1 /*
2 * Copyright 2009 Jerome Glisse.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sub license, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
12 *
13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19 * USE OR OTHER DEALINGS IN THE SOFTWARE.
20 *
21 * The above copyright notice and this permission notice (including the
22 * next paragraph) shall be included in all copies or substantial portions
23 * of the Software.
24 *
25 */
26 /*
27 * Authors:
28 * Jerome Glisse <glisse@freedesktop.org>
29 * Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
30 * Dave Airlie
31 */
32 #include <linux/list.h>
33 #include <linux/slab.h>
34 #include <linux/dma-buf.h>
35
36 #include <drm/drm_drv.h>
37 #include <drm/amdgpu_drm.h>
38 #include <drm/drm_cache.h>
39 #include "amdgpu.h"
40 #include "amdgpu_trace.h"
41 #include "amdgpu_amdkfd.h"
42
43 /**
44 * DOC: amdgpu_object
45 *
46 * This defines the interfaces to operate on an &amdgpu_bo buffer object which
47 * represents memory used by driver (VRAM, system memory, etc.). The driver
48 * provides DRM/GEM APIs to userspace. DRM/GEM APIs then use these interfaces
49 * to create/destroy/set buffer object which are then managed by the kernel TTM
50 * memory manager.
51 * The interfaces are also used internally by kernel clients, including gfx,
52 * uvd, etc. for kernel managed allocations used by the GPU.
53 *
54 */
55
amdgpu_bo_destroy(struct ttm_buffer_object * tbo)56 static void amdgpu_bo_destroy(struct ttm_buffer_object *tbo)
57 {
58 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
59
60 amdgpu_bo_kunmap(bo);
61
62 if (bo->tbo.base.import_attach)
63 drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
64 drm_gem_object_release(&bo->tbo.base);
65 amdgpu_bo_unref(&bo->parent);
66 kvfree(bo);
67 }
68
amdgpu_bo_user_destroy(struct ttm_buffer_object * tbo)69 static void amdgpu_bo_user_destroy(struct ttm_buffer_object *tbo)
70 {
71 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
72 struct amdgpu_bo_user *ubo;
73
74 ubo = to_amdgpu_bo_user(bo);
75 kfree(ubo->metadata);
76 amdgpu_bo_destroy(tbo);
77 }
78
amdgpu_bo_vm_destroy(struct ttm_buffer_object * tbo)79 static void amdgpu_bo_vm_destroy(struct ttm_buffer_object *tbo)
80 {
81 struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
82 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
83 struct amdgpu_bo_vm *vmbo;
84
85 vmbo = to_amdgpu_bo_vm(bo);
86 /* in case amdgpu_device_recover_vram got NULL of bo->parent */
87 if (!list_empty(&vmbo->shadow_list)) {
88 mutex_lock(&adev->shadow_list_lock);
89 list_del_init(&vmbo->shadow_list);
90 mutex_unlock(&adev->shadow_list_lock);
91 }
92
93 amdgpu_bo_destroy(tbo);
94 }
95
96 /**
97 * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
98 * @bo: buffer object to be checked
99 *
100 * Uses destroy function associated with the object to determine if this is
101 * an &amdgpu_bo.
102 *
103 * Returns:
104 * true if the object belongs to &amdgpu_bo, false if not.
105 */
amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object * bo)106 bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
107 {
108 if (bo->destroy == &amdgpu_bo_destroy ||
109 bo->destroy == &amdgpu_bo_user_destroy ||
110 bo->destroy == &amdgpu_bo_vm_destroy)
111 return true;
112
113 return false;
114 }
115
116 /**
117 * amdgpu_bo_placement_from_domain - set buffer's placement
118 * @abo: &amdgpu_bo buffer object whose placement is to be set
119 * @domain: requested domain
120 *
121 * Sets buffer's placement according to requested domain and the buffer's
122 * flags.
123 */
amdgpu_bo_placement_from_domain(struct amdgpu_bo * abo,u32 domain)124 void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
125 {
126 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
127 struct ttm_placement *placement = &abo->placement;
128 struct ttm_place *places = abo->placements;
129 u64 flags = abo->flags;
130 u32 c = 0;
131
132 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
133 unsigned visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
134
135 places[c].fpfn = 0;
136 places[c].lpfn = 0;
137 places[c].mem_type = TTM_PL_VRAM;
138 places[c].flags = 0;
139
140 if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
141 places[c].lpfn = visible_pfn;
142 else
143 places[c].flags |= TTM_PL_FLAG_TOPDOWN;
144
145 if (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
146 places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
147 c++;
148 }
149
150 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
151 places[c].fpfn = 0;
152 places[c].lpfn = 0;
153 places[c].mem_type =
154 abo->flags & AMDGPU_GEM_CREATE_PREEMPTIBLE ?
155 AMDGPU_PL_PREEMPT : TTM_PL_TT;
156 places[c].flags = 0;
157 c++;
158 }
159
160 if (domain & AMDGPU_GEM_DOMAIN_CPU) {
161 places[c].fpfn = 0;
162 places[c].lpfn = 0;
163 places[c].mem_type = TTM_PL_SYSTEM;
164 places[c].flags = 0;
165 c++;
166 }
167
168 if (domain & AMDGPU_GEM_DOMAIN_GDS) {
169 places[c].fpfn = 0;
170 places[c].lpfn = 0;
171 places[c].mem_type = AMDGPU_PL_GDS;
172 places[c].flags = 0;
173 c++;
174 }
175
176 if (domain & AMDGPU_GEM_DOMAIN_GWS) {
177 places[c].fpfn = 0;
178 places[c].lpfn = 0;
179 places[c].mem_type = AMDGPU_PL_GWS;
180 places[c].flags = 0;
181 c++;
182 }
183
184 if (domain & AMDGPU_GEM_DOMAIN_OA) {
185 places[c].fpfn = 0;
186 places[c].lpfn = 0;
187 places[c].mem_type = AMDGPU_PL_OA;
188 places[c].flags = 0;
189 c++;
190 }
191
192 if (!c) {
193 places[c].fpfn = 0;
194 places[c].lpfn = 0;
195 places[c].mem_type = TTM_PL_SYSTEM;
196 places[c].flags = 0;
197 c++;
198 }
199
200 BUG_ON(c > AMDGPU_BO_MAX_PLACEMENTS);
201
202 placement->num_placement = c;
203 placement->placement = places;
204
205 placement->num_busy_placement = c;
206 placement->busy_placement = places;
207 }
208
209 /**
210 * amdgpu_bo_create_reserved - create reserved BO for kernel use
211 *
212 * @adev: amdgpu device object
213 * @size: size for the new BO
214 * @align: alignment for the new BO
215 * @domain: where to place it
216 * @bo_ptr: used to initialize BOs in structures
217 * @gpu_addr: GPU addr of the pinned BO
218 * @cpu_addr: optional CPU address mapping
219 *
220 * Allocates and pins a BO for kernel internal use, and returns it still
221 * reserved.
222 *
223 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
224 *
225 * Returns:
226 * 0 on success, negative error code otherwise.
227 */
amdgpu_bo_create_reserved(struct amdgpu_device * adev,unsigned long size,int align,u32 domain,struct amdgpu_bo ** bo_ptr,u64 * gpu_addr,void ** cpu_addr)228 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
229 unsigned long size, int align,
230 u32 domain, struct amdgpu_bo **bo_ptr,
231 u64 *gpu_addr, void **cpu_addr)
232 {
233 struct amdgpu_bo_param bp;
234 bool free = false;
235 int r;
236
237 if (!size) {
238 amdgpu_bo_unref(bo_ptr);
239 return 0;
240 }
241
242 memset(&bp, 0, sizeof(bp));
243 bp.size = size;
244 bp.byte_align = align;
245 bp.domain = domain;
246 bp.flags = cpu_addr ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
247 : AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
248 bp.flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
249 bp.type = ttm_bo_type_kernel;
250 bp.resv = NULL;
251 bp.bo_ptr_size = sizeof(struct amdgpu_bo);
252
253 if (!*bo_ptr) {
254 r = amdgpu_bo_create(adev, &bp, bo_ptr);
255 if (r) {
256 dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
257 r);
258 return r;
259 }
260 free = true;
261 }
262
263 r = amdgpu_bo_reserve(*bo_ptr, false);
264 if (r) {
265 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
266 goto error_free;
267 }
268
269 r = amdgpu_bo_pin(*bo_ptr, domain);
270 if (r) {
271 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
272 goto error_unreserve;
273 }
274
275 r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
276 if (r) {
277 dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
278 goto error_unpin;
279 }
280
281 if (gpu_addr)
282 *gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
283
284 if (cpu_addr) {
285 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
286 if (r) {
287 dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
288 goto error_unpin;
289 }
290 }
291
292 return 0;
293
294 error_unpin:
295 amdgpu_bo_unpin(*bo_ptr);
296 error_unreserve:
297 amdgpu_bo_unreserve(*bo_ptr);
298
299 error_free:
300 if (free)
301 amdgpu_bo_unref(bo_ptr);
302
303 return r;
304 }
305
306 /**
307 * amdgpu_bo_create_kernel - create BO for kernel use
308 *
309 * @adev: amdgpu device object
310 * @size: size for the new BO
311 * @align: alignment for the new BO
312 * @domain: where to place it
313 * @bo_ptr: used to initialize BOs in structures
314 * @gpu_addr: GPU addr of the pinned BO
315 * @cpu_addr: optional CPU address mapping
316 *
317 * Allocates and pins a BO for kernel internal use.
318 *
319 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
320 *
321 * Returns:
322 * 0 on success, negative error code otherwise.
323 */
amdgpu_bo_create_kernel(struct amdgpu_device * adev,unsigned long size,int align,u32 domain,struct amdgpu_bo ** bo_ptr,u64 * gpu_addr,void ** cpu_addr)324 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
325 unsigned long size, int align,
326 u32 domain, struct amdgpu_bo **bo_ptr,
327 u64 *gpu_addr, void **cpu_addr)
328 {
329 int r;
330
331 r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
332 gpu_addr, cpu_addr);
333
334 if (r)
335 return r;
336
337 if (*bo_ptr)
338 amdgpu_bo_unreserve(*bo_ptr);
339
340 return 0;
341 }
342
343 /**
344 * amdgpu_bo_create_kernel_at - create BO for kernel use at specific location
345 *
346 * @adev: amdgpu device object
347 * @offset: offset of the BO
348 * @size: size of the BO
349 * @domain: where to place it
350 * @bo_ptr: used to initialize BOs in structures
351 * @cpu_addr: optional CPU address mapping
352 *
353 * Creates a kernel BO at a specific offset in the address space of the domain.
354 *
355 * Returns:
356 * 0 on success, negative error code otherwise.
357 */
amdgpu_bo_create_kernel_at(struct amdgpu_device * adev,uint64_t offset,uint64_t size,uint32_t domain,struct amdgpu_bo ** bo_ptr,void ** cpu_addr)358 int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev,
359 uint64_t offset, uint64_t size, uint32_t domain,
360 struct amdgpu_bo **bo_ptr, void **cpu_addr)
361 {
362 struct ttm_operation_ctx ctx = { false, false };
363 unsigned int i;
364 int r;
365
366 offset &= PAGE_MASK;
367 size = ALIGN(size, PAGE_SIZE);
368
369 r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE, domain, bo_ptr,
370 NULL, cpu_addr);
371 if (r)
372 return r;
373
374 if ((*bo_ptr) == NULL)
375 return 0;
376
377 /*
378 * Remove the original mem node and create a new one at the request
379 * position.
380 */
381 if (cpu_addr)
382 amdgpu_bo_kunmap(*bo_ptr);
383
384 ttm_resource_free(&(*bo_ptr)->tbo, &(*bo_ptr)->tbo.resource);
385
386 for (i = 0; i < (*bo_ptr)->placement.num_placement; ++i) {
387 (*bo_ptr)->placements[i].fpfn = offset >> PAGE_SHIFT;
388 (*bo_ptr)->placements[i].lpfn = (offset + size) >> PAGE_SHIFT;
389 }
390 r = ttm_bo_mem_space(&(*bo_ptr)->tbo, &(*bo_ptr)->placement,
391 &(*bo_ptr)->tbo.resource, &ctx);
392 if (r)
393 goto error;
394
395 if (cpu_addr) {
396 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
397 if (r)
398 goto error;
399 }
400
401 amdgpu_bo_unreserve(*bo_ptr);
402 return 0;
403
404 error:
405 amdgpu_bo_unreserve(*bo_ptr);
406 amdgpu_bo_unref(bo_ptr);
407 return r;
408 }
409
410 /**
411 * amdgpu_bo_free_kernel - free BO for kernel use
412 *
413 * @bo: amdgpu BO to free
414 * @gpu_addr: pointer to where the BO's GPU memory space address was stored
415 * @cpu_addr: pointer to where the BO's CPU memory space address was stored
416 *
417 * unmaps and unpin a BO for kernel internal use.
418 */
amdgpu_bo_free_kernel(struct amdgpu_bo ** bo,u64 * gpu_addr,void ** cpu_addr)419 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
420 void **cpu_addr)
421 {
422 if (*bo == NULL)
423 return;
424
425 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
426 if (cpu_addr)
427 amdgpu_bo_kunmap(*bo);
428
429 amdgpu_bo_unpin(*bo);
430 amdgpu_bo_unreserve(*bo);
431 }
432 amdgpu_bo_unref(bo);
433
434 if (gpu_addr)
435 *gpu_addr = 0;
436
437 if (cpu_addr)
438 *cpu_addr = NULL;
439 }
440
441 /* Validate bo size is bit bigger then the request domain */
amdgpu_bo_validate_size(struct amdgpu_device * adev,unsigned long size,u32 domain)442 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
443 unsigned long size, u32 domain)
444 {
445 struct ttm_resource_manager *man = NULL;
446
447 /*
448 * If GTT is part of requested domains the check must succeed to
449 * allow fall back to GTT
450 */
451 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
452 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_TT);
453
454 if (size < man->size)
455 return true;
456 else
457 goto fail;
458 }
459
460 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
461 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
462
463 if (size < man->size)
464 return true;
465 else
466 goto fail;
467 }
468
469
470 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */
471 return true;
472
473 fail:
474 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size,
475 man->size);
476 return false;
477 }
478
amdgpu_bo_support_uswc(u64 bo_flags)479 bool amdgpu_bo_support_uswc(u64 bo_flags)
480 {
481
482 #ifdef CONFIG_X86_32
483 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
484 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
485 */
486 return false;
487 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
488 /* Don't try to enable write-combining when it can't work, or things
489 * may be slow
490 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
491 */
492
493 #ifndef CONFIG_COMPILE_TEST
494 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
495 thanks to write-combining
496 #endif
497
498 if (bo_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
499 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
500 "better performance thanks to write-combining\n");
501 return false;
502 #else
503 /* For architectures that don't support WC memory,
504 * mask out the WC flag from the BO
505 */
506 if (!drm_arch_can_wc_memory())
507 return false;
508
509 return true;
510 #endif
511 }
512
513 /**
514 * amdgpu_bo_create - create an &amdgpu_bo buffer object
515 * @adev: amdgpu device object
516 * @bp: parameters to be used for the buffer object
517 * @bo_ptr: pointer to the buffer object pointer
518 *
519 * Creates an &amdgpu_bo buffer object.
520 *
521 * Returns:
522 * 0 for success or a negative error code on failure.
523 */
amdgpu_bo_create(struct amdgpu_device * adev,struct amdgpu_bo_param * bp,struct amdgpu_bo ** bo_ptr)524 int amdgpu_bo_create(struct amdgpu_device *adev,
525 struct amdgpu_bo_param *bp,
526 struct amdgpu_bo **bo_ptr)
527 {
528 struct ttm_operation_ctx ctx = {
529 .interruptible = (bp->type != ttm_bo_type_kernel),
530 .no_wait_gpu = bp->no_wait_gpu,
531 /* We opt to avoid OOM on system pages allocations */
532 .gfp_retry_mayfail = true,
533 .allow_res_evict = bp->type != ttm_bo_type_kernel,
534 .resv = bp->resv
535 };
536 struct amdgpu_bo *bo;
537 unsigned long page_align, size = bp->size;
538 int r;
539
540 /* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
541 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
542 /* GWS and OA don't need any alignment. */
543 page_align = bp->byte_align;
544 size <<= PAGE_SHIFT;
545 } else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
546 /* Both size and alignment must be a multiple of 4. */
547 page_align = ALIGN(bp->byte_align, 4);
548 size = ALIGN(size, 4) << PAGE_SHIFT;
549 } else {
550 /* Memory should be aligned at least to a page size. */
551 page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
552 size = ALIGN(size, PAGE_SIZE);
553 }
554
555 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
556 return -ENOMEM;
557
558 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo));
559
560 *bo_ptr = NULL;
561 bo = kvzalloc(bp->bo_ptr_size, GFP_KERNEL);
562 if (bo == NULL)
563 return -ENOMEM;
564 drm_gem_private_object_init(adev_to_drm(adev), &bo->tbo.base, size);
565 bo->vm_bo = NULL;
566 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
567 bp->domain;
568 bo->allowed_domains = bo->preferred_domains;
569 if (bp->type != ttm_bo_type_kernel &&
570 !(bp->flags & AMDGPU_GEM_CREATE_DISCARDABLE) &&
571 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
572 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
573
574 bo->flags = bp->flags;
575
576 if (!amdgpu_bo_support_uswc(bo->flags))
577 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
578
579 if (adev->ras_enabled)
580 bo->flags |= AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE;
581
582 bo->tbo.bdev = &adev->mman.bdev;
583 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
584 AMDGPU_GEM_DOMAIN_GDS))
585 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
586 else
587 amdgpu_bo_placement_from_domain(bo, bp->domain);
588 if (bp->type == ttm_bo_type_kernel)
589 bo->tbo.priority = 1;
590
591 if (!bp->destroy)
592 bp->destroy = &amdgpu_bo_destroy;
593
594 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, bp->type,
595 &bo->placement, page_align, &ctx, NULL,
596 bp->resv, bp->destroy);
597 if (unlikely(r != 0))
598 return r;
599
600 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
601 bo->tbo.resource->mem_type == TTM_PL_VRAM &&
602 bo->tbo.resource->start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
603 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
604 ctx.bytes_moved);
605 else
606 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
607
608 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
609 bo->tbo.resource->mem_type == TTM_PL_VRAM) {
610 struct dma_fence *fence;
611
612 r = amdgpu_fill_buffer(bo, 0, bo->tbo.base.resv, &fence);
613 if (unlikely(r))
614 goto fail_unreserve;
615
616 dma_resv_add_fence(bo->tbo.base.resv, fence,
617 DMA_RESV_USAGE_KERNEL);
618 dma_fence_put(fence);
619 }
620 if (!bp->resv)
621 amdgpu_bo_unreserve(bo);
622 *bo_ptr = bo;
623
624 trace_amdgpu_bo_create(bo);
625
626 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
627 if (bp->type == ttm_bo_type_device)
628 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
629
630 return 0;
631
632 fail_unreserve:
633 if (!bp->resv)
634 dma_resv_unlock(bo->tbo.base.resv);
635 amdgpu_bo_unref(&bo);
636 return r;
637 }
638
639 /**
640 * amdgpu_bo_create_user - create an &amdgpu_bo_user buffer object
641 * @adev: amdgpu device object
642 * @bp: parameters to be used for the buffer object
643 * @ubo_ptr: pointer to the buffer object pointer
644 *
645 * Create a BO to be used by user application;
646 *
647 * Returns:
648 * 0 for success or a negative error code on failure.
649 */
650
amdgpu_bo_create_user(struct amdgpu_device * adev,struct amdgpu_bo_param * bp,struct amdgpu_bo_user ** ubo_ptr)651 int amdgpu_bo_create_user(struct amdgpu_device *adev,
652 struct amdgpu_bo_param *bp,
653 struct amdgpu_bo_user **ubo_ptr)
654 {
655 struct amdgpu_bo *bo_ptr;
656 int r;
657
658 bp->bo_ptr_size = sizeof(struct amdgpu_bo_user);
659 bp->destroy = &amdgpu_bo_user_destroy;
660 r = amdgpu_bo_create(adev, bp, &bo_ptr);
661 if (r)
662 return r;
663
664 *ubo_ptr = to_amdgpu_bo_user(bo_ptr);
665 return r;
666 }
667
668 /**
669 * amdgpu_bo_create_vm - create an &amdgpu_bo_vm buffer object
670 * @adev: amdgpu device object
671 * @bp: parameters to be used for the buffer object
672 * @vmbo_ptr: pointer to the buffer object pointer
673 *
674 * Create a BO to be for GPUVM.
675 *
676 * Returns:
677 * 0 for success or a negative error code on failure.
678 */
679
amdgpu_bo_create_vm(struct amdgpu_device * adev,struct amdgpu_bo_param * bp,struct amdgpu_bo_vm ** vmbo_ptr)680 int amdgpu_bo_create_vm(struct amdgpu_device *adev,
681 struct amdgpu_bo_param *bp,
682 struct amdgpu_bo_vm **vmbo_ptr)
683 {
684 struct amdgpu_bo *bo_ptr;
685 int r;
686
687 /* bo_ptr_size will be determined by the caller and it depends on
688 * num of amdgpu_vm_pt entries.
689 */
690 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo_vm));
691 r = amdgpu_bo_create(adev, bp, &bo_ptr);
692 if (r)
693 return r;
694
695 *vmbo_ptr = to_amdgpu_bo_vm(bo_ptr);
696 INIT_LIST_HEAD(&(*vmbo_ptr)->shadow_list);
697 /* Set destroy callback to amdgpu_bo_vm_destroy after vmbo->shadow_list
698 * is initialized.
699 */
700 bo_ptr->tbo.destroy = &amdgpu_bo_vm_destroy;
701 return r;
702 }
703
704 /**
705 * amdgpu_bo_add_to_shadow_list - add a BO to the shadow list
706 *
707 * @vmbo: BO that will be inserted into the shadow list
708 *
709 * Insert a BO to the shadow list.
710 */
amdgpu_bo_add_to_shadow_list(struct amdgpu_bo_vm * vmbo)711 void amdgpu_bo_add_to_shadow_list(struct amdgpu_bo_vm *vmbo)
712 {
713 struct amdgpu_device *adev = amdgpu_ttm_adev(vmbo->bo.tbo.bdev);
714
715 mutex_lock(&adev->shadow_list_lock);
716 list_add_tail(&vmbo->shadow_list, &adev->shadow_list);
717 mutex_unlock(&adev->shadow_list_lock);
718 }
719
720 /**
721 * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow
722 *
723 * @shadow: &amdgpu_bo shadow to be restored
724 * @fence: dma_fence associated with the operation
725 *
726 * Copies a buffer object's shadow content back to the object.
727 * This is used for recovering a buffer from its shadow in case of a gpu
728 * reset where vram context may be lost.
729 *
730 * Returns:
731 * 0 for success or a negative error code on failure.
732 */
amdgpu_bo_restore_shadow(struct amdgpu_bo * shadow,struct dma_fence ** fence)733 int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence)
734
735 {
736 struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev);
737 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
738 uint64_t shadow_addr, parent_addr;
739
740 shadow_addr = amdgpu_bo_gpu_offset(shadow);
741 parent_addr = amdgpu_bo_gpu_offset(shadow->parent);
742
743 return amdgpu_copy_buffer(ring, shadow_addr, parent_addr,
744 amdgpu_bo_size(shadow), NULL, fence,
745 true, false, false);
746 }
747
748 /**
749 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
750 * @bo: &amdgpu_bo buffer object to be mapped
751 * @ptr: kernel virtual address to be returned
752 *
753 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
754 * amdgpu_bo_kptr() to get the kernel virtual address.
755 *
756 * Returns:
757 * 0 for success or a negative error code on failure.
758 */
amdgpu_bo_kmap(struct amdgpu_bo * bo,void ** ptr)759 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
760 {
761 void *kptr;
762 long r;
763
764 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
765 return -EPERM;
766
767 r = dma_resv_wait_timeout(bo->tbo.base.resv, DMA_RESV_USAGE_KERNEL,
768 false, MAX_SCHEDULE_TIMEOUT);
769 if (r < 0)
770 return r;
771
772 kptr = amdgpu_bo_kptr(bo);
773 if (kptr) {
774 if (ptr)
775 *ptr = kptr;
776 return 0;
777 }
778
779 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.resource->num_pages, &bo->kmap);
780 if (r)
781 return r;
782
783 if (ptr)
784 *ptr = amdgpu_bo_kptr(bo);
785
786 return 0;
787 }
788
789 /**
790 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
791 * @bo: &amdgpu_bo buffer object
792 *
793 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
794 *
795 * Returns:
796 * the virtual address of a buffer object area.
797 */
amdgpu_bo_kptr(struct amdgpu_bo * bo)798 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
799 {
800 bool is_iomem;
801
802 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
803 }
804
805 /**
806 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
807 * @bo: &amdgpu_bo buffer object to be unmapped
808 *
809 * Unmaps a kernel map set up by amdgpu_bo_kmap().
810 */
amdgpu_bo_kunmap(struct amdgpu_bo * bo)811 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
812 {
813 if (bo->kmap.bo)
814 ttm_bo_kunmap(&bo->kmap);
815 }
816
817 /**
818 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
819 * @bo: &amdgpu_bo buffer object
820 *
821 * References the contained &ttm_buffer_object.
822 *
823 * Returns:
824 * a refcounted pointer to the &amdgpu_bo buffer object.
825 */
amdgpu_bo_ref(struct amdgpu_bo * bo)826 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
827 {
828 if (bo == NULL)
829 return NULL;
830
831 ttm_bo_get(&bo->tbo);
832 return bo;
833 }
834
835 /**
836 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
837 * @bo: &amdgpu_bo buffer object
838 *
839 * Unreferences the contained &ttm_buffer_object and clear the pointer
840 */
amdgpu_bo_unref(struct amdgpu_bo ** bo)841 void amdgpu_bo_unref(struct amdgpu_bo **bo)
842 {
843 struct ttm_buffer_object *tbo;
844
845 if ((*bo) == NULL)
846 return;
847
848 tbo = &((*bo)->tbo);
849 ttm_bo_put(tbo);
850 *bo = NULL;
851 }
852
853 /**
854 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
855 * @bo: &amdgpu_bo buffer object to be pinned
856 * @domain: domain to be pinned to
857 * @min_offset: the start of requested address range
858 * @max_offset: the end of requested address range
859 *
860 * Pins the buffer object according to requested domain and address range. If
861 * the memory is unbound gart memory, binds the pages into gart table. Adjusts
862 * pin_count and pin_size accordingly.
863 *
864 * Pinning means to lock pages in memory along with keeping them at a fixed
865 * offset. It is required when a buffer can not be moved, for example, when
866 * a display buffer is being scanned out.
867 *
868 * Compared with amdgpu_bo_pin(), this function gives more flexibility on
869 * where to pin a buffer if there are specific restrictions on where a buffer
870 * must be located.
871 *
872 * Returns:
873 * 0 for success or a negative error code on failure.
874 */
amdgpu_bo_pin_restricted(struct amdgpu_bo * bo,u32 domain,u64 min_offset,u64 max_offset)875 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
876 u64 min_offset, u64 max_offset)
877 {
878 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
879 struct ttm_operation_ctx ctx = { false, false };
880 int r, i;
881
882 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
883 return -EPERM;
884
885 if (WARN_ON_ONCE(min_offset > max_offset))
886 return -EINVAL;
887
888 /* Check domain to be pinned to against preferred domains */
889 if (bo->preferred_domains & domain)
890 domain = bo->preferred_domains & domain;
891
892 /* A shared bo cannot be migrated to VRAM */
893 if (bo->tbo.base.import_attach) {
894 if (domain & AMDGPU_GEM_DOMAIN_GTT)
895 domain = AMDGPU_GEM_DOMAIN_GTT;
896 else
897 return -EINVAL;
898 }
899
900 if (bo->tbo.pin_count) {
901 uint32_t mem_type = bo->tbo.resource->mem_type;
902 uint32_t mem_flags = bo->tbo.resource->placement;
903
904 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
905 return -EINVAL;
906
907 if ((mem_type == TTM_PL_VRAM) &&
908 (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) &&
909 !(mem_flags & TTM_PL_FLAG_CONTIGUOUS))
910 return -EINVAL;
911
912 ttm_bo_pin(&bo->tbo);
913
914 if (max_offset != 0) {
915 u64 domain_start = amdgpu_ttm_domain_start(adev,
916 mem_type);
917 WARN_ON_ONCE(max_offset <
918 (amdgpu_bo_gpu_offset(bo) - domain_start));
919 }
920
921 return 0;
922 }
923
924 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
925 * See function amdgpu_display_supported_domains()
926 */
927 domain = amdgpu_bo_get_preferred_domain(adev, domain);
928
929 if (bo->tbo.base.import_attach)
930 dma_buf_pin(bo->tbo.base.import_attach);
931
932 /* force to pin into visible video ram */
933 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
934 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
935 amdgpu_bo_placement_from_domain(bo, domain);
936 for (i = 0; i < bo->placement.num_placement; i++) {
937 unsigned fpfn, lpfn;
938
939 fpfn = min_offset >> PAGE_SHIFT;
940 lpfn = max_offset >> PAGE_SHIFT;
941
942 if (fpfn > bo->placements[i].fpfn)
943 bo->placements[i].fpfn = fpfn;
944 if (!bo->placements[i].lpfn ||
945 (lpfn && lpfn < bo->placements[i].lpfn))
946 bo->placements[i].lpfn = lpfn;
947 }
948
949 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
950 if (unlikely(r)) {
951 dev_err(adev->dev, "%p pin failed\n", bo);
952 goto error;
953 }
954
955 ttm_bo_pin(&bo->tbo);
956
957 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type);
958 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
959 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
960 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
961 &adev->visible_pin_size);
962 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
963 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
964 }
965
966 error:
967 return r;
968 }
969
970 /**
971 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
972 * @bo: &amdgpu_bo buffer object to be pinned
973 * @domain: domain to be pinned to
974 *
975 * A simple wrapper to amdgpu_bo_pin_restricted().
976 * Provides a simpler API for buffers that do not have any strict restrictions
977 * on where a buffer must be located.
978 *
979 * Returns:
980 * 0 for success or a negative error code on failure.
981 */
amdgpu_bo_pin(struct amdgpu_bo * bo,u32 domain)982 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
983 {
984 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
985 return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
986 }
987
988 /**
989 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
990 * @bo: &amdgpu_bo buffer object to be unpinned
991 *
992 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
993 * Changes placement and pin size accordingly.
994 *
995 * Returns:
996 * 0 for success or a negative error code on failure.
997 */
amdgpu_bo_unpin(struct amdgpu_bo * bo)998 void amdgpu_bo_unpin(struct amdgpu_bo *bo)
999 {
1000 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1001
1002 ttm_bo_unpin(&bo->tbo);
1003 if (bo->tbo.pin_count)
1004 return;
1005
1006 if (bo->tbo.base.import_attach)
1007 dma_buf_unpin(bo->tbo.base.import_attach);
1008
1009 if (bo->tbo.resource->mem_type == TTM_PL_VRAM) {
1010 atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
1011 atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
1012 &adev->visible_pin_size);
1013 } else if (bo->tbo.resource->mem_type == TTM_PL_TT) {
1014 atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size);
1015 }
1016 }
1017
1018 static const char *amdgpu_vram_names[] = {
1019 "UNKNOWN",
1020 "GDDR1",
1021 "DDR2",
1022 "GDDR3",
1023 "GDDR4",
1024 "GDDR5",
1025 "HBM",
1026 "DDR3",
1027 "DDR4",
1028 "GDDR6",
1029 "DDR5",
1030 "LPDDR4",
1031 "LPDDR5"
1032 };
1033
1034 /**
1035 * amdgpu_bo_init - initialize memory manager
1036 * @adev: amdgpu device object
1037 *
1038 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
1039 *
1040 * Returns:
1041 * 0 for success or a negative error code on failure.
1042 */
amdgpu_bo_init(struct amdgpu_device * adev)1043 int amdgpu_bo_init(struct amdgpu_device *adev)
1044 {
1045 /* On A+A platform, VRAM can be mapped as WB */
1046 if (!adev->gmc.xgmi.connected_to_cpu) {
1047 /* reserve PAT memory space to WC for VRAM */
1048 int r = arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1049 adev->gmc.aper_size);
1050
1051 if (r) {
1052 DRM_ERROR("Unable to set WC memtype for the aperture base\n");
1053 return r;
1054 }
1055
1056 /* Add an MTRR for the VRAM */
1057 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1058 adev->gmc.aper_size);
1059 }
1060
1061 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1062 adev->gmc.mc_vram_size >> 20,
1063 (unsigned long long)adev->gmc.aper_size >> 20);
1064 DRM_INFO("RAM width %dbits %s\n",
1065 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1066 return amdgpu_ttm_init(adev);
1067 }
1068
1069 /**
1070 * amdgpu_bo_fini - tear down memory manager
1071 * @adev: amdgpu device object
1072 *
1073 * Reverses amdgpu_bo_init() to tear down memory manager.
1074 */
amdgpu_bo_fini(struct amdgpu_device * adev)1075 void amdgpu_bo_fini(struct amdgpu_device *adev)
1076 {
1077 int idx;
1078
1079 amdgpu_ttm_fini(adev);
1080
1081 if (drm_dev_enter(adev_to_drm(adev), &idx)) {
1082
1083 if (!adev->gmc.xgmi.connected_to_cpu) {
1084 arch_phys_wc_del(adev->gmc.vram_mtrr);
1085 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1086 }
1087 drm_dev_exit(idx);
1088 }
1089 }
1090
1091 /**
1092 * amdgpu_bo_set_tiling_flags - set tiling flags
1093 * @bo: &amdgpu_bo buffer object
1094 * @tiling_flags: new flags
1095 *
1096 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1097 * kernel driver to set the tiling flags on a buffer.
1098 *
1099 * Returns:
1100 * 0 for success or a negative error code on failure.
1101 */
amdgpu_bo_set_tiling_flags(struct amdgpu_bo * bo,u64 tiling_flags)1102 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1103 {
1104 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1105 struct amdgpu_bo_user *ubo;
1106
1107 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1108 if (adev->family <= AMDGPU_FAMILY_CZ &&
1109 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1110 return -EINVAL;
1111
1112 ubo = to_amdgpu_bo_user(bo);
1113 ubo->tiling_flags = tiling_flags;
1114 return 0;
1115 }
1116
1117 /**
1118 * amdgpu_bo_get_tiling_flags - get tiling flags
1119 * @bo: &amdgpu_bo buffer object
1120 * @tiling_flags: returned flags
1121 *
1122 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1123 * set the tiling flags on a buffer.
1124 */
amdgpu_bo_get_tiling_flags(struct amdgpu_bo * bo,u64 * tiling_flags)1125 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1126 {
1127 struct amdgpu_bo_user *ubo;
1128
1129 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1130 dma_resv_assert_held(bo->tbo.base.resv);
1131 ubo = to_amdgpu_bo_user(bo);
1132
1133 if (tiling_flags)
1134 *tiling_flags = ubo->tiling_flags;
1135 }
1136
1137 /**
1138 * amdgpu_bo_set_metadata - set metadata
1139 * @bo: &amdgpu_bo buffer object
1140 * @metadata: new metadata
1141 * @metadata_size: size of the new metadata
1142 * @flags: flags of the new metadata
1143 *
1144 * Sets buffer object's metadata, its size and flags.
1145 * Used via GEM ioctl.
1146 *
1147 * Returns:
1148 * 0 for success or a negative error code on failure.
1149 */
amdgpu_bo_set_metadata(struct amdgpu_bo * bo,void * metadata,uint32_t metadata_size,uint64_t flags)1150 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
1151 uint32_t metadata_size, uint64_t flags)
1152 {
1153 struct amdgpu_bo_user *ubo;
1154 void *buffer;
1155
1156 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1157 ubo = to_amdgpu_bo_user(bo);
1158 if (!metadata_size) {
1159 if (ubo->metadata_size) {
1160 kfree(ubo->metadata);
1161 ubo->metadata = NULL;
1162 ubo->metadata_size = 0;
1163 }
1164 return 0;
1165 }
1166
1167 if (metadata == NULL)
1168 return -EINVAL;
1169
1170 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1171 if (buffer == NULL)
1172 return -ENOMEM;
1173
1174 kfree(ubo->metadata);
1175 ubo->metadata_flags = flags;
1176 ubo->metadata = buffer;
1177 ubo->metadata_size = metadata_size;
1178
1179 return 0;
1180 }
1181
1182 /**
1183 * amdgpu_bo_get_metadata - get metadata
1184 * @bo: &amdgpu_bo buffer object
1185 * @buffer: returned metadata
1186 * @buffer_size: size of the buffer
1187 * @metadata_size: size of the returned metadata
1188 * @flags: flags of the returned metadata
1189 *
1190 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1191 * less than metadata_size.
1192 * Used via GEM ioctl.
1193 *
1194 * Returns:
1195 * 0 for success or a negative error code on failure.
1196 */
amdgpu_bo_get_metadata(struct amdgpu_bo * bo,void * buffer,size_t buffer_size,uint32_t * metadata_size,uint64_t * flags)1197 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1198 size_t buffer_size, uint32_t *metadata_size,
1199 uint64_t *flags)
1200 {
1201 struct amdgpu_bo_user *ubo;
1202
1203 if (!buffer && !metadata_size)
1204 return -EINVAL;
1205
1206 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1207 ubo = to_amdgpu_bo_user(bo);
1208 if (metadata_size)
1209 *metadata_size = ubo->metadata_size;
1210
1211 if (buffer) {
1212 if (buffer_size < ubo->metadata_size)
1213 return -EINVAL;
1214
1215 if (ubo->metadata_size)
1216 memcpy(buffer, ubo->metadata, ubo->metadata_size);
1217 }
1218
1219 if (flags)
1220 *flags = ubo->metadata_flags;
1221
1222 return 0;
1223 }
1224
1225 /**
1226 * amdgpu_bo_move_notify - notification about a memory move
1227 * @bo: pointer to a buffer object
1228 * @evict: if this move is evicting the buffer from the graphics address space
1229 * @new_mem: new information of the bufer object
1230 *
1231 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1232 * bookkeeping.
1233 * TTM driver callback which is called when ttm moves a buffer.
1234 */
amdgpu_bo_move_notify(struct ttm_buffer_object * bo,bool evict,struct ttm_resource * new_mem)1235 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1236 bool evict,
1237 struct ttm_resource *new_mem)
1238 {
1239 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1240 struct amdgpu_bo *abo;
1241 struct ttm_resource *old_mem = bo->resource;
1242
1243 if (!amdgpu_bo_is_amdgpu_bo(bo))
1244 return;
1245
1246 abo = ttm_to_amdgpu_bo(bo);
1247 amdgpu_vm_bo_invalidate(adev, abo, evict);
1248
1249 amdgpu_bo_kunmap(abo);
1250
1251 if (abo->tbo.base.dma_buf && !abo->tbo.base.import_attach &&
1252 bo->resource->mem_type != TTM_PL_SYSTEM)
1253 dma_buf_move_notify(abo->tbo.base.dma_buf);
1254
1255 /* remember the eviction */
1256 if (evict)
1257 atomic64_inc(&adev->num_evictions);
1258
1259 /* update statistics */
1260 if (!new_mem)
1261 return;
1262
1263 /* move_notify is called before move happens */
1264 trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
1265 }
1266
amdgpu_bo_get_memory(struct amdgpu_bo * bo,uint64_t * vram_mem,uint64_t * gtt_mem,uint64_t * cpu_mem)1267 void amdgpu_bo_get_memory(struct amdgpu_bo *bo, uint64_t *vram_mem,
1268 uint64_t *gtt_mem, uint64_t *cpu_mem)
1269 {
1270 unsigned int domain;
1271
1272 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type);
1273 switch (domain) {
1274 case AMDGPU_GEM_DOMAIN_VRAM:
1275 *vram_mem += amdgpu_bo_size(bo);
1276 break;
1277 case AMDGPU_GEM_DOMAIN_GTT:
1278 *gtt_mem += amdgpu_bo_size(bo);
1279 break;
1280 case AMDGPU_GEM_DOMAIN_CPU:
1281 default:
1282 *cpu_mem += amdgpu_bo_size(bo);
1283 break;
1284 }
1285 }
1286
1287 /**
1288 * amdgpu_bo_release_notify - notification about a BO being released
1289 * @bo: pointer to a buffer object
1290 *
1291 * Wipes VRAM buffers whose contents should not be leaked before the
1292 * memory is released.
1293 */
amdgpu_bo_release_notify(struct ttm_buffer_object * bo)1294 void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)
1295 {
1296 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1297 struct dma_fence *fence = NULL;
1298 struct amdgpu_bo *abo;
1299 int r;
1300
1301 if (!amdgpu_bo_is_amdgpu_bo(bo))
1302 return;
1303
1304 abo = ttm_to_amdgpu_bo(bo);
1305
1306 if (abo->kfd_bo)
1307 amdgpu_amdkfd_release_notify(abo);
1308
1309 /* We only remove the fence if the resv has individualized. */
1310 WARN_ON_ONCE(bo->type == ttm_bo_type_kernel
1311 && bo->base.resv != &bo->base._resv);
1312 if (bo->base.resv == &bo->base._resv)
1313 amdgpu_amdkfd_remove_fence_on_pt_pd_bos(abo);
1314
1315 if (!bo->resource || bo->resource->mem_type != TTM_PL_VRAM ||
1316 !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE) ||
1317 adev->in_suspend || adev->shutdown)
1318 return;
1319
1320 if (WARN_ON_ONCE(!dma_resv_trylock(bo->base.resv)))
1321 return;
1322
1323 r = amdgpu_fill_buffer(abo, AMDGPU_POISON, bo->base.resv, &fence);
1324 if (!WARN_ON(r)) {
1325 amdgpu_bo_fence(abo, fence, false);
1326 dma_fence_put(fence);
1327 }
1328
1329 dma_resv_unlock(bo->base.resv);
1330 }
1331
1332 /**
1333 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1334 * @bo: pointer to a buffer object
1335 *
1336 * Notifies the driver we are taking a fault on this BO and have reserved it,
1337 * also performs bookkeeping.
1338 * TTM driver callback for dealing with vm faults.
1339 *
1340 * Returns:
1341 * 0 for success or a negative error code on failure.
1342 */
amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object * bo)1343 vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1344 {
1345 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1346 struct ttm_operation_ctx ctx = { false, false };
1347 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1348 unsigned long offset;
1349 int r;
1350
1351 /* Remember that this BO was accessed by the CPU */
1352 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1353
1354 if (bo->resource->mem_type != TTM_PL_VRAM)
1355 return 0;
1356
1357 offset = bo->resource->start << PAGE_SHIFT;
1358 if ((offset + bo->base.size) <= adev->gmc.visible_vram_size)
1359 return 0;
1360
1361 /* Can't move a pinned BO to visible VRAM */
1362 if (abo->tbo.pin_count > 0)
1363 return VM_FAULT_SIGBUS;
1364
1365 /* hurrah the memory is not visible ! */
1366 atomic64_inc(&adev->num_vram_cpu_page_faults);
1367 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1368 AMDGPU_GEM_DOMAIN_GTT);
1369
1370 /* Avoid costly evictions; only set GTT as a busy placement */
1371 abo->placement.num_busy_placement = 1;
1372 abo->placement.busy_placement = &abo->placements[1];
1373
1374 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1375 if (unlikely(r == -EBUSY || r == -ERESTARTSYS))
1376 return VM_FAULT_NOPAGE;
1377 else if (unlikely(r))
1378 return VM_FAULT_SIGBUS;
1379
1380 offset = bo->resource->start << PAGE_SHIFT;
1381 /* this should never happen */
1382 if (bo->resource->mem_type == TTM_PL_VRAM &&
1383 (offset + bo->base.size) > adev->gmc.visible_vram_size)
1384 return VM_FAULT_SIGBUS;
1385
1386 ttm_bo_move_to_lru_tail_unlocked(bo);
1387 return 0;
1388 }
1389
1390 /**
1391 * amdgpu_bo_fence - add fence to buffer object
1392 *
1393 * @bo: buffer object in question
1394 * @fence: fence to add
1395 * @shared: true if fence should be added shared
1396 *
1397 */
amdgpu_bo_fence(struct amdgpu_bo * bo,struct dma_fence * fence,bool shared)1398 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1399 bool shared)
1400 {
1401 struct dma_resv *resv = bo->tbo.base.resv;
1402 int r;
1403
1404 r = dma_resv_reserve_fences(resv, 1);
1405 if (r) {
1406 /* As last resort on OOM we block for the fence */
1407 dma_fence_wait(fence, false);
1408 return;
1409 }
1410
1411 dma_resv_add_fence(resv, fence, shared ? DMA_RESV_USAGE_READ :
1412 DMA_RESV_USAGE_WRITE);
1413 }
1414
1415 /**
1416 * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences
1417 *
1418 * @adev: amdgpu device pointer
1419 * @resv: reservation object to sync to
1420 * @sync_mode: synchronization mode
1421 * @owner: fence owner
1422 * @intr: Whether the wait is interruptible
1423 *
1424 * Extract the fences from the reservation object and waits for them to finish.
1425 *
1426 * Returns:
1427 * 0 on success, errno otherwise.
1428 */
amdgpu_bo_sync_wait_resv(struct amdgpu_device * adev,struct dma_resv * resv,enum amdgpu_sync_mode sync_mode,void * owner,bool intr)1429 int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv,
1430 enum amdgpu_sync_mode sync_mode, void *owner,
1431 bool intr)
1432 {
1433 struct amdgpu_sync sync;
1434 int r;
1435
1436 amdgpu_sync_create(&sync);
1437 amdgpu_sync_resv(adev, &sync, resv, sync_mode, owner);
1438 r = amdgpu_sync_wait(&sync, intr);
1439 amdgpu_sync_free(&sync);
1440 return r;
1441 }
1442
1443 /**
1444 * amdgpu_bo_sync_wait - Wrapper for amdgpu_bo_sync_wait_resv
1445 * @bo: buffer object to wait for
1446 * @owner: fence owner
1447 * @intr: Whether the wait is interruptible
1448 *
1449 * Wrapper to wait for fences in a BO.
1450 * Returns:
1451 * 0 on success, errno otherwise.
1452 */
amdgpu_bo_sync_wait(struct amdgpu_bo * bo,void * owner,bool intr)1453 int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)
1454 {
1455 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1456
1457 return amdgpu_bo_sync_wait_resv(adev, bo->tbo.base.resv,
1458 AMDGPU_SYNC_NE_OWNER, owner, intr);
1459 }
1460
1461 /**
1462 * amdgpu_bo_gpu_offset - return GPU offset of bo
1463 * @bo: amdgpu object for which we query the offset
1464 *
1465 * Note: object should either be pinned or reserved when calling this
1466 * function, it might be useful to add check for this for debugging.
1467 *
1468 * Returns:
1469 * current GPU offset of the object.
1470 */
amdgpu_bo_gpu_offset(struct amdgpu_bo * bo)1471 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1472 {
1473 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_SYSTEM);
1474 WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) &&
1475 !bo->tbo.pin_count && bo->tbo.type != ttm_bo_type_kernel);
1476 WARN_ON_ONCE(bo->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET);
1477 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_VRAM &&
1478 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1479
1480 return amdgpu_bo_gpu_offset_no_check(bo);
1481 }
1482
1483 /**
1484 * amdgpu_bo_gpu_offset_no_check - return GPU offset of bo
1485 * @bo: amdgpu object for which we query the offset
1486 *
1487 * Returns:
1488 * current GPU offset of the object without raising warnings.
1489 */
amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo * bo)1490 u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo)
1491 {
1492 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1493 uint64_t offset;
1494
1495 offset = (bo->tbo.resource->start << PAGE_SHIFT) +
1496 amdgpu_ttm_domain_start(adev, bo->tbo.resource->mem_type);
1497
1498 return amdgpu_gmc_sign_extend(offset);
1499 }
1500
1501 /**
1502 * amdgpu_bo_get_preferred_domain - get preferred domain
1503 * @adev: amdgpu device object
1504 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1505 *
1506 * Returns:
1507 * Which of the allowed domains is preferred for allocating the BO.
1508 */
amdgpu_bo_get_preferred_domain(struct amdgpu_device * adev,uint32_t domain)1509 uint32_t amdgpu_bo_get_preferred_domain(struct amdgpu_device *adev,
1510 uint32_t domain)
1511 {
1512 if (domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) {
1513 domain = AMDGPU_GEM_DOMAIN_VRAM;
1514 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1515 domain = AMDGPU_GEM_DOMAIN_GTT;
1516 }
1517 return domain;
1518 }
1519
1520 #if defined(CONFIG_DEBUG_FS)
1521 #define amdgpu_bo_print_flag(m, bo, flag) \
1522 do { \
1523 if (bo->flags & (AMDGPU_GEM_CREATE_ ## flag)) { \
1524 seq_printf((m), " " #flag); \
1525 } \
1526 } while (0)
1527
1528 /**
1529 * amdgpu_bo_print_info - print BO info in debugfs file
1530 *
1531 * @id: Index or Id of the BO
1532 * @bo: Requested BO for printing info
1533 * @m: debugfs file
1534 *
1535 * Print BO information in debugfs file
1536 *
1537 * Returns:
1538 * Size of the BO in bytes.
1539 */
amdgpu_bo_print_info(int id,struct amdgpu_bo * bo,struct seq_file * m)1540 u64 amdgpu_bo_print_info(int id, struct amdgpu_bo *bo, struct seq_file *m)
1541 {
1542 struct dma_buf_attachment *attachment;
1543 struct dma_buf *dma_buf;
1544 unsigned int domain;
1545 const char *placement;
1546 unsigned int pin_count;
1547 u64 size;
1548
1549 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type);
1550 switch (domain) {
1551 case AMDGPU_GEM_DOMAIN_VRAM:
1552 placement = "VRAM";
1553 break;
1554 case AMDGPU_GEM_DOMAIN_GTT:
1555 placement = " GTT";
1556 break;
1557 case AMDGPU_GEM_DOMAIN_CPU:
1558 default:
1559 placement = " CPU";
1560 break;
1561 }
1562
1563 size = amdgpu_bo_size(bo);
1564 seq_printf(m, "\t\t0x%08x: %12lld byte %s",
1565 id, size, placement);
1566
1567 pin_count = READ_ONCE(bo->tbo.pin_count);
1568 if (pin_count)
1569 seq_printf(m, " pin count %d", pin_count);
1570
1571 dma_buf = READ_ONCE(bo->tbo.base.dma_buf);
1572 attachment = READ_ONCE(bo->tbo.base.import_attach);
1573
1574 if (attachment)
1575 seq_printf(m, " imported from %p", dma_buf);
1576 else if (dma_buf)
1577 seq_printf(m, " exported as %p", dma_buf);
1578
1579 amdgpu_bo_print_flag(m, bo, CPU_ACCESS_REQUIRED);
1580 amdgpu_bo_print_flag(m, bo, NO_CPU_ACCESS);
1581 amdgpu_bo_print_flag(m, bo, CPU_GTT_USWC);
1582 amdgpu_bo_print_flag(m, bo, VRAM_CLEARED);
1583 amdgpu_bo_print_flag(m, bo, VRAM_CONTIGUOUS);
1584 amdgpu_bo_print_flag(m, bo, VM_ALWAYS_VALID);
1585 amdgpu_bo_print_flag(m, bo, EXPLICIT_SYNC);
1586
1587 seq_puts(m, "\n");
1588
1589 return size;
1590 }
1591 #endif
1592