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))
90 ttm_bo_assign_mem(bo, new_mem);
93 EXPORT_SYMBOL(ttm_bo_move_ttm);
95 int ttm_mem_io_reserve(struct ttm_bo_device *bdev,
96 struct ttm_resource *mem)
98 if (mem->bus.offset || mem->bus.addr)
101 mem->bus.is_iomem = false;
102 if (!bdev->driver->io_mem_reserve)
105 return bdev->driver->io_mem_reserve(bdev, mem);
108 void ttm_mem_io_free(struct ttm_bo_device *bdev,
109 struct ttm_resource *mem)
111 if (!mem->bus.offset && !mem->bus.addr)
114 if (bdev->driver->io_mem_free)
115 bdev->driver->io_mem_free(bdev, mem);
118 mem->bus.addr = NULL;
121 static int ttm_resource_ioremap(struct ttm_bo_device *bdev,
122 struct ttm_resource *mem,
129 ret = ttm_mem_io_reserve(bdev, mem);
130 if (ret || !mem->bus.is_iomem)
134 addr = mem->bus.addr;
136 size_t bus_size = (size_t)mem->num_pages << PAGE_SHIFT;
138 if (mem->placement & TTM_PL_FLAG_WC)
139 addr = ioremap_wc(mem->bus.offset, bus_size);
141 addr = ioremap(mem->bus.offset, bus_size);
143 ttm_mem_io_free(bdev, mem);
151 static void ttm_resource_iounmap(struct ttm_bo_device *bdev,
152 struct ttm_resource *mem,
155 if (virtual && mem->bus.addr == NULL)
157 ttm_mem_io_free(bdev, mem);
160 static int ttm_copy_io_page(void *dst, void *src, unsigned long page)
163 (uint32_t *) ((unsigned long)dst + (page << PAGE_SHIFT));
165 (uint32_t *) ((unsigned long)src + (page << PAGE_SHIFT));
168 for (i = 0; i < PAGE_SIZE / sizeof(uint32_t); ++i)
169 iowrite32(ioread32(srcP++), dstP++);
173 static int ttm_copy_io_ttm_page(struct ttm_tt *ttm, void *src,
177 struct page *d = ttm->pages[page];
183 src = (void *)((unsigned long)src + (page << PAGE_SHIFT));
184 dst = kmap_atomic_prot(d, prot);
188 memcpy_fromio(dst, src, PAGE_SIZE);
195 static int ttm_copy_ttm_io_page(struct ttm_tt *ttm, void *dst,
199 struct page *s = ttm->pages[page];
205 dst = (void *)((unsigned long)dst + (page << PAGE_SHIFT));
206 src = kmap_atomic_prot(s, prot);
210 memcpy_toio(dst, src, PAGE_SIZE);
217 int ttm_bo_move_memcpy(struct ttm_buffer_object *bo,
218 struct ttm_operation_ctx *ctx,
219 struct ttm_resource *new_mem)
221 struct ttm_bo_device *bdev = bo->bdev;
222 struct ttm_resource_manager *man = ttm_manager_type(bdev, new_mem->mem_type);
223 struct ttm_tt *ttm = bo->ttm;
224 struct ttm_resource *old_mem = &bo->mem;
225 struct ttm_resource old_copy = *old_mem;
231 unsigned long add = 0;
234 ret = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu);
238 ret = ttm_resource_ioremap(bdev, old_mem, &old_iomap);
241 ret = ttm_resource_ioremap(bdev, new_mem, &new_iomap);
246 * Single TTM move. NOP.
248 if (old_iomap == NULL && new_iomap == NULL)
252 * Don't move nonexistent data. Clear destination instead.
254 if (old_iomap == NULL &&
255 (ttm == NULL || (!ttm_tt_is_populated(ttm) &&
256 !(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)))) {
257 memset_io(new_iomap, 0, new_mem->num_pages*PAGE_SIZE);
262 * TTM might be null for moves within the same region.
265 ret = ttm_tt_populate(bdev, ttm, ctx);
273 if ((old_mem->mem_type == new_mem->mem_type) &&
274 (new_mem->start < old_mem->start + old_mem->size)) {
276 add = new_mem->num_pages - 1;
279 for (i = 0; i < new_mem->num_pages; ++i) {
280 page = i * dir + add;
281 if (old_iomap == NULL) {
282 pgprot_t prot = ttm_io_prot(old_mem->placement,
284 ret = ttm_copy_ttm_io_page(ttm, new_iomap, page,
286 } else if (new_iomap == NULL) {
287 pgprot_t prot = ttm_io_prot(new_mem->placement,
289 ret = ttm_copy_io_ttm_page(ttm, old_iomap, page,
292 ret = ttm_copy_io_page(new_iomap, old_iomap, page);
301 ttm_bo_assign_mem(bo, new_mem);
304 ttm_bo_tt_destroy(bo);
307 ttm_resource_iounmap(bdev, old_mem, new_iomap);
309 ttm_resource_iounmap(bdev, &old_copy, old_iomap);
312 * On error, keep the mm node!
315 ttm_resource_free(bo, &old_copy);
318 EXPORT_SYMBOL(ttm_bo_move_memcpy);
320 static void ttm_transfered_destroy(struct ttm_buffer_object *bo)
322 struct ttm_transfer_obj *fbo;
324 fbo = container_of(bo, struct ttm_transfer_obj, base);
330 * ttm_buffer_object_transfer
332 * @bo: A pointer to a struct ttm_buffer_object.
333 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object,
334 * holding the data of @bo with the old placement.
336 * This is a utility function that may be called after an accelerated move
337 * has been scheduled. A new buffer object is created as a placeholder for
338 * the old data while it's being copied. When that buffer object is idle,
339 * it can be destroyed, releasing the space of the old placement.
344 static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
345 struct ttm_buffer_object **new_obj)
347 struct ttm_transfer_obj *fbo;
350 fbo = kmalloc(sizeof(*fbo), GFP_KERNEL);
355 fbo->base.mem.placement |= TTM_PL_FLAG_NO_EVICT;
361 * Fix up members that we shouldn't copy directly:
362 * TODO: Explicit member copy would probably be better here.
365 atomic_inc(&ttm_bo_glob.bo_count);
366 INIT_LIST_HEAD(&fbo->base.ddestroy);
367 INIT_LIST_HEAD(&fbo->base.lru);
368 INIT_LIST_HEAD(&fbo->base.swap);
369 fbo->base.moving = NULL;
370 drm_vma_node_reset(&fbo->base.base.vma_node);
372 kref_init(&fbo->base.kref);
373 fbo->base.destroy = &ttm_transfered_destroy;
374 fbo->base.acc_size = 0;
375 if (bo->type != ttm_bo_type_sg)
376 fbo->base.base.resv = &fbo->base.base._resv;
378 dma_resv_init(&fbo->base.base._resv);
379 fbo->base.base.dev = NULL;
380 ret = dma_resv_trylock(&fbo->base.base._resv);
383 *new_obj = &fbo->base;
387 pgprot_t ttm_io_prot(uint32_t caching_flags, pgprot_t tmp)
389 /* Cached mappings need no adjustment */
390 if (caching_flags & TTM_PL_FLAG_CACHED)
393 #if defined(__i386__) || defined(__x86_64__)
394 if (caching_flags & TTM_PL_FLAG_WC)
395 tmp = pgprot_writecombine(tmp);
396 else if (boot_cpu_data.x86 > 3)
397 tmp = pgprot_noncached(tmp);
399 #if defined(__ia64__) || defined(__arm__) || defined(__aarch64__) || \
400 defined(__powerpc__) || defined(__mips__)
401 if (caching_flags & TTM_PL_FLAG_WC)
402 tmp = pgprot_writecombine(tmp);
404 tmp = pgprot_noncached(tmp);
406 #if defined(__sparc__)
407 tmp = pgprot_noncached(tmp);
411 EXPORT_SYMBOL(ttm_io_prot);
413 static int ttm_bo_ioremap(struct ttm_buffer_object *bo,
414 unsigned long offset,
416 struct ttm_bo_kmap_obj *map)
418 struct ttm_resource *mem = &bo->mem;
420 if (bo->mem.bus.addr) {
421 map->bo_kmap_type = ttm_bo_map_premapped;
422 map->virtual = (void *)(((u8 *)bo->mem.bus.addr) + offset);
424 map->bo_kmap_type = ttm_bo_map_iomap;
425 if (mem->placement & TTM_PL_FLAG_WC)
426 map->virtual = ioremap_wc(bo->mem.bus.offset + offset,
429 map->virtual = ioremap(bo->mem.bus.offset + offset,
432 return (!map->virtual) ? -ENOMEM : 0;
435 static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo,
436 unsigned long start_page,
437 unsigned long num_pages,
438 struct ttm_bo_kmap_obj *map)
440 struct ttm_resource *mem = &bo->mem;
441 struct ttm_operation_ctx ctx = {
442 .interruptible = false,
445 struct ttm_tt *ttm = bo->ttm;
451 ret = ttm_tt_populate(bo->bdev, ttm, &ctx);
455 if (num_pages == 1 && (mem->placement & TTM_PL_FLAG_CACHED)) {
457 * We're mapping a single page, and the desired
458 * page protection is consistent with the bo.
461 map->bo_kmap_type = ttm_bo_map_kmap;
462 map->page = ttm->pages[start_page];
463 map->virtual = kmap(map->page);
466 * We need to use vmap to get the desired page protection
467 * or to make the buffer object look contiguous.
469 prot = ttm_io_prot(mem->placement, PAGE_KERNEL);
470 map->bo_kmap_type = ttm_bo_map_vmap;
471 map->virtual = vmap(ttm->pages + start_page, num_pages,
474 return (!map->virtual) ? -ENOMEM : 0;
477 int ttm_bo_kmap(struct ttm_buffer_object *bo,
478 unsigned long start_page, unsigned long num_pages,
479 struct ttm_bo_kmap_obj *map)
481 unsigned long offset, size;
486 if (num_pages > bo->num_pages)
488 if (start_page > bo->num_pages)
491 ret = ttm_mem_io_reserve(bo->bdev, &bo->mem);
494 if (!bo->mem.bus.is_iomem) {
495 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map);
497 offset = start_page << PAGE_SHIFT;
498 size = num_pages << PAGE_SHIFT;
499 return ttm_bo_ioremap(bo, offset, size, map);
502 EXPORT_SYMBOL(ttm_bo_kmap);
504 void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map)
508 switch (map->bo_kmap_type) {
509 case ttm_bo_map_iomap:
510 iounmap(map->virtual);
512 case ttm_bo_map_vmap:
513 vunmap(map->virtual);
515 case ttm_bo_map_kmap:
518 case ttm_bo_map_premapped:
523 ttm_mem_io_free(map->bo->bdev, &map->bo->mem);
527 EXPORT_SYMBOL(ttm_bo_kunmap);
529 static int ttm_bo_wait_free_node(struct ttm_buffer_object *bo,
533 ret = ttm_bo_wait(bo, false, false);
538 ttm_bo_tt_destroy(bo);
539 ttm_bo_free_old_node(bo);
543 static int ttm_bo_move_to_ghost(struct ttm_buffer_object *bo,
544 struct dma_fence *fence,
547 struct ttm_buffer_object *ghost_obj;
551 * This should help pipeline ordinary buffer moves.
553 * Hang old buffer memory on a new buffer object,
554 * and leave it to be released when the GPU
555 * operation has completed.
558 dma_fence_put(bo->moving);
559 bo->moving = dma_fence_get(fence);
561 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
565 dma_resv_add_excl_fence(&ghost_obj->base._resv, fence);
568 * If we're not moving to fixed memory, the TTM object
569 * needs to stay alive. Otherwhise hang it on the ghost
570 * bo to be unbound and destroyed.
574 ghost_obj->ttm = NULL;
578 dma_resv_unlock(&ghost_obj->base._resv);
579 ttm_bo_put(ghost_obj);
583 static void ttm_bo_move_pipeline_evict(struct ttm_buffer_object *bo,
584 struct dma_fence *fence)
586 struct ttm_bo_device *bdev = bo->bdev;
587 struct ttm_resource_manager *from = ttm_manager_type(bdev, bo->mem.mem_type);
590 * BO doesn't have a TTM we need to bind/unbind. Just remember
591 * this eviction and free up the allocation
593 spin_lock(&from->move_lock);
594 if (!from->move || dma_fence_is_later(fence, from->move)) {
595 dma_fence_put(from->move);
596 from->move = dma_fence_get(fence);
598 spin_unlock(&from->move_lock);
600 ttm_bo_free_old_node(bo);
602 dma_fence_put(bo->moving);
603 bo->moving = dma_fence_get(fence);
606 int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
607 struct dma_fence *fence,
610 struct ttm_resource *new_mem)
612 struct ttm_bo_device *bdev = bo->bdev;
613 struct ttm_resource_manager *from = ttm_manager_type(bdev, bo->mem.mem_type);
614 struct ttm_resource_manager *man = ttm_manager_type(bdev, new_mem->mem_type);
617 dma_resv_add_excl_fence(bo->base.resv, fence);
619 ret = ttm_bo_move_to_ghost(bo, fence, man->use_tt);
620 else if (!from->use_tt && pipeline)
621 ttm_bo_move_pipeline_evict(bo, fence);
623 ret = ttm_bo_wait_free_node(bo, man->use_tt);
628 ttm_bo_assign_mem(bo, new_mem);
632 EXPORT_SYMBOL(ttm_bo_move_accel_cleanup);
634 int ttm_bo_pipeline_gutting(struct ttm_buffer_object *bo)
636 struct ttm_buffer_object *ghost;
639 ret = ttm_buffer_object_transfer(bo, &ghost);
643 ret = dma_resv_copy_fences(&ghost->base._resv, bo->base.resv);
644 /* Last resort, wait for the BO to be idle when we are OOM */
646 ttm_bo_wait(bo, false, false);
648 memset(&bo->mem, 0, sizeof(bo->mem));
649 bo->mem.mem_type = TTM_PL_SYSTEM;
652 dma_resv_unlock(&ghost->base._resv);