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 (man && size < man->size)
455 return true;
456 else if (!man)
457 WARN_ON_ONCE("GTT domain requested but GTT mem manager uninitialized");
458 goto fail;
459 } else if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
460 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
461
462 if (man && size < man->size)
463 return true;
464 goto fail;
465 }
466
467 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */
468 return true;
469
470 fail:
471 if (man)
472 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size,
473 man->size);
474 return false;
475 }
476
amdgpu_bo_support_uswc(u64 bo_flags)477 bool amdgpu_bo_support_uswc(u64 bo_flags)
478 {
479
480 #ifdef CONFIG_X86_32
481 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
482 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
483 */
484 return false;
485 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
486 /* Don't try to enable write-combining when it can't work, or things
487 * may be slow
488 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
489 */
490
491 #ifndef CONFIG_COMPILE_TEST
492 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
493 thanks to write-combining
494 #endif
495
496 if (bo_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
497 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
498 "better performance thanks to write-combining\n");
499 return false;
500 #else
501 /* For architectures that don't support WC memory,
502 * mask out the WC flag from the BO
503 */
504 if (!drm_arch_can_wc_memory())
505 return false;
506
507 return true;
508 #endif
509 }
510
511 /**
512 * amdgpu_bo_create - create an &amdgpu_bo buffer object
513 * @adev: amdgpu device object
514 * @bp: parameters to be used for the buffer object
515 * @bo_ptr: pointer to the buffer object pointer
516 *
517 * Creates an &amdgpu_bo buffer object.
518 *
519 * Returns:
520 * 0 for success or a negative error code on failure.
521 */
amdgpu_bo_create(struct amdgpu_device * adev,struct amdgpu_bo_param * bp,struct amdgpu_bo ** bo_ptr)522 int amdgpu_bo_create(struct amdgpu_device *adev,
523 struct amdgpu_bo_param *bp,
524 struct amdgpu_bo **bo_ptr)
525 {
526 struct ttm_operation_ctx ctx = {
527 .interruptible = (bp->type != ttm_bo_type_kernel),
528 .no_wait_gpu = bp->no_wait_gpu,
529 /* We opt to avoid OOM on system pages allocations */
530 .gfp_retry_mayfail = true,
531 .allow_res_evict = bp->type != ttm_bo_type_kernel,
532 .resv = bp->resv
533 };
534 struct amdgpu_bo *bo;
535 unsigned long page_align, size = bp->size;
536 int r;
537
538 /* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
539 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
540 /* GWS and OA don't need any alignment. */
541 page_align = bp->byte_align;
542 size <<= PAGE_SHIFT;
543 } else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
544 /* Both size and alignment must be a multiple of 4. */
545 page_align = ALIGN(bp->byte_align, 4);
546 size = ALIGN(size, 4) << PAGE_SHIFT;
547 } else {
548 /* Memory should be aligned at least to a page size. */
549 page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
550 size = ALIGN(size, PAGE_SIZE);
551 }
552
553 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
554 return -ENOMEM;
555
556 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo));
557
558 *bo_ptr = NULL;
559 bo = kvzalloc(bp->bo_ptr_size, GFP_KERNEL);
560 if (bo == NULL)
561 return -ENOMEM;
562 drm_gem_private_object_init(adev_to_drm(adev), &bo->tbo.base, size);
563 bo->vm_bo = NULL;
564 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
565 bp->domain;
566 bo->allowed_domains = bo->preferred_domains;
567 if (bp->type != ttm_bo_type_kernel &&
568 !(bp->flags & AMDGPU_GEM_CREATE_DISCARDABLE) &&
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 if (adev->ras_enabled)
578 bo->flags |= AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE;
579
580 bo->tbo.bdev = &adev->mman.bdev;
581 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
582 AMDGPU_GEM_DOMAIN_GDS))
583 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
584 else
585 amdgpu_bo_placement_from_domain(bo, bp->domain);
586 if (bp->type == ttm_bo_type_kernel)
587 bo->tbo.priority = 1;
588
589 if (!bp->destroy)
590 bp->destroy = &amdgpu_bo_destroy;
591
592 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, bp->type,
593 &bo->placement, page_align, &ctx, NULL,
594 bp->resv, bp->destroy);
595 if (unlikely(r != 0))
596 return r;
597
598 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
599 bo->tbo.resource->mem_type == TTM_PL_VRAM &&
600 bo->tbo.resource->start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
601 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
602 ctx.bytes_moved);
603 else
604 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
605
606 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
607 bo->tbo.resource->mem_type == TTM_PL_VRAM) {
608 struct dma_fence *fence;
609
610 r = amdgpu_fill_buffer(bo, 0, bo->tbo.base.resv, &fence);
611 if (unlikely(r))
612 goto fail_unreserve;
613
614 dma_resv_add_fence(bo->tbo.base.resv, fence,
615 DMA_RESV_USAGE_KERNEL);
616 dma_fence_put(fence);
617 }
618 if (!bp->resv)
619 amdgpu_bo_unreserve(bo);
620 *bo_ptr = bo;
621
622 trace_amdgpu_bo_create(bo);
623
624 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
625 if (bp->type == ttm_bo_type_device)
626 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
627
628 return 0;
629
630 fail_unreserve:
631 if (!bp->resv)
632 dma_resv_unlock(bo->tbo.base.resv);
633 amdgpu_bo_unref(&bo);
634 return r;
635 }
636
637 /**
638 * amdgpu_bo_create_user - create an &amdgpu_bo_user buffer object
639 * @adev: amdgpu device object
640 * @bp: parameters to be used for the buffer object
641 * @ubo_ptr: pointer to the buffer object pointer
642 *
643 * Create a BO to be used by user application;
644 *
645 * Returns:
646 * 0 for success or a negative error code on failure.
647 */
648
amdgpu_bo_create_user(struct amdgpu_device * adev,struct amdgpu_bo_param * bp,struct amdgpu_bo_user ** ubo_ptr)649 int amdgpu_bo_create_user(struct amdgpu_device *adev,
650 struct amdgpu_bo_param *bp,
651 struct amdgpu_bo_user **ubo_ptr)
652 {
653 struct amdgpu_bo *bo_ptr;
654 int r;
655
656 bp->bo_ptr_size = sizeof(struct amdgpu_bo_user);
657 bp->destroy = &amdgpu_bo_user_destroy;
658 r = amdgpu_bo_create(adev, bp, &bo_ptr);
659 if (r)
660 return r;
661
662 *ubo_ptr = to_amdgpu_bo_user(bo_ptr);
663 return r;
664 }
665
666 /**
667 * amdgpu_bo_create_vm - create an &amdgpu_bo_vm buffer object
668 * @adev: amdgpu device object
669 * @bp: parameters to be used for the buffer object
670 * @vmbo_ptr: pointer to the buffer object pointer
671 *
672 * Create a BO to be for GPUVM.
673 *
674 * Returns:
675 * 0 for success or a negative error code on failure.
676 */
677
amdgpu_bo_create_vm(struct amdgpu_device * adev,struct amdgpu_bo_param * bp,struct amdgpu_bo_vm ** vmbo_ptr)678 int amdgpu_bo_create_vm(struct amdgpu_device *adev,
679 struct amdgpu_bo_param *bp,
680 struct amdgpu_bo_vm **vmbo_ptr)
681 {
682 struct amdgpu_bo *bo_ptr;
683 int r;
684
685 /* bo_ptr_size will be determined by the caller and it depends on
686 * num of amdgpu_vm_pt entries.
687 */
688 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo_vm));
689 r = amdgpu_bo_create(adev, bp, &bo_ptr);
690 if (r)
691 return r;
692
693 *vmbo_ptr = to_amdgpu_bo_vm(bo_ptr);
694 INIT_LIST_HEAD(&(*vmbo_ptr)->shadow_list);
695 /* Set destroy callback to amdgpu_bo_vm_destroy after vmbo->shadow_list
696 * is initialized.
697 */
698 bo_ptr->tbo.destroy = &amdgpu_bo_vm_destroy;
699 return r;
700 }
701
702 /**
703 * amdgpu_bo_add_to_shadow_list - add a BO to the shadow list
704 *
705 * @vmbo: BO that will be inserted into the shadow list
706 *
707 * Insert a BO to the shadow list.
708 */
amdgpu_bo_add_to_shadow_list(struct amdgpu_bo_vm * vmbo)709 void amdgpu_bo_add_to_shadow_list(struct amdgpu_bo_vm *vmbo)
710 {
711 struct amdgpu_device *adev = amdgpu_ttm_adev(vmbo->bo.tbo.bdev);
712
713 mutex_lock(&adev->shadow_list_lock);
714 list_add_tail(&vmbo->shadow_list, &adev->shadow_list);
715 mutex_unlock(&adev->shadow_list_lock);
716 }
717
718 /**
719 * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow
720 *
721 * @shadow: &amdgpu_bo shadow to be restored
722 * @fence: dma_fence associated with the operation
723 *
724 * Copies a buffer object's shadow content back to the object.
725 * This is used for recovering a buffer from its shadow in case of a gpu
726 * reset where vram context may be lost.
727 *
728 * Returns:
729 * 0 for success or a negative error code on failure.
730 */
amdgpu_bo_restore_shadow(struct amdgpu_bo * shadow,struct dma_fence ** fence)731 int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence)
732
733 {
734 struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev);
735 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
736 uint64_t shadow_addr, parent_addr;
737
738 shadow_addr = amdgpu_bo_gpu_offset(shadow);
739 parent_addr = amdgpu_bo_gpu_offset(shadow->parent);
740
741 return amdgpu_copy_buffer(ring, shadow_addr, parent_addr,
742 amdgpu_bo_size(shadow), NULL, fence,
743 true, false, false);
744 }
745
746 /**
747 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
748 * @bo: &amdgpu_bo buffer object to be mapped
749 * @ptr: kernel virtual address to be returned
750 *
751 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
752 * amdgpu_bo_kptr() to get the kernel virtual address.
753 *
754 * Returns:
755 * 0 for success or a negative error code on failure.
756 */
amdgpu_bo_kmap(struct amdgpu_bo * bo,void ** ptr)757 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
758 {
759 void *kptr;
760 long r;
761
762 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
763 return -EPERM;
764
765 r = dma_resv_wait_timeout(bo->tbo.base.resv, DMA_RESV_USAGE_KERNEL,
766 false, MAX_SCHEDULE_TIMEOUT);
767 if (r < 0)
768 return r;
769
770 kptr = amdgpu_bo_kptr(bo);
771 if (kptr) {
772 if (ptr)
773 *ptr = kptr;
774 return 0;
775 }
776
777 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.resource->num_pages, &bo->kmap);
778 if (r)
779 return r;
780
781 if (ptr)
782 *ptr = amdgpu_bo_kptr(bo);
783
784 return 0;
785 }
786
787 /**
788 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
789 * @bo: &amdgpu_bo buffer object
790 *
791 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
792 *
793 * Returns:
794 * the virtual address of a buffer object area.
795 */
amdgpu_bo_kptr(struct amdgpu_bo * bo)796 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
797 {
798 bool is_iomem;
799
800 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
801 }
802
803 /**
804 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
805 * @bo: &amdgpu_bo buffer object to be unmapped
806 *
807 * Unmaps a kernel map set up by amdgpu_bo_kmap().
808 */
amdgpu_bo_kunmap(struct amdgpu_bo * bo)809 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
810 {
811 if (bo->kmap.bo)
812 ttm_bo_kunmap(&bo->kmap);
813 }
814
815 /**
816 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
817 * @bo: &amdgpu_bo buffer object
818 *
819 * References the contained &ttm_buffer_object.
820 *
821 * Returns:
822 * a refcounted pointer to the &amdgpu_bo buffer object.
823 */
amdgpu_bo_ref(struct amdgpu_bo * bo)824 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
825 {
826 if (bo == NULL)
827 return NULL;
828
829 ttm_bo_get(&bo->tbo);
830 return bo;
831 }
832
833 /**
834 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
835 * @bo: &amdgpu_bo buffer object
836 *
837 * Unreferences the contained &ttm_buffer_object and clear the pointer
838 */
amdgpu_bo_unref(struct amdgpu_bo ** bo)839 void amdgpu_bo_unref(struct amdgpu_bo **bo)
840 {
841 struct ttm_buffer_object *tbo;
842
843 if ((*bo) == NULL)
844 return;
845
846 tbo = &((*bo)->tbo);
847 ttm_bo_put(tbo);
848 *bo = NULL;
849 }
850
851 /**
852 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object
853 * @bo: &amdgpu_bo buffer object to be pinned
854 * @domain: domain to be pinned to
855 * @min_offset: the start of requested address range
856 * @max_offset: the end of requested address range
857 *
858 * Pins the buffer object according to requested domain and address range. If
859 * the memory is unbound gart memory, binds the pages into gart table. Adjusts
860 * pin_count and pin_size accordingly.
861 *
862 * Pinning means to lock pages in memory along with keeping them at a fixed
863 * offset. It is required when a buffer can not be moved, for example, when
864 * a display buffer is being scanned out.
865 *
866 * Compared with amdgpu_bo_pin(), this function gives more flexibility on
867 * where to pin a buffer if there are specific restrictions on where a buffer
868 * must be located.
869 *
870 * Returns:
871 * 0 for success or a negative error code on failure.
872 */
amdgpu_bo_pin_restricted(struct amdgpu_bo * bo,u32 domain,u64 min_offset,u64 max_offset)873 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
874 u64 min_offset, u64 max_offset)
875 {
876 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
877 struct ttm_operation_ctx ctx = { false, false };
878 int r, i;
879
880 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
881 return -EPERM;
882
883 if (WARN_ON_ONCE(min_offset > max_offset))
884 return -EINVAL;
885
886 /* Check domain to be pinned to against preferred domains */
887 if (bo->preferred_domains & domain)
888 domain = bo->preferred_domains & domain;
889
890 /* A shared bo cannot be migrated to VRAM */
891 if (bo->tbo.base.import_attach) {
892 if (domain & AMDGPU_GEM_DOMAIN_GTT)
893 domain = AMDGPU_GEM_DOMAIN_GTT;
894 else
895 return -EINVAL;
896 }
897
898 if (bo->tbo.pin_count) {
899 uint32_t mem_type = bo->tbo.resource->mem_type;
900 uint32_t mem_flags = bo->tbo.resource->placement;
901
902 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
903 return -EINVAL;
904
905 if ((mem_type == TTM_PL_VRAM) &&
906 (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) &&
907 !(mem_flags & TTM_PL_FLAG_CONTIGUOUS))
908 return -EINVAL;
909
910 ttm_bo_pin(&bo->tbo);
911
912 if (max_offset != 0) {
913 u64 domain_start = amdgpu_ttm_domain_start(adev,
914 mem_type);
915 WARN_ON_ONCE(max_offset <
916 (amdgpu_bo_gpu_offset(bo) - domain_start));
917 }
918
919 return 0;
920 }
921
922 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
923 * See function amdgpu_display_supported_domains()
924 */
925 domain = amdgpu_bo_get_preferred_domain(adev, domain);
926
927 if (bo->tbo.base.import_attach)
928 dma_buf_pin(bo->tbo.base.import_attach);
929
930 /* force to pin into visible video ram */
931 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
932 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
933 amdgpu_bo_placement_from_domain(bo, domain);
934 for (i = 0; i < bo->placement.num_placement; i++) {
935 unsigned fpfn, lpfn;
936
937 fpfn = min_offset >> PAGE_SHIFT;
938 lpfn = max_offset >> PAGE_SHIFT;
939
940 if (fpfn > bo->placements[i].fpfn)
941 bo->placements[i].fpfn = fpfn;
942 if (!bo->placements[i].lpfn ||
943 (lpfn && lpfn < bo->placements[i].lpfn))
944 bo->placements[i].lpfn = lpfn;
945 }
946
947 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
948 if (unlikely(r)) {
949 dev_err(adev->dev, "%p pin failed\n", bo);
950 goto error;
951 }
952
953 ttm_bo_pin(&bo->tbo);
954
955 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type);
956 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
957 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
958 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
959 &adev->visible_pin_size);
960 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
961 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
962 }
963
964 error:
965 return r;
966 }
967
968 /**
969 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
970 * @bo: &amdgpu_bo buffer object to be pinned
971 * @domain: domain to be pinned to
972 *
973 * A simple wrapper to amdgpu_bo_pin_restricted().
974 * Provides a simpler API for buffers that do not have any strict restrictions
975 * on where a buffer must be located.
976 *
977 * Returns:
978 * 0 for success or a negative error code on failure.
979 */
amdgpu_bo_pin(struct amdgpu_bo * bo,u32 domain)980 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
981 {
982 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
983 return amdgpu_bo_pin_restricted(bo, domain, 0, 0);
984 }
985
986 /**
987 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
988 * @bo: &amdgpu_bo buffer object to be unpinned
989 *
990 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
991 * Changes placement and pin size accordingly.
992 *
993 * Returns:
994 * 0 for success or a negative error code on failure.
995 */
amdgpu_bo_unpin(struct amdgpu_bo * bo)996 void amdgpu_bo_unpin(struct amdgpu_bo *bo)
997 {
998 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
999
1000 ttm_bo_unpin(&bo->tbo);
1001 if (bo->tbo.pin_count)
1002 return;
1003
1004 if (bo->tbo.base.import_attach)
1005 dma_buf_unpin(bo->tbo.base.import_attach);
1006
1007 if (bo->tbo.resource->mem_type == TTM_PL_VRAM) {
1008 atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
1009 atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
1010 &adev->visible_pin_size);
1011 } else if (bo->tbo.resource->mem_type == TTM_PL_TT) {
1012 atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size);
1013 }
1014 }
1015
1016 static const char *amdgpu_vram_names[] = {
1017 "UNKNOWN",
1018 "GDDR1",
1019 "DDR2",
1020 "GDDR3",
1021 "GDDR4",
1022 "GDDR5",
1023 "HBM",
1024 "DDR3",
1025 "DDR4",
1026 "GDDR6",
1027 "DDR5",
1028 "LPDDR4",
1029 "LPDDR5"
1030 };
1031
1032 /**
1033 * amdgpu_bo_init - initialize memory manager
1034 * @adev: amdgpu device object
1035 *
1036 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
1037 *
1038 * Returns:
1039 * 0 for success or a negative error code on failure.
1040 */
amdgpu_bo_init(struct amdgpu_device * adev)1041 int amdgpu_bo_init(struct amdgpu_device *adev)
1042 {
1043 /* On A+A platform, VRAM can be mapped as WB */
1044 if (!adev->gmc.xgmi.connected_to_cpu) {
1045 /* reserve PAT memory space to WC for VRAM */
1046 int r = arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1047 adev->gmc.aper_size);
1048
1049 if (r) {
1050 DRM_ERROR("Unable to set WC memtype for the aperture base\n");
1051 return r;
1052 }
1053
1054 /* Add an MTRR for the VRAM */
1055 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1056 adev->gmc.aper_size);
1057 }
1058
1059 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1060 adev->gmc.mc_vram_size >> 20,
1061 (unsigned long long)adev->gmc.aper_size >> 20);
1062 DRM_INFO("RAM width %dbits %s\n",
1063 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1064 return amdgpu_ttm_init(adev);
1065 }
1066
1067 /**
1068 * amdgpu_bo_fini - tear down memory manager
1069 * @adev: amdgpu device object
1070 *
1071 * Reverses amdgpu_bo_init() to tear down memory manager.
1072 */
amdgpu_bo_fini(struct amdgpu_device * adev)1073 void amdgpu_bo_fini(struct amdgpu_device *adev)
1074 {
1075 int idx;
1076
1077 amdgpu_ttm_fini(adev);
1078
1079 if (drm_dev_enter(adev_to_drm(adev), &idx)) {
1080
1081 if (!adev->gmc.xgmi.connected_to_cpu) {
1082 arch_phys_wc_del(adev->gmc.vram_mtrr);
1083 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1084 }
1085 drm_dev_exit(idx);
1086 }
1087 }
1088
1089 /**
1090 * amdgpu_bo_set_tiling_flags - set tiling flags
1091 * @bo: &amdgpu_bo buffer object
1092 * @tiling_flags: new flags
1093 *
1094 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1095 * kernel driver to set the tiling flags on a buffer.
1096 *
1097 * Returns:
1098 * 0 for success or a negative error code on failure.
1099 */
amdgpu_bo_set_tiling_flags(struct amdgpu_bo * bo,u64 tiling_flags)1100 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1101 {
1102 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1103 struct amdgpu_bo_user *ubo;
1104
1105 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1106 if (adev->family <= AMDGPU_FAMILY_CZ &&
1107 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1108 return -EINVAL;
1109
1110 ubo = to_amdgpu_bo_user(bo);
1111 ubo->tiling_flags = tiling_flags;
1112 return 0;
1113 }
1114
1115 /**
1116 * amdgpu_bo_get_tiling_flags - get tiling flags
1117 * @bo: &amdgpu_bo buffer object
1118 * @tiling_flags: returned flags
1119 *
1120 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1121 * set the tiling flags on a buffer.
1122 */
amdgpu_bo_get_tiling_flags(struct amdgpu_bo * bo,u64 * tiling_flags)1123 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1124 {
1125 struct amdgpu_bo_user *ubo;
1126
1127 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1128 dma_resv_assert_held(bo->tbo.base.resv);
1129 ubo = to_amdgpu_bo_user(bo);
1130
1131 if (tiling_flags)
1132 *tiling_flags = ubo->tiling_flags;
1133 }
1134
1135 /**
1136 * amdgpu_bo_set_metadata - set metadata
1137 * @bo: &amdgpu_bo buffer object
1138 * @metadata: new metadata
1139 * @metadata_size: size of the new metadata
1140 * @flags: flags of the new metadata
1141 *
1142 * Sets buffer object's metadata, its size and flags.
1143 * Used via GEM ioctl.
1144 *
1145 * Returns:
1146 * 0 for success or a negative error code on failure.
1147 */
amdgpu_bo_set_metadata(struct amdgpu_bo * bo,void * metadata,uint32_t metadata_size,uint64_t flags)1148 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
1149 uint32_t metadata_size, uint64_t flags)
1150 {
1151 struct amdgpu_bo_user *ubo;
1152 void *buffer;
1153
1154 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1155 ubo = to_amdgpu_bo_user(bo);
1156 if (!metadata_size) {
1157 if (ubo->metadata_size) {
1158 kfree(ubo->metadata);
1159 ubo->metadata = NULL;
1160 ubo->metadata_size = 0;
1161 }
1162 return 0;
1163 }
1164
1165 if (metadata == NULL)
1166 return -EINVAL;
1167
1168 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1169 if (buffer == NULL)
1170 return -ENOMEM;
1171
1172 kfree(ubo->metadata);
1173 ubo->metadata_flags = flags;
1174 ubo->metadata = buffer;
1175 ubo->metadata_size = metadata_size;
1176
1177 return 0;
1178 }
1179
1180 /**
1181 * amdgpu_bo_get_metadata - get metadata
1182 * @bo: &amdgpu_bo buffer object
1183 * @buffer: returned metadata
1184 * @buffer_size: size of the buffer
1185 * @metadata_size: size of the returned metadata
1186 * @flags: flags of the returned metadata
1187 *
1188 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1189 * less than metadata_size.
1190 * Used via GEM ioctl.
1191 *
1192 * Returns:
1193 * 0 for success or a negative error code on failure.
1194 */
amdgpu_bo_get_metadata(struct amdgpu_bo * bo,void * buffer,size_t buffer_size,uint32_t * metadata_size,uint64_t * flags)1195 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1196 size_t buffer_size, uint32_t *metadata_size,
1197 uint64_t *flags)
1198 {
1199 struct amdgpu_bo_user *ubo;
1200
1201 if (!buffer && !metadata_size)
1202 return -EINVAL;
1203
1204 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1205 ubo = to_amdgpu_bo_user(bo);
1206 if (metadata_size)
1207 *metadata_size = ubo->metadata_size;
1208
1209 if (buffer) {
1210 if (buffer_size < ubo->metadata_size)
1211 return -EINVAL;
1212
1213 if (ubo->metadata_size)
1214 memcpy(buffer, ubo->metadata, ubo->metadata_size);
1215 }
1216
1217 if (flags)
1218 *flags = ubo->metadata_flags;
1219
1220 return 0;
1221 }
1222
1223 /**
1224 * amdgpu_bo_move_notify - notification about a memory move
1225 * @bo: pointer to a buffer object
1226 * @evict: if this move is evicting the buffer from the graphics address space
1227 * @new_mem: new information of the bufer object
1228 *
1229 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1230 * bookkeeping.
1231 * TTM driver callback which is called when ttm moves a buffer.
1232 */
amdgpu_bo_move_notify(struct ttm_buffer_object * bo,bool evict,struct ttm_resource * new_mem)1233 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1234 bool evict,
1235 struct ttm_resource *new_mem)
1236 {
1237 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1238 struct amdgpu_bo *abo;
1239 struct ttm_resource *old_mem = bo->resource;
1240
1241 if (!amdgpu_bo_is_amdgpu_bo(bo))
1242 return;
1243
1244 abo = ttm_to_amdgpu_bo(bo);
1245 amdgpu_vm_bo_invalidate(adev, abo, evict);
1246
1247 amdgpu_bo_kunmap(abo);
1248
1249 if (abo->tbo.base.dma_buf && !abo->tbo.base.import_attach &&
1250 bo->resource->mem_type != TTM_PL_SYSTEM)
1251 dma_buf_move_notify(abo->tbo.base.dma_buf);
1252
1253 /* remember the eviction */
1254 if (evict)
1255 atomic64_inc(&adev->num_evictions);
1256
1257 /* update statistics */
1258 if (!new_mem)
1259 return;
1260
1261 /* move_notify is called before move happens */
1262 trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
1263 }
1264
amdgpu_bo_get_memory(struct amdgpu_bo * bo,uint64_t * vram_mem,uint64_t * gtt_mem,uint64_t * cpu_mem)1265 void amdgpu_bo_get_memory(struct amdgpu_bo *bo, uint64_t *vram_mem,
1266 uint64_t *gtt_mem, uint64_t *cpu_mem)
1267 {
1268 unsigned int domain;
1269
1270 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type);
1271 switch (domain) {
1272 case AMDGPU_GEM_DOMAIN_VRAM:
1273 *vram_mem += amdgpu_bo_size(bo);
1274 break;
1275 case AMDGPU_GEM_DOMAIN_GTT:
1276 *gtt_mem += amdgpu_bo_size(bo);
1277 break;
1278 case AMDGPU_GEM_DOMAIN_CPU:
1279 default:
1280 *cpu_mem += amdgpu_bo_size(bo);
1281 break;
1282 }
1283 }
1284
1285 /**
1286 * amdgpu_bo_release_notify - notification about a BO being released
1287 * @bo: pointer to a buffer object
1288 *
1289 * Wipes VRAM buffers whose contents should not be leaked before the
1290 * memory is released.
1291 */
amdgpu_bo_release_notify(struct ttm_buffer_object * bo)1292 void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)
1293 {
1294 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1295 struct dma_fence *fence = NULL;
1296 struct amdgpu_bo *abo;
1297 int r;
1298
1299 if (!amdgpu_bo_is_amdgpu_bo(bo))
1300 return;
1301
1302 abo = ttm_to_amdgpu_bo(bo);
1303
1304 if (abo->kfd_bo)
1305 amdgpu_amdkfd_release_notify(abo);
1306
1307 /* We only remove the fence if the resv has individualized. */
1308 WARN_ON_ONCE(bo->type == ttm_bo_type_kernel
1309 && bo->base.resv != &bo->base._resv);
1310 if (bo->base.resv == &bo->base._resv)
1311 amdgpu_amdkfd_remove_fence_on_pt_pd_bos(abo);
1312
1313 if (!bo->resource || bo->resource->mem_type != TTM_PL_VRAM ||
1314 !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE) ||
1315 adev->in_suspend || adev->shutdown)
1316 return;
1317
1318 if (WARN_ON_ONCE(!dma_resv_trylock(bo->base.resv)))
1319 return;
1320
1321 r = amdgpu_fill_buffer(abo, AMDGPU_POISON, bo->base.resv, &fence);
1322 if (!WARN_ON(r)) {
1323 amdgpu_bo_fence(abo, fence, false);
1324 dma_fence_put(fence);
1325 }
1326
1327 dma_resv_unlock(bo->base.resv);
1328 }
1329
1330 /**
1331 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1332 * @bo: pointer to a buffer object
1333 *
1334 * Notifies the driver we are taking a fault on this BO and have reserved it,
1335 * also performs bookkeeping.
1336 * TTM driver callback for dealing with vm faults.
1337 *
1338 * Returns:
1339 * 0 for success or a negative error code on failure.
1340 */
amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object * bo)1341 vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1342 {
1343 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1344 struct ttm_operation_ctx ctx = { false, false };
1345 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1346 unsigned long offset;
1347 int r;
1348
1349 /* Remember that this BO was accessed by the CPU */
1350 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1351
1352 if (bo->resource->mem_type != TTM_PL_VRAM)
1353 return 0;
1354
1355 offset = bo->resource->start << PAGE_SHIFT;
1356 if ((offset + bo->base.size) <= adev->gmc.visible_vram_size)
1357 return 0;
1358
1359 /* Can't move a pinned BO to visible VRAM */
1360 if (abo->tbo.pin_count > 0)
1361 return VM_FAULT_SIGBUS;
1362
1363 /* hurrah the memory is not visible ! */
1364 atomic64_inc(&adev->num_vram_cpu_page_faults);
1365 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1366 AMDGPU_GEM_DOMAIN_GTT);
1367
1368 /* Avoid costly evictions; only set GTT as a busy placement */
1369 abo->placement.num_busy_placement = 1;
1370 abo->placement.busy_placement = &abo->placements[1];
1371
1372 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1373 if (unlikely(r == -EBUSY || r == -ERESTARTSYS))
1374 return VM_FAULT_NOPAGE;
1375 else if (unlikely(r))
1376 return VM_FAULT_SIGBUS;
1377
1378 offset = bo->resource->start << PAGE_SHIFT;
1379 /* this should never happen */
1380 if (bo->resource->mem_type == TTM_PL_VRAM &&
1381 (offset + bo->base.size) > adev->gmc.visible_vram_size)
1382 return VM_FAULT_SIGBUS;
1383
1384 ttm_bo_move_to_lru_tail_unlocked(bo);
1385 return 0;
1386 }
1387
1388 /**
1389 * amdgpu_bo_fence - add fence to buffer object
1390 *
1391 * @bo: buffer object in question
1392 * @fence: fence to add
1393 * @shared: true if fence should be added shared
1394 *
1395 */
amdgpu_bo_fence(struct amdgpu_bo * bo,struct dma_fence * fence,bool shared)1396 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1397 bool shared)
1398 {
1399 struct dma_resv *resv = bo->tbo.base.resv;
1400 int r;
1401
1402 r = dma_resv_reserve_fences(resv, 1);
1403 if (r) {
1404 /* As last resort on OOM we block for the fence */
1405 dma_fence_wait(fence, false);
1406 return;
1407 }
1408
1409 dma_resv_add_fence(resv, fence, shared ? DMA_RESV_USAGE_READ :
1410 DMA_RESV_USAGE_WRITE);
1411 }
1412
1413 /**
1414 * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences
1415 *
1416 * @adev: amdgpu device pointer
1417 * @resv: reservation object to sync to
1418 * @sync_mode: synchronization mode
1419 * @owner: fence owner
1420 * @intr: Whether the wait is interruptible
1421 *
1422 * Extract the fences from the reservation object and waits for them to finish.
1423 *
1424 * Returns:
1425 * 0 on success, errno otherwise.
1426 */
amdgpu_bo_sync_wait_resv(struct amdgpu_device * adev,struct dma_resv * resv,enum amdgpu_sync_mode sync_mode,void * owner,bool intr)1427 int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv,
1428 enum amdgpu_sync_mode sync_mode, void *owner,
1429 bool intr)
1430 {
1431 struct amdgpu_sync sync;
1432 int r;
1433
1434 amdgpu_sync_create(&sync);
1435 amdgpu_sync_resv(adev, &sync, resv, sync_mode, owner);
1436 r = amdgpu_sync_wait(&sync, intr);
1437 amdgpu_sync_free(&sync);
1438 return r;
1439 }
1440
1441 /**
1442 * amdgpu_bo_sync_wait - Wrapper for amdgpu_bo_sync_wait_resv
1443 * @bo: buffer object to wait for
1444 * @owner: fence owner
1445 * @intr: Whether the wait is interruptible
1446 *
1447 * Wrapper to wait for fences in a BO.
1448 * Returns:
1449 * 0 on success, errno otherwise.
1450 */
amdgpu_bo_sync_wait(struct amdgpu_bo * bo,void * owner,bool intr)1451 int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)
1452 {
1453 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1454
1455 return amdgpu_bo_sync_wait_resv(adev, bo->tbo.base.resv,
1456 AMDGPU_SYNC_NE_OWNER, owner, intr);
1457 }
1458
1459 /**
1460 * amdgpu_bo_gpu_offset - return GPU offset of bo
1461 * @bo: amdgpu object for which we query the offset
1462 *
1463 * Note: object should either be pinned or reserved when calling this
1464 * function, it might be useful to add check for this for debugging.
1465 *
1466 * Returns:
1467 * current GPU offset of the object.
1468 */
amdgpu_bo_gpu_offset(struct amdgpu_bo * bo)1469 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1470 {
1471 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_SYSTEM);
1472 WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) &&
1473 !bo->tbo.pin_count && bo->tbo.type != ttm_bo_type_kernel);
1474 WARN_ON_ONCE(bo->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET);
1475 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_VRAM &&
1476 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1477
1478 return amdgpu_bo_gpu_offset_no_check(bo);
1479 }
1480
1481 /**
1482 * amdgpu_bo_gpu_offset_no_check - return GPU offset of bo
1483 * @bo: amdgpu object for which we query the offset
1484 *
1485 * Returns:
1486 * current GPU offset of the object without raising warnings.
1487 */
amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo * bo)1488 u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo)
1489 {
1490 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1491 uint64_t offset;
1492
1493 offset = (bo->tbo.resource->start << PAGE_SHIFT) +
1494 amdgpu_ttm_domain_start(adev, bo->tbo.resource->mem_type);
1495
1496 return amdgpu_gmc_sign_extend(offset);
1497 }
1498
1499 /**
1500 * amdgpu_bo_get_preferred_domain - get preferred domain
1501 * @adev: amdgpu device object
1502 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1503 *
1504 * Returns:
1505 * Which of the allowed domains is preferred for allocating the BO.
1506 */
amdgpu_bo_get_preferred_domain(struct amdgpu_device * adev,uint32_t domain)1507 uint32_t amdgpu_bo_get_preferred_domain(struct amdgpu_device *adev,
1508 uint32_t domain)
1509 {
1510 if ((domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) &&
1511 ((adev->asic_type == CHIP_CARRIZO) || (adev->asic_type == CHIP_STONEY))) {
1512 domain = AMDGPU_GEM_DOMAIN_VRAM;
1513 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1514 domain = AMDGPU_GEM_DOMAIN_GTT;
1515 }
1516 return domain;
1517 }
1518
1519 #if defined(CONFIG_DEBUG_FS)
1520 #define amdgpu_bo_print_flag(m, bo, flag) \
1521 do { \
1522 if (bo->flags & (AMDGPU_GEM_CREATE_ ## flag)) { \
1523 seq_printf((m), " " #flag); \
1524 } \
1525 } while (0)
1526
1527 /**
1528 * amdgpu_bo_print_info - print BO info in debugfs file
1529 *
1530 * @id: Index or Id of the BO
1531 * @bo: Requested BO for printing info
1532 * @m: debugfs file
1533 *
1534 * Print BO information in debugfs file
1535 *
1536 * Returns:
1537 * Size of the BO in bytes.
1538 */
amdgpu_bo_print_info(int id,struct amdgpu_bo * bo,struct seq_file * m)1539 u64 amdgpu_bo_print_info(int id, struct amdgpu_bo *bo, struct seq_file *m)
1540 {
1541 struct dma_buf_attachment *attachment;
1542 struct dma_buf *dma_buf;
1543 unsigned int domain;
1544 const char *placement;
1545 unsigned int pin_count;
1546 u64 size;
1547
1548 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type);
1549 switch (domain) {
1550 case AMDGPU_GEM_DOMAIN_VRAM:
1551 placement = "VRAM";
1552 break;
1553 case AMDGPU_GEM_DOMAIN_GTT:
1554 placement = " GTT";
1555 break;
1556 case AMDGPU_GEM_DOMAIN_CPU:
1557 default:
1558 placement = " CPU";
1559 break;
1560 }
1561
1562 size = amdgpu_bo_size(bo);
1563 seq_printf(m, "\t\t0x%08x: %12lld byte %s",
1564 id, size, placement);
1565
1566 pin_count = READ_ONCE(bo->tbo.pin_count);
1567 if (pin_count)
1568 seq_printf(m, " pin count %d", pin_count);
1569
1570 dma_buf = READ_ONCE(bo->tbo.base.dma_buf);
1571 attachment = READ_ONCE(bo->tbo.base.import_attach);
1572
1573 if (attachment)
1574 seq_printf(m, " imported from %p", dma_buf);
1575 else if (dma_buf)
1576 seq_printf(m, " exported as %p", dma_buf);
1577
1578 amdgpu_bo_print_flag(m, bo, CPU_ACCESS_REQUIRED);
1579 amdgpu_bo_print_flag(m, bo, NO_CPU_ACCESS);
1580 amdgpu_bo_print_flag(m, bo, CPU_GTT_USWC);
1581 amdgpu_bo_print_flag(m, bo, VRAM_CLEARED);
1582 amdgpu_bo_print_flag(m, bo, VRAM_CONTIGUOUS);
1583 amdgpu_bo_print_flag(m, bo, VM_ALWAYS_VALID);
1584 amdgpu_bo_print_flag(m, bo, EXPLICIT_SYNC);
1585
1586 seq_puts(m, "\n");
1587
1588 return size;
1589 }
1590 #endif
1591