Linux 4.2.1
[linux/fpc-iii.git] / drivers / gpu / drm / radeon / radeon_ttm.c
blob06ac59fe332ab089d21b279aba092f615710c7ca
1 /*
2 * Copyright 2009 Jerome Glisse.
3 * All Rights Reserved.
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:
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.
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.
27 * Authors:
28 * Jerome Glisse <glisse@freedesktop.org>
29 * Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
30 * Dave Airlie
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"
48 #define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT)
50 static int radeon_ttm_debugfs_init(struct radeon_device *rdev);
51 static void radeon_ttm_debugfs_fini(struct radeon_device *rdev);
53 static struct radeon_device *radeon_get_rdev(struct ttm_bo_device *bdev)
55 struct radeon_mman *mman;
56 struct radeon_device *rdev;
58 mman = container_of(bdev, struct radeon_mman, bdev);
59 rdev = container_of(mman, struct radeon_device, mman);
60 return rdev;
65 * Global memory.
67 static int radeon_ttm_mem_global_init(struct drm_global_reference *ref)
69 return ttm_mem_global_init(ref->object);
72 static void radeon_ttm_mem_global_release(struct drm_global_reference *ref)
74 ttm_mem_global_release(ref->object);
77 static int radeon_ttm_global_init(struct radeon_device *rdev)
79 struct drm_global_reference *global_ref;
80 int r;
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;
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;
109 rdev->mman.mem_global_referenced = true;
110 return 0;
113 static void radeon_ttm_global_fini(struct radeon_device *rdev)
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;
122 static int radeon_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
124 return 0;
127 static int radeon_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
128 struct ttm_mem_type_manager *man)
130 struct radeon_device *rdev;
132 rdev = radeon_get_rdev(bdev);
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 __OS_HAS_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;
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;
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;
175 return 0;
178 static void radeon_evict_flags(struct ttm_buffer_object *bo,
179 struct ttm_placement *placement)
181 static struct ttm_place placements = {
182 .fpfn = 0,
183 .lpfn = 0,
184 .flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
187 struct radeon_bo *rbo;
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;
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;
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
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;
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);
231 *placement = rbo->placement;
234 static int radeon_verify_access(struct ttm_buffer_object *bo, struct file *filp)
236 struct radeon_bo *rbo = container_of(bo, struct radeon_bo, tbo);
238 return drm_vma_node_verify_access(&rbo->gem_base.vma_node, filp);
241 static void radeon_move_null(struct ttm_buffer_object *bo,
242 struct ttm_mem_reg *new_mem)
244 struct ttm_mem_reg *old_mem = &bo->mem;
246 BUG_ON(old_mem->mm_node != NULL);
247 *old_mem = *new_mem;
248 new_mem->mm_node = NULL;
251 static int radeon_move_blit(struct ttm_buffer_object *bo,
252 bool evict, bool no_wait_gpu,
253 struct ttm_mem_reg *new_mem,
254 struct ttm_mem_reg *old_mem)
256 struct radeon_device *rdev;
257 uint64_t old_start, new_start;
258 struct radeon_fence *fence;
259 unsigned num_pages;
260 int r, ridx;
262 rdev = radeon_get_rdev(bo->bdev);
263 ridx = radeon_copy_ring_index(rdev);
264 old_start = old_mem->start << PAGE_SHIFT;
265 new_start = new_mem->start << PAGE_SHIFT;
267 switch (old_mem->mem_type) {
268 case TTM_PL_VRAM:
269 old_start += rdev->mc.vram_start;
270 break;
271 case TTM_PL_TT:
272 old_start += rdev->mc.gtt_start;
273 break;
274 default:
275 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
276 return -EINVAL;
278 switch (new_mem->mem_type) {
279 case TTM_PL_VRAM:
280 new_start += rdev->mc.vram_start;
281 break;
282 case TTM_PL_TT:
283 new_start += rdev->mc.gtt_start;
284 break;
285 default:
286 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
287 return -EINVAL;
289 if (!rdev->ring[ridx].ready) {
290 DRM_ERROR("Trying to move memory with ring turned off.\n");
291 return -EINVAL;
294 BUILD_BUG_ON((PAGE_SIZE % RADEON_GPU_PAGE_SIZE) != 0);
296 num_pages = new_mem->num_pages * (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
297 fence = radeon_copy(rdev, old_start, new_start, num_pages, bo->resv);
298 if (IS_ERR(fence))
299 return PTR_ERR(fence);
301 r = ttm_bo_move_accel_cleanup(bo, &fence->base,
302 evict, no_wait_gpu, new_mem);
303 radeon_fence_unref(&fence);
304 return r;
307 static int radeon_move_vram_ram(struct ttm_buffer_object *bo,
308 bool evict, bool interruptible,
309 bool no_wait_gpu,
310 struct ttm_mem_reg *new_mem)
312 struct radeon_device *rdev;
313 struct ttm_mem_reg *old_mem = &bo->mem;
314 struct ttm_mem_reg tmp_mem;
315 struct ttm_place placements;
316 struct ttm_placement placement;
317 int r;
319 rdev = radeon_get_rdev(bo->bdev);
320 tmp_mem = *new_mem;
321 tmp_mem.mm_node = NULL;
322 placement.num_placement = 1;
323 placement.placement = &placements;
324 placement.num_busy_placement = 1;
325 placement.busy_placement = &placements;
326 placements.fpfn = 0;
327 placements.lpfn = 0;
328 placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
329 r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
330 interruptible, no_wait_gpu);
331 if (unlikely(r)) {
332 return r;
335 r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
336 if (unlikely(r)) {
337 goto out_cleanup;
340 r = ttm_tt_bind(bo->ttm, &tmp_mem);
341 if (unlikely(r)) {
342 goto out_cleanup;
344 r = radeon_move_blit(bo, true, no_wait_gpu, &tmp_mem, old_mem);
345 if (unlikely(r)) {
346 goto out_cleanup;
348 r = ttm_bo_move_ttm(bo, true, no_wait_gpu, new_mem);
349 out_cleanup:
350 ttm_bo_mem_put(bo, &tmp_mem);
351 return r;
354 static int radeon_move_ram_vram(struct ttm_buffer_object *bo,
355 bool evict, bool interruptible,
356 bool no_wait_gpu,
357 struct ttm_mem_reg *new_mem)
359 struct radeon_device *rdev;
360 struct ttm_mem_reg *old_mem = &bo->mem;
361 struct ttm_mem_reg tmp_mem;
362 struct ttm_placement placement;
363 struct ttm_place placements;
364 int r;
366 rdev = radeon_get_rdev(bo->bdev);
367 tmp_mem = *new_mem;
368 tmp_mem.mm_node = NULL;
369 placement.num_placement = 1;
370 placement.placement = &placements;
371 placement.num_busy_placement = 1;
372 placement.busy_placement = &placements;
373 placements.fpfn = 0;
374 placements.lpfn = 0;
375 placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
376 r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
377 interruptible, no_wait_gpu);
378 if (unlikely(r)) {
379 return r;
381 r = ttm_bo_move_ttm(bo, true, no_wait_gpu, &tmp_mem);
382 if (unlikely(r)) {
383 goto out_cleanup;
385 r = radeon_move_blit(bo, true, no_wait_gpu, new_mem, old_mem);
386 if (unlikely(r)) {
387 goto out_cleanup;
389 out_cleanup:
390 ttm_bo_mem_put(bo, &tmp_mem);
391 return r;
394 static int radeon_bo_move(struct ttm_buffer_object *bo,
395 bool evict, bool interruptible,
396 bool no_wait_gpu,
397 struct ttm_mem_reg *new_mem)
399 struct radeon_device *rdev;
400 struct ttm_mem_reg *old_mem = &bo->mem;
401 int r;
403 rdev = radeon_get_rdev(bo->bdev);
404 if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
405 radeon_move_null(bo, new_mem);
406 return 0;
408 if ((old_mem->mem_type == TTM_PL_TT &&
409 new_mem->mem_type == TTM_PL_SYSTEM) ||
410 (old_mem->mem_type == TTM_PL_SYSTEM &&
411 new_mem->mem_type == TTM_PL_TT)) {
412 /* bind is enough */
413 radeon_move_null(bo, new_mem);
414 return 0;
416 if (!rdev->ring[radeon_copy_ring_index(rdev)].ready ||
417 rdev->asic->copy.copy == NULL) {
418 /* use memcpy */
419 goto memcpy;
422 if (old_mem->mem_type == TTM_PL_VRAM &&
423 new_mem->mem_type == TTM_PL_SYSTEM) {
424 r = radeon_move_vram_ram(bo, evict, interruptible,
425 no_wait_gpu, new_mem);
426 } else if (old_mem->mem_type == TTM_PL_SYSTEM &&
427 new_mem->mem_type == TTM_PL_VRAM) {
428 r = radeon_move_ram_vram(bo, evict, interruptible,
429 no_wait_gpu, new_mem);
430 } else {
431 r = radeon_move_blit(bo, evict, no_wait_gpu, new_mem, old_mem);
434 if (r) {
435 memcpy:
436 r = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
437 if (r) {
438 return r;
442 /* update statistics */
443 atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &rdev->num_bytes_moved);
444 return 0;
447 static int radeon_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
449 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
450 struct radeon_device *rdev = radeon_get_rdev(bdev);
452 mem->bus.addr = NULL;
453 mem->bus.offset = 0;
454 mem->bus.size = mem->num_pages << PAGE_SHIFT;
455 mem->bus.base = 0;
456 mem->bus.is_iomem = false;
457 if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
458 return -EINVAL;
459 switch (mem->mem_type) {
460 case TTM_PL_SYSTEM:
461 /* system memory */
462 return 0;
463 case TTM_PL_TT:
464 #if __OS_HAS_AGP
465 if (rdev->flags & RADEON_IS_AGP) {
466 /* RADEON_IS_AGP is set only if AGP is active */
467 mem->bus.offset = mem->start << PAGE_SHIFT;
468 mem->bus.base = rdev->mc.agp_base;
469 mem->bus.is_iomem = !rdev->ddev->agp->cant_use_aperture;
471 #endif
472 break;
473 case TTM_PL_VRAM:
474 mem->bus.offset = mem->start << PAGE_SHIFT;
475 /* check if it's visible */
476 if ((mem->bus.offset + mem->bus.size) > rdev->mc.visible_vram_size)
477 return -EINVAL;
478 mem->bus.base = rdev->mc.aper_base;
479 mem->bus.is_iomem = true;
480 #ifdef __alpha__
482 * Alpha: use bus.addr to hold the ioremap() return,
483 * so we can modify bus.base below.
485 if (mem->placement & TTM_PL_FLAG_WC)
486 mem->bus.addr =
487 ioremap_wc(mem->bus.base + mem->bus.offset,
488 mem->bus.size);
489 else
490 mem->bus.addr =
491 ioremap_nocache(mem->bus.base + mem->bus.offset,
492 mem->bus.size);
495 * Alpha: Use just the bus offset plus
496 * the hose/domain memory base for bus.base.
497 * It then can be used to build PTEs for VRAM
498 * access, as done in ttm_bo_vm_fault().
500 mem->bus.base = (mem->bus.base & 0x0ffffffffUL) +
501 rdev->ddev->hose->dense_mem_base;
502 #endif
503 break;
504 default:
505 return -EINVAL;
507 return 0;
510 static void radeon_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
515 * TTM backend functions.
517 struct radeon_ttm_tt {
518 struct ttm_dma_tt ttm;
519 struct radeon_device *rdev;
520 u64 offset;
522 uint64_t userptr;
523 struct mm_struct *usermm;
524 uint32_t userflags;
527 /* prepare the sg table with the user pages */
528 static int radeon_ttm_tt_pin_userptr(struct ttm_tt *ttm)
530 struct radeon_device *rdev = radeon_get_rdev(ttm->bdev);
531 struct radeon_ttm_tt *gtt = (void *)ttm;
532 unsigned pinned = 0, nents;
533 int r;
535 int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
536 enum dma_data_direction direction = write ?
537 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
539 if (current->mm != gtt->usermm)
540 return -EPERM;
542 if (gtt->userflags & RADEON_GEM_USERPTR_ANONONLY) {
543 /* check that we only pin down anonymous memory
544 to prevent problems with writeback */
545 unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE;
546 struct vm_area_struct *vma;
547 vma = find_vma(gtt->usermm, gtt->userptr);
548 if (!vma || vma->vm_file || vma->vm_end < end)
549 return -EPERM;
552 do {
553 unsigned num_pages = ttm->num_pages - pinned;
554 uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
555 struct page **pages = ttm->pages + pinned;
557 r = get_user_pages(current, current->mm, userptr, num_pages,
558 write, 0, pages, NULL);
559 if (r < 0)
560 goto release_pages;
562 pinned += r;
564 } while (pinned < ttm->num_pages);
566 r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
567 ttm->num_pages << PAGE_SHIFT,
568 GFP_KERNEL);
569 if (r)
570 goto release_sg;
572 r = -ENOMEM;
573 nents = dma_map_sg(rdev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
574 if (nents != ttm->sg->nents)
575 goto release_sg;
577 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
578 gtt->ttm.dma_address, ttm->num_pages);
580 return 0;
582 release_sg:
583 kfree(ttm->sg);
585 release_pages:
586 release_pages(ttm->pages, pinned, 0);
587 return r;
590 static void radeon_ttm_tt_unpin_userptr(struct ttm_tt *ttm)
592 struct radeon_device *rdev = radeon_get_rdev(ttm->bdev);
593 struct radeon_ttm_tt *gtt = (void *)ttm;
594 struct sg_page_iter sg_iter;
596 int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
597 enum dma_data_direction direction = write ?
598 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
600 /* double check that we don't free the table twice */
601 if (!ttm->sg->sgl)
602 return;
604 /* free the sg table and pages again */
605 dma_unmap_sg(rdev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
607 for_each_sg_page(ttm->sg->sgl, &sg_iter, ttm->sg->nents, 0) {
608 struct page *page = sg_page_iter_page(&sg_iter);
609 if (!(gtt->userflags & RADEON_GEM_USERPTR_READONLY))
610 set_page_dirty(page);
612 mark_page_accessed(page);
613 page_cache_release(page);
616 sg_free_table(ttm->sg);
619 static int radeon_ttm_backend_bind(struct ttm_tt *ttm,
620 struct ttm_mem_reg *bo_mem)
622 struct radeon_ttm_tt *gtt = (void*)ttm;
623 uint32_t flags = RADEON_GART_PAGE_VALID | RADEON_GART_PAGE_READ |
624 RADEON_GART_PAGE_WRITE;
625 int r;
627 if (gtt->userptr) {
628 radeon_ttm_tt_pin_userptr(ttm);
629 flags &= ~RADEON_GART_PAGE_WRITE;
632 gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT);
633 if (!ttm->num_pages) {
634 WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
635 ttm->num_pages, bo_mem, ttm);
637 if (ttm->caching_state == tt_cached)
638 flags |= RADEON_GART_PAGE_SNOOP;
639 r = radeon_gart_bind(gtt->rdev, gtt->offset, ttm->num_pages,
640 ttm->pages, gtt->ttm.dma_address, flags);
641 if (r) {
642 DRM_ERROR("failed to bind %lu pages at 0x%08X\n",
643 ttm->num_pages, (unsigned)gtt->offset);
644 return r;
646 return 0;
649 static int radeon_ttm_backend_unbind(struct ttm_tt *ttm)
651 struct radeon_ttm_tt *gtt = (void *)ttm;
653 radeon_gart_unbind(gtt->rdev, gtt->offset, ttm->num_pages);
655 if (gtt->userptr)
656 radeon_ttm_tt_unpin_userptr(ttm);
658 return 0;
661 static void radeon_ttm_backend_destroy(struct ttm_tt *ttm)
663 struct radeon_ttm_tt *gtt = (void *)ttm;
665 ttm_dma_tt_fini(&gtt->ttm);
666 kfree(gtt);
669 static struct ttm_backend_func radeon_backend_func = {
670 .bind = &radeon_ttm_backend_bind,
671 .unbind = &radeon_ttm_backend_unbind,
672 .destroy = &radeon_ttm_backend_destroy,
675 static struct ttm_tt *radeon_ttm_tt_create(struct ttm_bo_device *bdev,
676 unsigned long size, uint32_t page_flags,
677 struct page *dummy_read_page)
679 struct radeon_device *rdev;
680 struct radeon_ttm_tt *gtt;
682 rdev = radeon_get_rdev(bdev);
683 #if __OS_HAS_AGP
684 if (rdev->flags & RADEON_IS_AGP) {
685 return ttm_agp_tt_create(bdev, rdev->ddev->agp->bridge,
686 size, page_flags, dummy_read_page);
688 #endif
690 gtt = kzalloc(sizeof(struct radeon_ttm_tt), GFP_KERNEL);
691 if (gtt == NULL) {
692 return NULL;
694 gtt->ttm.ttm.func = &radeon_backend_func;
695 gtt->rdev = rdev;
696 if (ttm_dma_tt_init(&gtt->ttm, bdev, size, page_flags, dummy_read_page)) {
697 kfree(gtt);
698 return NULL;
700 return &gtt->ttm.ttm;
703 static struct radeon_ttm_tt *radeon_ttm_tt_to_gtt(struct ttm_tt *ttm)
705 if (!ttm || ttm->func != &radeon_backend_func)
706 return NULL;
707 return (struct radeon_ttm_tt *)ttm;
710 static int radeon_ttm_tt_populate(struct ttm_tt *ttm)
712 struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
713 struct radeon_device *rdev;
714 unsigned i;
715 int r;
716 bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
718 if (ttm->state != tt_unpopulated)
719 return 0;
721 if (gtt && gtt->userptr) {
722 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
723 if (!ttm->sg)
724 return -ENOMEM;
726 ttm->page_flags |= TTM_PAGE_FLAG_SG;
727 ttm->state = tt_unbound;
728 return 0;
731 if (slave && ttm->sg) {
732 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
733 gtt->ttm.dma_address, ttm->num_pages);
734 ttm->state = tt_unbound;
735 return 0;
738 rdev = radeon_get_rdev(ttm->bdev);
739 #if __OS_HAS_AGP
740 if (rdev->flags & RADEON_IS_AGP) {
741 return ttm_agp_tt_populate(ttm);
743 #endif
745 #ifdef CONFIG_SWIOTLB
746 if (swiotlb_nr_tbl()) {
747 return ttm_dma_populate(&gtt->ttm, rdev->dev);
749 #endif
751 r = ttm_pool_populate(ttm);
752 if (r) {
753 return r;
756 for (i = 0; i < ttm->num_pages; i++) {
757 gtt->ttm.dma_address[i] = pci_map_page(rdev->pdev, ttm->pages[i],
758 0, PAGE_SIZE,
759 PCI_DMA_BIDIRECTIONAL);
760 if (pci_dma_mapping_error(rdev->pdev, gtt->ttm.dma_address[i])) {
761 while (--i) {
762 pci_unmap_page(rdev->pdev, gtt->ttm.dma_address[i],
763 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
764 gtt->ttm.dma_address[i] = 0;
766 ttm_pool_unpopulate(ttm);
767 return -EFAULT;
770 return 0;
773 static void radeon_ttm_tt_unpopulate(struct ttm_tt *ttm)
775 struct radeon_device *rdev;
776 struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
777 unsigned i;
778 bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
780 if (gtt && gtt->userptr) {
781 kfree(ttm->sg);
782 ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
783 return;
786 if (slave)
787 return;
789 rdev = radeon_get_rdev(ttm->bdev);
790 #if __OS_HAS_AGP
791 if (rdev->flags & RADEON_IS_AGP) {
792 ttm_agp_tt_unpopulate(ttm);
793 return;
795 #endif
797 #ifdef CONFIG_SWIOTLB
798 if (swiotlb_nr_tbl()) {
799 ttm_dma_unpopulate(&gtt->ttm, rdev->dev);
800 return;
802 #endif
804 for (i = 0; i < ttm->num_pages; i++) {
805 if (gtt->ttm.dma_address[i]) {
806 pci_unmap_page(rdev->pdev, gtt->ttm.dma_address[i],
807 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
811 ttm_pool_unpopulate(ttm);
814 int radeon_ttm_tt_set_userptr(struct ttm_tt *ttm, uint64_t addr,
815 uint32_t flags)
817 struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
819 if (gtt == NULL)
820 return -EINVAL;
822 gtt->userptr = addr;
823 gtt->usermm = current->mm;
824 gtt->userflags = flags;
825 return 0;
828 bool radeon_ttm_tt_has_userptr(struct ttm_tt *ttm)
830 struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
832 if (gtt == NULL)
833 return false;
835 return !!gtt->userptr;
838 bool radeon_ttm_tt_is_readonly(struct ttm_tt *ttm)
840 struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
842 if (gtt == NULL)
843 return false;
845 return !!(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
848 static struct ttm_bo_driver radeon_bo_driver = {
849 .ttm_tt_create = &radeon_ttm_tt_create,
850 .ttm_tt_populate = &radeon_ttm_tt_populate,
851 .ttm_tt_unpopulate = &radeon_ttm_tt_unpopulate,
852 .invalidate_caches = &radeon_invalidate_caches,
853 .init_mem_type = &radeon_init_mem_type,
854 .evict_flags = &radeon_evict_flags,
855 .move = &radeon_bo_move,
856 .verify_access = &radeon_verify_access,
857 .move_notify = &radeon_bo_move_notify,
858 .fault_reserve_notify = &radeon_bo_fault_reserve_notify,
859 .io_mem_reserve = &radeon_ttm_io_mem_reserve,
860 .io_mem_free = &radeon_ttm_io_mem_free,
863 int radeon_ttm_init(struct radeon_device *rdev)
865 int r;
867 r = radeon_ttm_global_init(rdev);
868 if (r) {
869 return r;
871 /* No others user of address space so set it to 0 */
872 r = ttm_bo_device_init(&rdev->mman.bdev,
873 rdev->mman.bo_global_ref.ref.object,
874 &radeon_bo_driver,
875 rdev->ddev->anon_inode->i_mapping,
876 DRM_FILE_PAGE_OFFSET,
877 rdev->need_dma32);
878 if (r) {
879 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
880 return r;
882 rdev->mman.initialized = true;
883 r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_VRAM,
884 rdev->mc.real_vram_size >> PAGE_SHIFT);
885 if (r) {
886 DRM_ERROR("Failed initializing VRAM heap.\n");
887 return r;
889 /* Change the size here instead of the init above so only lpfn is affected */
890 radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
892 r = radeon_bo_create(rdev, 256 * 1024, PAGE_SIZE, true,
893 RADEON_GEM_DOMAIN_VRAM, 0, NULL,
894 NULL, &rdev->stollen_vga_memory);
895 if (r) {
896 return r;
898 r = radeon_bo_reserve(rdev->stollen_vga_memory, false);
899 if (r)
900 return r;
901 r = radeon_bo_pin(rdev->stollen_vga_memory, RADEON_GEM_DOMAIN_VRAM, NULL);
902 radeon_bo_unreserve(rdev->stollen_vga_memory);
903 if (r) {
904 radeon_bo_unref(&rdev->stollen_vga_memory);
905 return r;
907 DRM_INFO("radeon: %uM of VRAM memory ready\n",
908 (unsigned) (rdev->mc.real_vram_size / (1024 * 1024)));
909 r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_TT,
910 rdev->mc.gtt_size >> PAGE_SHIFT);
911 if (r) {
912 DRM_ERROR("Failed initializing GTT heap.\n");
913 return r;
915 DRM_INFO("radeon: %uM of GTT memory ready.\n",
916 (unsigned)(rdev->mc.gtt_size / (1024 * 1024)));
918 r = radeon_ttm_debugfs_init(rdev);
919 if (r) {
920 DRM_ERROR("Failed to init debugfs\n");
921 return r;
923 return 0;
926 void radeon_ttm_fini(struct radeon_device *rdev)
928 int r;
930 if (!rdev->mman.initialized)
931 return;
932 radeon_ttm_debugfs_fini(rdev);
933 if (rdev->stollen_vga_memory) {
934 r = radeon_bo_reserve(rdev->stollen_vga_memory, false);
935 if (r == 0) {
936 radeon_bo_unpin(rdev->stollen_vga_memory);
937 radeon_bo_unreserve(rdev->stollen_vga_memory);
939 radeon_bo_unref(&rdev->stollen_vga_memory);
941 ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_VRAM);
942 ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_TT);
943 ttm_bo_device_release(&rdev->mman.bdev);
944 radeon_gart_fini(rdev);
945 radeon_ttm_global_fini(rdev);
946 rdev->mman.initialized = false;
947 DRM_INFO("radeon: ttm finalized\n");
950 /* this should only be called at bootup or when userspace
951 * isn't running */
952 void radeon_ttm_set_active_vram_size(struct radeon_device *rdev, u64 size)
954 struct ttm_mem_type_manager *man;
956 if (!rdev->mman.initialized)
957 return;
959 man = &rdev->mman.bdev.man[TTM_PL_VRAM];
960 /* this just adjusts TTM size idea, which sets lpfn to the correct value */
961 man->size = size >> PAGE_SHIFT;
964 static struct vm_operations_struct radeon_ttm_vm_ops;
965 static const struct vm_operations_struct *ttm_vm_ops = NULL;
967 static int radeon_ttm_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
969 struct ttm_buffer_object *bo;
970 struct radeon_device *rdev;
971 int r;
973 bo = (struct ttm_buffer_object *)vma->vm_private_data;
974 if (bo == NULL) {
975 return VM_FAULT_NOPAGE;
977 rdev = radeon_get_rdev(bo->bdev);
978 down_read(&rdev->pm.mclk_lock);
979 r = ttm_vm_ops->fault(vma, vmf);
980 up_read(&rdev->pm.mclk_lock);
981 return r;
984 int radeon_mmap(struct file *filp, struct vm_area_struct *vma)
986 struct drm_file *file_priv;
987 struct radeon_device *rdev;
988 int r;
990 if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET)) {
991 return -EINVAL;
994 file_priv = filp->private_data;
995 rdev = file_priv->minor->dev->dev_private;
996 if (rdev == NULL) {
997 return -EINVAL;
999 r = ttm_bo_mmap(filp, vma, &rdev->mman.bdev);
1000 if (unlikely(r != 0)) {
1001 return r;
1003 if (unlikely(ttm_vm_ops == NULL)) {
1004 ttm_vm_ops = vma->vm_ops;
1005 radeon_ttm_vm_ops = *ttm_vm_ops;
1006 radeon_ttm_vm_ops.fault = &radeon_ttm_fault;
1008 vma->vm_ops = &radeon_ttm_vm_ops;
1009 return 0;
1012 #if defined(CONFIG_DEBUG_FS)
1014 static int radeon_mm_dump_table(struct seq_file *m, void *data)
1016 struct drm_info_node *node = (struct drm_info_node *)m->private;
1017 unsigned ttm_pl = *(int *)node->info_ent->data;
1018 struct drm_device *dev = node->minor->dev;
1019 struct radeon_device *rdev = dev->dev_private;
1020 struct drm_mm *mm = (struct drm_mm *)rdev->mman.bdev.man[ttm_pl].priv;
1021 int ret;
1022 struct ttm_bo_global *glob = rdev->mman.bdev.glob;
1024 spin_lock(&glob->lru_lock);
1025 ret = drm_mm_dump_table(m, mm);
1026 spin_unlock(&glob->lru_lock);
1027 return ret;
1030 static int ttm_pl_vram = TTM_PL_VRAM;
1031 static int ttm_pl_tt = TTM_PL_TT;
1033 static struct drm_info_list radeon_ttm_debugfs_list[] = {
1034 {"radeon_vram_mm", radeon_mm_dump_table, 0, &ttm_pl_vram},
1035 {"radeon_gtt_mm", radeon_mm_dump_table, 0, &ttm_pl_tt},
1036 {"ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL},
1037 #ifdef CONFIG_SWIOTLB
1038 {"ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL}
1039 #endif
1042 static int radeon_ttm_vram_open(struct inode *inode, struct file *filep)
1044 struct radeon_device *rdev = inode->i_private;
1045 i_size_write(inode, rdev->mc.mc_vram_size);
1046 filep->private_data = inode->i_private;
1047 return 0;
1050 static ssize_t radeon_ttm_vram_read(struct file *f, char __user *buf,
1051 size_t size, loff_t *pos)
1053 struct radeon_device *rdev = f->private_data;
1054 ssize_t result = 0;
1055 int r;
1057 if (size & 0x3 || *pos & 0x3)
1058 return -EINVAL;
1060 while (size) {
1061 unsigned long flags;
1062 uint32_t value;
1064 if (*pos >= rdev->mc.mc_vram_size)
1065 return result;
1067 spin_lock_irqsave(&rdev->mmio_idx_lock, flags);
1068 WREG32(RADEON_MM_INDEX, ((uint32_t)*pos) | 0x80000000);
1069 if (rdev->family >= CHIP_CEDAR)
1070 WREG32(EVERGREEN_MM_INDEX_HI, *pos >> 31);
1071 value = RREG32(RADEON_MM_DATA);
1072 spin_unlock_irqrestore(&rdev->mmio_idx_lock, flags);
1074 r = put_user(value, (uint32_t *)buf);
1075 if (r)
1076 return r;
1078 result += 4;
1079 buf += 4;
1080 *pos += 4;
1081 size -= 4;
1084 return result;
1087 static const struct file_operations radeon_ttm_vram_fops = {
1088 .owner = THIS_MODULE,
1089 .open = radeon_ttm_vram_open,
1090 .read = radeon_ttm_vram_read,
1091 .llseek = default_llseek
1094 static int radeon_ttm_gtt_open(struct inode *inode, struct file *filep)
1096 struct radeon_device *rdev = inode->i_private;
1097 i_size_write(inode, rdev->mc.gtt_size);
1098 filep->private_data = inode->i_private;
1099 return 0;
1102 static ssize_t radeon_ttm_gtt_read(struct file *f, char __user *buf,
1103 size_t size, loff_t *pos)
1105 struct radeon_device *rdev = f->private_data;
1106 ssize_t result = 0;
1107 int r;
1109 while (size) {
1110 loff_t p = *pos / PAGE_SIZE;
1111 unsigned off = *pos & ~PAGE_MASK;
1112 size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
1113 struct page *page;
1114 void *ptr;
1116 if (p >= rdev->gart.num_cpu_pages)
1117 return result;
1119 page = rdev->gart.pages[p];
1120 if (page) {
1121 ptr = kmap(page);
1122 ptr += off;
1124 r = copy_to_user(buf, ptr, cur_size);
1125 kunmap(rdev->gart.pages[p]);
1126 } else
1127 r = clear_user(buf, cur_size);
1129 if (r)
1130 return -EFAULT;
1132 result += cur_size;
1133 buf += cur_size;
1134 *pos += cur_size;
1135 size -= cur_size;
1138 return result;
1141 static const struct file_operations radeon_ttm_gtt_fops = {
1142 .owner = THIS_MODULE,
1143 .open = radeon_ttm_gtt_open,
1144 .read = radeon_ttm_gtt_read,
1145 .llseek = default_llseek
1148 #endif
1150 static int radeon_ttm_debugfs_init(struct radeon_device *rdev)
1152 #if defined(CONFIG_DEBUG_FS)
1153 unsigned count;
1155 struct drm_minor *minor = rdev->ddev->primary;
1156 struct dentry *ent, *root = minor->debugfs_root;
1158 ent = debugfs_create_file("radeon_vram", S_IFREG | S_IRUGO, root,
1159 rdev, &radeon_ttm_vram_fops);
1160 if (IS_ERR(ent))
1161 return PTR_ERR(ent);
1162 rdev->mman.vram = ent;
1164 ent = debugfs_create_file("radeon_gtt", S_IFREG | S_IRUGO, root,
1165 rdev, &radeon_ttm_gtt_fops);
1166 if (IS_ERR(ent))
1167 return PTR_ERR(ent);
1168 rdev->mman.gtt = ent;
1170 count = ARRAY_SIZE(radeon_ttm_debugfs_list);
1172 #ifdef CONFIG_SWIOTLB
1173 if (!swiotlb_nr_tbl())
1174 --count;
1175 #endif
1177 return radeon_debugfs_add_files(rdev, radeon_ttm_debugfs_list, count);
1178 #else
1180 return 0;
1181 #endif
1184 static void radeon_ttm_debugfs_fini(struct radeon_device *rdev)
1186 #if defined(CONFIG_DEBUG_FS)
1188 debugfs_remove(rdev->mman.vram);
1189 rdev->mman.vram = NULL;
1191 debugfs_remove(rdev->mman.gtt);
1192 rdev->mman.gtt = NULL;
1193 #endif