1 /* SPDX-License-Identifier: GPL-2.0 OR MIT */
2 /**************************************************************************
4 * Copyright (c) 2007-2009 VMware, Inc., Palo Alto, CA., USA
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
22 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
23 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
24 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
25 * USE OR OTHER DEALINGS IN THE SOFTWARE.
27 **************************************************************************/
29 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
32 #include <linux/vmalloc.h>
34 #include <drm/ttm/ttm_bo.h>
35 #include <drm/ttm/ttm_placement.h>
36 #include <drm/ttm/ttm_tt.h>
38 #include <drm/drm_cache.h>
40 struct ttm_transfer_obj {
41 struct ttm_buffer_object base;
42 struct ttm_buffer_object *bo;
45 int ttm_mem_io_reserve(struct ttm_device *bdev,
46 struct ttm_resource *mem)
48 if (mem->bus.offset || mem->bus.addr)
51 mem->bus.is_iomem = false;
52 if (!bdev->funcs->io_mem_reserve)
55 return bdev->funcs->io_mem_reserve(bdev, mem);
58 void ttm_mem_io_free(struct ttm_device *bdev,
59 struct ttm_resource *mem)
64 if (!mem->bus.offset && !mem->bus.addr)
67 if (bdev->funcs->io_mem_free)
68 bdev->funcs->io_mem_free(bdev, mem);
75 * ttm_move_memcpy - Helper to perform a memcpy ttm move operation.
76 * @clear: Whether to clear rather than copy.
77 * @num_pages: Number of pages of the operation.
78 * @dst_iter: A struct ttm_kmap_iter representing the destination resource.
79 * @src_iter: A struct ttm_kmap_iter representing the source resource.
81 * This function is intended to be able to move out async under a
82 * dma-fence if desired.
84 void ttm_move_memcpy(bool clear,
86 struct ttm_kmap_iter *dst_iter,
87 struct ttm_kmap_iter *src_iter)
89 const struct ttm_kmap_iter_ops *dst_ops = dst_iter->ops;
90 const struct ttm_kmap_iter_ops *src_ops = src_iter->ops;
91 struct iosys_map src_map, dst_map;
94 /* Single TTM move. NOP */
95 if (dst_ops->maps_tt && src_ops->maps_tt)
98 /* Don't move nonexistent data. Clear destination instead. */
100 for (i = 0; i < num_pages; ++i) {
101 dst_ops->map_local(dst_iter, &dst_map, i);
102 if (dst_map.is_iomem)
103 memset_io(dst_map.vaddr_iomem, 0, PAGE_SIZE);
105 memset(dst_map.vaddr, 0, PAGE_SIZE);
106 if (dst_ops->unmap_local)
107 dst_ops->unmap_local(dst_iter, &dst_map);
112 for (i = 0; i < num_pages; ++i) {
113 dst_ops->map_local(dst_iter, &dst_map, i);
114 src_ops->map_local(src_iter, &src_map, i);
116 drm_memcpy_from_wc(&dst_map, &src_map, PAGE_SIZE);
118 if (src_ops->unmap_local)
119 src_ops->unmap_local(src_iter, &src_map);
120 if (dst_ops->unmap_local)
121 dst_ops->unmap_local(dst_iter, &dst_map);
124 EXPORT_SYMBOL(ttm_move_memcpy);
129 * @bo: A pointer to a struct ttm_buffer_object.
130 * @ctx: operation context
131 * @dst_mem: struct ttm_resource indicating where to move.
133 * Fallback move function for a mappable buffer object in mappable memory.
134 * The function will, if successful,
135 * free any old aperture space, and set (@new_mem)->mm_node to NULL,
136 * and update the (@bo)->mem placement flags. If unsuccessful, the old
137 * data remains untouched, and it's up to the caller to free the
138 * memory space indicated by @new_mem.
142 int ttm_bo_move_memcpy(struct ttm_buffer_object *bo,
143 struct ttm_operation_ctx *ctx,
144 struct ttm_resource *dst_mem)
146 struct ttm_device *bdev = bo->bdev;
147 struct ttm_resource_manager *dst_man =
148 ttm_manager_type(bo->bdev, dst_mem->mem_type);
149 struct ttm_tt *ttm = bo->ttm;
150 struct ttm_resource *src_mem = bo->resource;
151 struct ttm_resource_manager *src_man;
153 struct ttm_kmap_iter_tt tt;
154 struct ttm_kmap_iter_linear_io io;
155 } _dst_iter, _src_iter;
156 struct ttm_kmap_iter *dst_iter, *src_iter;
163 src_man = ttm_manager_type(bdev, src_mem->mem_type);
164 if (ttm && ((ttm->page_flags & TTM_TT_FLAG_SWAPPED) ||
166 ret = ttm_tt_populate(bdev, ttm, ctx);
171 dst_iter = ttm_kmap_iter_linear_io_init(&_dst_iter.io, bdev, dst_mem);
172 if (PTR_ERR(dst_iter) == -EINVAL && dst_man->use_tt)
173 dst_iter = ttm_kmap_iter_tt_init(&_dst_iter.tt, bo->ttm);
174 if (IS_ERR(dst_iter))
175 return PTR_ERR(dst_iter);
177 src_iter = ttm_kmap_iter_linear_io_init(&_src_iter.io, bdev, src_mem);
178 if (PTR_ERR(src_iter) == -EINVAL && src_man->use_tt)
179 src_iter = ttm_kmap_iter_tt_init(&_src_iter.tt, bo->ttm);
180 if (IS_ERR(src_iter)) {
181 ret = PTR_ERR(src_iter);
185 clear = src_iter->ops->maps_tt && (!ttm || !ttm_tt_is_populated(ttm));
186 if (!(clear && ttm && !(ttm->page_flags & TTM_TT_FLAG_ZERO_ALLOC)))
187 ttm_move_memcpy(clear, PFN_UP(dst_mem->size), dst_iter, src_iter);
189 if (!src_iter->ops->maps_tt)
190 ttm_kmap_iter_linear_io_fini(&_src_iter.io, bdev, src_mem);
191 ttm_bo_move_sync_cleanup(bo, dst_mem);
194 if (!dst_iter->ops->maps_tt)
195 ttm_kmap_iter_linear_io_fini(&_dst_iter.io, bdev, dst_mem);
199 EXPORT_SYMBOL(ttm_bo_move_memcpy);
201 static void ttm_transfered_destroy(struct ttm_buffer_object *bo)
203 struct ttm_transfer_obj *fbo;
205 fbo = container_of(bo, struct ttm_transfer_obj, base);
206 dma_resv_fini(&fbo->base.base._resv);
212 * ttm_buffer_object_transfer
214 * @bo: A pointer to a struct ttm_buffer_object.
215 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object,
216 * holding the data of @bo with the old placement.
218 * This is a utility function that may be called after an accelerated move
219 * has been scheduled. A new buffer object is created as a placeholder for
220 * the old data while it's being copied. When that buffer object is idle,
221 * it can be destroyed, releasing the space of the old placement.
226 static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
227 struct ttm_buffer_object **new_obj)
229 struct ttm_transfer_obj *fbo;
232 fbo = kmalloc(sizeof(*fbo), GFP_KERNEL);
239 * Fix up members that we shouldn't copy directly:
240 * TODO: Explicit member copy would probably be better here.
243 atomic_inc(&ttm_glob.bo_count);
244 drm_vma_node_reset(&fbo->base.base.vma_node);
246 kref_init(&fbo->base.kref);
247 fbo->base.destroy = &ttm_transfered_destroy;
248 fbo->base.pin_count = 0;
249 if (bo->type != ttm_bo_type_sg)
250 fbo->base.base.resv = &fbo->base.base._resv;
252 dma_resv_init(&fbo->base.base._resv);
253 fbo->base.base.dev = NULL;
254 ret = dma_resv_trylock(&fbo->base.base._resv);
257 if (fbo->base.resource) {
258 ttm_resource_set_bo(fbo->base.resource, &fbo->base);
260 ttm_bo_set_bulk_move(&fbo->base, NULL);
262 fbo->base.bulk_move = NULL;
265 ret = dma_resv_reserve_fences(&fbo->base.base._resv, 1);
274 ttm_bo_move_to_lru_tail_unlocked(&fbo->base);
276 *new_obj = &fbo->base;
283 * @bo: ttm buffer object
284 * @res: ttm resource object
285 * @tmp: Page protection flag for a normal, cached mapping.
287 * Utility function that returns the pgprot_t that should be used for
288 * setting up a PTE with the caching model indicated by @c_state.
290 pgprot_t ttm_io_prot(struct ttm_buffer_object *bo, struct ttm_resource *res,
293 struct ttm_resource_manager *man;
294 enum ttm_caching caching;
296 man = ttm_manager_type(bo->bdev, res->mem_type);
297 caching = man->use_tt ? bo->ttm->caching : res->bus.caching;
299 return ttm_prot_from_caching(caching, tmp);
301 EXPORT_SYMBOL(ttm_io_prot);
303 static int ttm_bo_ioremap(struct ttm_buffer_object *bo,
304 unsigned long offset,
306 struct ttm_bo_kmap_obj *map)
308 struct ttm_resource *mem = bo->resource;
310 if (bo->resource->bus.addr) {
311 map->bo_kmap_type = ttm_bo_map_premapped;
312 map->virtual = ((u8 *)bo->resource->bus.addr) + offset;
314 resource_size_t res = bo->resource->bus.offset + offset;
316 map->bo_kmap_type = ttm_bo_map_iomap;
317 if (mem->bus.caching == ttm_write_combined)
318 map->virtual = ioremap_wc(res, size);
320 else if (mem->bus.caching == ttm_cached)
321 map->virtual = ioremap_cache(res, size);
324 map->virtual = ioremap(res, size);
326 return (!map->virtual) ? -ENOMEM : 0;
329 static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo,
330 unsigned long start_page,
331 unsigned long num_pages,
332 struct ttm_bo_kmap_obj *map)
334 struct ttm_resource *mem = bo->resource;
335 struct ttm_operation_ctx ctx = {
336 .interruptible = false,
339 struct ttm_tt *ttm = bo->ttm;
345 ret = ttm_tt_populate(bo->bdev, ttm, &ctx);
349 if (num_pages == 1 && ttm->caching == ttm_cached) {
351 * We're mapping a single page, and the desired
352 * page protection is consistent with the bo.
355 map->bo_kmap_type = ttm_bo_map_kmap;
356 map->page = ttm->pages[start_page];
357 map->virtual = kmap(map->page);
360 * We need to use vmap to get the desired page protection
361 * or to make the buffer object look contiguous.
363 prot = ttm_io_prot(bo, mem, PAGE_KERNEL);
364 map->bo_kmap_type = ttm_bo_map_vmap;
365 map->virtual = vmap(ttm->pages + start_page, num_pages,
368 return (!map->virtual) ? -ENOMEM : 0;
374 * @bo: The buffer object.
375 * @start_page: The first page to map.
376 * @num_pages: Number of pages to map.
377 * @map: pointer to a struct ttm_bo_kmap_obj representing the map.
379 * Sets up a kernel virtual mapping, using ioremap, vmap or kmap to the
380 * data in the buffer object. The ttm_kmap_obj_virtual function can then be
381 * used to obtain a virtual address to the data.
384 * -ENOMEM: Out of memory.
385 * -EINVAL: Invalid range.
387 int ttm_bo_kmap(struct ttm_buffer_object *bo,
388 unsigned long start_page, unsigned long num_pages,
389 struct ttm_bo_kmap_obj *map)
391 unsigned long offset, size;
396 if (num_pages > PFN_UP(bo->resource->size))
398 if ((start_page + num_pages) > PFN_UP(bo->resource->size))
401 ret = ttm_mem_io_reserve(bo->bdev, bo->resource);
404 if (!bo->resource->bus.is_iomem) {
405 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map);
407 offset = start_page << PAGE_SHIFT;
408 size = num_pages << PAGE_SHIFT;
409 return ttm_bo_ioremap(bo, offset, size, map);
412 EXPORT_SYMBOL(ttm_bo_kmap);
417 * @map: Object describing the map to unmap.
419 * Unmaps a kernel map set up by ttm_bo_kmap.
421 void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map)
425 switch (map->bo_kmap_type) {
426 case ttm_bo_map_iomap:
427 iounmap(map->virtual);
429 case ttm_bo_map_vmap:
430 vunmap(map->virtual);
432 case ttm_bo_map_kmap:
435 case ttm_bo_map_premapped:
440 ttm_mem_io_free(map->bo->bdev, map->bo->resource);
444 EXPORT_SYMBOL(ttm_bo_kunmap);
449 * @bo: The buffer object.
450 * @map: pointer to a struct iosys_map representing the map.
452 * Sets up a kernel virtual mapping, using ioremap or vmap to the
453 * data in the buffer object. The parameter @map returns the virtual
454 * address as struct iosys_map. Unmap the buffer with ttm_bo_vunmap().
457 * -ENOMEM: Out of memory.
458 * -EINVAL: Invalid range.
460 int ttm_bo_vmap(struct ttm_buffer_object *bo, struct iosys_map *map)
462 struct ttm_resource *mem = bo->resource;
465 dma_resv_assert_held(bo->base.resv);
467 ret = ttm_mem_io_reserve(bo->bdev, mem);
471 if (mem->bus.is_iomem) {
472 void __iomem *vaddr_iomem;
475 vaddr_iomem = (void __iomem *)mem->bus.addr;
476 else if (mem->bus.caching == ttm_write_combined)
477 vaddr_iomem = ioremap_wc(mem->bus.offset,
480 else if (mem->bus.caching == ttm_cached)
481 vaddr_iomem = ioremap_cache(mem->bus.offset,
485 vaddr_iomem = ioremap(mem->bus.offset, bo->base.size);
490 iosys_map_set_vaddr_iomem(map, vaddr_iomem);
493 struct ttm_operation_ctx ctx = {
494 .interruptible = false,
497 struct ttm_tt *ttm = bo->ttm;
501 ret = ttm_tt_populate(bo->bdev, ttm, &ctx);
506 * We need to use vmap to get the desired page protection
507 * or to make the buffer object look contiguous.
509 prot = ttm_io_prot(bo, mem, PAGE_KERNEL);
510 vaddr = vmap(ttm->pages, ttm->num_pages, 0, prot);
514 iosys_map_set_vaddr(map, vaddr);
519 EXPORT_SYMBOL(ttm_bo_vmap);
524 * @bo: The buffer object.
525 * @map: Object describing the map to unmap.
527 * Unmaps a kernel map set up by ttm_bo_vmap().
529 void ttm_bo_vunmap(struct ttm_buffer_object *bo, struct iosys_map *map)
531 struct ttm_resource *mem = bo->resource;
533 dma_resv_assert_held(bo->base.resv);
535 if (iosys_map_is_null(map))
540 else if (!mem->bus.addr)
541 iounmap(map->vaddr_iomem);
542 iosys_map_clear(map);
544 ttm_mem_io_free(bo->bdev, bo->resource);
546 EXPORT_SYMBOL(ttm_bo_vunmap);
548 static int ttm_bo_wait_free_node(struct ttm_buffer_object *bo,
553 ret = dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP,
561 ttm_bo_tt_destroy(bo);
562 ttm_resource_free(bo, &bo->resource);
566 static int ttm_bo_move_to_ghost(struct ttm_buffer_object *bo,
567 struct dma_fence *fence,
570 struct ttm_buffer_object *ghost_obj;
574 * This should help pipeline ordinary buffer moves.
576 * Hang old buffer memory on a new buffer object,
577 * and leave it to be released when the GPU
578 * operation has completed.
581 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
585 dma_resv_add_fence(&ghost_obj->base._resv, fence,
586 DMA_RESV_USAGE_KERNEL);
589 * If we're not moving to fixed memory, the TTM object
590 * needs to stay alive. Otherwhise hang it on the ghost
591 * bo to be unbound and destroyed.
595 ghost_obj->ttm = NULL;
599 dma_resv_unlock(&ghost_obj->base._resv);
600 ttm_bo_put(ghost_obj);
604 static void ttm_bo_move_pipeline_evict(struct ttm_buffer_object *bo,
605 struct dma_fence *fence)
607 struct ttm_device *bdev = bo->bdev;
608 struct ttm_resource_manager *from;
610 from = ttm_manager_type(bdev, bo->resource->mem_type);
613 * BO doesn't have a TTM we need to bind/unbind. Just remember
614 * this eviction and free up the allocation
616 spin_lock(&from->move_lock);
617 if (!from->move || dma_fence_is_later(fence, from->move)) {
618 dma_fence_put(from->move);
619 from->move = dma_fence_get(fence);
621 spin_unlock(&from->move_lock);
623 ttm_resource_free(bo, &bo->resource);
627 * ttm_bo_move_accel_cleanup - cleanup helper for hw copies
629 * @bo: A pointer to a struct ttm_buffer_object.
630 * @fence: A fence object that signals when moving is complete.
631 * @evict: This is an evict move. Don't return until the buffer is idle.
632 * @pipeline: evictions are to be pipelined.
633 * @new_mem: struct ttm_resource indicating where to move.
635 * Accelerated move function to be called when an accelerated move
636 * has been scheduled. The function will create a new temporary buffer object
637 * representing the old placement, and put the sync object on both buffer
638 * objects. After that the newly created buffer object is unref'd to be
639 * destroyed when the move is complete. This will help pipeline
642 int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
643 struct dma_fence *fence,
646 struct ttm_resource *new_mem)
648 struct ttm_device *bdev = bo->bdev;
649 struct ttm_resource_manager *from = ttm_manager_type(bdev, bo->resource->mem_type);
650 struct ttm_resource_manager *man = ttm_manager_type(bdev, new_mem->mem_type);
653 dma_resv_add_fence(bo->base.resv, fence, DMA_RESV_USAGE_KERNEL);
655 ret = ttm_bo_move_to_ghost(bo, fence, man->use_tt);
656 else if (!from->use_tt && pipeline)
657 ttm_bo_move_pipeline_evict(bo, fence);
659 ret = ttm_bo_wait_free_node(bo, man->use_tt);
664 ttm_bo_assign_mem(bo, new_mem);
668 EXPORT_SYMBOL(ttm_bo_move_accel_cleanup);
671 * ttm_bo_move_sync_cleanup - cleanup by waiting for the move to finish
673 * @bo: A pointer to a struct ttm_buffer_object.
674 * @new_mem: struct ttm_resource indicating where to move.
676 * Special case of ttm_bo_move_accel_cleanup where the bo is guaranteed
677 * by the caller to be idle. Typically used after memcpy buffer moves.
679 void ttm_bo_move_sync_cleanup(struct ttm_buffer_object *bo,
680 struct ttm_resource *new_mem)
682 struct ttm_device *bdev = bo->bdev;
683 struct ttm_resource_manager *man = ttm_manager_type(bdev, new_mem->mem_type);
686 ret = ttm_bo_wait_free_node(bo, man->use_tt);
690 ttm_bo_assign_mem(bo, new_mem);
692 EXPORT_SYMBOL(ttm_bo_move_sync_cleanup);
695 * ttm_bo_pipeline_gutting - purge the contents of a bo
696 * @bo: The buffer object
698 * Purge the contents of a bo, async if the bo is not idle.
699 * After a successful call, the bo is left unpopulated in
700 * system placement. The function may wait uninterruptible
703 * Return: 0 if successful, negative error code on failure.
705 int ttm_bo_pipeline_gutting(struct ttm_buffer_object *bo)
707 static const struct ttm_place sys_mem = { .mem_type = TTM_PL_SYSTEM };
708 struct ttm_buffer_object *ghost;
709 struct ttm_resource *sys_res;
713 ret = ttm_resource_alloc(bo, &sys_mem, &sys_res);
717 /* If already idle, no need for ghost object dance. */
718 if (dma_resv_test_signaled(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP)) {
720 /* See comment below about clearing. */
721 ret = ttm_tt_create(bo, true);
723 goto error_free_sys_mem;
725 ttm_tt_unpopulate(bo->bdev, bo->ttm);
726 if (bo->type == ttm_bo_type_device)
727 ttm_tt_mark_for_clear(bo->ttm);
729 ttm_resource_free(bo, &bo->resource);
730 ttm_bo_assign_mem(bo, sys_res);
735 * We need an unpopulated ttm_tt after giving our current one,
736 * if any, to the ghost object. And we can't afford to fail
737 * creating one *after* the operation. If the bo subsequently gets
738 * resurrected, make sure it's cleared (if ttm_bo_type_device)
739 * to avoid leaking sensitive information to user-space.
744 ret = ttm_tt_create(bo, true);
747 goto error_free_sys_mem;
749 ret = ttm_buffer_object_transfer(bo, &ghost);
751 goto error_destroy_tt;
753 ret = dma_resv_copy_fences(&ghost->base._resv, bo->base.resv);
754 /* Last resort, wait for the BO to be idle when we are OOM */
756 dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP,
757 false, MAX_SCHEDULE_TIMEOUT);
760 dma_resv_unlock(&ghost->base._resv);
763 ttm_bo_assign_mem(bo, sys_res);
767 ttm_tt_destroy(bo->bdev, ttm);
770 ttm_resource_free(bo, &sys_res);