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 <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/**
56 * amdgpu_bo_subtract_pin_size - Remove BO from pin_size accounting
57 *
58 * @bo: &amdgpu_bo buffer object
59 *
60 * This function is called when a BO stops being pinned, and updates the
61 * &amdgpu_device pin_size values accordingly.
62 */
63static void amdgpu_bo_subtract_pin_size(struct amdgpu_bo *bo)
64{
65 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
66
67 if (bo->tbo.mem.mem_type == TTM_PL_VRAM) {
68 atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
69 atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
70 &adev->visible_pin_size);
71 } else if (bo->tbo.mem.mem_type == TTM_PL_TT) {
72 atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size);
73 }
74}
75
76static void amdgpu_bo_destroy(struct ttm_buffer_object *tbo)
77{
78 struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
79 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
80
81 if (bo->pin_count > 0)
82 amdgpu_bo_subtract_pin_size(bo);
83
84 amdgpu_bo_kunmap(bo);
85
86 if (bo->tbo.base.import_attach)
87 drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
88 drm_gem_object_release(&bo->tbo.base);
89 /* in case amdgpu_device_recover_vram got NULL of bo->parent */
90 if (!list_empty(&bo->shadow_list)) {
91 mutex_lock(&adev->shadow_list_lock);
92 list_del_init(&bo->shadow_list);
93 mutex_unlock(&adev->shadow_list_lock);
94 }
95 amdgpu_bo_unref(&bo->parent);
96
97 kfree(bo->metadata);
98 kfree(bo);
99}
100
101/**
102 * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
103 * @bo: buffer object to be checked
104 *
105 * Uses destroy function associated with the object to determine if this is
106 * an &amdgpu_bo.
107 *
108 * Returns:
109 * true if the object belongs to &amdgpu_bo, false if not.
110 */
111bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
112{
113 if (bo->destroy == &amdgpu_bo_destroy)
114 return true;
115 return false;
116}
117
118/**
119 * amdgpu_bo_placement_from_domain - set buffer's placement
120 * @abo: &amdgpu_bo buffer object whose placement is to be set
121 * @domain: requested domain
122 *
123 * Sets buffer's placement according to requested domain and the buffer's
124 * flags.
125 */
126void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
127{
128 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
129 struct ttm_placement *placement = &abo->placement;
130 struct ttm_place *places = abo->placements;
131 u64 flags = abo->flags;
132 u32 c = 0;
133
134 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
135 unsigned visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
136
137 places[c].fpfn = 0;
138 places[c].lpfn = 0;
139 places[c].flags = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
140 TTM_PL_FLAG_VRAM;
141
142 if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
143 places[c].lpfn = visible_pfn;
144 else
145 places[c].flags |= TTM_PL_FLAG_TOPDOWN;
146
147 if (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
148 places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
149 c++;
150 }
151
152 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
153 places[c].fpfn = 0;
154 places[c].lpfn = 0;
155 places[c].flags = TTM_PL_FLAG_TT;
156 if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
157 places[c].flags |= TTM_PL_FLAG_WC |
158 TTM_PL_FLAG_UNCACHED;
159 else
160 places[c].flags |= TTM_PL_FLAG_CACHED;
161 c++;
162 }
163
164 if (domain & AMDGPU_GEM_DOMAIN_CPU) {
165 places[c].fpfn = 0;
166 places[c].lpfn = 0;
167 places[c].flags = TTM_PL_FLAG_SYSTEM;
168 if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
169 places[c].flags |= TTM_PL_FLAG_WC |
170 TTM_PL_FLAG_UNCACHED;
171 else
172 places[c].flags |= TTM_PL_FLAG_CACHED;
173 c++;
174 }
175
176 if (domain & AMDGPU_GEM_DOMAIN_GDS) {
177 places[c].fpfn = 0;
178 places[c].lpfn = 0;
179 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GDS;
180 c++;
181 }
182
183 if (domain & AMDGPU_GEM_DOMAIN_GWS) {
184 places[c].fpfn = 0;
185 places[c].lpfn = 0;
186 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GWS;
187 c++;
188 }
189
190 if (domain & AMDGPU_GEM_DOMAIN_OA) {
191 places[c].fpfn = 0;
192 places[c].lpfn = 0;
193 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_OA;
194 c++;
195 }
196
197 if (!c) {
198 places[c].fpfn = 0;
199 places[c].lpfn = 0;
200 places[c].flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
201 c++;
202 }
203
204 BUG_ON(c >= AMDGPU_BO_MAX_PLACEMENTS);
205
206 placement->num_placement = c;
207 placement->placement = places;
208
209 placement->num_busy_placement = c;
210 placement->busy_placement = places;
211}
212
213/**
214 * amdgpu_bo_create_reserved - create reserved BO for kernel use
215 *
216 * @adev: amdgpu device object
217 * @size: size for the new BO
218 * @align: alignment for the new BO
219 * @domain: where to place it
220 * @bo_ptr: used to initialize BOs in structures
221 * @gpu_addr: GPU addr of the pinned BO
222 * @cpu_addr: optional CPU address mapping
223 *
224 * Allocates and pins a BO for kernel internal use, and returns it still
225 * reserved.
226 *
227 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
228 *
229 * Returns:
230 * 0 on success, negative error code otherwise.
231 */
232int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
233 unsigned long size, int align,
234 u32 domain, struct amdgpu_bo **bo_ptr,
235 u64 *gpu_addr, void **cpu_addr)
236{
237 struct amdgpu_bo_param bp;
238 bool free = false;
239 int r;
240
241 if (!size) {
242 amdgpu_bo_unref(bo_ptr);
243 return 0;
244 }
245
246 memset(&bp, 0, sizeof(bp));
247 bp.size = size;
248 bp.byte_align = align;
249 bp.domain = domain;
250 bp.flags = cpu_addr ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
251 : AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
252 bp.flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
253 bp.type = ttm_bo_type_kernel;
254 bp.resv = NULL;
255
256 if (!*bo_ptr) {
257 r = amdgpu_bo_create(adev, &bp, bo_ptr);
258 if (r) {
259 dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
260 r);
261 return r;
262 }
263 free = true;
264 }
265
266 r = amdgpu_bo_reserve(*bo_ptr, false);
267 if (r) {
268 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
269 goto error_free;
270 }
271
272 r = amdgpu_bo_pin(*bo_ptr, domain);
273 if (r) {
274 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
275 goto error_unreserve;
276 }
277
278 r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
279 if (r) {
280 dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
281 goto error_unpin;
282 }
283
284 if (gpu_addr)
285 *gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
286
287 if (cpu_addr) {
288 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
289 if (r) {
290 dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
291 goto error_unpin;
292 }
293 }
294
295 return 0;
296
297error_unpin:
298 amdgpu_bo_unpin(*bo_ptr);
299error_unreserve:
300 amdgpu_bo_unreserve(*bo_ptr);
301
302error_free:
303 if (free)
304 amdgpu_bo_unref(bo_ptr);
305
306 return r;
307}
308
309/**
310 * amdgpu_bo_create_kernel - create BO for kernel use
311 *
312 * @adev: amdgpu device object
313 * @size: size for the new BO
314 * @align: alignment for the new BO
315 * @domain: where to place it
316 * @bo_ptr: used to initialize BOs in structures
317 * @gpu_addr: GPU addr of the pinned BO
318 * @cpu_addr: optional CPU address mapping
319 *
320 * Allocates and pins a BO for kernel internal use.
321 *
322 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
323 *
324 * Returns:
325 * 0 on success, negative error code otherwise.
326 */
327int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
328 unsigned long size, int align,
329 u32 domain, struct amdgpu_bo **bo_ptr,
330 u64 *gpu_addr, void **cpu_addr)
331{
332 int r;
333
334 r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
335 gpu_addr, cpu_addr);
336
337 if (r)
338 return r;
339
340 if (*bo_ptr)
341 amdgpu_bo_unreserve(*bo_ptr);
342
343 return 0;
344}
345
346/**
347 * amdgpu_bo_create_kernel_at - create BO for kernel use at specific location
348 *
349 * @adev: amdgpu device object
350 * @offset: offset of the BO
351 * @size: size of the BO
352 * @domain: where to place it
353 * @bo_ptr: used to initialize BOs in structures
354 * @cpu_addr: optional CPU address mapping
355 *
356 * Creates a kernel BO at a specific offset in the address space of the domain.
357 *
358 * Returns:
359 * 0 on success, negative error code otherwise.
360 */
361int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev,
362 uint64_t offset, uint64_t size, uint32_t domain,
363 struct amdgpu_bo **bo_ptr, void **cpu_addr)
364{
365 struct ttm_operation_ctx ctx = { false, false };
366 unsigned int i;
367 int r;
368
369 offset &= PAGE_MASK;
370 size = ALIGN(size, PAGE_SIZE);
371
372 r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE, domain, bo_ptr,
373 NULL, cpu_addr);
374 if (r)
375 return r;
376
377 /*
378 * Remove the original mem node and create a new one at the request
379 * position.
380 */
381 if (cpu_addr)
382 amdgpu_bo_kunmap(*bo_ptr);
383
384 ttm_bo_mem_put(&(*bo_ptr)->tbo, &(*bo_ptr)->tbo.mem);
385
386 for (i = 0; i < (*bo_ptr)->placement.num_placement; ++i) {
387 (*bo_ptr)->placements[i].fpfn = offset >> PAGE_SHIFT;
388 (*bo_ptr)->placements[i].lpfn = (offset + size) >> PAGE_SHIFT;
389 }
390 r = ttm_bo_mem_space(&(*bo_ptr)->tbo, &(*bo_ptr)->placement,
391 &(*bo_ptr)->tbo.mem, &ctx);
392 if (r)
393 goto error;
394
395 if (cpu_addr) {
396 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
397 if (r)
398 goto error;
399 }
400
401 amdgpu_bo_unreserve(*bo_ptr);
402 return 0;
403
404error:
405 amdgpu_bo_unreserve(*bo_ptr);
406 amdgpu_bo_unref(bo_ptr);
407 return r;
408}
409
410/**
411 * amdgpu_bo_free_kernel - free BO for kernel use
412 *
413 * @bo: amdgpu BO to free
414 * @gpu_addr: pointer to where the BO's GPU memory space address was stored
415 * @cpu_addr: pointer to where the BO's CPU memory space address was stored
416 *
417 * unmaps and unpin a BO for kernel internal use.
418 */
419void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
420 void **cpu_addr)
421{
422 if (*bo == NULL)
423 return;
424
425 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
426 if (cpu_addr)
427 amdgpu_bo_kunmap(*bo);
428
429 amdgpu_bo_unpin(*bo);
430 amdgpu_bo_unreserve(*bo);
431 }
432 amdgpu_bo_unref(bo);
433
434 if (gpu_addr)
435 *gpu_addr = 0;
436
437 if (cpu_addr)
438 *cpu_addr = NULL;
439}
440
441/* Validate bo size is bit bigger then the request domain */
442static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
443 unsigned long size, u32 domain)
444{
445 struct ttm_mem_type_manager *man = NULL;
446
447 /*
448 * If GTT is part of requested domains the check must succeed to
449 * allow fall back to GTT
450 */
451 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
452 man = &adev->mman.bdev.man[TTM_PL_TT];
453
454 if (size < (man->size << PAGE_SHIFT))
455 return true;
456 else
457 goto fail;
458 }
459
460 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
461 man = &adev->mman.bdev.man[TTM_PL_VRAM];
462
463 if (size < (man->size << PAGE_SHIFT))
464 return true;
465 else
466 goto fail;
467 }
468
469
470 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */
471 return true;
472
473fail:
474 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size,
475 man->size << PAGE_SHIFT);
476 return false;
477}
478
479bool amdgpu_bo_support_uswc(u64 bo_flags)
480{
481
482#ifdef CONFIG_X86_32
483 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
484 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
485 */
486 return false;
487#elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
488 /* Don't try to enable write-combining when it can't work, or things
489 * may be slow
490 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
491 */
492
493#ifndef CONFIG_COMPILE_TEST
494#warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
495 thanks to write-combining
496#endif
497
498 if (bo_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
499 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
500 "better performance thanks to write-combining\n");
501 return false;
502#else
503 /* For architectures that don't support WC memory,
504 * mask out the WC flag from the BO
505 */
506 if (!drm_arch_can_wc_memory())
507 return false;
508
509 return true;
510#endif
511}
512
513static int amdgpu_bo_do_create(struct amdgpu_device *adev,
514 struct amdgpu_bo_param *bp,
515 struct amdgpu_bo **bo_ptr)
516{
517 struct ttm_operation_ctx ctx = {
518 .interruptible = (bp->type != ttm_bo_type_kernel),
519 .no_wait_gpu = bp->no_wait_gpu,
520 .resv = bp->resv,
521 .flags = bp->type != ttm_bo_type_kernel ?
522 TTM_OPT_FLAG_ALLOW_RES_EVICT : 0
523 };
524 struct amdgpu_bo *bo;
525 unsigned long page_align, size = bp->size;
526 size_t acc_size;
527 int r;
528
529 /* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
530 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
531 /* GWS and OA don't need any alignment. */
532 page_align = bp->byte_align;
533 size <<= PAGE_SHIFT;
534 } else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
535 /* Both size and alignment must be a multiple of 4. */
536 page_align = ALIGN(bp->byte_align, 4);
537 size = ALIGN(size, 4) << PAGE_SHIFT;
538 } else {
539 /* Memory should be aligned at least to a page size. */
540 page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
541 size = ALIGN(size, PAGE_SIZE);
542 }
543
544 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
545 return -ENOMEM;
546
547 *bo_ptr = NULL;
548
549 acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
550 sizeof(struct amdgpu_bo));
551
552 bo = kzalloc(sizeof(struct amdgpu_bo), GFP_KERNEL);
553 if (bo == NULL)
554 return -ENOMEM;
555 drm_gem_private_object_init(adev->ddev, &bo->tbo.base, size);
556 INIT_LIST_HEAD(&bo->shadow_list);
557 bo->vm_bo = NULL;
558 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
559 bp->domain;
560 bo->allowed_domains = bo->preferred_domains;
561 if (bp->type != ttm_bo_type_kernel &&
562 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
563 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
564
565 bo->flags = bp->flags;
566
567 if (!amdgpu_bo_support_uswc(bo->flags))
568 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
569
570 bo->tbo.bdev = &adev->mman.bdev;
571 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
572 AMDGPU_GEM_DOMAIN_GDS))
573 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
574 else
575 amdgpu_bo_placement_from_domain(bo, bp->domain);
576 if (bp->type == ttm_bo_type_kernel)
577 bo->tbo.priority = 1;
578
579 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, bp->type,
580 &bo->placement, page_align, &ctx, acc_size,
581 NULL, bp->resv, &amdgpu_bo_destroy);
582 if (unlikely(r != 0))
583 return r;
584
585 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
586 bo->tbo.mem.mem_type == TTM_PL_VRAM &&
587 bo->tbo.mem.start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
588 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
589 ctx.bytes_moved);
590 else
591 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
592
593 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
594 bo->tbo.mem.placement & TTM_PL_FLAG_VRAM) {
595 struct dma_fence *fence;
596
597 r = amdgpu_fill_buffer(bo, 0, bo->tbo.base.resv, &fence);
598 if (unlikely(r))
599 goto fail_unreserve;
600
601 amdgpu_bo_fence(bo, fence, false);
602 dma_fence_put(bo->tbo.moving);
603 bo->tbo.moving = dma_fence_get(fence);
604 dma_fence_put(fence);
605 }
606 if (!bp->resv)
607 amdgpu_bo_unreserve(bo);
608 *bo_ptr = bo;
609
610 trace_amdgpu_bo_create(bo);
611
612 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
613 if (bp->type == ttm_bo_type_device)
614 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
615
616 return 0;
617
618fail_unreserve:
619 if (!bp->resv)
620 dma_resv_unlock(bo->tbo.base.resv);
621 amdgpu_bo_unref(&bo);
622 return r;
623}
624
625static int amdgpu_bo_create_shadow(struct amdgpu_device *adev,
626 unsigned long size,
627 struct amdgpu_bo *bo)
628{
629 struct amdgpu_bo_param bp;
630 int r;
631
632 if (bo->shadow)
633 return 0;
634
635 memset(&bp, 0, sizeof(bp));
636 bp.size = size;
637 bp.domain = AMDGPU_GEM_DOMAIN_GTT;
638 bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC |
639 AMDGPU_GEM_CREATE_SHADOW;
640 bp.type = ttm_bo_type_kernel;
641 bp.resv = bo->tbo.base.resv;
642
643 r = amdgpu_bo_do_create(adev, &bp, &bo->shadow);
644 if (!r) {
645 bo->shadow->parent = amdgpu_bo_ref(bo);
646 mutex_lock(&adev->shadow_list_lock);
647 list_add_tail(&bo->shadow->shadow_list, &adev->shadow_list);
648 mutex_unlock(&adev->shadow_list_lock);
649 }
650
651 return r;
652}
653
654/**
655 * amdgpu_bo_create - create an &amdgpu_bo buffer object
656 * @adev: amdgpu device object
657 * @bp: parameters to be used for the buffer object
658 * @bo_ptr: pointer to the buffer object pointer
659 *
660 * Creates an &amdgpu_bo buffer object; and if requested, also creates a
661 * shadow object.
662 * Shadow object is used to backup the original buffer object, and is always
663 * in GTT.
664 *
665 * Returns:
666 * 0 for success or a negative error code on failure.
667 */
668int amdgpu_bo_create(struct amdgpu_device *adev,
669 struct amdgpu_bo_param *bp,
670 struct amdgpu_bo **bo_ptr)
671{
672 u64 flags = bp->flags;
673 int r;
674
675 bp->flags = bp->flags & ~AMDGPU_GEM_CREATE_SHADOW;
676 r = amdgpu_bo_do_create(adev, bp, bo_ptr);
677 if (r)
678 return r;
679
680 if ((flags & AMDGPU_GEM_CREATE_SHADOW) && !(adev->flags & AMD_IS_APU)) {
681 if (!bp->resv)
682 WARN_ON(dma_resv_lock((*bo_ptr)->tbo.base.resv,
683 NULL));
684
685 r = amdgpu_bo_create_shadow(adev, bp->size, *bo_ptr);
686
687 if (!bp->resv)
688 dma_resv_unlock((*bo_ptr)->tbo.base.resv);
689
690 if (r)
691 amdgpu_bo_unref(bo_ptr);
692 }
693
694 return r;
695}
696
697/**
698 * amdgpu_bo_validate - validate an &amdgpu_bo buffer object
699 * @bo: pointer to the buffer object
700 *
701 * Sets placement according to domain; and changes placement and caching
702 * policy of the buffer object according to the placement.
703 * This is used for validating shadow bos. It calls ttm_bo_validate() to
704 * make sure the buffer is resident where it needs to be.
705 *
706 * Returns:
707 * 0 for success or a negative error code on failure.
708 */
709int amdgpu_bo_validate(struct amdgpu_bo *bo)
710{
711 struct ttm_operation_ctx ctx = { false, false };
712 uint32_t domain;
713 int r;
714
715 if (bo->pin_count)
716 return 0;
717
718 domain = bo->preferred_domains;
719
720retry:
721 amdgpu_bo_placement_from_domain(bo, domain);
722 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
723 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
724 domain = bo->allowed_domains;
725 goto retry;
726 }
727
728 return r;
729}
730
731/**
732 * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow
733 *
734 * @shadow: &amdgpu_bo shadow to be restored
735 * @fence: dma_fence associated with the operation
736 *
737 * Copies a buffer object's shadow content back to the object.
738 * This is used for recovering a buffer from its shadow in case of a gpu
739 * reset where vram context may be lost.
740 *
741 * Returns:
742 * 0 for success or a negative error code on failure.
743 */
744int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence)
745
746{
747 struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev);
748 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
749 uint64_t shadow_addr, parent_addr;
750
751 shadow_addr = amdgpu_bo_gpu_offset(shadow);
752 parent_addr = amdgpu_bo_gpu_offset(shadow->parent);
753
754 return amdgpu_copy_buffer(ring, shadow_addr, parent_addr,
755 amdgpu_bo_size(shadow), NULL, fence,
756 true, false, false);
757}
758
759/**
760 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
761 * @bo: &amdgpu_bo buffer object to be mapped
762 * @ptr: kernel virtual address to be returned
763 *
764 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
765 * amdgpu_bo_kptr() to get the kernel virtual address.
766 *
767 * Returns:
768 * 0 for success or a negative error code on failure.
769 */
770int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
771{
772 void *kptr;
773 long r;
774
775 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
776 return -EPERM;
777
778 kptr = amdgpu_bo_kptr(bo);
779 if (kptr) {
780 if (ptr)
781 *ptr = kptr;
782 return 0;
783 }
784
785 r = dma_resv_wait_timeout_rcu(bo->tbo.base.resv, false, false,
786 MAX_SCHEDULE_TIMEOUT);
787 if (r < 0)
788 return r;
789
790 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
791 if (r)
792 return r;
793
794 if (ptr)
795 *ptr = amdgpu_bo_kptr(bo);
796
797 return 0;
798}
799
800/**
801 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
802 * @bo: &amdgpu_bo buffer object
803 *
804 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
805 *
806 * Returns:
807 * the virtual address of a buffer object area.
808 */
809void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
810{
811 bool is_iomem;
812
813 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
814}
815
816/**
817 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
818 * @bo: &amdgpu_bo buffer object to be unmapped
819 *
820 * Unmaps a kernel map set up by amdgpu_bo_kmap().
821 */
822void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
823{
824 if (bo->kmap.bo)
825 ttm_bo_kunmap(&bo->kmap);
826}
827
828/**
829 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
830 * @bo: &amdgpu_bo buffer object
831 *
832 * References the contained &ttm_buffer_object.
833 *
834 * Returns:
835 * a refcounted pointer to the &amdgpu_bo buffer object.
836 */
837struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
838{
839 if (bo == NULL)
840 return NULL;
841
842 ttm_bo_get(&bo->tbo);
843 return bo;
844}
845
846/**
847 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
848 * @bo: &amdgpu_bo buffer object
849 *
850 * Unreferences the contained &ttm_buffer_object and clear the pointer
851 */
852void amdgpu_bo_unref(struct amdgpu_bo **bo)
853{
854 struct ttm_buffer_object *tbo;
855
856 if ((*bo) == NULL)
857 return;
858
859 tbo = &((*bo)->tbo);
860 ttm_bo_put(tbo);
861 *bo = NULL;
862}
863
864/**
865 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
866 * @bo: &amdgpu_bo buffer object to be pinned
867 * @domain: domain to be pinned to
868 * @min_offset: the start of requested address range
869 * @max_offset: the end of requested address range
870 *
871 * Pins the buffer object according to requested domain and address range. If
872 * the memory is unbound gart memory, binds the pages into gart table. Adjusts
873 * pin_count and pin_size accordingly.
874 *
875 * Pinning means to lock pages in memory along with keeping them at a fixed
876 * offset. It is required when a buffer can not be moved, for example, when
877 * a display buffer is being scanned out.
878 *
879 * Compared with amdgpu_bo_pin(), this function gives more flexibility on
880 * where to pin a buffer if there are specific restrictions on where a buffer
881 * must be located.
882 *
883 * Returns:
884 * 0 for success or a negative error code on failure.
885 */
886int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
887 u64 min_offset, u64 max_offset)
888{
889 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
890 struct ttm_operation_ctx ctx = { false, false };
891 int r, i;
892
893 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
894 return -EPERM;
895
896 if (WARN_ON_ONCE(min_offset > max_offset))
897 return -EINVAL;
898
899 /* A shared bo cannot be migrated to VRAM */
900 if (bo->prime_shared_count) {
901 if (domain & AMDGPU_GEM_DOMAIN_GTT)
902 domain = AMDGPU_GEM_DOMAIN_GTT;
903 else
904 return -EINVAL;
905 }
906
907 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
908 * See function amdgpu_display_supported_domains()
909 */
910 domain = amdgpu_bo_get_preferred_pin_domain(adev, domain);
911
912 if (bo->pin_count) {
913 uint32_t mem_type = bo->tbo.mem.mem_type;
914
915 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
916 return -EINVAL;
917
918 bo->pin_count++;
919
920 if (max_offset != 0) {
921 u64 domain_start = amdgpu_ttm_domain_start(adev,
922 mem_type);
923 WARN_ON_ONCE(max_offset <
924 (amdgpu_bo_gpu_offset(bo) - domain_start));
925 }
926
927 return 0;
928 }
929
930 if (bo->tbo.base.import_attach)
931 dma_buf_pin(bo->tbo.base.import_attach);
932
933 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
934 /* force to pin into visible video ram */
935 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
936 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
937 amdgpu_bo_placement_from_domain(bo, domain);
938 for (i = 0; i < bo->placement.num_placement; i++) {
939 unsigned fpfn, lpfn;
940
941 fpfn = min_offset >> PAGE_SHIFT;
942 lpfn = max_offset >> PAGE_SHIFT;
943
944 if (fpfn > bo->placements[i].fpfn)
945 bo->placements[i].fpfn = fpfn;
946 if (!bo->placements[i].lpfn ||
947 (lpfn && lpfn < bo->placements[i].lpfn))
948 bo->placements[i].lpfn = lpfn;
949 bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
950 }
951
952 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
953 if (unlikely(r)) {
954 dev_err(adev->dev, "%p pin failed\n", bo);
955 goto error;
956 }
957
958 bo->pin_count = 1;
959
960 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
961 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
962 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
963 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
964 &adev->visible_pin_size);
965 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
966 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
967 }
968
969error:
970 return r;
971}
972
973/**
974 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
975 * @bo: &amdgpu_bo buffer object to be pinned
976 * @domain: domain to be pinned to
977 *
978 * A simple wrapper to amdgpu_bo_pin_restricted().
979 * Provides a simpler API for buffers that do not have any strict restrictions
980 * on where a buffer must be located.
981 *
982 * Returns:
983 * 0 for success or a negative error code on failure.
984 */
985int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
986{
987 return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
988}
989
990/**
991 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
992 * @bo: &amdgpu_bo buffer object to be unpinned
993 *
994 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
995 * Changes placement and pin size accordingly.
996 *
997 * Returns:
998 * 0 for success or a negative error code on failure.
999 */
1000int amdgpu_bo_unpin(struct amdgpu_bo *bo)
1001{
1002 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1003 struct ttm_operation_ctx ctx = { false, false };
1004 int r, i;
1005
1006 if (WARN_ON_ONCE(!bo->pin_count)) {
1007 dev_warn(adev->dev, "%p unpin not necessary\n", bo);
1008 return 0;
1009 }
1010 bo->pin_count--;
1011 if (bo->pin_count)
1012 return 0;
1013
1014 amdgpu_bo_subtract_pin_size(bo);
1015
1016 if (bo->tbo.base.import_attach)
1017 dma_buf_unpin(bo->tbo.base.import_attach);
1018
1019 for (i = 0; i < bo->placement.num_placement; i++) {
1020 bo->placements[i].lpfn = 0;
1021 bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
1022 }
1023 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1024 if (unlikely(r))
1025 dev_err(adev->dev, "%p validate failed for unpin\n", bo);
1026
1027 return r;
1028}
1029
1030/**
1031 * amdgpu_bo_evict_vram - evict VRAM buffers
1032 * @adev: amdgpu device object
1033 *
1034 * Evicts all VRAM buffers on the lru list of the memory type.
1035 * Mainly used for evicting vram at suspend time.
1036 *
1037 * Returns:
1038 * 0 for success or a negative error code on failure.
1039 */
1040int amdgpu_bo_evict_vram(struct amdgpu_device *adev)
1041{
1042 /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
1043#ifndef CONFIG_HIBERNATION
1044 if (adev->flags & AMD_IS_APU) {
1045 /* Useless to evict on IGP chips */
1046 return 0;
1047 }
1048#endif
1049 return ttm_bo_evict_mm(&adev->mman.bdev, TTM_PL_VRAM);
1050}
1051
1052static const char *amdgpu_vram_names[] = {
1053 "UNKNOWN",
1054 "GDDR1",
1055 "DDR2",
1056 "GDDR3",
1057 "GDDR4",
1058 "GDDR5",
1059 "HBM",
1060 "DDR3",
1061 "DDR4",
1062 "GDDR6",
1063};
1064
1065/**
1066 * amdgpu_bo_init - initialize memory manager
1067 * @adev: amdgpu device object
1068 *
1069 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
1070 *
1071 * Returns:
1072 * 0 for success or a negative error code on failure.
1073 */
1074int amdgpu_bo_init(struct amdgpu_device *adev)
1075{
1076 /* reserve PAT memory space to WC for VRAM */
1077 arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1078 adev->gmc.aper_size);
1079
1080 /* Add an MTRR for the VRAM */
1081 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1082 adev->gmc.aper_size);
1083 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1084 adev->gmc.mc_vram_size >> 20,
1085 (unsigned long long)adev->gmc.aper_size >> 20);
1086 DRM_INFO("RAM width %dbits %s\n",
1087 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1088 return amdgpu_ttm_init(adev);
1089}
1090
1091/**
1092 * amdgpu_bo_late_init - late init
1093 * @adev: amdgpu device object
1094 *
1095 * Calls amdgpu_ttm_late_init() to free resources used earlier during
1096 * initialization.
1097 *
1098 * Returns:
1099 * 0 for success or a negative error code on failure.
1100 */
1101int amdgpu_bo_late_init(struct amdgpu_device *adev)
1102{
1103 amdgpu_ttm_late_init(adev);
1104
1105 return 0;
1106}
1107
1108/**
1109 * amdgpu_bo_fini - tear down memory manager
1110 * @adev: amdgpu device object
1111 *
1112 * Reverses amdgpu_bo_init() to tear down memory manager.
1113 */
1114void amdgpu_bo_fini(struct amdgpu_device *adev)
1115{
1116 amdgpu_ttm_fini(adev);
1117 arch_phys_wc_del(adev->gmc.vram_mtrr);
1118 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1119}
1120
1121/**
1122 * amdgpu_bo_fbdev_mmap - mmap fbdev memory
1123 * @bo: &amdgpu_bo buffer object
1124 * @vma: vma as input from the fbdev mmap method
1125 *
1126 * Calls ttm_fbdev_mmap() to mmap fbdev memory if it is backed by a bo.
1127 *
1128 * Returns:
1129 * 0 for success or a negative error code on failure.
1130 */
1131int amdgpu_bo_fbdev_mmap(struct amdgpu_bo *bo,
1132 struct vm_area_struct *vma)
1133{
1134 if (vma->vm_pgoff != 0)
1135 return -EACCES;
1136
1137 return ttm_bo_mmap_obj(vma, &bo->tbo);
1138}
1139
1140/**
1141 * amdgpu_bo_set_tiling_flags - set tiling flags
1142 * @bo: &amdgpu_bo buffer object
1143 * @tiling_flags: new flags
1144 *
1145 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1146 * kernel driver to set the tiling flags on a buffer.
1147 *
1148 * Returns:
1149 * 0 for success or a negative error code on failure.
1150 */
1151int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1152{
1153 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1154
1155 if (adev->family <= AMDGPU_FAMILY_CZ &&
1156 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1157 return -EINVAL;
1158
1159 bo->tiling_flags = tiling_flags;
1160 return 0;
1161}
1162
1163/**
1164 * amdgpu_bo_get_tiling_flags - get tiling flags
1165 * @bo: &amdgpu_bo buffer object
1166 * @tiling_flags: returned flags
1167 *
1168 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1169 * set the tiling flags on a buffer.
1170 */
1171void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1172{
1173 dma_resv_assert_held(bo->tbo.base.resv);
1174
1175 if (tiling_flags)
1176 *tiling_flags = bo->tiling_flags;
1177}
1178
1179/**
1180 * amdgpu_bo_set_metadata - set metadata
1181 * @bo: &amdgpu_bo buffer object
1182 * @metadata: new metadata
1183 * @metadata_size: size of the new metadata
1184 * @flags: flags of the new metadata
1185 *
1186 * Sets buffer object's metadata, its size and flags.
1187 * Used via GEM ioctl.
1188 *
1189 * Returns:
1190 * 0 for success or a negative error code on failure.
1191 */
1192int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
1193 uint32_t metadata_size, uint64_t flags)
1194{
1195 void *buffer;
1196
1197 if (!metadata_size) {
1198 if (bo->metadata_size) {
1199 kfree(bo->metadata);
1200 bo->metadata = NULL;
1201 bo->metadata_size = 0;
1202 }
1203 return 0;
1204 }
1205
1206 if (metadata == NULL)
1207 return -EINVAL;
1208
1209 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1210 if (buffer == NULL)
1211 return -ENOMEM;
1212
1213 kfree(bo->metadata);
1214 bo->metadata_flags = flags;
1215 bo->metadata = buffer;
1216 bo->metadata_size = metadata_size;
1217
1218 return 0;
1219}
1220
1221/**
1222 * amdgpu_bo_get_metadata - get metadata
1223 * @bo: &amdgpu_bo buffer object
1224 * @buffer: returned metadata
1225 * @buffer_size: size of the buffer
1226 * @metadata_size: size of the returned metadata
1227 * @flags: flags of the returned metadata
1228 *
1229 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1230 * less than metadata_size.
1231 * Used via GEM ioctl.
1232 *
1233 * Returns:
1234 * 0 for success or a negative error code on failure.
1235 */
1236int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1237 size_t buffer_size, uint32_t *metadata_size,
1238 uint64_t *flags)
1239{
1240 if (!buffer && !metadata_size)
1241 return -EINVAL;
1242
1243 if (buffer) {
1244 if (buffer_size < bo->metadata_size)
1245 return -EINVAL;
1246
1247 if (bo->metadata_size)
1248 memcpy(buffer, bo->metadata, bo->metadata_size);
1249 }
1250
1251 if (metadata_size)
1252 *metadata_size = bo->metadata_size;
1253 if (flags)
1254 *flags = bo->metadata_flags;
1255
1256 return 0;
1257}
1258
1259/**
1260 * amdgpu_bo_move_notify - notification about a memory move
1261 * @bo: pointer to a buffer object
1262 * @evict: if this move is evicting the buffer from the graphics address space
1263 * @new_mem: new information of the bufer object
1264 *
1265 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1266 * bookkeeping.
1267 * TTM driver callback which is called when ttm moves a buffer.
1268 */
1269void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1270 bool evict,
1271 struct ttm_mem_reg *new_mem)
1272{
1273 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1274 struct amdgpu_bo *abo;
1275 struct ttm_mem_reg *old_mem = &bo->mem;
1276
1277 if (!amdgpu_bo_is_amdgpu_bo(bo))
1278 return;
1279
1280 abo = ttm_to_amdgpu_bo(bo);
1281 amdgpu_vm_bo_invalidate(adev, abo, evict);
1282
1283 amdgpu_bo_kunmap(abo);
1284
1285 if (abo->tbo.base.dma_buf && !abo->tbo.base.import_attach &&
1286 bo->mem.mem_type != TTM_PL_SYSTEM)
1287 dma_buf_move_notify(abo->tbo.base.dma_buf);
1288
1289 /* remember the eviction */
1290 if (evict)
1291 atomic64_inc(&adev->num_evictions);
1292
1293 /* update statistics */
1294 if (!new_mem)
1295 return;
1296
1297 /* move_notify is called before move happens */
1298 trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
1299}
1300
1301/**
1302 * amdgpu_bo_move_notify - notification about a BO being released
1303 * @bo: pointer to a buffer object
1304 *
1305 * Wipes VRAM buffers whose contents should not be leaked before the
1306 * memory is released.
1307 */
1308void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)
1309{
1310 struct dma_fence *fence = NULL;
1311 struct amdgpu_bo *abo;
1312 int r;
1313
1314 if (!amdgpu_bo_is_amdgpu_bo(bo))
1315 return;
1316
1317 abo = ttm_to_amdgpu_bo(bo);
1318
1319 if (abo->kfd_bo)
1320 amdgpu_amdkfd_unreserve_memory_limit(abo);
1321
1322 /* We only remove the fence if the resv has individualized. */
1323 WARN_ON_ONCE(bo->type == ttm_bo_type_kernel
1324 && bo->base.resv != &bo->base._resv);
1325 if (bo->base.resv == &bo->base._resv)
1326 amdgpu_amdkfd_remove_fence_on_pt_pd_bos(abo);
1327
1328 if (bo->mem.mem_type != TTM_PL_VRAM || !bo->mem.mm_node ||
1329 !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE))
1330 return;
1331
1332 dma_resv_lock(bo->base.resv, NULL);
1333
1334 r = amdgpu_fill_buffer(abo, AMDGPU_POISON, bo->base.resv, &fence);
1335 if (!WARN_ON(r)) {
1336 amdgpu_bo_fence(abo, fence, false);
1337 dma_fence_put(fence);
1338 }
1339
1340 dma_resv_unlock(bo->base.resv);
1341}
1342
1343/**
1344 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1345 * @bo: pointer to a buffer object
1346 *
1347 * Notifies the driver we are taking a fault on this BO and have reserved it,
1348 * also performs bookkeeping.
1349 * TTM driver callback for dealing with vm faults.
1350 *
1351 * Returns:
1352 * 0 for success or a negative error code on failure.
1353 */
1354int amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1355{
1356 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1357 struct ttm_operation_ctx ctx = { false, false };
1358 struct amdgpu_bo *abo;
1359 unsigned long offset, size;
1360 int r;
1361
1362 if (!amdgpu_bo_is_amdgpu_bo(bo))
1363 return 0;
1364
1365 abo = ttm_to_amdgpu_bo(bo);
1366
1367 /* Remember that this BO was accessed by the CPU */
1368 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1369
1370 if (bo->mem.mem_type != TTM_PL_VRAM)
1371 return 0;
1372
1373 size = bo->mem.num_pages << PAGE_SHIFT;
1374 offset = bo->mem.start << PAGE_SHIFT;
1375 if ((offset + size) <= adev->gmc.visible_vram_size)
1376 return 0;
1377
1378 /* Can't move a pinned BO to visible VRAM */
1379 if (abo->pin_count > 0)
1380 return -EINVAL;
1381
1382 /* hurrah the memory is not visible ! */
1383 atomic64_inc(&adev->num_vram_cpu_page_faults);
1384 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1385 AMDGPU_GEM_DOMAIN_GTT);
1386
1387 /* Avoid costly evictions; only set GTT as a busy placement */
1388 abo->placement.num_busy_placement = 1;
1389 abo->placement.busy_placement = &abo->placements[1];
1390
1391 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1392 if (unlikely(r != 0))
1393 return r;
1394
1395 offset = bo->mem.start << PAGE_SHIFT;
1396 /* this should never happen */
1397 if (bo->mem.mem_type == TTM_PL_VRAM &&
1398 (offset + size) > adev->gmc.visible_vram_size)
1399 return -EINVAL;
1400
1401 return 0;
1402}
1403
1404/**
1405 * amdgpu_bo_fence - add fence to buffer object
1406 *
1407 * @bo: buffer object in question
1408 * @fence: fence to add
1409 * @shared: true if fence should be added shared
1410 *
1411 */
1412void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1413 bool shared)
1414{
1415 struct dma_resv *resv = bo->tbo.base.resv;
1416
1417 if (shared)
1418 dma_resv_add_shared_fence(resv, fence);
1419 else
1420 dma_resv_add_excl_fence(resv, fence);
1421}
1422
1423/**
1424 * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences
1425 *
1426 * @adev: amdgpu device pointer
1427 * @resv: reservation object to sync to
1428 * @sync_mode: synchronization mode
1429 * @owner: fence owner
1430 * @intr: Whether the wait is interruptible
1431 *
1432 * Extract the fences from the reservation object and waits for them to finish.
1433 *
1434 * Returns:
1435 * 0 on success, errno otherwise.
1436 */
1437int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv,
1438 enum amdgpu_sync_mode sync_mode, void *owner,
1439 bool intr)
1440{
1441 struct amdgpu_sync sync;
1442 int r;
1443
1444 amdgpu_sync_create(&sync);
1445 amdgpu_sync_resv(adev, &sync, resv, sync_mode, owner);
1446 r = amdgpu_sync_wait(&sync, intr);
1447 amdgpu_sync_free(&sync);
1448 return r;
1449}
1450
1451/**
1452 * amdgpu_bo_sync_wait - Wrapper for amdgpu_bo_sync_wait_resv
1453 * @bo: buffer object to wait for
1454 * @owner: fence owner
1455 * @intr: Whether the wait is interruptible
1456 *
1457 * Wrapper to wait for fences in a BO.
1458 * Returns:
1459 * 0 on success, errno otherwise.
1460 */
1461int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)
1462{
1463 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1464
1465 return amdgpu_bo_sync_wait_resv(adev, bo->tbo.base.resv,
1466 AMDGPU_SYNC_NE_OWNER, owner, intr);
1467}
1468
1469/**
1470 * amdgpu_bo_gpu_offset - return GPU offset of bo
1471 * @bo: amdgpu object for which we query the offset
1472 *
1473 * Note: object should either be pinned or reserved when calling this
1474 * function, it might be useful to add check for this for debugging.
1475 *
1476 * Returns:
1477 * current GPU offset of the object.
1478 */
1479u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1480{
1481 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_SYSTEM);
1482 WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) &&
1483 !bo->pin_count && bo->tbo.type != ttm_bo_type_kernel);
1484 WARN_ON_ONCE(bo->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET);
1485 WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_VRAM &&
1486 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1487
1488 return amdgpu_bo_gpu_offset_no_check(bo);
1489}
1490
1491/**
1492 * amdgpu_bo_gpu_offset_no_check - return GPU offset of bo
1493 * @bo: amdgpu object for which we query the offset
1494 *
1495 * Returns:
1496 * current GPU offset of the object without raising warnings.
1497 */
1498u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo)
1499{
1500 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1501 uint64_t offset;
1502
1503 offset = (bo->tbo.mem.start << PAGE_SHIFT) +
1504 amdgpu_ttm_domain_start(adev, bo->tbo.mem.mem_type);
1505
1506 return amdgpu_gmc_sign_extend(offset);
1507}
1508
1509/**
1510 * amdgpu_bo_get_preferred_pin_domain - get preferred domain for scanout
1511 * @adev: amdgpu device object
1512 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1513 *
1514 * Returns:
1515 * Which of the allowed domains is preferred for pinning the BO for scanout.
1516 */
1517uint32_t amdgpu_bo_get_preferred_pin_domain(struct amdgpu_device *adev,
1518 uint32_t domain)
1519{
1520 if (domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) {
1521 domain = AMDGPU_GEM_DOMAIN_VRAM;
1522 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1523 domain = AMDGPU_GEM_DOMAIN_GTT;
1524 }
1525 return domain;
1526}
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
55static void amdgpu_bo_destroy(struct ttm_buffer_object *tbo)
56{
57 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
58
59 amdgpu_bo_kunmap(bo);
60
61 if (bo->tbo.base.import_attach)
62 drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
63 drm_gem_object_release(&bo->tbo.base);
64 amdgpu_bo_unref(&bo->parent);
65 kvfree(bo);
66}
67
68static void amdgpu_bo_user_destroy(struct ttm_buffer_object *tbo)
69{
70 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
71 struct amdgpu_bo_user *ubo;
72
73 ubo = to_amdgpu_bo_user(bo);
74 kfree(ubo->metadata);
75 amdgpu_bo_destroy(tbo);
76}
77
78static void amdgpu_bo_vm_destroy(struct ttm_buffer_object *tbo)
79{
80 struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
81 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
82 struct amdgpu_bo_vm *vmbo;
83
84 vmbo = to_amdgpu_bo_vm(bo);
85 /* in case amdgpu_device_recover_vram got NULL of bo->parent */
86 if (!list_empty(&vmbo->shadow_list)) {
87 mutex_lock(&adev->shadow_list_lock);
88 list_del_init(&vmbo->shadow_list);
89 mutex_unlock(&adev->shadow_list_lock);
90 }
91
92 amdgpu_bo_destroy(tbo);
93}
94
95/**
96 * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
97 * @bo: buffer object to be checked
98 *
99 * Uses destroy function associated with the object to determine if this is
100 * an &amdgpu_bo.
101 *
102 * Returns:
103 * true if the object belongs to &amdgpu_bo, false if not.
104 */
105bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
106{
107 if (bo->destroy == &amdgpu_bo_destroy ||
108 bo->destroy == &amdgpu_bo_user_destroy ||
109 bo->destroy == &amdgpu_bo_vm_destroy)
110 return true;
111
112 return false;
113}
114
115/**
116 * amdgpu_bo_placement_from_domain - set buffer's placement
117 * @abo: &amdgpu_bo buffer object whose placement is to be set
118 * @domain: requested domain
119 *
120 * Sets buffer's placement according to requested domain and the buffer's
121 * flags.
122 */
123void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
124{
125 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
126 struct ttm_placement *placement = &abo->placement;
127 struct ttm_place *places = abo->placements;
128 u64 flags = abo->flags;
129 u32 c = 0;
130
131 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
132 unsigned visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
133
134 places[c].fpfn = 0;
135 places[c].lpfn = 0;
136 places[c].mem_type = TTM_PL_VRAM;
137 places[c].flags = 0;
138
139 if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
140 places[c].lpfn = visible_pfn;
141 else
142 places[c].flags |= TTM_PL_FLAG_TOPDOWN;
143
144 if (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
145 places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
146 c++;
147 }
148
149 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
150 places[c].fpfn = 0;
151 places[c].lpfn = 0;
152 places[c].mem_type =
153 abo->flags & AMDGPU_GEM_CREATE_PREEMPTIBLE ?
154 AMDGPU_PL_PREEMPT : TTM_PL_TT;
155 places[c].flags = 0;
156 c++;
157 }
158
159 if (domain & AMDGPU_GEM_DOMAIN_CPU) {
160 places[c].fpfn = 0;
161 places[c].lpfn = 0;
162 places[c].mem_type = TTM_PL_SYSTEM;
163 places[c].flags = 0;
164 c++;
165 }
166
167 if (domain & AMDGPU_GEM_DOMAIN_GDS) {
168 places[c].fpfn = 0;
169 places[c].lpfn = 0;
170 places[c].mem_type = AMDGPU_PL_GDS;
171 places[c].flags = 0;
172 c++;
173 }
174
175 if (domain & AMDGPU_GEM_DOMAIN_GWS) {
176 places[c].fpfn = 0;
177 places[c].lpfn = 0;
178 places[c].mem_type = AMDGPU_PL_GWS;
179 places[c].flags = 0;
180 c++;
181 }
182
183 if (domain & AMDGPU_GEM_DOMAIN_OA) {
184 places[c].fpfn = 0;
185 places[c].lpfn = 0;
186 places[c].mem_type = AMDGPU_PL_OA;
187 places[c].flags = 0;
188 c++;
189 }
190
191 if (!c) {
192 places[c].fpfn = 0;
193 places[c].lpfn = 0;
194 places[c].mem_type = TTM_PL_SYSTEM;
195 places[c].flags = 0;
196 c++;
197 }
198
199 BUG_ON(c > AMDGPU_BO_MAX_PLACEMENTS);
200
201 placement->num_placement = c;
202 placement->placement = places;
203
204 placement->num_busy_placement = c;
205 placement->busy_placement = places;
206}
207
208/**
209 * amdgpu_bo_create_reserved - create reserved BO for kernel use
210 *
211 * @adev: amdgpu device object
212 * @size: size for the new BO
213 * @align: alignment for the new BO
214 * @domain: where to place it
215 * @bo_ptr: used to initialize BOs in structures
216 * @gpu_addr: GPU addr of the pinned BO
217 * @cpu_addr: optional CPU address mapping
218 *
219 * Allocates and pins a BO for kernel internal use, and returns it still
220 * reserved.
221 *
222 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
223 *
224 * Returns:
225 * 0 on success, negative error code otherwise.
226 */
227int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
228 unsigned long size, int align,
229 u32 domain, struct amdgpu_bo **bo_ptr,
230 u64 *gpu_addr, void **cpu_addr)
231{
232 struct amdgpu_bo_param bp;
233 bool free = false;
234 int r;
235
236 if (!size) {
237 amdgpu_bo_unref(bo_ptr);
238 return 0;
239 }
240
241 memset(&bp, 0, sizeof(bp));
242 bp.size = size;
243 bp.byte_align = align;
244 bp.domain = domain;
245 bp.flags = cpu_addr ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
246 : AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
247 bp.flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
248 bp.type = ttm_bo_type_kernel;
249 bp.resv = NULL;
250 bp.bo_ptr_size = sizeof(struct amdgpu_bo);
251
252 if (!*bo_ptr) {
253 r = amdgpu_bo_create(adev, &bp, bo_ptr);
254 if (r) {
255 dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
256 r);
257 return r;
258 }
259 free = true;
260 }
261
262 r = amdgpu_bo_reserve(*bo_ptr, false);
263 if (r) {
264 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
265 goto error_free;
266 }
267
268 r = amdgpu_bo_pin(*bo_ptr, domain);
269 if (r) {
270 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
271 goto error_unreserve;
272 }
273
274 r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
275 if (r) {
276 dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
277 goto error_unpin;
278 }
279
280 if (gpu_addr)
281 *gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
282
283 if (cpu_addr) {
284 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
285 if (r) {
286 dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
287 goto error_unpin;
288 }
289 }
290
291 return 0;
292
293error_unpin:
294 amdgpu_bo_unpin(*bo_ptr);
295error_unreserve:
296 amdgpu_bo_unreserve(*bo_ptr);
297
298error_free:
299 if (free)
300 amdgpu_bo_unref(bo_ptr);
301
302 return r;
303}
304
305/**
306 * amdgpu_bo_create_kernel - create BO for kernel use
307 *
308 * @adev: amdgpu device object
309 * @size: size for the new BO
310 * @align: alignment for the new BO
311 * @domain: where to place it
312 * @bo_ptr: used to initialize BOs in structures
313 * @gpu_addr: GPU addr of the pinned BO
314 * @cpu_addr: optional CPU address mapping
315 *
316 * Allocates and pins a BO for kernel internal use.
317 *
318 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
319 *
320 * Returns:
321 * 0 on success, negative error code otherwise.
322 */
323int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
324 unsigned long size, int align,
325 u32 domain, struct amdgpu_bo **bo_ptr,
326 u64 *gpu_addr, void **cpu_addr)
327{
328 int r;
329
330 r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
331 gpu_addr, cpu_addr);
332
333 if (r)
334 return r;
335
336 if (*bo_ptr)
337 amdgpu_bo_unreserve(*bo_ptr);
338
339 return 0;
340}
341
342/**
343 * amdgpu_bo_create_kernel_at - create BO for kernel use at specific location
344 *
345 * @adev: amdgpu device object
346 * @offset: offset of the BO
347 * @size: size of the BO
348 * @domain: where to place it
349 * @bo_ptr: used to initialize BOs in structures
350 * @cpu_addr: optional CPU address mapping
351 *
352 * Creates a kernel BO at a specific offset in the address space of the domain.
353 *
354 * Returns:
355 * 0 on success, negative error code otherwise.
356 */
357int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev,
358 uint64_t offset, uint64_t size, uint32_t domain,
359 struct amdgpu_bo **bo_ptr, void **cpu_addr)
360{
361 struct ttm_operation_ctx ctx = { false, false };
362 unsigned int i;
363 int r;
364
365 offset &= PAGE_MASK;
366 size = ALIGN(size, PAGE_SIZE);
367
368 r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE, domain, bo_ptr,
369 NULL, cpu_addr);
370 if (r)
371 return r;
372
373 if ((*bo_ptr) == NULL)
374 return 0;
375
376 /*
377 * Remove the original mem node and create a new one at the request
378 * position.
379 */
380 if (cpu_addr)
381 amdgpu_bo_kunmap(*bo_ptr);
382
383 ttm_resource_free(&(*bo_ptr)->tbo, &(*bo_ptr)->tbo.resource);
384
385 for (i = 0; i < (*bo_ptr)->placement.num_placement; ++i) {
386 (*bo_ptr)->placements[i].fpfn = offset >> PAGE_SHIFT;
387 (*bo_ptr)->placements[i].lpfn = (offset + size) >> PAGE_SHIFT;
388 }
389 r = ttm_bo_mem_space(&(*bo_ptr)->tbo, &(*bo_ptr)->placement,
390 &(*bo_ptr)->tbo.resource, &ctx);
391 if (r)
392 goto error;
393
394 if (cpu_addr) {
395 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
396 if (r)
397 goto error;
398 }
399
400 amdgpu_bo_unreserve(*bo_ptr);
401 return 0;
402
403error:
404 amdgpu_bo_unreserve(*bo_ptr);
405 amdgpu_bo_unref(bo_ptr);
406 return r;
407}
408
409/**
410 * amdgpu_bo_free_kernel - free BO for kernel use
411 *
412 * @bo: amdgpu BO to free
413 * @gpu_addr: pointer to where the BO's GPU memory space address was stored
414 * @cpu_addr: pointer to where the BO's CPU memory space address was stored
415 *
416 * unmaps and unpin a BO for kernel internal use.
417 */
418void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
419 void **cpu_addr)
420{
421 if (*bo == NULL)
422 return;
423
424 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
425 if (cpu_addr)
426 amdgpu_bo_kunmap(*bo);
427
428 amdgpu_bo_unpin(*bo);
429 amdgpu_bo_unreserve(*bo);
430 }
431 amdgpu_bo_unref(bo);
432
433 if (gpu_addr)
434 *gpu_addr = 0;
435
436 if (cpu_addr)
437 *cpu_addr = NULL;
438}
439
440/* Validate bo size is bit bigger then the request domain */
441static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
442 unsigned long size, u32 domain)
443{
444 struct ttm_resource_manager *man = NULL;
445
446 /*
447 * If GTT is part of requested domains the check must succeed to
448 * allow fall back to GTT
449 */
450 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
451 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_TT);
452
453 if (size < (man->size << PAGE_SHIFT))
454 return true;
455 else
456 goto fail;
457 }
458
459 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
460 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
461
462 if (size < (man->size << PAGE_SHIFT))
463 return true;
464 else
465 goto fail;
466 }
467
468
469 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */
470 return true;
471
472fail:
473 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size,
474 man->size << PAGE_SHIFT);
475 return false;
476}
477
478bool amdgpu_bo_support_uswc(u64 bo_flags)
479{
480
481#ifdef CONFIG_X86_32
482 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
483 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
484 */
485 return false;
486#elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
487 /* Don't try to enable write-combining when it can't work, or things
488 * may be slow
489 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
490 */
491
492#ifndef CONFIG_COMPILE_TEST
493#warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
494 thanks to write-combining
495#endif
496
497 if (bo_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
498 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
499 "better performance thanks to write-combining\n");
500 return false;
501#else
502 /* For architectures that don't support WC memory,
503 * mask out the WC flag from the BO
504 */
505 if (!drm_arch_can_wc_memory())
506 return false;
507
508 return true;
509#endif
510}
511
512/**
513 * amdgpu_bo_create - create an &amdgpu_bo buffer object
514 * @adev: amdgpu device object
515 * @bp: parameters to be used for the buffer object
516 * @bo_ptr: pointer to the buffer object pointer
517 *
518 * Creates an &amdgpu_bo buffer object.
519 *
520 * Returns:
521 * 0 for success or a negative error code on failure.
522 */
523int amdgpu_bo_create(struct amdgpu_device *adev,
524 struct amdgpu_bo_param *bp,
525 struct amdgpu_bo **bo_ptr)
526{
527 struct ttm_operation_ctx ctx = {
528 .interruptible = (bp->type != ttm_bo_type_kernel),
529 .no_wait_gpu = bp->no_wait_gpu,
530 /* We opt to avoid OOM on system pages allocations */
531 .gfp_retry_mayfail = true,
532 .allow_res_evict = bp->type != ttm_bo_type_kernel,
533 .resv = bp->resv
534 };
535 struct amdgpu_bo *bo;
536 unsigned long page_align, size = bp->size;
537 int r;
538
539 /* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
540 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
541 /* GWS and OA don't need any alignment. */
542 page_align = bp->byte_align;
543 size <<= PAGE_SHIFT;
544 } else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
545 /* Both size and alignment must be a multiple of 4. */
546 page_align = ALIGN(bp->byte_align, 4);
547 size = ALIGN(size, 4) << PAGE_SHIFT;
548 } else {
549 /* Memory should be aligned at least to a page size. */
550 page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
551 size = ALIGN(size, PAGE_SIZE);
552 }
553
554 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
555 return -ENOMEM;
556
557 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo));
558
559 *bo_ptr = NULL;
560 bo = kvzalloc(bp->bo_ptr_size, GFP_KERNEL);
561 if (bo == NULL)
562 return -ENOMEM;
563 drm_gem_private_object_init(adev_to_drm(adev), &bo->tbo.base, size);
564 bo->vm_bo = NULL;
565 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
566 bp->domain;
567 bo->allowed_domains = bo->preferred_domains;
568 if (bp->type != ttm_bo_type_kernel &&
569 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
570 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
571
572 bo->flags = bp->flags;
573
574 if (!amdgpu_bo_support_uswc(bo->flags))
575 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
576
577 bo->tbo.bdev = &adev->mman.bdev;
578 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
579 AMDGPU_GEM_DOMAIN_GDS))
580 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
581 else
582 amdgpu_bo_placement_from_domain(bo, bp->domain);
583 if (bp->type == ttm_bo_type_kernel)
584 bo->tbo.priority = 1;
585
586 if (!bp->destroy)
587 bp->destroy = &amdgpu_bo_destroy;
588
589 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, bp->type,
590 &bo->placement, page_align, &ctx, NULL,
591 bp->resv, bp->destroy);
592 if (unlikely(r != 0))
593 return r;
594
595 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
596 bo->tbo.resource->mem_type == TTM_PL_VRAM &&
597 bo->tbo.resource->start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
598 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
599 ctx.bytes_moved);
600 else
601 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
602
603 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
604 bo->tbo.resource->mem_type == TTM_PL_VRAM) {
605 struct dma_fence *fence;
606
607 r = amdgpu_fill_buffer(bo, 0, bo->tbo.base.resv, &fence);
608 if (unlikely(r))
609 goto fail_unreserve;
610
611 amdgpu_bo_fence(bo, fence, false);
612 dma_fence_put(bo->tbo.moving);
613 bo->tbo.moving = dma_fence_get(fence);
614 dma_fence_put(fence);
615 }
616 if (!bp->resv)
617 amdgpu_bo_unreserve(bo);
618 *bo_ptr = bo;
619
620 trace_amdgpu_bo_create(bo);
621
622 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
623 if (bp->type == ttm_bo_type_device)
624 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
625
626 return 0;
627
628fail_unreserve:
629 if (!bp->resv)
630 dma_resv_unlock(bo->tbo.base.resv);
631 amdgpu_bo_unref(&bo);
632 return r;
633}
634
635/**
636 * amdgpu_bo_create_user - create an &amdgpu_bo_user buffer object
637 * @adev: amdgpu device object
638 * @bp: parameters to be used for the buffer object
639 * @ubo_ptr: pointer to the buffer object pointer
640 *
641 * Create a BO to be used by user application;
642 *
643 * Returns:
644 * 0 for success or a negative error code on failure.
645 */
646
647int amdgpu_bo_create_user(struct amdgpu_device *adev,
648 struct amdgpu_bo_param *bp,
649 struct amdgpu_bo_user **ubo_ptr)
650{
651 struct amdgpu_bo *bo_ptr;
652 int r;
653
654 bp->bo_ptr_size = sizeof(struct amdgpu_bo_user);
655 bp->destroy = &amdgpu_bo_user_destroy;
656 r = amdgpu_bo_create(adev, bp, &bo_ptr);
657 if (r)
658 return r;
659
660 *ubo_ptr = to_amdgpu_bo_user(bo_ptr);
661 return r;
662}
663
664/**
665 * amdgpu_bo_create_vm - create an &amdgpu_bo_vm buffer object
666 * @adev: amdgpu device object
667 * @bp: parameters to be used for the buffer object
668 * @vmbo_ptr: pointer to the buffer object pointer
669 *
670 * Create a BO to be for GPUVM.
671 *
672 * Returns:
673 * 0 for success or a negative error code on failure.
674 */
675
676int amdgpu_bo_create_vm(struct amdgpu_device *adev,
677 struct amdgpu_bo_param *bp,
678 struct amdgpu_bo_vm **vmbo_ptr)
679{
680 struct amdgpu_bo *bo_ptr;
681 int r;
682
683 /* bo_ptr_size will be determined by the caller and it depends on
684 * num of amdgpu_vm_pt entries.
685 */
686 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo_vm));
687 bp->destroy = &amdgpu_bo_vm_destroy;
688 r = amdgpu_bo_create(adev, bp, &bo_ptr);
689 if (r)
690 return r;
691
692 *vmbo_ptr = to_amdgpu_bo_vm(bo_ptr);
693 INIT_LIST_HEAD(&(*vmbo_ptr)->shadow_list);
694 return r;
695}
696
697/**
698 * amdgpu_bo_validate - validate an &amdgpu_bo buffer object
699 * @bo: pointer to the buffer object
700 *
701 * Sets placement according to domain; and changes placement and caching
702 * policy of the buffer object according to the placement.
703 * This is used for validating shadow bos. It calls ttm_bo_validate() to
704 * make sure the buffer is resident where it needs to be.
705 *
706 * Returns:
707 * 0 for success or a negative error code on failure.
708 */
709int amdgpu_bo_validate(struct amdgpu_bo *bo)
710{
711 struct ttm_operation_ctx ctx = { false, false };
712 uint32_t domain;
713 int r;
714
715 if (bo->tbo.pin_count)
716 return 0;
717
718 domain = bo->preferred_domains;
719
720retry:
721 amdgpu_bo_placement_from_domain(bo, domain);
722 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
723 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
724 domain = bo->allowed_domains;
725 goto retry;
726 }
727
728 return r;
729}
730
731/**
732 * amdgpu_bo_add_to_shadow_list - add a BO to the shadow list
733 *
734 * @bo: BO that will be inserted into the shadow list
735 *
736 * Insert a BO to the shadow list.
737 */
738void amdgpu_bo_add_to_shadow_list(struct amdgpu_bo_vm *vmbo)
739{
740 struct amdgpu_device *adev = amdgpu_ttm_adev(vmbo->bo.tbo.bdev);
741
742 mutex_lock(&adev->shadow_list_lock);
743 list_add_tail(&vmbo->shadow_list, &adev->shadow_list);
744 mutex_unlock(&adev->shadow_list_lock);
745}
746
747/**
748 * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow
749 *
750 * @shadow: &amdgpu_bo shadow to be restored
751 * @fence: dma_fence associated with the operation
752 *
753 * Copies a buffer object's shadow content back to the object.
754 * This is used for recovering a buffer from its shadow in case of a gpu
755 * reset where vram context may be lost.
756 *
757 * Returns:
758 * 0 for success or a negative error code on failure.
759 */
760int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence)
761
762{
763 struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev);
764 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
765 uint64_t shadow_addr, parent_addr;
766
767 shadow_addr = amdgpu_bo_gpu_offset(shadow);
768 parent_addr = amdgpu_bo_gpu_offset(shadow->parent);
769
770 return amdgpu_copy_buffer(ring, shadow_addr, parent_addr,
771 amdgpu_bo_size(shadow), NULL, fence,
772 true, false, false);
773}
774
775/**
776 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
777 * @bo: &amdgpu_bo buffer object to be mapped
778 * @ptr: kernel virtual address to be returned
779 *
780 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
781 * amdgpu_bo_kptr() to get the kernel virtual address.
782 *
783 * Returns:
784 * 0 for success or a negative error code on failure.
785 */
786int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
787{
788 void *kptr;
789 long r;
790
791 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
792 return -EPERM;
793
794 kptr = amdgpu_bo_kptr(bo);
795 if (kptr) {
796 if (ptr)
797 *ptr = kptr;
798 return 0;
799 }
800
801 r = dma_resv_wait_timeout(bo->tbo.base.resv, false, false,
802 MAX_SCHEDULE_TIMEOUT);
803 if (r < 0)
804 return r;
805
806 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.resource->num_pages, &bo->kmap);
807 if (r)
808 return r;
809
810 if (ptr)
811 *ptr = amdgpu_bo_kptr(bo);
812
813 return 0;
814}
815
816/**
817 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
818 * @bo: &amdgpu_bo buffer object
819 *
820 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
821 *
822 * Returns:
823 * the virtual address of a buffer object area.
824 */
825void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
826{
827 bool is_iomem;
828
829 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
830}
831
832/**
833 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
834 * @bo: &amdgpu_bo buffer object to be unmapped
835 *
836 * Unmaps a kernel map set up by amdgpu_bo_kmap().
837 */
838void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
839{
840 if (bo->kmap.bo)
841 ttm_bo_kunmap(&bo->kmap);
842}
843
844/**
845 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
846 * @bo: &amdgpu_bo buffer object
847 *
848 * References the contained &ttm_buffer_object.
849 *
850 * Returns:
851 * a refcounted pointer to the &amdgpu_bo buffer object.
852 */
853struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
854{
855 if (bo == NULL)
856 return NULL;
857
858 ttm_bo_get(&bo->tbo);
859 return bo;
860}
861
862/**
863 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
864 * @bo: &amdgpu_bo buffer object
865 *
866 * Unreferences the contained &ttm_buffer_object and clear the pointer
867 */
868void amdgpu_bo_unref(struct amdgpu_bo **bo)
869{
870 struct ttm_buffer_object *tbo;
871
872 if ((*bo) == NULL)
873 return;
874
875 tbo = &((*bo)->tbo);
876 ttm_bo_put(tbo);
877 *bo = NULL;
878}
879
880/**
881 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
882 * @bo: &amdgpu_bo buffer object to be pinned
883 * @domain: domain to be pinned to
884 * @min_offset: the start of requested address range
885 * @max_offset: the end of requested address range
886 *
887 * Pins the buffer object according to requested domain and address range. If
888 * the memory is unbound gart memory, binds the pages into gart table. Adjusts
889 * pin_count and pin_size accordingly.
890 *
891 * Pinning means to lock pages in memory along with keeping them at a fixed
892 * offset. It is required when a buffer can not be moved, for example, when
893 * a display buffer is being scanned out.
894 *
895 * Compared with amdgpu_bo_pin(), this function gives more flexibility on
896 * where to pin a buffer if there are specific restrictions on where a buffer
897 * must be located.
898 *
899 * Returns:
900 * 0 for success or a negative error code on failure.
901 */
902int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
903 u64 min_offset, u64 max_offset)
904{
905 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
906 struct ttm_operation_ctx ctx = { false, false };
907 int r, i;
908
909 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
910 return -EPERM;
911
912 if (WARN_ON_ONCE(min_offset > max_offset))
913 return -EINVAL;
914
915 /* A shared bo cannot be migrated to VRAM */
916 if (bo->prime_shared_count || bo->tbo.base.import_attach) {
917 if (domain & AMDGPU_GEM_DOMAIN_GTT)
918 domain = AMDGPU_GEM_DOMAIN_GTT;
919 else
920 return -EINVAL;
921 }
922
923 if (bo->tbo.pin_count) {
924 uint32_t mem_type = bo->tbo.resource->mem_type;
925 uint32_t mem_flags = bo->tbo.resource->placement;
926
927 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
928 return -EINVAL;
929
930 if ((mem_type == TTM_PL_VRAM) &&
931 (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) &&
932 !(mem_flags & TTM_PL_FLAG_CONTIGUOUS))
933 return -EINVAL;
934
935 ttm_bo_pin(&bo->tbo);
936
937 if (max_offset != 0) {
938 u64 domain_start = amdgpu_ttm_domain_start(adev,
939 mem_type);
940 WARN_ON_ONCE(max_offset <
941 (amdgpu_bo_gpu_offset(bo) - domain_start));
942 }
943
944 return 0;
945 }
946
947 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
948 * See function amdgpu_display_supported_domains()
949 */
950 domain = amdgpu_bo_get_preferred_pin_domain(adev, domain);
951
952 if (bo->tbo.base.import_attach)
953 dma_buf_pin(bo->tbo.base.import_attach);
954
955 /* force to pin into visible video ram */
956 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
957 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
958 amdgpu_bo_placement_from_domain(bo, domain);
959 for (i = 0; i < bo->placement.num_placement; i++) {
960 unsigned fpfn, lpfn;
961
962 fpfn = min_offset >> PAGE_SHIFT;
963 lpfn = max_offset >> PAGE_SHIFT;
964
965 if (fpfn > bo->placements[i].fpfn)
966 bo->placements[i].fpfn = fpfn;
967 if (!bo->placements[i].lpfn ||
968 (lpfn && lpfn < bo->placements[i].lpfn))
969 bo->placements[i].lpfn = lpfn;
970 }
971
972 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
973 if (unlikely(r)) {
974 dev_err(adev->dev, "%p pin failed\n", bo);
975 goto error;
976 }
977
978 ttm_bo_pin(&bo->tbo);
979
980 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type);
981 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
982 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
983 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
984 &adev->visible_pin_size);
985 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
986 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
987 }
988
989error:
990 return r;
991}
992
993/**
994 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
995 * @bo: &amdgpu_bo buffer object to be pinned
996 * @domain: domain to be pinned to
997 *
998 * A simple wrapper to amdgpu_bo_pin_restricted().
999 * Provides a simpler API for buffers that do not have any strict restrictions
1000 * on where a buffer must be located.
1001 *
1002 * Returns:
1003 * 0 for success or a negative error code on failure.
1004 */
1005int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
1006{
1007 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1008 return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
1009}
1010
1011/**
1012 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
1013 * @bo: &amdgpu_bo buffer object to be unpinned
1014 *
1015 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
1016 * Changes placement and pin size accordingly.
1017 *
1018 * Returns:
1019 * 0 for success or a negative error code on failure.
1020 */
1021void amdgpu_bo_unpin(struct amdgpu_bo *bo)
1022{
1023 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1024
1025 ttm_bo_unpin(&bo->tbo);
1026 if (bo->tbo.pin_count)
1027 return;
1028
1029 if (bo->tbo.base.import_attach)
1030 dma_buf_unpin(bo->tbo.base.import_attach);
1031
1032 if (bo->tbo.resource->mem_type == TTM_PL_VRAM) {
1033 atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
1034 atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
1035 &adev->visible_pin_size);
1036 } else if (bo->tbo.resource->mem_type == TTM_PL_TT) {
1037 atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size);
1038 }
1039}
1040
1041/**
1042 * amdgpu_bo_evict_vram - evict VRAM buffers
1043 * @adev: amdgpu device object
1044 *
1045 * Evicts all VRAM buffers on the lru list of the memory type.
1046 * Mainly used for evicting vram at suspend time.
1047 *
1048 * Returns:
1049 * 0 for success or a negative error code on failure.
1050 */
1051int amdgpu_bo_evict_vram(struct amdgpu_device *adev)
1052{
1053 struct ttm_resource_manager *man;
1054
1055 if (adev->in_s3 && (adev->flags & AMD_IS_APU)) {
1056 /* No need to evict vram on APUs for suspend to ram */
1057 return 0;
1058 }
1059
1060 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
1061 return ttm_resource_manager_evict_all(&adev->mman.bdev, man);
1062}
1063
1064static const char *amdgpu_vram_names[] = {
1065 "UNKNOWN",
1066 "GDDR1",
1067 "DDR2",
1068 "GDDR3",
1069 "GDDR4",
1070 "GDDR5",
1071 "HBM",
1072 "DDR3",
1073 "DDR4",
1074 "GDDR6",
1075 "DDR5"
1076};
1077
1078/**
1079 * amdgpu_bo_init - initialize memory manager
1080 * @adev: amdgpu device object
1081 *
1082 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
1083 *
1084 * Returns:
1085 * 0 for success or a negative error code on failure.
1086 */
1087int amdgpu_bo_init(struct amdgpu_device *adev)
1088{
1089 /* On A+A platform, VRAM can be mapped as WB */
1090 if (!adev->gmc.xgmi.connected_to_cpu) {
1091 /* reserve PAT memory space to WC for VRAM */
1092 arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1093 adev->gmc.aper_size);
1094
1095 /* Add an MTRR for the VRAM */
1096 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1097 adev->gmc.aper_size);
1098 }
1099
1100 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1101 adev->gmc.mc_vram_size >> 20,
1102 (unsigned long long)adev->gmc.aper_size >> 20);
1103 DRM_INFO("RAM width %dbits %s\n",
1104 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1105 return amdgpu_ttm_init(adev);
1106}
1107
1108/**
1109 * amdgpu_bo_fini - tear down memory manager
1110 * @adev: amdgpu device object
1111 *
1112 * Reverses amdgpu_bo_init() to tear down memory manager.
1113 */
1114void amdgpu_bo_fini(struct amdgpu_device *adev)
1115{
1116 amdgpu_ttm_fini(adev);
1117}
1118
1119/**
1120 * amdgpu_bo_set_tiling_flags - set tiling flags
1121 * @bo: &amdgpu_bo buffer object
1122 * @tiling_flags: new flags
1123 *
1124 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1125 * kernel driver to set the tiling flags on a buffer.
1126 *
1127 * Returns:
1128 * 0 for success or a negative error code on failure.
1129 */
1130int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1131{
1132 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1133 struct amdgpu_bo_user *ubo;
1134
1135 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1136 if (adev->family <= AMDGPU_FAMILY_CZ &&
1137 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1138 return -EINVAL;
1139
1140 ubo = to_amdgpu_bo_user(bo);
1141 ubo->tiling_flags = tiling_flags;
1142 return 0;
1143}
1144
1145/**
1146 * amdgpu_bo_get_tiling_flags - get tiling flags
1147 * @bo: &amdgpu_bo buffer object
1148 * @tiling_flags: returned flags
1149 *
1150 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1151 * set the tiling flags on a buffer.
1152 */
1153void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1154{
1155 struct amdgpu_bo_user *ubo;
1156
1157 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1158 dma_resv_assert_held(bo->tbo.base.resv);
1159 ubo = to_amdgpu_bo_user(bo);
1160
1161 if (tiling_flags)
1162 *tiling_flags = ubo->tiling_flags;
1163}
1164
1165/**
1166 * amdgpu_bo_set_metadata - set metadata
1167 * @bo: &amdgpu_bo buffer object
1168 * @metadata: new metadata
1169 * @metadata_size: size of the new metadata
1170 * @flags: flags of the new metadata
1171 *
1172 * Sets buffer object's metadata, its size and flags.
1173 * Used via GEM ioctl.
1174 *
1175 * Returns:
1176 * 0 for success or a negative error code on failure.
1177 */
1178int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
1179 uint32_t metadata_size, uint64_t flags)
1180{
1181 struct amdgpu_bo_user *ubo;
1182 void *buffer;
1183
1184 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1185 ubo = to_amdgpu_bo_user(bo);
1186 if (!metadata_size) {
1187 if (ubo->metadata_size) {
1188 kfree(ubo->metadata);
1189 ubo->metadata = NULL;
1190 ubo->metadata_size = 0;
1191 }
1192 return 0;
1193 }
1194
1195 if (metadata == NULL)
1196 return -EINVAL;
1197
1198 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1199 if (buffer == NULL)
1200 return -ENOMEM;
1201
1202 kfree(ubo->metadata);
1203 ubo->metadata_flags = flags;
1204 ubo->metadata = buffer;
1205 ubo->metadata_size = metadata_size;
1206
1207 return 0;
1208}
1209
1210/**
1211 * amdgpu_bo_get_metadata - get metadata
1212 * @bo: &amdgpu_bo buffer object
1213 * @buffer: returned metadata
1214 * @buffer_size: size of the buffer
1215 * @metadata_size: size of the returned metadata
1216 * @flags: flags of the returned metadata
1217 *
1218 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1219 * less than metadata_size.
1220 * Used via GEM ioctl.
1221 *
1222 * Returns:
1223 * 0 for success or a negative error code on failure.
1224 */
1225int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1226 size_t buffer_size, uint32_t *metadata_size,
1227 uint64_t *flags)
1228{
1229 struct amdgpu_bo_user *ubo;
1230
1231 if (!buffer && !metadata_size)
1232 return -EINVAL;
1233
1234 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1235 ubo = to_amdgpu_bo_user(bo);
1236 if (metadata_size)
1237 *metadata_size = ubo->metadata_size;
1238
1239 if (buffer) {
1240 if (buffer_size < ubo->metadata_size)
1241 return -EINVAL;
1242
1243 if (ubo->metadata_size)
1244 memcpy(buffer, ubo->metadata, ubo->metadata_size);
1245 }
1246
1247 if (flags)
1248 *flags = ubo->metadata_flags;
1249
1250 return 0;
1251}
1252
1253/**
1254 * amdgpu_bo_move_notify - notification about a memory move
1255 * @bo: pointer to a buffer object
1256 * @evict: if this move is evicting the buffer from the graphics address space
1257 * @new_mem: new information of the bufer object
1258 *
1259 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1260 * bookkeeping.
1261 * TTM driver callback which is called when ttm moves a buffer.
1262 */
1263void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1264 bool evict,
1265 struct ttm_resource *new_mem)
1266{
1267 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1268 struct amdgpu_bo *abo;
1269 struct ttm_resource *old_mem = bo->resource;
1270
1271 if (!amdgpu_bo_is_amdgpu_bo(bo))
1272 return;
1273
1274 abo = ttm_to_amdgpu_bo(bo);
1275 amdgpu_vm_bo_invalidate(adev, abo, evict);
1276
1277 amdgpu_bo_kunmap(abo);
1278
1279 if (abo->tbo.base.dma_buf && !abo->tbo.base.import_attach &&
1280 bo->resource->mem_type != TTM_PL_SYSTEM)
1281 dma_buf_move_notify(abo->tbo.base.dma_buf);
1282
1283 /* remember the eviction */
1284 if (evict)
1285 atomic64_inc(&adev->num_evictions);
1286
1287 /* update statistics */
1288 if (!new_mem)
1289 return;
1290
1291 /* move_notify is called before move happens */
1292 trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
1293}
1294
1295void amdgpu_bo_get_memory(struct amdgpu_bo *bo, uint64_t *vram_mem,
1296 uint64_t *gtt_mem, uint64_t *cpu_mem)
1297{
1298 unsigned int domain;
1299
1300 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type);
1301 switch (domain) {
1302 case AMDGPU_GEM_DOMAIN_VRAM:
1303 *vram_mem += amdgpu_bo_size(bo);
1304 break;
1305 case AMDGPU_GEM_DOMAIN_GTT:
1306 *gtt_mem += amdgpu_bo_size(bo);
1307 break;
1308 case AMDGPU_GEM_DOMAIN_CPU:
1309 default:
1310 *cpu_mem += amdgpu_bo_size(bo);
1311 break;
1312 }
1313}
1314
1315/**
1316 * amdgpu_bo_release_notify - notification about a BO being released
1317 * @bo: pointer to a buffer object
1318 *
1319 * Wipes VRAM buffers whose contents should not be leaked before the
1320 * memory is released.
1321 */
1322void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)
1323{
1324 struct dma_fence *fence = NULL;
1325 struct amdgpu_bo *abo;
1326 int r;
1327
1328 if (!amdgpu_bo_is_amdgpu_bo(bo))
1329 return;
1330
1331 abo = ttm_to_amdgpu_bo(bo);
1332
1333 if (abo->kfd_bo)
1334 amdgpu_amdkfd_unreserve_memory_limit(abo);
1335
1336 /* We only remove the fence if the resv has individualized. */
1337 WARN_ON_ONCE(bo->type == ttm_bo_type_kernel
1338 && bo->base.resv != &bo->base._resv);
1339 if (bo->base.resv == &bo->base._resv)
1340 amdgpu_amdkfd_remove_fence_on_pt_pd_bos(abo);
1341
1342 if (bo->resource->mem_type != TTM_PL_VRAM ||
1343 !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE))
1344 return;
1345
1346 dma_resv_lock(bo->base.resv, NULL);
1347
1348 r = amdgpu_fill_buffer(abo, AMDGPU_POISON, bo->base.resv, &fence);
1349 if (!WARN_ON(r)) {
1350 amdgpu_bo_fence(abo, fence, false);
1351 dma_fence_put(fence);
1352 }
1353
1354 dma_resv_unlock(bo->base.resv);
1355}
1356
1357/**
1358 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1359 * @bo: pointer to a buffer object
1360 *
1361 * Notifies the driver we are taking a fault on this BO and have reserved it,
1362 * also performs bookkeeping.
1363 * TTM driver callback for dealing with vm faults.
1364 *
1365 * Returns:
1366 * 0 for success or a negative error code on failure.
1367 */
1368vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1369{
1370 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1371 struct ttm_operation_ctx ctx = { false, false };
1372 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1373 unsigned long offset;
1374 int r;
1375
1376 /* Remember that this BO was accessed by the CPU */
1377 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1378
1379 if (bo->resource->mem_type != TTM_PL_VRAM)
1380 return 0;
1381
1382 offset = bo->resource->start << PAGE_SHIFT;
1383 if ((offset + bo->base.size) <= adev->gmc.visible_vram_size)
1384 return 0;
1385
1386 /* Can't move a pinned BO to visible VRAM */
1387 if (abo->tbo.pin_count > 0)
1388 return VM_FAULT_SIGBUS;
1389
1390 /* hurrah the memory is not visible ! */
1391 atomic64_inc(&adev->num_vram_cpu_page_faults);
1392 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1393 AMDGPU_GEM_DOMAIN_GTT);
1394
1395 /* Avoid costly evictions; only set GTT as a busy placement */
1396 abo->placement.num_busy_placement = 1;
1397 abo->placement.busy_placement = &abo->placements[1];
1398
1399 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1400 if (unlikely(r == -EBUSY || r == -ERESTARTSYS))
1401 return VM_FAULT_NOPAGE;
1402 else if (unlikely(r))
1403 return VM_FAULT_SIGBUS;
1404
1405 offset = bo->resource->start << PAGE_SHIFT;
1406 /* this should never happen */
1407 if (bo->resource->mem_type == TTM_PL_VRAM &&
1408 (offset + bo->base.size) > adev->gmc.visible_vram_size)
1409 return VM_FAULT_SIGBUS;
1410
1411 ttm_bo_move_to_lru_tail_unlocked(bo);
1412 return 0;
1413}
1414
1415/**
1416 * amdgpu_bo_fence - add fence to buffer object
1417 *
1418 * @bo: buffer object in question
1419 * @fence: fence to add
1420 * @shared: true if fence should be added shared
1421 *
1422 */
1423void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1424 bool shared)
1425{
1426 struct dma_resv *resv = bo->tbo.base.resv;
1427
1428 if (shared)
1429 dma_resv_add_shared_fence(resv, fence);
1430 else
1431 dma_resv_add_excl_fence(resv, fence);
1432}
1433
1434/**
1435 * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences
1436 *
1437 * @adev: amdgpu device pointer
1438 * @resv: reservation object to sync to
1439 * @sync_mode: synchronization mode
1440 * @owner: fence owner
1441 * @intr: Whether the wait is interruptible
1442 *
1443 * Extract the fences from the reservation object and waits for them to finish.
1444 *
1445 * Returns:
1446 * 0 on success, errno otherwise.
1447 */
1448int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv,
1449 enum amdgpu_sync_mode sync_mode, void *owner,
1450 bool intr)
1451{
1452 struct amdgpu_sync sync;
1453 int r;
1454
1455 amdgpu_sync_create(&sync);
1456 amdgpu_sync_resv(adev, &sync, resv, sync_mode, owner);
1457 r = amdgpu_sync_wait(&sync, intr);
1458 amdgpu_sync_free(&sync);
1459 return r;
1460}
1461
1462/**
1463 * amdgpu_bo_sync_wait - Wrapper for amdgpu_bo_sync_wait_resv
1464 * @bo: buffer object to wait for
1465 * @owner: fence owner
1466 * @intr: Whether the wait is interruptible
1467 *
1468 * Wrapper to wait for fences in a BO.
1469 * Returns:
1470 * 0 on success, errno otherwise.
1471 */
1472int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)
1473{
1474 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1475
1476 return amdgpu_bo_sync_wait_resv(adev, bo->tbo.base.resv,
1477 AMDGPU_SYNC_NE_OWNER, owner, intr);
1478}
1479
1480/**
1481 * amdgpu_bo_gpu_offset - return GPU offset of bo
1482 * @bo: amdgpu object for which we query the offset
1483 *
1484 * Note: object should either be pinned or reserved when calling this
1485 * function, it might be useful to add check for this for debugging.
1486 *
1487 * Returns:
1488 * current GPU offset of the object.
1489 */
1490u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1491{
1492 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_SYSTEM);
1493 WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) &&
1494 !bo->tbo.pin_count && bo->tbo.type != ttm_bo_type_kernel);
1495 WARN_ON_ONCE(bo->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET);
1496 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_VRAM &&
1497 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1498
1499 return amdgpu_bo_gpu_offset_no_check(bo);
1500}
1501
1502/**
1503 * amdgpu_bo_gpu_offset_no_check - return GPU offset of bo
1504 * @bo: amdgpu object for which we query the offset
1505 *
1506 * Returns:
1507 * current GPU offset of the object without raising warnings.
1508 */
1509u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo)
1510{
1511 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1512 uint64_t offset;
1513
1514 offset = (bo->tbo.resource->start << PAGE_SHIFT) +
1515 amdgpu_ttm_domain_start(adev, bo->tbo.resource->mem_type);
1516
1517 return amdgpu_gmc_sign_extend(offset);
1518}
1519
1520/**
1521 * amdgpu_bo_get_preferred_pin_domain - get preferred domain for scanout
1522 * @adev: amdgpu device object
1523 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1524 *
1525 * Returns:
1526 * Which of the allowed domains is preferred for pinning the BO for scanout.
1527 */
1528uint32_t amdgpu_bo_get_preferred_pin_domain(struct amdgpu_device *adev,
1529 uint32_t domain)
1530{
1531 if (domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) {
1532 domain = AMDGPU_GEM_DOMAIN_VRAM;
1533 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1534 domain = AMDGPU_GEM_DOMAIN_GTT;
1535 }
1536 return domain;
1537}
1538
1539#if defined(CONFIG_DEBUG_FS)
1540#define amdgpu_bo_print_flag(m, bo, flag) \
1541 do { \
1542 if (bo->flags & (AMDGPU_GEM_CREATE_ ## flag)) { \
1543 seq_printf((m), " " #flag); \
1544 } \
1545 } while (0)
1546
1547/**
1548 * amdgpu_bo_print_info - print BO info in debugfs file
1549 *
1550 * @id: Index or Id of the BO
1551 * @bo: Requested BO for printing info
1552 * @m: debugfs file
1553 *
1554 * Print BO information in debugfs file
1555 *
1556 * Returns:
1557 * Size of the BO in bytes.
1558 */
1559u64 amdgpu_bo_print_info(int id, struct amdgpu_bo *bo, struct seq_file *m)
1560{
1561 struct dma_buf_attachment *attachment;
1562 struct dma_buf *dma_buf;
1563 unsigned int domain;
1564 const char *placement;
1565 unsigned int pin_count;
1566 u64 size;
1567
1568 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type);
1569 switch (domain) {
1570 case AMDGPU_GEM_DOMAIN_VRAM:
1571 placement = "VRAM";
1572 break;
1573 case AMDGPU_GEM_DOMAIN_GTT:
1574 placement = " GTT";
1575 break;
1576 case AMDGPU_GEM_DOMAIN_CPU:
1577 default:
1578 placement = " CPU";
1579 break;
1580 }
1581
1582 size = amdgpu_bo_size(bo);
1583 seq_printf(m, "\t\t0x%08x: %12lld byte %s",
1584 id, size, placement);
1585
1586 pin_count = READ_ONCE(bo->tbo.pin_count);
1587 if (pin_count)
1588 seq_printf(m, " pin count %d", pin_count);
1589
1590 dma_buf = READ_ONCE(bo->tbo.base.dma_buf);
1591 attachment = READ_ONCE(bo->tbo.base.import_attach);
1592
1593 if (attachment)
1594 seq_printf(m, " imported from %p", dma_buf);
1595 else if (dma_buf)
1596 seq_printf(m, " exported as %p", dma_buf);
1597
1598 amdgpu_bo_print_flag(m, bo, CPU_ACCESS_REQUIRED);
1599 amdgpu_bo_print_flag(m, bo, NO_CPU_ACCESS);
1600 amdgpu_bo_print_flag(m, bo, CPU_GTT_USWC);
1601 amdgpu_bo_print_flag(m, bo, VRAM_CLEARED);
1602 amdgpu_bo_print_flag(m, bo, VRAM_CONTIGUOUS);
1603 amdgpu_bo_print_flag(m, bo, VM_ALWAYS_VALID);
1604 amdgpu_bo_print_flag(m, bo, EXPLICIT_SYNC);
1605
1606 seq_puts(m, "\n");
1607
1608 return size;
1609}
1610#endif