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>
34 #include <linux/highmem.h>
35 #include <linux/wait.h>
36 #include <linux/slab.h>
37 #include <linux/vmalloc.h>
38 #include <linux/module.h>
40 void ttm_bo_free_old_node(struct ttm_buffer_object *bo)
42 ttm_bo_mem_put(bo, &bo->mem);
45 int ttm_bo_move_ttm(struct ttm_buffer_object *bo,
47 bool no_wait_gpu, struct ttm_mem_reg *new_mem)
49 struct ttm_tt *ttm = bo->ttm;
50 struct ttm_mem_reg *old_mem = &bo->mem;
53 if (old_mem->mem_type != TTM_PL_SYSTEM) {
55 ttm_bo_free_old_node(bo);
56 ttm_flag_masked(&old_mem->placement, TTM_PL_FLAG_SYSTEM,
58 old_mem->mem_type = TTM_PL_SYSTEM;
61 ret = ttm_tt_set_placement_caching(ttm, new_mem->placement);
62 if (unlikely(ret != 0))
65 if (new_mem->mem_type != TTM_PL_SYSTEM) {
66 ret = ttm_tt_bind(ttm, new_mem);
67 if (unlikely(ret != 0))
72 new_mem->mm_node = NULL;
76 EXPORT_SYMBOL(ttm_bo_move_ttm);
78 int ttm_mem_io_lock(struct ttm_mem_type_manager *man, bool interruptible)
80 if (likely(man->io_reserve_fastpath))
84 return mutex_lock_interruptible(&man->io_reserve_mutex);
86 mutex_lock(&man->io_reserve_mutex);
89 EXPORT_SYMBOL(ttm_mem_io_lock);
91 void ttm_mem_io_unlock(struct ttm_mem_type_manager *man)
93 if (likely(man->io_reserve_fastpath))
96 mutex_unlock(&man->io_reserve_mutex);
98 EXPORT_SYMBOL(ttm_mem_io_unlock);
100 static int ttm_mem_io_evict(struct ttm_mem_type_manager *man)
102 struct ttm_buffer_object *bo;
104 if (!man->use_io_reserve_lru || list_empty(&man->io_reserve_lru))
107 bo = list_first_entry(&man->io_reserve_lru,
108 struct ttm_buffer_object,
110 list_del_init(&bo->io_reserve_lru);
111 ttm_bo_unmap_virtual_locked(bo);
117 int ttm_mem_io_reserve(struct ttm_bo_device *bdev,
118 struct ttm_mem_reg *mem)
120 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
123 if (!bdev->driver->io_mem_reserve)
125 if (likely(man->io_reserve_fastpath))
126 return bdev->driver->io_mem_reserve(bdev, mem);
128 if (bdev->driver->io_mem_reserve &&
129 mem->bus.io_reserved_count++ == 0) {
131 ret = bdev->driver->io_mem_reserve(bdev, mem);
132 if (ret == -EAGAIN) {
133 ret = ttm_mem_io_evict(man);
140 EXPORT_SYMBOL(ttm_mem_io_reserve);
142 void ttm_mem_io_free(struct ttm_bo_device *bdev,
143 struct ttm_mem_reg *mem)
145 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
147 if (likely(man->io_reserve_fastpath))
150 if (bdev->driver->io_mem_reserve &&
151 --mem->bus.io_reserved_count == 0 &&
152 bdev->driver->io_mem_free)
153 bdev->driver->io_mem_free(bdev, mem);
156 EXPORT_SYMBOL(ttm_mem_io_free);
158 int ttm_mem_io_reserve_vm(struct ttm_buffer_object *bo)
160 struct ttm_mem_reg *mem = &bo->mem;
163 if (!mem->bus.io_reserved_vm) {
164 struct ttm_mem_type_manager *man =
165 &bo->bdev->man[mem->mem_type];
167 ret = ttm_mem_io_reserve(bo->bdev, mem);
168 if (unlikely(ret != 0))
170 mem->bus.io_reserved_vm = true;
171 if (man->use_io_reserve_lru)
172 list_add_tail(&bo->io_reserve_lru,
173 &man->io_reserve_lru);
178 void ttm_mem_io_free_vm(struct ttm_buffer_object *bo)
180 struct ttm_mem_reg *mem = &bo->mem;
182 if (mem->bus.io_reserved_vm) {
183 mem->bus.io_reserved_vm = false;
184 list_del_init(&bo->io_reserve_lru);
185 ttm_mem_io_free(bo->bdev, mem);
189 int ttm_mem_reg_ioremap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem,
192 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
197 (void) ttm_mem_io_lock(man, false);
198 ret = ttm_mem_io_reserve(bdev, mem);
199 ttm_mem_io_unlock(man);
200 if (ret || !mem->bus.is_iomem)
204 addr = mem->bus.addr;
206 if (mem->placement & TTM_PL_FLAG_WC)
207 addr = ioremap_wc(mem->bus.base + mem->bus.offset, mem->bus.size);
209 addr = ioremap_nocache(mem->bus.base + mem->bus.offset, mem->bus.size);
211 (void) ttm_mem_io_lock(man, false);
212 ttm_mem_io_free(bdev, mem);
213 ttm_mem_io_unlock(man);
221 void ttm_mem_reg_iounmap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem,
224 struct ttm_mem_type_manager *man;
226 man = &bdev->man[mem->mem_type];
228 if (virtual && mem->bus.addr == NULL)
230 (void) ttm_mem_io_lock(man, false);
231 ttm_mem_io_free(bdev, mem);
232 ttm_mem_io_unlock(man);
235 static int ttm_copy_io_page(void *dst, void *src, unsigned long page)
238 (uint32_t *) ((unsigned long)dst + (page << PAGE_SHIFT));
240 (uint32_t *) ((unsigned long)src + (page << PAGE_SHIFT));
243 for (i = 0; i < PAGE_SIZE / sizeof(uint32_t); ++i)
244 iowrite32(ioread32(srcP++), dstP++);
248 static int ttm_copy_io_ttm_page(struct ttm_tt *ttm, void *src,
252 struct page *d = ttm->pages[page];
258 src = (void *)((unsigned long)src + (page << PAGE_SHIFT));
261 dst = kmap_atomic_prot(d, prot);
263 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
264 dst = vmap(&d, 1, 0, prot);
271 memcpy_fromio(dst, src, PAGE_SIZE);
276 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
285 static int ttm_copy_ttm_io_page(struct ttm_tt *ttm, void *dst,
289 struct page *s = ttm->pages[page];
295 dst = (void *)((unsigned long)dst + (page << PAGE_SHIFT));
297 src = kmap_atomic_prot(s, prot);
299 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
300 src = vmap(&s, 1, 0, prot);
307 memcpy_toio(dst, src, PAGE_SIZE);
312 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
321 int ttm_bo_move_memcpy(struct ttm_buffer_object *bo,
322 bool evict, bool no_wait_gpu,
323 struct ttm_mem_reg *new_mem)
325 struct ttm_bo_device *bdev = bo->bdev;
326 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
327 struct ttm_tt *ttm = bo->ttm;
328 struct ttm_mem_reg *old_mem = &bo->mem;
329 struct ttm_mem_reg old_copy = *old_mem;
335 unsigned long add = 0;
338 ret = ttm_mem_reg_ioremap(bdev, old_mem, &old_iomap);
341 ret = ttm_mem_reg_ioremap(bdev, new_mem, &new_iomap);
345 if (old_iomap == NULL && new_iomap == NULL)
347 if (old_iomap == NULL && ttm == NULL)
350 if (ttm->state == tt_unpopulated) {
351 ret = ttm->bdev->driver->ttm_tt_populate(ttm);
353 /* if we fail here don't nuke the mm node
354 * as the bo still owns it */
355 old_copy.mm_node = NULL;
363 if ((old_mem->mem_type == new_mem->mem_type) &&
364 (new_mem->start < old_mem->start + old_mem->size)) {
366 add = new_mem->num_pages - 1;
369 for (i = 0; i < new_mem->num_pages; ++i) {
370 page = i * dir + add;
371 if (old_iomap == NULL) {
372 pgprot_t prot = ttm_io_prot(old_mem->placement,
374 ret = ttm_copy_ttm_io_page(ttm, new_iomap, page,
376 } else if (new_iomap == NULL) {
377 pgprot_t prot = ttm_io_prot(new_mem->placement,
379 ret = ttm_copy_io_ttm_page(ttm, old_iomap, page,
382 ret = ttm_copy_io_page(new_iomap, old_iomap, page);
384 /* failing here, means keep old copy as-is */
385 old_copy.mm_node = NULL;
393 new_mem->mm_node = NULL;
395 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) && (ttm != NULL)) {
402 ttm_mem_reg_iounmap(bdev, old_mem, new_iomap);
404 ttm_mem_reg_iounmap(bdev, &old_copy, old_iomap);
405 ttm_bo_mem_put(bo, &old_copy);
408 EXPORT_SYMBOL(ttm_bo_move_memcpy);
410 static void ttm_transfered_destroy(struct ttm_buffer_object *bo)
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_buffer_object *fbo;
434 struct ttm_bo_device *bdev = bo->bdev;
435 struct ttm_bo_driver *driver = bdev->driver;
438 fbo = kmalloc(sizeof(*fbo), GFP_KERNEL);
445 * Fix up members that we shouldn't copy directly:
446 * TODO: Explicit member copy would probably be better here.
449 INIT_LIST_HEAD(&fbo->ddestroy);
450 INIT_LIST_HEAD(&fbo->lru);
451 INIT_LIST_HEAD(&fbo->swap);
452 INIT_LIST_HEAD(&fbo->io_reserve_lru);
454 atomic_set(&fbo->cpu_writers, 0);
456 spin_lock(&bdev->fence_lock);
458 fbo->sync_obj = driver->sync_obj_ref(bo->sync_obj);
460 fbo->sync_obj = NULL;
461 spin_unlock(&bdev->fence_lock);
462 kref_init(&fbo->list_kref);
463 kref_init(&fbo->kref);
464 fbo->destroy = &ttm_transfered_destroy;
466 fbo->resv = &fbo->ttm_resv;
467 reservation_object_init(fbo->resv);
468 ret = ww_mutex_trylock(&fbo->resv->lock);
475 pgprot_t ttm_io_prot(uint32_t caching_flags, pgprot_t tmp)
477 #if defined(__i386__) || defined(__x86_64__)
478 if (caching_flags & TTM_PL_FLAG_WC)
479 tmp = pgprot_writecombine(tmp);
480 else if (boot_cpu_data.x86 > 3)
481 tmp = pgprot_noncached(tmp);
483 #elif defined(__powerpc__)
484 if (!(caching_flags & TTM_PL_FLAG_CACHED)) {
485 pgprot_val(tmp) |= _PAGE_NO_CACHE;
486 if (caching_flags & TTM_PL_FLAG_UNCACHED)
487 pgprot_val(tmp) |= _PAGE_GUARDED;
490 #if defined(__ia64__)
491 if (caching_flags & TTM_PL_FLAG_WC)
492 tmp = pgprot_writecombine(tmp);
494 tmp = pgprot_noncached(tmp);
496 #if defined(__sparc__) || defined(__mips__)
497 if (!(caching_flags & TTM_PL_FLAG_CACHED))
498 tmp = pgprot_noncached(tmp);
502 EXPORT_SYMBOL(ttm_io_prot);
504 static int ttm_bo_ioremap(struct ttm_buffer_object *bo,
505 unsigned long offset,
507 struct ttm_bo_kmap_obj *map)
509 struct ttm_mem_reg *mem = &bo->mem;
511 if (bo->mem.bus.addr) {
512 map->bo_kmap_type = ttm_bo_map_premapped;
513 map->virtual = (void *)(((u8 *)bo->mem.bus.addr) + offset);
515 map->bo_kmap_type = ttm_bo_map_iomap;
516 if (mem->placement & TTM_PL_FLAG_WC)
517 map->virtual = ioremap_wc(bo->mem.bus.base + bo->mem.bus.offset + offset,
520 map->virtual = ioremap_nocache(bo->mem.bus.base + bo->mem.bus.offset + offset,
523 return (!map->virtual) ? -ENOMEM : 0;
526 static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo,
527 unsigned long start_page,
528 unsigned long num_pages,
529 struct ttm_bo_kmap_obj *map)
531 struct ttm_mem_reg *mem = &bo->mem; pgprot_t prot;
532 struct ttm_tt *ttm = bo->ttm;
537 if (ttm->state == tt_unpopulated) {
538 ret = ttm->bdev->driver->ttm_tt_populate(ttm);
543 if (num_pages == 1 && (mem->placement & TTM_PL_FLAG_CACHED)) {
545 * We're mapping a single page, and the desired
546 * page protection is consistent with the bo.
549 map->bo_kmap_type = ttm_bo_map_kmap;
550 map->page = ttm->pages[start_page];
551 map->virtual = kmap(map->page);
554 * We need to use vmap to get the desired page protection
555 * or to make the buffer object look contiguous.
557 prot = (mem->placement & TTM_PL_FLAG_CACHED) ?
559 ttm_io_prot(mem->placement, PAGE_KERNEL);
560 map->bo_kmap_type = ttm_bo_map_vmap;
561 map->virtual = vmap(ttm->pages + start_page, num_pages,
564 return (!map->virtual) ? -ENOMEM : 0;
567 int ttm_bo_kmap(struct ttm_buffer_object *bo,
568 unsigned long start_page, unsigned long num_pages,
569 struct ttm_bo_kmap_obj *map)
571 struct ttm_mem_type_manager *man =
572 &bo->bdev->man[bo->mem.mem_type];
573 unsigned long offset, size;
576 BUG_ON(!list_empty(&bo->swap));
579 if (num_pages > bo->num_pages)
581 if (start_page > bo->num_pages)
584 if (num_pages > 1 && !DRM_SUSER(DRM_CURPROC))
587 (void) ttm_mem_io_lock(man, false);
588 ret = ttm_mem_io_reserve(bo->bdev, &bo->mem);
589 ttm_mem_io_unlock(man);
592 if (!bo->mem.bus.is_iomem) {
593 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map);
595 offset = start_page << PAGE_SHIFT;
596 size = num_pages << PAGE_SHIFT;
597 return ttm_bo_ioremap(bo, offset, size, map);
600 EXPORT_SYMBOL(ttm_bo_kmap);
602 void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map)
604 struct ttm_buffer_object *bo = map->bo;
605 struct ttm_mem_type_manager *man =
606 &bo->bdev->man[bo->mem.mem_type];
610 switch (map->bo_kmap_type) {
611 case ttm_bo_map_iomap:
612 iounmap(map->virtual);
614 case ttm_bo_map_vmap:
615 vunmap(map->virtual);
617 case ttm_bo_map_kmap:
620 case ttm_bo_map_premapped:
625 (void) ttm_mem_io_lock(man, false);
626 ttm_mem_io_free(map->bo->bdev, &map->bo->mem);
627 ttm_mem_io_unlock(man);
631 EXPORT_SYMBOL(ttm_bo_kunmap);
633 int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
637 struct ttm_mem_reg *new_mem)
639 struct ttm_bo_device *bdev = bo->bdev;
640 struct ttm_bo_driver *driver = bdev->driver;
641 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
642 struct ttm_mem_reg *old_mem = &bo->mem;
644 struct ttm_buffer_object *ghost_obj;
645 void *tmp_obj = NULL;
647 spin_lock(&bdev->fence_lock);
649 tmp_obj = bo->sync_obj;
652 bo->sync_obj = driver->sync_obj_ref(sync_obj);
654 ret = ttm_bo_wait(bo, false, false, false);
655 spin_unlock(&bdev->fence_lock);
657 driver->sync_obj_unref(&tmp_obj);
661 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
663 ttm_tt_unbind(bo->ttm);
664 ttm_tt_destroy(bo->ttm);
667 ttm_bo_free_old_node(bo);
670 * This should help pipeline ordinary buffer moves.
672 * Hang old buffer memory on a new buffer object,
673 * and leave it to be released when the GPU
674 * operation has completed.
677 set_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
678 spin_unlock(&bdev->fence_lock);
680 driver->sync_obj_unref(&tmp_obj);
682 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
687 * If we're not moving to fixed memory, the TTM object
688 * needs to stay alive. Otherwhise hang it on the ghost
689 * bo to be unbound and destroyed.
692 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED))
693 ghost_obj->ttm = NULL;
697 ttm_bo_unreserve(ghost_obj);
698 ttm_bo_unref(&ghost_obj);
702 new_mem->mm_node = NULL;
706 EXPORT_SYMBOL(ttm_bo_move_accel_cleanup);