2 * Copyright 2009 Jerome Glisse.
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
28 * Jerome Glisse <glisse@freedesktop.org>
29 * Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
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>
38 #include <drm/amdgpu_drm.h>
39 #include <linux/seq_file.h>
40 #include <linux/slab.h>
41 #include <linux/swiotlb.h>
42 #include <linux/swap.h>
43 #include <linux/pagemap.h>
44 #include <linux/debugfs.h>
46 #include "bif/bif_4_1_d.h"
48 #define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT)
50 static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev);
51 static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev);
53 static struct amdgpu_device *amdgpu_get_adev(struct ttm_bo_device *bdev)
55 struct amdgpu_mman *mman;
56 struct amdgpu_device *adev;
58 mman = container_of(bdev, struct amdgpu_mman, bdev);
59 adev = container_of(mman, struct amdgpu_device, mman);
67 static int amdgpu_ttm_mem_global_init(struct drm_global_reference *ref)
69 return ttm_mem_global_init(ref->object);
72 static void amdgpu_ttm_mem_global_release(struct drm_global_reference *ref)
74 ttm_mem_global_release(ref->object);
77 static int amdgpu_ttm_global_init(struct amdgpu_device *adev)
79 struct drm_global_reference *global_ref;
80 struct amdgpu_ring *ring;
81 struct amd_sched_rq *rq;
84 adev->mman.mem_global_referenced = false;
85 global_ref = &adev->mman.mem_global_ref;
86 global_ref->global_type = DRM_GLOBAL_TTM_MEM;
87 global_ref->size = sizeof(struct ttm_mem_global);
88 global_ref->init = &amdgpu_ttm_mem_global_init;
89 global_ref->release = &amdgpu_ttm_mem_global_release;
90 r = drm_global_item_ref(global_ref);
92 DRM_ERROR("Failed setting up TTM memory accounting "
97 adev->mman.bo_global_ref.mem_glob =
98 adev->mman.mem_global_ref.object;
99 global_ref = &adev->mman.bo_global_ref.ref;
100 global_ref->global_type = DRM_GLOBAL_TTM_BO;
101 global_ref->size = sizeof(struct ttm_bo_global);
102 global_ref->init = &ttm_bo_global_init;
103 global_ref->release = &ttm_bo_global_release;
104 r = drm_global_item_ref(global_ref);
106 DRM_ERROR("Failed setting up TTM BO subsystem.\n");
107 drm_global_item_unref(&adev->mman.mem_global_ref);
111 ring = adev->mman.buffer_funcs_ring;
112 rq = &ring->sched.sched_rq[AMD_SCHED_PRIORITY_KERNEL];
113 r = amd_sched_entity_init(&ring->sched, &adev->mman.entity,
114 rq, amdgpu_sched_jobs);
116 DRM_ERROR("Failed setting up TTM BO move run queue.\n");
117 drm_global_item_unref(&adev->mman.mem_global_ref);
118 drm_global_item_unref(&adev->mman.bo_global_ref.ref);
122 adev->mman.mem_global_referenced = true;
127 static void amdgpu_ttm_global_fini(struct amdgpu_device *adev)
129 if (adev->mman.mem_global_referenced) {
130 amd_sched_entity_fini(adev->mman.entity.sched,
132 drm_global_item_unref(&adev->mman.bo_global_ref.ref);
133 drm_global_item_unref(&adev->mman.mem_global_ref);
134 adev->mman.mem_global_referenced = false;
138 static int amdgpu_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
143 static int amdgpu_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
144 struct ttm_mem_type_manager *man)
146 struct amdgpu_device *adev;
148 adev = amdgpu_get_adev(bdev);
153 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
154 man->available_caching = TTM_PL_MASK_CACHING;
155 man->default_caching = TTM_PL_FLAG_CACHED;
158 man->func = &ttm_bo_manager_func;
159 man->gpu_offset = adev->mc.gtt_start;
160 man->available_caching = TTM_PL_MASK_CACHING;
161 man->default_caching = TTM_PL_FLAG_CACHED;
162 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE | TTM_MEMTYPE_FLAG_CMA;
165 /* "On-card" video ram */
166 man->func = &ttm_bo_manager_func;
167 man->gpu_offset = adev->mc.vram_start;
168 man->flags = TTM_MEMTYPE_FLAG_FIXED |
169 TTM_MEMTYPE_FLAG_MAPPABLE;
170 man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
171 man->default_caching = TTM_PL_FLAG_WC;
176 /* On-chip GDS memory*/
177 man->func = &ttm_bo_manager_func;
179 man->flags = TTM_MEMTYPE_FLAG_FIXED | TTM_MEMTYPE_FLAG_CMA;
180 man->available_caching = TTM_PL_FLAG_UNCACHED;
181 man->default_caching = TTM_PL_FLAG_UNCACHED;
184 DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
190 static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
191 struct ttm_placement *placement)
193 struct amdgpu_bo *rbo;
194 static struct ttm_place placements = {
197 .flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
200 if (!amdgpu_ttm_bo_is_amdgpu_bo(bo)) {
201 placement->placement = &placements;
202 placement->busy_placement = &placements;
203 placement->num_placement = 1;
204 placement->num_busy_placement = 1;
207 rbo = container_of(bo, struct amdgpu_bo, tbo);
208 switch (bo->mem.mem_type) {
210 if (rbo->adev->mman.buffer_funcs_ring->ready == false)
211 amdgpu_ttm_placement_from_domain(rbo, AMDGPU_GEM_DOMAIN_CPU);
213 amdgpu_ttm_placement_from_domain(rbo, AMDGPU_GEM_DOMAIN_GTT);
217 amdgpu_ttm_placement_from_domain(rbo, AMDGPU_GEM_DOMAIN_CPU);
219 *placement = rbo->placement;
222 static int amdgpu_verify_access(struct ttm_buffer_object *bo, struct file *filp)
224 struct amdgpu_bo *rbo = container_of(bo, struct amdgpu_bo, tbo);
226 if (amdgpu_ttm_tt_get_usermm(bo->ttm))
228 return drm_vma_node_verify_access(&rbo->gem_base.vma_node, filp);
231 static void amdgpu_move_null(struct ttm_buffer_object *bo,
232 struct ttm_mem_reg *new_mem)
234 struct ttm_mem_reg *old_mem = &bo->mem;
236 BUG_ON(old_mem->mm_node != NULL);
238 new_mem->mm_node = NULL;
241 static int amdgpu_move_blit(struct ttm_buffer_object *bo,
242 bool evict, bool no_wait_gpu,
243 struct ttm_mem_reg *new_mem,
244 struct ttm_mem_reg *old_mem)
246 struct amdgpu_device *adev;
247 struct amdgpu_ring *ring;
248 uint64_t old_start, new_start;
252 adev = amdgpu_get_adev(bo->bdev);
253 ring = adev->mman.buffer_funcs_ring;
254 old_start = old_mem->start << PAGE_SHIFT;
255 new_start = new_mem->start << PAGE_SHIFT;
257 switch (old_mem->mem_type) {
259 old_start += adev->mc.vram_start;
262 old_start += adev->mc.gtt_start;
265 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
268 switch (new_mem->mem_type) {
270 new_start += adev->mc.vram_start;
273 new_start += adev->mc.gtt_start;
276 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
280 DRM_ERROR("Trying to move memory with ring turned off.\n");
284 BUILD_BUG_ON((PAGE_SIZE % AMDGPU_GPU_PAGE_SIZE) != 0);
286 r = amdgpu_copy_buffer(ring, old_start, new_start,
287 new_mem->num_pages * PAGE_SIZE, /* bytes */
289 /* FIXME: handle copy error */
290 r = ttm_bo_move_accel_cleanup(bo, fence,
291 evict, no_wait_gpu, new_mem);
296 static int amdgpu_move_vram_ram(struct ttm_buffer_object *bo,
297 bool evict, bool interruptible,
299 struct ttm_mem_reg *new_mem)
301 struct amdgpu_device *adev;
302 struct ttm_mem_reg *old_mem = &bo->mem;
303 struct ttm_mem_reg tmp_mem;
304 struct ttm_place placements;
305 struct ttm_placement placement;
308 adev = amdgpu_get_adev(bo->bdev);
310 tmp_mem.mm_node = NULL;
311 placement.num_placement = 1;
312 placement.placement = &placements;
313 placement.num_busy_placement = 1;
314 placement.busy_placement = &placements;
317 placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
318 r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
319 interruptible, no_wait_gpu);
324 r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
329 r = ttm_tt_bind(bo->ttm, &tmp_mem);
333 r = amdgpu_move_blit(bo, true, no_wait_gpu, &tmp_mem, old_mem);
337 r = ttm_bo_move_ttm(bo, true, no_wait_gpu, new_mem);
339 ttm_bo_mem_put(bo, &tmp_mem);
343 static int amdgpu_move_ram_vram(struct ttm_buffer_object *bo,
344 bool evict, bool interruptible,
346 struct ttm_mem_reg *new_mem)
348 struct amdgpu_device *adev;
349 struct ttm_mem_reg *old_mem = &bo->mem;
350 struct ttm_mem_reg tmp_mem;
351 struct ttm_placement placement;
352 struct ttm_place placements;
355 adev = amdgpu_get_adev(bo->bdev);
357 tmp_mem.mm_node = NULL;
358 placement.num_placement = 1;
359 placement.placement = &placements;
360 placement.num_busy_placement = 1;
361 placement.busy_placement = &placements;
364 placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
365 r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
366 interruptible, no_wait_gpu);
370 r = ttm_bo_move_ttm(bo, true, no_wait_gpu, &tmp_mem);
374 r = amdgpu_move_blit(bo, true, no_wait_gpu, new_mem, old_mem);
379 ttm_bo_mem_put(bo, &tmp_mem);
383 static int amdgpu_bo_move(struct ttm_buffer_object *bo,
384 bool evict, bool interruptible,
386 struct ttm_mem_reg *new_mem)
388 struct amdgpu_device *adev;
389 struct amdgpu_bo *abo;
390 struct ttm_mem_reg *old_mem = &bo->mem;
393 /* Can't move a pinned BO */
394 abo = container_of(bo, struct amdgpu_bo, tbo);
395 if (WARN_ON_ONCE(abo->pin_count > 0))
398 adev = amdgpu_get_adev(bo->bdev);
399 if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
400 amdgpu_move_null(bo, new_mem);
403 if ((old_mem->mem_type == TTM_PL_TT &&
404 new_mem->mem_type == TTM_PL_SYSTEM) ||
405 (old_mem->mem_type == TTM_PL_SYSTEM &&
406 new_mem->mem_type == TTM_PL_TT)) {
408 amdgpu_move_null(bo, new_mem);
411 if (adev->mman.buffer_funcs == NULL ||
412 adev->mman.buffer_funcs_ring == NULL ||
413 !adev->mman.buffer_funcs_ring->ready) {
418 if (old_mem->mem_type == TTM_PL_VRAM &&
419 new_mem->mem_type == TTM_PL_SYSTEM) {
420 r = amdgpu_move_vram_ram(bo, evict, interruptible,
421 no_wait_gpu, new_mem);
422 } else if (old_mem->mem_type == TTM_PL_SYSTEM &&
423 new_mem->mem_type == TTM_PL_VRAM) {
424 r = amdgpu_move_ram_vram(bo, evict, interruptible,
425 no_wait_gpu, new_mem);
427 r = amdgpu_move_blit(bo, evict, no_wait_gpu, new_mem, old_mem);
432 r = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
438 /* update statistics */
439 atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &adev->num_bytes_moved);
443 static int amdgpu_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
445 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
446 struct amdgpu_device *adev = amdgpu_get_adev(bdev);
448 mem->bus.addr = NULL;
450 mem->bus.size = mem->num_pages << PAGE_SHIFT;
452 mem->bus.is_iomem = false;
453 if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
455 switch (mem->mem_type) {
462 mem->bus.offset = mem->start << PAGE_SHIFT;
463 /* check if it's visible */
464 if ((mem->bus.offset + mem->bus.size) > adev->mc.visible_vram_size)
466 mem->bus.base = adev->mc.aper_base;
467 mem->bus.is_iomem = true;
470 * Alpha: use bus.addr to hold the ioremap() return,
471 * so we can modify bus.base below.
473 if (mem->placement & TTM_PL_FLAG_WC)
475 ioremap_wc(mem->bus.base + mem->bus.offset,
479 ioremap_nocache(mem->bus.base + mem->bus.offset,
483 * Alpha: Use just the bus offset plus
484 * the hose/domain memory base for bus.base.
485 * It then can be used to build PTEs for VRAM
486 * access, as done in ttm_bo_vm_fault().
488 mem->bus.base = (mem->bus.base & 0x0ffffffffUL) +
489 adev->ddev->hose->dense_mem_base;
498 static void amdgpu_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
503 * TTM backend functions.
505 struct amdgpu_ttm_gup_task_list {
506 struct list_head list;
507 struct task_struct *task;
510 struct amdgpu_ttm_tt {
511 struct ttm_dma_tt ttm;
512 struct amdgpu_device *adev;
515 struct mm_struct *usermm;
517 spinlock_t guptasklock;
518 struct list_head guptasks;
519 atomic_t mmu_invalidations;
522 int amdgpu_ttm_tt_get_user_pages(struct ttm_tt *ttm, struct page **pages)
524 struct amdgpu_ttm_tt *gtt = (void *)ttm;
525 int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
529 if (gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) {
530 /* check that we only use anonymous memory
531 to prevent problems with writeback */
532 unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE;
533 struct vm_area_struct *vma;
535 vma = find_vma(gtt->usermm, gtt->userptr);
536 if (!vma || vma->vm_file || vma->vm_end < end)
541 unsigned num_pages = ttm->num_pages - pinned;
542 uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
543 struct page **p = pages + pinned;
544 struct amdgpu_ttm_gup_task_list guptask;
546 guptask.task = current;
547 spin_lock(>t->guptasklock);
548 list_add(&guptask.list, >t->guptasks);
549 spin_unlock(>t->guptasklock);
551 r = get_user_pages(userptr, num_pages, write, 0, p, NULL);
553 spin_lock(>t->guptasklock);
554 list_del(&guptask.list);
555 spin_unlock(>t->guptasklock);
562 } while (pinned < ttm->num_pages);
567 release_pages(pages, pinned, 0);
571 /* prepare the sg table with the user pages */
572 static int amdgpu_ttm_tt_pin_userptr(struct ttm_tt *ttm)
574 struct amdgpu_device *adev = amdgpu_get_adev(ttm->bdev);
575 struct amdgpu_ttm_tt *gtt = (void *)ttm;
579 int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
580 enum dma_data_direction direction = write ?
581 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
583 r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
584 ttm->num_pages << PAGE_SHIFT,
590 nents = dma_map_sg(adev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
591 if (nents != ttm->sg->nents)
594 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
595 gtt->ttm.dma_address, ttm->num_pages);
604 static void amdgpu_ttm_tt_unpin_userptr(struct ttm_tt *ttm)
606 struct amdgpu_device *adev = amdgpu_get_adev(ttm->bdev);
607 struct amdgpu_ttm_tt *gtt = (void *)ttm;
608 struct sg_page_iter sg_iter;
610 int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
611 enum dma_data_direction direction = write ?
612 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
614 /* double check that we don't free the table twice */
618 /* free the sg table and pages again */
619 dma_unmap_sg(adev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
621 for_each_sg_page(ttm->sg->sgl, &sg_iter, ttm->sg->nents, 0) {
622 struct page *page = sg_page_iter_page(&sg_iter);
623 if (!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY))
624 set_page_dirty(page);
626 mark_page_accessed(page);
630 sg_free_table(ttm->sg);
633 static int amdgpu_ttm_backend_bind(struct ttm_tt *ttm,
634 struct ttm_mem_reg *bo_mem)
636 struct amdgpu_ttm_tt *gtt = (void*)ttm;
637 uint32_t flags = amdgpu_ttm_tt_pte_flags(gtt->adev, ttm, bo_mem);
641 r = amdgpu_ttm_tt_pin_userptr(ttm);
643 DRM_ERROR("failed to pin userptr\n");
647 gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT);
648 if (!ttm->num_pages) {
649 WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
650 ttm->num_pages, bo_mem, ttm);
653 if (bo_mem->mem_type == AMDGPU_PL_GDS ||
654 bo_mem->mem_type == AMDGPU_PL_GWS ||
655 bo_mem->mem_type == AMDGPU_PL_OA)
658 r = amdgpu_gart_bind(gtt->adev, gtt->offset, ttm->num_pages,
659 ttm->pages, gtt->ttm.dma_address, flags);
662 DRM_ERROR("failed to bind %lu pages at 0x%08X\n",
663 ttm->num_pages, (unsigned)gtt->offset);
669 static int amdgpu_ttm_backend_unbind(struct ttm_tt *ttm)
671 struct amdgpu_ttm_tt *gtt = (void *)ttm;
673 /* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */
674 if (gtt->adev->gart.ready)
675 amdgpu_gart_unbind(gtt->adev, gtt->offset, ttm->num_pages);
678 amdgpu_ttm_tt_unpin_userptr(ttm);
683 static void amdgpu_ttm_backend_destroy(struct ttm_tt *ttm)
685 struct amdgpu_ttm_tt *gtt = (void *)ttm;
687 ttm_dma_tt_fini(>t->ttm);
691 static struct ttm_backend_func amdgpu_backend_func = {
692 .bind = &amdgpu_ttm_backend_bind,
693 .unbind = &amdgpu_ttm_backend_unbind,
694 .destroy = &amdgpu_ttm_backend_destroy,
697 static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_bo_device *bdev,
698 unsigned long size, uint32_t page_flags,
699 struct page *dummy_read_page)
701 struct amdgpu_device *adev;
702 struct amdgpu_ttm_tt *gtt;
704 adev = amdgpu_get_adev(bdev);
706 gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL);
710 gtt->ttm.ttm.func = &amdgpu_backend_func;
712 if (ttm_dma_tt_init(>t->ttm, bdev, size, page_flags, dummy_read_page)) {
716 return >t->ttm.ttm;
719 static int amdgpu_ttm_tt_populate(struct ttm_tt *ttm)
721 struct amdgpu_device *adev;
722 struct amdgpu_ttm_tt *gtt = (void *)ttm;
725 bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
727 if (ttm->state != tt_unpopulated)
730 if (gtt && gtt->userptr) {
731 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
735 ttm->page_flags |= TTM_PAGE_FLAG_SG;
736 ttm->state = tt_unbound;
740 if (slave && ttm->sg) {
741 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
742 gtt->ttm.dma_address, ttm->num_pages);
743 ttm->state = tt_unbound;
747 adev = amdgpu_get_adev(ttm->bdev);
749 #ifdef CONFIG_SWIOTLB
750 if (swiotlb_nr_tbl()) {
751 return ttm_dma_populate(>t->ttm, adev->dev);
755 r = ttm_pool_populate(ttm);
760 for (i = 0; i < ttm->num_pages; i++) {
761 gtt->ttm.dma_address[i] = pci_map_page(adev->pdev, ttm->pages[i],
763 PCI_DMA_BIDIRECTIONAL);
764 if (pci_dma_mapping_error(adev->pdev, gtt->ttm.dma_address[i])) {
766 pci_unmap_page(adev->pdev, gtt->ttm.dma_address[i],
767 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
768 gtt->ttm.dma_address[i] = 0;
770 ttm_pool_unpopulate(ttm);
777 static void amdgpu_ttm_tt_unpopulate(struct ttm_tt *ttm)
779 struct amdgpu_device *adev;
780 struct amdgpu_ttm_tt *gtt = (void *)ttm;
782 bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
784 if (gtt && gtt->userptr) {
786 ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
793 adev = amdgpu_get_adev(ttm->bdev);
795 #ifdef CONFIG_SWIOTLB
796 if (swiotlb_nr_tbl()) {
797 ttm_dma_unpopulate(>t->ttm, adev->dev);
802 for (i = 0; i < ttm->num_pages; i++) {
803 if (gtt->ttm.dma_address[i]) {
804 pci_unmap_page(adev->pdev, gtt->ttm.dma_address[i],
805 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
809 ttm_pool_unpopulate(ttm);
812 int amdgpu_ttm_tt_set_userptr(struct ttm_tt *ttm, uint64_t addr,
815 struct amdgpu_ttm_tt *gtt = (void *)ttm;
821 gtt->usermm = current->mm;
822 gtt->userflags = flags;
823 spin_lock_init(>t->guptasklock);
824 INIT_LIST_HEAD(>t->guptasks);
825 atomic_set(>t->mmu_invalidations, 0);
830 struct mm_struct *amdgpu_ttm_tt_get_usermm(struct ttm_tt *ttm)
832 struct amdgpu_ttm_tt *gtt = (void *)ttm;
840 bool amdgpu_ttm_tt_affect_userptr(struct ttm_tt *ttm, unsigned long start,
843 struct amdgpu_ttm_tt *gtt = (void *)ttm;
844 struct amdgpu_ttm_gup_task_list *entry;
847 if (gtt == NULL || !gtt->userptr)
850 size = (unsigned long)gtt->ttm.ttm.num_pages * PAGE_SIZE;
851 if (gtt->userptr > end || gtt->userptr + size <= start)
854 spin_lock(>t->guptasklock);
855 list_for_each_entry(entry, >t->guptasks, list) {
856 if (entry->task == current) {
857 spin_unlock(>t->guptasklock);
861 spin_unlock(>t->guptasklock);
863 atomic_inc(>t->mmu_invalidations);
868 bool amdgpu_ttm_tt_userptr_invalidated(struct ttm_tt *ttm,
869 int *last_invalidated)
871 struct amdgpu_ttm_tt *gtt = (void *)ttm;
872 int prev_invalidated = *last_invalidated;
874 *last_invalidated = atomic_read(>t->mmu_invalidations);
875 return prev_invalidated != *last_invalidated;
878 bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm)
880 struct amdgpu_ttm_tt *gtt = (void *)ttm;
885 return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
888 uint32_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
889 struct ttm_mem_reg *mem)
893 if (mem && mem->mem_type != TTM_PL_SYSTEM)
894 flags |= AMDGPU_PTE_VALID;
896 if (mem && mem->mem_type == TTM_PL_TT) {
897 flags |= AMDGPU_PTE_SYSTEM;
899 if (ttm->caching_state == tt_cached)
900 flags |= AMDGPU_PTE_SNOOPED;
903 if (adev->asic_type >= CHIP_TONGA)
904 flags |= AMDGPU_PTE_EXECUTABLE;
906 flags |= AMDGPU_PTE_READABLE;
908 if (!amdgpu_ttm_tt_is_readonly(ttm))
909 flags |= AMDGPU_PTE_WRITEABLE;
914 static struct ttm_bo_driver amdgpu_bo_driver = {
915 .ttm_tt_create = &amdgpu_ttm_tt_create,
916 .ttm_tt_populate = &amdgpu_ttm_tt_populate,
917 .ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate,
918 .invalidate_caches = &amdgpu_invalidate_caches,
919 .init_mem_type = &amdgpu_init_mem_type,
920 .evict_flags = &amdgpu_evict_flags,
921 .move = &amdgpu_bo_move,
922 .verify_access = &amdgpu_verify_access,
923 .move_notify = &amdgpu_bo_move_notify,
924 .fault_reserve_notify = &amdgpu_bo_fault_reserve_notify,
925 .io_mem_reserve = &amdgpu_ttm_io_mem_reserve,
926 .io_mem_free = &amdgpu_ttm_io_mem_free,
929 int amdgpu_ttm_init(struct amdgpu_device *adev)
933 r = amdgpu_ttm_global_init(adev);
937 /* No others user of address space so set it to 0 */
938 r = ttm_bo_device_init(&adev->mman.bdev,
939 adev->mman.bo_global_ref.ref.object,
941 adev->ddev->anon_inode->i_mapping,
942 DRM_FILE_PAGE_OFFSET,
945 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
948 adev->mman.initialized = true;
949 r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_VRAM,
950 adev->mc.real_vram_size >> PAGE_SHIFT);
952 DRM_ERROR("Failed initializing VRAM heap.\n");
955 /* Change the size here instead of the init above so only lpfn is affected */
956 amdgpu_ttm_set_active_vram_size(adev, adev->mc.visible_vram_size);
958 r = amdgpu_bo_create(adev, 256 * 1024, PAGE_SIZE, true,
959 AMDGPU_GEM_DOMAIN_VRAM,
960 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
961 NULL, NULL, &adev->stollen_vga_memory);
965 r = amdgpu_bo_reserve(adev->stollen_vga_memory, false);
968 r = amdgpu_bo_pin(adev->stollen_vga_memory, AMDGPU_GEM_DOMAIN_VRAM, NULL);
969 amdgpu_bo_unreserve(adev->stollen_vga_memory);
971 amdgpu_bo_unref(&adev->stollen_vga_memory);
974 DRM_INFO("amdgpu: %uM of VRAM memory ready\n",
975 (unsigned) (adev->mc.real_vram_size / (1024 * 1024)));
976 r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_TT,
977 adev->mc.gtt_size >> PAGE_SHIFT);
979 DRM_ERROR("Failed initializing GTT heap.\n");
982 DRM_INFO("amdgpu: %uM of GTT memory ready.\n",
983 (unsigned)(adev->mc.gtt_size / (1024 * 1024)));
985 adev->gds.mem.total_size = adev->gds.mem.total_size << AMDGPU_GDS_SHIFT;
986 adev->gds.mem.gfx_partition_size = adev->gds.mem.gfx_partition_size << AMDGPU_GDS_SHIFT;
987 adev->gds.mem.cs_partition_size = adev->gds.mem.cs_partition_size << AMDGPU_GDS_SHIFT;
988 adev->gds.gws.total_size = adev->gds.gws.total_size << AMDGPU_GWS_SHIFT;
989 adev->gds.gws.gfx_partition_size = adev->gds.gws.gfx_partition_size << AMDGPU_GWS_SHIFT;
990 adev->gds.gws.cs_partition_size = adev->gds.gws.cs_partition_size << AMDGPU_GWS_SHIFT;
991 adev->gds.oa.total_size = adev->gds.oa.total_size << AMDGPU_OA_SHIFT;
992 adev->gds.oa.gfx_partition_size = adev->gds.oa.gfx_partition_size << AMDGPU_OA_SHIFT;
993 adev->gds.oa.cs_partition_size = adev->gds.oa.cs_partition_size << AMDGPU_OA_SHIFT;
995 r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_GDS,
996 adev->gds.mem.total_size >> PAGE_SHIFT);
998 DRM_ERROR("Failed initializing GDS heap.\n");
1003 r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_GWS,
1004 adev->gds.gws.total_size >> PAGE_SHIFT);
1006 DRM_ERROR("Failed initializing gws heap.\n");
1011 r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_OA,
1012 adev->gds.oa.total_size >> PAGE_SHIFT);
1014 DRM_ERROR("Failed initializing oa heap.\n");
1018 r = amdgpu_ttm_debugfs_init(adev);
1020 DRM_ERROR("Failed to init debugfs\n");
1026 void amdgpu_ttm_fini(struct amdgpu_device *adev)
1030 if (!adev->mman.initialized)
1032 amdgpu_ttm_debugfs_fini(adev);
1033 if (adev->stollen_vga_memory) {
1034 r = amdgpu_bo_reserve(adev->stollen_vga_memory, false);
1036 amdgpu_bo_unpin(adev->stollen_vga_memory);
1037 amdgpu_bo_unreserve(adev->stollen_vga_memory);
1039 amdgpu_bo_unref(&adev->stollen_vga_memory);
1041 ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_VRAM);
1042 ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_TT);
1043 ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_GDS);
1044 ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_GWS);
1045 ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_OA);
1046 ttm_bo_device_release(&adev->mman.bdev);
1047 amdgpu_gart_fini(adev);
1048 amdgpu_ttm_global_fini(adev);
1049 adev->mman.initialized = false;
1050 DRM_INFO("amdgpu: ttm finalized\n");
1053 /* this should only be called at bootup or when userspace
1055 void amdgpu_ttm_set_active_vram_size(struct amdgpu_device *adev, u64 size)
1057 struct ttm_mem_type_manager *man;
1059 if (!adev->mman.initialized)
1062 man = &adev->mman.bdev.man[TTM_PL_VRAM];
1063 /* this just adjusts TTM size idea, which sets lpfn to the correct value */
1064 man->size = size >> PAGE_SHIFT;
1067 int amdgpu_mmap(struct file *filp, struct vm_area_struct *vma)
1069 struct drm_file *file_priv;
1070 struct amdgpu_device *adev;
1072 if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET))
1075 file_priv = filp->private_data;
1076 adev = file_priv->minor->dev->dev_private;
1080 return ttm_bo_mmap(filp, vma, &adev->mman.bdev);
1083 int amdgpu_copy_buffer(struct amdgpu_ring *ring,
1084 uint64_t src_offset,
1085 uint64_t dst_offset,
1086 uint32_t byte_count,
1087 struct reservation_object *resv,
1088 struct fence **fence)
1090 struct amdgpu_device *adev = ring->adev;
1091 struct amdgpu_job *job;
1094 unsigned num_loops, num_dw;
1098 max_bytes = adev->mman.buffer_funcs->copy_max_bytes;
1099 num_loops = DIV_ROUND_UP(byte_count, max_bytes);
1100 num_dw = num_loops * adev->mman.buffer_funcs->copy_num_dw;
1102 /* for IB padding */
1103 while (num_dw & 0x7)
1106 r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, &job);
1111 r = amdgpu_sync_resv(adev, &job->sync, resv,
1112 AMDGPU_FENCE_OWNER_UNDEFINED);
1114 DRM_ERROR("sync failed (%d).\n", r);
1119 for (i = 0; i < num_loops; i++) {
1120 uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
1122 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_offset,
1123 dst_offset, cur_size_in_bytes);
1125 src_offset += cur_size_in_bytes;
1126 dst_offset += cur_size_in_bytes;
1127 byte_count -= cur_size_in_bytes;
1130 amdgpu_ring_pad_ib(ring, &job->ibs[0]);
1131 WARN_ON(job->ibs[0].length_dw > num_dw);
1132 r = amdgpu_job_submit(job, ring, &adev->mman.entity,
1133 AMDGPU_FENCE_OWNER_UNDEFINED, fence);
1140 amdgpu_job_free(job);
1144 #if defined(CONFIG_DEBUG_FS)
1146 static int amdgpu_mm_dump_table(struct seq_file *m, void *data)
1148 struct drm_info_node *node = (struct drm_info_node *)m->private;
1149 unsigned ttm_pl = *(int *)node->info_ent->data;
1150 struct drm_device *dev = node->minor->dev;
1151 struct amdgpu_device *adev = dev->dev_private;
1152 struct drm_mm *mm = (struct drm_mm *)adev->mman.bdev.man[ttm_pl].priv;
1154 struct ttm_bo_global *glob = adev->mman.bdev.glob;
1156 spin_lock(&glob->lru_lock);
1157 ret = drm_mm_dump_table(m, mm);
1158 spin_unlock(&glob->lru_lock);
1159 if (ttm_pl == TTM_PL_VRAM)
1160 seq_printf(m, "man size:%llu pages, ram usage:%lluMB, vis usage:%lluMB\n",
1161 adev->mman.bdev.man[ttm_pl].size,
1162 (u64)atomic64_read(&adev->vram_usage) >> 20,
1163 (u64)atomic64_read(&adev->vram_vis_usage) >> 20);
1167 static int ttm_pl_vram = TTM_PL_VRAM;
1168 static int ttm_pl_tt = TTM_PL_TT;
1170 static struct drm_info_list amdgpu_ttm_debugfs_list[] = {
1171 {"amdgpu_vram_mm", amdgpu_mm_dump_table, 0, &ttm_pl_vram},
1172 {"amdgpu_gtt_mm", amdgpu_mm_dump_table, 0, &ttm_pl_tt},
1173 {"ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL},
1174 #ifdef CONFIG_SWIOTLB
1175 {"ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL}
1179 static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf,
1180 size_t size, loff_t *pos)
1182 struct amdgpu_device *adev = f->f_inode->i_private;
1186 if (size & 0x3 || *pos & 0x3)
1190 unsigned long flags;
1193 if (*pos >= adev->mc.mc_vram_size)
1196 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
1197 WREG32(mmMM_INDEX, ((uint32_t)*pos) | 0x80000000);
1198 WREG32(mmMM_INDEX_HI, *pos >> 31);
1199 value = RREG32(mmMM_DATA);
1200 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
1202 r = put_user(value, (uint32_t *)buf);
1215 static const struct file_operations amdgpu_ttm_vram_fops = {
1216 .owner = THIS_MODULE,
1217 .read = amdgpu_ttm_vram_read,
1218 .llseek = default_llseek
1221 static ssize_t amdgpu_ttm_gtt_read(struct file *f, char __user *buf,
1222 size_t size, loff_t *pos)
1224 struct amdgpu_device *adev = f->f_inode->i_private;
1229 loff_t p = *pos / PAGE_SIZE;
1230 unsigned off = *pos & ~PAGE_MASK;
1231 size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
1235 if (p >= adev->gart.num_cpu_pages)
1238 page = adev->gart.pages[p];
1243 r = copy_to_user(buf, ptr, cur_size);
1244 kunmap(adev->gart.pages[p]);
1246 r = clear_user(buf, cur_size);
1260 static const struct file_operations amdgpu_ttm_gtt_fops = {
1261 .owner = THIS_MODULE,
1262 .read = amdgpu_ttm_gtt_read,
1263 .llseek = default_llseek
1268 static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev)
1270 #if defined(CONFIG_DEBUG_FS)
1273 struct drm_minor *minor = adev->ddev->primary;
1274 struct dentry *ent, *root = minor->debugfs_root;
1276 ent = debugfs_create_file("amdgpu_vram", S_IFREG | S_IRUGO, root,
1277 adev, &amdgpu_ttm_vram_fops);
1279 return PTR_ERR(ent);
1280 i_size_write(ent->d_inode, adev->mc.mc_vram_size);
1281 adev->mman.vram = ent;
1283 ent = debugfs_create_file("amdgpu_gtt", S_IFREG | S_IRUGO, root,
1284 adev, &amdgpu_ttm_gtt_fops);
1286 return PTR_ERR(ent);
1287 i_size_write(ent->d_inode, adev->mc.gtt_size);
1288 adev->mman.gtt = ent;
1290 count = ARRAY_SIZE(amdgpu_ttm_debugfs_list);
1292 #ifdef CONFIG_SWIOTLB
1293 if (!swiotlb_nr_tbl())
1297 return amdgpu_debugfs_add_files(adev, amdgpu_ttm_debugfs_list, count);
1304 static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev)
1306 #if defined(CONFIG_DEBUG_FS)
1308 debugfs_remove(adev->mman.vram);
1309 adev->mman.vram = NULL;
1311 debugfs_remove(adev->mman.gtt);
1312 adev->mman.gtt = NULL;