1 /**************************************************************************
3 * Copyright (c) 2007-2009 VMware, Inc., Palo Alto, CA., USA
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
21 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
22 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
23 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
24 * USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **************************************************************************/
28 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
31 #include <drm/ttm/ttm_bo_driver.h>
32 #include <drm/ttm/ttm_placement.h>
33 #include <drm/drm_vma_manager.h>
35 #include <linux/highmem.h>
36 #include <linux/wait.h>
37 #include <linux/slab.h>
38 #include <linux/vmalloc.h>
39 #include <linux/module.h>
40 #include <linux/reservation.h>
42 void ttm_bo_free_old_node(struct ttm_buffer_object *bo)
44 ttm_bo_mem_put(bo, &bo->mem);
47 int ttm_bo_move_ttm(struct ttm_buffer_object *bo,
48 bool interruptible, bool no_wait_gpu,
49 struct ttm_mem_reg *new_mem)
51 struct ttm_tt *ttm = bo->ttm;
52 struct ttm_mem_reg *old_mem = &bo->mem;
55 if (old_mem->mem_type != TTM_PL_SYSTEM) {
56 ret = ttm_bo_wait(bo, interruptible, no_wait_gpu);
58 if (unlikely(ret != 0)) {
59 if (ret != -ERESTARTSYS)
60 pr_err("Failed to expire sync object before unbinding TTM\n");
65 ttm_bo_free_old_node(bo);
66 ttm_flag_masked(&old_mem->placement, TTM_PL_FLAG_SYSTEM,
68 old_mem->mem_type = TTM_PL_SYSTEM;
71 ret = ttm_tt_set_placement_caching(ttm, new_mem->placement);
72 if (unlikely(ret != 0))
75 if (new_mem->mem_type != TTM_PL_SYSTEM) {
76 ret = ttm_tt_bind(ttm, new_mem);
77 if (unlikely(ret != 0))
82 new_mem->mm_node = NULL;
86 EXPORT_SYMBOL(ttm_bo_move_ttm);
88 int ttm_mem_io_lock(struct ttm_mem_type_manager *man, bool interruptible)
90 if (likely(man->io_reserve_fastpath))
94 return mutex_lock_interruptible(&man->io_reserve_mutex);
96 mutex_lock(&man->io_reserve_mutex);
99 EXPORT_SYMBOL(ttm_mem_io_lock);
101 void ttm_mem_io_unlock(struct ttm_mem_type_manager *man)
103 if (likely(man->io_reserve_fastpath))
106 mutex_unlock(&man->io_reserve_mutex);
108 EXPORT_SYMBOL(ttm_mem_io_unlock);
110 static int ttm_mem_io_evict(struct ttm_mem_type_manager *man)
112 struct ttm_buffer_object *bo;
114 if (!man->use_io_reserve_lru || list_empty(&man->io_reserve_lru))
117 bo = list_first_entry(&man->io_reserve_lru,
118 struct ttm_buffer_object,
120 list_del_init(&bo->io_reserve_lru);
121 ttm_bo_unmap_virtual_locked(bo);
127 int ttm_mem_io_reserve(struct ttm_bo_device *bdev,
128 struct ttm_mem_reg *mem)
130 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
133 if (!bdev->driver->io_mem_reserve)
135 if (likely(man->io_reserve_fastpath))
136 return bdev->driver->io_mem_reserve(bdev, mem);
138 if (bdev->driver->io_mem_reserve &&
139 mem->bus.io_reserved_count++ == 0) {
141 ret = bdev->driver->io_mem_reserve(bdev, mem);
142 if (ret == -EAGAIN) {
143 ret = ttm_mem_io_evict(man);
150 EXPORT_SYMBOL(ttm_mem_io_reserve);
152 void ttm_mem_io_free(struct ttm_bo_device *bdev,
153 struct ttm_mem_reg *mem)
155 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
157 if (likely(man->io_reserve_fastpath))
160 if (bdev->driver->io_mem_reserve &&
161 --mem->bus.io_reserved_count == 0 &&
162 bdev->driver->io_mem_free)
163 bdev->driver->io_mem_free(bdev, mem);
166 EXPORT_SYMBOL(ttm_mem_io_free);
168 int ttm_mem_io_reserve_vm(struct ttm_buffer_object *bo)
170 struct ttm_mem_reg *mem = &bo->mem;
173 if (!mem->bus.io_reserved_vm) {
174 struct ttm_mem_type_manager *man =
175 &bo->bdev->man[mem->mem_type];
177 ret = ttm_mem_io_reserve(bo->bdev, mem);
178 if (unlikely(ret != 0))
180 mem->bus.io_reserved_vm = true;
181 if (man->use_io_reserve_lru)
182 list_add_tail(&bo->io_reserve_lru,
183 &man->io_reserve_lru);
188 void ttm_mem_io_free_vm(struct ttm_buffer_object *bo)
190 struct ttm_mem_reg *mem = &bo->mem;
192 if (mem->bus.io_reserved_vm) {
193 mem->bus.io_reserved_vm = false;
194 list_del_init(&bo->io_reserve_lru);
195 ttm_mem_io_free(bo->bdev, mem);
199 static int ttm_mem_reg_ioremap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem,
202 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
207 (void) ttm_mem_io_lock(man, false);
208 ret = ttm_mem_io_reserve(bdev, mem);
209 ttm_mem_io_unlock(man);
210 if (ret || !mem->bus.is_iomem)
214 addr = mem->bus.addr;
216 if (mem->placement & TTM_PL_FLAG_WC)
217 addr = ioremap_wc(mem->bus.base + mem->bus.offset, mem->bus.size);
219 addr = ioremap_nocache(mem->bus.base + mem->bus.offset, mem->bus.size);
221 (void) ttm_mem_io_lock(man, false);
222 ttm_mem_io_free(bdev, mem);
223 ttm_mem_io_unlock(man);
231 static void ttm_mem_reg_iounmap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem,
234 struct ttm_mem_type_manager *man;
236 man = &bdev->man[mem->mem_type];
238 if (virtual && mem->bus.addr == NULL)
240 (void) ttm_mem_io_lock(man, false);
241 ttm_mem_io_free(bdev, mem);
242 ttm_mem_io_unlock(man);
245 static int ttm_copy_io_page(void *dst, void *src, unsigned long page)
248 (uint32_t *) ((unsigned long)dst + (page << PAGE_SHIFT));
250 (uint32_t *) ((unsigned long)src + (page << PAGE_SHIFT));
253 for (i = 0; i < PAGE_SIZE / sizeof(uint32_t); ++i)
254 iowrite32(ioread32(srcP++), dstP++);
258 static int ttm_copy_io_ttm_page(struct ttm_tt *ttm, void *src,
262 struct page *d = ttm->pages[page];
268 src = (void *)((unsigned long)src + (page << PAGE_SHIFT));
271 dst = kmap_atomic_prot(d, prot);
273 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
274 dst = vmap(&d, 1, 0, prot);
281 memcpy_fromio(dst, src, PAGE_SIZE);
286 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
295 static int ttm_copy_ttm_io_page(struct ttm_tt *ttm, void *dst,
299 struct page *s = ttm->pages[page];
305 dst = (void *)((unsigned long)dst + (page << PAGE_SHIFT));
307 src = kmap_atomic_prot(s, prot);
309 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
310 src = vmap(&s, 1, 0, prot);
317 memcpy_toio(dst, src, PAGE_SIZE);
322 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
331 int ttm_bo_move_memcpy(struct ttm_buffer_object *bo,
332 bool interruptible, bool no_wait_gpu,
333 struct ttm_mem_reg *new_mem)
335 struct ttm_bo_device *bdev = bo->bdev;
336 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
337 struct ttm_tt *ttm = bo->ttm;
338 struct ttm_mem_reg *old_mem = &bo->mem;
339 struct ttm_mem_reg old_copy = *old_mem;
345 unsigned long add = 0;
348 ret = ttm_bo_wait(bo, interruptible, no_wait_gpu);
352 ret = ttm_mem_reg_ioremap(bdev, old_mem, &old_iomap);
355 ret = ttm_mem_reg_ioremap(bdev, new_mem, &new_iomap);
360 * Single TTM move. NOP.
362 if (old_iomap == NULL && new_iomap == NULL)
366 * Don't move nonexistent data. Clear destination instead.
368 if (old_iomap == NULL &&
369 (ttm == NULL || (ttm->state == tt_unpopulated &&
370 !(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)))) {
371 memset_io(new_iomap, 0, new_mem->num_pages*PAGE_SIZE);
376 * TTM might be null for moves within the same region.
378 if (ttm && ttm->state == tt_unpopulated) {
379 ret = ttm->bdev->driver->ttm_tt_populate(ttm);
387 if ((old_mem->mem_type == new_mem->mem_type) &&
388 (new_mem->start < old_mem->start + old_mem->size)) {
390 add = new_mem->num_pages - 1;
393 for (i = 0; i < new_mem->num_pages; ++i) {
394 page = i * dir + add;
395 if (old_iomap == NULL) {
396 pgprot_t prot = ttm_io_prot(old_mem->placement,
398 ret = ttm_copy_ttm_io_page(ttm, new_iomap, page,
400 } else if (new_iomap == NULL) {
401 pgprot_t prot = ttm_io_prot(new_mem->placement,
403 ret = ttm_copy_io_ttm_page(ttm, old_iomap, page,
406 ret = ttm_copy_io_page(new_iomap, old_iomap, page);
414 new_mem->mm_node = NULL;
416 if (man->flags & TTM_MEMTYPE_FLAG_FIXED) {
422 ttm_mem_reg_iounmap(bdev, old_mem, new_iomap);
424 ttm_mem_reg_iounmap(bdev, &old_copy, old_iomap);
427 * On error, keep the mm node!
430 ttm_bo_mem_put(bo, &old_copy);
433 EXPORT_SYMBOL(ttm_bo_move_memcpy);
435 static void ttm_transfered_destroy(struct ttm_buffer_object *bo)
441 * ttm_buffer_object_transfer
443 * @bo: A pointer to a struct ttm_buffer_object.
444 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object,
445 * holding the data of @bo with the old placement.
447 * This is a utility function that may be called after an accelerated move
448 * has been scheduled. A new buffer object is created as a placeholder for
449 * the old data while it's being copied. When that buffer object is idle,
450 * it can be destroyed, releasing the space of the old placement.
455 static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
456 struct ttm_buffer_object **new_obj)
458 struct ttm_buffer_object *fbo;
461 fbo = kmalloc(sizeof(*fbo), GFP_KERNEL);
468 * Fix up members that we shouldn't copy directly:
469 * TODO: Explicit member copy would probably be better here.
472 atomic_inc(&bo->glob->bo_count);
473 INIT_LIST_HEAD(&fbo->ddestroy);
474 INIT_LIST_HEAD(&fbo->lru);
475 INIT_LIST_HEAD(&fbo->swap);
476 INIT_LIST_HEAD(&fbo->io_reserve_lru);
477 mutex_init(&fbo->wu_mutex);
479 drm_vma_node_reset(&fbo->vma_node);
480 atomic_set(&fbo->cpu_writers, 0);
482 kref_init(&fbo->list_kref);
483 kref_init(&fbo->kref);
484 fbo->destroy = &ttm_transfered_destroy;
486 fbo->resv = &fbo->ttm_resv;
487 reservation_object_init(fbo->resv);
488 ret = ww_mutex_trylock(&fbo->resv->lock);
495 pgprot_t ttm_io_prot(uint32_t caching_flags, pgprot_t tmp)
497 /* Cached mappings need no adjustment */
498 if (caching_flags & TTM_PL_FLAG_CACHED)
501 #if defined(__i386__) || defined(__x86_64__)
502 if (caching_flags & TTM_PL_FLAG_WC)
503 tmp = pgprot_writecombine(tmp);
504 else if (boot_cpu_data.x86 > 3)
505 tmp = pgprot_noncached(tmp);
507 #if defined(__ia64__) || defined(__arm__) || defined(__aarch64__) || \
509 if (caching_flags & TTM_PL_FLAG_WC)
510 tmp = pgprot_writecombine(tmp);
512 tmp = pgprot_noncached(tmp);
514 #if defined(__sparc__) || defined(__mips__)
515 tmp = pgprot_noncached(tmp);
519 EXPORT_SYMBOL(ttm_io_prot);
521 static int ttm_bo_ioremap(struct ttm_buffer_object *bo,
522 unsigned long offset,
524 struct ttm_bo_kmap_obj *map)
526 struct ttm_mem_reg *mem = &bo->mem;
528 if (bo->mem.bus.addr) {
529 map->bo_kmap_type = ttm_bo_map_premapped;
530 map->virtual = (void *)(((u8 *)bo->mem.bus.addr) + offset);
532 map->bo_kmap_type = ttm_bo_map_iomap;
533 if (mem->placement & TTM_PL_FLAG_WC)
534 map->virtual = ioremap_wc(bo->mem.bus.base + bo->mem.bus.offset + offset,
537 map->virtual = ioremap_nocache(bo->mem.bus.base + bo->mem.bus.offset + offset,
540 return (!map->virtual) ? -ENOMEM : 0;
543 static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo,
544 unsigned long start_page,
545 unsigned long num_pages,
546 struct ttm_bo_kmap_obj *map)
548 struct ttm_mem_reg *mem = &bo->mem; pgprot_t prot;
549 struct ttm_tt *ttm = bo->ttm;
554 if (ttm->state == tt_unpopulated) {
555 ret = ttm->bdev->driver->ttm_tt_populate(ttm);
560 if (num_pages == 1 && (mem->placement & TTM_PL_FLAG_CACHED)) {
562 * We're mapping a single page, and the desired
563 * page protection is consistent with the bo.
566 map->bo_kmap_type = ttm_bo_map_kmap;
567 map->page = ttm->pages[start_page];
568 map->virtual = kmap(map->page);
571 * We need to use vmap to get the desired page protection
572 * or to make the buffer object look contiguous.
574 prot = ttm_io_prot(mem->placement, PAGE_KERNEL);
575 map->bo_kmap_type = ttm_bo_map_vmap;
576 map->virtual = vmap(ttm->pages + start_page, num_pages,
579 return (!map->virtual) ? -ENOMEM : 0;
582 int ttm_bo_kmap(struct ttm_buffer_object *bo,
583 unsigned long start_page, unsigned long num_pages,
584 struct ttm_bo_kmap_obj *map)
586 struct ttm_mem_type_manager *man =
587 &bo->bdev->man[bo->mem.mem_type];
588 unsigned long offset, size;
591 BUG_ON(!list_empty(&bo->swap));
594 if (num_pages > bo->num_pages)
596 if (start_page > bo->num_pages)
599 if (num_pages > 1 && !capable(CAP_SYS_ADMIN))
602 (void) ttm_mem_io_lock(man, false);
603 ret = ttm_mem_io_reserve(bo->bdev, &bo->mem);
604 ttm_mem_io_unlock(man);
607 if (!bo->mem.bus.is_iomem) {
608 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map);
610 offset = start_page << PAGE_SHIFT;
611 size = num_pages << PAGE_SHIFT;
612 return ttm_bo_ioremap(bo, offset, size, map);
615 EXPORT_SYMBOL(ttm_bo_kmap);
617 void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map)
619 struct ttm_buffer_object *bo = map->bo;
620 struct ttm_mem_type_manager *man =
621 &bo->bdev->man[bo->mem.mem_type];
625 switch (map->bo_kmap_type) {
626 case ttm_bo_map_iomap:
627 iounmap(map->virtual);
629 case ttm_bo_map_vmap:
630 vunmap(map->virtual);
632 case ttm_bo_map_kmap:
635 case ttm_bo_map_premapped:
640 (void) ttm_mem_io_lock(man, false);
641 ttm_mem_io_free(map->bo->bdev, &map->bo->mem);
642 ttm_mem_io_unlock(man);
646 EXPORT_SYMBOL(ttm_bo_kunmap);
648 int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
649 struct dma_fence *fence,
651 struct ttm_mem_reg *new_mem)
653 struct ttm_bo_device *bdev = bo->bdev;
654 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
655 struct ttm_mem_reg *old_mem = &bo->mem;
657 struct ttm_buffer_object *ghost_obj;
659 reservation_object_add_excl_fence(bo->resv, fence);
661 ret = ttm_bo_wait(bo, false, false);
665 if (man->flags & TTM_MEMTYPE_FLAG_FIXED) {
666 ttm_tt_destroy(bo->ttm);
669 ttm_bo_free_old_node(bo);
672 * This should help pipeline ordinary buffer moves.
674 * Hang old buffer memory on a new buffer object,
675 * and leave it to be released when the GPU
676 * operation has completed.
679 dma_fence_put(bo->moving);
680 bo->moving = dma_fence_get(fence);
682 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
686 reservation_object_add_excl_fence(ghost_obj->resv, fence);
689 * If we're not moving to fixed memory, the TTM object
690 * needs to stay alive. Otherwhise hang it on the ghost
691 * bo to be unbound and destroyed.
694 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED))
695 ghost_obj->ttm = NULL;
699 ttm_bo_unreserve(ghost_obj);
700 ttm_bo_unref(&ghost_obj);
704 new_mem->mm_node = NULL;
708 EXPORT_SYMBOL(ttm_bo_move_accel_cleanup);
710 int ttm_bo_pipeline_move(struct ttm_buffer_object *bo,
711 struct dma_fence *fence, bool evict,
712 struct ttm_mem_reg *new_mem)
714 struct ttm_bo_device *bdev = bo->bdev;
715 struct ttm_mem_reg *old_mem = &bo->mem;
717 struct ttm_mem_type_manager *from = &bdev->man[old_mem->mem_type];
718 struct ttm_mem_type_manager *to = &bdev->man[new_mem->mem_type];
722 reservation_object_add_excl_fence(bo->resv, fence);
725 struct ttm_buffer_object *ghost_obj;
728 * This should help pipeline ordinary buffer moves.
730 * Hang old buffer memory on a new buffer object,
731 * and leave it to be released when the GPU
732 * operation has completed.
735 dma_fence_put(bo->moving);
736 bo->moving = dma_fence_get(fence);
738 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
742 reservation_object_add_excl_fence(ghost_obj->resv, fence);
745 * If we're not moving to fixed memory, the TTM object
746 * needs to stay alive. Otherwhise hang it on the ghost
747 * bo to be unbound and destroyed.
750 if (!(to->flags & TTM_MEMTYPE_FLAG_FIXED))
751 ghost_obj->ttm = NULL;
755 ttm_bo_unreserve(ghost_obj);
756 ttm_bo_unref(&ghost_obj);
758 } else if (from->flags & TTM_MEMTYPE_FLAG_FIXED) {
761 * BO doesn't have a TTM we need to bind/unbind. Just remember
762 * this eviction and free up the allocation
765 spin_lock(&from->move_lock);
766 if (!from->move || dma_fence_is_later(fence, from->move)) {
767 dma_fence_put(from->move);
768 from->move = dma_fence_get(fence);
770 spin_unlock(&from->move_lock);
772 ttm_bo_free_old_node(bo);
774 dma_fence_put(bo->moving);
775 bo->moving = dma_fence_get(fence);
779 * Last resort, wait for the move to be completed.
781 * Should never happen in pratice.
784 ret = ttm_bo_wait(bo, false, false);
788 if (to->flags & TTM_MEMTYPE_FLAG_FIXED) {
789 ttm_tt_destroy(bo->ttm);
792 ttm_bo_free_old_node(bo);
796 new_mem->mm_node = NULL;
800 EXPORT_SYMBOL(ttm_bo_pipeline_move);