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 <drm/ttm/ttm_bo_driver.h>
33 #include <drm/ttm/ttm_placement.h>
34 #include <drm/drm_cache.h>
35 #include <drm/drm_vma_manager.h>
36 #include <linux/iosys-map.h>
38 #include <linux/highmem.h>
39 #include <linux/wait.h>
40 #include <linux/slab.h>
41 #include <linux/vmalloc.h>
42 #include <linux/module.h>
43 #include <linux/dma-resv.h>
45 struct ttm_transfer_obj {
46 struct ttm_buffer_object base;
47 struct ttm_buffer_object *bo;
50 int ttm_mem_io_reserve(struct ttm_device *bdev,
51 struct ttm_resource *mem)
53 if (mem->bus.offset || mem->bus.addr)
56 mem->bus.is_iomem = false;
57 if (!bdev->funcs->io_mem_reserve)
60 return bdev->funcs->io_mem_reserve(bdev, mem);
63 void ttm_mem_io_free(struct ttm_device *bdev,
64 struct ttm_resource *mem)
69 if (!mem->bus.offset && !mem->bus.addr)
72 if (bdev->funcs->io_mem_free)
73 bdev->funcs->io_mem_free(bdev, mem);
80 * ttm_move_memcpy - Helper to perform a memcpy ttm move operation.
81 * @clear: Whether to clear rather than copy.
82 * @num_pages: Number of pages of the operation.
83 * @dst_iter: A struct ttm_kmap_iter representing the destination resource.
84 * @src_iter: A struct ttm_kmap_iter representing the source resource.
86 * This function is intended to be able to move out async under a
87 * dma-fence if desired.
89 void ttm_move_memcpy(bool clear,
91 struct ttm_kmap_iter *dst_iter,
92 struct ttm_kmap_iter *src_iter)
94 const struct ttm_kmap_iter_ops *dst_ops = dst_iter->ops;
95 const struct ttm_kmap_iter_ops *src_ops = src_iter->ops;
96 struct iosys_map src_map, dst_map;
99 /* Single TTM move. NOP */
100 if (dst_ops->maps_tt && src_ops->maps_tt)
103 /* Don't move nonexistent data. Clear destination instead. */
105 for (i = 0; i < num_pages; ++i) {
106 dst_ops->map_local(dst_iter, &dst_map, i);
107 if (dst_map.is_iomem)
108 memset_io(dst_map.vaddr_iomem, 0, PAGE_SIZE);
110 memset(dst_map.vaddr, 0, PAGE_SIZE);
111 if (dst_ops->unmap_local)
112 dst_ops->unmap_local(dst_iter, &dst_map);
117 for (i = 0; i < num_pages; ++i) {
118 dst_ops->map_local(dst_iter, &dst_map, i);
119 src_ops->map_local(src_iter, &src_map, i);
121 drm_memcpy_from_wc(&dst_map, &src_map, PAGE_SIZE);
123 if (src_ops->unmap_local)
124 src_ops->unmap_local(src_iter, &src_map);
125 if (dst_ops->unmap_local)
126 dst_ops->unmap_local(dst_iter, &dst_map);
129 EXPORT_SYMBOL(ttm_move_memcpy);
131 int ttm_bo_move_memcpy(struct ttm_buffer_object *bo,
132 struct ttm_operation_ctx *ctx,
133 struct ttm_resource *dst_mem)
135 struct ttm_device *bdev = bo->bdev;
136 struct ttm_resource_manager *dst_man =
137 ttm_manager_type(bo->bdev, dst_mem->mem_type);
138 struct ttm_tt *ttm = bo->ttm;
139 struct ttm_resource *src_mem = bo->resource;
140 struct ttm_resource_manager *src_man;
142 struct ttm_kmap_iter_tt tt;
143 struct ttm_kmap_iter_linear_io io;
144 } _dst_iter, _src_iter;
145 struct ttm_kmap_iter *dst_iter, *src_iter;
152 src_man = ttm_manager_type(bdev, src_mem->mem_type);
153 if (ttm && ((ttm->page_flags & TTM_TT_FLAG_SWAPPED) ||
155 ret = ttm_tt_populate(bdev, ttm, ctx);
160 dst_iter = ttm_kmap_iter_linear_io_init(&_dst_iter.io, bdev, dst_mem);
161 if (PTR_ERR(dst_iter) == -EINVAL && dst_man->use_tt)
162 dst_iter = ttm_kmap_iter_tt_init(&_dst_iter.tt, bo->ttm);
163 if (IS_ERR(dst_iter))
164 return PTR_ERR(dst_iter);
166 src_iter = ttm_kmap_iter_linear_io_init(&_src_iter.io, bdev, src_mem);
167 if (PTR_ERR(src_iter) == -EINVAL && src_man->use_tt)
168 src_iter = ttm_kmap_iter_tt_init(&_src_iter.tt, bo->ttm);
169 if (IS_ERR(src_iter)) {
170 ret = PTR_ERR(src_iter);
174 clear = src_iter->ops->maps_tt && (!ttm || !ttm_tt_is_populated(ttm));
175 if (!(clear && ttm && !(ttm->page_flags & TTM_TT_FLAG_ZERO_ALLOC)))
176 ttm_move_memcpy(clear, ttm->num_pages, dst_iter, src_iter);
178 if (!src_iter->ops->maps_tt)
179 ttm_kmap_iter_linear_io_fini(&_src_iter.io, bdev, src_mem);
180 ttm_bo_move_sync_cleanup(bo, dst_mem);
183 if (!dst_iter->ops->maps_tt)
184 ttm_kmap_iter_linear_io_fini(&_dst_iter.io, bdev, dst_mem);
188 EXPORT_SYMBOL(ttm_bo_move_memcpy);
190 static void ttm_transfered_destroy(struct ttm_buffer_object *bo)
192 struct ttm_transfer_obj *fbo;
194 fbo = container_of(bo, struct ttm_transfer_obj, base);
195 dma_resv_fini(&fbo->base.base._resv);
201 * ttm_buffer_object_transfer
203 * @bo: A pointer to a struct ttm_buffer_object.
204 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object,
205 * holding the data of @bo with the old placement.
207 * This is a utility function that may be called after an accelerated move
208 * has been scheduled. A new buffer object is created as a placeholder for
209 * the old data while it's being copied. When that buffer object is idle,
210 * it can be destroyed, releasing the space of the old placement.
215 static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
216 struct ttm_buffer_object **new_obj)
218 struct ttm_transfer_obj *fbo;
221 fbo = kmalloc(sizeof(*fbo), GFP_KERNEL);
228 * Fix up members that we shouldn't copy directly:
229 * TODO: Explicit member copy would probably be better here.
232 atomic_inc(&ttm_glob.bo_count);
233 INIT_LIST_HEAD(&fbo->base.ddestroy);
234 drm_vma_node_reset(&fbo->base.base.vma_node);
236 kref_init(&fbo->base.kref);
237 fbo->base.destroy = &ttm_transfered_destroy;
238 fbo->base.pin_count = 0;
239 if (bo->type != ttm_bo_type_sg)
240 fbo->base.base.resv = &fbo->base.base._resv;
242 dma_resv_init(&fbo->base.base._resv);
243 fbo->base.base.dev = NULL;
244 ret = dma_resv_trylock(&fbo->base.base._resv);
247 if (fbo->base.resource) {
248 ttm_resource_set_bo(fbo->base.resource, &fbo->base);
250 ttm_bo_set_bulk_move(&fbo->base, NULL);
252 fbo->base.bulk_move = NULL;
255 ret = dma_resv_reserve_fences(&fbo->base.base._resv, 1);
264 ttm_bo_move_to_lru_tail_unlocked(&fbo->base);
266 *new_obj = &fbo->base;
270 pgprot_t ttm_io_prot(struct ttm_buffer_object *bo, struct ttm_resource *res,
273 struct ttm_resource_manager *man;
274 enum ttm_caching caching;
276 man = ttm_manager_type(bo->bdev, res->mem_type);
277 caching = man->use_tt ? bo->ttm->caching : res->bus.caching;
279 return ttm_prot_from_caching(caching, tmp);
281 EXPORT_SYMBOL(ttm_io_prot);
283 static int ttm_bo_ioremap(struct ttm_buffer_object *bo,
284 unsigned long offset,
286 struct ttm_bo_kmap_obj *map)
288 struct ttm_resource *mem = bo->resource;
290 if (bo->resource->bus.addr) {
291 map->bo_kmap_type = ttm_bo_map_premapped;
292 map->virtual = ((u8 *)bo->resource->bus.addr) + offset;
294 resource_size_t res = bo->resource->bus.offset + offset;
296 map->bo_kmap_type = ttm_bo_map_iomap;
297 if (mem->bus.caching == ttm_write_combined)
298 map->virtual = ioremap_wc(res, size);
300 else if (mem->bus.caching == ttm_cached)
301 map->virtual = ioremap_cache(res, size);
304 map->virtual = ioremap(res, size);
306 return (!map->virtual) ? -ENOMEM : 0;
309 static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo,
310 unsigned long start_page,
311 unsigned long num_pages,
312 struct ttm_bo_kmap_obj *map)
314 struct ttm_resource *mem = bo->resource;
315 struct ttm_operation_ctx ctx = {
316 .interruptible = false,
319 struct ttm_tt *ttm = bo->ttm;
325 ret = ttm_tt_populate(bo->bdev, ttm, &ctx);
329 if (num_pages == 1 && ttm->caching == ttm_cached) {
331 * We're mapping a single page, and the desired
332 * page protection is consistent with the bo.
335 map->bo_kmap_type = ttm_bo_map_kmap;
336 map->page = ttm->pages[start_page];
337 map->virtual = kmap(map->page);
340 * We need to use vmap to get the desired page protection
341 * or to make the buffer object look contiguous.
343 prot = ttm_io_prot(bo, mem, PAGE_KERNEL);
344 map->bo_kmap_type = ttm_bo_map_vmap;
345 map->virtual = vmap(ttm->pages + start_page, num_pages,
348 return (!map->virtual) ? -ENOMEM : 0;
351 int ttm_bo_kmap(struct ttm_buffer_object *bo,
352 unsigned long start_page, unsigned long num_pages,
353 struct ttm_bo_kmap_obj *map)
355 unsigned long offset, size;
360 if (num_pages > PFN_UP(bo->resource->size))
362 if ((start_page + num_pages) > PFN_UP(bo->resource->size))
365 ret = ttm_mem_io_reserve(bo->bdev, bo->resource);
368 if (!bo->resource->bus.is_iomem) {
369 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map);
371 offset = start_page << PAGE_SHIFT;
372 size = num_pages << PAGE_SHIFT;
373 return ttm_bo_ioremap(bo, offset, size, map);
376 EXPORT_SYMBOL(ttm_bo_kmap);
378 void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map)
382 switch (map->bo_kmap_type) {
383 case ttm_bo_map_iomap:
384 iounmap(map->virtual);
386 case ttm_bo_map_vmap:
387 vunmap(map->virtual);
389 case ttm_bo_map_kmap:
392 case ttm_bo_map_premapped:
397 ttm_mem_io_free(map->bo->bdev, map->bo->resource);
401 EXPORT_SYMBOL(ttm_bo_kunmap);
403 int ttm_bo_vmap(struct ttm_buffer_object *bo, struct iosys_map *map)
405 struct ttm_resource *mem = bo->resource;
408 dma_resv_assert_held(bo->base.resv);
410 ret = ttm_mem_io_reserve(bo->bdev, mem);
414 if (mem->bus.is_iomem) {
415 void __iomem *vaddr_iomem;
418 vaddr_iomem = (void __iomem *)mem->bus.addr;
419 else if (mem->bus.caching == ttm_write_combined)
420 vaddr_iomem = ioremap_wc(mem->bus.offset,
423 else if (mem->bus.caching == ttm_cached)
424 vaddr_iomem = ioremap_cache(mem->bus.offset,
428 vaddr_iomem = ioremap(mem->bus.offset, bo->base.size);
433 iosys_map_set_vaddr_iomem(map, vaddr_iomem);
436 struct ttm_operation_ctx ctx = {
437 .interruptible = false,
440 struct ttm_tt *ttm = bo->ttm;
444 ret = ttm_tt_populate(bo->bdev, ttm, &ctx);
449 * We need to use vmap to get the desired page protection
450 * or to make the buffer object look contiguous.
452 prot = ttm_io_prot(bo, mem, PAGE_KERNEL);
453 vaddr = vmap(ttm->pages, ttm->num_pages, 0, prot);
457 iosys_map_set_vaddr(map, vaddr);
462 EXPORT_SYMBOL(ttm_bo_vmap);
464 void ttm_bo_vunmap(struct ttm_buffer_object *bo, struct iosys_map *map)
466 struct ttm_resource *mem = bo->resource;
468 dma_resv_assert_held(bo->base.resv);
470 if (iosys_map_is_null(map))
475 else if (!mem->bus.addr)
476 iounmap(map->vaddr_iomem);
477 iosys_map_clear(map);
479 ttm_mem_io_free(bo->bdev, bo->resource);
481 EXPORT_SYMBOL(ttm_bo_vunmap);
483 static int ttm_bo_wait_free_node(struct ttm_buffer_object *bo,
487 ret = ttm_bo_wait(bo, false, false);
492 ttm_bo_tt_destroy(bo);
493 ttm_resource_free(bo, &bo->resource);
497 static int ttm_bo_move_to_ghost(struct ttm_buffer_object *bo,
498 struct dma_fence *fence,
501 struct ttm_buffer_object *ghost_obj;
505 * This should help pipeline ordinary buffer moves.
507 * Hang old buffer memory on a new buffer object,
508 * and leave it to be released when the GPU
509 * operation has completed.
512 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
516 dma_resv_add_fence(&ghost_obj->base._resv, fence,
517 DMA_RESV_USAGE_KERNEL);
520 * If we're not moving to fixed memory, the TTM object
521 * needs to stay alive. Otherwhise hang it on the ghost
522 * bo to be unbound and destroyed.
526 ghost_obj->ttm = NULL;
530 dma_resv_unlock(&ghost_obj->base._resv);
531 ttm_bo_put(ghost_obj);
535 static void ttm_bo_move_pipeline_evict(struct ttm_buffer_object *bo,
536 struct dma_fence *fence)
538 struct ttm_device *bdev = bo->bdev;
539 struct ttm_resource_manager *from;
541 from = ttm_manager_type(bdev, bo->resource->mem_type);
544 * BO doesn't have a TTM we need to bind/unbind. Just remember
545 * this eviction and free up the allocation
547 spin_lock(&from->move_lock);
548 if (!from->move || dma_fence_is_later(fence, from->move)) {
549 dma_fence_put(from->move);
550 from->move = dma_fence_get(fence);
552 spin_unlock(&from->move_lock);
554 ttm_resource_free(bo, &bo->resource);
557 int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
558 struct dma_fence *fence,
561 struct ttm_resource *new_mem)
563 struct ttm_device *bdev = bo->bdev;
564 struct ttm_resource_manager *from = ttm_manager_type(bdev, bo->resource->mem_type);
565 struct ttm_resource_manager *man = ttm_manager_type(bdev, new_mem->mem_type);
568 dma_resv_add_fence(bo->base.resv, fence, DMA_RESV_USAGE_KERNEL);
570 ret = ttm_bo_move_to_ghost(bo, fence, man->use_tt);
571 else if (!from->use_tt && pipeline)
572 ttm_bo_move_pipeline_evict(bo, fence);
574 ret = ttm_bo_wait_free_node(bo, man->use_tt);
579 ttm_bo_assign_mem(bo, new_mem);
583 EXPORT_SYMBOL(ttm_bo_move_accel_cleanup);
585 void ttm_bo_move_sync_cleanup(struct ttm_buffer_object *bo,
586 struct ttm_resource *new_mem)
588 struct ttm_device *bdev = bo->bdev;
589 struct ttm_resource_manager *man = ttm_manager_type(bdev, new_mem->mem_type);
592 ret = ttm_bo_wait_free_node(bo, man->use_tt);
596 ttm_bo_assign_mem(bo, new_mem);
598 EXPORT_SYMBOL(ttm_bo_move_sync_cleanup);
601 * ttm_bo_pipeline_gutting - purge the contents of a bo
602 * @bo: The buffer object
604 * Purge the contents of a bo, async if the bo is not idle.
605 * After a successful call, the bo is left unpopulated in
606 * system placement. The function may wait uninterruptible
609 * Return: 0 if successful, negative error code on failure.
611 int ttm_bo_pipeline_gutting(struct ttm_buffer_object *bo)
613 static const struct ttm_place sys_mem = { .mem_type = TTM_PL_SYSTEM };
614 struct ttm_buffer_object *ghost;
615 struct ttm_resource *sys_res;
619 ret = ttm_resource_alloc(bo, &sys_mem, &sys_res);
623 /* If already idle, no need for ghost object dance. */
624 ret = ttm_bo_wait(bo, false, true);
627 /* See comment below about clearing. */
628 ret = ttm_tt_create(bo, true);
630 goto error_free_sys_mem;
632 ttm_tt_unpopulate(bo->bdev, bo->ttm);
633 if (bo->type == ttm_bo_type_device)
634 ttm_tt_mark_for_clear(bo->ttm);
636 ttm_resource_free(bo, &bo->resource);
637 ttm_bo_assign_mem(bo, sys_res);
642 * We need an unpopulated ttm_tt after giving our current one,
643 * if any, to the ghost object. And we can't afford to fail
644 * creating one *after* the operation. If the bo subsequently gets
645 * resurrected, make sure it's cleared (if ttm_bo_type_device)
646 * to avoid leaking sensitive information to user-space.
651 ret = ttm_tt_create(bo, true);
654 goto error_free_sys_mem;
656 ret = ttm_buffer_object_transfer(bo, &ghost);
658 goto error_destroy_tt;
660 ret = dma_resv_copy_fences(&ghost->base._resv, bo->base.resv);
661 /* Last resort, wait for the BO to be idle when we are OOM */
663 ttm_bo_wait(bo, false, false);
665 dma_resv_unlock(&ghost->base._resv);
668 ttm_bo_assign_mem(bo, sys_res);
672 ttm_tt_destroy(bo->bdev, ttm);
675 ttm_resource_free(bo, &sys_res);