->mman.bdev.man[TTM_PL_VRAM];
397
398 if (size < (man->size << PAGE_SHIFT))
399 return true;
400 else
401 goto fail;
402 }
403
404
405 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */
406 return true;
407
408fail:
409 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size,
410 man->size << PAGE_SHIFT);
411 return false;
412}
413
414bool amdgpu_bo_support_uswc(u64 bo_flags)
415{
416
417#ifdef CONFIG_X86_32
418 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
419 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
420 */
421 return false;
422#elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
423 /* Don't try to enable write-combining when it can't work, or things
424 * may be slow
425 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
426 */
427
428#ifndef CONFIG_COMPILE_TEST
429#warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
430 thanks to write-combining
431#endif
432
433 if (bo_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
434 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
435 "better performance thanks to write-combining\n");
436 return false;
437#else
438 /* For architectures that don't support WC memory,
439 * mask out the WC flag from the BO
440 */
441 if (!drm_arch_can_wc_memory())
442 return false;
443
444 return true;
445#endif
446}
447
448static int amdgpu_bo_do_create(struct amdgpu_device *adev,
449 struct amdgpu_bo_param *bp,
450 struct amdgpu_bo **bo_ptr)
451{
452 struct ttm_operation_ctx ctx = {
453 .interruptible = (bp->type != ttm_bo_type_kernel),
454 .no_wait_gpu = false,
455 .resv = bp->resv,
456 .flags = bp->type != ttm_bo_type_kernel ?
457 TTM_OPT_FLAG_ALLOW_RES_EVICT : 0
458 };
459 struct amdgpu_bo *bo;
460 unsigned long page_align, size = bp->size;
461 size_t acc_size;
462 int r;
463
464 /* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
465 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
466 /* GWS and OA don't need any alignment. */
467 page_align = bp->byte_align;
468 size <<= PAGE_SHIFT;
469 } else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
470 /* Both size and alignment must be a multiple of 4. */
471 page_align = ALIGN(bp->byte_align, 4);
472 size = ALIGN(size, 4) << PAGE_SHIFT;
473 } else {
474 /* Memory should be aligned at least to a page size. */
475 page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
476 size = ALIGN(size, PAGE_SIZE);
477 }
478
479 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
480 return -ENOMEM;
481
482 *bo_ptr = NULL;
483
484 acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
485 sizeof(struct amdgpu_bo));
486
487 bo = kzalloc(sizeof(struct amdgpu_bo), GFP_KERNEL);
488 if (bo == NULL)
489 return -ENOMEM;
490 drm_gem_private_object_init(adev->ddev, &bo->tbo.base, size);
491 INIT_LIST_HEAD(&bo->shadow_list);
492 bo->vm_bo = NULL;
493 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
494 bp->domain;
495 bo->allowed_domains = bo->preferred_domains;
496 if (bp->type != ttm_bo_type_kernel &&
497 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
498 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
499
500 bo->flags = bp->flags;
501
502 if (!amdgpu_bo_support_uswc(bo->flags))
503 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
504
505 bo->tbo.bdev = &adev->mman.bdev;
506 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
507 AMDGPU_GEM_DOMAIN_GDS))
508 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
509 else
510 amdgpu_bo_placement_from_domain(bo, bp->domain);
511 if (bp->type == ttm_bo_type_kernel)
512 bo->tbo.priority = 1;
513
514 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, bp->type,
515 &bo->placement, page_align, &ctx, acc_size,
516 NULL, bp->resv, &amdgpu_bo_destroy);
517 if (unlikely(r != 0))
518 return r;
519
520 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
521 bo->tbo.mem.mem_type == TTM_PL_VRAM &&
522 bo->tbo.mem.start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
523 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
524 ctx.bytes_moved);
525 else
526 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
527
528 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
529 bo->tbo.mem.placement & TTM_PL_FLAG_VRAM) {
530 struct dma_fence *fence;
531
532 r = amdgpu_fill_buffer(bo, 0, bo->tbo.base.resv, &fence);
533 if (unlikely(r))
534 goto fail_unreserve;
535
536 amdgpu_bo_fence(bo, fence, false);
537 dma_fence_put(bo->tbo.moving);
538 bo->tbo.moving = dma_fence_get(fence);
539 dma_fence_put(fence);
540 }
541 if (!bp->resv)
542 amdgpu_bo_unreserve(bo);
543 *bo_ptr = bo;
544
545 trace_amdgpu_bo_create(bo);
546
547 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
548 if (bp->type == ttm_bo_type_device)
549 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
550
551 return 0;
552
553fail_unreserve:
554 if (!bp->resv)
555 dma_resv_unlock(bo->tbo.base.resv);
556 amdgpu_bo_unref(&bo);
557 return r;
558}
559
560static int amdgpu_bo_create_shadow(struct amdgpu_device *adev,
561 unsigned long size,
562 struct amdgpu_bo *bo)
563{
564 struct amdgpu_bo_param bp;
565 int r;
566
567 if (bo->shadow)
568 return 0;
569
570 memset(&bp, 0, sizeof(bp));
571 bp.size = size;
572 bp.domain = AMDGPU_GEM_DOMAIN_GTT;
573 bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC |
574 AMDGPU_GEM_CREATE_SHADOW;
575 bp.type = ttm_bo_type_kernel;
576 bp.resv = bo->tbo.base.resv;
577
578 r = amdgpu_bo_do_create(adev, &bp, &bo->shadow);
579 if (!r) {
580 bo->shadow->parent = amdgpu_bo_ref(bo);
581 mutex_lock(&adev->shadow_list_lock);
582 list_add_tail(&bo->shadow->shadow_list, &adev->shadow_list);
583 mutex_unlock(&adev->shadow_list_lock);
584 }
585
586 return r;
587}
588
589/**
590 * amdgpu_bo_create - create an &amdgpu_bo buffer object
591 * @adev: amdgpu device object
592 * @bp: parameters to be used for the buffer object
593 * @bo_ptr: pointer to the buffer object pointer
594 *
595 * Creates an &amdgpu_bo buffer object; and if requested, also creates a
596 * shadow object.
597 * Shadow object is used to backup the original buffer object, and is always
598 * in GTT.
599 *
600 * Returns:
601 * 0 for success or a negative error code on failure.
602 */
603int amdgpu_bo_create(struct amdgpu_device *adev,
604 struct amdgpu_bo_param *bp,
605 struct amdgpu_bo **bo_ptr)
606{
607 u64 flags = bp->flags;
608 int r;
609
610 bp->flags = bp->flags & ~AMDGPU_GEM_CREATE_SHADOW;
611 r = amdgpu_bo_do_create(adev, bp, bo_ptr);
612 if (r)
613 return r;
614
615 if ((flags & AMDGPU_GEM_CREATE_SHADOW) && !(adev->flags & AMD_IS_APU)) {
616 if (!bp->resv)
617 WARN_ON(dma_resv_lock((*bo_ptr)->tbo.base.resv,
618 NULL));
619
620 r = amdgpu_bo_create_shadow(adev, bp->size, *bo_ptr);
621
622 if (!bp->resv)
623 dma_resv_unlock((*bo_ptr)->tbo.base.resv);
624
625 if (r)
626 amdgpu_bo_unref(bo_ptr);
627 }
628
629 return r;
630}
631
632/**
633 * amdgpu_bo_validate - validate an &amdgpu_bo buffer object
634 * @bo: pointer to the buffer object
635 *
636 * Sets placement according to domain; and changes placement and caching
637 * policy of the buffer object according to the placement.
638 * This is used for validating shadow bos. It calls ttm_bo_validate() to
639 * make sure the buffer is resident where it needs to be.
640 *
641 * Returns:
642 * 0 for success or a negative error code on failure.
643 */
644int amdgpu_bo_validate(struct amdgpu_bo *bo)
645{
646 struct ttm_operation_ctx ctx = { false, false };
647 uint32_t domain;
648 int r;
649
650 if (bo->pin_count)
651 return 0;
652
653 domain = bo->preferred_domains;
654
655retry:
656 amdgpu_bo_placement_from_domain(bo, domain);
657 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
658 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
659 domain = bo->allowed_domains;
660 goto retry;
661 }
662
663 return r;
664}
665
666/**
667 * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow
668 *
669 * @shadow: &amdgpu_bo shadow to be restored
670 * @fence: dma_fence associated with the operation
671 *
672 * Copies a buffer object's shadow content back to the object.
673 * This is used for recovering a buffer from its shadow in case of a gpu
674 * reset where vram context may be lost.
675 *
676 * Returns:
677 * 0 for success or a negative error code on failure.
678 */
679int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence)
680
681{
682 struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev);
683 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
684 uint64_t shadow_addr, parent_addr;
685
686 shadow_addr = amdgpu_bo_gpu_offset(shadow);
687 parent_addr = amdgpu_bo_gpu_offset(shadow->parent);
688
689 return amdgpu_copy_buffer(ring, shadow_addr, parent_addr,
690 amdgpu_bo_size(shadow), NULL, fence,
691 true, false);
692}
693
694/**
695 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
696 * @bo: &amdgpu_bo buffer object to be mapped
697 * @ptr: kernel virtual address to be returned
698 *
699 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
700 * amdgpu_bo_kptr() to get the kernel virtual address.
701 *
702 * Returns:
703 * 0 for success or a negative error code on failure.
704 */
705int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
706{
707 void *kptr;
708 long r;
709
710 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
711 return -EPERM;
712
713 kptr = amdgpu_bo_kptr(bo);
714 if (kptr) {
715 if (ptr)
716 *ptr = kptr;
717 return 0;
718 }
719
720 r = dma_resv_wait_timeout_rcu(bo->tbo.base.resv, false, false,
721 MAX_SCHEDULE_TIMEOUT);
722 if (r < 0)
723 return r;
724
725 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
726 if (r)
727 return r;
728
729 if (ptr)
730 *ptr = amdgpu_bo_kptr(bo);
731
732 return 0;
733}
734
735/**
736 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
737 * @bo: &amdgpu_bo buffer object
738 *
739 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
740 *
741 * Returns:
742 * the virtual address of a buffer object area.
743 */
744void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
745{
746 bool is_iomem;
747
748 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
749}
750
751/**
752 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
753 * @bo: &amdgpu_bo buffer object to be unmapped
754 *
755 * Unmaps a kernel map set up by amdgpu_bo_kmap().
756 */
757void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
758{
759 if (bo->kmap.bo)
760 ttm_bo_kunmap(&bo->kmap);
761}
762
763/**
764 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
765 * @bo: &amdgpu_bo buffer object
766 *
767 * References the contained &ttm_buffer_object.
768 *
769 * Returns:
770 * a refcounted pointer to the &amdgpu_bo buffer object.
771 */
772struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
773{
774 if (bo == NULL)
775 return NULL;
776
777 ttm_bo_get(&bo->tbo);
778 return bo;
779}
780
781/**
782 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
783 * @bo: &amdgpu_bo buffer object
784 *
785 * Unreferences the contained &ttm_buffer_object and clear the pointer
786 */
787void amdgpu_bo_unref(struct amdgpu_bo **bo)
788{
789 struct ttm_buffer_object *tbo;
790
791 if ((*bo) == NULL)
792 return;
793
794 tbo = &((*bo)->tbo);
795 ttm_bo_put(tbo);
796 *bo = NULL;
797}
798
799/**
800 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
801 * @bo: &amdgpu_bo buffer object to be pinned
802 * @domain: domain to be pinned to
803 * @min_offset: the start of requested address range
804 * @max_offset: the end of requested address range
805 *
806 * Pins the buffer object according to requested domain and address range. If
807 * the memory is unbound gart memory, binds the pages into gart table. Adjusts
808 * pin_count and pin_size accordingly.
809 *
810 * Pinning means to lock pages in memory along with keeping them at a fixed
811 * offset. It is required when a buffer can not be moved, for example, when
812 * a display buffer is being scanned out.
813 *
814 * Compared with amdgpu_bo_pin(), this function gives more flexibility on
815 * where to pin a buffer if there are specific restrictions on where a buffer
816 * must be located.
817 *
818 * Returns:
819 * 0 for success or a negative error code on failure.
820 */
821int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
822 u64 min_offset, u64 max_offset)
823{
824 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
825 struct ttm_operation_ctx ctx = { false, false };
826 int r, i;
827
828 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
829 return -EPERM;
830
831 if (WARN_ON_ONCE(min_offset > max_offset))
832 return -EINVAL;
833
834 /* A shared bo cannot be migrated to VRAM */
835 if (bo->prime_shared_count) {
836 if (domain & AMDGPU_GEM_DOMAIN_GTT)
837 domain = AMDGPU_GEM_DOMAIN_GTT;
838 else
839 return -EINVAL;
840 }
841
842 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
843 * See function amdgpu_display_supported_domains()
844 */
845 domain = amdgpu_bo_get_preferred_pin_domain(adev, domain);
846
847 if (bo->pin_count) {
848 uint32_t mem_type = bo->tbo.mem.mem_type;
849
850 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
851 return -EINVAL;
852
853 bo->pin_count++;
854
855 if (max_offset != 0) {
856 u64 domain_start = bo->tbo.bdev->man[mem_type].gpu_offset;
857 WARN_ON_ONCE(max_offset <
858 (amdgpu_bo_gpu_offset(bo) - domain_start));
859 }
860
861 return 0;
862 }
863
864 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
865 /* force to pin into visible video ram */
866 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
867 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
868 amdgpu_bo_placement_from_domain(bo, domain);
869 for (i = 0; i < bo->placement.num_placement; i++) {
870 unsigned fpfn, lpfn;
871
872 fpfn = min_offset >> PAGE_SHIFT;
873 lpfn = max_offset >> PAGE_SHIFT;
874
875 if (fpfn > bo->placements[i].fpfn)
876 bo->placements[i].fpfn = fpfn;
877 if (!bo->placements[i].lpfn ||
878 (lpfn && lpfn < bo->placements[i].lpfn))
879 bo->placements[i].lpfn = lpfn;
880 bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
881 }
882
883 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
884 if (unlikely(r)) {
885 dev_err(adev->dev, "%p pin failed\n", bo);
886 goto error;
887 }
888
889 bo->pin_count = 1;
890
891 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
892 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
893 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
894 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
895 &adev->visible_pin_size);
896 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
897 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
898 }
899
900error:
901 return r;
902}
903
904/**
905 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
906 * @bo: &amdgpu_bo buffer object to be pinned
907 * @domain: domain to be pinned to
908 *
909 * A simple wrapper to amdgpu_bo_pin_restricted().
910 * Provides a simpler API for buffers that do not have any strict restrictions
911 * on where a buffer must be located.
912 *
913 * Returns:
914 * 0 for success or a negative error code on failure.
915 */
916int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
917{
918 return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
919}
920
921/**
922 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
923 * @bo: &amdgpu_bo buffer object to be unpinned
924 *
925 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
926 * Changes placement and pin size accordingly.
927 *
928 * Returns:
929 * 0 for success or a negative error code on failure.
930 */
931int amdgpu_bo_unpin(struct amdgpu_bo *bo)
932{
933 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
934 struct ttm_operation_ctx ctx = { false, false };
935 int r, i;
936
937 if (WARN_ON_ONCE(!bo->pin_count)) {
938 dev_warn(adev->dev, "%p unpin not necessary\n", bo);
939 return 0;
940 }
941 bo->pin_count--;
942 if (bo->pin_count)
943 return 0;
944
945 amdgpu_bo_subtract_pin_size(bo);
946
947 for (i = 0; i < bo->placement.num_placement; i++) {
948 bo->placements[i].lpfn = 0;
949 bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
950 }
951 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
952 if (unlikely(r))
953 dev_err(adev->dev, "%p validate failed for unpin\n", bo);
954
955 return r;
956}
957
958/**
959 * amdgpu_bo_evict_vram - evict VRAM buffers
960 * @adev: amdgpu device object
961 *
962 * Evicts all VRAM buffers on the lru list of the memory type.
963 * Mainly used for evicting vram at suspend time.
964 *
965 * Returns:
966 * 0 for success or a negative error code on failure.
967 */
968int amdgpu_bo_evict_vram(struct amdgpu_device *adev)
969{
970 /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
971#ifndef CONFIG_HIBERNATION
972 if (adev->flags & AMD_IS_APU) {
973 /* Useless to evict on IGP chips */
974 return 0;
975 }
976#endif
977 return ttm_bo_evict_mm(&adev->mman.bdev, TTM_PL_VRAM);
978}
979
980static const char *amdgpu_vram_names[] = {
981 "UNKNOWN",
982 "GDDR1",
983 "DDR2",
984 "GDDR3",
985 "GDDR4",
986 "GDDR5",
987 "HBM",
988 "DDR3",
989 "DDR4",
990 "GDDR6",
991};
992
993/**
994 * amdgpu_bo_init - initialize memory manager
995 * @adev: amdgpu device object
996 *
997 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
998 *
999 * Returns:
1000 * 0 for success or a negative error code on failure.
1001 */
1002int amdgpu_bo_init(struct amdgpu_device *adev)
1003{
1004 /* reserve PAT memory space to WC for VRAM */
1005 arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1006 adev->gmc.aper_size);
1007
1008 /* Add an MTRR for the VRAM */
1009 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1010 adev->gmc.aper_size);
1011 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1012 adev->gmc.mc_vram_size >> 20,
1013 (unsigned long long)adev->gmc.aper_size >> 20);
1014 DRM_INFO("RAM width %dbits %s\n",
1015 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1016 return amdgpu_ttm_init(adev);
1017}
1018
1019/**
1020 * amdgpu_bo_late_init - late init
1021 * @adev: amdgpu device object
1022 *
1023 * Calls amdgpu_ttm_late_init() to free resources used earlier during
1024 * initialization.
1025 *
1026 * Returns:
1027 * 0 for success or a negative error code on failure.
1028 */
1029int amdgpu_bo_late_init(struct amdgpu_device *adev)
1030{
1031 amdgpu_ttm_late_init(adev);
1032
1033 return 0;
1034}
1035
1036/**
1037 * amdgpu_bo_fini - tear down memory manager
1038 * @adev: amdgpu device object
1039 *
1040 * Reverses amdgpu_bo_init() to tear down memory manager.
1041 */
1042void amdgpu_bo_fini(struct amdgpu_device *adev)
1043{
1044 amdgpu_ttm_fini(adev);
1045 arch_phys_wc_del(adev->gmc.vram_mtrr);
1046 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1047}
1048
1049/**
1050 * amdgpu_bo_fbdev_mmap - mmap fbdev memory
1051 * @bo: &amdgpu_bo buffer object
1052 * @vma: vma as input from the fbdev mmap method
1053 *
1054 * Calls ttm_fbdev_mmap() to mmap fbdev memory if it is backed by a bo.
1055 *
1056 * Returns:
1057 * 0 for success or a negative error code on failure.
1058 */
1059int amdgpu_bo_fbdev_mmap(struct amdgpu_bo *bo,
1060 struct vm_area_struct *vma)
1061{
1062 return ttm_fbdev_mmap(vma, &bo->tbo);
1063}
1064
1065/**
1066 * amdgpu_bo_set_tiling_flags - set tiling flags
1067 * @bo: &amdgpu_bo buffer object
1068 * @tiling_flags: new flags
1069 *
1070 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1071 * kernel driver to set the tiling flags on a buffer.
1072 *
1073 * Returns:
1074 * 0 for success or a negative error code on failure.
1075 */
1076int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1077{
1078 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1079
1080 if (adev->family <= AMDGPU_FAMILY_CZ &&
1081 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1082 return -EINVAL;
1083
1084 bo->tiling_flags = tiling_flags;
1085 return 0;
1086}
1087
1088/**
1089 * amdgpu_bo_get_tiling_flags - get tiling flags
1090 * @bo: &amdgpu_bo buffer object
1091 * @tiling_flags: returned flags
1092 *
1093 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1094 * set the tiling flags on a buffer.
1095 */
1096void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1097{
1098 dma_resv_assert_held(bo->tbo.base.resv);
1099
1100 if (tiling_flags)
1101 *tiling_flags = bo->tiling_flags;
1102}
1103
1104/**
1105 * amdgpu_bo_set_metadata - set metadata
1106 * @bo: &amdgpu_bo buffer object
1107 * @metadata: new metadata
1108 * @metadata_size: size of the new metadata
1109 * @flags: flags of the new metadata
1110 *
1111 * Sets buffer object's metadata, its size and flags.
1112 * Used via GEM ioctl.
1113 *
1114 * Returns:
1115 * 0 for success or a negative error code on failure.
1116 */
1117int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
1118 uint32_t metadata_size, uint64_t flags)
1119{
1120 void *buffer;
1121
1122 if (!metadata_size) {
1123 if (bo->metadata_size) {
1124 kfree(bo->metadata);
1125 bo->metadata = NULL;
1126 bo->metadata_size = 0;
1127 }
1128 return 0;
1129 }
1130
1131 if (metadata == NULL)
1132 return -EINVAL;
1133
1134 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1135 if (buffer == NULL)
1136 return -ENOMEM;
1137
1138 kfree(bo->metadata);
1139 bo->metadata_flags = flags;
1140 bo->metadata = buffer;
1141 bo->metadata_size = metadata_size;
1142
1143 return 0;
1144}
1145
1146/**
1147 * amdgpu_bo_get_metadata - get metadata
1148 * @bo: &amdgpu_bo buffer object
1149 * @buffer: returned metadata
1150 * @buffer_size: size of the buffer
1151 * @metadata_size: size of the returned metadata
1152 * @flags: flags of the returned metadata
1153 *
1154 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1155 * less than metadata_size.
1156 * Used via GEM ioctl.
1157 *
1158 * Returns:
1159 * 0 for success or a negative error code on failure.
1160 */
1161int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1162 size_t buffer_size, uint32_t *metadata_size,
1163 uint64_t *flags)
1164{
1165 if (!buffer && !metadata_size)
1166 return -EINVAL;
1167
1168 if (buffer) {
1169 if (buffer_size < bo->metadata_size)
1170 return -EINVAL;
1171
1172 if (bo->metadata_size)
1173 memcpy(buffer, bo->metadata, bo->metadata_size);
1174 }
1175
1176 if (metadata_size)
1177 *metadata_size = bo->metadata_size;
1178 if (flags)
1179 *flags = bo->metadata_flags;
1180
1181 return 0;
1182}
1183
1184/**
1185 * amdgpu_bo_move_notify - notification about a memory move
1186 * @bo: pointer to a buffer object
1187 * @evict: if this move is evicting the buffer from the graphics address space
1188 * @new_mem: new information of the bufer object
1189 *
1190 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1191 * bookkeeping.
1192 * TTM driver callback which is called when ttm moves a buffer.
1193 */
1194void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1195 bool evict,
1196 struct ttm_mem_reg *new_mem)
1197{
1198 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1199 struct amdgpu_bo *abo;
1200 struct ttm_mem_reg *old_mem = &bo->mem;
1201
1202 if (!amdgpu_bo_is_amdgpu_bo(bo))
1203 return;
1204
1205 abo = ttm_to_amdgpu_bo(bo);
1206 amdgpu_vm_bo_invalidate(adev, abo, evict);
1207
1208 amdgpu_bo_kunmap(abo);
1209
1210 /* remember the eviction */
1211 if (evict)
1212 atomic64_inc(&adev->num_evictions);
1213
1214 /* update statistics */
1215 if (!new_mem)
1216 return;
1217
1218 /* move_notify is called before move happens */
1219 trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
1220}
1221
1222/**
1223 * amdgpu_bo_move_notify - notification about a BO being released
1224 * @bo: pointer to a buffer object
1225 *
1226 * Wipes VRAM buffers whose contents should not be leaked before the
1227 * memory is released.
1228 */
1229void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)
1230{
1231 struct dma_fence *fence = NULL;
1232 struct amdgpu_bo *abo;
1233 int r;
1234
1235 if (!amdgpu_bo_is_amdgpu_bo(bo))
1236 return;
1237
1238 abo = ttm_to_amdgpu_bo(bo);
1239
1240 if (abo->kfd_bo)
1241 amdgpu_amdkfd_unreserve_memory_limit(abo);
1242
1243 if (bo->mem.mem_type != TTM_PL_VRAM || !bo->mem.mm_node ||
1244 !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE))
1245 return;
1246
1247 dma_resv_lock(bo->base.resv, NULL);
1248
1249 r = amdgpu_fill_buffer(abo, AMDGPU_POISON, bo->base.resv, &fence);
1250 if (!WARN_ON(r)) {
1251 amdgpu_bo_fence(abo, fence, false);
1252 dma_fence_put(fence);
1253 }
1254
1255 dma_resv_unlock(bo->base.resv);
1256}
1257
1258/**
1259 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1260 * @bo: pointer to a buffer object
1261 *
1262 * Notifies the driver we are taking a fault on this BO and have reserved it,
1263 * also performs bookkeeping.
1264 * TTM driver callback for dealing with vm faults.
1265 *
1266 * Returns:
1267 * 0 for success or a negative error code on failure.
1268 */
1269int amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1270{
1271 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1272 struct ttm_operation_ctx ctx = { false, false };
1273 struct amdgpu_bo *abo;
1274 unsigned long offset, size;
1275 int r;
1276
1277 if (!amdgpu_bo_is_amdgpu_bo(bo))
1278 return 0;
1279
1280 abo = ttm_to_amdgpu_bo(bo);
1281
1282 /* Remember that this BO was accessed by the CPU */
1283 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1284
1285 if (bo->mem.mem_type != TTM_PL_VRAM)
1286 return 0;
1287
1288 size = bo->mem.num_pages << PAGE_SHIFT;
1289 offset = bo->mem.start << PAGE_SHIFT;
1290 if ((offset + size) <= adev->gmc.visible_vram_size)
1291 return 0;
1292
1293 /* Can't move a pinned BO to visible VRAM */
1294 if (abo->pin_count > 0)
1295 return -EINVAL;
1296
1297 /* hurrah the memory is not visible ! */
1298 atomic64_inc(&adev->num_vram_cpu_page_faults);
1299 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1300 AMDGPU_GEM_DOMAIN_GTT);
1301
1302 /* Avoid costly evictions; only set GTT as a busy placement */
1303 abo->placement.num_busy_placement = 1;
1304 abo->placement.busy_placement = &abo->placements[1];
1305
1306 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1307 if (unlikely(r != 0))
1308 return r;
1309
1310 offset = bo->mem.start << PAGE_SHIFT;
1311 /* this should never happen */
1312 if (bo->mem.mem_type == TTM_PL_VRAM &&
1313 (offset + size) > adev->gmc.visible_vram_size)
1314 return -EINVAL;
1315
1316 return 0;
1317}
1318
1319/**
1320 * amdgpu_bo_fence - add fence to buffer object
1321 *
1322 * @bo: buffer object in question
1323 * @fence: fence to add
1324 * @shared: true if fence should be added shared
1325 *
1326 */
1327void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1328 bool shared)
1329{
1330 struct dma_resv *resv = bo->tbo.base.resv;
1331
1332 if (shared)
1333 dma_resv_add_shared_fence(resv, fence);
1334 else
1335 dma_resv_add_excl_fence(resv, fence);
1336}
1337
1338/**
1339 * amdgpu_sync_wait_resv - Wait for BO reservation fences
1340 *
1341 * @bo: buffer object
1342 * @owner: fence owner
1343 * @intr: Whether the wait is interruptible
1344 *
1345 * Returns:
1346 * 0 on success, errno otherwise.
1347 */
1348int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)
1349{
1350 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1351 struct amdgpu_sync sync;
1352 int r;
1353
1354 amdgpu_sync_create(&sync);
1355 amdgpu_sync_resv(adev, &sync, bo->tbo.base.resv, owner, false);
1356 r = amdgpu_sync_wait(&sync, intr);
1357 amdgpu_sync_free(&sync);
1358
1359 return r;
1360}
1361
1362/**
1363 * amdgpu_bo_gpu_offset - return GPU offset of bo
1364 * @bo: amdgpu object for which we query the offset
1365 *
1366 * Note: object should either be pinned or reserved when calling this
1367 * function, it might be useful to add check for this for debugging.
1368 *
1369 * Returns:
1370 * current GPU offset of the object.
1371 */
1372u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1373{
1374 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_SYSTEM);
1375 WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) &&
1376 !bo->pin_count && bo->tbo.type != ttm_bo_type_kernel);
1377 WARN_ON_ONCE(bo->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET);
1378 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_VRAM &&
1379 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1380
1381 return amdgpu_gmc_sign_extend(bo->tbo.offset);
1382}
1383
1384/**
1385 * amdgpu_bo_get_preferred_pin_domain - get preferred domain for scanout
1386 * @adev: amdgpu device object
1387 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1388 *
1389 * Returns:
1390 * Which of the allowed domains is preferred for pinning the BO for scanout.
1391 */
1392uint32_t amdgpu_bo_get_preferred_pin_domain(struct amdgpu_device *adev,
1393 uint32_t domain)
1394{
1395 if (domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) {
1396 domain = AMDGPU_GEM_DOMAIN_VRAM;
1397 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1398 domain = AMDGPU_GEM_DOMAIN_GTT;
1399 }
1400 return domain;
1401}