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 ttm_flag_masked(&old_mem->placement, TTM_PL_FLAG_SYSTEM,
74 old_mem->mem_type = TTM_PL_SYSTEM;
77 ret = ttm_tt_set_placement_caching(ttm, new_mem->placement);
78 if (unlikely(ret != 0))
81 if (new_mem->mem_type != TTM_PL_SYSTEM) {
82 ret = ttm_tt_bind(ttm, new_mem, ctx);
83 if (unlikely(ret != 0))
88 new_mem->mm_node = NULL;
92 EXPORT_SYMBOL(ttm_bo_move_ttm);
94 int ttm_mem_io_lock(struct ttm_resource_manager *man, bool interruptible)
96 if (likely(!man->use_io_reserve_lru))
100 return mutex_lock_interruptible(&man->io_reserve_mutex);
102 mutex_lock(&man->io_reserve_mutex);
106 void ttm_mem_io_unlock(struct ttm_resource_manager *man)
108 if (likely(!man->use_io_reserve_lru))
111 mutex_unlock(&man->io_reserve_mutex);
114 static int ttm_mem_io_evict(struct ttm_resource_manager *man)
116 struct ttm_buffer_object *bo;
118 bo = list_first_entry_or_null(&man->io_reserve_lru,
119 struct ttm_buffer_object,
124 list_del_init(&bo->io_reserve_lru);
125 ttm_bo_unmap_virtual_locked(bo);
129 int ttm_mem_io_reserve(struct ttm_bo_device *bdev,
130 struct ttm_resource *mem)
132 struct ttm_resource_manager *man = ttm_manager_type(bdev, mem->mem_type);
135 if (mem->bus.io_reserved_count++)
138 if (!bdev->driver->io_mem_reserve)
142 ret = bdev->driver->io_mem_reserve(bdev, mem);
143 if (ret == -ENOSPC) {
144 ret = ttm_mem_io_evict(man);
151 void ttm_mem_io_free(struct ttm_bo_device *bdev,
152 struct ttm_resource *mem)
154 if (--mem->bus.io_reserved_count)
157 if (!bdev->driver->io_mem_free)
160 bdev->driver->io_mem_free(bdev, mem);
163 int ttm_mem_io_reserve_vm(struct ttm_buffer_object *bo)
165 struct ttm_resource_manager *man = ttm_manager_type(bo->bdev, bo->mem.mem_type);
166 struct ttm_resource *mem = &bo->mem;
169 if (mem->bus.io_reserved_vm)
172 ret = ttm_mem_io_reserve(bo->bdev, mem);
173 if (unlikely(ret != 0))
175 mem->bus.io_reserved_vm = true;
176 if (man->use_io_reserve_lru)
177 list_add_tail(&bo->io_reserve_lru,
178 &man->io_reserve_lru);
182 void ttm_mem_io_free_vm(struct ttm_buffer_object *bo)
184 struct ttm_resource *mem = &bo->mem;
186 if (!mem->bus.io_reserved_vm)
189 mem->bus.io_reserved_vm = false;
190 list_del_init(&bo->io_reserve_lru);
191 ttm_mem_io_free(bo->bdev, mem);
194 static int ttm_resource_ioremap(struct ttm_bo_device *bdev,
195 struct ttm_resource *mem,
198 struct ttm_resource_manager *man = ttm_manager_type(bdev, mem->mem_type);
203 (void) ttm_mem_io_lock(man, false);
204 ret = ttm_mem_io_reserve(bdev, mem);
205 ttm_mem_io_unlock(man);
206 if (ret || !mem->bus.is_iomem)
210 addr = mem->bus.addr;
212 if (mem->placement & TTM_PL_FLAG_WC)
213 addr = ioremap_wc(mem->bus.base + mem->bus.offset,
216 addr = ioremap(mem->bus.base + mem->bus.offset,
219 (void) ttm_mem_io_lock(man, false);
220 ttm_mem_io_free(bdev, mem);
221 ttm_mem_io_unlock(man);
229 static void ttm_resource_iounmap(struct ttm_bo_device *bdev,
230 struct ttm_resource *mem,
233 struct ttm_resource_manager *man;
235 man = ttm_manager_type(bdev, mem->mem_type);
237 if (virtual && mem->bus.addr == NULL)
239 (void) ttm_mem_io_lock(man, false);
240 ttm_mem_io_free(bdev, mem);
241 ttm_mem_io_unlock(man);
244 static int ttm_copy_io_page(void *dst, void *src, unsigned long page)
247 (uint32_t *) ((unsigned long)dst + (page << PAGE_SHIFT));
249 (uint32_t *) ((unsigned long)src + (page << PAGE_SHIFT));
252 for (i = 0; i < PAGE_SIZE / sizeof(uint32_t); ++i)
253 iowrite32(ioread32(srcP++), dstP++);
257 static int ttm_copy_io_ttm_page(struct ttm_tt *ttm, void *src,
261 struct page *d = ttm->pages[page];
267 src = (void *)((unsigned long)src + (page << PAGE_SHIFT));
268 dst = kmap_atomic_prot(d, prot);
272 memcpy_fromio(dst, src, PAGE_SIZE);
279 static int ttm_copy_ttm_io_page(struct ttm_tt *ttm, void *dst,
283 struct page *s = ttm->pages[page];
289 dst = (void *)((unsigned long)dst + (page << PAGE_SHIFT));
290 src = kmap_atomic_prot(s, prot);
294 memcpy_toio(dst, src, PAGE_SIZE);
301 int ttm_bo_move_memcpy(struct ttm_buffer_object *bo,
302 struct ttm_operation_ctx *ctx,
303 struct ttm_resource *new_mem)
305 struct ttm_bo_device *bdev = bo->bdev;
306 struct ttm_resource_manager *man = ttm_manager_type(bdev, new_mem->mem_type);
307 struct ttm_tt *ttm = bo->ttm;
308 struct ttm_resource *old_mem = &bo->mem;
309 struct ttm_resource old_copy = *old_mem;
315 unsigned long add = 0;
318 ret = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu);
322 ret = ttm_resource_ioremap(bdev, old_mem, &old_iomap);
325 ret = ttm_resource_ioremap(bdev, new_mem, &new_iomap);
330 * Single TTM move. NOP.
332 if (old_iomap == NULL && new_iomap == NULL)
336 * Don't move nonexistent data. Clear destination instead.
338 if (old_iomap == NULL &&
339 (ttm == NULL || (ttm->state == tt_unpopulated &&
340 !(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)))) {
341 memset_io(new_iomap, 0, new_mem->num_pages*PAGE_SIZE);
346 * TTM might be null for moves within the same region.
349 ret = ttm_tt_populate(ttm, ctx);
357 if ((old_mem->mem_type == new_mem->mem_type) &&
358 (new_mem->start < old_mem->start + old_mem->size)) {
360 add = new_mem->num_pages - 1;
363 for (i = 0; i < new_mem->num_pages; ++i) {
364 page = i * dir + add;
365 if (old_iomap == NULL) {
366 pgprot_t prot = ttm_io_prot(old_mem->placement,
368 ret = ttm_copy_ttm_io_page(ttm, new_iomap, page,
370 } else if (new_iomap == NULL) {
371 pgprot_t prot = ttm_io_prot(new_mem->placement,
373 ret = ttm_copy_io_ttm_page(ttm, old_iomap, page,
376 ret = ttm_copy_io_page(new_iomap, old_iomap, page);
385 new_mem->mm_node = NULL;
393 ttm_resource_iounmap(bdev, old_mem, new_iomap);
395 ttm_resource_iounmap(bdev, &old_copy, old_iomap);
398 * On error, keep the mm node!
401 ttm_resource_free(bo, &old_copy);
404 EXPORT_SYMBOL(ttm_bo_move_memcpy);
406 static void ttm_transfered_destroy(struct ttm_buffer_object *bo)
408 struct ttm_transfer_obj *fbo;
410 fbo = container_of(bo, struct ttm_transfer_obj, base);
416 * ttm_buffer_object_transfer
418 * @bo: A pointer to a struct ttm_buffer_object.
419 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object,
420 * holding the data of @bo with the old placement.
422 * This is a utility function that may be called after an accelerated move
423 * has been scheduled. A new buffer object is created as a placeholder for
424 * the old data while it's being copied. When that buffer object is idle,
425 * it can be destroyed, releasing the space of the old placement.
430 static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
431 struct ttm_buffer_object **new_obj)
433 struct ttm_transfer_obj *fbo;
436 fbo = kmalloc(sizeof(*fbo), GFP_KERNEL);
441 fbo->base.mem.placement |= TTM_PL_FLAG_NO_EVICT;
447 * Fix up members that we shouldn't copy directly:
448 * TODO: Explicit member copy would probably be better here.
451 atomic_inc(&ttm_bo_glob.bo_count);
452 INIT_LIST_HEAD(&fbo->base.ddestroy);
453 INIT_LIST_HEAD(&fbo->base.lru);
454 INIT_LIST_HEAD(&fbo->base.swap);
455 INIT_LIST_HEAD(&fbo->base.io_reserve_lru);
456 fbo->base.moving = NULL;
457 drm_vma_node_reset(&fbo->base.base.vma_node);
459 kref_init(&fbo->base.kref);
460 fbo->base.destroy = &ttm_transfered_destroy;
461 fbo->base.acc_size = 0;
462 if (bo->type != ttm_bo_type_sg)
463 fbo->base.base.resv = &fbo->base.base._resv;
465 dma_resv_init(&fbo->base.base._resv);
466 fbo->base.base.dev = NULL;
467 ret = dma_resv_trylock(&fbo->base.base._resv);
470 *new_obj = &fbo->base;
474 pgprot_t ttm_io_prot(uint32_t caching_flags, pgprot_t tmp)
476 /* Cached mappings need no adjustment */
477 if (caching_flags & TTM_PL_FLAG_CACHED)
480 #if defined(__i386__) || defined(__x86_64__)
481 if (caching_flags & TTM_PL_FLAG_WC)
482 tmp = pgprot_writecombine(tmp);
483 else if (boot_cpu_data.x86 > 3)
484 tmp = pgprot_noncached(tmp);
486 #if defined(__ia64__) || defined(__arm__) || defined(__aarch64__) || \
487 defined(__powerpc__) || defined(__mips__)
488 if (caching_flags & TTM_PL_FLAG_WC)
489 tmp = pgprot_writecombine(tmp);
491 tmp = pgprot_noncached(tmp);
493 #if defined(__sparc__)
494 tmp = pgprot_noncached(tmp);
498 EXPORT_SYMBOL(ttm_io_prot);
500 static int ttm_bo_ioremap(struct ttm_buffer_object *bo,
501 unsigned long offset,
503 struct ttm_bo_kmap_obj *map)
505 struct ttm_resource *mem = &bo->mem;
507 if (bo->mem.bus.addr) {
508 map->bo_kmap_type = ttm_bo_map_premapped;
509 map->virtual = (void *)(((u8 *)bo->mem.bus.addr) + offset);
511 map->bo_kmap_type = ttm_bo_map_iomap;
512 if (mem->placement & TTM_PL_FLAG_WC)
513 map->virtual = ioremap_wc(bo->mem.bus.base +
514 bo->mem.bus.offset + offset,
517 map->virtual = ioremap(bo->mem.bus.base +
518 bo->mem.bus.offset + offset,
521 return (!map->virtual) ? -ENOMEM : 0;
524 static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo,
525 unsigned long start_page,
526 unsigned long num_pages,
527 struct ttm_bo_kmap_obj *map)
529 struct ttm_resource *mem = &bo->mem;
530 struct ttm_operation_ctx ctx = {
531 .interruptible = false,
538 ret = ttm_tt_create(bo, true);
543 ret = ttm_tt_populate(ttm, &ctx);
547 if (num_pages == 1 && (mem->placement & TTM_PL_FLAG_CACHED)) {
549 * We're mapping a single page, and the desired
550 * page protection is consistent with the bo.
553 map->bo_kmap_type = ttm_bo_map_kmap;
554 map->page = ttm->pages[start_page];
555 map->virtual = kmap(map->page);
558 * We need to use vmap to get the desired page protection
559 * or to make the buffer object look contiguous.
561 prot = ttm_io_prot(mem->placement, PAGE_KERNEL);
562 map->bo_kmap_type = ttm_bo_map_vmap;
563 map->virtual = vmap(ttm->pages + start_page, num_pages,
566 return (!map->virtual) ? -ENOMEM : 0;
569 int ttm_bo_kmap(struct ttm_buffer_object *bo,
570 unsigned long start_page, unsigned long num_pages,
571 struct ttm_bo_kmap_obj *map)
573 struct ttm_resource_manager *man =
574 ttm_manager_type(bo->bdev, bo->mem.mem_type);
575 unsigned long offset, size;
580 if (num_pages > bo->num_pages)
582 if (start_page > bo->num_pages)
585 (void) ttm_mem_io_lock(man, false);
586 ret = ttm_mem_io_reserve(bo->bdev, &bo->mem);
587 ttm_mem_io_unlock(man);
590 if (!bo->mem.bus.is_iomem) {
591 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map);
593 offset = start_page << PAGE_SHIFT;
594 size = num_pages << PAGE_SHIFT;
595 return ttm_bo_ioremap(bo, offset, size, map);
598 EXPORT_SYMBOL(ttm_bo_kmap);
600 void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map)
602 struct ttm_buffer_object *bo = map->bo;
603 struct ttm_resource_manager *man =
604 ttm_manager_type(bo->bdev, bo->mem.mem_type);
608 switch (map->bo_kmap_type) {
609 case ttm_bo_map_iomap:
610 iounmap(map->virtual);
612 case ttm_bo_map_vmap:
613 vunmap(map->virtual);
615 case ttm_bo_map_kmap:
618 case ttm_bo_map_premapped:
623 (void) ttm_mem_io_lock(man, false);
624 ttm_mem_io_free(map->bo->bdev, &map->bo->mem);
625 ttm_mem_io_unlock(man);
629 EXPORT_SYMBOL(ttm_bo_kunmap);
631 int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
632 struct dma_fence *fence,
634 struct ttm_resource *new_mem)
636 struct ttm_bo_device *bdev = bo->bdev;
637 struct ttm_resource_manager *man = ttm_manager_type(bdev, new_mem->mem_type);
638 struct ttm_resource *old_mem = &bo->mem;
640 struct ttm_buffer_object *ghost_obj;
642 dma_resv_add_excl_fence(bo->base.resv, fence);
644 ret = ttm_bo_wait(bo, false, false);
649 ttm_tt_destroy(bo->ttm);
652 ttm_bo_free_old_node(bo);
655 * This should help pipeline ordinary buffer moves.
657 * Hang old buffer memory on a new buffer object,
658 * and leave it to be released when the GPU
659 * operation has completed.
662 dma_fence_put(bo->moving);
663 bo->moving = dma_fence_get(fence);
665 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
669 dma_resv_add_excl_fence(&ghost_obj->base._resv, fence);
672 * If we're not moving to fixed memory, the TTM object
673 * needs to stay alive. Otherwhise hang it on the ghost
674 * bo to be unbound and destroyed.
678 ghost_obj->ttm = NULL;
682 dma_resv_unlock(&ghost_obj->base._resv);
683 ttm_bo_put(ghost_obj);
687 new_mem->mm_node = NULL;
691 EXPORT_SYMBOL(ttm_bo_move_accel_cleanup);
693 int ttm_bo_pipeline_move(struct ttm_buffer_object *bo,
694 struct dma_fence *fence, bool evict,
695 struct ttm_resource *new_mem)
697 struct ttm_bo_device *bdev = bo->bdev;
698 struct ttm_resource *old_mem = &bo->mem;
700 struct ttm_resource_manager *from = ttm_manager_type(bdev, old_mem->mem_type);
701 struct ttm_resource_manager *to = ttm_manager_type(bdev, new_mem->mem_type);
705 dma_resv_add_excl_fence(bo->base.resv, fence);
708 struct ttm_buffer_object *ghost_obj;
711 * This should help pipeline ordinary buffer moves.
713 * Hang old buffer memory on a new buffer object,
714 * and leave it to be released when the GPU
715 * operation has completed.
718 dma_fence_put(bo->moving);
719 bo->moving = dma_fence_get(fence);
721 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
725 dma_resv_add_excl_fence(&ghost_obj->base._resv, fence);
728 * If we're not moving to fixed memory, the TTM object
729 * needs to stay alive. Otherwhise hang it on the ghost
730 * bo to be unbound and destroyed.
734 ghost_obj->ttm = NULL;
738 dma_resv_unlock(&ghost_obj->base._resv);
739 ttm_bo_put(ghost_obj);
741 } else if (!from->use_tt) {
744 * BO doesn't have a TTM we need to bind/unbind. Just remember
745 * this eviction and free up the allocation
748 spin_lock(&from->move_lock);
749 if (!from->move || dma_fence_is_later(fence, from->move)) {
750 dma_fence_put(from->move);
751 from->move = dma_fence_get(fence);
753 spin_unlock(&from->move_lock);
755 ttm_bo_free_old_node(bo);
757 dma_fence_put(bo->moving);
758 bo->moving = dma_fence_get(fence);
762 * Last resort, wait for the move to be completed.
764 * Should never happen in pratice.
767 ret = ttm_bo_wait(bo, false, false);
772 ttm_tt_destroy(bo->ttm);
775 ttm_bo_free_old_node(bo);
779 new_mem->mm_node = NULL;
783 EXPORT_SYMBOL(ttm_bo_pipeline_move);
785 int ttm_bo_pipeline_gutting(struct ttm_buffer_object *bo)
787 struct ttm_buffer_object *ghost;
790 ret = ttm_buffer_object_transfer(bo, &ghost);
794 ret = dma_resv_copy_fences(&ghost->base._resv, bo->base.resv);
795 /* Last resort, wait for the BO to be idle when we are OOM */
797 ttm_bo_wait(bo, false, false);
799 memset(&bo->mem, 0, sizeof(bo->mem));
800 bo->mem.mem_type = TTM_PL_SYSTEM;
803 dma_resv_unlock(&ghost->base._resv);