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_vma_manager.h>
36 #include <linux/highmem.h>
37 #include <linux/wait.h>
38 #include <linux/slab.h>
39 #include <linux/vmalloc.h>
40 #include <linux/module.h>
41 #include <linux/dma-resv.h>
43 struct ttm_transfer_obj {
44 struct ttm_buffer_object base;
45 struct ttm_buffer_object *bo;
48 void ttm_bo_free_old_node(struct ttm_buffer_object *bo)
50 ttm_resource_free(bo, &bo->mem);
53 int ttm_bo_move_ttm(struct ttm_buffer_object *bo,
54 struct ttm_operation_ctx *ctx,
55 struct ttm_resource *new_mem)
57 struct ttm_tt *ttm = bo->ttm;
58 struct ttm_resource *old_mem = &bo->mem;
61 if (old_mem->mem_type != TTM_PL_SYSTEM) {
62 ret = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu);
64 if (unlikely(ret != 0)) {
65 if (ret != -ERESTARTSYS)
66 pr_err("Failed to expire sync object before unbinding TTM\n");
71 ttm_bo_free_old_node(bo);
72 old_mem->mem_type = TTM_PL_SYSTEM;
75 ret = ttm_tt_set_placement_caching(ttm, new_mem->placement);
76 if (unlikely(ret != 0))
79 if (new_mem->mem_type != TTM_PL_SYSTEM) {
81 ret = ttm_tt_populate(bo->bdev, ttm, ctx);
82 if (unlikely(ret != 0))
85 ret = ttm_bo_tt_bind(bo, new_mem);
86 if (unlikely(ret != 0))
91 new_mem->mm_node = NULL;
95 EXPORT_SYMBOL(ttm_bo_move_ttm);
97 int ttm_mem_io_reserve(struct ttm_bo_device *bdev,
98 struct ttm_resource *mem)
100 if (mem->bus.offset || mem->bus.addr)
103 mem->bus.is_iomem = false;
104 if (!bdev->driver->io_mem_reserve)
107 return bdev->driver->io_mem_reserve(bdev, mem);
110 void ttm_mem_io_free(struct ttm_bo_device *bdev,
111 struct ttm_resource *mem)
113 if (!mem->bus.offset && !mem->bus.addr)
116 if (bdev->driver->io_mem_free)
117 bdev->driver->io_mem_free(bdev, mem);
120 mem->bus.addr = NULL;
123 static int ttm_resource_ioremap(struct ttm_bo_device *bdev,
124 struct ttm_resource *mem,
131 ret = ttm_mem_io_reserve(bdev, mem);
132 if (ret || !mem->bus.is_iomem)
136 addr = mem->bus.addr;
138 size_t bus_size = (size_t)mem->num_pages << PAGE_SHIFT;
140 if (mem->placement & TTM_PL_FLAG_WC)
141 addr = ioremap_wc(mem->bus.offset, bus_size);
143 addr = ioremap(mem->bus.offset, bus_size);
145 ttm_mem_io_free(bdev, mem);
153 static void ttm_resource_iounmap(struct ttm_bo_device *bdev,
154 struct ttm_resource *mem,
157 if (virtual && mem->bus.addr == NULL)
159 ttm_mem_io_free(bdev, mem);
162 static int ttm_copy_io_page(void *dst, void *src, unsigned long page)
165 (uint32_t *) ((unsigned long)dst + (page << PAGE_SHIFT));
167 (uint32_t *) ((unsigned long)src + (page << PAGE_SHIFT));
170 for (i = 0; i < PAGE_SIZE / sizeof(uint32_t); ++i)
171 iowrite32(ioread32(srcP++), dstP++);
175 static int ttm_copy_io_ttm_page(struct ttm_tt *ttm, void *src,
179 struct page *d = ttm->pages[page];
185 src = (void *)((unsigned long)src + (page << PAGE_SHIFT));
186 dst = kmap_atomic_prot(d, prot);
190 memcpy_fromio(dst, src, PAGE_SIZE);
197 static int ttm_copy_ttm_io_page(struct ttm_tt *ttm, void *dst,
201 struct page *s = ttm->pages[page];
207 dst = (void *)((unsigned long)dst + (page << PAGE_SHIFT));
208 src = kmap_atomic_prot(s, prot);
212 memcpy_toio(dst, src, PAGE_SIZE);
219 int ttm_bo_move_memcpy(struct ttm_buffer_object *bo,
220 struct ttm_operation_ctx *ctx,
221 struct ttm_resource *new_mem)
223 struct ttm_bo_device *bdev = bo->bdev;
224 struct ttm_resource_manager *man = ttm_manager_type(bdev, new_mem->mem_type);
225 struct ttm_tt *ttm = bo->ttm;
226 struct ttm_resource *old_mem = &bo->mem;
227 struct ttm_resource old_copy = *old_mem;
233 unsigned long add = 0;
236 ret = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu);
240 ret = ttm_resource_ioremap(bdev, old_mem, &old_iomap);
243 ret = ttm_resource_ioremap(bdev, new_mem, &new_iomap);
248 * Single TTM move. NOP.
250 if (old_iomap == NULL && new_iomap == NULL)
254 * Don't move nonexistent data. Clear destination instead.
256 if (old_iomap == NULL &&
257 (ttm == NULL || (!ttm_tt_is_populated(ttm) &&
258 !(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)))) {
259 memset_io(new_iomap, 0, new_mem->num_pages*PAGE_SIZE);
264 * TTM might be null for moves within the same region.
267 ret = ttm_tt_populate(bdev, ttm, ctx);
275 if ((old_mem->mem_type == new_mem->mem_type) &&
276 (new_mem->start < old_mem->start + old_mem->size)) {
278 add = new_mem->num_pages - 1;
281 for (i = 0; i < new_mem->num_pages; ++i) {
282 page = i * dir + add;
283 if (old_iomap == NULL) {
284 pgprot_t prot = ttm_io_prot(old_mem->placement,
286 ret = ttm_copy_ttm_io_page(ttm, new_iomap, page,
288 } else if (new_iomap == NULL) {
289 pgprot_t prot = ttm_io_prot(new_mem->placement,
291 ret = ttm_copy_io_ttm_page(ttm, old_iomap, page,
294 ret = ttm_copy_io_page(new_iomap, old_iomap, page);
303 new_mem->mm_node = NULL;
306 ttm_bo_tt_destroy(bo);
309 ttm_resource_iounmap(bdev, old_mem, new_iomap);
311 ttm_resource_iounmap(bdev, &old_copy, old_iomap);
314 * On error, keep the mm node!
317 ttm_resource_free(bo, &old_copy);
320 EXPORT_SYMBOL(ttm_bo_move_memcpy);
322 static void ttm_transfered_destroy(struct ttm_buffer_object *bo)
324 struct ttm_transfer_obj *fbo;
326 fbo = container_of(bo, struct ttm_transfer_obj, base);
332 * ttm_buffer_object_transfer
334 * @bo: A pointer to a struct ttm_buffer_object.
335 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object,
336 * holding the data of @bo with the old placement.
338 * This is a utility function that may be called after an accelerated move
339 * has been scheduled. A new buffer object is created as a placeholder for
340 * the old data while it's being copied. When that buffer object is idle,
341 * it can be destroyed, releasing the space of the old placement.
346 static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
347 struct ttm_buffer_object **new_obj)
349 struct ttm_transfer_obj *fbo;
352 fbo = kmalloc(sizeof(*fbo), GFP_KERNEL);
357 fbo->base.mem.placement |= TTM_PL_FLAG_NO_EVICT;
363 * Fix up members that we shouldn't copy directly:
364 * TODO: Explicit member copy would probably be better here.
367 atomic_inc(&ttm_bo_glob.bo_count);
368 INIT_LIST_HEAD(&fbo->base.ddestroy);
369 INIT_LIST_HEAD(&fbo->base.lru);
370 INIT_LIST_HEAD(&fbo->base.swap);
371 fbo->base.moving = NULL;
372 drm_vma_node_reset(&fbo->base.base.vma_node);
374 kref_init(&fbo->base.kref);
375 fbo->base.destroy = &ttm_transfered_destroy;
376 fbo->base.acc_size = 0;
377 if (bo->type != ttm_bo_type_sg)
378 fbo->base.base.resv = &fbo->base.base._resv;
380 dma_resv_init(&fbo->base.base._resv);
381 fbo->base.base.dev = NULL;
382 ret = dma_resv_trylock(&fbo->base.base._resv);
385 *new_obj = &fbo->base;
389 pgprot_t ttm_io_prot(uint32_t caching_flags, pgprot_t tmp)
391 /* Cached mappings need no adjustment */
392 if (caching_flags & TTM_PL_FLAG_CACHED)
395 #if defined(__i386__) || defined(__x86_64__)
396 if (caching_flags & TTM_PL_FLAG_WC)
397 tmp = pgprot_writecombine(tmp);
398 else if (boot_cpu_data.x86 > 3)
399 tmp = pgprot_noncached(tmp);
401 #if defined(__ia64__) || defined(__arm__) || defined(__aarch64__) || \
402 defined(__powerpc__) || defined(__mips__)
403 if (caching_flags & TTM_PL_FLAG_WC)
404 tmp = pgprot_writecombine(tmp);
406 tmp = pgprot_noncached(tmp);
408 #if defined(__sparc__)
409 tmp = pgprot_noncached(tmp);
413 EXPORT_SYMBOL(ttm_io_prot);
415 static int ttm_bo_ioremap(struct ttm_buffer_object *bo,
416 unsigned long offset,
418 struct ttm_bo_kmap_obj *map)
420 struct ttm_resource *mem = &bo->mem;
422 if (bo->mem.bus.addr) {
423 map->bo_kmap_type = ttm_bo_map_premapped;
424 map->virtual = (void *)(((u8 *)bo->mem.bus.addr) + offset);
426 map->bo_kmap_type = ttm_bo_map_iomap;
427 if (mem->placement & TTM_PL_FLAG_WC)
428 map->virtual = ioremap_wc(bo->mem.bus.offset + offset,
431 map->virtual = ioremap(bo->mem.bus.offset + offset,
434 return (!map->virtual) ? -ENOMEM : 0;
437 static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo,
438 unsigned long start_page,
439 unsigned long num_pages,
440 struct ttm_bo_kmap_obj *map)
442 struct ttm_resource *mem = &bo->mem;
443 struct ttm_operation_ctx ctx = {
444 .interruptible = false,
447 struct ttm_tt *ttm = bo->ttm;
453 ret = ttm_tt_populate(bo->bdev, ttm, &ctx);
457 if (num_pages == 1 && (mem->placement & TTM_PL_FLAG_CACHED)) {
459 * We're mapping a single page, and the desired
460 * page protection is consistent with the bo.
463 map->bo_kmap_type = ttm_bo_map_kmap;
464 map->page = ttm->pages[start_page];
465 map->virtual = kmap(map->page);
468 * We need to use vmap to get the desired page protection
469 * or to make the buffer object look contiguous.
471 prot = ttm_io_prot(mem->placement, PAGE_KERNEL);
472 map->bo_kmap_type = ttm_bo_map_vmap;
473 map->virtual = vmap(ttm->pages + start_page, num_pages,
476 return (!map->virtual) ? -ENOMEM : 0;
479 int ttm_bo_kmap(struct ttm_buffer_object *bo,
480 unsigned long start_page, unsigned long num_pages,
481 struct ttm_bo_kmap_obj *map)
483 unsigned long offset, size;
488 if (num_pages > bo->num_pages)
490 if (start_page > bo->num_pages)
493 ret = ttm_mem_io_reserve(bo->bdev, &bo->mem);
496 if (!bo->mem.bus.is_iomem) {
497 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map);
499 offset = start_page << PAGE_SHIFT;
500 size = num_pages << PAGE_SHIFT;
501 return ttm_bo_ioremap(bo, offset, size, map);
504 EXPORT_SYMBOL(ttm_bo_kmap);
506 void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map)
510 switch (map->bo_kmap_type) {
511 case ttm_bo_map_iomap:
512 iounmap(map->virtual);
514 case ttm_bo_map_vmap:
515 vunmap(map->virtual);
517 case ttm_bo_map_kmap:
520 case ttm_bo_map_premapped:
525 ttm_mem_io_free(map->bo->bdev, &map->bo->mem);
529 EXPORT_SYMBOL(ttm_bo_kunmap);
531 int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
532 struct dma_fence *fence,
534 struct ttm_resource *new_mem)
536 struct ttm_bo_device *bdev = bo->bdev;
537 struct ttm_resource_manager *man = ttm_manager_type(bdev, new_mem->mem_type);
538 struct ttm_resource *old_mem = &bo->mem;
540 struct ttm_buffer_object *ghost_obj;
542 dma_resv_add_excl_fence(bo->base.resv, fence);
544 ret = ttm_bo_wait(bo, false, false);
549 ttm_bo_tt_destroy(bo);
550 ttm_bo_free_old_node(bo);
553 * This should help pipeline ordinary buffer moves.
555 * Hang old buffer memory on a new buffer object,
556 * and leave it to be released when the GPU
557 * operation has completed.
560 dma_fence_put(bo->moving);
561 bo->moving = dma_fence_get(fence);
563 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
567 dma_resv_add_excl_fence(&ghost_obj->base._resv, fence);
570 * If we're not moving to fixed memory, the TTM object
571 * needs to stay alive. Otherwhise hang it on the ghost
572 * bo to be unbound and destroyed.
576 ghost_obj->ttm = NULL;
580 dma_resv_unlock(&ghost_obj->base._resv);
581 ttm_bo_put(ghost_obj);
585 new_mem->mm_node = NULL;
589 EXPORT_SYMBOL(ttm_bo_move_accel_cleanup);
591 int ttm_bo_pipeline_move(struct ttm_buffer_object *bo,
592 struct dma_fence *fence, bool evict,
593 struct ttm_resource *new_mem)
595 struct ttm_bo_device *bdev = bo->bdev;
596 struct ttm_resource *old_mem = &bo->mem;
598 struct ttm_resource_manager *from = ttm_manager_type(bdev, old_mem->mem_type);
599 struct ttm_resource_manager *to = ttm_manager_type(bdev, new_mem->mem_type);
603 dma_resv_add_excl_fence(bo->base.resv, fence);
606 struct ttm_buffer_object *ghost_obj;
609 * This should help pipeline ordinary buffer moves.
611 * Hang old buffer memory on a new buffer object,
612 * and leave it to be released when the GPU
613 * operation has completed.
616 dma_fence_put(bo->moving);
617 bo->moving = dma_fence_get(fence);
619 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
623 dma_resv_add_excl_fence(&ghost_obj->base._resv, fence);
626 * If we're not moving to fixed memory, the TTM object
627 * needs to stay alive. Otherwhise hang it on the ghost
628 * bo to be unbound and destroyed.
632 ghost_obj->ttm = NULL;
636 dma_resv_unlock(&ghost_obj->base._resv);
637 ttm_bo_put(ghost_obj);
639 } else if (!from->use_tt) {
642 * BO doesn't have a TTM we need to bind/unbind. Just remember
643 * this eviction and free up the allocation
646 spin_lock(&from->move_lock);
647 if (!from->move || dma_fence_is_later(fence, from->move)) {
648 dma_fence_put(from->move);
649 from->move = dma_fence_get(fence);
651 spin_unlock(&from->move_lock);
653 ttm_bo_free_old_node(bo);
655 dma_fence_put(bo->moving);
656 bo->moving = dma_fence_get(fence);
660 * Last resort, wait for the move to be completed.
662 * Should never happen in pratice.
665 ret = ttm_bo_wait(bo, false, false);
670 ttm_bo_tt_destroy(bo);
671 ttm_bo_free_old_node(bo);
675 new_mem->mm_node = NULL;
679 EXPORT_SYMBOL(ttm_bo_pipeline_move);
681 int ttm_bo_pipeline_gutting(struct ttm_buffer_object *bo)
683 struct ttm_buffer_object *ghost;
686 ret = ttm_buffer_object_transfer(bo, &ghost);
690 ret = dma_resv_copy_fences(&ghost->base._resv, bo->base.resv);
691 /* Last resort, wait for the BO to be idle when we are OOM */
693 ttm_bo_wait(bo, false, false);
695 memset(&bo->mem, 0, sizeof(bo->mem));
696 bo->mem.mem_type = TTM_PL_SYSTEM;
699 dma_resv_unlock(&ghost->base._resv);