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 #include "amdgpu_vram_mgr.h"
43 #include "amdgpu_vm.h"
44
45 /**
46 * DOC: amdgpu_object
47 *
48 * This defines the interfaces to operate on an &amdgpu_bo buffer object which
49 * represents memory used by driver (VRAM, system memory, etc.). The driver
50 * provides DRM/GEM APIs to userspace. DRM/GEM APIs then use these interfaces
51 * to create/destroy/set buffer object which are then managed by the kernel TTM
52 * memory manager.
53 * The interfaces are also used internally by kernel clients, including gfx,
54 * uvd, etc. for kernel managed allocations used by the GPU.
55 *
56 */
57
amdgpu_bo_destroy(struct ttm_buffer_object * tbo)58 static void amdgpu_bo_destroy(struct ttm_buffer_object *tbo)
59 {
60 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
61
62 amdgpu_bo_kunmap(bo);
63
64 if (bo->tbo.base.import_attach)
65 drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
66 drm_gem_object_release(&bo->tbo.base);
67 amdgpu_bo_unref(&bo->parent);
68 kvfree(bo);
69 }
70
amdgpu_bo_user_destroy(struct ttm_buffer_object * tbo)71 static void amdgpu_bo_user_destroy(struct ttm_buffer_object *tbo)
72 {
73 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
74 struct amdgpu_bo_user *ubo;
75
76 ubo = to_amdgpu_bo_user(bo);
77 kfree(ubo->metadata);
78 amdgpu_bo_destroy(tbo);
79 }
80
81 /**
82 * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
83 * @bo: buffer object to be checked
84 *
85 * Uses destroy function associated with the object to determine if this is
86 * an &amdgpu_bo.
87 *
88 * Returns:
89 * true if the object belongs to &amdgpu_bo, false if not.
90 */
amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object * bo)91 bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
92 {
93 if (bo->destroy == &amdgpu_bo_destroy ||
94 bo->destroy == &amdgpu_bo_user_destroy)
95 return true;
96
97 return false;
98 }
99
100 /**
101 * amdgpu_bo_placement_from_domain - set buffer's placement
102 * @abo: &amdgpu_bo buffer object whose placement is to be set
103 * @domain: requested domain
104 *
105 * Sets buffer's placement according to requested domain and the buffer's
106 * flags.
107 */
amdgpu_bo_placement_from_domain(struct amdgpu_bo * abo,u32 domain)108 void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
109 {
110 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
111 struct ttm_placement *placement = &abo->placement;
112 struct ttm_place *places = abo->placements;
113 u64 flags = abo->flags;
114 u32 c = 0;
115
116 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
117 unsigned int visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
118 int8_t mem_id = KFD_XCP_MEM_ID(adev, abo->xcp_id);
119
120 if (adev->gmc.mem_partitions && mem_id >= 0) {
121 places[c].fpfn = adev->gmc.mem_partitions[mem_id].range.fpfn;
122 /*
123 * memory partition range lpfn is inclusive start + size - 1
124 * TTM place lpfn is exclusive start + size
125 */
126 places[c].lpfn = adev->gmc.mem_partitions[mem_id].range.lpfn + 1;
127 } else {
128 places[c].fpfn = 0;
129 places[c].lpfn = 0;
130 }
131 places[c].mem_type = TTM_PL_VRAM;
132 places[c].flags = 0;
133
134 if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
135 places[c].lpfn = min_not_zero(places[c].lpfn, visible_pfn);
136 else
137 places[c].flags |= TTM_PL_FLAG_TOPDOWN;
138
139 if (abo->tbo.type == ttm_bo_type_kernel &&
140 flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
141 places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
142
143 c++;
144 }
145
146 if (domain & AMDGPU_GEM_DOMAIN_DOORBELL) {
147 places[c].fpfn = 0;
148 places[c].lpfn = 0;
149 places[c].mem_type = AMDGPU_PL_DOORBELL;
150 places[c].flags = 0;
151 c++;
152 }
153
154 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
155 places[c].fpfn = 0;
156 places[c].lpfn = 0;
157 places[c].mem_type =
158 abo->flags & AMDGPU_GEM_CREATE_PREEMPTIBLE ?
159 AMDGPU_PL_PREEMPT : TTM_PL_TT;
160 places[c].flags = 0;
161 /*
162 * When GTT is just an alternative to VRAM make sure that we
163 * only use it as fallback and still try to fill up VRAM first.
164 */
165 if (domain & abo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM &&
166 !(adev->flags & AMD_IS_APU))
167 places[c].flags |= TTM_PL_FLAG_FALLBACK;
168 c++;
169 }
170
171 if (domain & AMDGPU_GEM_DOMAIN_CPU) {
172 places[c].fpfn = 0;
173 places[c].lpfn = 0;
174 places[c].mem_type = TTM_PL_SYSTEM;
175 places[c].flags = 0;
176 c++;
177 }
178
179 if (domain & AMDGPU_GEM_DOMAIN_GDS) {
180 places[c].fpfn = 0;
181 places[c].lpfn = 0;
182 places[c].mem_type = AMDGPU_PL_GDS;
183 places[c].flags = 0;
184 c++;
185 }
186
187 if (domain & AMDGPU_GEM_DOMAIN_GWS) {
188 places[c].fpfn = 0;
189 places[c].lpfn = 0;
190 places[c].mem_type = AMDGPU_PL_GWS;
191 places[c].flags = 0;
192 c++;
193 }
194
195 if (domain & AMDGPU_GEM_DOMAIN_OA) {
196 places[c].fpfn = 0;
197 places[c].lpfn = 0;
198 places[c].mem_type = AMDGPU_PL_OA;
199 places[c].flags = 0;
200 c++;
201 }
202
203 if (!c) {
204 places[c].fpfn = 0;
205 places[c].lpfn = 0;
206 places[c].mem_type = TTM_PL_SYSTEM;
207 places[c].flags = 0;
208 c++;
209 }
210
211 BUG_ON(c > AMDGPU_BO_MAX_PLACEMENTS);
212
213 placement->num_placement = c;
214 placement->placement = places;
215 }
216
217 /**
218 * amdgpu_bo_create_reserved - create reserved BO for kernel use
219 *
220 * @adev: amdgpu device object
221 * @size: size for the new BO
222 * @align: alignment for the new BO
223 * @domain: where to place it
224 * @bo_ptr: used to initialize BOs in structures
225 * @gpu_addr: GPU addr of the pinned BO
226 * @cpu_addr: optional CPU address mapping
227 *
228 * Allocates and pins a BO for kernel internal use, and returns it still
229 * reserved.
230 *
231 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
232 *
233 * Returns:
234 * 0 on success, negative error code otherwise.
235 */
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)236 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
237 unsigned long size, int align,
238 u32 domain, struct amdgpu_bo **bo_ptr,
239 u64 *gpu_addr, void **cpu_addr)
240 {
241 struct amdgpu_bo_param bp;
242 bool free = false;
243 int r;
244
245 if (!size) {
246 amdgpu_bo_unref(bo_ptr);
247 return 0;
248 }
249
250 memset(&bp, 0, sizeof(bp));
251 bp.size = size;
252 bp.byte_align = align;
253 bp.domain = domain;
254 bp.flags = cpu_addr ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
255 : AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
256 bp.flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
257 bp.type = ttm_bo_type_kernel;
258 bp.resv = NULL;
259 bp.bo_ptr_size = sizeof(struct amdgpu_bo);
260
261 if (!*bo_ptr) {
262 r = amdgpu_bo_create(adev, &bp, bo_ptr);
263 if (r) {
264 dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
265 r);
266 return r;
267 }
268 free = true;
269 }
270
271 r = amdgpu_bo_reserve(*bo_ptr, false);
272 if (r) {
273 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
274 goto error_free;
275 }
276
277 r = amdgpu_bo_pin(*bo_ptr, domain);
278 if (r) {
279 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
280 goto error_unreserve;
281 }
282
283 r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
284 if (r) {
285 dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
286 goto error_unpin;
287 }
288
289 if (gpu_addr)
290 *gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
291
292 if (cpu_addr) {
293 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
294 if (r) {
295 dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
296 goto error_unpin;
297 }
298 }
299
300 return 0;
301
302 error_unpin:
303 amdgpu_bo_unpin(*bo_ptr);
304 error_unreserve:
305 amdgpu_bo_unreserve(*bo_ptr);
306
307 error_free:
308 if (free)
309 amdgpu_bo_unref(bo_ptr);
310
311 return r;
312 }
313
314 /**
315 * amdgpu_bo_create_kernel - create BO for kernel use
316 *
317 * @adev: amdgpu device object
318 * @size: size for the new BO
319 * @align: alignment for the new BO
320 * @domain: where to place it
321 * @bo_ptr: used to initialize BOs in structures
322 * @gpu_addr: GPU addr of the pinned BO
323 * @cpu_addr: optional CPU address mapping
324 *
325 * Allocates and pins a BO for kernel internal use.
326 *
327 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
328 *
329 * Returns:
330 * 0 on success, negative error code otherwise.
331 */
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)332 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
333 unsigned long size, int align,
334 u32 domain, struct amdgpu_bo **bo_ptr,
335 u64 *gpu_addr, void **cpu_addr)
336 {
337 int r;
338
339 r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
340 gpu_addr, cpu_addr);
341
342 if (r)
343 return r;
344
345 if (*bo_ptr)
346 amdgpu_bo_unreserve(*bo_ptr);
347
348 return 0;
349 }
350
351 /**
352 * amdgpu_bo_create_kernel_at - create BO for kernel use at specific location
353 *
354 * @adev: amdgpu device object
355 * @offset: offset of the BO
356 * @size: size of the BO
357 * @bo_ptr: used to initialize BOs in structures
358 * @cpu_addr: optional CPU address mapping
359 *
360 * Creates a kernel BO at a specific offset in VRAM.
361 *
362 * Returns:
363 * 0 on success, negative error code otherwise.
364 */
amdgpu_bo_create_kernel_at(struct amdgpu_device * adev,uint64_t offset,uint64_t size,struct amdgpu_bo ** bo_ptr,void ** cpu_addr)365 int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev,
366 uint64_t offset, uint64_t size,
367 struct amdgpu_bo **bo_ptr, void **cpu_addr)
368 {
369 struct ttm_operation_ctx ctx = { false, false };
370 unsigned int i;
371 int r;
372
373 offset &= PAGE_MASK;
374 size = ALIGN(size, PAGE_SIZE);
375
376 r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE,
377 AMDGPU_GEM_DOMAIN_VRAM, bo_ptr, NULL,
378 cpu_addr);
379 if (r)
380 return r;
381
382 if ((*bo_ptr) == NULL)
383 return 0;
384
385 /*
386 * Remove the original mem node and create a new one at the request
387 * position.
388 */
389 if (cpu_addr)
390 amdgpu_bo_kunmap(*bo_ptr);
391
392 ttm_resource_free(&(*bo_ptr)->tbo, &(*bo_ptr)->tbo.resource);
393
394 for (i = 0; i < (*bo_ptr)->placement.num_placement; ++i) {
395 (*bo_ptr)->placements[i].fpfn = offset >> PAGE_SHIFT;
396 (*bo_ptr)->placements[i].lpfn = (offset + size) >> PAGE_SHIFT;
397 }
398 r = ttm_bo_mem_space(&(*bo_ptr)->tbo, &(*bo_ptr)->placement,
399 &(*bo_ptr)->tbo.resource, &ctx);
400 if (r)
401 goto error;
402
403 if (cpu_addr) {
404 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
405 if (r)
406 goto error;
407 }
408
409 amdgpu_bo_unreserve(*bo_ptr);
410 return 0;
411
412 error:
413 amdgpu_bo_unreserve(*bo_ptr);
414 amdgpu_bo_unref(bo_ptr);
415 return r;
416 }
417
418 /**
419 * amdgpu_bo_free_kernel - free BO for kernel use
420 *
421 * @bo: amdgpu BO to free
422 * @gpu_addr: pointer to where the BO's GPU memory space address was stored
423 * @cpu_addr: pointer to where the BO's CPU memory space address was stored
424 *
425 * unmaps and unpin a BO for kernel internal use.
426 */
amdgpu_bo_free_kernel(struct amdgpu_bo ** bo,u64 * gpu_addr,void ** cpu_addr)427 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
428 void **cpu_addr)
429 {
430 if (*bo == NULL)
431 return;
432
433 WARN_ON(amdgpu_ttm_adev((*bo)->tbo.bdev)->in_suspend);
434
435 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
436 if (cpu_addr)
437 amdgpu_bo_kunmap(*bo);
438
439 amdgpu_bo_unpin(*bo);
440 amdgpu_bo_unreserve(*bo);
441 }
442 amdgpu_bo_unref(bo);
443
444 if (gpu_addr)
445 *gpu_addr = 0;
446
447 if (cpu_addr)
448 *cpu_addr = NULL;
449 }
450
451 /* Validate bo size is bit bigger than the request domain */
amdgpu_bo_validate_size(struct amdgpu_device * adev,unsigned long size,u32 domain)452 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
453 unsigned long size, u32 domain)
454 {
455 struct ttm_resource_manager *man = NULL;
456
457 /*
458 * If GTT is part of requested domains the check must succeed to
459 * allow fall back to GTT.
460 */
461 if (domain & AMDGPU_GEM_DOMAIN_GTT)
462 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_TT);
463 else if (domain & AMDGPU_GEM_DOMAIN_VRAM)
464 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
465 else
466 return true;
467
468 if (!man) {
469 if (domain & AMDGPU_GEM_DOMAIN_GTT)
470 WARN_ON_ONCE("GTT domain requested but GTT mem manager uninitialized");
471 return false;
472 }
473
474 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU, _DOMAIN_DOORBELL */
475 if (size < man->size)
476 return true;
477
478 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size, man->size);
479 return false;
480 }
481
amdgpu_bo_support_uswc(u64 bo_flags)482 bool amdgpu_bo_support_uswc(u64 bo_flags)
483 {
484
485 #ifdef CONFIG_X86_32
486 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
487 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
488 */
489 return false;
490 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
491 /* Don't try to enable write-combining when it can't work, or things
492 * may be slow
493 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
494 */
495
496 #ifndef CONFIG_COMPILE_TEST
497 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
498 thanks to write-combining
499 #endif
500
501 if (bo_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
502 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
503 "better performance thanks to write-combining\n");
504 return false;
505 #else
506 /* For architectures that don't support WC memory,
507 * mask out the WC flag from the BO
508 */
509 if (!drm_arch_can_wc_memory())
510 return false;
511
512 return true;
513 #endif
514 }
515
516 /**
517 * amdgpu_bo_create - create an &amdgpu_bo buffer object
518 * @adev: amdgpu device object
519 * @bp: parameters to be used for the buffer object
520 * @bo_ptr: pointer to the buffer object pointer
521 *
522 * Creates an &amdgpu_bo buffer object.
523 *
524 * Returns:
525 * 0 for success or a negative error code on failure.
526 */
amdgpu_bo_create(struct amdgpu_device * adev,struct amdgpu_bo_param * bp,struct amdgpu_bo ** bo_ptr)527 int amdgpu_bo_create(struct amdgpu_device *adev,
528 struct amdgpu_bo_param *bp,
529 struct amdgpu_bo **bo_ptr)
530 {
531 struct ttm_operation_ctx ctx = {
532 .interruptible = (bp->type != ttm_bo_type_kernel),
533 .no_wait_gpu = bp->no_wait_gpu,
534 /* We opt to avoid OOM on system pages allocations */
535 .gfp_retry_mayfail = true,
536 .allow_res_evict = bp->type != ttm_bo_type_kernel,
537 .resv = bp->resv
538 };
539 struct amdgpu_bo *bo;
540 unsigned long page_align, size = bp->size;
541 int r;
542
543 /* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
544 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
545 /* GWS and OA don't need any alignment. */
546 page_align = bp->byte_align;
547 size <<= PAGE_SHIFT;
548
549 } else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
550 /* Both size and alignment must be a multiple of 4. */
551 page_align = ALIGN(bp->byte_align, 4);
552 size = ALIGN(size, 4) << PAGE_SHIFT;
553 } else {
554 /* Memory should be aligned at least to a page size. */
555 page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
556 size = ALIGN(size, PAGE_SIZE);
557 }
558
559 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
560 return -ENOMEM;
561
562 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo));
563
564 *bo_ptr = NULL;
565 bo = kvzalloc(bp->bo_ptr_size, GFP_KERNEL);
566 if (bo == NULL)
567 return -ENOMEM;
568 drm_gem_private_object_init(adev_to_drm(adev), &bo->tbo.base, size);
569 bo->tbo.base.funcs = &amdgpu_gem_object_funcs;
570 bo->vm_bo = NULL;
571 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
572 bp->domain;
573 bo->allowed_domains = bo->preferred_domains;
574 if (bp->type != ttm_bo_type_kernel &&
575 !(bp->flags & AMDGPU_GEM_CREATE_DISCARDABLE) &&
576 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
577 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
578
579 bo->flags = bp->flags;
580
581 if (adev->gmc.mem_partitions)
582 /* For GPUs with spatial partitioning, bo->xcp_id=-1 means any partition */
583 bo->xcp_id = bp->xcp_id_plus1 - 1;
584 else
585 /* For GPUs without spatial partitioning */
586 bo->xcp_id = 0;
587
588 if (!amdgpu_bo_support_uswc(bo->flags))
589 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
590
591 bo->tbo.bdev = &adev->mman.bdev;
592 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
593 AMDGPU_GEM_DOMAIN_GDS))
594 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
595 else
596 amdgpu_bo_placement_from_domain(bo, bp->domain);
597 if (bp->type == ttm_bo_type_kernel)
598 bo->tbo.priority = 2;
599 else if (!(bp->flags & AMDGPU_GEM_CREATE_DISCARDABLE))
600 bo->tbo.priority = 1;
601
602 if (!bp->destroy)
603 bp->destroy = &amdgpu_bo_destroy;
604
605 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, bp->type,
606 &bo->placement, page_align, &ctx, NULL,
607 bp->resv, bp->destroy);
608 if (unlikely(r != 0))
609 return r;
610
611 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
612 amdgpu_res_cpu_visible(adev, bo->tbo.resource))
613 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
614 ctx.bytes_moved);
615 else
616 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
617
618 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
619 bo->tbo.resource->mem_type == TTM_PL_VRAM) {
620 struct dma_fence *fence;
621
622 r = amdgpu_ttm_clear_buffer(bo, bo->tbo.base.resv, &fence);
623 if (unlikely(r))
624 goto fail_unreserve;
625
626 dma_resv_add_fence(bo->tbo.base.resv, fence,
627 DMA_RESV_USAGE_KERNEL);
628 dma_fence_put(fence);
629 }
630 if (!bp->resv)
631 amdgpu_bo_unreserve(bo);
632 *bo_ptr = bo;
633
634 trace_amdgpu_bo_create(bo);
635
636 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
637 if (bp->type == ttm_bo_type_device)
638 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
639
640 return 0;
641
642 fail_unreserve:
643 if (!bp->resv)
644 dma_resv_unlock(bo->tbo.base.resv);
645 amdgpu_bo_unref(&bo);
646 return r;
647 }
648
649 /**
650 * amdgpu_bo_create_user - create an &amdgpu_bo_user buffer object
651 * @adev: amdgpu device object
652 * @bp: parameters to be used for the buffer object
653 * @ubo_ptr: pointer to the buffer object pointer
654 *
655 * Create a BO to be used by user application;
656 *
657 * Returns:
658 * 0 for success or a negative error code on failure.
659 */
660
amdgpu_bo_create_user(struct amdgpu_device * adev,struct amdgpu_bo_param * bp,struct amdgpu_bo_user ** ubo_ptr)661 int amdgpu_bo_create_user(struct amdgpu_device *adev,
662 struct amdgpu_bo_param *bp,
663 struct amdgpu_bo_user **ubo_ptr)
664 {
665 struct amdgpu_bo *bo_ptr;
666 int r;
667
668 bp->bo_ptr_size = sizeof(struct amdgpu_bo_user);
669 bp->destroy = &amdgpu_bo_user_destroy;
670 r = amdgpu_bo_create(adev, bp, &bo_ptr);
671 if (r)
672 return r;
673
674 *ubo_ptr = to_amdgpu_bo_user(bo_ptr);
675 return r;
676 }
677
678 /**
679 * amdgpu_bo_create_vm - create an &amdgpu_bo_vm buffer object
680 * @adev: amdgpu device object
681 * @bp: parameters to be used for the buffer object
682 * @vmbo_ptr: pointer to the buffer object pointer
683 *
684 * Create a BO to be for GPUVM.
685 *
686 * Returns:
687 * 0 for success or a negative error code on failure.
688 */
689
amdgpu_bo_create_vm(struct amdgpu_device * adev,struct amdgpu_bo_param * bp,struct amdgpu_bo_vm ** vmbo_ptr)690 int amdgpu_bo_create_vm(struct amdgpu_device *adev,
691 struct amdgpu_bo_param *bp,
692 struct amdgpu_bo_vm **vmbo_ptr)
693 {
694 struct amdgpu_bo *bo_ptr;
695 int r;
696
697 /* bo_ptr_size will be determined by the caller and it depends on
698 * num of amdgpu_vm_pt entries.
699 */
700 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo_vm));
701 r = amdgpu_bo_create(adev, bp, &bo_ptr);
702 if (r)
703 return r;
704
705 *vmbo_ptr = to_amdgpu_bo_vm(bo_ptr);
706 return r;
707 }
708
709 /**
710 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
711 * @bo: &amdgpu_bo buffer object to be mapped
712 * @ptr: kernel virtual address to be returned
713 *
714 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
715 * amdgpu_bo_kptr() to get the kernel virtual address.
716 *
717 * Returns:
718 * 0 for success or a negative error code on failure.
719 */
amdgpu_bo_kmap(struct amdgpu_bo * bo,void ** ptr)720 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
721 {
722 void *kptr;
723 long r;
724
725 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
726 return -EPERM;
727
728 r = dma_resv_wait_timeout(bo->tbo.base.resv, DMA_RESV_USAGE_KERNEL,
729 false, MAX_SCHEDULE_TIMEOUT);
730 if (r < 0)
731 return r;
732
733 kptr = amdgpu_bo_kptr(bo);
734 if (kptr) {
735 if (ptr)
736 *ptr = kptr;
737 return 0;
738 }
739
740 r = ttm_bo_kmap(&bo->tbo, 0, PFN_UP(bo->tbo.base.size), &bo->kmap);
741 if (r)
742 return r;
743
744 if (ptr)
745 *ptr = amdgpu_bo_kptr(bo);
746
747 return 0;
748 }
749
750 /**
751 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
752 * @bo: &amdgpu_bo buffer object
753 *
754 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
755 *
756 * Returns:
757 * the virtual address of a buffer object area.
758 */
amdgpu_bo_kptr(struct amdgpu_bo * bo)759 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
760 {
761 bool is_iomem;
762
763 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
764 }
765
766 /**
767 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
768 * @bo: &amdgpu_bo buffer object to be unmapped
769 *
770 * Unmaps a kernel map set up by amdgpu_bo_kmap().
771 */
amdgpu_bo_kunmap(struct amdgpu_bo * bo)772 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
773 {
774 if (bo->kmap.bo)
775 ttm_bo_kunmap(&bo->kmap);
776 }
777
778 /**
779 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
780 * @bo: &amdgpu_bo buffer object
781 *
782 * References the contained &ttm_buffer_object.
783 *
784 * Returns:
785 * a refcounted pointer to the &amdgpu_bo buffer object.
786 */
amdgpu_bo_ref(struct amdgpu_bo * bo)787 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
788 {
789 if (bo == NULL)
790 return NULL;
791
792 drm_gem_object_get(&bo->tbo.base);
793 return bo;
794 }
795
796 /**
797 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
798 * @bo: &amdgpu_bo buffer object
799 *
800 * Unreferences the contained &ttm_buffer_object and clear the pointer
801 */
amdgpu_bo_unref(struct amdgpu_bo ** bo)802 void amdgpu_bo_unref(struct amdgpu_bo **bo)
803 {
804 if ((*bo) == NULL)
805 return;
806
807 drm_gem_object_put(&(*bo)->tbo.base);
808 *bo = NULL;
809 }
810
811 /**
812 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
813 * @bo: &amdgpu_bo buffer object to be pinned
814 * @domain: domain to be pinned to
815 *
816 * Pins the buffer object according to requested domain. If the memory is
817 * unbound gart memory, binds the pages into gart table. Adjusts pin_count and
818 * pin_size accordingly.
819 *
820 * Pinning means to lock pages in memory along with keeping them at a fixed
821 * offset. It is required when a buffer can not be moved, for example, when
822 * a display buffer is being scanned out.
823 *
824 * Returns:
825 * 0 for success or a negative error code on failure.
826 */
amdgpu_bo_pin(struct amdgpu_bo * bo,u32 domain)827 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
828 {
829 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
830 struct ttm_operation_ctx ctx = { false, false };
831 int r, i;
832
833 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
834 return -EPERM;
835
836 /* Check domain to be pinned to against preferred domains */
837 if (bo->preferred_domains & domain)
838 domain = bo->preferred_domains & domain;
839
840 /* A shared bo cannot be migrated to VRAM */
841 if (bo->tbo.base.import_attach) {
842 if (domain & AMDGPU_GEM_DOMAIN_GTT)
843 domain = AMDGPU_GEM_DOMAIN_GTT;
844 else
845 return -EINVAL;
846 }
847
848 if (bo->tbo.pin_count) {
849 uint32_t mem_type = bo->tbo.resource->mem_type;
850 uint32_t mem_flags = bo->tbo.resource->placement;
851
852 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
853 return -EINVAL;
854
855 if ((mem_type == TTM_PL_VRAM) &&
856 (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) &&
857 !(mem_flags & TTM_PL_FLAG_CONTIGUOUS))
858 return -EINVAL;
859
860 ttm_bo_pin(&bo->tbo);
861 return 0;
862 }
863
864 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
865 * See function amdgpu_display_supported_domains()
866 */
867 domain = amdgpu_bo_get_preferred_domain(adev, domain);
868
869 if (bo->tbo.base.import_attach)
870 dma_buf_pin(bo->tbo.base.import_attach);
871
872 /* force to pin into visible video ram */
873 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
874 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
875 amdgpu_bo_placement_from_domain(bo, domain);
876 for (i = 0; i < bo->placement.num_placement; i++) {
877 if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS &&
878 bo->placements[i].mem_type == TTM_PL_VRAM)
879 bo->placements[i].flags |= TTM_PL_FLAG_CONTIGUOUS;
880 }
881
882 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
883 if (unlikely(r)) {
884 dev_err(adev->dev, "%p pin failed\n", bo);
885 goto error;
886 }
887
888 ttm_bo_pin(&bo->tbo);
889
890 if (bo->tbo.resource->mem_type == TTM_PL_VRAM) {
891 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
892 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
893 &adev->visible_pin_size);
894 } else if (bo->tbo.resource->mem_type == TTM_PL_TT) {
895 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
896 }
897
898 error:
899 return r;
900 }
901
902 /**
903 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
904 * @bo: &amdgpu_bo buffer object to be unpinned
905 *
906 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
907 * Changes placement and pin size accordingly.
908 *
909 * Returns:
910 * 0 for success or a negative error code on failure.
911 */
amdgpu_bo_unpin(struct amdgpu_bo * bo)912 void amdgpu_bo_unpin(struct amdgpu_bo *bo)
913 {
914 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
915
916 ttm_bo_unpin(&bo->tbo);
917 if (bo->tbo.pin_count)
918 return;
919
920 if (bo->tbo.base.import_attach)
921 dma_buf_unpin(bo->tbo.base.import_attach);
922
923 if (bo->tbo.resource->mem_type == TTM_PL_VRAM) {
924 atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
925 atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
926 &adev->visible_pin_size);
927 } else if (bo->tbo.resource->mem_type == TTM_PL_TT) {
928 atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size);
929 }
930
931 }
932
933 static const char * const amdgpu_vram_names[] = {
934 "UNKNOWN",
935 "GDDR1",
936 "DDR2",
937 "GDDR3",
938 "GDDR4",
939 "GDDR5",
940 "HBM",
941 "DDR3",
942 "DDR4",
943 "GDDR6",
944 "DDR5",
945 "LPDDR4",
946 "LPDDR5"
947 };
948
949 /**
950 * amdgpu_bo_init - initialize memory manager
951 * @adev: amdgpu device object
952 *
953 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
954 *
955 * Returns:
956 * 0 for success or a negative error code on failure.
957 */
amdgpu_bo_init(struct amdgpu_device * adev)958 int amdgpu_bo_init(struct amdgpu_device *adev)
959 {
960 /* On A+A platform, VRAM can be mapped as WB */
961 if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) {
962 /* reserve PAT memory space to WC for VRAM */
963 int r = arch_io_reserve_memtype_wc(adev->gmc.aper_base,
964 adev->gmc.aper_size);
965
966 if (r) {
967 DRM_ERROR("Unable to set WC memtype for the aperture base\n");
968 return r;
969 }
970
971 /* Add an MTRR for the VRAM */
972 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
973 adev->gmc.aper_size);
974 }
975
976 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
977 adev->gmc.mc_vram_size >> 20,
978 (unsigned long long)adev->gmc.aper_size >> 20);
979 DRM_INFO("RAM width %dbits %s\n",
980 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
981 return amdgpu_ttm_init(adev);
982 }
983
984 /**
985 * amdgpu_bo_fini - tear down memory manager
986 * @adev: amdgpu device object
987 *
988 * Reverses amdgpu_bo_init() to tear down memory manager.
989 */
amdgpu_bo_fini(struct amdgpu_device * adev)990 void amdgpu_bo_fini(struct amdgpu_device *adev)
991 {
992 int idx;
993
994 amdgpu_ttm_fini(adev);
995
996 if (drm_dev_enter(adev_to_drm(adev), &idx)) {
997 if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) {
998 arch_phys_wc_del(adev->gmc.vram_mtrr);
999 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1000 }
1001 drm_dev_exit(idx);
1002 }
1003 }
1004
1005 /**
1006 * amdgpu_bo_set_tiling_flags - set tiling flags
1007 * @bo: &amdgpu_bo buffer object
1008 * @tiling_flags: new flags
1009 *
1010 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1011 * kernel driver to set the tiling flags on a buffer.
1012 *
1013 * Returns:
1014 * 0 for success or a negative error code on failure.
1015 */
amdgpu_bo_set_tiling_flags(struct amdgpu_bo * bo,u64 tiling_flags)1016 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1017 {
1018 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1019 struct amdgpu_bo_user *ubo;
1020
1021 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1022 if (adev->family <= AMDGPU_FAMILY_CZ &&
1023 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1024 return -EINVAL;
1025
1026 ubo = to_amdgpu_bo_user(bo);
1027 ubo->tiling_flags = tiling_flags;
1028 return 0;
1029 }
1030
1031 /**
1032 * amdgpu_bo_get_tiling_flags - get tiling flags
1033 * @bo: &amdgpu_bo buffer object
1034 * @tiling_flags: returned flags
1035 *
1036 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1037 * set the tiling flags on a buffer.
1038 */
amdgpu_bo_get_tiling_flags(struct amdgpu_bo * bo,u64 * tiling_flags)1039 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1040 {
1041 struct amdgpu_bo_user *ubo;
1042
1043 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1044 dma_resv_assert_held(bo->tbo.base.resv);
1045 ubo = to_amdgpu_bo_user(bo);
1046
1047 if (tiling_flags)
1048 *tiling_flags = ubo->tiling_flags;
1049 }
1050
1051 /**
1052 * amdgpu_bo_set_metadata - set metadata
1053 * @bo: &amdgpu_bo buffer object
1054 * @metadata: new metadata
1055 * @metadata_size: size of the new metadata
1056 * @flags: flags of the new metadata
1057 *
1058 * Sets buffer object's metadata, its size and flags.
1059 * Used via GEM ioctl.
1060 *
1061 * Returns:
1062 * 0 for success or a negative error code on failure.
1063 */
amdgpu_bo_set_metadata(struct amdgpu_bo * bo,void * metadata,u32 metadata_size,uint64_t flags)1064 int amdgpu_bo_set_metadata(struct amdgpu_bo *bo, void *metadata,
1065 u32 metadata_size, uint64_t flags)
1066 {
1067 struct amdgpu_bo_user *ubo;
1068 void *buffer;
1069
1070 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1071 ubo = to_amdgpu_bo_user(bo);
1072 if (!metadata_size) {
1073 if (ubo->metadata_size) {
1074 kfree(ubo->metadata);
1075 ubo->metadata = NULL;
1076 ubo->metadata_size = 0;
1077 }
1078 return 0;
1079 }
1080
1081 if (metadata == NULL)
1082 return -EINVAL;
1083
1084 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1085 if (buffer == NULL)
1086 return -ENOMEM;
1087
1088 kfree(ubo->metadata);
1089 ubo->metadata_flags = flags;
1090 ubo->metadata = buffer;
1091 ubo->metadata_size = metadata_size;
1092
1093 return 0;
1094 }
1095
1096 /**
1097 * amdgpu_bo_get_metadata - get metadata
1098 * @bo: &amdgpu_bo buffer object
1099 * @buffer: returned metadata
1100 * @buffer_size: size of the buffer
1101 * @metadata_size: size of the returned metadata
1102 * @flags: flags of the returned metadata
1103 *
1104 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1105 * less than metadata_size.
1106 * Used via GEM ioctl.
1107 *
1108 * Returns:
1109 * 0 for success or a negative error code on failure.
1110 */
amdgpu_bo_get_metadata(struct amdgpu_bo * bo,void * buffer,size_t buffer_size,uint32_t * metadata_size,uint64_t * flags)1111 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1112 size_t buffer_size, uint32_t *metadata_size,
1113 uint64_t *flags)
1114 {
1115 struct amdgpu_bo_user *ubo;
1116
1117 if (!buffer && !metadata_size)
1118 return -EINVAL;
1119
1120 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1121 ubo = to_amdgpu_bo_user(bo);
1122 if (metadata_size)
1123 *metadata_size = ubo->metadata_size;
1124
1125 if (buffer) {
1126 if (buffer_size < ubo->metadata_size)
1127 return -EINVAL;
1128
1129 if (ubo->metadata_size)
1130 memcpy(buffer, ubo->metadata, ubo->metadata_size);
1131 }
1132
1133 if (flags)
1134 *flags = ubo->metadata_flags;
1135
1136 return 0;
1137 }
1138
1139 /**
1140 * amdgpu_bo_move_notify - notification about a memory move
1141 * @bo: pointer to a buffer object
1142 * @evict: if this move is evicting the buffer from the graphics address space
1143 * @new_mem: new resource for backing the BO
1144 *
1145 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1146 * bookkeeping.
1147 * TTM driver callback which is called when ttm moves a buffer.
1148 */
amdgpu_bo_move_notify(struct ttm_buffer_object * bo,bool evict,struct ttm_resource * new_mem)1149 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1150 bool evict,
1151 struct ttm_resource *new_mem)
1152 {
1153 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1154 struct ttm_resource *old_mem = bo->resource;
1155 struct amdgpu_bo *abo;
1156
1157 if (!amdgpu_bo_is_amdgpu_bo(bo))
1158 return;
1159
1160 abo = ttm_to_amdgpu_bo(bo);
1161 amdgpu_vm_bo_invalidate(adev, abo, evict);
1162
1163 amdgpu_bo_kunmap(abo);
1164
1165 if (abo->tbo.base.dma_buf && !abo->tbo.base.import_attach &&
1166 old_mem && old_mem->mem_type != TTM_PL_SYSTEM)
1167 dma_buf_move_notify(abo->tbo.base.dma_buf);
1168
1169 /* move_notify is called before move happens */
1170 trace_amdgpu_bo_move(abo, new_mem ? new_mem->mem_type : -1,
1171 old_mem ? old_mem->mem_type : -1);
1172 }
1173
amdgpu_bo_get_memory(struct amdgpu_bo * bo,struct amdgpu_mem_stats * stats,unsigned int sz)1174 void amdgpu_bo_get_memory(struct amdgpu_bo *bo,
1175 struct amdgpu_mem_stats *stats,
1176 unsigned int sz)
1177 {
1178 const unsigned int domain_to_pl[] = {
1179 [ilog2(AMDGPU_GEM_DOMAIN_CPU)] = TTM_PL_SYSTEM,
1180 [ilog2(AMDGPU_GEM_DOMAIN_GTT)] = TTM_PL_TT,
1181 [ilog2(AMDGPU_GEM_DOMAIN_VRAM)] = TTM_PL_VRAM,
1182 [ilog2(AMDGPU_GEM_DOMAIN_GDS)] = AMDGPU_PL_GDS,
1183 [ilog2(AMDGPU_GEM_DOMAIN_GWS)] = AMDGPU_PL_GWS,
1184 [ilog2(AMDGPU_GEM_DOMAIN_OA)] = AMDGPU_PL_OA,
1185 [ilog2(AMDGPU_GEM_DOMAIN_DOORBELL)] = AMDGPU_PL_DOORBELL,
1186 };
1187 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1188 struct ttm_resource *res = bo->tbo.resource;
1189 struct drm_gem_object *obj = &bo->tbo.base;
1190 uint64_t size = amdgpu_bo_size(bo);
1191 unsigned int type;
1192
1193 if (!res) {
1194 /*
1195 * If no backing store use one of the preferred domain for basic
1196 * stats. We take the MSB since that should give a reasonable
1197 * view.
1198 */
1199 BUILD_BUG_ON(TTM_PL_VRAM < TTM_PL_TT ||
1200 TTM_PL_VRAM < TTM_PL_SYSTEM);
1201 type = fls(bo->preferred_domains & AMDGPU_GEM_DOMAIN_MASK);
1202 if (!type)
1203 return;
1204 type--;
1205 if (drm_WARN_ON_ONCE(&adev->ddev,
1206 type >= ARRAY_SIZE(domain_to_pl)))
1207 return;
1208 type = domain_to_pl[type];
1209 } else {
1210 type = res->mem_type;
1211 }
1212
1213 if (drm_WARN_ON_ONCE(&adev->ddev, type >= sz))
1214 return;
1215
1216 /* DRM stats common fields: */
1217
1218 if (drm_gem_object_is_shared_for_memory_stats(obj))
1219 stats[type].drm.shared += size;
1220 else
1221 stats[type].drm.private += size;
1222
1223 if (res) {
1224 stats[type].drm.resident += size;
1225
1226 if (!dma_resv_test_signaled(obj->resv, DMA_RESV_USAGE_BOOKKEEP))
1227 stats[type].drm.active += size;
1228 else if (bo->flags & AMDGPU_GEM_CREATE_DISCARDABLE)
1229 stats[type].drm.purgeable += size;
1230 }
1231
1232 /* amdgpu specific stats: */
1233
1234 if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) {
1235 stats[TTM_PL_VRAM].requested += size;
1236 if (type != TTM_PL_VRAM)
1237 stats[TTM_PL_VRAM].evicted += size;
1238 } else if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_GTT) {
1239 stats[TTM_PL_TT].requested += size;
1240 }
1241 }
1242
1243 /**
1244 * amdgpu_bo_release_notify - notification about a BO being released
1245 * @bo: pointer to a buffer object
1246 *
1247 * Wipes VRAM buffers whose contents should not be leaked before the
1248 * memory is released.
1249 */
amdgpu_bo_release_notify(struct ttm_buffer_object * bo)1250 void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)
1251 {
1252 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1253 struct dma_fence *fence = NULL;
1254 struct amdgpu_bo *abo;
1255 int r;
1256
1257 if (!amdgpu_bo_is_amdgpu_bo(bo))
1258 return;
1259
1260 abo = ttm_to_amdgpu_bo(bo);
1261
1262 WARN_ON(abo->vm_bo);
1263
1264 if (abo->kfd_bo)
1265 amdgpu_amdkfd_release_notify(abo);
1266
1267 /* We only remove the fence if the resv has individualized. */
1268 WARN_ON_ONCE(bo->type == ttm_bo_type_kernel
1269 && bo->base.resv != &bo->base._resv);
1270 if (bo->base.resv == &bo->base._resv)
1271 amdgpu_amdkfd_remove_fence_on_pt_pd_bos(abo);
1272
1273 if (!bo->resource || bo->resource->mem_type != TTM_PL_VRAM ||
1274 !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE) ||
1275 adev->in_suspend || drm_dev_is_unplugged(adev_to_drm(adev)))
1276 return;
1277
1278 if (WARN_ON_ONCE(!dma_resv_trylock(bo->base.resv)))
1279 return;
1280
1281 r = amdgpu_fill_buffer(abo, 0, bo->base.resv, &fence, true);
1282 if (!WARN_ON(r)) {
1283 amdgpu_vram_mgr_set_cleared(bo->resource);
1284 amdgpu_bo_fence(abo, fence, false);
1285 dma_fence_put(fence);
1286 }
1287
1288 dma_resv_unlock(bo->base.resv);
1289 }
1290
1291 /**
1292 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1293 * @bo: pointer to a buffer object
1294 *
1295 * Notifies the driver we are taking a fault on this BO and have reserved it,
1296 * also performs bookkeeping.
1297 * TTM driver callback for dealing with vm faults.
1298 *
1299 * Returns:
1300 * 0 for success or a negative error code on failure.
1301 */
amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object * bo)1302 vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1303 {
1304 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1305 struct ttm_operation_ctx ctx = { false, false };
1306 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1307 int r;
1308
1309 /* Remember that this BO was accessed by the CPU */
1310 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1311
1312 if (amdgpu_res_cpu_visible(adev, bo->resource))
1313 return 0;
1314
1315 /* Can't move a pinned BO to visible VRAM */
1316 if (abo->tbo.pin_count > 0)
1317 return VM_FAULT_SIGBUS;
1318
1319 /* hurrah the memory is not visible ! */
1320 atomic64_inc(&adev->num_vram_cpu_page_faults);
1321 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1322 AMDGPU_GEM_DOMAIN_GTT);
1323
1324 /* Avoid costly evictions; only set GTT as a busy placement */
1325 abo->placements[0].flags |= TTM_PL_FLAG_DESIRED;
1326
1327 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1328 if (unlikely(r == -EBUSY || r == -ERESTARTSYS))
1329 return VM_FAULT_NOPAGE;
1330 else if (unlikely(r))
1331 return VM_FAULT_SIGBUS;
1332
1333 /* this should never happen */
1334 if (bo->resource->mem_type == TTM_PL_VRAM &&
1335 !amdgpu_res_cpu_visible(adev, bo->resource))
1336 return VM_FAULT_SIGBUS;
1337
1338 ttm_bo_move_to_lru_tail_unlocked(bo);
1339 return 0;
1340 }
1341
1342 /**
1343 * amdgpu_bo_fence - add fence to buffer object
1344 *
1345 * @bo: buffer object in question
1346 * @fence: fence to add
1347 * @shared: true if fence should be added shared
1348 *
1349 */
amdgpu_bo_fence(struct amdgpu_bo * bo,struct dma_fence * fence,bool shared)1350 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1351 bool shared)
1352 {
1353 struct dma_resv *resv = bo->tbo.base.resv;
1354 int r;
1355
1356 r = dma_resv_reserve_fences(resv, 1);
1357 if (r) {
1358 /* As last resort on OOM we block for the fence */
1359 dma_fence_wait(fence, false);
1360 return;
1361 }
1362
1363 dma_resv_add_fence(resv, fence, shared ? DMA_RESV_USAGE_READ :
1364 DMA_RESV_USAGE_WRITE);
1365 }
1366
1367 /**
1368 * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences
1369 *
1370 * @adev: amdgpu device pointer
1371 * @resv: reservation object to sync to
1372 * @sync_mode: synchronization mode
1373 * @owner: fence owner
1374 * @intr: Whether the wait is interruptible
1375 *
1376 * Extract the fences from the reservation object and waits for them to finish.
1377 *
1378 * Returns:
1379 * 0 on success, errno otherwise.
1380 */
amdgpu_bo_sync_wait_resv(struct amdgpu_device * adev,struct dma_resv * resv,enum amdgpu_sync_mode sync_mode,void * owner,bool intr)1381 int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv,
1382 enum amdgpu_sync_mode sync_mode, void *owner,
1383 bool intr)
1384 {
1385 struct amdgpu_sync sync;
1386 int r;
1387
1388 amdgpu_sync_create(&sync);
1389 amdgpu_sync_resv(adev, &sync, resv, sync_mode, owner);
1390 r = amdgpu_sync_wait(&sync, intr);
1391 amdgpu_sync_free(&sync);
1392 return r;
1393 }
1394
1395 /**
1396 * amdgpu_bo_sync_wait - Wrapper for amdgpu_bo_sync_wait_resv
1397 * @bo: buffer object to wait for
1398 * @owner: fence owner
1399 * @intr: Whether the wait is interruptible
1400 *
1401 * Wrapper to wait for fences in a BO.
1402 * Returns:
1403 * 0 on success, errno otherwise.
1404 */
amdgpu_bo_sync_wait(struct amdgpu_bo * bo,void * owner,bool intr)1405 int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)
1406 {
1407 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1408
1409 return amdgpu_bo_sync_wait_resv(adev, bo->tbo.base.resv,
1410 AMDGPU_SYNC_NE_OWNER, owner, intr);
1411 }
1412
1413 /**
1414 * amdgpu_bo_gpu_offset - return GPU offset of bo
1415 * @bo: amdgpu object for which we query the offset
1416 *
1417 * Note: object should either be pinned or reserved when calling this
1418 * function, it might be useful to add check for this for debugging.
1419 *
1420 * Returns:
1421 * current GPU offset of the object.
1422 */
amdgpu_bo_gpu_offset(struct amdgpu_bo * bo)1423 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1424 {
1425 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_SYSTEM);
1426 WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) &&
1427 !bo->tbo.pin_count && bo->tbo.type != ttm_bo_type_kernel);
1428 WARN_ON_ONCE(bo->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET);
1429 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_VRAM &&
1430 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1431
1432 return amdgpu_bo_gpu_offset_no_check(bo);
1433 }
1434
1435 /**
1436 * amdgpu_bo_gpu_offset_no_check - return GPU offset of bo
1437 * @bo: amdgpu object for which we query the offset
1438 *
1439 * Returns:
1440 * current GPU offset of the object without raising warnings.
1441 */
amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo * bo)1442 u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo)
1443 {
1444 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1445 uint64_t offset = AMDGPU_BO_INVALID_OFFSET;
1446
1447 if (bo->tbo.resource->mem_type == TTM_PL_TT)
1448 offset = amdgpu_gmc_agp_addr(&bo->tbo);
1449
1450 if (offset == AMDGPU_BO_INVALID_OFFSET)
1451 offset = (bo->tbo.resource->start << PAGE_SHIFT) +
1452 amdgpu_ttm_domain_start(adev, bo->tbo.resource->mem_type);
1453
1454 return amdgpu_gmc_sign_extend(offset);
1455 }
1456
1457 /**
1458 * amdgpu_bo_get_preferred_domain - get preferred domain
1459 * @adev: amdgpu device object
1460 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1461 *
1462 * Returns:
1463 * Which of the allowed domains is preferred for allocating the BO.
1464 */
amdgpu_bo_get_preferred_domain(struct amdgpu_device * adev,uint32_t domain)1465 uint32_t amdgpu_bo_get_preferred_domain(struct amdgpu_device *adev,
1466 uint32_t domain)
1467 {
1468 if ((domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) &&
1469 ((adev->asic_type == CHIP_CARRIZO) || (adev->asic_type == CHIP_STONEY))) {
1470 domain = AMDGPU_GEM_DOMAIN_VRAM;
1471 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1472 domain = AMDGPU_GEM_DOMAIN_GTT;
1473 }
1474 return domain;
1475 }
1476
1477 #if defined(CONFIG_DEBUG_FS)
1478 #define amdgpu_bo_print_flag(m, bo, flag) \
1479 do { \
1480 if (bo->flags & (AMDGPU_GEM_CREATE_ ## flag)) { \
1481 seq_printf((m), " " #flag); \
1482 } \
1483 } while (0)
1484
1485 /**
1486 * amdgpu_bo_print_info - print BO info in debugfs file
1487 *
1488 * @id: Index or Id of the BO
1489 * @bo: Requested BO for printing info
1490 * @m: debugfs file
1491 *
1492 * Print BO information in debugfs file
1493 *
1494 * Returns:
1495 * Size of the BO in bytes.
1496 */
amdgpu_bo_print_info(int id,struct amdgpu_bo * bo,struct seq_file * m)1497 u64 amdgpu_bo_print_info(int id, struct amdgpu_bo *bo, struct seq_file *m)
1498 {
1499 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1500 struct dma_buf_attachment *attachment;
1501 struct dma_buf *dma_buf;
1502 const char *placement;
1503 unsigned int pin_count;
1504 u64 size;
1505
1506 if (dma_resv_trylock(bo->tbo.base.resv)) {
1507 if (!bo->tbo.resource) {
1508 placement = "NONE";
1509 } else {
1510 switch (bo->tbo.resource->mem_type) {
1511 case TTM_PL_VRAM:
1512 if (amdgpu_res_cpu_visible(adev, bo->tbo.resource))
1513 placement = "VRAM VISIBLE";
1514 else
1515 placement = "VRAM";
1516 break;
1517 case TTM_PL_TT:
1518 placement = "GTT";
1519 break;
1520 case AMDGPU_PL_GDS:
1521 placement = "GDS";
1522 break;
1523 case AMDGPU_PL_GWS:
1524 placement = "GWS";
1525 break;
1526 case AMDGPU_PL_OA:
1527 placement = "OA";
1528 break;
1529 case AMDGPU_PL_PREEMPT:
1530 placement = "PREEMPTIBLE";
1531 break;
1532 case AMDGPU_PL_DOORBELL:
1533 placement = "DOORBELL";
1534 break;
1535 case TTM_PL_SYSTEM:
1536 default:
1537 placement = "CPU";
1538 break;
1539 }
1540 }
1541 dma_resv_unlock(bo->tbo.base.resv);
1542 } else {
1543 placement = "UNKNOWN";
1544 }
1545
1546 size = amdgpu_bo_size(bo);
1547 seq_printf(m, "\t\t0x%08x: %12lld byte %s",
1548 id, size, placement);
1549
1550 pin_count = READ_ONCE(bo->tbo.pin_count);
1551 if (pin_count)
1552 seq_printf(m, " pin count %d", pin_count);
1553
1554 dma_buf = READ_ONCE(bo->tbo.base.dma_buf);
1555 attachment = READ_ONCE(bo->tbo.base.import_attach);
1556
1557 if (attachment)
1558 seq_printf(m, " imported from ino:%lu", file_inode(dma_buf->file)->i_ino);
1559 else if (dma_buf)
1560 seq_printf(m, " exported as ino:%lu", file_inode(dma_buf->file)->i_ino);
1561
1562 amdgpu_bo_print_flag(m, bo, CPU_ACCESS_REQUIRED);
1563 amdgpu_bo_print_flag(m, bo, NO_CPU_ACCESS);
1564 amdgpu_bo_print_flag(m, bo, CPU_GTT_USWC);
1565 amdgpu_bo_print_flag(m, bo, VRAM_CLEARED);
1566 amdgpu_bo_print_flag(m, bo, VRAM_CONTIGUOUS);
1567 amdgpu_bo_print_flag(m, bo, VM_ALWAYS_VALID);
1568 amdgpu_bo_print_flag(m, bo, EXPLICIT_SYNC);
1569
1570 seq_puts(m, "\n");
1571
1572 return size;
1573 }
1574 #endif
1575