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 <ttm/ttm_bo_api.h>
33#include <ttm/ttm_bo_driver.h>
34#include <ttm/ttm_placement.h>
35#include <ttm/ttm_module.h>
36#include <ttm/ttm_page_alloc.h>
37#include <drm/drmP.h>
38#include <drm/radeon_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 "radeon_reg.h"
46#include "radeon.h"
47
48#define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT)
49
50static int radeon_ttm_debugfs_init(struct radeon_device *rdev);
51static void radeon_ttm_debugfs_fini(struct radeon_device *rdev);
52
53static struct radeon_device *radeon_get_rdev(struct ttm_bo_device *bdev)
54{
55 struct radeon_mman *mman;
56 struct radeon_device *rdev;
57
58 mman = container_of(bdev, struct radeon_mman, bdev);
59 rdev = container_of(mman, struct radeon_device, mman);
60 return rdev;
61}
62
63
64/*
65 * Global memory.
66 */
67static int radeon_ttm_mem_global_init(struct drm_global_reference *ref)
68{
69 return ttm_mem_global_init(ref->object);
70}
71
72static void radeon_ttm_mem_global_release(struct drm_global_reference *ref)
73{
74 ttm_mem_global_release(ref->object);
75}
76
77static int radeon_ttm_global_init(struct radeon_device *rdev)
78{
79 struct drm_global_reference *global_ref;
80 int r;
81
82 rdev->mman.mem_global_referenced = false;
83 global_ref = &rdev->mman.mem_global_ref;
84 global_ref->global_type = DRM_GLOBAL_TTM_MEM;
85 global_ref->size = sizeof(struct ttm_mem_global);
86 global_ref->init = &radeon_ttm_mem_global_init;
87 global_ref->release = &radeon_ttm_mem_global_release;
88 r = drm_global_item_ref(global_ref);
89 if (r != 0) {
90 DRM_ERROR("Failed setting up TTM memory accounting "
91 "subsystem.\n");
92 return r;
93 }
94
95 rdev->mman.bo_global_ref.mem_glob =
96 rdev->mman.mem_global_ref.object;
97 global_ref = &rdev->mman.bo_global_ref.ref;
98 global_ref->global_type = DRM_GLOBAL_TTM_BO;
99 global_ref->size = sizeof(struct ttm_bo_global);
100 global_ref->init = &ttm_bo_global_init;
101 global_ref->release = &ttm_bo_global_release;
102 r = drm_global_item_ref(global_ref);
103 if (r != 0) {
104 DRM_ERROR("Failed setting up TTM BO subsystem.\n");
105 drm_global_item_unref(&rdev->mman.mem_global_ref);
106 return r;
107 }
108
109 rdev->mman.mem_global_referenced = true;
110 return 0;
111}
112
113static void radeon_ttm_global_fini(struct radeon_device *rdev)
114{
115 if (rdev->mman.mem_global_referenced) {
116 drm_global_item_unref(&rdev->mman.bo_global_ref.ref);
117 drm_global_item_unref(&rdev->mman.mem_global_ref);
118 rdev->mman.mem_global_referenced = false;
119 }
120}
121
122static int radeon_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
123{
124 return 0;
125}
126
127static int radeon_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
128 struct ttm_mem_type_manager *man)
129{
130 struct radeon_device *rdev;
131
132 rdev = radeon_get_rdev(bdev);
133
134 switch (type) {
135 case TTM_PL_SYSTEM:
136 /* System memory */
137 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
138 man->available_caching = TTM_PL_MASK_CACHING;
139 man->default_caching = TTM_PL_FLAG_CACHED;
140 break;
141 case TTM_PL_TT:
142 man->func = &ttm_bo_manager_func;
143 man->gpu_offset = rdev->mc.gtt_start;
144 man->available_caching = TTM_PL_MASK_CACHING;
145 man->default_caching = TTM_PL_FLAG_CACHED;
146 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE | TTM_MEMTYPE_FLAG_CMA;
147#if IS_ENABLED(CONFIG_AGP)
148 if (rdev->flags & RADEON_IS_AGP) {
149 if (!rdev->ddev->agp) {
150 DRM_ERROR("AGP is not enabled for memory type %u\n",
151 (unsigned)type);
152 return -EINVAL;
153 }
154 if (!rdev->ddev->agp->cant_use_aperture)
155 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
156 man->available_caching = TTM_PL_FLAG_UNCACHED |
157 TTM_PL_FLAG_WC;
158 man->default_caching = TTM_PL_FLAG_WC;
159 }
160#endif
161 break;
162 case TTM_PL_VRAM:
163 /* "On-card" video ram */
164 man->func = &ttm_bo_manager_func;
165 man->gpu_offset = rdev->mc.vram_start;
166 man->flags = TTM_MEMTYPE_FLAG_FIXED |
167 TTM_MEMTYPE_FLAG_MAPPABLE;
168 man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
169 man->default_caching = TTM_PL_FLAG_WC;
170 break;
171 default:
172 DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
173 return -EINVAL;
174 }
175 return 0;
176}
177
178static void radeon_evict_flags(struct ttm_buffer_object *bo,
179 struct ttm_placement *placement)
180{
181 static struct ttm_place placements = {
182 .fpfn = 0,
183 .lpfn = 0,
184 .flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
185 };
186
187 struct radeon_bo *rbo;
188
189 if (!radeon_ttm_bo_is_radeon_bo(bo)) {
190 placement->placement = &placements;
191 placement->busy_placement = &placements;
192 placement->num_placement = 1;
193 placement->num_busy_placement = 1;
194 return;
195 }
196 rbo = container_of(bo, struct radeon_bo, tbo);
197 switch (bo->mem.mem_type) {
198 case TTM_PL_VRAM:
199 if (rbo->rdev->ring[radeon_copy_ring_index(rbo->rdev)].ready == false)
200 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
201 else if (rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size &&
202 bo->mem.start < (rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT)) {
203 unsigned fpfn = rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
204 int i;
205
206 /* Try evicting to the CPU inaccessible part of VRAM
207 * first, but only set GTT as busy placement, so this
208 * BO will be evicted to GTT rather than causing other
209 * BOs to be evicted from VRAM
210 */
211 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM |
212 RADEON_GEM_DOMAIN_GTT);
213 rbo->placement.num_busy_placement = 0;
214 for (i = 0; i < rbo->placement.num_placement; i++) {
215 if (rbo->placements[i].flags & TTM_PL_FLAG_VRAM) {
216 if (rbo->placements[0].fpfn < fpfn)
217 rbo->placements[0].fpfn = fpfn;
218 } else {
219 rbo->placement.busy_placement =
220 &rbo->placements[i];
221 rbo->placement.num_busy_placement = 1;
222 }
223 }
224 } else
225 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
226 break;
227 case TTM_PL_TT:
228 default:
229 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
230 }
231 *placement = rbo->placement;
232}
233
234static int radeon_verify_access(struct ttm_buffer_object *bo, struct file *filp)
235{
236 struct radeon_bo *rbo = container_of(bo, struct radeon_bo, tbo);
237
238 if (radeon_ttm_tt_has_userptr(bo->ttm))
239 return -EPERM;
240 return drm_vma_node_verify_access(&rbo->gem_base.vma_node, filp);
241}
242
243static void radeon_move_null(struct ttm_buffer_object *bo,
244 struct ttm_mem_reg *new_mem)
245{
246 struct ttm_mem_reg *old_mem = &bo->mem;
247
248 BUG_ON(old_mem->mm_node != NULL);
249 *old_mem = *new_mem;
250 new_mem->mm_node = NULL;
251}
252
253static int radeon_move_blit(struct ttm_buffer_object *bo,
254 bool evict, bool no_wait_gpu,
255 struct ttm_mem_reg *new_mem,
256 struct ttm_mem_reg *old_mem)
257{
258 struct radeon_device *rdev;
259 uint64_t old_start, new_start;
260 struct radeon_fence *fence;
261 unsigned num_pages;
262 int r, ridx;
263
264 rdev = radeon_get_rdev(bo->bdev);
265 ridx = radeon_copy_ring_index(rdev);
266 old_start = old_mem->start << PAGE_SHIFT;
267 new_start = new_mem->start << PAGE_SHIFT;
268
269 switch (old_mem->mem_type) {
270 case TTM_PL_VRAM:
271 old_start += rdev->mc.vram_start;
272 break;
273 case TTM_PL_TT:
274 old_start += rdev->mc.gtt_start;
275 break;
276 default:
277 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
278 return -EINVAL;
279 }
280 switch (new_mem->mem_type) {
281 case TTM_PL_VRAM:
282 new_start += rdev->mc.vram_start;
283 break;
284 case TTM_PL_TT:
285 new_start += rdev->mc.gtt_start;
286 break;
287 default:
288 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
289 return -EINVAL;
290 }
291 if (!rdev->ring[ridx].ready) {
292 DRM_ERROR("Trying to move memory with ring turned off.\n");
293 return -EINVAL;
294 }
295
296 BUILD_BUG_ON((PAGE_SIZE % RADEON_GPU_PAGE_SIZE) != 0);
297
298 num_pages = new_mem->num_pages * (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
299 fence = radeon_copy(rdev, old_start, new_start, num_pages, bo->resv);
300 if (IS_ERR(fence))
301 return PTR_ERR(fence);
302
303 r = ttm_bo_move_accel_cleanup(bo, &fence->base,
304 evict, no_wait_gpu, new_mem);
305 radeon_fence_unref(&fence);
306 return r;
307}
308
309static int radeon_move_vram_ram(struct ttm_buffer_object *bo,
310 bool evict, bool interruptible,
311 bool no_wait_gpu,
312 struct ttm_mem_reg *new_mem)
313{
314 struct radeon_device *rdev;
315 struct ttm_mem_reg *old_mem = &bo->mem;
316 struct ttm_mem_reg tmp_mem;
317 struct ttm_place placements;
318 struct ttm_placement placement;
319 int r;
320
321 rdev = radeon_get_rdev(bo->bdev);
322 tmp_mem = *new_mem;
323 tmp_mem.mm_node = NULL;
324 placement.num_placement = 1;
325 placement.placement = &placements;
326 placement.num_busy_placement = 1;
327 placement.busy_placement = &placements;
328 placements.fpfn = 0;
329 placements.lpfn = 0;
330 placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
331 r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
332 interruptible, no_wait_gpu);
333 if (unlikely(r)) {
334 return r;
335 }
336
337 r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
338 if (unlikely(r)) {
339 goto out_cleanup;
340 }
341
342 r = ttm_tt_bind(bo->ttm, &tmp_mem);
343 if (unlikely(r)) {
344 goto out_cleanup;
345 }
346 r = radeon_move_blit(bo, true, no_wait_gpu, &tmp_mem, old_mem);
347 if (unlikely(r)) {
348 goto out_cleanup;
349 }
350 r = ttm_bo_move_ttm(bo, true, no_wait_gpu, new_mem);
351out_cleanup:
352 ttm_bo_mem_put(bo, &tmp_mem);
353 return r;
354}
355
356static int radeon_move_ram_vram(struct ttm_buffer_object *bo,
357 bool evict, bool interruptible,
358 bool no_wait_gpu,
359 struct ttm_mem_reg *new_mem)
360{
361 struct radeon_device *rdev;
362 struct ttm_mem_reg *old_mem = &bo->mem;
363 struct ttm_mem_reg tmp_mem;
364 struct ttm_placement placement;
365 struct ttm_place placements;
366 int r;
367
368 rdev = radeon_get_rdev(bo->bdev);
369 tmp_mem = *new_mem;
370 tmp_mem.mm_node = NULL;
371 placement.num_placement = 1;
372 placement.placement = &placements;
373 placement.num_busy_placement = 1;
374 placement.busy_placement = &placements;
375 placements.fpfn = 0;
376 placements.lpfn = 0;
377 placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
378 r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
379 interruptible, no_wait_gpu);
380 if (unlikely(r)) {
381 return r;
382 }
383 r = ttm_bo_move_ttm(bo, true, no_wait_gpu, &tmp_mem);
384 if (unlikely(r)) {
385 goto out_cleanup;
386 }
387 r = radeon_move_blit(bo, true, no_wait_gpu, new_mem, old_mem);
388 if (unlikely(r)) {
389 goto out_cleanup;
390 }
391out_cleanup:
392 ttm_bo_mem_put(bo, &tmp_mem);
393 return r;
394}
395
396static int radeon_bo_move(struct ttm_buffer_object *bo,
397 bool evict, bool interruptible,
398 bool no_wait_gpu,
399 struct ttm_mem_reg *new_mem)
400{
401 struct radeon_device *rdev;
402 struct radeon_bo *rbo;
403 struct ttm_mem_reg *old_mem = &bo->mem;
404 int r;
405
406 /* Can't move a pinned BO */
407 rbo = container_of(bo, struct radeon_bo, tbo);
408 if (WARN_ON_ONCE(rbo->pin_count > 0))
409 return -EINVAL;
410
411 rdev = radeon_get_rdev(bo->bdev);
412 if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
413 radeon_move_null(bo, new_mem);
414 return 0;
415 }
416 if ((old_mem->mem_type == TTM_PL_TT &&
417 new_mem->mem_type == TTM_PL_SYSTEM) ||
418 (old_mem->mem_type == TTM_PL_SYSTEM &&
419 new_mem->mem_type == TTM_PL_TT)) {
420 /* bind is enough */
421 radeon_move_null(bo, new_mem);
422 return 0;
423 }
424 if (!rdev->ring[radeon_copy_ring_index(rdev)].ready ||
425 rdev->asic->copy.copy == NULL) {
426 /* use memcpy */
427 goto memcpy;
428 }
429
430 if (old_mem->mem_type == TTM_PL_VRAM &&
431 new_mem->mem_type == TTM_PL_SYSTEM) {
432 r = radeon_move_vram_ram(bo, evict, interruptible,
433 no_wait_gpu, new_mem);
434 } else if (old_mem->mem_type == TTM_PL_SYSTEM &&
435 new_mem->mem_type == TTM_PL_VRAM) {
436 r = radeon_move_ram_vram(bo, evict, interruptible,
437 no_wait_gpu, new_mem);
438 } else {
439 r = radeon_move_blit(bo, evict, no_wait_gpu, new_mem, old_mem);
440 }
441
442 if (r) {
443memcpy:
444 r = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
445 if (r) {
446 return r;
447 }
448 }
449
450 /* update statistics */
451 atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &rdev->num_bytes_moved);
452 return 0;
453}
454
455static int radeon_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
456{
457 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
458 struct radeon_device *rdev = radeon_get_rdev(bdev);
459
460 mem->bus.addr = NULL;
461 mem->bus.offset = 0;
462 mem->bus.size = mem->num_pages << PAGE_SHIFT;
463 mem->bus.base = 0;
464 mem->bus.is_iomem = false;
465 if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
466 return -EINVAL;
467 switch (mem->mem_type) {
468 case TTM_PL_SYSTEM:
469 /* system memory */
470 return 0;
471 case TTM_PL_TT:
472#if IS_ENABLED(CONFIG_AGP)
473 if (rdev->flags & RADEON_IS_AGP) {
474 /* RADEON_IS_AGP is set only if AGP is active */
475 mem->bus.offset = mem->start << PAGE_SHIFT;
476 mem->bus.base = rdev->mc.agp_base;
477 mem->bus.is_iomem = !rdev->ddev->agp->cant_use_aperture;
478 }
479#endif
480 break;
481 case TTM_PL_VRAM:
482 mem->bus.offset = mem->start << PAGE_SHIFT;
483 /* check if it's visible */
484 if ((mem->bus.offset + mem->bus.size) > rdev->mc.visible_vram_size)
485 return -EINVAL;
486 mem->bus.base = rdev->mc.aper_base;
487 mem->bus.is_iomem = true;
488#ifdef __alpha__
489 /*
490 * Alpha: use bus.addr to hold the ioremap() return,
491 * so we can modify bus.base below.
492 */
493 if (mem->placement & TTM_PL_FLAG_WC)
494 mem->bus.addr =
495 ioremap_wc(mem->bus.base + mem->bus.offset,
496 mem->bus.size);
497 else
498 mem->bus.addr =
499 ioremap_nocache(mem->bus.base + mem->bus.offset,
500 mem->bus.size);
501
502 /*
503 * Alpha: Use just the bus offset plus
504 * the hose/domain memory base for bus.base.
505 * It then can be used to build PTEs for VRAM
506 * access, as done in ttm_bo_vm_fault().
507 */
508 mem->bus.base = (mem->bus.base & 0x0ffffffffUL) +
509 rdev->ddev->hose->dense_mem_base;
510#endif
511 break;
512 default:
513 return -EINVAL;
514 }
515 return 0;
516}
517
518static void radeon_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
519{
520}
521
522/*
523 * TTM backend functions.
524 */
525struct radeon_ttm_tt {
526 struct ttm_dma_tt ttm;
527 struct radeon_device *rdev;
528 u64 offset;
529
530 uint64_t userptr;
531 struct mm_struct *usermm;
532 uint32_t userflags;
533};
534
535/* prepare the sg table with the user pages */
536static int radeon_ttm_tt_pin_userptr(struct ttm_tt *ttm)
537{
538 struct radeon_device *rdev = radeon_get_rdev(ttm->bdev);
539 struct radeon_ttm_tt *gtt = (void *)ttm;
540 unsigned pinned = 0, nents;
541 int r;
542
543 int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
544 enum dma_data_direction direction = write ?
545 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
546
547 if (current->mm != gtt->usermm)
548 return -EPERM;
549
550 if (gtt->userflags & RADEON_GEM_USERPTR_ANONONLY) {
551 /* check that we only pin down anonymous memory
552 to prevent problems with writeback */
553 unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE;
554 struct vm_area_struct *vma;
555 vma = find_vma(gtt->usermm, gtt->userptr);
556 if (!vma || vma->vm_file || vma->vm_end < end)
557 return -EPERM;
558 }
559
560 do {
561 unsigned num_pages = ttm->num_pages - pinned;
562 uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
563 struct page **pages = ttm->pages + pinned;
564
565 r = get_user_pages(userptr, num_pages, write, 0, pages, NULL);
566 if (r < 0)
567 goto release_pages;
568
569 pinned += r;
570
571 } while (pinned < ttm->num_pages);
572
573 r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
574 ttm->num_pages << PAGE_SHIFT,
575 GFP_KERNEL);
576 if (r)
577 goto release_sg;
578
579 r = -ENOMEM;
580 nents = dma_map_sg(rdev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
581 if (nents != ttm->sg->nents)
582 goto release_sg;
583
584 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
585 gtt->ttm.dma_address, ttm->num_pages);
586
587 return 0;
588
589release_sg:
590 kfree(ttm->sg);
591
592release_pages:
593 release_pages(ttm->pages, pinned, 0);
594 return r;
595}
596
597static void radeon_ttm_tt_unpin_userptr(struct ttm_tt *ttm)
598{
599 struct radeon_device *rdev = radeon_get_rdev(ttm->bdev);
600 struct radeon_ttm_tt *gtt = (void *)ttm;
601 struct sg_page_iter sg_iter;
602
603 int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
604 enum dma_data_direction direction = write ?
605 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
606
607 /* double check that we don't free the table twice */
608 if (!ttm->sg->sgl)
609 return;
610
611 /* free the sg table and pages again */
612 dma_unmap_sg(rdev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
613
614 for_each_sg_page(ttm->sg->sgl, &sg_iter, ttm->sg->nents, 0) {
615 struct page *page = sg_page_iter_page(&sg_iter);
616 if (!(gtt->userflags & RADEON_GEM_USERPTR_READONLY))
617 set_page_dirty(page);
618
619 mark_page_accessed(page);
620 put_page(page);
621 }
622
623 sg_free_table(ttm->sg);
624}
625
626static int radeon_ttm_backend_bind(struct ttm_tt *ttm,
627 struct ttm_mem_reg *bo_mem)
628{
629 struct radeon_ttm_tt *gtt = (void*)ttm;
630 uint32_t flags = RADEON_GART_PAGE_VALID | RADEON_GART_PAGE_READ |
631 RADEON_GART_PAGE_WRITE;
632 int r;
633
634 if (gtt->userptr) {
635 radeon_ttm_tt_pin_userptr(ttm);
636 flags &= ~RADEON_GART_PAGE_WRITE;
637 }
638
639 gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT);
640 if (!ttm->num_pages) {
641 WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
642 ttm->num_pages, bo_mem, ttm);
643 }
644 if (ttm->caching_state == tt_cached)
645 flags |= RADEON_GART_PAGE_SNOOP;
646 r = radeon_gart_bind(gtt->rdev, gtt->offset, ttm->num_pages,
647 ttm->pages, gtt->ttm.dma_address, flags);
648 if (r) {
649 DRM_ERROR("failed to bind %lu pages at 0x%08X\n",
650 ttm->num_pages, (unsigned)gtt->offset);
651 return r;
652 }
653 return 0;
654}
655
656static int radeon_ttm_backend_unbind(struct ttm_tt *ttm)
657{
658 struct radeon_ttm_tt *gtt = (void *)ttm;
659
660 radeon_gart_unbind(gtt->rdev, gtt->offset, ttm->num_pages);
661
662 if (gtt->userptr)
663 radeon_ttm_tt_unpin_userptr(ttm);
664
665 return 0;
666}
667
668static void radeon_ttm_backend_destroy(struct ttm_tt *ttm)
669{
670 struct radeon_ttm_tt *gtt = (void *)ttm;
671
672 ttm_dma_tt_fini(>t->ttm);
673 kfree(gtt);
674}
675
676static struct ttm_backend_func radeon_backend_func = {
677 .bind = &radeon_ttm_backend_bind,
678 .unbind = &radeon_ttm_backend_unbind,
679 .destroy = &radeon_ttm_backend_destroy,
680};
681
682static struct ttm_tt *radeon_ttm_tt_create(struct ttm_bo_device *bdev,
683 unsigned long size, uint32_t page_flags,
684 struct page *dummy_read_page)
685{
686 struct radeon_device *rdev;
687 struct radeon_ttm_tt *gtt;
688
689 rdev = radeon_get_rdev(bdev);
690#if IS_ENABLED(CONFIG_AGP)
691 if (rdev->flags & RADEON_IS_AGP) {
692 return ttm_agp_tt_create(bdev, rdev->ddev->agp->bridge,
693 size, page_flags, dummy_read_page);
694 }
695#endif
696
697 gtt = kzalloc(sizeof(struct radeon_ttm_tt), GFP_KERNEL);
698 if (gtt == NULL) {
699 return NULL;
700 }
701 gtt->ttm.ttm.func = &radeon_backend_func;
702 gtt->rdev = rdev;
703 if (ttm_dma_tt_init(>t->ttm, bdev, size, page_flags, dummy_read_page)) {
704 kfree(gtt);
705 return NULL;
706 }
707 return >t->ttm.ttm;
708}
709
710static struct radeon_ttm_tt *radeon_ttm_tt_to_gtt(struct ttm_tt *ttm)
711{
712 if (!ttm || ttm->func != &radeon_backend_func)
713 return NULL;
714 return (struct radeon_ttm_tt *)ttm;
715}
716
717static int radeon_ttm_tt_populate(struct ttm_tt *ttm)
718{
719 struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
720 struct radeon_device *rdev;
721 unsigned i;
722 int r;
723 bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
724
725 if (ttm->state != tt_unpopulated)
726 return 0;
727
728 if (gtt && gtt->userptr) {
729 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
730 if (!ttm->sg)
731 return -ENOMEM;
732
733 ttm->page_flags |= TTM_PAGE_FLAG_SG;
734 ttm->state = tt_unbound;
735 return 0;
736 }
737
738 if (slave && ttm->sg) {
739 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
740 gtt->ttm.dma_address, ttm->num_pages);
741 ttm->state = tt_unbound;
742 return 0;
743 }
744
745 rdev = radeon_get_rdev(ttm->bdev);
746#if IS_ENABLED(CONFIG_AGP)
747 if (rdev->flags & RADEON_IS_AGP) {
748 return ttm_agp_tt_populate(ttm);
749 }
750#endif
751
752#ifdef CONFIG_SWIOTLB
753 if (swiotlb_nr_tbl()) {
754 return ttm_dma_populate(>t->ttm, rdev->dev);
755 }
756#endif
757
758 r = ttm_pool_populate(ttm);
759 if (r) {
760 return r;
761 }
762
763 for (i = 0; i < ttm->num_pages; i++) {
764 gtt->ttm.dma_address[i] = pci_map_page(rdev->pdev, ttm->pages[i],
765 0, PAGE_SIZE,
766 PCI_DMA_BIDIRECTIONAL);
767 if (pci_dma_mapping_error(rdev->pdev, gtt->ttm.dma_address[i])) {
768 while (i--) {
769 pci_unmap_page(rdev->pdev, gtt->ttm.dma_address[i],
770 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
771 gtt->ttm.dma_address[i] = 0;
772 }
773 ttm_pool_unpopulate(ttm);
774 return -EFAULT;
775 }
776 }
777 return 0;
778}
779
780static void radeon_ttm_tt_unpopulate(struct ttm_tt *ttm)
781{
782 struct radeon_device *rdev;
783 struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
784 unsigned i;
785 bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
786
787 if (gtt && gtt->userptr) {
788 kfree(ttm->sg);
789 ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
790 return;
791 }
792
793 if (slave)
794 return;
795
796 rdev = radeon_get_rdev(ttm->bdev);
797#if IS_ENABLED(CONFIG_AGP)
798 if (rdev->flags & RADEON_IS_AGP) {
799 ttm_agp_tt_unpopulate(ttm);
800 return;
801 }
802#endif
803
804#ifdef CONFIG_SWIOTLB
805 if (swiotlb_nr_tbl()) {
806 ttm_dma_unpopulate(>t->ttm, rdev->dev);
807 return;
808 }
809#endif
810
811 for (i = 0; i < ttm->num_pages; i++) {
812 if (gtt->ttm.dma_address[i]) {
813 pci_unmap_page(rdev->pdev, gtt->ttm.dma_address[i],
814 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
815 }
816 }
817
818 ttm_pool_unpopulate(ttm);
819}
820
821int radeon_ttm_tt_set_userptr(struct ttm_tt *ttm, uint64_t addr,
822 uint32_t flags)
823{
824 struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
825
826 if (gtt == NULL)
827 return -EINVAL;
828
829 gtt->userptr = addr;
830 gtt->usermm = current->mm;
831 gtt->userflags = flags;
832 return 0;
833}
834
835bool radeon_ttm_tt_has_userptr(struct ttm_tt *ttm)
836{
837 struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
838
839 if (gtt == NULL)
840 return false;
841
842 return !!gtt->userptr;
843}
844
845bool radeon_ttm_tt_is_readonly(struct ttm_tt *ttm)
846{
847 struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
848
849 if (gtt == NULL)
850 return false;
851
852 return !!(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
853}
854
855static struct ttm_bo_driver radeon_bo_driver = {
856 .ttm_tt_create = &radeon_ttm_tt_create,
857 .ttm_tt_populate = &radeon_ttm_tt_populate,
858 .ttm_tt_unpopulate = &radeon_ttm_tt_unpopulate,
859 .invalidate_caches = &radeon_invalidate_caches,
860 .init_mem_type = &radeon_init_mem_type,
861 .evict_flags = &radeon_evict_flags,
862 .move = &radeon_bo_move,
863 .verify_access = &radeon_verify_access,
864 .move_notify = &radeon_bo_move_notify,
865 .fault_reserve_notify = &radeon_bo_fault_reserve_notify,
866 .io_mem_reserve = &radeon_ttm_io_mem_reserve,
867 .io_mem_free = &radeon_ttm_io_mem_free,
868};
869
870int radeon_ttm_init(struct radeon_device *rdev)
871{
872 int r;
873
874 r = radeon_ttm_global_init(rdev);
875 if (r) {
876 return r;
877 }
878 /* No others user of address space so set it to 0 */
879 r = ttm_bo_device_init(&rdev->mman.bdev,
880 rdev->mman.bo_global_ref.ref.object,
881 &radeon_bo_driver,
882 rdev->ddev->anon_inode->i_mapping,
883 DRM_FILE_PAGE_OFFSET,
884 rdev->need_dma32);
885 if (r) {
886 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
887 return r;
888 }
889 rdev->mman.initialized = true;
890 r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_VRAM,
891 rdev->mc.real_vram_size >> PAGE_SHIFT);
892 if (r) {
893 DRM_ERROR("Failed initializing VRAM heap.\n");
894 return r;
895 }
896 /* Change the size here instead of the init above so only lpfn is affected */
897 radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
898
899 r = radeon_bo_create(rdev, 256 * 1024, PAGE_SIZE, true,
900 RADEON_GEM_DOMAIN_VRAM, 0, NULL,
901 NULL, &rdev->stollen_vga_memory);
902 if (r) {
903 return r;
904 }
905 r = radeon_bo_reserve(rdev->stollen_vga_memory, false);
906 if (r)
907 return r;
908 r = radeon_bo_pin(rdev->stollen_vga_memory, RADEON_GEM_DOMAIN_VRAM, NULL);
909 radeon_bo_unreserve(rdev->stollen_vga_memory);
910 if (r) {
911 radeon_bo_unref(&rdev->stollen_vga_memory);
912 return r;
913 }
914 DRM_INFO("radeon: %uM of VRAM memory ready\n",
915 (unsigned) (rdev->mc.real_vram_size / (1024 * 1024)));
916 r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_TT,
917 rdev->mc.gtt_size >> PAGE_SHIFT);
918 if (r) {
919 DRM_ERROR("Failed initializing GTT heap.\n");
920 return r;
921 }
922 DRM_INFO("radeon: %uM of GTT memory ready.\n",
923 (unsigned)(rdev->mc.gtt_size / (1024 * 1024)));
924
925 r = radeon_ttm_debugfs_init(rdev);
926 if (r) {
927 DRM_ERROR("Failed to init debugfs\n");
928 return r;
929 }
930 return 0;
931}
932
933void radeon_ttm_fini(struct radeon_device *rdev)
934{
935 int r;
936
937 if (!rdev->mman.initialized)
938 return;
939 radeon_ttm_debugfs_fini(rdev);
940 if (rdev->stollen_vga_memory) {
941 r = radeon_bo_reserve(rdev->stollen_vga_memory, false);
942 if (r == 0) {
943 radeon_bo_unpin(rdev->stollen_vga_memory);
944 radeon_bo_unreserve(rdev->stollen_vga_memory);
945 }
946 radeon_bo_unref(&rdev->stollen_vga_memory);
947 }
948 ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_VRAM);
949 ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_TT);
950 ttm_bo_device_release(&rdev->mman.bdev);
951 radeon_gart_fini(rdev);
952 radeon_ttm_global_fini(rdev);
953 rdev->mman.initialized = false;
954 DRM_INFO("radeon: ttm finalized\n");
955}
956
957/* this should only be called at bootup or when userspace
958 * isn't running */
959void radeon_ttm_set_active_vram_size(struct radeon_device *rdev, u64 size)
960{
961 struct ttm_mem_type_manager *man;
962
963 if (!rdev->mman.initialized)
964 return;
965
966 man = &rdev->mman.bdev.man[TTM_PL_VRAM];
967 /* this just adjusts TTM size idea, which sets lpfn to the correct value */
968 man->size = size >> PAGE_SHIFT;
969}
970
971static struct vm_operations_struct radeon_ttm_vm_ops;
972static const struct vm_operations_struct *ttm_vm_ops = NULL;
973
974static int radeon_ttm_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
975{
976 struct ttm_buffer_object *bo;
977 struct radeon_device *rdev;
978 int r;
979
980 bo = (struct ttm_buffer_object *)vma->vm_private_data;
981 if (bo == NULL) {
982 return VM_FAULT_NOPAGE;
983 }
984 rdev = radeon_get_rdev(bo->bdev);
985 down_read(&rdev->pm.mclk_lock);
986 r = ttm_vm_ops->fault(vma, vmf);
987 up_read(&rdev->pm.mclk_lock);
988 return r;
989}
990
991int radeon_mmap(struct file *filp, struct vm_area_struct *vma)
992{
993 struct drm_file *file_priv;
994 struct radeon_device *rdev;
995 int r;
996
997 if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET)) {
998 return -EINVAL;
999 }
1000
1001 file_priv = filp->private_data;
1002 rdev = file_priv->minor->dev->dev_private;
1003 if (rdev == NULL) {
1004 return -EINVAL;
1005 }
1006 r = ttm_bo_mmap(filp, vma, &rdev->mman.bdev);
1007 if (unlikely(r != 0)) {
1008 return r;
1009 }
1010 if (unlikely(ttm_vm_ops == NULL)) {
1011 ttm_vm_ops = vma->vm_ops;
1012 radeon_ttm_vm_ops = *ttm_vm_ops;
1013 radeon_ttm_vm_ops.fault = &radeon_ttm_fault;
1014 }
1015 vma->vm_ops = &radeon_ttm_vm_ops;
1016 return 0;
1017}
1018
1019#if defined(CONFIG_DEBUG_FS)
1020
1021static int radeon_mm_dump_table(struct seq_file *m, void *data)
1022{
1023 struct drm_info_node *node = (struct drm_info_node *)m->private;
1024 unsigned ttm_pl = *(int *)node->info_ent->data;
1025 struct drm_device *dev = node->minor->dev;
1026 struct radeon_device *rdev = dev->dev_private;
1027 struct drm_mm *mm = (struct drm_mm *)rdev->mman.bdev.man[ttm_pl].priv;
1028 int ret;
1029 struct ttm_bo_global *glob = rdev->mman.bdev.glob;
1030
1031 spin_lock(&glob->lru_lock);
1032 ret = drm_mm_dump_table(m, mm);
1033 spin_unlock(&glob->lru_lock);
1034 return ret;
1035}
1036
1037static int ttm_pl_vram = TTM_PL_VRAM;
1038static int ttm_pl_tt = TTM_PL_TT;
1039
1040static struct drm_info_list radeon_ttm_debugfs_list[] = {
1041 {"radeon_vram_mm", radeon_mm_dump_table, 0, &ttm_pl_vram},
1042 {"radeon_gtt_mm", radeon_mm_dump_table, 0, &ttm_pl_tt},
1043 {"ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL},
1044#ifdef CONFIG_SWIOTLB
1045 {"ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL}
1046#endif
1047};
1048
1049static int radeon_ttm_vram_open(struct inode *inode, struct file *filep)
1050{
1051 struct radeon_device *rdev = inode->i_private;
1052 i_size_write(inode, rdev->mc.mc_vram_size);
1053 filep->private_data = inode->i_private;
1054 return 0;
1055}
1056
1057static ssize_t radeon_ttm_vram_read(struct file *f, char __user *buf,
1058 size_t size, loff_t *pos)
1059{
1060 struct radeon_device *rdev = f->private_data;
1061 ssize_t result = 0;
1062 int r;
1063
1064 if (size & 0x3 || *pos & 0x3)
1065 return -EINVAL;
1066
1067 while (size) {
1068 unsigned long flags;
1069 uint32_t value;
1070
1071 if (*pos >= rdev->mc.mc_vram_size)
1072 return result;
1073
1074 spin_lock_irqsave(&rdev->mmio_idx_lock, flags);
1075 WREG32(RADEON_MM_INDEX, ((uint32_t)*pos) | 0x80000000);
1076 if (rdev->family >= CHIP_CEDAR)
1077 WREG32(EVERGREEN_MM_INDEX_HI, *pos >> 31);
1078 value = RREG32(RADEON_MM_DATA);
1079 spin_unlock_irqrestore(&rdev->mmio_idx_lock, flags);
1080
1081 r = put_user(value, (uint32_t *)buf);
1082 if (r)
1083 return r;
1084
1085 result += 4;
1086 buf += 4;
1087 *pos += 4;
1088 size -= 4;
1089 }
1090
1091 return result;
1092}
1093
1094static const struct file_operations radeon_ttm_vram_fops = {
1095 .owner = THIS_MODULE,
1096 .open = radeon_ttm_vram_open,
1097 .read = radeon_ttm_vram_read,
1098 .llseek = default_llseek
1099};
1100
1101static int radeon_ttm_gtt_open(struct inode *inode, struct file *filep)
1102{
1103 struct radeon_device *rdev = inode->i_private;
1104 i_size_write(inode, rdev->mc.gtt_size);
1105 filep->private_data = inode->i_private;
1106 return 0;
1107}
1108
1109static ssize_t radeon_ttm_gtt_read(struct file *f, char __user *buf,
1110 size_t size, loff_t *pos)
1111{
1112 struct radeon_device *rdev = f->private_data;
1113 ssize_t result = 0;
1114 int r;
1115
1116 while (size) {
1117 loff_t p = *pos / PAGE_SIZE;
1118 unsigned off = *pos & ~PAGE_MASK;
1119 size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
1120 struct page *page;
1121 void *ptr;
1122
1123 if (p >= rdev->gart.num_cpu_pages)
1124 return result;
1125
1126 page = rdev->gart.pages[p];
1127 if (page) {
1128 ptr = kmap(page);
1129 ptr += off;
1130
1131 r = copy_to_user(buf, ptr, cur_size);
1132 kunmap(rdev->gart.pages[p]);
1133 } else
1134 r = clear_user(buf, cur_size);
1135
1136 if (r)
1137 return -EFAULT;
1138
1139 result += cur_size;
1140 buf += cur_size;
1141 *pos += cur_size;
1142 size -= cur_size;
1143 }
1144
1145 return result;
1146}
1147
1148static const struct file_operations radeon_ttm_gtt_fops = {
1149 .owner = THIS_MODULE,
1150 .open = radeon_ttm_gtt_open,
1151 .read = radeon_ttm_gtt_read,
1152 .llseek = default_llseek
1153};
1154
1155#endif
1156
1157static int radeon_ttm_debugfs_init(struct radeon_device *rdev)
1158{
1159#if defined(CONFIG_DEBUG_FS)
1160 unsigned count;
1161
1162 struct drm_minor *minor = rdev->ddev->primary;
1163 struct dentry *ent, *root = minor->debugfs_root;
1164
1165 ent = debugfs_create_file("radeon_vram", S_IFREG | S_IRUGO, root,
1166 rdev, &radeon_ttm_vram_fops);
1167 if (IS_ERR(ent))
1168 return PTR_ERR(ent);
1169 rdev->mman.vram = ent;
1170
1171 ent = debugfs_create_file("radeon_gtt", S_IFREG | S_IRUGO, root,
1172 rdev, &radeon_ttm_gtt_fops);
1173 if (IS_ERR(ent))
1174 return PTR_ERR(ent);
1175 rdev->mman.gtt = ent;
1176
1177 count = ARRAY_SIZE(radeon_ttm_debugfs_list);
1178
1179#ifdef CONFIG_SWIOTLB
1180 if (!swiotlb_nr_tbl())
1181 --count;
1182#endif
1183
1184 return radeon_debugfs_add_files(rdev, radeon_ttm_debugfs_list, count);
1185#else
1186
1187 return 0;
1188#endif
1189}
1190
1191static void radeon_ttm_debugfs_fini(struct radeon_device *rdev)
1192{
1193#if defined(CONFIG_DEBUG_FS)
1194
1195 debugfs_remove(rdev->mman.vram);
1196 rdev->mman.vram = NULL;
1197
1198 debugfs_remove(rdev->mman.gtt);
1199 rdev->mman.gtt = NULL;
1200#endif
1201}
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/pagemap.h>
35#include <linux/pci.h>
36#include <linux/seq_file.h>
37#include <linux/slab.h>
38#include <linux/swap.h>
39#include <linux/swiotlb.h>
40
41#include <drm/drm_device.h>
42#include <drm/drm_file.h>
43#include <drm/drm_prime.h>
44#include <drm/radeon_drm.h>
45#include <drm/ttm/ttm_bo_api.h>
46#include <drm/ttm/ttm_bo_driver.h>
47#include <drm/ttm/ttm_placement.h>
48#include <drm/ttm/ttm_range_manager.h>
49
50#include "radeon_reg.h"
51#include "radeon.h"
52#include "radeon_ttm.h"
53
54static void radeon_ttm_debugfs_init(struct radeon_device *rdev);
55
56static int radeon_ttm_tt_bind(struct ttm_device *bdev, struct ttm_tt *ttm,
57 struct ttm_resource *bo_mem);
58static void radeon_ttm_tt_unbind(struct ttm_device *bdev, struct ttm_tt *ttm);
59
60struct radeon_device *radeon_get_rdev(struct ttm_device *bdev)
61{
62 struct radeon_mman *mman;
63 struct radeon_device *rdev;
64
65 mman = container_of(bdev, struct radeon_mman, bdev);
66 rdev = container_of(mman, struct radeon_device, mman);
67 return rdev;
68}
69
70static int radeon_ttm_init_vram(struct radeon_device *rdev)
71{
72 return ttm_range_man_init(&rdev->mman.bdev, TTM_PL_VRAM,
73 false, rdev->mc.real_vram_size >> PAGE_SHIFT);
74}
75
76static int radeon_ttm_init_gtt(struct radeon_device *rdev)
77{
78 return ttm_range_man_init(&rdev->mman.bdev, TTM_PL_TT,
79 true, rdev->mc.gtt_size >> PAGE_SHIFT);
80}
81
82static void radeon_evict_flags(struct ttm_buffer_object *bo,
83 struct ttm_placement *placement)
84{
85 static const struct ttm_place placements = {
86 .fpfn = 0,
87 .lpfn = 0,
88 .mem_type = TTM_PL_SYSTEM,
89 .flags = 0
90 };
91
92 struct radeon_bo *rbo;
93
94 if (!radeon_ttm_bo_is_radeon_bo(bo)) {
95 placement->placement = &placements;
96 placement->busy_placement = &placements;
97 placement->num_placement = 1;
98 placement->num_busy_placement = 1;
99 return;
100 }
101 rbo = container_of(bo, struct radeon_bo, tbo);
102 switch (bo->resource->mem_type) {
103 case TTM_PL_VRAM:
104 if (rbo->rdev->ring[radeon_copy_ring_index(rbo->rdev)].ready == false)
105 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
106 else if (rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size &&
107 bo->resource->start < (rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT)) {
108 unsigned fpfn = rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
109 int i;
110
111 /* Try evicting to the CPU inaccessible part of VRAM
112 * first, but only set GTT as busy placement, so this
113 * BO will be evicted to GTT rather than causing other
114 * BOs to be evicted from VRAM
115 */
116 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM |
117 RADEON_GEM_DOMAIN_GTT);
118 rbo->placement.num_busy_placement = 0;
119 for (i = 0; i < rbo->placement.num_placement; i++) {
120 if (rbo->placements[i].mem_type == TTM_PL_VRAM) {
121 if (rbo->placements[i].fpfn < fpfn)
122 rbo->placements[i].fpfn = fpfn;
123 } else {
124 rbo->placement.busy_placement =
125 &rbo->placements[i];
126 rbo->placement.num_busy_placement = 1;
127 }
128 }
129 } else
130 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
131 break;
132 case TTM_PL_TT:
133 default:
134 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
135 }
136 *placement = rbo->placement;
137}
138
139static int radeon_move_blit(struct ttm_buffer_object *bo,
140 bool evict,
141 struct ttm_resource *new_mem,
142 struct ttm_resource *old_mem)
143{
144 struct radeon_device *rdev;
145 uint64_t old_start, new_start;
146 struct radeon_fence *fence;
147 unsigned num_pages;
148 int r, ridx;
149
150 rdev = radeon_get_rdev(bo->bdev);
151 ridx = radeon_copy_ring_index(rdev);
152 old_start = (u64)old_mem->start << PAGE_SHIFT;
153 new_start = (u64)new_mem->start << PAGE_SHIFT;
154
155 switch (old_mem->mem_type) {
156 case TTM_PL_VRAM:
157 old_start += rdev->mc.vram_start;
158 break;
159 case TTM_PL_TT:
160 old_start += rdev->mc.gtt_start;
161 break;
162 default:
163 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
164 return -EINVAL;
165 }
166 switch (new_mem->mem_type) {
167 case TTM_PL_VRAM:
168 new_start += rdev->mc.vram_start;
169 break;
170 case TTM_PL_TT:
171 new_start += rdev->mc.gtt_start;
172 break;
173 default:
174 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
175 return -EINVAL;
176 }
177 if (!rdev->ring[ridx].ready) {
178 DRM_ERROR("Trying to move memory with ring turned off.\n");
179 return -EINVAL;
180 }
181
182 BUILD_BUG_ON((PAGE_SIZE % RADEON_GPU_PAGE_SIZE) != 0);
183
184 num_pages = PFN_UP(new_mem->size) * (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
185 fence = radeon_copy(rdev, old_start, new_start, num_pages, bo->base.resv);
186 if (IS_ERR(fence))
187 return PTR_ERR(fence);
188
189 r = ttm_bo_move_accel_cleanup(bo, &fence->base, evict, false, new_mem);
190 radeon_fence_unref(&fence);
191 return r;
192}
193
194static int radeon_bo_move(struct ttm_buffer_object *bo, bool evict,
195 struct ttm_operation_ctx *ctx,
196 struct ttm_resource *new_mem,
197 struct ttm_place *hop)
198{
199 struct ttm_resource *old_mem = bo->resource;
200 struct radeon_device *rdev;
201 struct radeon_bo *rbo;
202 int r;
203
204 if (new_mem->mem_type == TTM_PL_TT) {
205 r = radeon_ttm_tt_bind(bo->bdev, bo->ttm, new_mem);
206 if (r)
207 return r;
208 }
209
210 r = ttm_bo_wait_ctx(bo, ctx);
211 if (r)
212 return r;
213
214 /* Can't move a pinned BO */
215 rbo = container_of(bo, struct radeon_bo, tbo);
216 if (WARN_ON_ONCE(rbo->tbo.pin_count > 0))
217 return -EINVAL;
218
219 rdev = radeon_get_rdev(bo->bdev);
220 if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
221 ttm_bo_move_null(bo, new_mem);
222 goto out;
223 }
224 if (old_mem->mem_type == TTM_PL_SYSTEM &&
225 new_mem->mem_type == TTM_PL_TT) {
226 ttm_bo_move_null(bo, new_mem);
227 goto out;
228 }
229
230 if (old_mem->mem_type == TTM_PL_TT &&
231 new_mem->mem_type == TTM_PL_SYSTEM) {
232 radeon_ttm_tt_unbind(bo->bdev, bo->ttm);
233 ttm_resource_free(bo, &bo->resource);
234 ttm_bo_assign_mem(bo, new_mem);
235 goto out;
236 }
237 if (rdev->ring[radeon_copy_ring_index(rdev)].ready &&
238 rdev->asic->copy.copy != NULL) {
239 if ((old_mem->mem_type == TTM_PL_SYSTEM &&
240 new_mem->mem_type == TTM_PL_VRAM) ||
241 (old_mem->mem_type == TTM_PL_VRAM &&
242 new_mem->mem_type == TTM_PL_SYSTEM)) {
243 hop->fpfn = 0;
244 hop->lpfn = 0;
245 hop->mem_type = TTM_PL_TT;
246 hop->flags = 0;
247 return -EMULTIHOP;
248 }
249
250 r = radeon_move_blit(bo, evict, new_mem, old_mem);
251 } else {
252 r = -ENODEV;
253 }
254
255 if (r) {
256 r = ttm_bo_move_memcpy(bo, ctx, new_mem);
257 if (r)
258 return r;
259 }
260
261out:
262 /* update statistics */
263 atomic64_add(bo->base.size, &rdev->num_bytes_moved);
264 radeon_bo_move_notify(bo);
265 return 0;
266}
267
268static int radeon_ttm_io_mem_reserve(struct ttm_device *bdev, struct ttm_resource *mem)
269{
270 struct radeon_device *rdev = radeon_get_rdev(bdev);
271 size_t bus_size = (size_t)mem->size;
272
273 switch (mem->mem_type) {
274 case TTM_PL_SYSTEM:
275 /* system memory */
276 return 0;
277 case TTM_PL_TT:
278#if IS_ENABLED(CONFIG_AGP)
279 if (rdev->flags & RADEON_IS_AGP) {
280 /* RADEON_IS_AGP is set only if AGP is active */
281 mem->bus.offset = (mem->start << PAGE_SHIFT) +
282 rdev->mc.agp_base;
283 mem->bus.is_iomem = !rdev->agp->cant_use_aperture;
284 mem->bus.caching = ttm_write_combined;
285 }
286#endif
287 break;
288 case TTM_PL_VRAM:
289 mem->bus.offset = mem->start << PAGE_SHIFT;
290 /* check if it's visible */
291 if ((mem->bus.offset + bus_size) > rdev->mc.visible_vram_size)
292 return -EINVAL;
293 mem->bus.offset += rdev->mc.aper_base;
294 mem->bus.is_iomem = true;
295 mem->bus.caching = ttm_write_combined;
296#ifdef __alpha__
297 /*
298 * Alpha: use bus.addr to hold the ioremap() return,
299 * so we can modify bus.base below.
300 */
301 mem->bus.addr = ioremap_wc(mem->bus.offset, bus_size);
302 if (!mem->bus.addr)
303 return -ENOMEM;
304
305 /*
306 * Alpha: Use just the bus offset plus
307 * the hose/domain memory base for bus.base.
308 * It then can be used to build PTEs for VRAM
309 * access, as done in ttm_bo_vm_fault().
310 */
311 mem->bus.offset = (mem->bus.offset & 0x0ffffffffUL) +
312 rdev->hose->dense_mem_base;
313#endif
314 break;
315 default:
316 return -EINVAL;
317 }
318 return 0;
319}
320
321/*
322 * TTM backend functions.
323 */
324struct radeon_ttm_tt {
325 struct ttm_tt ttm;
326 u64 offset;
327
328 uint64_t userptr;
329 struct mm_struct *usermm;
330 uint32_t userflags;
331 bool bound;
332};
333
334/* prepare the sg table with the user pages */
335static int radeon_ttm_tt_pin_userptr(struct ttm_device *bdev, struct ttm_tt *ttm)
336{
337 struct radeon_device *rdev = radeon_get_rdev(bdev);
338 struct radeon_ttm_tt *gtt = (void *)ttm;
339 unsigned pinned = 0;
340 int r;
341
342 int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
343 enum dma_data_direction direction = write ?
344 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
345
346 if (current->mm != gtt->usermm)
347 return -EPERM;
348
349 if (gtt->userflags & RADEON_GEM_USERPTR_ANONONLY) {
350 /* check that we only pin down anonymous memory
351 to prevent problems with writeback */
352 unsigned long end = gtt->userptr + (u64)ttm->num_pages * PAGE_SIZE;
353 struct vm_area_struct *vma;
354 vma = find_vma(gtt->usermm, gtt->userptr);
355 if (!vma || vma->vm_file || vma->vm_end < end)
356 return -EPERM;
357 }
358
359 do {
360 unsigned num_pages = ttm->num_pages - pinned;
361 uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
362 struct page **pages = ttm->pages + pinned;
363
364 r = get_user_pages(userptr, num_pages, write ? FOLL_WRITE : 0,
365 pages, NULL);
366 if (r < 0)
367 goto release_pages;
368
369 pinned += r;
370
371 } while (pinned < ttm->num_pages);
372
373 r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
374 (u64)ttm->num_pages << PAGE_SHIFT,
375 GFP_KERNEL);
376 if (r)
377 goto release_sg;
378
379 r = dma_map_sgtable(rdev->dev, ttm->sg, direction, 0);
380 if (r)
381 goto release_sg;
382
383 drm_prime_sg_to_dma_addr_array(ttm->sg, gtt->ttm.dma_address,
384 ttm->num_pages);
385
386 return 0;
387
388release_sg:
389 kfree(ttm->sg);
390
391release_pages:
392 release_pages(ttm->pages, pinned);
393 return r;
394}
395
396static void radeon_ttm_tt_unpin_userptr(struct ttm_device *bdev, struct ttm_tt *ttm)
397{
398 struct radeon_device *rdev = radeon_get_rdev(bdev);
399 struct radeon_ttm_tt *gtt = (void *)ttm;
400 struct sg_page_iter sg_iter;
401
402 int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
403 enum dma_data_direction direction = write ?
404 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
405
406 /* double check that we don't free the table twice */
407 if (!ttm->sg || !ttm->sg->sgl)
408 return;
409
410 /* free the sg table and pages again */
411 dma_unmap_sgtable(rdev->dev, ttm->sg, direction, 0);
412
413 for_each_sgtable_page(ttm->sg, &sg_iter, 0) {
414 struct page *page = sg_page_iter_page(&sg_iter);
415 if (!(gtt->userflags & RADEON_GEM_USERPTR_READONLY))
416 set_page_dirty(page);
417
418 mark_page_accessed(page);
419 put_page(page);
420 }
421
422 sg_free_table(ttm->sg);
423}
424
425static bool radeon_ttm_backend_is_bound(struct ttm_tt *ttm)
426{
427 struct radeon_ttm_tt *gtt = (void*)ttm;
428
429 return (gtt->bound);
430}
431
432static int radeon_ttm_backend_bind(struct ttm_device *bdev,
433 struct ttm_tt *ttm,
434 struct ttm_resource *bo_mem)
435{
436 struct radeon_ttm_tt *gtt = (void*)ttm;
437 struct radeon_device *rdev = radeon_get_rdev(bdev);
438 uint32_t flags = RADEON_GART_PAGE_VALID | RADEON_GART_PAGE_READ |
439 RADEON_GART_PAGE_WRITE;
440 int r;
441
442 if (gtt->bound)
443 return 0;
444
445 if (gtt->userptr) {
446 radeon_ttm_tt_pin_userptr(bdev, ttm);
447 flags &= ~RADEON_GART_PAGE_WRITE;
448 }
449
450 gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT);
451 if (!ttm->num_pages) {
452 WARN(1, "nothing to bind %u pages for mreg %p back %p!\n",
453 ttm->num_pages, bo_mem, ttm);
454 }
455 if (ttm->caching == ttm_cached)
456 flags |= RADEON_GART_PAGE_SNOOP;
457 r = radeon_gart_bind(rdev, gtt->offset, ttm->num_pages,
458 ttm->pages, gtt->ttm.dma_address, flags);
459 if (r) {
460 DRM_ERROR("failed to bind %u pages at 0x%08X\n",
461 ttm->num_pages, (unsigned)gtt->offset);
462 return r;
463 }
464 gtt->bound = true;
465 return 0;
466}
467
468static void radeon_ttm_backend_unbind(struct ttm_device *bdev, struct ttm_tt *ttm)
469{
470 struct radeon_ttm_tt *gtt = (void *)ttm;
471 struct radeon_device *rdev = radeon_get_rdev(bdev);
472
473 if (gtt->userptr)
474 radeon_ttm_tt_unpin_userptr(bdev, ttm);
475
476 if (!gtt->bound)
477 return;
478
479 radeon_gart_unbind(rdev, gtt->offset, ttm->num_pages);
480
481 gtt->bound = false;
482}
483
484static void radeon_ttm_backend_destroy(struct ttm_device *bdev, struct ttm_tt *ttm)
485{
486 struct radeon_ttm_tt *gtt = (void *)ttm;
487
488 ttm_tt_fini(>t->ttm);
489 kfree(gtt);
490}
491
492static struct ttm_tt *radeon_ttm_tt_create(struct ttm_buffer_object *bo,
493 uint32_t page_flags)
494{
495 struct radeon_ttm_tt *gtt;
496 enum ttm_caching caching;
497 struct radeon_bo *rbo;
498#if IS_ENABLED(CONFIG_AGP)
499 struct radeon_device *rdev = radeon_get_rdev(bo->bdev);
500
501 if (rdev->flags & RADEON_IS_AGP) {
502 return ttm_agp_tt_create(bo, rdev->agp->bridge, page_flags);
503 }
504#endif
505 rbo = container_of(bo, struct radeon_bo, tbo);
506
507 gtt = kzalloc(sizeof(struct radeon_ttm_tt), GFP_KERNEL);
508 if (gtt == NULL) {
509 return NULL;
510 }
511
512 if (rbo->flags & RADEON_GEM_GTT_UC)
513 caching = ttm_uncached;
514 else if (rbo->flags & RADEON_GEM_GTT_WC)
515 caching = ttm_write_combined;
516 else
517 caching = ttm_cached;
518
519 if (ttm_sg_tt_init(>t->ttm, bo, page_flags, caching)) {
520 kfree(gtt);
521 return NULL;
522 }
523 return >t->ttm;
524}
525
526static struct radeon_ttm_tt *radeon_ttm_tt_to_gtt(struct radeon_device *rdev,
527 struct ttm_tt *ttm)
528{
529#if IS_ENABLED(CONFIG_AGP)
530 if (rdev->flags & RADEON_IS_AGP)
531 return NULL;
532#endif
533
534 if (!ttm)
535 return NULL;
536 return container_of(ttm, struct radeon_ttm_tt, ttm);
537}
538
539static int radeon_ttm_tt_populate(struct ttm_device *bdev,
540 struct ttm_tt *ttm,
541 struct ttm_operation_ctx *ctx)
542{
543 struct radeon_device *rdev = radeon_get_rdev(bdev);
544 struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm);
545 bool slave = !!(ttm->page_flags & TTM_TT_FLAG_EXTERNAL);
546
547 if (gtt && gtt->userptr) {
548 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
549 if (!ttm->sg)
550 return -ENOMEM;
551
552 ttm->page_flags |= TTM_TT_FLAG_EXTERNAL;
553 return 0;
554 }
555
556 if (slave && ttm->sg) {
557 drm_prime_sg_to_dma_addr_array(ttm->sg, gtt->ttm.dma_address,
558 ttm->num_pages);
559 return 0;
560 }
561
562 return ttm_pool_alloc(&rdev->mman.bdev.pool, ttm, ctx);
563}
564
565static void radeon_ttm_tt_unpopulate(struct ttm_device *bdev, struct ttm_tt *ttm)
566{
567 struct radeon_device *rdev = radeon_get_rdev(bdev);
568 struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm);
569 bool slave = !!(ttm->page_flags & TTM_TT_FLAG_EXTERNAL);
570
571 radeon_ttm_tt_unbind(bdev, ttm);
572
573 if (gtt && gtt->userptr) {
574 kfree(ttm->sg);
575 ttm->page_flags &= ~TTM_TT_FLAG_EXTERNAL;
576 return;
577 }
578
579 if (slave)
580 return;
581
582 return ttm_pool_free(&rdev->mman.bdev.pool, ttm);
583}
584
585int radeon_ttm_tt_set_userptr(struct radeon_device *rdev,
586 struct ttm_tt *ttm, uint64_t addr,
587 uint32_t flags)
588{
589 struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm);
590
591 if (gtt == NULL)
592 return -EINVAL;
593
594 gtt->userptr = addr;
595 gtt->usermm = current->mm;
596 gtt->userflags = flags;
597 return 0;
598}
599
600bool radeon_ttm_tt_is_bound(struct ttm_device *bdev,
601 struct ttm_tt *ttm)
602{
603#if IS_ENABLED(CONFIG_AGP)
604 struct radeon_device *rdev = radeon_get_rdev(bdev);
605 if (rdev->flags & RADEON_IS_AGP)
606 return ttm_agp_is_bound(ttm);
607#endif
608 return radeon_ttm_backend_is_bound(ttm);
609}
610
611static int radeon_ttm_tt_bind(struct ttm_device *bdev,
612 struct ttm_tt *ttm,
613 struct ttm_resource *bo_mem)
614{
615#if IS_ENABLED(CONFIG_AGP)
616 struct radeon_device *rdev = radeon_get_rdev(bdev);
617#endif
618
619 if (!bo_mem)
620 return -EINVAL;
621#if IS_ENABLED(CONFIG_AGP)
622 if (rdev->flags & RADEON_IS_AGP)
623 return ttm_agp_bind(ttm, bo_mem);
624#endif
625
626 return radeon_ttm_backend_bind(bdev, ttm, bo_mem);
627}
628
629static void radeon_ttm_tt_unbind(struct ttm_device *bdev,
630 struct ttm_tt *ttm)
631{
632#if IS_ENABLED(CONFIG_AGP)
633 struct radeon_device *rdev = radeon_get_rdev(bdev);
634
635 if (rdev->flags & RADEON_IS_AGP) {
636 ttm_agp_unbind(ttm);
637 return;
638 }
639#endif
640 radeon_ttm_backend_unbind(bdev, ttm);
641}
642
643static void radeon_ttm_tt_destroy(struct ttm_device *bdev,
644 struct ttm_tt *ttm)
645{
646#if IS_ENABLED(CONFIG_AGP)
647 struct radeon_device *rdev = radeon_get_rdev(bdev);
648
649 if (rdev->flags & RADEON_IS_AGP) {
650 ttm_agp_destroy(ttm);
651 return;
652 }
653#endif
654 radeon_ttm_backend_destroy(bdev, ttm);
655}
656
657bool radeon_ttm_tt_has_userptr(struct radeon_device *rdev,
658 struct ttm_tt *ttm)
659{
660 struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm);
661
662 if (gtt == NULL)
663 return false;
664
665 return !!gtt->userptr;
666}
667
668bool radeon_ttm_tt_is_readonly(struct radeon_device *rdev,
669 struct ttm_tt *ttm)
670{
671 struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm);
672
673 if (gtt == NULL)
674 return false;
675
676 return !!(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
677}
678
679static struct ttm_device_funcs radeon_bo_driver = {
680 .ttm_tt_create = &radeon_ttm_tt_create,
681 .ttm_tt_populate = &radeon_ttm_tt_populate,
682 .ttm_tt_unpopulate = &radeon_ttm_tt_unpopulate,
683 .ttm_tt_destroy = &radeon_ttm_tt_destroy,
684 .eviction_valuable = ttm_bo_eviction_valuable,
685 .evict_flags = &radeon_evict_flags,
686 .move = &radeon_bo_move,
687 .io_mem_reserve = &radeon_ttm_io_mem_reserve,
688};
689
690int radeon_ttm_init(struct radeon_device *rdev)
691{
692 int r;
693
694 /* No others user of address space so set it to 0 */
695 r = ttm_device_init(&rdev->mman.bdev, &radeon_bo_driver, rdev->dev,
696 rdev->ddev->anon_inode->i_mapping,
697 rdev->ddev->vma_offset_manager,
698 rdev->need_swiotlb,
699 dma_addressing_limited(&rdev->pdev->dev));
700 if (r) {
701 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
702 return r;
703 }
704 rdev->mman.initialized = true;
705
706 r = radeon_ttm_init_vram(rdev);
707 if (r) {
708 DRM_ERROR("Failed initializing VRAM heap.\n");
709 return r;
710 }
711 /* Change the size here instead of the init above so only lpfn is affected */
712 radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
713
714 r = radeon_bo_create(rdev, 256 * 1024, PAGE_SIZE, true,
715 RADEON_GEM_DOMAIN_VRAM, 0, NULL,
716 NULL, &rdev->stolen_vga_memory);
717 if (r) {
718 return r;
719 }
720 r = radeon_bo_reserve(rdev->stolen_vga_memory, false);
721 if (r)
722 return r;
723 r = radeon_bo_pin(rdev->stolen_vga_memory, RADEON_GEM_DOMAIN_VRAM, NULL);
724 radeon_bo_unreserve(rdev->stolen_vga_memory);
725 if (r) {
726 radeon_bo_unref(&rdev->stolen_vga_memory);
727 return r;
728 }
729 DRM_INFO("radeon: %uM of VRAM memory ready\n",
730 (unsigned) (rdev->mc.real_vram_size / (1024 * 1024)));
731
732 r = radeon_ttm_init_gtt(rdev);
733 if (r) {
734 DRM_ERROR("Failed initializing GTT heap.\n");
735 return r;
736 }
737 DRM_INFO("radeon: %uM of GTT memory ready.\n",
738 (unsigned)(rdev->mc.gtt_size / (1024 * 1024)));
739
740 radeon_ttm_debugfs_init(rdev);
741
742 return 0;
743}
744
745void radeon_ttm_fini(struct radeon_device *rdev)
746{
747 int r;
748
749 if (!rdev->mman.initialized)
750 return;
751
752 if (rdev->stolen_vga_memory) {
753 r = radeon_bo_reserve(rdev->stolen_vga_memory, false);
754 if (r == 0) {
755 radeon_bo_unpin(rdev->stolen_vga_memory);
756 radeon_bo_unreserve(rdev->stolen_vga_memory);
757 }
758 radeon_bo_unref(&rdev->stolen_vga_memory);
759 }
760 ttm_range_man_fini(&rdev->mman.bdev, TTM_PL_VRAM);
761 ttm_range_man_fini(&rdev->mman.bdev, TTM_PL_TT);
762 ttm_device_fini(&rdev->mman.bdev);
763 radeon_gart_fini(rdev);
764 rdev->mman.initialized = false;
765 DRM_INFO("radeon: ttm finalized\n");
766}
767
768/* this should only be called at bootup or when userspace
769 * isn't running */
770void radeon_ttm_set_active_vram_size(struct radeon_device *rdev, u64 size)
771{
772 struct ttm_resource_manager *man;
773
774 if (!rdev->mman.initialized)
775 return;
776
777 man = ttm_manager_type(&rdev->mman.bdev, TTM_PL_VRAM);
778 /* this just adjusts TTM size idea, which sets lpfn to the correct value */
779 man->size = size >> PAGE_SHIFT;
780}
781
782#if defined(CONFIG_DEBUG_FS)
783
784static int radeon_ttm_page_pool_show(struct seq_file *m, void *data)
785{
786 struct radeon_device *rdev = (struct radeon_device *)m->private;
787
788 return ttm_pool_debugfs(&rdev->mman.bdev.pool, m);
789}
790
791DEFINE_SHOW_ATTRIBUTE(radeon_ttm_page_pool);
792
793static int radeon_ttm_vram_open(struct inode *inode, struct file *filep)
794{
795 struct radeon_device *rdev = inode->i_private;
796 i_size_write(inode, rdev->mc.mc_vram_size);
797 filep->private_data = inode->i_private;
798 return 0;
799}
800
801static ssize_t radeon_ttm_vram_read(struct file *f, char __user *buf,
802 size_t size, loff_t *pos)
803{
804 struct radeon_device *rdev = f->private_data;
805 ssize_t result = 0;
806 int r;
807
808 if (size & 0x3 || *pos & 0x3)
809 return -EINVAL;
810
811 while (size) {
812 unsigned long flags;
813 uint32_t value;
814
815 if (*pos >= rdev->mc.mc_vram_size)
816 return result;
817
818 spin_lock_irqsave(&rdev->mmio_idx_lock, flags);
819 WREG32(RADEON_MM_INDEX, ((uint32_t)*pos) | 0x80000000);
820 if (rdev->family >= CHIP_CEDAR)
821 WREG32(EVERGREEN_MM_INDEX_HI, *pos >> 31);
822 value = RREG32(RADEON_MM_DATA);
823 spin_unlock_irqrestore(&rdev->mmio_idx_lock, flags);
824
825 r = put_user(value, (uint32_t __user *)buf);
826 if (r)
827 return r;
828
829 result += 4;
830 buf += 4;
831 *pos += 4;
832 size -= 4;
833 }
834
835 return result;
836}
837
838static const struct file_operations radeon_ttm_vram_fops = {
839 .owner = THIS_MODULE,
840 .open = radeon_ttm_vram_open,
841 .read = radeon_ttm_vram_read,
842 .llseek = default_llseek
843};
844
845static int radeon_ttm_gtt_open(struct inode *inode, struct file *filep)
846{
847 struct radeon_device *rdev = inode->i_private;
848 i_size_write(inode, rdev->mc.gtt_size);
849 filep->private_data = inode->i_private;
850 return 0;
851}
852
853static ssize_t radeon_ttm_gtt_read(struct file *f, char __user *buf,
854 size_t size, loff_t *pos)
855{
856 struct radeon_device *rdev = f->private_data;
857 ssize_t result = 0;
858 int r;
859
860 while (size) {
861 loff_t p = *pos / PAGE_SIZE;
862 unsigned off = *pos & ~PAGE_MASK;
863 size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
864 struct page *page;
865 void *ptr;
866
867 if (p >= rdev->gart.num_cpu_pages)
868 return result;
869
870 page = rdev->gart.pages[p];
871 if (page) {
872 ptr = kmap_local_page(page);
873 ptr += off;
874
875 r = copy_to_user(buf, ptr, cur_size);
876 kunmap_local(ptr);
877 } else
878 r = clear_user(buf, cur_size);
879
880 if (r)
881 return -EFAULT;
882
883 result += cur_size;
884 buf += cur_size;
885 *pos += cur_size;
886 size -= cur_size;
887 }
888
889 return result;
890}
891
892static const struct file_operations radeon_ttm_gtt_fops = {
893 .owner = THIS_MODULE,
894 .open = radeon_ttm_gtt_open,
895 .read = radeon_ttm_gtt_read,
896 .llseek = default_llseek
897};
898
899#endif
900
901static void radeon_ttm_debugfs_init(struct radeon_device *rdev)
902{
903#if defined(CONFIG_DEBUG_FS)
904 struct drm_minor *minor = rdev->ddev->primary;
905 struct dentry *root = minor->debugfs_root;
906
907 debugfs_create_file("radeon_vram", 0444, root, rdev,
908 &radeon_ttm_vram_fops);
909 debugfs_create_file("radeon_gtt", 0444, root, rdev,
910 &radeon_ttm_gtt_fops);
911 debugfs_create_file("ttm_page_pool", 0444, root, rdev,
912 &radeon_ttm_page_pool_fops);
913 ttm_resource_manager_create_debugfs(ttm_manager_type(&rdev->mman.bdev,
914 TTM_PL_VRAM),
915 root, "radeon_vram_mm");
916 ttm_resource_manager_create_debugfs(ttm_manager_type(&rdev->mman.bdev,
917 TTM_PL_TT),
918 root, "radeon_gtt_mm");
919#endif
920}