Loading...
1/*
2 * Copyright 2009 Jerome Glisse.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sub license, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
12 *
13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19 * USE OR OTHER DEALINGS IN THE SOFTWARE.
20 *
21 * The above copyright notice and this permission notice (including the
22 * next paragraph) shall be included in all copies or substantial portions
23 * of the Software.
24 *
25 */
26/*
27 * Authors:
28 * Jerome Glisse <glisse@freedesktop.org>
29 * Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
30 * Dave Airlie
31 */
32#include <drm/ttm/ttm_bo_api.h>
33#include <drm/ttm/ttm_bo_driver.h>
34#include <drm/ttm/ttm_placement.h>
35#include <drm/ttm/ttm_module.h>
36#include <drm/ttm/ttm_page_alloc.h>
37#include <drm/drmP.h>
38#include <drm/amdgpu_drm.h>
39#include <linux/seq_file.h>
40#include <linux/slab.h>
41#include <linux/swiotlb.h>
42#include <linux/swap.h>
43#include <linux/pagemap.h>
44#include <linux/debugfs.h>
45#include <linux/iommu.h>
46#include "amdgpu.h"
47#include "amdgpu_object.h"
48#include "amdgpu_trace.h"
49#include "amdgpu_amdkfd.h"
50#include "bif/bif_4_1_d.h"
51
52#define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT)
53
54static int amdgpu_map_buffer(struct ttm_buffer_object *bo,
55 struct ttm_mem_reg *mem, unsigned num_pages,
56 uint64_t offset, unsigned window,
57 struct amdgpu_ring *ring,
58 uint64_t *addr);
59
60static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev);
61static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev);
62
63/*
64 * Global memory.
65 */
66static int amdgpu_ttm_mem_global_init(struct drm_global_reference *ref)
67{
68 return ttm_mem_global_init(ref->object);
69}
70
71static void amdgpu_ttm_mem_global_release(struct drm_global_reference *ref)
72{
73 ttm_mem_global_release(ref->object);
74}
75
76static int amdgpu_ttm_global_init(struct amdgpu_device *adev)
77{
78 struct drm_global_reference *global_ref;
79 struct amdgpu_ring *ring;
80 struct drm_sched_rq *rq;
81 int r;
82
83 adev->mman.mem_global_referenced = false;
84 global_ref = &adev->mman.mem_global_ref;
85 global_ref->global_type = DRM_GLOBAL_TTM_MEM;
86 global_ref->size = sizeof(struct ttm_mem_global);
87 global_ref->init = &amdgpu_ttm_mem_global_init;
88 global_ref->release = &amdgpu_ttm_mem_global_release;
89 r = drm_global_item_ref(global_ref);
90 if (r) {
91 DRM_ERROR("Failed setting up TTM memory accounting "
92 "subsystem.\n");
93 goto error_mem;
94 }
95
96 adev->mman.bo_global_ref.mem_glob =
97 adev->mman.mem_global_ref.object;
98 global_ref = &adev->mman.bo_global_ref.ref;
99 global_ref->global_type = DRM_GLOBAL_TTM_BO;
100 global_ref->size = sizeof(struct ttm_bo_global);
101 global_ref->init = &ttm_bo_global_init;
102 global_ref->release = &ttm_bo_global_release;
103 r = drm_global_item_ref(global_ref);
104 if (r) {
105 DRM_ERROR("Failed setting up TTM BO subsystem.\n");
106 goto error_bo;
107 }
108
109 mutex_init(&adev->mman.gtt_window_lock);
110
111 ring = adev->mman.buffer_funcs_ring;
112 rq = &ring->sched.sched_rq[DRM_SCHED_PRIORITY_KERNEL];
113 r = drm_sched_entity_init(&ring->sched, &adev->mman.entity,
114 rq, amdgpu_sched_jobs, NULL);
115 if (r) {
116 DRM_ERROR("Failed setting up TTM BO move run queue.\n");
117 goto error_entity;
118 }
119
120 adev->mman.mem_global_referenced = true;
121
122 return 0;
123
124error_entity:
125 drm_global_item_unref(&adev->mman.bo_global_ref.ref);
126error_bo:
127 drm_global_item_unref(&adev->mman.mem_global_ref);
128error_mem:
129 return r;
130}
131
132static void amdgpu_ttm_global_fini(struct amdgpu_device *adev)
133{
134 if (adev->mman.mem_global_referenced) {
135 drm_sched_entity_fini(adev->mman.entity.sched,
136 &adev->mman.entity);
137 mutex_destroy(&adev->mman.gtt_window_lock);
138 drm_global_item_unref(&adev->mman.bo_global_ref.ref);
139 drm_global_item_unref(&adev->mman.mem_global_ref);
140 adev->mman.mem_global_referenced = false;
141 }
142}
143
144static int amdgpu_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
145{
146 return 0;
147}
148
149static int amdgpu_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
150 struct ttm_mem_type_manager *man)
151{
152 struct amdgpu_device *adev;
153
154 adev = amdgpu_ttm_adev(bdev);
155
156 switch (type) {
157 case TTM_PL_SYSTEM:
158 /* System memory */
159 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
160 man->available_caching = TTM_PL_MASK_CACHING;
161 man->default_caching = TTM_PL_FLAG_CACHED;
162 break;
163 case TTM_PL_TT:
164 man->func = &amdgpu_gtt_mgr_func;
165 man->gpu_offset = adev->gmc.gart_start;
166 man->available_caching = TTM_PL_MASK_CACHING;
167 man->default_caching = TTM_PL_FLAG_CACHED;
168 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE | TTM_MEMTYPE_FLAG_CMA;
169 break;
170 case TTM_PL_VRAM:
171 /* "On-card" video ram */
172 man->func = &amdgpu_vram_mgr_func;
173 man->gpu_offset = adev->gmc.vram_start;
174 man->flags = TTM_MEMTYPE_FLAG_FIXED |
175 TTM_MEMTYPE_FLAG_MAPPABLE;
176 man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
177 man->default_caching = TTM_PL_FLAG_WC;
178 break;
179 case AMDGPU_PL_GDS:
180 case AMDGPU_PL_GWS:
181 case AMDGPU_PL_OA:
182 /* On-chip GDS memory*/
183 man->func = &ttm_bo_manager_func;
184 man->gpu_offset = 0;
185 man->flags = TTM_MEMTYPE_FLAG_FIXED | TTM_MEMTYPE_FLAG_CMA;
186 man->available_caching = TTM_PL_FLAG_UNCACHED;
187 man->default_caching = TTM_PL_FLAG_UNCACHED;
188 break;
189 default:
190 DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
191 return -EINVAL;
192 }
193 return 0;
194}
195
196static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
197 struct ttm_placement *placement)
198{
199 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
200 struct amdgpu_bo *abo;
201 static const struct ttm_place placements = {
202 .fpfn = 0,
203 .lpfn = 0,
204 .flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
205 };
206
207 if (bo->type == ttm_bo_type_sg) {
208 placement->num_placement = 0;
209 placement->num_busy_placement = 0;
210 return;
211 }
212
213 if (!amdgpu_ttm_bo_is_amdgpu_bo(bo)) {
214 placement->placement = &placements;
215 placement->busy_placement = &placements;
216 placement->num_placement = 1;
217 placement->num_busy_placement = 1;
218 return;
219 }
220 abo = ttm_to_amdgpu_bo(bo);
221 switch (bo->mem.mem_type) {
222 case TTM_PL_VRAM:
223 if (!adev->mman.buffer_funcs_enabled) {
224 amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
225 } else if (adev->gmc.visible_vram_size < adev->gmc.real_vram_size &&
226 !(abo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)) {
227 unsigned fpfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
228 struct drm_mm_node *node = bo->mem.mm_node;
229 unsigned long pages_left;
230
231 for (pages_left = bo->mem.num_pages;
232 pages_left;
233 pages_left -= node->size, node++) {
234 if (node->start < fpfn)
235 break;
236 }
237
238 if (!pages_left)
239 goto gtt;
240
241 /* Try evicting to the CPU inaccessible part of VRAM
242 * first, but only set GTT as busy placement, so this
243 * BO will be evicted to GTT rather than causing other
244 * BOs to be evicted from VRAM
245 */
246 amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
247 AMDGPU_GEM_DOMAIN_GTT);
248 abo->placements[0].fpfn = fpfn;
249 abo->placements[0].lpfn = 0;
250 abo->placement.busy_placement = &abo->placements[1];
251 abo->placement.num_busy_placement = 1;
252 } else {
253gtt:
254 amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_GTT);
255 }
256 break;
257 case TTM_PL_TT:
258 default:
259 amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
260 }
261 *placement = abo->placement;
262}
263
264static int amdgpu_verify_access(struct ttm_buffer_object *bo, struct file *filp)
265{
266 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
267
268 /*
269 * Don't verify access for KFD BOs. They don't have a GEM
270 * object associated with them.
271 */
272 if (abo->kfd_bo)
273 return 0;
274
275 if (amdgpu_ttm_tt_get_usermm(bo->ttm))
276 return -EPERM;
277 return drm_vma_node_verify_access(&abo->gem_base.vma_node,
278 filp->private_data);
279}
280
281static void amdgpu_move_null(struct ttm_buffer_object *bo,
282 struct ttm_mem_reg *new_mem)
283{
284 struct ttm_mem_reg *old_mem = &bo->mem;
285
286 BUG_ON(old_mem->mm_node != NULL);
287 *old_mem = *new_mem;
288 new_mem->mm_node = NULL;
289}
290
291static uint64_t amdgpu_mm_node_addr(struct ttm_buffer_object *bo,
292 struct drm_mm_node *mm_node,
293 struct ttm_mem_reg *mem)
294{
295 uint64_t addr = 0;
296
297 if (mem->mem_type != TTM_PL_TT || amdgpu_gtt_mgr_has_gart_addr(mem)) {
298 addr = mm_node->start << PAGE_SHIFT;
299 addr += bo->bdev->man[mem->mem_type].gpu_offset;
300 }
301 return addr;
302}
303
304/**
305 * amdgpu_find_mm_node - Helper function finds the drm_mm_node
306 * corresponding to @offset. It also modifies the offset to be
307 * within the drm_mm_node returned
308 */
309static struct drm_mm_node *amdgpu_find_mm_node(struct ttm_mem_reg *mem,
310 unsigned long *offset)
311{
312 struct drm_mm_node *mm_node = mem->mm_node;
313
314 while (*offset >= (mm_node->size << PAGE_SHIFT)) {
315 *offset -= (mm_node->size << PAGE_SHIFT);
316 ++mm_node;
317 }
318 return mm_node;
319}
320
321/**
322 * amdgpu_copy_ttm_mem_to_mem - Helper function for copy
323 *
324 * The function copies @size bytes from {src->mem + src->offset} to
325 * {dst->mem + dst->offset}. src->bo and dst->bo could be same BO for a
326 * move and different for a BO to BO copy.
327 *
328 * @f: Returns the last fence if multiple jobs are submitted.
329 */
330int amdgpu_ttm_copy_mem_to_mem(struct amdgpu_device *adev,
331 struct amdgpu_copy_mem *src,
332 struct amdgpu_copy_mem *dst,
333 uint64_t size,
334 struct reservation_object *resv,
335 struct dma_fence **f)
336{
337 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
338 struct drm_mm_node *src_mm, *dst_mm;
339 uint64_t src_node_start, dst_node_start, src_node_size,
340 dst_node_size, src_page_offset, dst_page_offset;
341 struct dma_fence *fence = NULL;
342 int r = 0;
343 const uint64_t GTT_MAX_BYTES = (AMDGPU_GTT_MAX_TRANSFER_SIZE *
344 AMDGPU_GPU_PAGE_SIZE);
345
346 if (!adev->mman.buffer_funcs_enabled) {
347 DRM_ERROR("Trying to move memory with ring turned off.\n");
348 return -EINVAL;
349 }
350
351 src_mm = amdgpu_find_mm_node(src->mem, &src->offset);
352 src_node_start = amdgpu_mm_node_addr(src->bo, src_mm, src->mem) +
353 src->offset;
354 src_node_size = (src_mm->size << PAGE_SHIFT) - src->offset;
355 src_page_offset = src_node_start & (PAGE_SIZE - 1);
356
357 dst_mm = amdgpu_find_mm_node(dst->mem, &dst->offset);
358 dst_node_start = amdgpu_mm_node_addr(dst->bo, dst_mm, dst->mem) +
359 dst->offset;
360 dst_node_size = (dst_mm->size << PAGE_SHIFT) - dst->offset;
361 dst_page_offset = dst_node_start & (PAGE_SIZE - 1);
362
363 mutex_lock(&adev->mman.gtt_window_lock);
364
365 while (size) {
366 unsigned long cur_size;
367 uint64_t from = src_node_start, to = dst_node_start;
368 struct dma_fence *next;
369
370 /* Copy size cannot exceed GTT_MAX_BYTES. So if src or dst
371 * begins at an offset, then adjust the size accordingly
372 */
373 cur_size = min3(min(src_node_size, dst_node_size), size,
374 GTT_MAX_BYTES);
375 if (cur_size + src_page_offset > GTT_MAX_BYTES ||
376 cur_size + dst_page_offset > GTT_MAX_BYTES)
377 cur_size -= max(src_page_offset, dst_page_offset);
378
379 /* Map only what needs to be accessed. Map src to window 0 and
380 * dst to window 1
381 */
382 if (src->mem->mem_type == TTM_PL_TT &&
383 !amdgpu_gtt_mgr_has_gart_addr(src->mem)) {
384 r = amdgpu_map_buffer(src->bo, src->mem,
385 PFN_UP(cur_size + src_page_offset),
386 src_node_start, 0, ring,
387 &from);
388 if (r)
389 goto error;
390 /* Adjust the offset because amdgpu_map_buffer returns
391 * start of mapped page
392 */
393 from += src_page_offset;
394 }
395
396 if (dst->mem->mem_type == TTM_PL_TT &&
397 !amdgpu_gtt_mgr_has_gart_addr(dst->mem)) {
398 r = amdgpu_map_buffer(dst->bo, dst->mem,
399 PFN_UP(cur_size + dst_page_offset),
400 dst_node_start, 1, ring,
401 &to);
402 if (r)
403 goto error;
404 to += dst_page_offset;
405 }
406
407 r = amdgpu_copy_buffer(ring, from, to, cur_size,
408 resv, &next, false, true);
409 if (r)
410 goto error;
411
412 dma_fence_put(fence);
413 fence = next;
414
415 size -= cur_size;
416 if (!size)
417 break;
418
419 src_node_size -= cur_size;
420 if (!src_node_size) {
421 src_node_start = amdgpu_mm_node_addr(src->bo, ++src_mm,
422 src->mem);
423 src_node_size = (src_mm->size << PAGE_SHIFT);
424 } else {
425 src_node_start += cur_size;
426 src_page_offset = src_node_start & (PAGE_SIZE - 1);
427 }
428 dst_node_size -= cur_size;
429 if (!dst_node_size) {
430 dst_node_start = amdgpu_mm_node_addr(dst->bo, ++dst_mm,
431 dst->mem);
432 dst_node_size = (dst_mm->size << PAGE_SHIFT);
433 } else {
434 dst_node_start += cur_size;
435 dst_page_offset = dst_node_start & (PAGE_SIZE - 1);
436 }
437 }
438error:
439 mutex_unlock(&adev->mman.gtt_window_lock);
440 if (f)
441 *f = dma_fence_get(fence);
442 dma_fence_put(fence);
443 return r;
444}
445
446
447static int amdgpu_move_blit(struct ttm_buffer_object *bo,
448 bool evict, bool no_wait_gpu,
449 struct ttm_mem_reg *new_mem,
450 struct ttm_mem_reg *old_mem)
451{
452 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
453 struct amdgpu_copy_mem src, dst;
454 struct dma_fence *fence = NULL;
455 int r;
456
457 src.bo = bo;
458 dst.bo = bo;
459 src.mem = old_mem;
460 dst.mem = new_mem;
461 src.offset = 0;
462 dst.offset = 0;
463
464 r = amdgpu_ttm_copy_mem_to_mem(adev, &src, &dst,
465 new_mem->num_pages << PAGE_SHIFT,
466 bo->resv, &fence);
467 if (r)
468 goto error;
469
470 r = ttm_bo_pipeline_move(bo, fence, evict, new_mem);
471 dma_fence_put(fence);
472 return r;
473
474error:
475 if (fence)
476 dma_fence_wait(fence, false);
477 dma_fence_put(fence);
478 return r;
479}
480
481static int amdgpu_move_vram_ram(struct ttm_buffer_object *bo, bool evict,
482 struct ttm_operation_ctx *ctx,
483 struct ttm_mem_reg *new_mem)
484{
485 struct amdgpu_device *adev;
486 struct ttm_mem_reg *old_mem = &bo->mem;
487 struct ttm_mem_reg tmp_mem;
488 struct ttm_place placements;
489 struct ttm_placement placement;
490 int r;
491
492 adev = amdgpu_ttm_adev(bo->bdev);
493 tmp_mem = *new_mem;
494 tmp_mem.mm_node = NULL;
495 placement.num_placement = 1;
496 placement.placement = &placements;
497 placement.num_busy_placement = 1;
498 placement.busy_placement = &placements;
499 placements.fpfn = 0;
500 placements.lpfn = 0;
501 placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
502 r = ttm_bo_mem_space(bo, &placement, &tmp_mem, ctx);
503 if (unlikely(r)) {
504 return r;
505 }
506
507 r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
508 if (unlikely(r)) {
509 goto out_cleanup;
510 }
511
512 r = ttm_tt_bind(bo->ttm, &tmp_mem, ctx);
513 if (unlikely(r)) {
514 goto out_cleanup;
515 }
516 r = amdgpu_move_blit(bo, true, ctx->no_wait_gpu, &tmp_mem, old_mem);
517 if (unlikely(r)) {
518 goto out_cleanup;
519 }
520 r = ttm_bo_move_ttm(bo, ctx, new_mem);
521out_cleanup:
522 ttm_bo_mem_put(bo, &tmp_mem);
523 return r;
524}
525
526static int amdgpu_move_ram_vram(struct ttm_buffer_object *bo, bool evict,
527 struct ttm_operation_ctx *ctx,
528 struct ttm_mem_reg *new_mem)
529{
530 struct amdgpu_device *adev;
531 struct ttm_mem_reg *old_mem = &bo->mem;
532 struct ttm_mem_reg tmp_mem;
533 struct ttm_placement placement;
534 struct ttm_place placements;
535 int r;
536
537 adev = amdgpu_ttm_adev(bo->bdev);
538 tmp_mem = *new_mem;
539 tmp_mem.mm_node = NULL;
540 placement.num_placement = 1;
541 placement.placement = &placements;
542 placement.num_busy_placement = 1;
543 placement.busy_placement = &placements;
544 placements.fpfn = 0;
545 placements.lpfn = 0;
546 placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
547 r = ttm_bo_mem_space(bo, &placement, &tmp_mem, ctx);
548 if (unlikely(r)) {
549 return r;
550 }
551 r = ttm_bo_move_ttm(bo, ctx, &tmp_mem);
552 if (unlikely(r)) {
553 goto out_cleanup;
554 }
555 r = amdgpu_move_blit(bo, true, ctx->no_wait_gpu, new_mem, old_mem);
556 if (unlikely(r)) {
557 goto out_cleanup;
558 }
559out_cleanup:
560 ttm_bo_mem_put(bo, &tmp_mem);
561 return r;
562}
563
564static int amdgpu_bo_move(struct ttm_buffer_object *bo, bool evict,
565 struct ttm_operation_ctx *ctx,
566 struct ttm_mem_reg *new_mem)
567{
568 struct amdgpu_device *adev;
569 struct amdgpu_bo *abo;
570 struct ttm_mem_reg *old_mem = &bo->mem;
571 int r;
572
573 /* Can't move a pinned BO */
574 abo = ttm_to_amdgpu_bo(bo);
575 if (WARN_ON_ONCE(abo->pin_count > 0))
576 return -EINVAL;
577
578 adev = amdgpu_ttm_adev(bo->bdev);
579
580 if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
581 amdgpu_move_null(bo, new_mem);
582 return 0;
583 }
584 if ((old_mem->mem_type == TTM_PL_TT &&
585 new_mem->mem_type == TTM_PL_SYSTEM) ||
586 (old_mem->mem_type == TTM_PL_SYSTEM &&
587 new_mem->mem_type == TTM_PL_TT)) {
588 /* bind is enough */
589 amdgpu_move_null(bo, new_mem);
590 return 0;
591 }
592
593 if (!adev->mman.buffer_funcs_enabled)
594 goto memcpy;
595
596 if (old_mem->mem_type == TTM_PL_VRAM &&
597 new_mem->mem_type == TTM_PL_SYSTEM) {
598 r = amdgpu_move_vram_ram(bo, evict, ctx, new_mem);
599 } else if (old_mem->mem_type == TTM_PL_SYSTEM &&
600 new_mem->mem_type == TTM_PL_VRAM) {
601 r = amdgpu_move_ram_vram(bo, evict, ctx, new_mem);
602 } else {
603 r = amdgpu_move_blit(bo, evict, ctx->no_wait_gpu,
604 new_mem, old_mem);
605 }
606
607 if (r) {
608memcpy:
609 r = ttm_bo_move_memcpy(bo, ctx, new_mem);
610 if (r) {
611 return r;
612 }
613 }
614
615 if (bo->type == ttm_bo_type_device &&
616 new_mem->mem_type == TTM_PL_VRAM &&
617 old_mem->mem_type != TTM_PL_VRAM) {
618 /* amdgpu_bo_fault_reserve_notify will re-set this if the CPU
619 * accesses the BO after it's moved.
620 */
621 abo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
622 }
623
624 /* update statistics */
625 atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &adev->num_bytes_moved);
626 return 0;
627}
628
629static int amdgpu_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
630{
631 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
632 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
633 struct drm_mm_node *mm_node = mem->mm_node;
634
635 mem->bus.addr = NULL;
636 mem->bus.offset = 0;
637 mem->bus.size = mem->num_pages << PAGE_SHIFT;
638 mem->bus.base = 0;
639 mem->bus.is_iomem = false;
640 if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
641 return -EINVAL;
642 switch (mem->mem_type) {
643 case TTM_PL_SYSTEM:
644 /* system memory */
645 return 0;
646 case TTM_PL_TT:
647 break;
648 case TTM_PL_VRAM:
649 mem->bus.offset = mem->start << PAGE_SHIFT;
650 /* check if it's visible */
651 if ((mem->bus.offset + mem->bus.size) > adev->gmc.visible_vram_size)
652 return -EINVAL;
653 /* Only physically contiguous buffers apply. In a contiguous
654 * buffer, size of the first mm_node would match the number of
655 * pages in ttm_mem_reg.
656 */
657 if (adev->mman.aper_base_kaddr &&
658 (mm_node->size == mem->num_pages))
659 mem->bus.addr = (u8 *)adev->mman.aper_base_kaddr +
660 mem->bus.offset;
661
662 mem->bus.base = adev->gmc.aper_base;
663 mem->bus.is_iomem = true;
664 break;
665 default:
666 return -EINVAL;
667 }
668 return 0;
669}
670
671static void amdgpu_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
672{
673}
674
675static unsigned long amdgpu_ttm_io_mem_pfn(struct ttm_buffer_object *bo,
676 unsigned long page_offset)
677{
678 struct drm_mm_node *mm;
679 unsigned long offset = (page_offset << PAGE_SHIFT);
680
681 mm = amdgpu_find_mm_node(&bo->mem, &offset);
682 return (bo->mem.bus.base >> PAGE_SHIFT) + mm->start +
683 (offset >> PAGE_SHIFT);
684}
685
686/*
687 * TTM backend functions.
688 */
689struct amdgpu_ttm_gup_task_list {
690 struct list_head list;
691 struct task_struct *task;
692};
693
694struct amdgpu_ttm_tt {
695 struct ttm_dma_tt ttm;
696 u64 offset;
697 uint64_t userptr;
698 struct mm_struct *usermm;
699 uint32_t userflags;
700 spinlock_t guptasklock;
701 struct list_head guptasks;
702 atomic_t mmu_invalidations;
703 uint32_t last_set_pages;
704};
705
706int amdgpu_ttm_tt_get_user_pages(struct ttm_tt *ttm, struct page **pages)
707{
708 struct amdgpu_ttm_tt *gtt = (void *)ttm;
709 unsigned int flags = 0;
710 unsigned pinned = 0;
711 int r;
712
713 if (!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY))
714 flags |= FOLL_WRITE;
715
716 down_read(¤t->mm->mmap_sem);
717
718 if (gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) {
719 /* check that we only use anonymous memory
720 to prevent problems with writeback */
721 unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE;
722 struct vm_area_struct *vma;
723
724 vma = find_vma(gtt->usermm, gtt->userptr);
725 if (!vma || vma->vm_file || vma->vm_end < end) {
726 up_read(¤t->mm->mmap_sem);
727 return -EPERM;
728 }
729 }
730
731 do {
732 unsigned num_pages = ttm->num_pages - pinned;
733 uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
734 struct page **p = pages + pinned;
735 struct amdgpu_ttm_gup_task_list guptask;
736
737 guptask.task = current;
738 spin_lock(>t->guptasklock);
739 list_add(&guptask.list, >t->guptasks);
740 spin_unlock(>t->guptasklock);
741
742 r = get_user_pages(userptr, num_pages, flags, p, NULL);
743
744 spin_lock(>t->guptasklock);
745 list_del(&guptask.list);
746 spin_unlock(>t->guptasklock);
747
748 if (r < 0)
749 goto release_pages;
750
751 pinned += r;
752
753 } while (pinned < ttm->num_pages);
754
755 up_read(¤t->mm->mmap_sem);
756 return 0;
757
758release_pages:
759 release_pages(pages, pinned);
760 up_read(¤t->mm->mmap_sem);
761 return r;
762}
763
764void amdgpu_ttm_tt_set_user_pages(struct ttm_tt *ttm, struct page **pages)
765{
766 struct amdgpu_ttm_tt *gtt = (void *)ttm;
767 unsigned i;
768
769 gtt->last_set_pages = atomic_read(>t->mmu_invalidations);
770 for (i = 0; i < ttm->num_pages; ++i) {
771 if (ttm->pages[i])
772 put_page(ttm->pages[i]);
773
774 ttm->pages[i] = pages ? pages[i] : NULL;
775 }
776}
777
778void amdgpu_ttm_tt_mark_user_pages(struct ttm_tt *ttm)
779{
780 struct amdgpu_ttm_tt *gtt = (void *)ttm;
781 unsigned i;
782
783 for (i = 0; i < ttm->num_pages; ++i) {
784 struct page *page = ttm->pages[i];
785
786 if (!page)
787 continue;
788
789 if (!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY))
790 set_page_dirty(page);
791
792 mark_page_accessed(page);
793 }
794}
795
796/* prepare the sg table with the user pages */
797static int amdgpu_ttm_tt_pin_userptr(struct ttm_tt *ttm)
798{
799 struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
800 struct amdgpu_ttm_tt *gtt = (void *)ttm;
801 unsigned nents;
802 int r;
803
804 int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
805 enum dma_data_direction direction = write ?
806 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
807
808 r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
809 ttm->num_pages << PAGE_SHIFT,
810 GFP_KERNEL);
811 if (r)
812 goto release_sg;
813
814 r = -ENOMEM;
815 nents = dma_map_sg(adev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
816 if (nents != ttm->sg->nents)
817 goto release_sg;
818
819 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
820 gtt->ttm.dma_address, ttm->num_pages);
821
822 return 0;
823
824release_sg:
825 kfree(ttm->sg);
826 return r;
827}
828
829static void amdgpu_ttm_tt_unpin_userptr(struct ttm_tt *ttm)
830{
831 struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
832 struct amdgpu_ttm_tt *gtt = (void *)ttm;
833
834 int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
835 enum dma_data_direction direction = write ?
836 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
837
838 /* double check that we don't free the table twice */
839 if (!ttm->sg->sgl)
840 return;
841
842 /* free the sg table and pages again */
843 dma_unmap_sg(adev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
844
845 amdgpu_ttm_tt_mark_user_pages(ttm);
846
847 sg_free_table(ttm->sg);
848}
849
850static int amdgpu_ttm_backend_bind(struct ttm_tt *ttm,
851 struct ttm_mem_reg *bo_mem)
852{
853 struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
854 struct amdgpu_ttm_tt *gtt = (void*)ttm;
855 uint64_t flags;
856 int r = 0;
857
858 if (gtt->userptr) {
859 r = amdgpu_ttm_tt_pin_userptr(ttm);
860 if (r) {
861 DRM_ERROR("failed to pin userptr\n");
862 return r;
863 }
864 }
865 if (!ttm->num_pages) {
866 WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
867 ttm->num_pages, bo_mem, ttm);
868 }
869
870 if (bo_mem->mem_type == AMDGPU_PL_GDS ||
871 bo_mem->mem_type == AMDGPU_PL_GWS ||
872 bo_mem->mem_type == AMDGPU_PL_OA)
873 return -EINVAL;
874
875 if (!amdgpu_gtt_mgr_has_gart_addr(bo_mem)) {
876 gtt->offset = AMDGPU_BO_INVALID_OFFSET;
877 return 0;
878 }
879
880 flags = amdgpu_ttm_tt_pte_flags(adev, ttm, bo_mem);
881 gtt->offset = (u64)bo_mem->start << PAGE_SHIFT;
882 r = amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
883 ttm->pages, gtt->ttm.dma_address, flags);
884
885 if (r)
886 DRM_ERROR("failed to bind %lu pages at 0x%08llX\n",
887 ttm->num_pages, gtt->offset);
888 return r;
889}
890
891int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo)
892{
893 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
894 struct ttm_operation_ctx ctx = { false, false };
895 struct amdgpu_ttm_tt *gtt = (void*)bo->ttm;
896 struct ttm_mem_reg tmp;
897 struct ttm_placement placement;
898 struct ttm_place placements;
899 uint64_t flags;
900 int r;
901
902 if (bo->mem.mem_type != TTM_PL_TT ||
903 amdgpu_gtt_mgr_has_gart_addr(&bo->mem))
904 return 0;
905
906 tmp = bo->mem;
907 tmp.mm_node = NULL;
908 placement.num_placement = 1;
909 placement.placement = &placements;
910 placement.num_busy_placement = 1;
911 placement.busy_placement = &placements;
912 placements.fpfn = 0;
913 placements.lpfn = adev->gmc.gart_size >> PAGE_SHIFT;
914 placements.flags = (bo->mem.placement & ~TTM_PL_MASK_MEM) |
915 TTM_PL_FLAG_TT;
916
917 r = ttm_bo_mem_space(bo, &placement, &tmp, &ctx);
918 if (unlikely(r))
919 return r;
920
921 flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, &tmp);
922 gtt->offset = (u64)tmp.start << PAGE_SHIFT;
923 r = amdgpu_gart_bind(adev, gtt->offset, bo->ttm->num_pages,
924 bo->ttm->pages, gtt->ttm.dma_address, flags);
925 if (unlikely(r)) {
926 ttm_bo_mem_put(bo, &tmp);
927 return r;
928 }
929
930 ttm_bo_mem_put(bo, &bo->mem);
931 bo->mem = tmp;
932 bo->offset = (bo->mem.start << PAGE_SHIFT) +
933 bo->bdev->man[bo->mem.mem_type].gpu_offset;
934
935 return 0;
936}
937
938int amdgpu_ttm_recover_gart(struct ttm_buffer_object *tbo)
939{
940 struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
941 struct amdgpu_ttm_tt *gtt = (void *)tbo->ttm;
942 uint64_t flags;
943 int r;
944
945 if (!gtt)
946 return 0;
947
948 flags = amdgpu_ttm_tt_pte_flags(adev, >t->ttm.ttm, &tbo->mem);
949 r = amdgpu_gart_bind(adev, gtt->offset, gtt->ttm.ttm.num_pages,
950 gtt->ttm.ttm.pages, gtt->ttm.dma_address, flags);
951 if (r)
952 DRM_ERROR("failed to bind %lu pages at 0x%08llX\n",
953 gtt->ttm.ttm.num_pages, gtt->offset);
954 return r;
955}
956
957static int amdgpu_ttm_backend_unbind(struct ttm_tt *ttm)
958{
959 struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
960 struct amdgpu_ttm_tt *gtt = (void *)ttm;
961 int r;
962
963 if (gtt->userptr)
964 amdgpu_ttm_tt_unpin_userptr(ttm);
965
966 if (gtt->offset == AMDGPU_BO_INVALID_OFFSET)
967 return 0;
968
969 /* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */
970 r = amdgpu_gart_unbind(adev, gtt->offset, ttm->num_pages);
971 if (r)
972 DRM_ERROR("failed to unbind %lu pages at 0x%08llX\n",
973 gtt->ttm.ttm.num_pages, gtt->offset);
974 return r;
975}
976
977static void amdgpu_ttm_backend_destroy(struct ttm_tt *ttm)
978{
979 struct amdgpu_ttm_tt *gtt = (void *)ttm;
980
981 ttm_dma_tt_fini(>t->ttm);
982 kfree(gtt);
983}
984
985static struct ttm_backend_func amdgpu_backend_func = {
986 .bind = &amdgpu_ttm_backend_bind,
987 .unbind = &amdgpu_ttm_backend_unbind,
988 .destroy = &amdgpu_ttm_backend_destroy,
989};
990
991static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_buffer_object *bo,
992 uint32_t page_flags)
993{
994 struct amdgpu_device *adev;
995 struct amdgpu_ttm_tt *gtt;
996
997 adev = amdgpu_ttm_adev(bo->bdev);
998
999 gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL);
1000 if (gtt == NULL) {
1001 return NULL;
1002 }
1003 gtt->ttm.ttm.func = &amdgpu_backend_func;
1004 if (ttm_sg_tt_init(>t->ttm, bo, page_flags)) {
1005 kfree(gtt);
1006 return NULL;
1007 }
1008 return >t->ttm.ttm;
1009}
1010
1011static int amdgpu_ttm_tt_populate(struct ttm_tt *ttm,
1012 struct ttm_operation_ctx *ctx)
1013{
1014 struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
1015 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1016 bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1017
1018 if (gtt && gtt->userptr) {
1019 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
1020 if (!ttm->sg)
1021 return -ENOMEM;
1022
1023 ttm->page_flags |= TTM_PAGE_FLAG_SG;
1024 ttm->state = tt_unbound;
1025 return 0;
1026 }
1027
1028 if (slave && ttm->sg) {
1029 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
1030 gtt->ttm.dma_address,
1031 ttm->num_pages);
1032 ttm->state = tt_unbound;
1033 return 0;
1034 }
1035
1036#ifdef CONFIG_SWIOTLB
1037 if (adev->need_swiotlb && swiotlb_nr_tbl()) {
1038 return ttm_dma_populate(>t->ttm, adev->dev, ctx);
1039 }
1040#endif
1041
1042 return ttm_populate_and_map_pages(adev->dev, >t->ttm, ctx);
1043}
1044
1045static void amdgpu_ttm_tt_unpopulate(struct ttm_tt *ttm)
1046{
1047 struct amdgpu_device *adev;
1048 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1049 bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1050
1051 if (gtt && gtt->userptr) {
1052 amdgpu_ttm_tt_set_user_pages(ttm, NULL);
1053 kfree(ttm->sg);
1054 ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
1055 return;
1056 }
1057
1058 if (slave)
1059 return;
1060
1061 adev = amdgpu_ttm_adev(ttm->bdev);
1062
1063#ifdef CONFIG_SWIOTLB
1064 if (adev->need_swiotlb && swiotlb_nr_tbl()) {
1065 ttm_dma_unpopulate(>t->ttm, adev->dev);
1066 return;
1067 }
1068#endif
1069
1070 ttm_unmap_and_unpopulate_pages(adev->dev, >t->ttm);
1071}
1072
1073int amdgpu_ttm_tt_set_userptr(struct ttm_tt *ttm, uint64_t addr,
1074 uint32_t flags)
1075{
1076 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1077
1078 if (gtt == NULL)
1079 return -EINVAL;
1080
1081 gtt->userptr = addr;
1082 gtt->usermm = current->mm;
1083 gtt->userflags = flags;
1084 spin_lock_init(>t->guptasklock);
1085 INIT_LIST_HEAD(>t->guptasks);
1086 atomic_set(>t->mmu_invalidations, 0);
1087 gtt->last_set_pages = 0;
1088
1089 return 0;
1090}
1091
1092struct mm_struct *amdgpu_ttm_tt_get_usermm(struct ttm_tt *ttm)
1093{
1094 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1095
1096 if (gtt == NULL)
1097 return NULL;
1098
1099 return gtt->usermm;
1100}
1101
1102bool amdgpu_ttm_tt_affect_userptr(struct ttm_tt *ttm, unsigned long start,
1103 unsigned long end)
1104{
1105 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1106 struct amdgpu_ttm_gup_task_list *entry;
1107 unsigned long size;
1108
1109 if (gtt == NULL || !gtt->userptr)
1110 return false;
1111
1112 size = (unsigned long)gtt->ttm.ttm.num_pages * PAGE_SIZE;
1113 if (gtt->userptr > end || gtt->userptr + size <= start)
1114 return false;
1115
1116 spin_lock(>t->guptasklock);
1117 list_for_each_entry(entry, >t->guptasks, list) {
1118 if (entry->task == current) {
1119 spin_unlock(>t->guptasklock);
1120 return false;
1121 }
1122 }
1123 spin_unlock(>t->guptasklock);
1124
1125 atomic_inc(>t->mmu_invalidations);
1126
1127 return true;
1128}
1129
1130bool amdgpu_ttm_tt_userptr_invalidated(struct ttm_tt *ttm,
1131 int *last_invalidated)
1132{
1133 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1134 int prev_invalidated = *last_invalidated;
1135
1136 *last_invalidated = atomic_read(>t->mmu_invalidations);
1137 return prev_invalidated != *last_invalidated;
1138}
1139
1140bool amdgpu_ttm_tt_userptr_needs_pages(struct ttm_tt *ttm)
1141{
1142 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1143
1144 if (gtt == NULL || !gtt->userptr)
1145 return false;
1146
1147 return atomic_read(>t->mmu_invalidations) != gtt->last_set_pages;
1148}
1149
1150bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm)
1151{
1152 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1153
1154 if (gtt == NULL)
1155 return false;
1156
1157 return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
1158}
1159
1160uint64_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
1161 struct ttm_mem_reg *mem)
1162{
1163 uint64_t flags = 0;
1164
1165 if (mem && mem->mem_type != TTM_PL_SYSTEM)
1166 flags |= AMDGPU_PTE_VALID;
1167
1168 if (mem && mem->mem_type == TTM_PL_TT) {
1169 flags |= AMDGPU_PTE_SYSTEM;
1170
1171 if (ttm->caching_state == tt_cached)
1172 flags |= AMDGPU_PTE_SNOOPED;
1173 }
1174
1175 flags |= adev->gart.gart_pte_flags;
1176 flags |= AMDGPU_PTE_READABLE;
1177
1178 if (!amdgpu_ttm_tt_is_readonly(ttm))
1179 flags |= AMDGPU_PTE_WRITEABLE;
1180
1181 return flags;
1182}
1183
1184static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
1185 const struct ttm_place *place)
1186{
1187 unsigned long num_pages = bo->mem.num_pages;
1188 struct drm_mm_node *node = bo->mem.mm_node;
1189 struct reservation_object_list *flist;
1190 struct dma_fence *f;
1191 int i;
1192
1193 /* If bo is a KFD BO, check if the bo belongs to the current process.
1194 * If true, then return false as any KFD process needs all its BOs to
1195 * be resident to run successfully
1196 */
1197 flist = reservation_object_get_list(bo->resv);
1198 if (flist) {
1199 for (i = 0; i < flist->shared_count; ++i) {
1200 f = rcu_dereference_protected(flist->shared[i],
1201 reservation_object_held(bo->resv));
1202 if (amdkfd_fence_check_mm(f, current->mm))
1203 return false;
1204 }
1205 }
1206
1207 switch (bo->mem.mem_type) {
1208 case TTM_PL_TT:
1209 return true;
1210
1211 case TTM_PL_VRAM:
1212 /* Check each drm MM node individually */
1213 while (num_pages) {
1214 if (place->fpfn < (node->start + node->size) &&
1215 !(place->lpfn && place->lpfn <= node->start))
1216 return true;
1217
1218 num_pages -= node->size;
1219 ++node;
1220 }
1221 return false;
1222
1223 default:
1224 break;
1225 }
1226
1227 return ttm_bo_eviction_valuable(bo, place);
1228}
1229
1230static int amdgpu_ttm_access_memory(struct ttm_buffer_object *bo,
1231 unsigned long offset,
1232 void *buf, int len, int write)
1233{
1234 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1235 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
1236 struct drm_mm_node *nodes;
1237 uint32_t value = 0;
1238 int ret = 0;
1239 uint64_t pos;
1240 unsigned long flags;
1241
1242 if (bo->mem.mem_type != TTM_PL_VRAM)
1243 return -EIO;
1244
1245 nodes = amdgpu_find_mm_node(&abo->tbo.mem, &offset);
1246 pos = (nodes->start << PAGE_SHIFT) + offset;
1247
1248 while (len && pos < adev->gmc.mc_vram_size) {
1249 uint64_t aligned_pos = pos & ~(uint64_t)3;
1250 uint32_t bytes = 4 - (pos & 3);
1251 uint32_t shift = (pos & 3) * 8;
1252 uint32_t mask = 0xffffffff << shift;
1253
1254 if (len < bytes) {
1255 mask &= 0xffffffff >> (bytes - len) * 8;
1256 bytes = len;
1257 }
1258
1259 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
1260 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)aligned_pos) | 0x80000000);
1261 WREG32_NO_KIQ(mmMM_INDEX_HI, aligned_pos >> 31);
1262 if (!write || mask != 0xffffffff)
1263 value = RREG32_NO_KIQ(mmMM_DATA);
1264 if (write) {
1265 value &= ~mask;
1266 value |= (*(uint32_t *)buf << shift) & mask;
1267 WREG32_NO_KIQ(mmMM_DATA, value);
1268 }
1269 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
1270 if (!write) {
1271 value = (value & mask) >> shift;
1272 memcpy(buf, &value, bytes);
1273 }
1274
1275 ret += bytes;
1276 buf = (uint8_t *)buf + bytes;
1277 pos += bytes;
1278 len -= bytes;
1279 if (pos >= (nodes->start + nodes->size) << PAGE_SHIFT) {
1280 ++nodes;
1281 pos = (nodes->start << PAGE_SHIFT);
1282 }
1283 }
1284
1285 return ret;
1286}
1287
1288static struct ttm_bo_driver amdgpu_bo_driver = {
1289 .ttm_tt_create = &amdgpu_ttm_tt_create,
1290 .ttm_tt_populate = &amdgpu_ttm_tt_populate,
1291 .ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate,
1292 .invalidate_caches = &amdgpu_invalidate_caches,
1293 .init_mem_type = &amdgpu_init_mem_type,
1294 .eviction_valuable = amdgpu_ttm_bo_eviction_valuable,
1295 .evict_flags = &amdgpu_evict_flags,
1296 .move = &amdgpu_bo_move,
1297 .verify_access = &amdgpu_verify_access,
1298 .move_notify = &amdgpu_bo_move_notify,
1299 .fault_reserve_notify = &amdgpu_bo_fault_reserve_notify,
1300 .io_mem_reserve = &amdgpu_ttm_io_mem_reserve,
1301 .io_mem_free = &amdgpu_ttm_io_mem_free,
1302 .io_mem_pfn = amdgpu_ttm_io_mem_pfn,
1303 .access_memory = &amdgpu_ttm_access_memory
1304};
1305
1306/*
1307 * Firmware Reservation functions
1308 */
1309/**
1310 * amdgpu_ttm_fw_reserve_vram_fini - free fw reserved vram
1311 *
1312 * @adev: amdgpu_device pointer
1313 *
1314 * free fw reserved vram if it has been reserved.
1315 */
1316static void amdgpu_ttm_fw_reserve_vram_fini(struct amdgpu_device *adev)
1317{
1318 amdgpu_bo_free_kernel(&adev->fw_vram_usage.reserved_bo,
1319 NULL, &adev->fw_vram_usage.va);
1320}
1321
1322/**
1323 * amdgpu_ttm_fw_reserve_vram_init - create bo vram reservation from fw
1324 *
1325 * @adev: amdgpu_device pointer
1326 *
1327 * create bo vram reservation from fw.
1328 */
1329static int amdgpu_ttm_fw_reserve_vram_init(struct amdgpu_device *adev)
1330{
1331 struct ttm_operation_ctx ctx = { false, false };
1332 int r = 0;
1333 int i;
1334 u64 vram_size = adev->gmc.visible_vram_size;
1335 u64 offset = adev->fw_vram_usage.start_offset;
1336 u64 size = adev->fw_vram_usage.size;
1337 struct amdgpu_bo *bo;
1338
1339 adev->fw_vram_usage.va = NULL;
1340 adev->fw_vram_usage.reserved_bo = NULL;
1341
1342 if (adev->fw_vram_usage.size > 0 &&
1343 adev->fw_vram_usage.size <= vram_size) {
1344
1345 r = amdgpu_bo_create(adev, adev->fw_vram_usage.size, PAGE_SIZE,
1346 AMDGPU_GEM_DOMAIN_VRAM,
1347 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
1348 AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS,
1349 ttm_bo_type_kernel, NULL,
1350 &adev->fw_vram_usage.reserved_bo);
1351 if (r)
1352 goto error_create;
1353
1354 r = amdgpu_bo_reserve(adev->fw_vram_usage.reserved_bo, false);
1355 if (r)
1356 goto error_reserve;
1357
1358 /* remove the original mem node and create a new one at the
1359 * request position
1360 */
1361 bo = adev->fw_vram_usage.reserved_bo;
1362 offset = ALIGN(offset, PAGE_SIZE);
1363 for (i = 0; i < bo->placement.num_placement; ++i) {
1364 bo->placements[i].fpfn = offset >> PAGE_SHIFT;
1365 bo->placements[i].lpfn = (offset + size) >> PAGE_SHIFT;
1366 }
1367
1368 ttm_bo_mem_put(&bo->tbo, &bo->tbo.mem);
1369 r = ttm_bo_mem_space(&bo->tbo, &bo->placement,
1370 &bo->tbo.mem, &ctx);
1371 if (r)
1372 goto error_pin;
1373
1374 r = amdgpu_bo_pin_restricted(adev->fw_vram_usage.reserved_bo,
1375 AMDGPU_GEM_DOMAIN_VRAM,
1376 adev->fw_vram_usage.start_offset,
1377 (adev->fw_vram_usage.start_offset +
1378 adev->fw_vram_usage.size), NULL);
1379 if (r)
1380 goto error_pin;
1381 r = amdgpu_bo_kmap(adev->fw_vram_usage.reserved_bo,
1382 &adev->fw_vram_usage.va);
1383 if (r)
1384 goto error_kmap;
1385
1386 amdgpu_bo_unreserve(adev->fw_vram_usage.reserved_bo);
1387 }
1388 return r;
1389
1390error_kmap:
1391 amdgpu_bo_unpin(adev->fw_vram_usage.reserved_bo);
1392error_pin:
1393 amdgpu_bo_unreserve(adev->fw_vram_usage.reserved_bo);
1394error_reserve:
1395 amdgpu_bo_unref(&adev->fw_vram_usage.reserved_bo);
1396error_create:
1397 adev->fw_vram_usage.va = NULL;
1398 adev->fw_vram_usage.reserved_bo = NULL;
1399 return r;
1400}
1401
1402int amdgpu_ttm_init(struct amdgpu_device *adev)
1403{
1404 uint64_t gtt_size;
1405 int r;
1406 u64 vis_vram_limit;
1407
1408 r = amdgpu_ttm_global_init(adev);
1409 if (r) {
1410 return r;
1411 }
1412 /* No others user of address space so set it to 0 */
1413 r = ttm_bo_device_init(&adev->mman.bdev,
1414 adev->mman.bo_global_ref.ref.object,
1415 &amdgpu_bo_driver,
1416 adev->ddev->anon_inode->i_mapping,
1417 DRM_FILE_PAGE_OFFSET,
1418 adev->need_dma32);
1419 if (r) {
1420 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
1421 return r;
1422 }
1423 adev->mman.initialized = true;
1424
1425 /* We opt to avoid OOM on system pages allocations */
1426 adev->mman.bdev.no_retry = true;
1427
1428 r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_VRAM,
1429 adev->gmc.real_vram_size >> PAGE_SHIFT);
1430 if (r) {
1431 DRM_ERROR("Failed initializing VRAM heap.\n");
1432 return r;
1433 }
1434
1435 /* Reduce size of CPU-visible VRAM if requested */
1436 vis_vram_limit = (u64)amdgpu_vis_vram_limit * 1024 * 1024;
1437 if (amdgpu_vis_vram_limit > 0 &&
1438 vis_vram_limit <= adev->gmc.visible_vram_size)
1439 adev->gmc.visible_vram_size = vis_vram_limit;
1440
1441 /* Change the size here instead of the init above so only lpfn is affected */
1442 amdgpu_ttm_set_buffer_funcs_status(adev, false);
1443#ifdef CONFIG_64BIT
1444 adev->mman.aper_base_kaddr = ioremap_wc(adev->gmc.aper_base,
1445 adev->gmc.visible_vram_size);
1446#endif
1447
1448 /*
1449 *The reserved vram for firmware must be pinned to the specified
1450 *place on the VRAM, so reserve it early.
1451 */
1452 r = amdgpu_ttm_fw_reserve_vram_init(adev);
1453 if (r) {
1454 return r;
1455 }
1456
1457 r = amdgpu_bo_create_kernel(adev, adev->gmc.stolen_size, PAGE_SIZE,
1458 AMDGPU_GEM_DOMAIN_VRAM,
1459 &adev->stolen_vga_memory,
1460 NULL, NULL);
1461 if (r)
1462 return r;
1463 DRM_INFO("amdgpu: %uM of VRAM memory ready\n",
1464 (unsigned) (adev->gmc.real_vram_size / (1024 * 1024)));
1465
1466 if (amdgpu_gtt_size == -1) {
1467 struct sysinfo si;
1468
1469 si_meminfo(&si);
1470 gtt_size = min(max((AMDGPU_DEFAULT_GTT_SIZE_MB << 20),
1471 adev->gmc.mc_vram_size),
1472 ((uint64_t)si.totalram * si.mem_unit * 3/4));
1473 }
1474 else
1475 gtt_size = (uint64_t)amdgpu_gtt_size << 20;
1476 r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_TT, gtt_size >> PAGE_SHIFT);
1477 if (r) {
1478 DRM_ERROR("Failed initializing GTT heap.\n");
1479 return r;
1480 }
1481 DRM_INFO("amdgpu: %uM of GTT memory ready.\n",
1482 (unsigned)(gtt_size / (1024 * 1024)));
1483
1484 adev->gds.mem.total_size = adev->gds.mem.total_size << AMDGPU_GDS_SHIFT;
1485 adev->gds.mem.gfx_partition_size = adev->gds.mem.gfx_partition_size << AMDGPU_GDS_SHIFT;
1486 adev->gds.mem.cs_partition_size = adev->gds.mem.cs_partition_size << AMDGPU_GDS_SHIFT;
1487 adev->gds.gws.total_size = adev->gds.gws.total_size << AMDGPU_GWS_SHIFT;
1488 adev->gds.gws.gfx_partition_size = adev->gds.gws.gfx_partition_size << AMDGPU_GWS_SHIFT;
1489 adev->gds.gws.cs_partition_size = adev->gds.gws.cs_partition_size << AMDGPU_GWS_SHIFT;
1490 adev->gds.oa.total_size = adev->gds.oa.total_size << AMDGPU_OA_SHIFT;
1491 adev->gds.oa.gfx_partition_size = adev->gds.oa.gfx_partition_size << AMDGPU_OA_SHIFT;
1492 adev->gds.oa.cs_partition_size = adev->gds.oa.cs_partition_size << AMDGPU_OA_SHIFT;
1493 /* GDS Memory */
1494 if (adev->gds.mem.total_size) {
1495 r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_GDS,
1496 adev->gds.mem.total_size >> PAGE_SHIFT);
1497 if (r) {
1498 DRM_ERROR("Failed initializing GDS heap.\n");
1499 return r;
1500 }
1501 }
1502
1503 /* GWS */
1504 if (adev->gds.gws.total_size) {
1505 r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_GWS,
1506 adev->gds.gws.total_size >> PAGE_SHIFT);
1507 if (r) {
1508 DRM_ERROR("Failed initializing gws heap.\n");
1509 return r;
1510 }
1511 }
1512
1513 /* OA */
1514 if (adev->gds.oa.total_size) {
1515 r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_OA,
1516 adev->gds.oa.total_size >> PAGE_SHIFT);
1517 if (r) {
1518 DRM_ERROR("Failed initializing oa heap.\n");
1519 return r;
1520 }
1521 }
1522
1523 r = amdgpu_ttm_debugfs_init(adev);
1524 if (r) {
1525 DRM_ERROR("Failed to init debugfs\n");
1526 return r;
1527 }
1528 return 0;
1529}
1530
1531void amdgpu_ttm_fini(struct amdgpu_device *adev)
1532{
1533 if (!adev->mman.initialized)
1534 return;
1535
1536 amdgpu_ttm_debugfs_fini(adev);
1537 amdgpu_bo_free_kernel(&adev->stolen_vga_memory, NULL, NULL);
1538 amdgpu_ttm_fw_reserve_vram_fini(adev);
1539 if (adev->mman.aper_base_kaddr)
1540 iounmap(adev->mman.aper_base_kaddr);
1541 adev->mman.aper_base_kaddr = NULL;
1542
1543 ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_VRAM);
1544 ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_TT);
1545 if (adev->gds.mem.total_size)
1546 ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_GDS);
1547 if (adev->gds.gws.total_size)
1548 ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_GWS);
1549 if (adev->gds.oa.total_size)
1550 ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_OA);
1551 ttm_bo_device_release(&adev->mman.bdev);
1552 amdgpu_ttm_global_fini(adev);
1553 adev->mman.initialized = false;
1554 DRM_INFO("amdgpu: ttm finalized\n");
1555}
1556
1557/**
1558 * amdgpu_ttm_set_buffer_funcs_status - enable/disable use of buffer functions
1559 *
1560 * @adev: amdgpu_device pointer
1561 * @enable: true when we can use buffer functions.
1562 *
1563 * Enable/disable use of buffer functions during suspend/resume. This should
1564 * only be called at bootup or when userspace isn't running.
1565 */
1566void amdgpu_ttm_set_buffer_funcs_status(struct amdgpu_device *adev, bool enable)
1567{
1568 struct ttm_mem_type_manager *man = &adev->mman.bdev.man[TTM_PL_VRAM];
1569 uint64_t size;
1570
1571 if (!adev->mman.initialized || adev->in_gpu_reset)
1572 return;
1573
1574 /* this just adjusts TTM size idea, which sets lpfn to the correct value */
1575 if (enable)
1576 size = adev->gmc.real_vram_size;
1577 else
1578 size = adev->gmc.visible_vram_size;
1579 man->size = size >> PAGE_SHIFT;
1580 adev->mman.buffer_funcs_enabled = enable;
1581}
1582
1583int amdgpu_mmap(struct file *filp, struct vm_area_struct *vma)
1584{
1585 struct drm_file *file_priv;
1586 struct amdgpu_device *adev;
1587
1588 if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET))
1589 return -EINVAL;
1590
1591 file_priv = filp->private_data;
1592 adev = file_priv->minor->dev->dev_private;
1593 if (adev == NULL)
1594 return -EINVAL;
1595
1596 return ttm_bo_mmap(filp, vma, &adev->mman.bdev);
1597}
1598
1599static int amdgpu_map_buffer(struct ttm_buffer_object *bo,
1600 struct ttm_mem_reg *mem, unsigned num_pages,
1601 uint64_t offset, unsigned window,
1602 struct amdgpu_ring *ring,
1603 uint64_t *addr)
1604{
1605 struct amdgpu_ttm_tt *gtt = (void *)bo->ttm;
1606 struct amdgpu_device *adev = ring->adev;
1607 struct ttm_tt *ttm = bo->ttm;
1608 struct amdgpu_job *job;
1609 unsigned num_dw, num_bytes;
1610 dma_addr_t *dma_address;
1611 struct dma_fence *fence;
1612 uint64_t src_addr, dst_addr;
1613 uint64_t flags;
1614 int r;
1615
1616 BUG_ON(adev->mman.buffer_funcs->copy_max_bytes <
1617 AMDGPU_GTT_MAX_TRANSFER_SIZE * 8);
1618
1619 *addr = adev->gmc.gart_start;
1620 *addr += (u64)window * AMDGPU_GTT_MAX_TRANSFER_SIZE *
1621 AMDGPU_GPU_PAGE_SIZE;
1622
1623 num_dw = adev->mman.buffer_funcs->copy_num_dw;
1624 while (num_dw & 0x7)
1625 num_dw++;
1626
1627 num_bytes = num_pages * 8;
1628
1629 r = amdgpu_job_alloc_with_ib(adev, num_dw * 4 + num_bytes, &job);
1630 if (r)
1631 return r;
1632
1633 src_addr = num_dw * 4;
1634 src_addr += job->ibs[0].gpu_addr;
1635
1636 dst_addr = adev->gart.table_addr;
1637 dst_addr += window * AMDGPU_GTT_MAX_TRANSFER_SIZE * 8;
1638 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr,
1639 dst_addr, num_bytes);
1640
1641 amdgpu_ring_pad_ib(ring, &job->ibs[0]);
1642 WARN_ON(job->ibs[0].length_dw > num_dw);
1643
1644 dma_address = >t->ttm.dma_address[offset >> PAGE_SHIFT];
1645 flags = amdgpu_ttm_tt_pte_flags(adev, ttm, mem);
1646 r = amdgpu_gart_map(adev, 0, num_pages, dma_address, flags,
1647 &job->ibs[0].ptr[num_dw]);
1648 if (r)
1649 goto error_free;
1650
1651 r = amdgpu_job_submit(job, ring, &adev->mman.entity,
1652 AMDGPU_FENCE_OWNER_UNDEFINED, &fence);
1653 if (r)
1654 goto error_free;
1655
1656 dma_fence_put(fence);
1657
1658 return r;
1659
1660error_free:
1661 amdgpu_job_free(job);
1662 return r;
1663}
1664
1665int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
1666 uint64_t dst_offset, uint32_t byte_count,
1667 struct reservation_object *resv,
1668 struct dma_fence **fence, bool direct_submit,
1669 bool vm_needs_flush)
1670{
1671 struct amdgpu_device *adev = ring->adev;
1672 struct amdgpu_job *job;
1673
1674 uint32_t max_bytes;
1675 unsigned num_loops, num_dw;
1676 unsigned i;
1677 int r;
1678
1679 if (direct_submit && !ring->ready) {
1680 DRM_ERROR("Trying to move memory with ring turned off.\n");
1681 return -EINVAL;
1682 }
1683
1684 max_bytes = adev->mman.buffer_funcs->copy_max_bytes;
1685 num_loops = DIV_ROUND_UP(byte_count, max_bytes);
1686 num_dw = num_loops * adev->mman.buffer_funcs->copy_num_dw;
1687
1688 /* for IB padding */
1689 while (num_dw & 0x7)
1690 num_dw++;
1691
1692 r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, &job);
1693 if (r)
1694 return r;
1695
1696 job->vm_needs_flush = vm_needs_flush;
1697 if (resv) {
1698 r = amdgpu_sync_resv(adev, &job->sync, resv,
1699 AMDGPU_FENCE_OWNER_UNDEFINED,
1700 false);
1701 if (r) {
1702 DRM_ERROR("sync failed (%d).\n", r);
1703 goto error_free;
1704 }
1705 }
1706
1707 for (i = 0; i < num_loops; i++) {
1708 uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
1709
1710 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_offset,
1711 dst_offset, cur_size_in_bytes);
1712
1713 src_offset += cur_size_in_bytes;
1714 dst_offset += cur_size_in_bytes;
1715 byte_count -= cur_size_in_bytes;
1716 }
1717
1718 amdgpu_ring_pad_ib(ring, &job->ibs[0]);
1719 WARN_ON(job->ibs[0].length_dw > num_dw);
1720 if (direct_submit) {
1721 r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
1722 NULL, fence);
1723 job->fence = dma_fence_get(*fence);
1724 if (r)
1725 DRM_ERROR("Error scheduling IBs (%d)\n", r);
1726 amdgpu_job_free(job);
1727 } else {
1728 r = amdgpu_job_submit(job, ring, &adev->mman.entity,
1729 AMDGPU_FENCE_OWNER_UNDEFINED, fence);
1730 if (r)
1731 goto error_free;
1732 }
1733
1734 return r;
1735
1736error_free:
1737 amdgpu_job_free(job);
1738 return r;
1739}
1740
1741int amdgpu_fill_buffer(struct amdgpu_bo *bo,
1742 uint32_t src_data,
1743 struct reservation_object *resv,
1744 struct dma_fence **fence)
1745{
1746 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1747 uint32_t max_bytes = adev->mman.buffer_funcs->fill_max_bytes;
1748 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
1749
1750 struct drm_mm_node *mm_node;
1751 unsigned long num_pages;
1752 unsigned int num_loops, num_dw;
1753
1754 struct amdgpu_job *job;
1755 int r;
1756
1757 if (!adev->mman.buffer_funcs_enabled) {
1758 DRM_ERROR("Trying to clear memory with ring turned off.\n");
1759 return -EINVAL;
1760 }
1761
1762 if (bo->tbo.mem.mem_type == TTM_PL_TT) {
1763 r = amdgpu_ttm_alloc_gart(&bo->tbo);
1764 if (r)
1765 return r;
1766 }
1767
1768 num_pages = bo->tbo.num_pages;
1769 mm_node = bo->tbo.mem.mm_node;
1770 num_loops = 0;
1771 while (num_pages) {
1772 uint32_t byte_count = mm_node->size << PAGE_SHIFT;
1773
1774 num_loops += DIV_ROUND_UP(byte_count, max_bytes);
1775 num_pages -= mm_node->size;
1776 ++mm_node;
1777 }
1778 num_dw = num_loops * adev->mman.buffer_funcs->fill_num_dw;
1779
1780 /* for IB padding */
1781 num_dw += 64;
1782
1783 r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, &job);
1784 if (r)
1785 return r;
1786
1787 if (resv) {
1788 r = amdgpu_sync_resv(adev, &job->sync, resv,
1789 AMDGPU_FENCE_OWNER_UNDEFINED, false);
1790 if (r) {
1791 DRM_ERROR("sync failed (%d).\n", r);
1792 goto error_free;
1793 }
1794 }
1795
1796 num_pages = bo->tbo.num_pages;
1797 mm_node = bo->tbo.mem.mm_node;
1798
1799 while (num_pages) {
1800 uint32_t byte_count = mm_node->size << PAGE_SHIFT;
1801 uint64_t dst_addr;
1802
1803 dst_addr = amdgpu_mm_node_addr(&bo->tbo, mm_node, &bo->tbo.mem);
1804 while (byte_count) {
1805 uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
1806
1807 amdgpu_emit_fill_buffer(adev, &job->ibs[0], src_data,
1808 dst_addr, cur_size_in_bytes);
1809
1810 dst_addr += cur_size_in_bytes;
1811 byte_count -= cur_size_in_bytes;
1812 }
1813
1814 num_pages -= mm_node->size;
1815 ++mm_node;
1816 }
1817
1818 amdgpu_ring_pad_ib(ring, &job->ibs[0]);
1819 WARN_ON(job->ibs[0].length_dw > num_dw);
1820 r = amdgpu_job_submit(job, ring, &adev->mman.entity,
1821 AMDGPU_FENCE_OWNER_UNDEFINED, fence);
1822 if (r)
1823 goto error_free;
1824
1825 return 0;
1826
1827error_free:
1828 amdgpu_job_free(job);
1829 return r;
1830}
1831
1832#if defined(CONFIG_DEBUG_FS)
1833
1834static int amdgpu_mm_dump_table(struct seq_file *m, void *data)
1835{
1836 struct drm_info_node *node = (struct drm_info_node *)m->private;
1837 unsigned ttm_pl = *(int *)node->info_ent->data;
1838 struct drm_device *dev = node->minor->dev;
1839 struct amdgpu_device *adev = dev->dev_private;
1840 struct ttm_mem_type_manager *man = &adev->mman.bdev.man[ttm_pl];
1841 struct drm_printer p = drm_seq_file_printer(m);
1842
1843 man->func->debug(man, &p);
1844 return 0;
1845}
1846
1847static int ttm_pl_vram = TTM_PL_VRAM;
1848static int ttm_pl_tt = TTM_PL_TT;
1849
1850static const struct drm_info_list amdgpu_ttm_debugfs_list[] = {
1851 {"amdgpu_vram_mm", amdgpu_mm_dump_table, 0, &ttm_pl_vram},
1852 {"amdgpu_gtt_mm", amdgpu_mm_dump_table, 0, &ttm_pl_tt},
1853 {"ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL},
1854#ifdef CONFIG_SWIOTLB
1855 {"ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL}
1856#endif
1857};
1858
1859static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf,
1860 size_t size, loff_t *pos)
1861{
1862 struct amdgpu_device *adev = file_inode(f)->i_private;
1863 ssize_t result = 0;
1864 int r;
1865
1866 if (size & 0x3 || *pos & 0x3)
1867 return -EINVAL;
1868
1869 if (*pos >= adev->gmc.mc_vram_size)
1870 return -ENXIO;
1871
1872 while (size) {
1873 unsigned long flags;
1874 uint32_t value;
1875
1876 if (*pos >= adev->gmc.mc_vram_size)
1877 return result;
1878
1879 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
1880 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)*pos) | 0x80000000);
1881 WREG32_NO_KIQ(mmMM_INDEX_HI, *pos >> 31);
1882 value = RREG32_NO_KIQ(mmMM_DATA);
1883 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
1884
1885 r = put_user(value, (uint32_t *)buf);
1886 if (r)
1887 return r;
1888
1889 result += 4;
1890 buf += 4;
1891 *pos += 4;
1892 size -= 4;
1893 }
1894
1895 return result;
1896}
1897
1898static ssize_t amdgpu_ttm_vram_write(struct file *f, const char __user *buf,
1899 size_t size, loff_t *pos)
1900{
1901 struct amdgpu_device *adev = file_inode(f)->i_private;
1902 ssize_t result = 0;
1903 int r;
1904
1905 if (size & 0x3 || *pos & 0x3)
1906 return -EINVAL;
1907
1908 if (*pos >= adev->gmc.mc_vram_size)
1909 return -ENXIO;
1910
1911 while (size) {
1912 unsigned long flags;
1913 uint32_t value;
1914
1915 if (*pos >= adev->gmc.mc_vram_size)
1916 return result;
1917
1918 r = get_user(value, (uint32_t *)buf);
1919 if (r)
1920 return r;
1921
1922 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
1923 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)*pos) | 0x80000000);
1924 WREG32_NO_KIQ(mmMM_INDEX_HI, *pos >> 31);
1925 WREG32_NO_KIQ(mmMM_DATA, value);
1926 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
1927
1928 result += 4;
1929 buf += 4;
1930 *pos += 4;
1931 size -= 4;
1932 }
1933
1934 return result;
1935}
1936
1937static const struct file_operations amdgpu_ttm_vram_fops = {
1938 .owner = THIS_MODULE,
1939 .read = amdgpu_ttm_vram_read,
1940 .write = amdgpu_ttm_vram_write,
1941 .llseek = default_llseek,
1942};
1943
1944#ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
1945
1946static ssize_t amdgpu_ttm_gtt_read(struct file *f, char __user *buf,
1947 size_t size, loff_t *pos)
1948{
1949 struct amdgpu_device *adev = file_inode(f)->i_private;
1950 ssize_t result = 0;
1951 int r;
1952
1953 while (size) {
1954 loff_t p = *pos / PAGE_SIZE;
1955 unsigned off = *pos & ~PAGE_MASK;
1956 size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
1957 struct page *page;
1958 void *ptr;
1959
1960 if (p >= adev->gart.num_cpu_pages)
1961 return result;
1962
1963 page = adev->gart.pages[p];
1964 if (page) {
1965 ptr = kmap(page);
1966 ptr += off;
1967
1968 r = copy_to_user(buf, ptr, cur_size);
1969 kunmap(adev->gart.pages[p]);
1970 } else
1971 r = clear_user(buf, cur_size);
1972
1973 if (r)
1974 return -EFAULT;
1975
1976 result += cur_size;
1977 buf += cur_size;
1978 *pos += cur_size;
1979 size -= cur_size;
1980 }
1981
1982 return result;
1983}
1984
1985static const struct file_operations amdgpu_ttm_gtt_fops = {
1986 .owner = THIS_MODULE,
1987 .read = amdgpu_ttm_gtt_read,
1988 .llseek = default_llseek
1989};
1990
1991#endif
1992
1993static ssize_t amdgpu_iomem_read(struct file *f, char __user *buf,
1994 size_t size, loff_t *pos)
1995{
1996 struct amdgpu_device *adev = file_inode(f)->i_private;
1997 struct iommu_domain *dom;
1998 ssize_t result = 0;
1999 int r;
2000
2001 dom = iommu_get_domain_for_dev(adev->dev);
2002
2003 while (size) {
2004 phys_addr_t addr = *pos & PAGE_MASK;
2005 loff_t off = *pos & ~PAGE_MASK;
2006 size_t bytes = PAGE_SIZE - off;
2007 unsigned long pfn;
2008 struct page *p;
2009 void *ptr;
2010
2011 bytes = bytes < size ? bytes : size;
2012
2013 addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2014
2015 pfn = addr >> PAGE_SHIFT;
2016 if (!pfn_valid(pfn))
2017 return -EPERM;
2018
2019 p = pfn_to_page(pfn);
2020 if (p->mapping != adev->mman.bdev.dev_mapping)
2021 return -EPERM;
2022
2023 ptr = kmap(p);
2024 r = copy_to_user(buf, ptr + off, bytes);
2025 kunmap(p);
2026 if (r)
2027 return -EFAULT;
2028
2029 size -= bytes;
2030 *pos += bytes;
2031 result += bytes;
2032 }
2033
2034 return result;
2035}
2036
2037static ssize_t amdgpu_iomem_write(struct file *f, const char __user *buf,
2038 size_t size, loff_t *pos)
2039{
2040 struct amdgpu_device *adev = file_inode(f)->i_private;
2041 struct iommu_domain *dom;
2042 ssize_t result = 0;
2043 int r;
2044
2045 dom = iommu_get_domain_for_dev(adev->dev);
2046
2047 while (size) {
2048 phys_addr_t addr = *pos & PAGE_MASK;
2049 loff_t off = *pos & ~PAGE_MASK;
2050 size_t bytes = PAGE_SIZE - off;
2051 unsigned long pfn;
2052 struct page *p;
2053 void *ptr;
2054
2055 bytes = bytes < size ? bytes : size;
2056
2057 addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2058
2059 pfn = addr >> PAGE_SHIFT;
2060 if (!pfn_valid(pfn))
2061 return -EPERM;
2062
2063 p = pfn_to_page(pfn);
2064 if (p->mapping != adev->mman.bdev.dev_mapping)
2065 return -EPERM;
2066
2067 ptr = kmap(p);
2068 r = copy_from_user(ptr + off, buf, bytes);
2069 kunmap(p);
2070 if (r)
2071 return -EFAULT;
2072
2073 size -= bytes;
2074 *pos += bytes;
2075 result += bytes;
2076 }
2077
2078 return result;
2079}
2080
2081static const struct file_operations amdgpu_ttm_iomem_fops = {
2082 .owner = THIS_MODULE,
2083 .read = amdgpu_iomem_read,
2084 .write = amdgpu_iomem_write,
2085 .llseek = default_llseek
2086};
2087
2088static const struct {
2089 char *name;
2090 const struct file_operations *fops;
2091 int domain;
2092} ttm_debugfs_entries[] = {
2093 { "amdgpu_vram", &amdgpu_ttm_vram_fops, TTM_PL_VRAM },
2094#ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
2095 { "amdgpu_gtt", &amdgpu_ttm_gtt_fops, TTM_PL_TT },
2096#endif
2097 { "amdgpu_iomem", &amdgpu_ttm_iomem_fops, TTM_PL_SYSTEM },
2098};
2099
2100#endif
2101
2102static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev)
2103{
2104#if defined(CONFIG_DEBUG_FS)
2105 unsigned count;
2106
2107 struct drm_minor *minor = adev->ddev->primary;
2108 struct dentry *ent, *root = minor->debugfs_root;
2109
2110 for (count = 0; count < ARRAY_SIZE(ttm_debugfs_entries); count++) {
2111 ent = debugfs_create_file(
2112 ttm_debugfs_entries[count].name,
2113 S_IFREG | S_IRUGO, root,
2114 adev,
2115 ttm_debugfs_entries[count].fops);
2116 if (IS_ERR(ent))
2117 return PTR_ERR(ent);
2118 if (ttm_debugfs_entries[count].domain == TTM_PL_VRAM)
2119 i_size_write(ent->d_inode, adev->gmc.mc_vram_size);
2120 else if (ttm_debugfs_entries[count].domain == TTM_PL_TT)
2121 i_size_write(ent->d_inode, adev->gmc.gart_size);
2122 adev->mman.debugfs_entries[count] = ent;
2123 }
2124
2125 count = ARRAY_SIZE(amdgpu_ttm_debugfs_list);
2126
2127#ifdef CONFIG_SWIOTLB
2128 if (!(adev->need_swiotlb && swiotlb_nr_tbl()))
2129 --count;
2130#endif
2131
2132 return amdgpu_debugfs_add_files(adev, amdgpu_ttm_debugfs_list, count);
2133#else
2134 return 0;
2135#endif
2136}
2137
2138static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev)
2139{
2140#if defined(CONFIG_DEBUG_FS)
2141 unsigned i;
2142
2143 for (i = 0; i < ARRAY_SIZE(ttm_debugfs_entries); i++)
2144 debugfs_remove(adev->mman.debugfs_entries[i]);
2145#endif
2146}
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
33#include <linux/dma-mapping.h>
34#include <linux/iommu.h>
35#include <linux/pagemap.h>
36#include <linux/sched/task.h>
37#include <linux/sched/mm.h>
38#include <linux/seq_file.h>
39#include <linux/slab.h>
40#include <linux/swap.h>
41#include <linux/dma-buf.h>
42#include <linux/sizes.h>
43#include <linux/module.h>
44
45#include <drm/drm_drv.h>
46#include <drm/ttm/ttm_bo.h>
47#include <drm/ttm/ttm_placement.h>
48#include <drm/ttm/ttm_range_manager.h>
49#include <drm/ttm/ttm_tt.h>
50
51#include <drm/amdgpu_drm.h>
52
53#include "amdgpu.h"
54#include "amdgpu_object.h"
55#include "amdgpu_trace.h"
56#include "amdgpu_amdkfd.h"
57#include "amdgpu_sdma.h"
58#include "amdgpu_ras.h"
59#include "amdgpu_hmm.h"
60#include "amdgpu_atomfirmware.h"
61#include "amdgpu_res_cursor.h"
62#include "bif/bif_4_1_d.h"
63
64MODULE_IMPORT_NS(DMA_BUF);
65
66#define AMDGPU_TTM_VRAM_MAX_DW_READ ((size_t)128)
67
68static int amdgpu_ttm_backend_bind(struct ttm_device *bdev,
69 struct ttm_tt *ttm,
70 struct ttm_resource *bo_mem);
71static void amdgpu_ttm_backend_unbind(struct ttm_device *bdev,
72 struct ttm_tt *ttm);
73
74static int amdgpu_ttm_init_on_chip(struct amdgpu_device *adev,
75 unsigned int type,
76 uint64_t size_in_page)
77{
78 return ttm_range_man_init(&adev->mman.bdev, type,
79 false, size_in_page);
80}
81
82/**
83 * amdgpu_evict_flags - Compute placement flags
84 *
85 * @bo: The buffer object to evict
86 * @placement: Possible destination(s) for evicted BO
87 *
88 * Fill in placement data when ttm_bo_evict() is called
89 */
90static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
91 struct ttm_placement *placement)
92{
93 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
94 struct amdgpu_bo *abo;
95 static const struct ttm_place placements = {
96 .fpfn = 0,
97 .lpfn = 0,
98 .mem_type = TTM_PL_SYSTEM,
99 .flags = 0
100 };
101
102 /* Don't handle scatter gather BOs */
103 if (bo->type == ttm_bo_type_sg) {
104 placement->num_placement = 0;
105 placement->num_busy_placement = 0;
106 return;
107 }
108
109 /* Object isn't an AMDGPU object so ignore */
110 if (!amdgpu_bo_is_amdgpu_bo(bo)) {
111 placement->placement = &placements;
112 placement->busy_placement = &placements;
113 placement->num_placement = 1;
114 placement->num_busy_placement = 1;
115 return;
116 }
117
118 abo = ttm_to_amdgpu_bo(bo);
119 if (abo->flags & AMDGPU_GEM_CREATE_DISCARDABLE) {
120 placement->num_placement = 0;
121 placement->num_busy_placement = 0;
122 return;
123 }
124
125 switch (bo->resource->mem_type) {
126 case AMDGPU_PL_GDS:
127 case AMDGPU_PL_GWS:
128 case AMDGPU_PL_OA:
129 case AMDGPU_PL_DOORBELL:
130 placement->num_placement = 0;
131 placement->num_busy_placement = 0;
132 return;
133
134 case TTM_PL_VRAM:
135 if (!adev->mman.buffer_funcs_enabled) {
136 /* Move to system memory */
137 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
138 } else if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
139 !(abo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) &&
140 amdgpu_bo_in_cpu_visible_vram(abo)) {
141
142 /* Try evicting to the CPU inaccessible part of VRAM
143 * first, but only set GTT as busy placement, so this
144 * BO will be evicted to GTT rather than causing other
145 * BOs to be evicted from VRAM
146 */
147 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
148 AMDGPU_GEM_DOMAIN_GTT |
149 AMDGPU_GEM_DOMAIN_CPU);
150 abo->placements[0].fpfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
151 abo->placements[0].lpfn = 0;
152 abo->placement.busy_placement = &abo->placements[1];
153 abo->placement.num_busy_placement = 1;
154 } else {
155 /* Move to GTT memory */
156 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_GTT |
157 AMDGPU_GEM_DOMAIN_CPU);
158 }
159 break;
160 case TTM_PL_TT:
161 case AMDGPU_PL_PREEMPT:
162 default:
163 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
164 break;
165 }
166 *placement = abo->placement;
167}
168
169/**
170 * amdgpu_ttm_map_buffer - Map memory into the GART windows
171 * @bo: buffer object to map
172 * @mem: memory object to map
173 * @mm_cur: range to map
174 * @window: which GART window to use
175 * @ring: DMA ring to use for the copy
176 * @tmz: if we should setup a TMZ enabled mapping
177 * @size: in number of bytes to map, out number of bytes mapped
178 * @addr: resulting address inside the MC address space
179 *
180 * Setup one of the GART windows to access a specific piece of memory or return
181 * the physical address for local memory.
182 */
183static int amdgpu_ttm_map_buffer(struct ttm_buffer_object *bo,
184 struct ttm_resource *mem,
185 struct amdgpu_res_cursor *mm_cur,
186 unsigned int window, struct amdgpu_ring *ring,
187 bool tmz, uint64_t *size, uint64_t *addr)
188{
189 struct amdgpu_device *adev = ring->adev;
190 unsigned int offset, num_pages, num_dw, num_bytes;
191 uint64_t src_addr, dst_addr;
192 struct amdgpu_job *job;
193 void *cpu_addr;
194 uint64_t flags;
195 unsigned int i;
196 int r;
197
198 BUG_ON(adev->mman.buffer_funcs->copy_max_bytes <
199 AMDGPU_GTT_MAX_TRANSFER_SIZE * 8);
200
201 if (WARN_ON(mem->mem_type == AMDGPU_PL_PREEMPT))
202 return -EINVAL;
203
204 /* Map only what can't be accessed directly */
205 if (!tmz && mem->start != AMDGPU_BO_INVALID_OFFSET) {
206 *addr = amdgpu_ttm_domain_start(adev, mem->mem_type) +
207 mm_cur->start;
208 return 0;
209 }
210
211
212 /*
213 * If start begins at an offset inside the page, then adjust the size
214 * and addr accordingly
215 */
216 offset = mm_cur->start & ~PAGE_MASK;
217
218 num_pages = PFN_UP(*size + offset);
219 num_pages = min_t(uint32_t, num_pages, AMDGPU_GTT_MAX_TRANSFER_SIZE);
220
221 *size = min(*size, (uint64_t)num_pages * PAGE_SIZE - offset);
222
223 *addr = adev->gmc.gart_start;
224 *addr += (u64)window * AMDGPU_GTT_MAX_TRANSFER_SIZE *
225 AMDGPU_GPU_PAGE_SIZE;
226 *addr += offset;
227
228 num_dw = ALIGN(adev->mman.buffer_funcs->copy_num_dw, 8);
229 num_bytes = num_pages * 8 * AMDGPU_GPU_PAGES_IN_CPU_PAGE;
230
231 r = amdgpu_job_alloc_with_ib(adev, &adev->mman.high_pr,
232 AMDGPU_FENCE_OWNER_UNDEFINED,
233 num_dw * 4 + num_bytes,
234 AMDGPU_IB_POOL_DELAYED, &job);
235 if (r)
236 return r;
237
238 src_addr = num_dw * 4;
239 src_addr += job->ibs[0].gpu_addr;
240
241 dst_addr = amdgpu_bo_gpu_offset(adev->gart.bo);
242 dst_addr += window * AMDGPU_GTT_MAX_TRANSFER_SIZE * 8;
243 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr,
244 dst_addr, num_bytes, false);
245
246 amdgpu_ring_pad_ib(ring, &job->ibs[0]);
247 WARN_ON(job->ibs[0].length_dw > num_dw);
248
249 flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, mem);
250 if (tmz)
251 flags |= AMDGPU_PTE_TMZ;
252
253 cpu_addr = &job->ibs[0].ptr[num_dw];
254
255 if (mem->mem_type == TTM_PL_TT) {
256 dma_addr_t *dma_addr;
257
258 dma_addr = &bo->ttm->dma_address[mm_cur->start >> PAGE_SHIFT];
259 amdgpu_gart_map(adev, 0, num_pages, dma_addr, flags, cpu_addr);
260 } else {
261 dma_addr_t dma_address;
262
263 dma_address = mm_cur->start;
264 dma_address += adev->vm_manager.vram_base_offset;
265
266 for (i = 0; i < num_pages; ++i) {
267 amdgpu_gart_map(adev, i << PAGE_SHIFT, 1, &dma_address,
268 flags, cpu_addr);
269 dma_address += PAGE_SIZE;
270 }
271 }
272
273 dma_fence_put(amdgpu_job_submit(job));
274 return 0;
275}
276
277/**
278 * amdgpu_ttm_copy_mem_to_mem - Helper function for copy
279 * @adev: amdgpu device
280 * @src: buffer/address where to read from
281 * @dst: buffer/address where to write to
282 * @size: number of bytes to copy
283 * @tmz: if a secure copy should be used
284 * @resv: resv object to sync to
285 * @f: Returns the last fence if multiple jobs are submitted.
286 *
287 * The function copies @size bytes from {src->mem + src->offset} to
288 * {dst->mem + dst->offset}. src->bo and dst->bo could be same BO for a
289 * move and different for a BO to BO copy.
290 *
291 */
292int amdgpu_ttm_copy_mem_to_mem(struct amdgpu_device *adev,
293 const struct amdgpu_copy_mem *src,
294 const struct amdgpu_copy_mem *dst,
295 uint64_t size, bool tmz,
296 struct dma_resv *resv,
297 struct dma_fence **f)
298{
299 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
300 struct amdgpu_res_cursor src_mm, dst_mm;
301 struct dma_fence *fence = NULL;
302 int r = 0;
303
304 if (!adev->mman.buffer_funcs_enabled) {
305 DRM_ERROR("Trying to move memory with ring turned off.\n");
306 return -EINVAL;
307 }
308
309 amdgpu_res_first(src->mem, src->offset, size, &src_mm);
310 amdgpu_res_first(dst->mem, dst->offset, size, &dst_mm);
311
312 mutex_lock(&adev->mman.gtt_window_lock);
313 while (src_mm.remaining) {
314 uint64_t from, to, cur_size;
315 struct dma_fence *next;
316
317 /* Never copy more than 256MiB at once to avoid a timeout */
318 cur_size = min3(src_mm.size, dst_mm.size, 256ULL << 20);
319
320 /* Map src to window 0 and dst to window 1. */
321 r = amdgpu_ttm_map_buffer(src->bo, src->mem, &src_mm,
322 0, ring, tmz, &cur_size, &from);
323 if (r)
324 goto error;
325
326 r = amdgpu_ttm_map_buffer(dst->bo, dst->mem, &dst_mm,
327 1, ring, tmz, &cur_size, &to);
328 if (r)
329 goto error;
330
331 r = amdgpu_copy_buffer(ring, from, to, cur_size,
332 resv, &next, false, true, tmz);
333 if (r)
334 goto error;
335
336 dma_fence_put(fence);
337 fence = next;
338
339 amdgpu_res_next(&src_mm, cur_size);
340 amdgpu_res_next(&dst_mm, cur_size);
341 }
342error:
343 mutex_unlock(&adev->mman.gtt_window_lock);
344 if (f)
345 *f = dma_fence_get(fence);
346 dma_fence_put(fence);
347 return r;
348}
349
350/*
351 * amdgpu_move_blit - Copy an entire buffer to another buffer
352 *
353 * This is a helper called by amdgpu_bo_move() and amdgpu_move_vram_ram() to
354 * help move buffers to and from VRAM.
355 */
356static int amdgpu_move_blit(struct ttm_buffer_object *bo,
357 bool evict,
358 struct ttm_resource *new_mem,
359 struct ttm_resource *old_mem)
360{
361 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
362 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
363 struct amdgpu_copy_mem src, dst;
364 struct dma_fence *fence = NULL;
365 int r;
366
367 src.bo = bo;
368 dst.bo = bo;
369 src.mem = old_mem;
370 dst.mem = new_mem;
371 src.offset = 0;
372 dst.offset = 0;
373
374 r = amdgpu_ttm_copy_mem_to_mem(adev, &src, &dst,
375 new_mem->size,
376 amdgpu_bo_encrypted(abo),
377 bo->base.resv, &fence);
378 if (r)
379 goto error;
380
381 /* clear the space being freed */
382 if (old_mem->mem_type == TTM_PL_VRAM &&
383 (abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE)) {
384 struct dma_fence *wipe_fence = NULL;
385
386 r = amdgpu_fill_buffer(abo, AMDGPU_POISON, NULL, &wipe_fence,
387 false);
388 if (r) {
389 goto error;
390 } else if (wipe_fence) {
391 dma_fence_put(fence);
392 fence = wipe_fence;
393 }
394 }
395
396 /* Always block for VM page tables before committing the new location */
397 if (bo->type == ttm_bo_type_kernel)
398 r = ttm_bo_move_accel_cleanup(bo, fence, true, false, new_mem);
399 else
400 r = ttm_bo_move_accel_cleanup(bo, fence, evict, true, new_mem);
401 dma_fence_put(fence);
402 return r;
403
404error:
405 if (fence)
406 dma_fence_wait(fence, false);
407 dma_fence_put(fence);
408 return r;
409}
410
411/*
412 * amdgpu_mem_visible - Check that memory can be accessed by ttm_bo_move_memcpy
413 *
414 * Called by amdgpu_bo_move()
415 */
416static bool amdgpu_mem_visible(struct amdgpu_device *adev,
417 struct ttm_resource *mem)
418{
419 u64 mem_size = (u64)mem->size;
420 struct amdgpu_res_cursor cursor;
421 u64 end;
422
423 if (mem->mem_type == TTM_PL_SYSTEM ||
424 mem->mem_type == TTM_PL_TT)
425 return true;
426 if (mem->mem_type != TTM_PL_VRAM)
427 return false;
428
429 amdgpu_res_first(mem, 0, mem_size, &cursor);
430 end = cursor.start + cursor.size;
431 while (cursor.remaining) {
432 amdgpu_res_next(&cursor, cursor.size);
433
434 if (!cursor.remaining)
435 break;
436
437 /* ttm_resource_ioremap only supports contiguous memory */
438 if (end != cursor.start)
439 return false;
440
441 end = cursor.start + cursor.size;
442 }
443
444 return end <= adev->gmc.visible_vram_size;
445}
446
447/*
448 * amdgpu_bo_move - Move a buffer object to a new memory location
449 *
450 * Called by ttm_bo_handle_move_mem()
451 */
452static int amdgpu_bo_move(struct ttm_buffer_object *bo, bool evict,
453 struct ttm_operation_ctx *ctx,
454 struct ttm_resource *new_mem,
455 struct ttm_place *hop)
456{
457 struct amdgpu_device *adev;
458 struct amdgpu_bo *abo;
459 struct ttm_resource *old_mem = bo->resource;
460 int r;
461
462 if (new_mem->mem_type == TTM_PL_TT ||
463 new_mem->mem_type == AMDGPU_PL_PREEMPT) {
464 r = amdgpu_ttm_backend_bind(bo->bdev, bo->ttm, new_mem);
465 if (r)
466 return r;
467 }
468
469 abo = ttm_to_amdgpu_bo(bo);
470 adev = amdgpu_ttm_adev(bo->bdev);
471
472 if (!old_mem || (old_mem->mem_type == TTM_PL_SYSTEM &&
473 bo->ttm == NULL)) {
474 ttm_bo_move_null(bo, new_mem);
475 goto out;
476 }
477 if (old_mem->mem_type == TTM_PL_SYSTEM &&
478 (new_mem->mem_type == TTM_PL_TT ||
479 new_mem->mem_type == AMDGPU_PL_PREEMPT)) {
480 ttm_bo_move_null(bo, new_mem);
481 goto out;
482 }
483 if ((old_mem->mem_type == TTM_PL_TT ||
484 old_mem->mem_type == AMDGPU_PL_PREEMPT) &&
485 new_mem->mem_type == TTM_PL_SYSTEM) {
486 r = ttm_bo_wait_ctx(bo, ctx);
487 if (r)
488 return r;
489
490 amdgpu_ttm_backend_unbind(bo->bdev, bo->ttm);
491 ttm_resource_free(bo, &bo->resource);
492 ttm_bo_assign_mem(bo, new_mem);
493 goto out;
494 }
495
496 if (old_mem->mem_type == AMDGPU_PL_GDS ||
497 old_mem->mem_type == AMDGPU_PL_GWS ||
498 old_mem->mem_type == AMDGPU_PL_OA ||
499 old_mem->mem_type == AMDGPU_PL_DOORBELL ||
500 new_mem->mem_type == AMDGPU_PL_GDS ||
501 new_mem->mem_type == AMDGPU_PL_GWS ||
502 new_mem->mem_type == AMDGPU_PL_OA ||
503 new_mem->mem_type == AMDGPU_PL_DOORBELL) {
504 /* Nothing to save here */
505 ttm_bo_move_null(bo, new_mem);
506 goto out;
507 }
508
509 if (bo->type == ttm_bo_type_device &&
510 new_mem->mem_type == TTM_PL_VRAM &&
511 old_mem->mem_type != TTM_PL_VRAM) {
512 /* amdgpu_bo_fault_reserve_notify will re-set this if the CPU
513 * accesses the BO after it's moved.
514 */
515 abo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
516 }
517
518 if (adev->mman.buffer_funcs_enabled) {
519 if (((old_mem->mem_type == TTM_PL_SYSTEM &&
520 new_mem->mem_type == TTM_PL_VRAM) ||
521 (old_mem->mem_type == TTM_PL_VRAM &&
522 new_mem->mem_type == TTM_PL_SYSTEM))) {
523 hop->fpfn = 0;
524 hop->lpfn = 0;
525 hop->mem_type = TTM_PL_TT;
526 hop->flags = TTM_PL_FLAG_TEMPORARY;
527 return -EMULTIHOP;
528 }
529
530 r = amdgpu_move_blit(bo, evict, new_mem, old_mem);
531 } else {
532 r = -ENODEV;
533 }
534
535 if (r) {
536 /* Check that all memory is CPU accessible */
537 if (!amdgpu_mem_visible(adev, old_mem) ||
538 !amdgpu_mem_visible(adev, new_mem)) {
539 pr_err("Move buffer fallback to memcpy unavailable\n");
540 return r;
541 }
542
543 r = ttm_bo_move_memcpy(bo, ctx, new_mem);
544 if (r)
545 return r;
546 }
547
548 trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
549out:
550 /* update statistics */
551 atomic64_add(bo->base.size, &adev->num_bytes_moved);
552 amdgpu_bo_move_notify(bo, evict);
553 return 0;
554}
555
556/*
557 * amdgpu_ttm_io_mem_reserve - Reserve a block of memory during a fault
558 *
559 * Called by ttm_mem_io_reserve() ultimately via ttm_bo_vm_fault()
560 */
561static int amdgpu_ttm_io_mem_reserve(struct ttm_device *bdev,
562 struct ttm_resource *mem)
563{
564 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
565 size_t bus_size = (size_t)mem->size;
566
567 switch (mem->mem_type) {
568 case TTM_PL_SYSTEM:
569 /* system memory */
570 return 0;
571 case TTM_PL_TT:
572 case AMDGPU_PL_PREEMPT:
573 break;
574 case TTM_PL_VRAM:
575 mem->bus.offset = mem->start << PAGE_SHIFT;
576 /* check if it's visible */
577 if ((mem->bus.offset + bus_size) > adev->gmc.visible_vram_size)
578 return -EINVAL;
579
580 if (adev->mman.aper_base_kaddr &&
581 mem->placement & TTM_PL_FLAG_CONTIGUOUS)
582 mem->bus.addr = (u8 *)adev->mman.aper_base_kaddr +
583 mem->bus.offset;
584
585 mem->bus.offset += adev->gmc.aper_base;
586 mem->bus.is_iomem = true;
587 break;
588 case AMDGPU_PL_DOORBELL:
589 mem->bus.offset = mem->start << PAGE_SHIFT;
590 mem->bus.offset += adev->doorbell.base;
591 mem->bus.is_iomem = true;
592 mem->bus.caching = ttm_uncached;
593 break;
594 default:
595 return -EINVAL;
596 }
597 return 0;
598}
599
600static unsigned long amdgpu_ttm_io_mem_pfn(struct ttm_buffer_object *bo,
601 unsigned long page_offset)
602{
603 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
604 struct amdgpu_res_cursor cursor;
605
606 amdgpu_res_first(bo->resource, (u64)page_offset << PAGE_SHIFT, 0,
607 &cursor);
608
609 if (bo->resource->mem_type == AMDGPU_PL_DOORBELL)
610 return ((uint64_t)(adev->doorbell.base + cursor.start)) >> PAGE_SHIFT;
611
612 return (adev->gmc.aper_base + cursor.start) >> PAGE_SHIFT;
613}
614
615/**
616 * amdgpu_ttm_domain_start - Returns GPU start address
617 * @adev: amdgpu device object
618 * @type: type of the memory
619 *
620 * Returns:
621 * GPU start address of a memory domain
622 */
623
624uint64_t amdgpu_ttm_domain_start(struct amdgpu_device *adev, uint32_t type)
625{
626 switch (type) {
627 case TTM_PL_TT:
628 return adev->gmc.gart_start;
629 case TTM_PL_VRAM:
630 return adev->gmc.vram_start;
631 }
632
633 return 0;
634}
635
636/*
637 * TTM backend functions.
638 */
639struct amdgpu_ttm_tt {
640 struct ttm_tt ttm;
641 struct drm_gem_object *gobj;
642 u64 offset;
643 uint64_t userptr;
644 struct task_struct *usertask;
645 uint32_t userflags;
646 bool bound;
647 int32_t pool_id;
648};
649
650#define ttm_to_amdgpu_ttm_tt(ptr) container_of(ptr, struct amdgpu_ttm_tt, ttm)
651
652#ifdef CONFIG_DRM_AMDGPU_USERPTR
653/*
654 * amdgpu_ttm_tt_get_user_pages - get device accessible pages that back user
655 * memory and start HMM tracking CPU page table update
656 *
657 * Calling function must call amdgpu_ttm_tt_userptr_range_done() once and only
658 * once afterwards to stop HMM tracking
659 */
660int amdgpu_ttm_tt_get_user_pages(struct amdgpu_bo *bo, struct page **pages,
661 struct hmm_range **range)
662{
663 struct ttm_tt *ttm = bo->tbo.ttm;
664 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
665 unsigned long start = gtt->userptr;
666 struct vm_area_struct *vma;
667 struct mm_struct *mm;
668 bool readonly;
669 int r = 0;
670
671 /* Make sure get_user_pages_done() can cleanup gracefully */
672 *range = NULL;
673
674 mm = bo->notifier.mm;
675 if (unlikely(!mm)) {
676 DRM_DEBUG_DRIVER("BO is not registered?\n");
677 return -EFAULT;
678 }
679
680 if (!mmget_not_zero(mm)) /* Happens during process shutdown */
681 return -ESRCH;
682
683 mmap_read_lock(mm);
684 vma = vma_lookup(mm, start);
685 if (unlikely(!vma)) {
686 r = -EFAULT;
687 goto out_unlock;
688 }
689 if (unlikely((gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) &&
690 vma->vm_file)) {
691 r = -EPERM;
692 goto out_unlock;
693 }
694
695 readonly = amdgpu_ttm_tt_is_readonly(ttm);
696 r = amdgpu_hmm_range_get_pages(&bo->notifier, start, ttm->num_pages,
697 readonly, NULL, pages, range);
698out_unlock:
699 mmap_read_unlock(mm);
700 if (r)
701 pr_debug("failed %d to get user pages 0x%lx\n", r, start);
702
703 mmput(mm);
704
705 return r;
706}
707
708/* amdgpu_ttm_tt_discard_user_pages - Discard range and pfn array allocations
709 */
710void amdgpu_ttm_tt_discard_user_pages(struct ttm_tt *ttm,
711 struct hmm_range *range)
712{
713 struct amdgpu_ttm_tt *gtt = (void *)ttm;
714
715 if (gtt && gtt->userptr && range)
716 amdgpu_hmm_range_get_pages_done(range);
717}
718
719/*
720 * amdgpu_ttm_tt_get_user_pages_done - stop HMM track the CPU page table change
721 * Check if the pages backing this ttm range have been invalidated
722 *
723 * Returns: true if pages are still valid
724 */
725bool amdgpu_ttm_tt_get_user_pages_done(struct ttm_tt *ttm,
726 struct hmm_range *range)
727{
728 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
729
730 if (!gtt || !gtt->userptr || !range)
731 return false;
732
733 DRM_DEBUG_DRIVER("user_pages_done 0x%llx pages 0x%x\n",
734 gtt->userptr, ttm->num_pages);
735
736 WARN_ONCE(!range->hmm_pfns, "No user pages to check\n");
737
738 return !amdgpu_hmm_range_get_pages_done(range);
739}
740#endif
741
742/*
743 * amdgpu_ttm_tt_set_user_pages - Copy pages in, putting old pages as necessary.
744 *
745 * Called by amdgpu_cs_list_validate(). This creates the page list
746 * that backs user memory and will ultimately be mapped into the device
747 * address space.
748 */
749void amdgpu_ttm_tt_set_user_pages(struct ttm_tt *ttm, struct page **pages)
750{
751 unsigned long i;
752
753 for (i = 0; i < ttm->num_pages; ++i)
754 ttm->pages[i] = pages ? pages[i] : NULL;
755}
756
757/*
758 * amdgpu_ttm_tt_pin_userptr - prepare the sg table with the user pages
759 *
760 * Called by amdgpu_ttm_backend_bind()
761 **/
762static int amdgpu_ttm_tt_pin_userptr(struct ttm_device *bdev,
763 struct ttm_tt *ttm)
764{
765 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
766 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
767 int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
768 enum dma_data_direction direction = write ?
769 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
770 int r;
771
772 /* Allocate an SG array and squash pages into it */
773 r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
774 (u64)ttm->num_pages << PAGE_SHIFT,
775 GFP_KERNEL);
776 if (r)
777 goto release_sg;
778
779 /* Map SG to device */
780 r = dma_map_sgtable(adev->dev, ttm->sg, direction, 0);
781 if (r)
782 goto release_sg;
783
784 /* convert SG to linear array of pages and dma addresses */
785 drm_prime_sg_to_dma_addr_array(ttm->sg, gtt->ttm.dma_address,
786 ttm->num_pages);
787
788 return 0;
789
790release_sg:
791 kfree(ttm->sg);
792 ttm->sg = NULL;
793 return r;
794}
795
796/*
797 * amdgpu_ttm_tt_unpin_userptr - Unpin and unmap userptr pages
798 */
799static void amdgpu_ttm_tt_unpin_userptr(struct ttm_device *bdev,
800 struct ttm_tt *ttm)
801{
802 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
803 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
804 int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
805 enum dma_data_direction direction = write ?
806 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
807
808 /* double check that we don't free the table twice */
809 if (!ttm->sg || !ttm->sg->sgl)
810 return;
811
812 /* unmap the pages mapped to the device */
813 dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
814 sg_free_table(ttm->sg);
815}
816
817/*
818 * total_pages is constructed as MQD0+CtrlStack0 + MQD1+CtrlStack1 + ...
819 * MQDn+CtrlStackn where n is the number of XCCs per partition.
820 * pages_per_xcc is the size of one MQD+CtrlStack. The first page is MQD
821 * and uses memory type default, UC. The rest of pages_per_xcc are
822 * Ctrl stack and modify their memory type to NC.
823 */
824static void amdgpu_ttm_gart_bind_gfx9_mqd(struct amdgpu_device *adev,
825 struct ttm_tt *ttm, uint64_t flags)
826{
827 struct amdgpu_ttm_tt *gtt = (void *)ttm;
828 uint64_t total_pages = ttm->num_pages;
829 int num_xcc = max(1U, adev->gfx.num_xcc_per_xcp);
830 uint64_t page_idx, pages_per_xcc;
831 int i;
832 uint64_t ctrl_flags = (flags & ~AMDGPU_PTE_MTYPE_VG10_MASK) |
833 AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_NC);
834
835 pages_per_xcc = total_pages;
836 do_div(pages_per_xcc, num_xcc);
837
838 for (i = 0, page_idx = 0; i < num_xcc; i++, page_idx += pages_per_xcc) {
839 /* MQD page: use default flags */
840 amdgpu_gart_bind(adev,
841 gtt->offset + (page_idx << PAGE_SHIFT),
842 1, >t->ttm.dma_address[page_idx], flags);
843 /*
844 * Ctrl pages - modify the memory type to NC (ctrl_flags) from
845 * the second page of the BO onward.
846 */
847 amdgpu_gart_bind(adev,
848 gtt->offset + ((page_idx + 1) << PAGE_SHIFT),
849 pages_per_xcc - 1,
850 >t->ttm.dma_address[page_idx + 1],
851 ctrl_flags);
852 }
853}
854
855static void amdgpu_ttm_gart_bind(struct amdgpu_device *adev,
856 struct ttm_buffer_object *tbo,
857 uint64_t flags)
858{
859 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(tbo);
860 struct ttm_tt *ttm = tbo->ttm;
861 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
862
863 if (amdgpu_bo_encrypted(abo))
864 flags |= AMDGPU_PTE_TMZ;
865
866 if (abo->flags & AMDGPU_GEM_CREATE_CP_MQD_GFX9) {
867 amdgpu_ttm_gart_bind_gfx9_mqd(adev, ttm, flags);
868 } else {
869 amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
870 gtt->ttm.dma_address, flags);
871 }
872}
873
874/*
875 * amdgpu_ttm_backend_bind - Bind GTT memory
876 *
877 * Called by ttm_tt_bind() on behalf of ttm_bo_handle_move_mem().
878 * This handles binding GTT memory to the device address space.
879 */
880static int amdgpu_ttm_backend_bind(struct ttm_device *bdev,
881 struct ttm_tt *ttm,
882 struct ttm_resource *bo_mem)
883{
884 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
885 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
886 uint64_t flags;
887 int r;
888
889 if (!bo_mem)
890 return -EINVAL;
891
892 if (gtt->bound)
893 return 0;
894
895 if (gtt->userptr) {
896 r = amdgpu_ttm_tt_pin_userptr(bdev, ttm);
897 if (r) {
898 DRM_ERROR("failed to pin userptr\n");
899 return r;
900 }
901 } else if (ttm->page_flags & TTM_TT_FLAG_EXTERNAL) {
902 if (!ttm->sg) {
903 struct dma_buf_attachment *attach;
904 struct sg_table *sgt;
905
906 attach = gtt->gobj->import_attach;
907 sgt = dma_buf_map_attachment(attach, DMA_BIDIRECTIONAL);
908 if (IS_ERR(sgt))
909 return PTR_ERR(sgt);
910
911 ttm->sg = sgt;
912 }
913
914 drm_prime_sg_to_dma_addr_array(ttm->sg, gtt->ttm.dma_address,
915 ttm->num_pages);
916 }
917
918 if (!ttm->num_pages) {
919 WARN(1, "nothing to bind %u pages for mreg %p back %p!\n",
920 ttm->num_pages, bo_mem, ttm);
921 }
922
923 if (bo_mem->mem_type != TTM_PL_TT ||
924 !amdgpu_gtt_mgr_has_gart_addr(bo_mem)) {
925 gtt->offset = AMDGPU_BO_INVALID_OFFSET;
926 return 0;
927 }
928
929 /* compute PTE flags relevant to this BO memory */
930 flags = amdgpu_ttm_tt_pte_flags(adev, ttm, bo_mem);
931
932 /* bind pages into GART page tables */
933 gtt->offset = (u64)bo_mem->start << PAGE_SHIFT;
934 amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
935 gtt->ttm.dma_address, flags);
936 gtt->bound = true;
937 return 0;
938}
939
940/*
941 * amdgpu_ttm_alloc_gart - Make sure buffer object is accessible either
942 * through AGP or GART aperture.
943 *
944 * If bo is accessible through AGP aperture, then use AGP aperture
945 * to access bo; otherwise allocate logical space in GART aperture
946 * and map bo to GART aperture.
947 */
948int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo)
949{
950 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
951 struct ttm_operation_ctx ctx = { false, false };
952 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(bo->ttm);
953 struct ttm_placement placement;
954 struct ttm_place placements;
955 struct ttm_resource *tmp;
956 uint64_t addr, flags;
957 int r;
958
959 if (bo->resource->start != AMDGPU_BO_INVALID_OFFSET)
960 return 0;
961
962 addr = amdgpu_gmc_agp_addr(bo);
963 if (addr != AMDGPU_BO_INVALID_OFFSET)
964 return 0;
965
966 /* allocate GART space */
967 placement.num_placement = 1;
968 placement.placement = &placements;
969 placement.num_busy_placement = 1;
970 placement.busy_placement = &placements;
971 placements.fpfn = 0;
972 placements.lpfn = adev->gmc.gart_size >> PAGE_SHIFT;
973 placements.mem_type = TTM_PL_TT;
974 placements.flags = bo->resource->placement;
975
976 r = ttm_bo_mem_space(bo, &placement, &tmp, &ctx);
977 if (unlikely(r))
978 return r;
979
980 /* compute PTE flags for this buffer object */
981 flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, tmp);
982
983 /* Bind pages */
984 gtt->offset = (u64)tmp->start << PAGE_SHIFT;
985 amdgpu_ttm_gart_bind(adev, bo, flags);
986 amdgpu_gart_invalidate_tlb(adev);
987 ttm_resource_free(bo, &bo->resource);
988 ttm_bo_assign_mem(bo, tmp);
989
990 return 0;
991}
992
993/*
994 * amdgpu_ttm_recover_gart - Rebind GTT pages
995 *
996 * Called by amdgpu_gtt_mgr_recover() from amdgpu_device_reset() to
997 * rebind GTT pages during a GPU reset.
998 */
999void amdgpu_ttm_recover_gart(struct ttm_buffer_object *tbo)
1000{
1001 struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
1002 uint64_t flags;
1003
1004 if (!tbo->ttm)
1005 return;
1006
1007 flags = amdgpu_ttm_tt_pte_flags(adev, tbo->ttm, tbo->resource);
1008 amdgpu_ttm_gart_bind(adev, tbo, flags);
1009}
1010
1011/*
1012 * amdgpu_ttm_backend_unbind - Unbind GTT mapped pages
1013 *
1014 * Called by ttm_tt_unbind() on behalf of ttm_bo_move_ttm() and
1015 * ttm_tt_destroy().
1016 */
1017static void amdgpu_ttm_backend_unbind(struct ttm_device *bdev,
1018 struct ttm_tt *ttm)
1019{
1020 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
1021 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1022
1023 /* if the pages have userptr pinning then clear that first */
1024 if (gtt->userptr) {
1025 amdgpu_ttm_tt_unpin_userptr(bdev, ttm);
1026 } else if (ttm->sg && gtt->gobj->import_attach) {
1027 struct dma_buf_attachment *attach;
1028
1029 attach = gtt->gobj->import_attach;
1030 dma_buf_unmap_attachment(attach, ttm->sg, DMA_BIDIRECTIONAL);
1031 ttm->sg = NULL;
1032 }
1033
1034 if (!gtt->bound)
1035 return;
1036
1037 if (gtt->offset == AMDGPU_BO_INVALID_OFFSET)
1038 return;
1039
1040 /* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */
1041 amdgpu_gart_unbind(adev, gtt->offset, ttm->num_pages);
1042 gtt->bound = false;
1043}
1044
1045static void amdgpu_ttm_backend_destroy(struct ttm_device *bdev,
1046 struct ttm_tt *ttm)
1047{
1048 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1049
1050 if (gtt->usertask)
1051 put_task_struct(gtt->usertask);
1052
1053 ttm_tt_fini(>t->ttm);
1054 kfree(gtt);
1055}
1056
1057/**
1058 * amdgpu_ttm_tt_create - Create a ttm_tt object for a given BO
1059 *
1060 * @bo: The buffer object to create a GTT ttm_tt object around
1061 * @page_flags: Page flags to be added to the ttm_tt object
1062 *
1063 * Called by ttm_tt_create().
1064 */
1065static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_buffer_object *bo,
1066 uint32_t page_flags)
1067{
1068 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1069 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1070 struct amdgpu_ttm_tt *gtt;
1071 enum ttm_caching caching;
1072
1073 gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL);
1074 if (!gtt)
1075 return NULL;
1076
1077 gtt->gobj = &bo->base;
1078 if (adev->gmc.mem_partitions && abo->xcp_id >= 0)
1079 gtt->pool_id = KFD_XCP_MEM_ID(adev, abo->xcp_id);
1080 else
1081 gtt->pool_id = abo->xcp_id;
1082
1083 if (abo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
1084 caching = ttm_write_combined;
1085 else
1086 caching = ttm_cached;
1087
1088 /* allocate space for the uninitialized page entries */
1089 if (ttm_sg_tt_init(>t->ttm, bo, page_flags, caching)) {
1090 kfree(gtt);
1091 return NULL;
1092 }
1093 return >t->ttm;
1094}
1095
1096/*
1097 * amdgpu_ttm_tt_populate - Map GTT pages visible to the device
1098 *
1099 * Map the pages of a ttm_tt object to an address space visible
1100 * to the underlying device.
1101 */
1102static int amdgpu_ttm_tt_populate(struct ttm_device *bdev,
1103 struct ttm_tt *ttm,
1104 struct ttm_operation_ctx *ctx)
1105{
1106 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
1107 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1108 struct ttm_pool *pool;
1109 pgoff_t i;
1110 int ret;
1111
1112 /* user pages are bound by amdgpu_ttm_tt_pin_userptr() */
1113 if (gtt->userptr) {
1114 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
1115 if (!ttm->sg)
1116 return -ENOMEM;
1117 return 0;
1118 }
1119
1120 if (ttm->page_flags & TTM_TT_FLAG_EXTERNAL)
1121 return 0;
1122
1123 if (adev->mman.ttm_pools && gtt->pool_id >= 0)
1124 pool = &adev->mman.ttm_pools[gtt->pool_id];
1125 else
1126 pool = &adev->mman.bdev.pool;
1127 ret = ttm_pool_alloc(pool, ttm, ctx);
1128 if (ret)
1129 return ret;
1130
1131 for (i = 0; i < ttm->num_pages; ++i)
1132 ttm->pages[i]->mapping = bdev->dev_mapping;
1133
1134 return 0;
1135}
1136
1137/*
1138 * amdgpu_ttm_tt_unpopulate - unmap GTT pages and unpopulate page arrays
1139 *
1140 * Unmaps pages of a ttm_tt object from the device address space and
1141 * unpopulates the page array backing it.
1142 */
1143static void amdgpu_ttm_tt_unpopulate(struct ttm_device *bdev,
1144 struct ttm_tt *ttm)
1145{
1146 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1147 struct amdgpu_device *adev;
1148 struct ttm_pool *pool;
1149 pgoff_t i;
1150
1151 amdgpu_ttm_backend_unbind(bdev, ttm);
1152
1153 if (gtt->userptr) {
1154 amdgpu_ttm_tt_set_user_pages(ttm, NULL);
1155 kfree(ttm->sg);
1156 ttm->sg = NULL;
1157 return;
1158 }
1159
1160 if (ttm->page_flags & TTM_TT_FLAG_EXTERNAL)
1161 return;
1162
1163 for (i = 0; i < ttm->num_pages; ++i)
1164 ttm->pages[i]->mapping = NULL;
1165
1166 adev = amdgpu_ttm_adev(bdev);
1167
1168 if (adev->mman.ttm_pools && gtt->pool_id >= 0)
1169 pool = &adev->mman.ttm_pools[gtt->pool_id];
1170 else
1171 pool = &adev->mman.bdev.pool;
1172
1173 return ttm_pool_free(pool, ttm);
1174}
1175
1176/**
1177 * amdgpu_ttm_tt_get_userptr - Return the userptr GTT ttm_tt for the current
1178 * task
1179 *
1180 * @tbo: The ttm_buffer_object that contains the userptr
1181 * @user_addr: The returned value
1182 */
1183int amdgpu_ttm_tt_get_userptr(const struct ttm_buffer_object *tbo,
1184 uint64_t *user_addr)
1185{
1186 struct amdgpu_ttm_tt *gtt;
1187
1188 if (!tbo->ttm)
1189 return -EINVAL;
1190
1191 gtt = (void *)tbo->ttm;
1192 *user_addr = gtt->userptr;
1193 return 0;
1194}
1195
1196/**
1197 * amdgpu_ttm_tt_set_userptr - Initialize userptr GTT ttm_tt for the current
1198 * task
1199 *
1200 * @bo: The ttm_buffer_object to bind this userptr to
1201 * @addr: The address in the current tasks VM space to use
1202 * @flags: Requirements of userptr object.
1203 *
1204 * Called by amdgpu_gem_userptr_ioctl() and kfd_ioctl_alloc_memory_of_gpu() to
1205 * bind userptr pages to current task and by kfd_ioctl_acquire_vm() to
1206 * initialize GPU VM for a KFD process.
1207 */
1208int amdgpu_ttm_tt_set_userptr(struct ttm_buffer_object *bo,
1209 uint64_t addr, uint32_t flags)
1210{
1211 struct amdgpu_ttm_tt *gtt;
1212
1213 if (!bo->ttm) {
1214 /* TODO: We want a separate TTM object type for userptrs */
1215 bo->ttm = amdgpu_ttm_tt_create(bo, 0);
1216 if (bo->ttm == NULL)
1217 return -ENOMEM;
1218 }
1219
1220 /* Set TTM_TT_FLAG_EXTERNAL before populate but after create. */
1221 bo->ttm->page_flags |= TTM_TT_FLAG_EXTERNAL;
1222
1223 gtt = ttm_to_amdgpu_ttm_tt(bo->ttm);
1224 gtt->userptr = addr;
1225 gtt->userflags = flags;
1226
1227 if (gtt->usertask)
1228 put_task_struct(gtt->usertask);
1229 gtt->usertask = current->group_leader;
1230 get_task_struct(gtt->usertask);
1231
1232 return 0;
1233}
1234
1235/*
1236 * amdgpu_ttm_tt_get_usermm - Return memory manager for ttm_tt object
1237 */
1238struct mm_struct *amdgpu_ttm_tt_get_usermm(struct ttm_tt *ttm)
1239{
1240 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1241
1242 if (gtt == NULL)
1243 return NULL;
1244
1245 if (gtt->usertask == NULL)
1246 return NULL;
1247
1248 return gtt->usertask->mm;
1249}
1250
1251/*
1252 * amdgpu_ttm_tt_affect_userptr - Determine if a ttm_tt object lays inside an
1253 * address range for the current task.
1254 *
1255 */
1256bool amdgpu_ttm_tt_affect_userptr(struct ttm_tt *ttm, unsigned long start,
1257 unsigned long end, unsigned long *userptr)
1258{
1259 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1260 unsigned long size;
1261
1262 if (gtt == NULL || !gtt->userptr)
1263 return false;
1264
1265 /* Return false if no part of the ttm_tt object lies within
1266 * the range
1267 */
1268 size = (unsigned long)gtt->ttm.num_pages * PAGE_SIZE;
1269 if (gtt->userptr > end || gtt->userptr + size <= start)
1270 return false;
1271
1272 if (userptr)
1273 *userptr = gtt->userptr;
1274 return true;
1275}
1276
1277/*
1278 * amdgpu_ttm_tt_is_userptr - Have the pages backing by userptr?
1279 */
1280bool amdgpu_ttm_tt_is_userptr(struct ttm_tt *ttm)
1281{
1282 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1283
1284 if (gtt == NULL || !gtt->userptr)
1285 return false;
1286
1287 return true;
1288}
1289
1290/*
1291 * amdgpu_ttm_tt_is_readonly - Is the ttm_tt object read only?
1292 */
1293bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm)
1294{
1295 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1296
1297 if (gtt == NULL)
1298 return false;
1299
1300 return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
1301}
1302
1303/**
1304 * amdgpu_ttm_tt_pde_flags - Compute PDE flags for ttm_tt object
1305 *
1306 * @ttm: The ttm_tt object to compute the flags for
1307 * @mem: The memory registry backing this ttm_tt object
1308 *
1309 * Figure out the flags to use for a VM PDE (Page Directory Entry).
1310 */
1311uint64_t amdgpu_ttm_tt_pde_flags(struct ttm_tt *ttm, struct ttm_resource *mem)
1312{
1313 uint64_t flags = 0;
1314
1315 if (mem && mem->mem_type != TTM_PL_SYSTEM)
1316 flags |= AMDGPU_PTE_VALID;
1317
1318 if (mem && (mem->mem_type == TTM_PL_TT ||
1319 mem->mem_type == AMDGPU_PL_DOORBELL ||
1320 mem->mem_type == AMDGPU_PL_PREEMPT)) {
1321 flags |= AMDGPU_PTE_SYSTEM;
1322
1323 if (ttm->caching == ttm_cached)
1324 flags |= AMDGPU_PTE_SNOOPED;
1325 }
1326
1327 if (mem && mem->mem_type == TTM_PL_VRAM &&
1328 mem->bus.caching == ttm_cached)
1329 flags |= AMDGPU_PTE_SNOOPED;
1330
1331 return flags;
1332}
1333
1334/**
1335 * amdgpu_ttm_tt_pte_flags - Compute PTE flags for ttm_tt object
1336 *
1337 * @adev: amdgpu_device pointer
1338 * @ttm: The ttm_tt object to compute the flags for
1339 * @mem: The memory registry backing this ttm_tt object
1340 *
1341 * Figure out the flags to use for a VM PTE (Page Table Entry).
1342 */
1343uint64_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
1344 struct ttm_resource *mem)
1345{
1346 uint64_t flags = amdgpu_ttm_tt_pde_flags(ttm, mem);
1347
1348 flags |= adev->gart.gart_pte_flags;
1349 flags |= AMDGPU_PTE_READABLE;
1350
1351 if (!amdgpu_ttm_tt_is_readonly(ttm))
1352 flags |= AMDGPU_PTE_WRITEABLE;
1353
1354 return flags;
1355}
1356
1357/*
1358 * amdgpu_ttm_bo_eviction_valuable - Check to see if we can evict a buffer
1359 * object.
1360 *
1361 * Return true if eviction is sensible. Called by ttm_mem_evict_first() on
1362 * behalf of ttm_bo_mem_force_space() which tries to evict buffer objects until
1363 * it can find space for a new object and by ttm_bo_force_list_clean() which is
1364 * used to clean out a memory space.
1365 */
1366static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
1367 const struct ttm_place *place)
1368{
1369 struct dma_resv_iter resv_cursor;
1370 struct dma_fence *f;
1371
1372 if (!amdgpu_bo_is_amdgpu_bo(bo))
1373 return ttm_bo_eviction_valuable(bo, place);
1374
1375 /* Swapout? */
1376 if (bo->resource->mem_type == TTM_PL_SYSTEM)
1377 return true;
1378
1379 if (bo->type == ttm_bo_type_kernel &&
1380 !amdgpu_vm_evictable(ttm_to_amdgpu_bo(bo)))
1381 return false;
1382
1383 /* If bo is a KFD BO, check if the bo belongs to the current process.
1384 * If true, then return false as any KFD process needs all its BOs to
1385 * be resident to run successfully
1386 */
1387 dma_resv_for_each_fence(&resv_cursor, bo->base.resv,
1388 DMA_RESV_USAGE_BOOKKEEP, f) {
1389 if (amdkfd_fence_check_mm(f, current->mm))
1390 return false;
1391 }
1392
1393 /* Preemptible BOs don't own system resources managed by the
1394 * driver (pages, VRAM, GART space). They point to resources
1395 * owned by someone else (e.g. pageable memory in user mode
1396 * or a DMABuf). They are used in a preemptible context so we
1397 * can guarantee no deadlocks and good QoS in case of MMU
1398 * notifiers or DMABuf move notifiers from the resource owner.
1399 */
1400 if (bo->resource->mem_type == AMDGPU_PL_PREEMPT)
1401 return false;
1402
1403 if (bo->resource->mem_type == TTM_PL_TT &&
1404 amdgpu_bo_encrypted(ttm_to_amdgpu_bo(bo)))
1405 return false;
1406
1407 return ttm_bo_eviction_valuable(bo, place);
1408}
1409
1410static void amdgpu_ttm_vram_mm_access(struct amdgpu_device *adev, loff_t pos,
1411 void *buf, size_t size, bool write)
1412{
1413 while (size) {
1414 uint64_t aligned_pos = ALIGN_DOWN(pos, 4);
1415 uint64_t bytes = 4 - (pos & 0x3);
1416 uint32_t shift = (pos & 0x3) * 8;
1417 uint32_t mask = 0xffffffff << shift;
1418 uint32_t value = 0;
1419
1420 if (size < bytes) {
1421 mask &= 0xffffffff >> (bytes - size) * 8;
1422 bytes = size;
1423 }
1424
1425 if (mask != 0xffffffff) {
1426 amdgpu_device_mm_access(adev, aligned_pos, &value, 4, false);
1427 if (write) {
1428 value &= ~mask;
1429 value |= (*(uint32_t *)buf << shift) & mask;
1430 amdgpu_device_mm_access(adev, aligned_pos, &value, 4, true);
1431 } else {
1432 value = (value & mask) >> shift;
1433 memcpy(buf, &value, bytes);
1434 }
1435 } else {
1436 amdgpu_device_mm_access(adev, aligned_pos, buf, 4, write);
1437 }
1438
1439 pos += bytes;
1440 buf += bytes;
1441 size -= bytes;
1442 }
1443}
1444
1445static int amdgpu_ttm_access_memory_sdma(struct ttm_buffer_object *bo,
1446 unsigned long offset, void *buf,
1447 int len, int write)
1448{
1449 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1450 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
1451 struct amdgpu_res_cursor src_mm;
1452 struct amdgpu_job *job;
1453 struct dma_fence *fence;
1454 uint64_t src_addr, dst_addr;
1455 unsigned int num_dw;
1456 int r, idx;
1457
1458 if (len != PAGE_SIZE)
1459 return -EINVAL;
1460
1461 if (!adev->mman.sdma_access_ptr)
1462 return -EACCES;
1463
1464 if (!drm_dev_enter(adev_to_drm(adev), &idx))
1465 return -ENODEV;
1466
1467 if (write)
1468 memcpy(adev->mman.sdma_access_ptr, buf, len);
1469
1470 num_dw = ALIGN(adev->mman.buffer_funcs->copy_num_dw, 8);
1471 r = amdgpu_job_alloc_with_ib(adev, &adev->mman.high_pr,
1472 AMDGPU_FENCE_OWNER_UNDEFINED,
1473 num_dw * 4, AMDGPU_IB_POOL_DELAYED,
1474 &job);
1475 if (r)
1476 goto out;
1477
1478 amdgpu_res_first(abo->tbo.resource, offset, len, &src_mm);
1479 src_addr = amdgpu_ttm_domain_start(adev, bo->resource->mem_type) +
1480 src_mm.start;
1481 dst_addr = amdgpu_bo_gpu_offset(adev->mman.sdma_access_bo);
1482 if (write)
1483 swap(src_addr, dst_addr);
1484
1485 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr, dst_addr,
1486 PAGE_SIZE, false);
1487
1488 amdgpu_ring_pad_ib(adev->mman.buffer_funcs_ring, &job->ibs[0]);
1489 WARN_ON(job->ibs[0].length_dw > num_dw);
1490
1491 fence = amdgpu_job_submit(job);
1492
1493 if (!dma_fence_wait_timeout(fence, false, adev->sdma_timeout))
1494 r = -ETIMEDOUT;
1495 dma_fence_put(fence);
1496
1497 if (!(r || write))
1498 memcpy(buf, adev->mman.sdma_access_ptr, len);
1499out:
1500 drm_dev_exit(idx);
1501 return r;
1502}
1503
1504/**
1505 * amdgpu_ttm_access_memory - Read or Write memory that backs a buffer object.
1506 *
1507 * @bo: The buffer object to read/write
1508 * @offset: Offset into buffer object
1509 * @buf: Secondary buffer to write/read from
1510 * @len: Length in bytes of access
1511 * @write: true if writing
1512 *
1513 * This is used to access VRAM that backs a buffer object via MMIO
1514 * access for debugging purposes.
1515 */
1516static int amdgpu_ttm_access_memory(struct ttm_buffer_object *bo,
1517 unsigned long offset, void *buf, int len,
1518 int write)
1519{
1520 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1521 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
1522 struct amdgpu_res_cursor cursor;
1523 int ret = 0;
1524
1525 if (bo->resource->mem_type != TTM_PL_VRAM)
1526 return -EIO;
1527
1528 if (amdgpu_device_has_timeouts_enabled(adev) &&
1529 !amdgpu_ttm_access_memory_sdma(bo, offset, buf, len, write))
1530 return len;
1531
1532 amdgpu_res_first(bo->resource, offset, len, &cursor);
1533 while (cursor.remaining) {
1534 size_t count, size = cursor.size;
1535 loff_t pos = cursor.start;
1536
1537 count = amdgpu_device_aper_access(adev, pos, buf, size, write);
1538 size -= count;
1539 if (size) {
1540 /* using MM to access rest vram and handle un-aligned address */
1541 pos += count;
1542 buf += count;
1543 amdgpu_ttm_vram_mm_access(adev, pos, buf, size, write);
1544 }
1545
1546 ret += cursor.size;
1547 buf += cursor.size;
1548 amdgpu_res_next(&cursor, cursor.size);
1549 }
1550
1551 return ret;
1552}
1553
1554static void
1555amdgpu_bo_delete_mem_notify(struct ttm_buffer_object *bo)
1556{
1557 amdgpu_bo_move_notify(bo, false);
1558}
1559
1560static struct ttm_device_funcs amdgpu_bo_driver = {
1561 .ttm_tt_create = &amdgpu_ttm_tt_create,
1562 .ttm_tt_populate = &amdgpu_ttm_tt_populate,
1563 .ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate,
1564 .ttm_tt_destroy = &amdgpu_ttm_backend_destroy,
1565 .eviction_valuable = amdgpu_ttm_bo_eviction_valuable,
1566 .evict_flags = &amdgpu_evict_flags,
1567 .move = &amdgpu_bo_move,
1568 .delete_mem_notify = &amdgpu_bo_delete_mem_notify,
1569 .release_notify = &amdgpu_bo_release_notify,
1570 .io_mem_reserve = &amdgpu_ttm_io_mem_reserve,
1571 .io_mem_pfn = amdgpu_ttm_io_mem_pfn,
1572 .access_memory = &amdgpu_ttm_access_memory,
1573};
1574
1575/*
1576 * Firmware Reservation functions
1577 */
1578/**
1579 * amdgpu_ttm_fw_reserve_vram_fini - free fw reserved vram
1580 *
1581 * @adev: amdgpu_device pointer
1582 *
1583 * free fw reserved vram if it has been reserved.
1584 */
1585static void amdgpu_ttm_fw_reserve_vram_fini(struct amdgpu_device *adev)
1586{
1587 amdgpu_bo_free_kernel(&adev->mman.fw_vram_usage_reserved_bo,
1588 NULL, &adev->mman.fw_vram_usage_va);
1589}
1590
1591/*
1592 * Driver Reservation functions
1593 */
1594/**
1595 * amdgpu_ttm_drv_reserve_vram_fini - free drv reserved vram
1596 *
1597 * @adev: amdgpu_device pointer
1598 *
1599 * free drv reserved vram if it has been reserved.
1600 */
1601static void amdgpu_ttm_drv_reserve_vram_fini(struct amdgpu_device *adev)
1602{
1603 amdgpu_bo_free_kernel(&adev->mman.drv_vram_usage_reserved_bo,
1604 NULL,
1605 &adev->mman.drv_vram_usage_va);
1606}
1607
1608/**
1609 * amdgpu_ttm_fw_reserve_vram_init - create bo vram reservation from fw
1610 *
1611 * @adev: amdgpu_device pointer
1612 *
1613 * create bo vram reservation from fw.
1614 */
1615static int amdgpu_ttm_fw_reserve_vram_init(struct amdgpu_device *adev)
1616{
1617 uint64_t vram_size = adev->gmc.visible_vram_size;
1618
1619 adev->mman.fw_vram_usage_va = NULL;
1620 adev->mman.fw_vram_usage_reserved_bo = NULL;
1621
1622 if (adev->mman.fw_vram_usage_size == 0 ||
1623 adev->mman.fw_vram_usage_size > vram_size)
1624 return 0;
1625
1626 return amdgpu_bo_create_kernel_at(adev,
1627 adev->mman.fw_vram_usage_start_offset,
1628 adev->mman.fw_vram_usage_size,
1629 &adev->mman.fw_vram_usage_reserved_bo,
1630 &adev->mman.fw_vram_usage_va);
1631}
1632
1633/**
1634 * amdgpu_ttm_drv_reserve_vram_init - create bo vram reservation from driver
1635 *
1636 * @adev: amdgpu_device pointer
1637 *
1638 * create bo vram reservation from drv.
1639 */
1640static int amdgpu_ttm_drv_reserve_vram_init(struct amdgpu_device *adev)
1641{
1642 u64 vram_size = adev->gmc.visible_vram_size;
1643
1644 adev->mman.drv_vram_usage_va = NULL;
1645 adev->mman.drv_vram_usage_reserved_bo = NULL;
1646
1647 if (adev->mman.drv_vram_usage_size == 0 ||
1648 adev->mman.drv_vram_usage_size > vram_size)
1649 return 0;
1650
1651 return amdgpu_bo_create_kernel_at(adev,
1652 adev->mman.drv_vram_usage_start_offset,
1653 adev->mman.drv_vram_usage_size,
1654 &adev->mman.drv_vram_usage_reserved_bo,
1655 &adev->mman.drv_vram_usage_va);
1656}
1657
1658/*
1659 * Memoy training reservation functions
1660 */
1661
1662/**
1663 * amdgpu_ttm_training_reserve_vram_fini - free memory training reserved vram
1664 *
1665 * @adev: amdgpu_device pointer
1666 *
1667 * free memory training reserved vram if it has been reserved.
1668 */
1669static int amdgpu_ttm_training_reserve_vram_fini(struct amdgpu_device *adev)
1670{
1671 struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1672
1673 ctx->init = PSP_MEM_TRAIN_NOT_SUPPORT;
1674 amdgpu_bo_free_kernel(&ctx->c2p_bo, NULL, NULL);
1675 ctx->c2p_bo = NULL;
1676
1677 return 0;
1678}
1679
1680static void amdgpu_ttm_training_data_block_init(struct amdgpu_device *adev,
1681 uint32_t reserve_size)
1682{
1683 struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1684
1685 memset(ctx, 0, sizeof(*ctx));
1686
1687 ctx->c2p_train_data_offset =
1688 ALIGN((adev->gmc.mc_vram_size - reserve_size - SZ_1M), SZ_1M);
1689 ctx->p2c_train_data_offset =
1690 (adev->gmc.mc_vram_size - GDDR6_MEM_TRAINING_OFFSET);
1691 ctx->train_data_size =
1692 GDDR6_MEM_TRAINING_DATA_SIZE_IN_BYTES;
1693
1694 DRM_DEBUG("train_data_size:%llx,p2c_train_data_offset:%llx,c2p_train_data_offset:%llx.\n",
1695 ctx->train_data_size,
1696 ctx->p2c_train_data_offset,
1697 ctx->c2p_train_data_offset);
1698}
1699
1700/*
1701 * reserve TMR memory at the top of VRAM which holds
1702 * IP Discovery data and is protected by PSP.
1703 */
1704static int amdgpu_ttm_reserve_tmr(struct amdgpu_device *adev)
1705{
1706 struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1707 bool mem_train_support = false;
1708 uint32_t reserve_size = 0;
1709 int ret;
1710
1711 if (adev->bios && !amdgpu_sriov_vf(adev)) {
1712 if (amdgpu_atomfirmware_mem_training_supported(adev))
1713 mem_train_support = true;
1714 else
1715 DRM_DEBUG("memory training does not support!\n");
1716 }
1717
1718 /*
1719 * Query reserved tmr size through atom firmwareinfo for Sienna_Cichlid and onwards for all
1720 * the use cases (IP discovery/G6 memory training/profiling/diagnostic data.etc)
1721 *
1722 * Otherwise, fallback to legacy approach to check and reserve tmr block for ip
1723 * discovery data and G6 memory training data respectively
1724 */
1725 if (adev->bios)
1726 reserve_size =
1727 amdgpu_atomfirmware_get_fw_reserved_fb_size(adev);
1728
1729 if (!adev->bios &&
1730 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3))
1731 reserve_size = max(reserve_size, (uint32_t)280 << 20);
1732 else if (!reserve_size)
1733 reserve_size = DISCOVERY_TMR_OFFSET;
1734
1735 if (mem_train_support) {
1736 /* reserve vram for mem train according to TMR location */
1737 amdgpu_ttm_training_data_block_init(adev, reserve_size);
1738 ret = amdgpu_bo_create_kernel_at(adev,
1739 ctx->c2p_train_data_offset,
1740 ctx->train_data_size,
1741 &ctx->c2p_bo,
1742 NULL);
1743 if (ret) {
1744 DRM_ERROR("alloc c2p_bo failed(%d)!\n", ret);
1745 amdgpu_ttm_training_reserve_vram_fini(adev);
1746 return ret;
1747 }
1748 ctx->init = PSP_MEM_TRAIN_RESERVE_SUCCESS;
1749 }
1750
1751 if (!adev->gmc.is_app_apu) {
1752 ret = amdgpu_bo_create_kernel_at(
1753 adev, adev->gmc.real_vram_size - reserve_size,
1754 reserve_size, &adev->mman.fw_reserved_memory, NULL);
1755 if (ret) {
1756 DRM_ERROR("alloc tmr failed(%d)!\n", ret);
1757 amdgpu_bo_free_kernel(&adev->mman.fw_reserved_memory,
1758 NULL, NULL);
1759 return ret;
1760 }
1761 } else {
1762 DRM_DEBUG_DRIVER("backdoor fw loading path for PSP TMR, no reservation needed\n");
1763 }
1764
1765 return 0;
1766}
1767
1768static int amdgpu_ttm_pools_init(struct amdgpu_device *adev)
1769{
1770 int i;
1771
1772 if (!adev->gmc.is_app_apu || !adev->gmc.num_mem_partitions)
1773 return 0;
1774
1775 adev->mman.ttm_pools = kcalloc(adev->gmc.num_mem_partitions,
1776 sizeof(*adev->mman.ttm_pools),
1777 GFP_KERNEL);
1778 if (!adev->mman.ttm_pools)
1779 return -ENOMEM;
1780
1781 for (i = 0; i < adev->gmc.num_mem_partitions; i++) {
1782 ttm_pool_init(&adev->mman.ttm_pools[i], adev->dev,
1783 adev->gmc.mem_partitions[i].numa.node,
1784 false, false);
1785 }
1786 return 0;
1787}
1788
1789static void amdgpu_ttm_pools_fini(struct amdgpu_device *adev)
1790{
1791 int i;
1792
1793 if (!adev->gmc.is_app_apu || !adev->mman.ttm_pools)
1794 return;
1795
1796 for (i = 0; i < adev->gmc.num_mem_partitions; i++)
1797 ttm_pool_fini(&adev->mman.ttm_pools[i]);
1798
1799 kfree(adev->mman.ttm_pools);
1800 adev->mman.ttm_pools = NULL;
1801}
1802
1803/*
1804 * amdgpu_ttm_init - Init the memory management (ttm) as well as various
1805 * gtt/vram related fields.
1806 *
1807 * This initializes all of the memory space pools that the TTM layer
1808 * will need such as the GTT space (system memory mapped to the device),
1809 * VRAM (on-board memory), and on-chip memories (GDS, GWS, OA) which
1810 * can be mapped per VMID.
1811 */
1812int amdgpu_ttm_init(struct amdgpu_device *adev)
1813{
1814 uint64_t gtt_size;
1815 int r;
1816
1817 mutex_init(&adev->mman.gtt_window_lock);
1818
1819 /* No others user of address space so set it to 0 */
1820 r = ttm_device_init(&adev->mman.bdev, &amdgpu_bo_driver, adev->dev,
1821 adev_to_drm(adev)->anon_inode->i_mapping,
1822 adev_to_drm(adev)->vma_offset_manager,
1823 adev->need_swiotlb,
1824 dma_addressing_limited(adev->dev));
1825 if (r) {
1826 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
1827 return r;
1828 }
1829
1830 r = amdgpu_ttm_pools_init(adev);
1831 if (r) {
1832 DRM_ERROR("failed to init ttm pools(%d).\n", r);
1833 return r;
1834 }
1835 adev->mman.initialized = true;
1836
1837 /* Initialize VRAM pool with all of VRAM divided into pages */
1838 r = amdgpu_vram_mgr_init(adev);
1839 if (r) {
1840 DRM_ERROR("Failed initializing VRAM heap.\n");
1841 return r;
1842 }
1843
1844 /* Change the size here instead of the init above so only lpfn is affected */
1845 amdgpu_ttm_set_buffer_funcs_status(adev, false);
1846#ifdef CONFIG_64BIT
1847#ifdef CONFIG_X86
1848 if (adev->gmc.xgmi.connected_to_cpu)
1849 adev->mman.aper_base_kaddr = ioremap_cache(adev->gmc.aper_base,
1850 adev->gmc.visible_vram_size);
1851
1852 else if (adev->gmc.is_app_apu)
1853 DRM_DEBUG_DRIVER(
1854 "No need to ioremap when real vram size is 0\n");
1855 else
1856#endif
1857 adev->mman.aper_base_kaddr = ioremap_wc(adev->gmc.aper_base,
1858 adev->gmc.visible_vram_size);
1859#endif
1860
1861 /*
1862 *The reserved vram for firmware must be pinned to the specified
1863 *place on the VRAM, so reserve it early.
1864 */
1865 r = amdgpu_ttm_fw_reserve_vram_init(adev);
1866 if (r)
1867 return r;
1868
1869 /*
1870 *The reserved vram for driver must be pinned to the specified
1871 *place on the VRAM, so reserve it early.
1872 */
1873 r = amdgpu_ttm_drv_reserve_vram_init(adev);
1874 if (r)
1875 return r;
1876
1877 /*
1878 * only NAVI10 and onwards ASIC support for IP discovery.
1879 * If IP discovery enabled, a block of memory should be
1880 * reserved for IP discovey.
1881 */
1882 if (adev->mman.discovery_bin) {
1883 r = amdgpu_ttm_reserve_tmr(adev);
1884 if (r)
1885 return r;
1886 }
1887
1888 /* allocate memory as required for VGA
1889 * This is used for VGA emulation and pre-OS scanout buffers to
1890 * avoid display artifacts while transitioning between pre-OS
1891 * and driver.
1892 */
1893 if (!adev->gmc.is_app_apu) {
1894 r = amdgpu_bo_create_kernel_at(adev, 0,
1895 adev->mman.stolen_vga_size,
1896 &adev->mman.stolen_vga_memory,
1897 NULL);
1898 if (r)
1899 return r;
1900
1901 r = amdgpu_bo_create_kernel_at(adev, adev->mman.stolen_vga_size,
1902 adev->mman.stolen_extended_size,
1903 &adev->mman.stolen_extended_memory,
1904 NULL);
1905
1906 if (r)
1907 return r;
1908
1909 r = amdgpu_bo_create_kernel_at(adev,
1910 adev->mman.stolen_reserved_offset,
1911 adev->mman.stolen_reserved_size,
1912 &adev->mman.stolen_reserved_memory,
1913 NULL);
1914 if (r)
1915 return r;
1916 } else {
1917 DRM_DEBUG_DRIVER("Skipped stolen memory reservation\n");
1918 }
1919
1920 DRM_INFO("amdgpu: %uM of VRAM memory ready\n",
1921 (unsigned int)(adev->gmc.real_vram_size / (1024 * 1024)));
1922
1923 /* Compute GTT size, either based on TTM limit
1924 * or whatever the user passed on module init.
1925 */
1926 if (amdgpu_gtt_size == -1)
1927 gtt_size = ttm_tt_pages_limit() << PAGE_SHIFT;
1928 else
1929 gtt_size = (uint64_t)amdgpu_gtt_size << 20;
1930
1931 /* Initialize GTT memory pool */
1932 r = amdgpu_gtt_mgr_init(adev, gtt_size);
1933 if (r) {
1934 DRM_ERROR("Failed initializing GTT heap.\n");
1935 return r;
1936 }
1937 DRM_INFO("amdgpu: %uM of GTT memory ready.\n",
1938 (unsigned int)(gtt_size / (1024 * 1024)));
1939
1940 /* Initiailize doorbell pool on PCI BAR */
1941 r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_DOORBELL, adev->doorbell.size / PAGE_SIZE);
1942 if (r) {
1943 DRM_ERROR("Failed initializing doorbell heap.\n");
1944 return r;
1945 }
1946
1947 /* Create a boorbell page for kernel usages */
1948 r = amdgpu_doorbell_create_kernel_doorbells(adev);
1949 if (r) {
1950 DRM_ERROR("Failed to initialize kernel doorbells.\n");
1951 return r;
1952 }
1953
1954 /* Initialize preemptible memory pool */
1955 r = amdgpu_preempt_mgr_init(adev);
1956 if (r) {
1957 DRM_ERROR("Failed initializing PREEMPT heap.\n");
1958 return r;
1959 }
1960
1961 /* Initialize various on-chip memory pools */
1962 r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GDS, adev->gds.gds_size);
1963 if (r) {
1964 DRM_ERROR("Failed initializing GDS heap.\n");
1965 return r;
1966 }
1967
1968 r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GWS, adev->gds.gws_size);
1969 if (r) {
1970 DRM_ERROR("Failed initializing gws heap.\n");
1971 return r;
1972 }
1973
1974 r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_OA, adev->gds.oa_size);
1975 if (r) {
1976 DRM_ERROR("Failed initializing oa heap.\n");
1977 return r;
1978 }
1979 if (amdgpu_bo_create_kernel(adev, PAGE_SIZE, PAGE_SIZE,
1980 AMDGPU_GEM_DOMAIN_GTT,
1981 &adev->mman.sdma_access_bo, NULL,
1982 &adev->mman.sdma_access_ptr))
1983 DRM_WARN("Debug VRAM access will use slowpath MM access\n");
1984
1985 return 0;
1986}
1987
1988/*
1989 * amdgpu_ttm_fini - De-initialize the TTM memory pools
1990 */
1991void amdgpu_ttm_fini(struct amdgpu_device *adev)
1992{
1993 int idx;
1994
1995 if (!adev->mman.initialized)
1996 return;
1997
1998 amdgpu_ttm_pools_fini(adev);
1999
2000 amdgpu_ttm_training_reserve_vram_fini(adev);
2001 /* return the stolen vga memory back to VRAM */
2002 if (!adev->gmc.is_app_apu) {
2003 amdgpu_bo_free_kernel(&adev->mman.stolen_vga_memory, NULL, NULL);
2004 amdgpu_bo_free_kernel(&adev->mman.stolen_extended_memory, NULL, NULL);
2005 /* return the FW reserved memory back to VRAM */
2006 amdgpu_bo_free_kernel(&adev->mman.fw_reserved_memory, NULL,
2007 NULL);
2008 if (adev->mman.stolen_reserved_size)
2009 amdgpu_bo_free_kernel(&adev->mman.stolen_reserved_memory,
2010 NULL, NULL);
2011 }
2012 amdgpu_bo_free_kernel(&adev->mman.sdma_access_bo, NULL,
2013 &adev->mman.sdma_access_ptr);
2014 amdgpu_ttm_fw_reserve_vram_fini(adev);
2015 amdgpu_ttm_drv_reserve_vram_fini(adev);
2016
2017 if (drm_dev_enter(adev_to_drm(adev), &idx)) {
2018
2019 if (adev->mman.aper_base_kaddr)
2020 iounmap(adev->mman.aper_base_kaddr);
2021 adev->mman.aper_base_kaddr = NULL;
2022
2023 drm_dev_exit(idx);
2024 }
2025
2026 amdgpu_vram_mgr_fini(adev);
2027 amdgpu_gtt_mgr_fini(adev);
2028 amdgpu_preempt_mgr_fini(adev);
2029 ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GDS);
2030 ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GWS);
2031 ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_OA);
2032 ttm_device_fini(&adev->mman.bdev);
2033 adev->mman.initialized = false;
2034 DRM_INFO("amdgpu: ttm finalized\n");
2035}
2036
2037/**
2038 * amdgpu_ttm_set_buffer_funcs_status - enable/disable use of buffer functions
2039 *
2040 * @adev: amdgpu_device pointer
2041 * @enable: true when we can use buffer functions.
2042 *
2043 * Enable/disable use of buffer functions during suspend/resume. This should
2044 * only be called at bootup or when userspace isn't running.
2045 */
2046void amdgpu_ttm_set_buffer_funcs_status(struct amdgpu_device *adev, bool enable)
2047{
2048 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
2049 uint64_t size;
2050 int r;
2051
2052 if (!adev->mman.initialized || amdgpu_in_reset(adev) ||
2053 adev->mman.buffer_funcs_enabled == enable || adev->gmc.is_app_apu)
2054 return;
2055
2056 if (enable) {
2057 struct amdgpu_ring *ring;
2058 struct drm_gpu_scheduler *sched;
2059
2060 ring = adev->mman.buffer_funcs_ring;
2061 sched = &ring->sched;
2062 r = drm_sched_entity_init(&adev->mman.high_pr,
2063 DRM_SCHED_PRIORITY_KERNEL, &sched,
2064 1, NULL);
2065 if (r) {
2066 DRM_ERROR("Failed setting up TTM BO move entity (%d)\n",
2067 r);
2068 return;
2069 }
2070
2071 r = drm_sched_entity_init(&adev->mman.low_pr,
2072 DRM_SCHED_PRIORITY_NORMAL, &sched,
2073 1, NULL);
2074 if (r) {
2075 DRM_ERROR("Failed setting up TTM BO move entity (%d)\n",
2076 r);
2077 goto error_free_entity;
2078 }
2079 } else {
2080 drm_sched_entity_destroy(&adev->mman.high_pr);
2081 drm_sched_entity_destroy(&adev->mman.low_pr);
2082 dma_fence_put(man->move);
2083 man->move = NULL;
2084 }
2085
2086 /* this just adjusts TTM size idea, which sets lpfn to the correct value */
2087 if (enable)
2088 size = adev->gmc.real_vram_size;
2089 else
2090 size = adev->gmc.visible_vram_size;
2091 man->size = size;
2092 adev->mman.buffer_funcs_enabled = enable;
2093
2094 return;
2095
2096error_free_entity:
2097 drm_sched_entity_destroy(&adev->mman.high_pr);
2098}
2099
2100static int amdgpu_ttm_prepare_job(struct amdgpu_device *adev,
2101 bool direct_submit,
2102 unsigned int num_dw,
2103 struct dma_resv *resv,
2104 bool vm_needs_flush,
2105 struct amdgpu_job **job,
2106 bool delayed)
2107{
2108 enum amdgpu_ib_pool_type pool = direct_submit ?
2109 AMDGPU_IB_POOL_DIRECT :
2110 AMDGPU_IB_POOL_DELAYED;
2111 int r;
2112 struct drm_sched_entity *entity = delayed ? &adev->mman.low_pr :
2113 &adev->mman.high_pr;
2114 r = amdgpu_job_alloc_with_ib(adev, entity,
2115 AMDGPU_FENCE_OWNER_UNDEFINED,
2116 num_dw * 4, pool, job);
2117 if (r)
2118 return r;
2119
2120 if (vm_needs_flush) {
2121 (*job)->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gmc.pdb0_bo ?
2122 adev->gmc.pdb0_bo :
2123 adev->gart.bo);
2124 (*job)->vm_needs_flush = true;
2125 }
2126 if (!resv)
2127 return 0;
2128
2129 return drm_sched_job_add_resv_dependencies(&(*job)->base, resv,
2130 DMA_RESV_USAGE_BOOKKEEP);
2131}
2132
2133int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
2134 uint64_t dst_offset, uint32_t byte_count,
2135 struct dma_resv *resv,
2136 struct dma_fence **fence, bool direct_submit,
2137 bool vm_needs_flush, bool tmz)
2138{
2139 struct amdgpu_device *adev = ring->adev;
2140 unsigned int num_loops, num_dw;
2141 struct amdgpu_job *job;
2142 uint32_t max_bytes;
2143 unsigned int i;
2144 int r;
2145
2146 if (!direct_submit && !ring->sched.ready) {
2147 DRM_ERROR("Trying to move memory with ring turned off.\n");
2148 return -EINVAL;
2149 }
2150
2151 max_bytes = adev->mman.buffer_funcs->copy_max_bytes;
2152 num_loops = DIV_ROUND_UP(byte_count, max_bytes);
2153 num_dw = ALIGN(num_loops * adev->mman.buffer_funcs->copy_num_dw, 8);
2154 r = amdgpu_ttm_prepare_job(adev, direct_submit, num_dw,
2155 resv, vm_needs_flush, &job, false);
2156 if (r)
2157 return r;
2158
2159 for (i = 0; i < num_loops; i++) {
2160 uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
2161
2162 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_offset,
2163 dst_offset, cur_size_in_bytes, tmz);
2164
2165 src_offset += cur_size_in_bytes;
2166 dst_offset += cur_size_in_bytes;
2167 byte_count -= cur_size_in_bytes;
2168 }
2169
2170 amdgpu_ring_pad_ib(ring, &job->ibs[0]);
2171 WARN_ON(job->ibs[0].length_dw > num_dw);
2172 if (direct_submit)
2173 r = amdgpu_job_submit_direct(job, ring, fence);
2174 else
2175 *fence = amdgpu_job_submit(job);
2176 if (r)
2177 goto error_free;
2178
2179 return r;
2180
2181error_free:
2182 amdgpu_job_free(job);
2183 DRM_ERROR("Error scheduling IBs (%d)\n", r);
2184 return r;
2185}
2186
2187static int amdgpu_ttm_fill_mem(struct amdgpu_ring *ring, uint32_t src_data,
2188 uint64_t dst_addr, uint32_t byte_count,
2189 struct dma_resv *resv,
2190 struct dma_fence **fence,
2191 bool vm_needs_flush, bool delayed)
2192{
2193 struct amdgpu_device *adev = ring->adev;
2194 unsigned int num_loops, num_dw;
2195 struct amdgpu_job *job;
2196 uint32_t max_bytes;
2197 unsigned int i;
2198 int r;
2199
2200 max_bytes = adev->mman.buffer_funcs->fill_max_bytes;
2201 num_loops = DIV_ROUND_UP_ULL(byte_count, max_bytes);
2202 num_dw = ALIGN(num_loops * adev->mman.buffer_funcs->fill_num_dw, 8);
2203 r = amdgpu_ttm_prepare_job(adev, false, num_dw, resv, vm_needs_flush,
2204 &job, delayed);
2205 if (r)
2206 return r;
2207
2208 for (i = 0; i < num_loops; i++) {
2209 uint32_t cur_size = min(byte_count, max_bytes);
2210
2211 amdgpu_emit_fill_buffer(adev, &job->ibs[0], src_data, dst_addr,
2212 cur_size);
2213
2214 dst_addr += cur_size;
2215 byte_count -= cur_size;
2216 }
2217
2218 amdgpu_ring_pad_ib(ring, &job->ibs[0]);
2219 WARN_ON(job->ibs[0].length_dw > num_dw);
2220 *fence = amdgpu_job_submit(job);
2221 return 0;
2222}
2223
2224int amdgpu_fill_buffer(struct amdgpu_bo *bo,
2225 uint32_t src_data,
2226 struct dma_resv *resv,
2227 struct dma_fence **f,
2228 bool delayed)
2229{
2230 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
2231 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
2232 struct dma_fence *fence = NULL;
2233 struct amdgpu_res_cursor dst;
2234 int r;
2235
2236 if (!adev->mman.buffer_funcs_enabled) {
2237 DRM_ERROR("Trying to clear memory with ring turned off.\n");
2238 return -EINVAL;
2239 }
2240
2241 amdgpu_res_first(bo->tbo.resource, 0, amdgpu_bo_size(bo), &dst);
2242
2243 mutex_lock(&adev->mman.gtt_window_lock);
2244 while (dst.remaining) {
2245 struct dma_fence *next;
2246 uint64_t cur_size, to;
2247
2248 /* Never fill more than 256MiB at once to avoid timeouts */
2249 cur_size = min(dst.size, 256ULL << 20);
2250
2251 r = amdgpu_ttm_map_buffer(&bo->tbo, bo->tbo.resource, &dst,
2252 1, ring, false, &cur_size, &to);
2253 if (r)
2254 goto error;
2255
2256 r = amdgpu_ttm_fill_mem(ring, src_data, to, cur_size, resv,
2257 &next, true, delayed);
2258 if (r)
2259 goto error;
2260
2261 dma_fence_put(fence);
2262 fence = next;
2263
2264 amdgpu_res_next(&dst, cur_size);
2265 }
2266error:
2267 mutex_unlock(&adev->mman.gtt_window_lock);
2268 if (f)
2269 *f = dma_fence_get(fence);
2270 dma_fence_put(fence);
2271 return r;
2272}
2273
2274/**
2275 * amdgpu_ttm_evict_resources - evict memory buffers
2276 * @adev: amdgpu device object
2277 * @mem_type: evicted BO's memory type
2278 *
2279 * Evicts all @mem_type buffers on the lru list of the memory type.
2280 *
2281 * Returns:
2282 * 0 for success or a negative error code on failure.
2283 */
2284int amdgpu_ttm_evict_resources(struct amdgpu_device *adev, int mem_type)
2285{
2286 struct ttm_resource_manager *man;
2287
2288 switch (mem_type) {
2289 case TTM_PL_VRAM:
2290 case TTM_PL_TT:
2291 case AMDGPU_PL_GWS:
2292 case AMDGPU_PL_GDS:
2293 case AMDGPU_PL_OA:
2294 man = ttm_manager_type(&adev->mman.bdev, mem_type);
2295 break;
2296 default:
2297 DRM_ERROR("Trying to evict invalid memory type\n");
2298 return -EINVAL;
2299 }
2300
2301 return ttm_resource_manager_evict_all(&adev->mman.bdev, man);
2302}
2303
2304#if defined(CONFIG_DEBUG_FS)
2305
2306static int amdgpu_ttm_page_pool_show(struct seq_file *m, void *unused)
2307{
2308 struct amdgpu_device *adev = m->private;
2309
2310 return ttm_pool_debugfs(&adev->mman.bdev.pool, m);
2311}
2312
2313DEFINE_SHOW_ATTRIBUTE(amdgpu_ttm_page_pool);
2314
2315/*
2316 * amdgpu_ttm_vram_read - Linear read access to VRAM
2317 *
2318 * Accesses VRAM via MMIO for debugging purposes.
2319 */
2320static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf,
2321 size_t size, loff_t *pos)
2322{
2323 struct amdgpu_device *adev = file_inode(f)->i_private;
2324 ssize_t result = 0;
2325
2326 if (size & 0x3 || *pos & 0x3)
2327 return -EINVAL;
2328
2329 if (*pos >= adev->gmc.mc_vram_size)
2330 return -ENXIO;
2331
2332 size = min(size, (size_t)(adev->gmc.mc_vram_size - *pos));
2333 while (size) {
2334 size_t bytes = min(size, AMDGPU_TTM_VRAM_MAX_DW_READ * 4);
2335 uint32_t value[AMDGPU_TTM_VRAM_MAX_DW_READ];
2336
2337 amdgpu_device_vram_access(adev, *pos, value, bytes, false);
2338 if (copy_to_user(buf, value, bytes))
2339 return -EFAULT;
2340
2341 result += bytes;
2342 buf += bytes;
2343 *pos += bytes;
2344 size -= bytes;
2345 }
2346
2347 return result;
2348}
2349
2350/*
2351 * amdgpu_ttm_vram_write - Linear write access to VRAM
2352 *
2353 * Accesses VRAM via MMIO for debugging purposes.
2354 */
2355static ssize_t amdgpu_ttm_vram_write(struct file *f, const char __user *buf,
2356 size_t size, loff_t *pos)
2357{
2358 struct amdgpu_device *adev = file_inode(f)->i_private;
2359 ssize_t result = 0;
2360 int r;
2361
2362 if (size & 0x3 || *pos & 0x3)
2363 return -EINVAL;
2364
2365 if (*pos >= adev->gmc.mc_vram_size)
2366 return -ENXIO;
2367
2368 while (size) {
2369 uint32_t value;
2370
2371 if (*pos >= adev->gmc.mc_vram_size)
2372 return result;
2373
2374 r = get_user(value, (uint32_t *)buf);
2375 if (r)
2376 return r;
2377
2378 amdgpu_device_mm_access(adev, *pos, &value, 4, true);
2379
2380 result += 4;
2381 buf += 4;
2382 *pos += 4;
2383 size -= 4;
2384 }
2385
2386 return result;
2387}
2388
2389static const struct file_operations amdgpu_ttm_vram_fops = {
2390 .owner = THIS_MODULE,
2391 .read = amdgpu_ttm_vram_read,
2392 .write = amdgpu_ttm_vram_write,
2393 .llseek = default_llseek,
2394};
2395
2396/*
2397 * amdgpu_iomem_read - Virtual read access to GPU mapped memory
2398 *
2399 * This function is used to read memory that has been mapped to the
2400 * GPU and the known addresses are not physical addresses but instead
2401 * bus addresses (e.g., what you'd put in an IB or ring buffer).
2402 */
2403static ssize_t amdgpu_iomem_read(struct file *f, char __user *buf,
2404 size_t size, loff_t *pos)
2405{
2406 struct amdgpu_device *adev = file_inode(f)->i_private;
2407 struct iommu_domain *dom;
2408 ssize_t result = 0;
2409 int r;
2410
2411 /* retrieve the IOMMU domain if any for this device */
2412 dom = iommu_get_domain_for_dev(adev->dev);
2413
2414 while (size) {
2415 phys_addr_t addr = *pos & PAGE_MASK;
2416 loff_t off = *pos & ~PAGE_MASK;
2417 size_t bytes = PAGE_SIZE - off;
2418 unsigned long pfn;
2419 struct page *p;
2420 void *ptr;
2421
2422 bytes = min(bytes, size);
2423
2424 /* Translate the bus address to a physical address. If
2425 * the domain is NULL it means there is no IOMMU active
2426 * and the address translation is the identity
2427 */
2428 addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2429
2430 pfn = addr >> PAGE_SHIFT;
2431 if (!pfn_valid(pfn))
2432 return -EPERM;
2433
2434 p = pfn_to_page(pfn);
2435 if (p->mapping != adev->mman.bdev.dev_mapping)
2436 return -EPERM;
2437
2438 ptr = kmap_local_page(p);
2439 r = copy_to_user(buf, ptr + off, bytes);
2440 kunmap_local(ptr);
2441 if (r)
2442 return -EFAULT;
2443
2444 size -= bytes;
2445 *pos += bytes;
2446 result += bytes;
2447 }
2448
2449 return result;
2450}
2451
2452/*
2453 * amdgpu_iomem_write - Virtual write access to GPU mapped memory
2454 *
2455 * This function is used to write memory that has been mapped to the
2456 * GPU and the known addresses are not physical addresses but instead
2457 * bus addresses (e.g., what you'd put in an IB or ring buffer).
2458 */
2459static ssize_t amdgpu_iomem_write(struct file *f, const char __user *buf,
2460 size_t size, loff_t *pos)
2461{
2462 struct amdgpu_device *adev = file_inode(f)->i_private;
2463 struct iommu_domain *dom;
2464 ssize_t result = 0;
2465 int r;
2466
2467 dom = iommu_get_domain_for_dev(adev->dev);
2468
2469 while (size) {
2470 phys_addr_t addr = *pos & PAGE_MASK;
2471 loff_t off = *pos & ~PAGE_MASK;
2472 size_t bytes = PAGE_SIZE - off;
2473 unsigned long pfn;
2474 struct page *p;
2475 void *ptr;
2476
2477 bytes = min(bytes, size);
2478
2479 addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2480
2481 pfn = addr >> PAGE_SHIFT;
2482 if (!pfn_valid(pfn))
2483 return -EPERM;
2484
2485 p = pfn_to_page(pfn);
2486 if (p->mapping != adev->mman.bdev.dev_mapping)
2487 return -EPERM;
2488
2489 ptr = kmap_local_page(p);
2490 r = copy_from_user(ptr + off, buf, bytes);
2491 kunmap_local(ptr);
2492 if (r)
2493 return -EFAULT;
2494
2495 size -= bytes;
2496 *pos += bytes;
2497 result += bytes;
2498 }
2499
2500 return result;
2501}
2502
2503static const struct file_operations amdgpu_ttm_iomem_fops = {
2504 .owner = THIS_MODULE,
2505 .read = amdgpu_iomem_read,
2506 .write = amdgpu_iomem_write,
2507 .llseek = default_llseek
2508};
2509
2510#endif
2511
2512void amdgpu_ttm_debugfs_init(struct amdgpu_device *adev)
2513{
2514#if defined(CONFIG_DEBUG_FS)
2515 struct drm_minor *minor = adev_to_drm(adev)->primary;
2516 struct dentry *root = minor->debugfs_root;
2517
2518 debugfs_create_file_size("amdgpu_vram", 0444, root, adev,
2519 &amdgpu_ttm_vram_fops, adev->gmc.mc_vram_size);
2520 debugfs_create_file("amdgpu_iomem", 0444, root, adev,
2521 &amdgpu_ttm_iomem_fops);
2522 debugfs_create_file("ttm_page_pool", 0444, root, adev,
2523 &amdgpu_ttm_page_pool_fops);
2524 ttm_resource_manager_create_debugfs(ttm_manager_type(&adev->mman.bdev,
2525 TTM_PL_VRAM),
2526 root, "amdgpu_vram_mm");
2527 ttm_resource_manager_create_debugfs(ttm_manager_type(&adev->mman.bdev,
2528 TTM_PL_TT),
2529 root, "amdgpu_gtt_mm");
2530 ttm_resource_manager_create_debugfs(ttm_manager_type(&adev->mman.bdev,
2531 AMDGPU_PL_GDS),
2532 root, "amdgpu_gds_mm");
2533 ttm_resource_manager_create_debugfs(ttm_manager_type(&adev->mman.bdev,
2534 AMDGPU_PL_GWS),
2535 root, "amdgpu_gws_mm");
2536 ttm_resource_manager_create_debugfs(ttm_manager_type(&adev->mman.bdev,
2537 AMDGPU_PL_OA),
2538 root, "amdgpu_oa_mm");
2539
2540#endif
2541}