1 /**************************************************************************
3 * Copyright (c) 2007-2009 VMware, Inc., Palo Alto, CA., USA
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
21 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
22 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
23 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
24 * USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **************************************************************************/
28 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
31 #include <drm/ttm/ttm_bo_driver.h>
32 #include <drm/ttm/ttm_placement.h>
33 #include <drm/drm_vma_manager.h>
35 #include <linux/highmem.h>
36 #include <linux/wait.h>
37 #include <linux/slab.h>
38 #include <linux/vmalloc.h>
39 #include <linux/module.h>
41 void ttm_bo_free_old_node(struct ttm_buffer_object *bo)
43 ttm_bo_mem_put(bo, &bo->mem);
46 int ttm_bo_move_ttm(struct ttm_buffer_object *bo,
48 bool no_wait_gpu, struct ttm_mem_reg *new_mem)
50 struct ttm_tt *ttm = bo->ttm;
51 struct ttm_mem_reg *old_mem = &bo->mem;
54 if (old_mem->mem_type != TTM_PL_SYSTEM) {
56 ttm_bo_free_old_node(bo);
57 ttm_flag_masked(&old_mem->placement, TTM_PL_FLAG_SYSTEM,
59 old_mem->mem_type = TTM_PL_SYSTEM;
62 ret = ttm_tt_set_placement_caching(ttm, new_mem->placement);
63 if (unlikely(ret != 0))
66 if (new_mem->mem_type != TTM_PL_SYSTEM) {
67 ret = ttm_tt_bind(ttm, new_mem);
68 if (unlikely(ret != 0))
73 new_mem->mm_node = NULL;
77 EXPORT_SYMBOL(ttm_bo_move_ttm);
79 int ttm_mem_io_lock(struct ttm_mem_type_manager *man, bool interruptible)
81 if (likely(man->io_reserve_fastpath))
85 return mutex_lock_interruptible(&man->io_reserve_mutex);
87 mutex_lock(&man->io_reserve_mutex);
90 EXPORT_SYMBOL(ttm_mem_io_lock);
92 void ttm_mem_io_unlock(struct ttm_mem_type_manager *man)
94 if (likely(man->io_reserve_fastpath))
97 mutex_unlock(&man->io_reserve_mutex);
99 EXPORT_SYMBOL(ttm_mem_io_unlock);
101 static int ttm_mem_io_evict(struct ttm_mem_type_manager *man)
103 struct ttm_buffer_object *bo;
105 if (!man->use_io_reserve_lru || list_empty(&man->io_reserve_lru))
108 bo = list_first_entry(&man->io_reserve_lru,
109 struct ttm_buffer_object,
111 list_del_init(&bo->io_reserve_lru);
112 ttm_bo_unmap_virtual_locked(bo);
118 int ttm_mem_io_reserve(struct ttm_bo_device *bdev,
119 struct ttm_mem_reg *mem)
121 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
124 if (!bdev->driver->io_mem_reserve)
126 if (likely(man->io_reserve_fastpath))
127 return bdev->driver->io_mem_reserve(bdev, mem);
129 if (bdev->driver->io_mem_reserve &&
130 mem->bus.io_reserved_count++ == 0) {
132 ret = bdev->driver->io_mem_reserve(bdev, mem);
133 if (ret == -EAGAIN) {
134 ret = ttm_mem_io_evict(man);
141 EXPORT_SYMBOL(ttm_mem_io_reserve);
143 void ttm_mem_io_free(struct ttm_bo_device *bdev,
144 struct ttm_mem_reg *mem)
146 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
148 if (likely(man->io_reserve_fastpath))
151 if (bdev->driver->io_mem_reserve &&
152 --mem->bus.io_reserved_count == 0 &&
153 bdev->driver->io_mem_free)
154 bdev->driver->io_mem_free(bdev, mem);
157 EXPORT_SYMBOL(ttm_mem_io_free);
159 int ttm_mem_io_reserve_vm(struct ttm_buffer_object *bo)
161 struct ttm_mem_reg *mem = &bo->mem;
164 if (!mem->bus.io_reserved_vm) {
165 struct ttm_mem_type_manager *man =
166 &bo->bdev->man[mem->mem_type];
168 ret = ttm_mem_io_reserve(bo->bdev, mem);
169 if (unlikely(ret != 0))
171 mem->bus.io_reserved_vm = true;
172 if (man->use_io_reserve_lru)
173 list_add_tail(&bo->io_reserve_lru,
174 &man->io_reserve_lru);
179 void ttm_mem_io_free_vm(struct ttm_buffer_object *bo)
181 struct ttm_mem_reg *mem = &bo->mem;
183 if (mem->bus.io_reserved_vm) {
184 mem->bus.io_reserved_vm = false;
185 list_del_init(&bo->io_reserve_lru);
186 ttm_mem_io_free(bo->bdev, mem);
190 int ttm_mem_reg_ioremap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem,
193 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
198 (void) ttm_mem_io_lock(man, false);
199 ret = ttm_mem_io_reserve(bdev, mem);
200 ttm_mem_io_unlock(man);
201 if (ret || !mem->bus.is_iomem)
205 addr = mem->bus.addr;
207 if (mem->placement & TTM_PL_FLAG_WC)
208 addr = ioremap_wc(mem->bus.base + mem->bus.offset, mem->bus.size);
210 addr = ioremap_nocache(mem->bus.base + mem->bus.offset, mem->bus.size);
212 (void) ttm_mem_io_lock(man, false);
213 ttm_mem_io_free(bdev, mem);
214 ttm_mem_io_unlock(man);
222 void ttm_mem_reg_iounmap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem,
225 struct ttm_mem_type_manager *man;
227 man = &bdev->man[mem->mem_type];
229 if (virtual && mem->bus.addr == NULL)
231 (void) ttm_mem_io_lock(man, false);
232 ttm_mem_io_free(bdev, mem);
233 ttm_mem_io_unlock(man);
236 static int ttm_copy_io_page(void *dst, void *src, unsigned long page)
239 (uint32_t *) ((unsigned long)dst + (page << PAGE_SHIFT));
241 (uint32_t *) ((unsigned long)src + (page << PAGE_SHIFT));
244 for (i = 0; i < PAGE_SIZE / sizeof(uint32_t); ++i)
245 iowrite32(ioread32(srcP++), dstP++);
249 static int ttm_copy_io_ttm_page(struct ttm_tt *ttm, void *src,
253 struct page *d = ttm->pages[page];
259 src = (void *)((unsigned long)src + (page << PAGE_SHIFT));
262 dst = kmap_atomic_prot(d, prot);
264 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
265 dst = vmap(&d, 1, 0, prot);
272 memcpy_fromio(dst, src, PAGE_SIZE);
277 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
286 static int ttm_copy_ttm_io_page(struct ttm_tt *ttm, void *dst,
290 struct page *s = ttm->pages[page];
296 dst = (void *)((unsigned long)dst + (page << PAGE_SHIFT));
298 src = kmap_atomic_prot(s, prot);
300 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
301 src = vmap(&s, 1, 0, prot);
308 memcpy_toio(dst, src, PAGE_SIZE);
313 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
322 int ttm_bo_move_memcpy(struct ttm_buffer_object *bo,
323 bool evict, bool no_wait_gpu,
324 struct ttm_mem_reg *new_mem)
326 struct ttm_bo_device *bdev = bo->bdev;
327 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
328 struct ttm_tt *ttm = bo->ttm;
329 struct ttm_mem_reg *old_mem = &bo->mem;
330 struct ttm_mem_reg old_copy = *old_mem;
336 unsigned long add = 0;
339 ret = ttm_mem_reg_ioremap(bdev, old_mem, &old_iomap);
342 ret = ttm_mem_reg_ioremap(bdev, new_mem, &new_iomap);
347 * Single TTM move. NOP.
349 if (old_iomap == NULL && new_iomap == NULL)
353 * Don't move nonexistent data. Clear destination instead.
355 if (old_iomap == NULL &&
356 (ttm == NULL || ttm->state == tt_unpopulated)) {
357 memset_io(new_iomap, 0, new_mem->num_pages*PAGE_SIZE);
362 * TTM might be null for moves within the same region.
364 if (ttm && ttm->state == tt_unpopulated) {
365 ret = ttm->bdev->driver->ttm_tt_populate(ttm);
373 if ((old_mem->mem_type == new_mem->mem_type) &&
374 (new_mem->start < old_mem->start + old_mem->size)) {
376 add = new_mem->num_pages - 1;
379 for (i = 0; i < new_mem->num_pages; ++i) {
380 page = i * dir + add;
381 if (old_iomap == NULL) {
382 pgprot_t prot = ttm_io_prot(old_mem->placement,
384 ret = ttm_copy_ttm_io_page(ttm, new_iomap, page,
386 } else if (new_iomap == NULL) {
387 pgprot_t prot = ttm_io_prot(new_mem->placement,
389 ret = ttm_copy_io_ttm_page(ttm, old_iomap, page,
392 ret = ttm_copy_io_page(new_iomap, old_iomap, page);
400 new_mem->mm_node = NULL;
402 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) && (ttm != NULL)) {
409 ttm_mem_reg_iounmap(bdev, old_mem, new_iomap);
411 ttm_mem_reg_iounmap(bdev, &old_copy, old_iomap);
414 * On error, keep the mm node!
417 ttm_bo_mem_put(bo, &old_copy);
420 EXPORT_SYMBOL(ttm_bo_move_memcpy);
422 static void ttm_transfered_destroy(struct ttm_buffer_object *bo)
428 * ttm_buffer_object_transfer
430 * @bo: A pointer to a struct ttm_buffer_object.
431 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object,
432 * holding the data of @bo with the old placement.
434 * This is a utility function that may be called after an accelerated move
435 * has been scheduled. A new buffer object is created as a placeholder for
436 * the old data while it's being copied. When that buffer object is idle,
437 * it can be destroyed, releasing the space of the old placement.
442 static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
443 struct ttm_buffer_object **new_obj)
445 struct ttm_buffer_object *fbo;
446 struct ttm_bo_device *bdev = bo->bdev;
447 struct ttm_bo_driver *driver = bdev->driver;
450 fbo = kmalloc(sizeof(*fbo), GFP_KERNEL);
457 * Fix up members that we shouldn't copy directly:
458 * TODO: Explicit member copy would probably be better here.
461 INIT_LIST_HEAD(&fbo->ddestroy);
462 INIT_LIST_HEAD(&fbo->lru);
463 INIT_LIST_HEAD(&fbo->swap);
464 INIT_LIST_HEAD(&fbo->io_reserve_lru);
465 drm_vma_node_reset(&fbo->vma_node);
466 atomic_set(&fbo->cpu_writers, 0);
468 spin_lock(&bdev->fence_lock);
470 fbo->sync_obj = driver->sync_obj_ref(bo->sync_obj);
472 fbo->sync_obj = NULL;
473 spin_unlock(&bdev->fence_lock);
474 kref_init(&fbo->list_kref);
475 kref_init(&fbo->kref);
476 fbo->destroy = &ttm_transfered_destroy;
478 fbo->resv = &fbo->ttm_resv;
479 reservation_object_init(fbo->resv);
480 ret = ww_mutex_trylock(&fbo->resv->lock);
487 pgprot_t ttm_io_prot(uint32_t caching_flags, pgprot_t tmp)
489 #if defined(__i386__) || defined(__x86_64__)
490 if (caching_flags & TTM_PL_FLAG_WC)
491 tmp = pgprot_writecombine(tmp);
492 else if (boot_cpu_data.x86 > 3)
493 tmp = pgprot_noncached(tmp);
495 #elif defined(__powerpc__)
496 if (!(caching_flags & TTM_PL_FLAG_CACHED)) {
497 pgprot_val(tmp) |= _PAGE_NO_CACHE;
498 if (caching_flags & TTM_PL_FLAG_UNCACHED)
499 pgprot_val(tmp) |= _PAGE_GUARDED;
502 #if defined(__ia64__)
503 if (caching_flags & TTM_PL_FLAG_WC)
504 tmp = pgprot_writecombine(tmp);
506 tmp = pgprot_noncached(tmp);
508 #if defined(__sparc__) || defined(__mips__)
509 if (!(caching_flags & TTM_PL_FLAG_CACHED))
510 tmp = pgprot_noncached(tmp);
514 EXPORT_SYMBOL(ttm_io_prot);
516 static int ttm_bo_ioremap(struct ttm_buffer_object *bo,
517 unsigned long offset,
519 struct ttm_bo_kmap_obj *map)
521 struct ttm_mem_reg *mem = &bo->mem;
523 if (bo->mem.bus.addr) {
524 map->bo_kmap_type = ttm_bo_map_premapped;
525 map->virtual = (void *)(((u8 *)bo->mem.bus.addr) + offset);
527 map->bo_kmap_type = ttm_bo_map_iomap;
528 if (mem->placement & TTM_PL_FLAG_WC)
529 map->virtual = ioremap_wc(bo->mem.bus.base + bo->mem.bus.offset + offset,
532 map->virtual = ioremap_nocache(bo->mem.bus.base + bo->mem.bus.offset + offset,
535 return (!map->virtual) ? -ENOMEM : 0;
538 static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo,
539 unsigned long start_page,
540 unsigned long num_pages,
541 struct ttm_bo_kmap_obj *map)
543 struct ttm_mem_reg *mem = &bo->mem; pgprot_t prot;
544 struct ttm_tt *ttm = bo->ttm;
549 if (ttm->state == tt_unpopulated) {
550 ret = ttm->bdev->driver->ttm_tt_populate(ttm);
555 if (num_pages == 1 && (mem->placement & TTM_PL_FLAG_CACHED)) {
557 * We're mapping a single page, and the desired
558 * page protection is consistent with the bo.
561 map->bo_kmap_type = ttm_bo_map_kmap;
562 map->page = ttm->pages[start_page];
563 map->virtual = kmap(map->page);
566 * We need to use vmap to get the desired page protection
567 * or to make the buffer object look contiguous.
569 prot = (mem->placement & TTM_PL_FLAG_CACHED) ?
571 ttm_io_prot(mem->placement, PAGE_KERNEL);
572 map->bo_kmap_type = ttm_bo_map_vmap;
573 map->virtual = vmap(ttm->pages + start_page, num_pages,
576 return (!map->virtual) ? -ENOMEM : 0;
579 int ttm_bo_kmap(struct ttm_buffer_object *bo,
580 unsigned long start_page, unsigned long num_pages,
581 struct ttm_bo_kmap_obj *map)
583 struct ttm_mem_type_manager *man =
584 &bo->bdev->man[bo->mem.mem_type];
585 unsigned long offset, size;
588 BUG_ON(!list_empty(&bo->swap));
591 if (num_pages > bo->num_pages)
593 if (start_page > bo->num_pages)
596 if (num_pages > 1 && !DRM_SUSER(DRM_CURPROC))
599 (void) ttm_mem_io_lock(man, false);
600 ret = ttm_mem_io_reserve(bo->bdev, &bo->mem);
601 ttm_mem_io_unlock(man);
604 if (!bo->mem.bus.is_iomem) {
605 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map);
607 offset = start_page << PAGE_SHIFT;
608 size = num_pages << PAGE_SHIFT;
609 return ttm_bo_ioremap(bo, offset, size, map);
612 EXPORT_SYMBOL(ttm_bo_kmap);
614 void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map)
616 struct ttm_buffer_object *bo = map->bo;
617 struct ttm_mem_type_manager *man =
618 &bo->bdev->man[bo->mem.mem_type];
622 switch (map->bo_kmap_type) {
623 case ttm_bo_map_iomap:
624 iounmap(map->virtual);
626 case ttm_bo_map_vmap:
627 vunmap(map->virtual);
629 case ttm_bo_map_kmap:
632 case ttm_bo_map_premapped:
637 (void) ttm_mem_io_lock(man, false);
638 ttm_mem_io_free(map->bo->bdev, &map->bo->mem);
639 ttm_mem_io_unlock(man);
643 EXPORT_SYMBOL(ttm_bo_kunmap);
645 int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
649 struct ttm_mem_reg *new_mem)
651 struct ttm_bo_device *bdev = bo->bdev;
652 struct ttm_bo_driver *driver = bdev->driver;
653 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
654 struct ttm_mem_reg *old_mem = &bo->mem;
656 struct ttm_buffer_object *ghost_obj;
657 void *tmp_obj = NULL;
659 spin_lock(&bdev->fence_lock);
661 tmp_obj = bo->sync_obj;
664 bo->sync_obj = driver->sync_obj_ref(sync_obj);
666 ret = ttm_bo_wait(bo, false, false, false);
667 spin_unlock(&bdev->fence_lock);
669 driver->sync_obj_unref(&tmp_obj);
673 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
675 ttm_tt_unbind(bo->ttm);
676 ttm_tt_destroy(bo->ttm);
679 ttm_bo_free_old_node(bo);
682 * This should help pipeline ordinary buffer moves.
684 * Hang old buffer memory on a new buffer object,
685 * and leave it to be released when the GPU
686 * operation has completed.
689 set_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
690 spin_unlock(&bdev->fence_lock);
692 driver->sync_obj_unref(&tmp_obj);
694 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
699 * If we're not moving to fixed memory, the TTM object
700 * needs to stay alive. Otherwhise hang it on the ghost
701 * bo to be unbound and destroyed.
704 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED))
705 ghost_obj->ttm = NULL;
709 ttm_bo_unreserve(ghost_obj);
710 ttm_bo_unref(&ghost_obj);
714 new_mem->mm_node = NULL;
718 EXPORT_SYMBOL(ttm_bo_move_accel_cleanup);