1 /**************************************************************************
3 * Copyright (c) 2006-2009 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 "ttm/ttm_module.h"
32 #include "ttm/ttm_bo_driver.h"
33 #include "ttm/ttm_placement.h"
34 #include <linux/jiffies.h>
35 #include <linux/slab.h>
36 #include <linux/sched.h>
38 #include <linux/file.h>
39 #include <linux/module.h>
40 #include <linux/atomic.h>
42 #define TTM_ASSERT_LOCKED(param)
43 #define TTM_DEBUG(fmt, arg...)
44 #define TTM_BO_HASH_ORDER 13
46 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo);
47 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink);
48 static void ttm_bo_global_kobj_release(struct kobject *kobj);
50 static struct attribute ttm_bo_count = {
55 static inline int ttm_mem_type_from_flags(uint32_t flags, uint32_t *mem_type)
59 for (i = 0; i <= TTM_PL_PRIV5; i++)
60 if (flags & (1 << i)) {
67 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type)
69 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
71 printk(KERN_ERR TTM_PFX " has_type: %d\n", man->has_type);
72 printk(KERN_ERR TTM_PFX " use_type: %d\n", man->use_type);
73 printk(KERN_ERR TTM_PFX " flags: 0x%08X\n", man->flags);
74 printk(KERN_ERR TTM_PFX " gpu_offset: 0x%08lX\n", man->gpu_offset);
75 printk(KERN_ERR TTM_PFX " size: %llu\n", man->size);
76 printk(KERN_ERR TTM_PFX " available_caching: 0x%08X\n",
77 man->available_caching);
78 printk(KERN_ERR TTM_PFX " default_caching: 0x%08X\n",
79 man->default_caching);
80 if (mem_type != TTM_PL_SYSTEM)
81 (*man->func->debug)(man, TTM_PFX);
84 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
85 struct ttm_placement *placement)
89 printk(KERN_ERR TTM_PFX "No space for %p (%lu pages, %luK, %luM)\n",
90 bo, bo->mem.num_pages, bo->mem.size >> 10,
92 for (i = 0; i < placement->num_placement; i++) {
93 ret = ttm_mem_type_from_flags(placement->placement[i],
97 printk(KERN_ERR TTM_PFX " placement[%d]=0x%08X (%d)\n",
98 i, placement->placement[i], mem_type);
99 ttm_mem_type_debug(bo->bdev, mem_type);
103 static ssize_t ttm_bo_global_show(struct kobject *kobj,
104 struct attribute *attr,
107 struct ttm_bo_global *glob =
108 container_of(kobj, struct ttm_bo_global, kobj);
110 return snprintf(buffer, PAGE_SIZE, "%lu\n",
111 (unsigned long) atomic_read(&glob->bo_count));
114 static struct attribute *ttm_bo_global_attrs[] = {
119 static const struct sysfs_ops ttm_bo_global_ops = {
120 .show = &ttm_bo_global_show
123 static struct kobj_type ttm_bo_glob_kobj_type = {
124 .release = &ttm_bo_global_kobj_release,
125 .sysfs_ops = &ttm_bo_global_ops,
126 .default_attrs = ttm_bo_global_attrs
130 static inline uint32_t ttm_bo_type_flags(unsigned type)
135 static void ttm_bo_release_list(struct kref *list_kref)
137 struct ttm_buffer_object *bo =
138 container_of(list_kref, struct ttm_buffer_object, list_kref);
139 struct ttm_bo_device *bdev = bo->bdev;
141 BUG_ON(atomic_read(&bo->list_kref.refcount));
142 BUG_ON(atomic_read(&bo->kref.refcount));
143 BUG_ON(atomic_read(&bo->cpu_writers));
144 BUG_ON(bo->sync_obj != NULL);
145 BUG_ON(bo->mem.mm_node != NULL);
146 BUG_ON(!list_empty(&bo->lru));
147 BUG_ON(!list_empty(&bo->ddestroy));
150 ttm_tt_destroy(bo->ttm);
151 atomic_dec(&bo->glob->bo_count);
155 ttm_mem_global_free(bdev->glob->mem_glob, bo->acc_size);
160 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo, bool interruptible)
163 return wait_event_interruptible(bo->event_queue,
164 atomic_read(&bo->reserved) == 0);
166 wait_event(bo->event_queue, atomic_read(&bo->reserved) == 0);
170 EXPORT_SYMBOL(ttm_bo_wait_unreserved);
172 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
174 struct ttm_bo_device *bdev = bo->bdev;
175 struct ttm_mem_type_manager *man;
177 BUG_ON(!atomic_read(&bo->reserved));
179 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
181 BUG_ON(!list_empty(&bo->lru));
183 man = &bdev->man[bo->mem.mem_type];
184 list_add_tail(&bo->lru, &man->lru);
185 kref_get(&bo->list_kref);
187 if (bo->ttm != NULL) {
188 list_add_tail(&bo->swap, &bo->glob->swap_lru);
189 kref_get(&bo->list_kref);
194 int ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
198 if (!list_empty(&bo->swap)) {
199 list_del_init(&bo->swap);
202 if (!list_empty(&bo->lru)) {
203 list_del_init(&bo->lru);
208 * TODO: Add a driver hook to delete from
209 * driver-specific LRU's here.
215 int ttm_bo_reserve_locked(struct ttm_buffer_object *bo,
217 bool no_wait, bool use_sequence, uint32_t sequence)
219 struct ttm_bo_global *glob = bo->glob;
222 while (unlikely(atomic_cmpxchg(&bo->reserved, 0, 1) != 0)) {
224 * Deadlock avoidance for multi-bo reserving.
226 if (use_sequence && bo->seq_valid) {
228 * We've already reserved this one.
230 if (unlikely(sequence == bo->val_seq))
233 * Already reserved by a thread that will not back
234 * off for us. We need to back off.
236 if (unlikely(sequence - bo->val_seq < (1 << 31)))
243 spin_unlock(&glob->lru_lock);
244 ret = ttm_bo_wait_unreserved(bo, interruptible);
245 spin_lock(&glob->lru_lock);
253 * Wake up waiters that may need to recheck for deadlock,
254 * if we decreased the sequence number.
256 if (unlikely((bo->val_seq - sequence < (1 << 31))
258 wake_up_all(&bo->event_queue);
260 bo->val_seq = sequence;
261 bo->seq_valid = true;
263 bo->seq_valid = false;
268 EXPORT_SYMBOL(ttm_bo_reserve);
270 static void ttm_bo_ref_bug(struct kref *list_kref)
275 void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count,
278 kref_sub(&bo->list_kref, count,
279 (never_free) ? ttm_bo_ref_bug : ttm_bo_release_list);
282 int ttm_bo_reserve(struct ttm_buffer_object *bo,
284 bool no_wait, bool use_sequence, uint32_t sequence)
286 struct ttm_bo_global *glob = bo->glob;
290 spin_lock(&glob->lru_lock);
291 ret = ttm_bo_reserve_locked(bo, interruptible, no_wait, use_sequence,
293 if (likely(ret == 0))
294 put_count = ttm_bo_del_from_lru(bo);
295 spin_unlock(&glob->lru_lock);
297 ttm_bo_list_ref_sub(bo, put_count, true);
302 void ttm_bo_unreserve_locked(struct ttm_buffer_object *bo)
304 ttm_bo_add_to_lru(bo);
305 atomic_set(&bo->reserved, 0);
306 wake_up_all(&bo->event_queue);
309 void ttm_bo_unreserve(struct ttm_buffer_object *bo)
311 struct ttm_bo_global *glob = bo->glob;
313 spin_lock(&glob->lru_lock);
314 ttm_bo_unreserve_locked(bo);
315 spin_unlock(&glob->lru_lock);
317 EXPORT_SYMBOL(ttm_bo_unreserve);
320 * Call bo->mutex locked.
322 static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
324 struct ttm_bo_device *bdev = bo->bdev;
325 struct ttm_bo_global *glob = bo->glob;
327 uint32_t page_flags = 0;
329 TTM_ASSERT_LOCKED(&bo->mutex);
332 if (bdev->need_dma32)
333 page_flags |= TTM_PAGE_FLAG_DMA32;
336 case ttm_bo_type_device:
338 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
339 case ttm_bo_type_kernel:
340 bo->ttm = ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
341 page_flags, glob->dummy_read_page);
342 if (unlikely(bo->ttm == NULL))
346 printk(KERN_ERR TTM_PFX "Illegal buffer object type\n");
354 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
355 struct ttm_mem_reg *mem,
356 bool evict, bool interruptible,
357 bool no_wait_reserve, bool no_wait_gpu)
359 struct ttm_bo_device *bdev = bo->bdev;
360 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
361 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
362 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
363 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
366 if (old_is_pci || new_is_pci ||
367 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
368 ret = ttm_mem_io_lock(old_man, true);
369 if (unlikely(ret != 0))
371 ttm_bo_unmap_virtual_locked(bo);
372 ttm_mem_io_unlock(old_man);
376 * Create and bind a ttm if required.
379 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
380 if (bo->ttm == NULL) {
381 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
382 ret = ttm_bo_add_ttm(bo, zero);
387 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
391 if (mem->mem_type != TTM_PL_SYSTEM) {
392 ret = ttm_tt_bind(bo->ttm, mem);
397 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
398 if (bdev->driver->move_notify)
399 bdev->driver->move_notify(bo, mem);
406 if (bdev->driver->move_notify)
407 bdev->driver->move_notify(bo, mem);
409 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
410 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
411 ret = ttm_bo_move_ttm(bo, evict, no_wait_reserve, no_wait_gpu, mem);
412 else if (bdev->driver->move)
413 ret = bdev->driver->move(bo, evict, interruptible,
414 no_wait_reserve, no_wait_gpu, mem);
416 ret = ttm_bo_move_memcpy(bo, evict, no_wait_reserve, no_wait_gpu, mem);
423 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
425 printk(KERN_ERR TTM_PFX "Can not flush read caches\n");
429 if (bo->mem.mm_node) {
430 bo->offset = (bo->mem.start << PAGE_SHIFT) +
431 bdev->man[bo->mem.mem_type].gpu_offset;
432 bo->cur_placement = bo->mem.placement;
439 new_man = &bdev->man[bo->mem.mem_type];
440 if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) {
441 ttm_tt_unbind(bo->ttm);
442 ttm_tt_destroy(bo->ttm);
451 * Will release GPU memory type usage on destruction.
452 * This is the place to put in driver specific hooks to release
453 * driver private resources.
454 * Will release the bo::reserved lock.
457 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
460 ttm_tt_unbind(bo->ttm);
461 ttm_tt_destroy(bo->ttm);
464 ttm_bo_mem_put(bo, &bo->mem);
466 atomic_set(&bo->reserved, 0);
469 * Make processes trying to reserve really pick it up.
471 smp_mb__after_atomic_dec();
472 wake_up_all(&bo->event_queue);
475 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
477 struct ttm_bo_device *bdev = bo->bdev;
478 struct ttm_bo_global *glob = bo->glob;
479 struct ttm_bo_driver *driver;
480 void *sync_obj = NULL;
485 spin_lock(&bdev->fence_lock);
486 (void) ttm_bo_wait(bo, false, false, true);
489 spin_lock(&glob->lru_lock);
492 * Lock inversion between bo:reserve and bdev::fence_lock here,
493 * but that's OK, since we're only trylocking.
496 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
498 if (unlikely(ret == -EBUSY))
501 spin_unlock(&bdev->fence_lock);
502 put_count = ttm_bo_del_from_lru(bo);
504 spin_unlock(&glob->lru_lock);
505 ttm_bo_cleanup_memtype_use(bo);
507 ttm_bo_list_ref_sub(bo, put_count, true);
511 spin_lock(&glob->lru_lock);
514 driver = bdev->driver;
516 sync_obj = driver->sync_obj_ref(bo->sync_obj);
517 sync_obj_arg = bo->sync_obj_arg;
519 kref_get(&bo->list_kref);
520 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
521 spin_unlock(&glob->lru_lock);
522 spin_unlock(&bdev->fence_lock);
525 driver->sync_obj_flush(sync_obj, sync_obj_arg);
526 driver->sync_obj_unref(&sync_obj);
528 schedule_delayed_work(&bdev->wq,
529 ((HZ / 100) < 1) ? 1 : HZ / 100);
533 * function ttm_bo_cleanup_refs
534 * If bo idle, remove from delayed- and lru lists, and unref.
535 * If not idle, do nothing.
537 * @interruptible Any sleeps should occur interruptibly.
538 * @no_wait_reserve Never wait for reserve. Return -EBUSY instead.
539 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
542 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
544 bool no_wait_reserve,
547 struct ttm_bo_device *bdev = bo->bdev;
548 struct ttm_bo_global *glob = bo->glob;
553 spin_lock(&bdev->fence_lock);
554 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
555 spin_unlock(&bdev->fence_lock);
557 if (unlikely(ret != 0))
560 spin_lock(&glob->lru_lock);
561 ret = ttm_bo_reserve_locked(bo, interruptible,
562 no_wait_reserve, false, 0);
564 if (unlikely(ret != 0) || list_empty(&bo->ddestroy)) {
565 spin_unlock(&glob->lru_lock);
570 * We can re-check for sync object without taking
571 * the bo::lock since setting the sync object requires
572 * also bo::reserved. A busy object at this point may
573 * be caused by another thread recently starting an accelerated
577 if (unlikely(bo->sync_obj)) {
578 atomic_set(&bo->reserved, 0);
579 wake_up_all(&bo->event_queue);
580 spin_unlock(&glob->lru_lock);
584 put_count = ttm_bo_del_from_lru(bo);
585 list_del_init(&bo->ddestroy);
588 spin_unlock(&glob->lru_lock);
589 ttm_bo_cleanup_memtype_use(bo);
591 ttm_bo_list_ref_sub(bo, put_count, true);
597 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
598 * encountered buffers.
601 static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
603 struct ttm_bo_global *glob = bdev->glob;
604 struct ttm_buffer_object *entry = NULL;
607 spin_lock(&glob->lru_lock);
608 if (list_empty(&bdev->ddestroy))
611 entry = list_first_entry(&bdev->ddestroy,
612 struct ttm_buffer_object, ddestroy);
613 kref_get(&entry->list_kref);
616 struct ttm_buffer_object *nentry = NULL;
618 if (entry->ddestroy.next != &bdev->ddestroy) {
619 nentry = list_first_entry(&entry->ddestroy,
620 struct ttm_buffer_object, ddestroy);
621 kref_get(&nentry->list_kref);
624 spin_unlock(&glob->lru_lock);
625 ret = ttm_bo_cleanup_refs(entry, false, !remove_all,
627 kref_put(&entry->list_kref, ttm_bo_release_list);
633 spin_lock(&glob->lru_lock);
634 if (list_empty(&entry->ddestroy))
639 spin_unlock(&glob->lru_lock);
642 kref_put(&entry->list_kref, ttm_bo_release_list);
646 static void ttm_bo_delayed_workqueue(struct work_struct *work)
648 struct ttm_bo_device *bdev =
649 container_of(work, struct ttm_bo_device, wq.work);
651 if (ttm_bo_delayed_delete(bdev, false)) {
652 schedule_delayed_work(&bdev->wq,
653 ((HZ / 100) < 1) ? 1 : HZ / 100);
657 static void ttm_bo_release(struct kref *kref)
659 struct ttm_buffer_object *bo =
660 container_of(kref, struct ttm_buffer_object, kref);
661 struct ttm_bo_device *bdev = bo->bdev;
662 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
664 if (likely(bo->vm_node != NULL)) {
665 rb_erase(&bo->vm_rb, &bdev->addr_space_rb);
666 drm_mm_put_block(bo->vm_node);
669 write_unlock(&bdev->vm_lock);
670 ttm_mem_io_lock(man, false);
671 ttm_mem_io_free_vm(bo);
672 ttm_mem_io_unlock(man);
673 ttm_bo_cleanup_refs_or_queue(bo);
674 kref_put(&bo->list_kref, ttm_bo_release_list);
675 write_lock(&bdev->vm_lock);
678 void ttm_bo_unref(struct ttm_buffer_object **p_bo)
680 struct ttm_buffer_object *bo = *p_bo;
681 struct ttm_bo_device *bdev = bo->bdev;
684 write_lock(&bdev->vm_lock);
685 kref_put(&bo->kref, ttm_bo_release);
686 write_unlock(&bdev->vm_lock);
688 EXPORT_SYMBOL(ttm_bo_unref);
690 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
692 return cancel_delayed_work_sync(&bdev->wq);
694 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
696 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
699 schedule_delayed_work(&bdev->wq,
700 ((HZ / 100) < 1) ? 1 : HZ / 100);
702 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
704 static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
705 bool no_wait_reserve, bool no_wait_gpu)
707 struct ttm_bo_device *bdev = bo->bdev;
708 struct ttm_mem_reg evict_mem;
709 struct ttm_placement placement;
712 spin_lock(&bdev->fence_lock);
713 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
714 spin_unlock(&bdev->fence_lock);
716 if (unlikely(ret != 0)) {
717 if (ret != -ERESTARTSYS) {
718 printk(KERN_ERR TTM_PFX
719 "Failed to expire sync object before "
720 "buffer eviction.\n");
725 BUG_ON(!atomic_read(&bo->reserved));
728 evict_mem.mm_node = NULL;
729 evict_mem.bus.io_reserved_vm = false;
730 evict_mem.bus.io_reserved_count = 0;
734 placement.num_placement = 0;
735 placement.num_busy_placement = 0;
736 bdev->driver->evict_flags(bo, &placement);
737 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
738 no_wait_reserve, no_wait_gpu);
740 if (ret != -ERESTARTSYS) {
741 printk(KERN_ERR TTM_PFX
742 "Failed to find memory space for "
743 "buffer 0x%p eviction.\n", bo);
744 ttm_bo_mem_space_debug(bo, &placement);
749 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
750 no_wait_reserve, no_wait_gpu);
752 if (ret != -ERESTARTSYS)
753 printk(KERN_ERR TTM_PFX "Buffer eviction failed\n");
754 ttm_bo_mem_put(bo, &evict_mem);
762 static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
764 bool interruptible, bool no_wait_reserve,
767 struct ttm_bo_global *glob = bdev->glob;
768 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
769 struct ttm_buffer_object *bo;
770 int ret, put_count = 0;
773 spin_lock(&glob->lru_lock);
774 if (list_empty(&man->lru)) {
775 spin_unlock(&glob->lru_lock);
779 bo = list_first_entry(&man->lru, struct ttm_buffer_object, lru);
780 kref_get(&bo->list_kref);
782 if (!list_empty(&bo->ddestroy)) {
783 spin_unlock(&glob->lru_lock);
784 ret = ttm_bo_cleanup_refs(bo, interruptible,
785 no_wait_reserve, no_wait_gpu);
786 kref_put(&bo->list_kref, ttm_bo_release_list);
788 if (likely(ret == 0 || ret == -ERESTARTSYS))
794 ret = ttm_bo_reserve_locked(bo, false, no_wait_reserve, false, 0);
796 if (unlikely(ret == -EBUSY)) {
797 spin_unlock(&glob->lru_lock);
798 if (likely(!no_wait_gpu))
799 ret = ttm_bo_wait_unreserved(bo, interruptible);
801 kref_put(&bo->list_kref, ttm_bo_release_list);
804 * We *need* to retry after releasing the lru lock.
807 if (unlikely(ret != 0))
812 put_count = ttm_bo_del_from_lru(bo);
813 spin_unlock(&glob->lru_lock);
817 ttm_bo_list_ref_sub(bo, put_count, true);
819 ret = ttm_bo_evict(bo, interruptible, no_wait_reserve, no_wait_gpu);
820 ttm_bo_unreserve(bo);
822 kref_put(&bo->list_kref, ttm_bo_release_list);
826 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
828 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
831 (*man->func->put_node)(man, mem);
833 EXPORT_SYMBOL(ttm_bo_mem_put);
836 * Repeatedly evict memory from the LRU for @mem_type until we create enough
837 * space, or we've evicted everything and there isn't enough space.
839 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
841 struct ttm_placement *placement,
842 struct ttm_mem_reg *mem,
844 bool no_wait_reserve,
847 struct ttm_bo_device *bdev = bo->bdev;
848 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
852 ret = (*man->func->get_node)(man, bo, placement, mem);
853 if (unlikely(ret != 0))
857 ret = ttm_mem_evict_first(bdev, mem_type, interruptible,
858 no_wait_reserve, no_wait_gpu);
859 if (unlikely(ret != 0))
862 if (mem->mm_node == NULL)
864 mem->mem_type = mem_type;
868 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
869 uint32_t cur_placement,
870 uint32_t proposed_placement)
872 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
873 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
876 * Keep current caching if possible.
879 if ((cur_placement & caching) != 0)
880 result |= (cur_placement & caching);
881 else if ((man->default_caching & caching) != 0)
882 result |= man->default_caching;
883 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
884 result |= TTM_PL_FLAG_CACHED;
885 else if ((TTM_PL_FLAG_WC & caching) != 0)
886 result |= TTM_PL_FLAG_WC;
887 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
888 result |= TTM_PL_FLAG_UNCACHED;
893 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
895 uint32_t proposed_placement,
896 uint32_t *masked_placement)
898 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
900 if ((cur_flags & proposed_placement & TTM_PL_MASK_MEM) == 0)
903 if ((proposed_placement & man->available_caching) == 0)
906 cur_flags |= (proposed_placement & man->available_caching);
908 *masked_placement = cur_flags;
913 * Creates space for memory region @mem according to its type.
915 * This function first searches for free space in compatible memory types in
916 * the priority order defined by the driver. If free space isn't found, then
917 * ttm_bo_mem_force_space is attempted in priority order to evict and find
920 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
921 struct ttm_placement *placement,
922 struct ttm_mem_reg *mem,
923 bool interruptible, bool no_wait_reserve,
926 struct ttm_bo_device *bdev = bo->bdev;
927 struct ttm_mem_type_manager *man;
928 uint32_t mem_type = TTM_PL_SYSTEM;
929 uint32_t cur_flags = 0;
930 bool type_found = false;
931 bool type_ok = false;
932 bool has_erestartsys = false;
936 for (i = 0; i < placement->num_placement; ++i) {
937 ret = ttm_mem_type_from_flags(placement->placement[i],
941 man = &bdev->man[mem_type];
943 type_ok = ttm_bo_mt_compatible(man,
945 placement->placement[i],
951 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
954 * Use the access and other non-mapping-related flag bits from
955 * the memory placement flags to the current flags
957 ttm_flag_masked(&cur_flags, placement->placement[i],
958 ~TTM_PL_MASK_MEMTYPE);
960 if (mem_type == TTM_PL_SYSTEM)
963 if (man->has_type && man->use_type) {
965 ret = (*man->func->get_node)(man, bo, placement, mem);
973 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
974 mem->mem_type = mem_type;
975 mem->placement = cur_flags;
982 for (i = 0; i < placement->num_busy_placement; ++i) {
983 ret = ttm_mem_type_from_flags(placement->busy_placement[i],
987 man = &bdev->man[mem_type];
990 if (!ttm_bo_mt_compatible(man,
992 placement->busy_placement[i],
996 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
999 * Use the access and other non-mapping-related flag bits from
1000 * the memory placement flags to the current flags
1002 ttm_flag_masked(&cur_flags, placement->busy_placement[i],
1003 ~TTM_PL_MASK_MEMTYPE);
1006 if (mem_type == TTM_PL_SYSTEM) {
1007 mem->mem_type = mem_type;
1008 mem->placement = cur_flags;
1009 mem->mm_node = NULL;
1013 ret = ttm_bo_mem_force_space(bo, mem_type, placement, mem,
1014 interruptible, no_wait_reserve, no_wait_gpu);
1015 if (ret == 0 && mem->mm_node) {
1016 mem->placement = cur_flags;
1019 if (ret == -ERESTARTSYS)
1020 has_erestartsys = true;
1022 ret = (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
1025 EXPORT_SYMBOL(ttm_bo_mem_space);
1027 int ttm_bo_wait_cpu(struct ttm_buffer_object *bo, bool no_wait)
1029 if ((atomic_read(&bo->cpu_writers) > 0) && no_wait)
1032 return wait_event_interruptible(bo->event_queue,
1033 atomic_read(&bo->cpu_writers) == 0);
1035 EXPORT_SYMBOL(ttm_bo_wait_cpu);
1037 int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1038 struct ttm_placement *placement,
1039 bool interruptible, bool no_wait_reserve,
1043 struct ttm_mem_reg mem;
1044 struct ttm_bo_device *bdev = bo->bdev;
1046 BUG_ON(!atomic_read(&bo->reserved));
1049 * FIXME: It's possible to pipeline buffer moves.
1050 * Have the driver move function wait for idle when necessary,
1051 * instead of doing it here.
1053 spin_lock(&bdev->fence_lock);
1054 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
1055 spin_unlock(&bdev->fence_lock);
1058 mem.num_pages = bo->num_pages;
1059 mem.size = mem.num_pages << PAGE_SHIFT;
1060 mem.page_alignment = bo->mem.page_alignment;
1061 mem.bus.io_reserved_vm = false;
1062 mem.bus.io_reserved_count = 0;
1064 * Determine where to move the buffer.
1066 ret = ttm_bo_mem_space(bo, placement, &mem, interruptible, no_wait_reserve, no_wait_gpu);
1069 ret = ttm_bo_handle_move_mem(bo, &mem, false, interruptible, no_wait_reserve, no_wait_gpu);
1071 if (ret && mem.mm_node)
1072 ttm_bo_mem_put(bo, &mem);
1076 static int ttm_bo_mem_compat(struct ttm_placement *placement,
1077 struct ttm_mem_reg *mem)
1081 if (mem->mm_node && placement->lpfn != 0 &&
1082 (mem->start < placement->fpfn ||
1083 mem->start + mem->num_pages > placement->lpfn))
1086 for (i = 0; i < placement->num_placement; i++) {
1087 if ((placement->placement[i] & mem->placement &
1088 TTM_PL_MASK_CACHING) &&
1089 (placement->placement[i] & mem->placement &
1096 int ttm_bo_validate(struct ttm_buffer_object *bo,
1097 struct ttm_placement *placement,
1098 bool interruptible, bool no_wait_reserve,
1103 BUG_ON(!atomic_read(&bo->reserved));
1104 /* Check that range is valid */
1105 if (placement->lpfn || placement->fpfn)
1106 if (placement->fpfn > placement->lpfn ||
1107 (placement->lpfn - placement->fpfn) < bo->num_pages)
1110 * Check whether we need to move buffer.
1112 ret = ttm_bo_mem_compat(placement, &bo->mem);
1114 ret = ttm_bo_move_buffer(bo, placement, interruptible, no_wait_reserve, no_wait_gpu);
1119 * Use the access and other non-mapping-related flag bits from
1120 * the compatible memory placement flags to the active flags
1122 ttm_flag_masked(&bo->mem.placement, placement->placement[ret],
1123 ~TTM_PL_MASK_MEMTYPE);
1126 * We might need to add a TTM.
1128 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1129 ret = ttm_bo_add_ttm(bo, true);
1135 EXPORT_SYMBOL(ttm_bo_validate);
1137 int ttm_bo_check_placement(struct ttm_buffer_object *bo,
1138 struct ttm_placement *placement)
1140 BUG_ON((placement->fpfn || placement->lpfn) &&
1141 (bo->mem.num_pages > (placement->lpfn - placement->fpfn)));
1146 int ttm_bo_init(struct ttm_bo_device *bdev,
1147 struct ttm_buffer_object *bo,
1149 enum ttm_bo_type type,
1150 struct ttm_placement *placement,
1151 uint32_t page_alignment,
1152 unsigned long buffer_start,
1154 struct file *persistent_swap_storage,
1156 void (*destroy) (struct ttm_buffer_object *))
1159 unsigned long num_pages;
1161 size += buffer_start & ~PAGE_MASK;
1162 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1163 if (num_pages == 0) {
1164 printk(KERN_ERR TTM_PFX "Illegal buffer object size.\n");
1171 bo->destroy = destroy;
1173 kref_init(&bo->kref);
1174 kref_init(&bo->list_kref);
1175 atomic_set(&bo->cpu_writers, 0);
1176 atomic_set(&bo->reserved, 1);
1177 init_waitqueue_head(&bo->event_queue);
1178 INIT_LIST_HEAD(&bo->lru);
1179 INIT_LIST_HEAD(&bo->ddestroy);
1180 INIT_LIST_HEAD(&bo->swap);
1181 INIT_LIST_HEAD(&bo->io_reserve_lru);
1183 bo->glob = bdev->glob;
1185 bo->num_pages = num_pages;
1186 bo->mem.size = num_pages << PAGE_SHIFT;
1187 bo->mem.mem_type = TTM_PL_SYSTEM;
1188 bo->mem.num_pages = bo->num_pages;
1189 bo->mem.mm_node = NULL;
1190 bo->mem.page_alignment = page_alignment;
1191 bo->mem.bus.io_reserved_vm = false;
1192 bo->mem.bus.io_reserved_count = 0;
1193 bo->buffer_start = buffer_start & PAGE_MASK;
1195 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1196 bo->seq_valid = false;
1197 bo->persistent_swap_storage = persistent_swap_storage;
1198 bo->acc_size = acc_size;
1199 atomic_inc(&bo->glob->bo_count);
1201 ret = ttm_bo_check_placement(bo, placement);
1202 if (unlikely(ret != 0))
1206 * For ttm_bo_type_device buffers, allocate
1207 * address space from the device.
1209 if (bo->type == ttm_bo_type_device) {
1210 ret = ttm_bo_setup_vm(bo);
1215 ret = ttm_bo_validate(bo, placement, interruptible, false, false);
1219 ttm_bo_unreserve(bo);
1223 ttm_bo_unreserve(bo);
1228 EXPORT_SYMBOL(ttm_bo_init);
1230 static inline size_t ttm_bo_size(struct ttm_bo_global *glob,
1231 unsigned long num_pages)
1233 size_t page_array_size = (num_pages * sizeof(void *) + PAGE_SIZE - 1) &
1236 return glob->ttm_bo_size + 2 * page_array_size;
1239 int ttm_bo_create(struct ttm_bo_device *bdev,
1241 enum ttm_bo_type type,
1242 struct ttm_placement *placement,
1243 uint32_t page_alignment,
1244 unsigned long buffer_start,
1246 struct file *persistent_swap_storage,
1247 struct ttm_buffer_object **p_bo)
1249 struct ttm_buffer_object *bo;
1250 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1254 ttm_bo_size(bdev->glob, (size + PAGE_SIZE - 1) >> PAGE_SHIFT);
1255 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1256 if (unlikely(ret != 0))
1259 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1261 if (unlikely(bo == NULL)) {
1262 ttm_mem_global_free(mem_glob, acc_size);
1266 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1267 buffer_start, interruptible,
1268 persistent_swap_storage, acc_size, NULL);
1269 if (likely(ret == 0))
1274 EXPORT_SYMBOL(ttm_bo_create);
1276 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1277 unsigned mem_type, bool allow_errors)
1279 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1280 struct ttm_bo_global *glob = bdev->glob;
1284 * Can't use standard list traversal since we're unlocking.
1287 spin_lock(&glob->lru_lock);
1288 while (!list_empty(&man->lru)) {
1289 spin_unlock(&glob->lru_lock);
1290 ret = ttm_mem_evict_first(bdev, mem_type, false, false, false);
1295 printk(KERN_ERR TTM_PFX
1296 "Cleanup eviction failed\n");
1299 spin_lock(&glob->lru_lock);
1301 spin_unlock(&glob->lru_lock);
1305 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1307 struct ttm_mem_type_manager *man;
1310 if (mem_type >= TTM_NUM_MEM_TYPES) {
1311 printk(KERN_ERR TTM_PFX "Illegal memory type %d\n", mem_type);
1314 man = &bdev->man[mem_type];
1316 if (!man->has_type) {
1317 printk(KERN_ERR TTM_PFX "Trying to take down uninitialized "
1318 "memory manager type %u\n", mem_type);
1322 man->use_type = false;
1323 man->has_type = false;
1327 ttm_bo_force_list_clean(bdev, mem_type, false);
1329 ret = (*man->func->takedown)(man);
1334 EXPORT_SYMBOL(ttm_bo_clean_mm);
1336 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1338 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1340 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1341 printk(KERN_ERR TTM_PFX
1342 "Illegal memory manager memory type %u.\n",
1347 if (!man->has_type) {
1348 printk(KERN_ERR TTM_PFX
1349 "Memory type %u has not been initialized.\n",
1354 return ttm_bo_force_list_clean(bdev, mem_type, true);
1356 EXPORT_SYMBOL(ttm_bo_evict_mm);
1358 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1359 unsigned long p_size)
1362 struct ttm_mem_type_manager *man;
1364 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1365 man = &bdev->man[type];
1366 BUG_ON(man->has_type);
1367 man->io_reserve_fastpath = true;
1368 man->use_io_reserve_lru = false;
1369 mutex_init(&man->io_reserve_mutex);
1370 INIT_LIST_HEAD(&man->io_reserve_lru);
1372 ret = bdev->driver->init_mem_type(bdev, type, man);
1378 if (type != TTM_PL_SYSTEM) {
1379 ret = (*man->func->init)(man, p_size);
1383 man->has_type = true;
1384 man->use_type = true;
1387 INIT_LIST_HEAD(&man->lru);
1391 EXPORT_SYMBOL(ttm_bo_init_mm);
1393 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1395 struct ttm_bo_global *glob =
1396 container_of(kobj, struct ttm_bo_global, kobj);
1398 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
1399 __free_page(glob->dummy_read_page);
1403 void ttm_bo_global_release(struct drm_global_reference *ref)
1405 struct ttm_bo_global *glob = ref->object;
1407 kobject_del(&glob->kobj);
1408 kobject_put(&glob->kobj);
1410 EXPORT_SYMBOL(ttm_bo_global_release);
1412 int ttm_bo_global_init(struct drm_global_reference *ref)
1414 struct ttm_bo_global_ref *bo_ref =
1415 container_of(ref, struct ttm_bo_global_ref, ref);
1416 struct ttm_bo_global *glob = ref->object;
1419 mutex_init(&glob->device_list_mutex);
1420 spin_lock_init(&glob->lru_lock);
1421 glob->mem_glob = bo_ref->mem_glob;
1422 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1424 if (unlikely(glob->dummy_read_page == NULL)) {
1429 INIT_LIST_HEAD(&glob->swap_lru);
1430 INIT_LIST_HEAD(&glob->device_list);
1432 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
1433 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
1434 if (unlikely(ret != 0)) {
1435 printk(KERN_ERR TTM_PFX
1436 "Could not register buffer object swapout.\n");
1440 glob->ttm_bo_extra_size =
1441 ttm_round_pot(sizeof(struct ttm_tt)) +
1442 ttm_round_pot(sizeof(struct ttm_backend));
1444 glob->ttm_bo_size = glob->ttm_bo_extra_size +
1445 ttm_round_pot(sizeof(struct ttm_buffer_object));
1447 atomic_set(&glob->bo_count, 0);
1449 ret = kobject_init_and_add(
1450 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1451 if (unlikely(ret != 0))
1452 kobject_put(&glob->kobj);
1455 __free_page(glob->dummy_read_page);
1460 EXPORT_SYMBOL(ttm_bo_global_init);
1463 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1466 unsigned i = TTM_NUM_MEM_TYPES;
1467 struct ttm_mem_type_manager *man;
1468 struct ttm_bo_global *glob = bdev->glob;
1471 man = &bdev->man[i];
1472 if (man->has_type) {
1473 man->use_type = false;
1474 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1476 printk(KERN_ERR TTM_PFX
1477 "DRM memory manager type %d "
1478 "is not clean.\n", i);
1480 man->has_type = false;
1484 mutex_lock(&glob->device_list_mutex);
1485 list_del(&bdev->device_list);
1486 mutex_unlock(&glob->device_list_mutex);
1488 cancel_delayed_work_sync(&bdev->wq);
1490 while (ttm_bo_delayed_delete(bdev, true))
1493 spin_lock(&glob->lru_lock);
1494 if (list_empty(&bdev->ddestroy))
1495 TTM_DEBUG("Delayed destroy list was clean\n");
1497 if (list_empty(&bdev->man[0].lru))
1498 TTM_DEBUG("Swap list was clean\n");
1499 spin_unlock(&glob->lru_lock);
1501 BUG_ON(!drm_mm_clean(&bdev->addr_space_mm));
1502 write_lock(&bdev->vm_lock);
1503 drm_mm_takedown(&bdev->addr_space_mm);
1504 write_unlock(&bdev->vm_lock);
1508 EXPORT_SYMBOL(ttm_bo_device_release);
1510 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1511 struct ttm_bo_global *glob,
1512 struct ttm_bo_driver *driver,
1513 uint64_t file_page_offset,
1518 rwlock_init(&bdev->vm_lock);
1519 bdev->driver = driver;
1521 memset(bdev->man, 0, sizeof(bdev->man));
1524 * Initialize the system memory buffer type.
1525 * Other types need to be driver / IOCTL initialized.
1527 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1528 if (unlikely(ret != 0))
1531 bdev->addr_space_rb = RB_ROOT;
1532 ret = drm_mm_init(&bdev->addr_space_mm, file_page_offset, 0x10000000);
1533 if (unlikely(ret != 0))
1534 goto out_no_addr_mm;
1536 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1537 bdev->nice_mode = true;
1538 INIT_LIST_HEAD(&bdev->ddestroy);
1539 bdev->dev_mapping = NULL;
1541 bdev->need_dma32 = need_dma32;
1543 spin_lock_init(&bdev->fence_lock);
1544 mutex_lock(&glob->device_list_mutex);
1545 list_add_tail(&bdev->device_list, &glob->device_list);
1546 mutex_unlock(&glob->device_list_mutex);
1550 ttm_bo_clean_mm(bdev, 0);
1554 EXPORT_SYMBOL(ttm_bo_device_init);
1557 * buffer object vm functions.
1560 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1562 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1564 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1565 if (mem->mem_type == TTM_PL_SYSTEM)
1568 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1571 if (mem->placement & TTM_PL_FLAG_CACHED)
1577 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1579 struct ttm_bo_device *bdev = bo->bdev;
1580 loff_t offset = (loff_t) bo->addr_space_offset;
1581 loff_t holelen = ((loff_t) bo->mem.num_pages) << PAGE_SHIFT;
1583 if (!bdev->dev_mapping)
1585 unmap_mapping_range(bdev->dev_mapping, offset, holelen, 1);
1586 ttm_mem_io_free_vm(bo);
1589 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1591 struct ttm_bo_device *bdev = bo->bdev;
1592 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1594 ttm_mem_io_lock(man, false);
1595 ttm_bo_unmap_virtual_locked(bo);
1596 ttm_mem_io_unlock(man);
1600 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1602 static void ttm_bo_vm_insert_rb(struct ttm_buffer_object *bo)
1604 struct ttm_bo_device *bdev = bo->bdev;
1605 struct rb_node **cur = &bdev->addr_space_rb.rb_node;
1606 struct rb_node *parent = NULL;
1607 struct ttm_buffer_object *cur_bo;
1608 unsigned long offset = bo->vm_node->start;
1609 unsigned long cur_offset;
1613 cur_bo = rb_entry(parent, struct ttm_buffer_object, vm_rb);
1614 cur_offset = cur_bo->vm_node->start;
1615 if (offset < cur_offset)
1616 cur = &parent->rb_left;
1617 else if (offset > cur_offset)
1618 cur = &parent->rb_right;
1623 rb_link_node(&bo->vm_rb, parent, cur);
1624 rb_insert_color(&bo->vm_rb, &bdev->addr_space_rb);
1630 * @bo: the buffer to allocate address space for
1632 * Allocate address space in the drm device so that applications
1633 * can mmap the buffer and access the contents. This only
1634 * applies to ttm_bo_type_device objects as others are not
1635 * placed in the drm device address space.
1638 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo)
1640 struct ttm_bo_device *bdev = bo->bdev;
1644 ret = drm_mm_pre_get(&bdev->addr_space_mm);
1645 if (unlikely(ret != 0))
1648 write_lock(&bdev->vm_lock);
1649 bo->vm_node = drm_mm_search_free(&bdev->addr_space_mm,
1650 bo->mem.num_pages, 0, 0);
1652 if (unlikely(bo->vm_node == NULL)) {
1657 bo->vm_node = drm_mm_get_block_atomic(bo->vm_node,
1658 bo->mem.num_pages, 0);
1660 if (unlikely(bo->vm_node == NULL)) {
1661 write_unlock(&bdev->vm_lock);
1665 ttm_bo_vm_insert_rb(bo);
1666 write_unlock(&bdev->vm_lock);
1667 bo->addr_space_offset = ((uint64_t) bo->vm_node->start) << PAGE_SHIFT;
1671 write_unlock(&bdev->vm_lock);
1675 int ttm_bo_wait(struct ttm_buffer_object *bo,
1676 bool lazy, bool interruptible, bool no_wait)
1678 struct ttm_bo_driver *driver = bo->bdev->driver;
1679 struct ttm_bo_device *bdev = bo->bdev;
1684 if (likely(bo->sync_obj == NULL))
1687 while (bo->sync_obj) {
1689 if (driver->sync_obj_signaled(bo->sync_obj, bo->sync_obj_arg)) {
1690 void *tmp_obj = bo->sync_obj;
1691 bo->sync_obj = NULL;
1692 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
1693 spin_unlock(&bdev->fence_lock);
1694 driver->sync_obj_unref(&tmp_obj);
1695 spin_lock(&bdev->fence_lock);
1702 sync_obj = driver->sync_obj_ref(bo->sync_obj);
1703 sync_obj_arg = bo->sync_obj_arg;
1704 spin_unlock(&bdev->fence_lock);
1705 ret = driver->sync_obj_wait(sync_obj, sync_obj_arg,
1706 lazy, interruptible);
1707 if (unlikely(ret != 0)) {
1708 driver->sync_obj_unref(&sync_obj);
1709 spin_lock(&bdev->fence_lock);
1712 spin_lock(&bdev->fence_lock);
1713 if (likely(bo->sync_obj == sync_obj &&
1714 bo->sync_obj_arg == sync_obj_arg)) {
1715 void *tmp_obj = bo->sync_obj;
1716 bo->sync_obj = NULL;
1717 clear_bit(TTM_BO_PRIV_FLAG_MOVING,
1719 spin_unlock(&bdev->fence_lock);
1720 driver->sync_obj_unref(&sync_obj);
1721 driver->sync_obj_unref(&tmp_obj);
1722 spin_lock(&bdev->fence_lock);
1724 spin_unlock(&bdev->fence_lock);
1725 driver->sync_obj_unref(&sync_obj);
1726 spin_lock(&bdev->fence_lock);
1731 EXPORT_SYMBOL(ttm_bo_wait);
1733 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1735 struct ttm_bo_device *bdev = bo->bdev;
1739 * Using ttm_bo_reserve makes sure the lru lists are updated.
1742 ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
1743 if (unlikely(ret != 0))
1745 spin_lock(&bdev->fence_lock);
1746 ret = ttm_bo_wait(bo, false, true, no_wait);
1747 spin_unlock(&bdev->fence_lock);
1748 if (likely(ret == 0))
1749 atomic_inc(&bo->cpu_writers);
1750 ttm_bo_unreserve(bo);
1753 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1755 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1757 if (atomic_dec_and_test(&bo->cpu_writers))
1758 wake_up_all(&bo->event_queue);
1760 EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1763 * A buffer object shrink method that tries to swap out the first
1764 * buffer object on the bo_global::swap_lru list.
1767 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1769 struct ttm_bo_global *glob =
1770 container_of(shrink, struct ttm_bo_global, shrink);
1771 struct ttm_buffer_object *bo;
1774 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM);
1776 spin_lock(&glob->lru_lock);
1777 while (ret == -EBUSY) {
1778 if (unlikely(list_empty(&glob->swap_lru))) {
1779 spin_unlock(&glob->lru_lock);
1783 bo = list_first_entry(&glob->swap_lru,
1784 struct ttm_buffer_object, swap);
1785 kref_get(&bo->list_kref);
1787 if (!list_empty(&bo->ddestroy)) {
1788 spin_unlock(&glob->lru_lock);
1789 (void) ttm_bo_cleanup_refs(bo, false, false, false);
1790 kref_put(&bo->list_kref, ttm_bo_release_list);
1795 * Reserve buffer. Since we unlock while sleeping, we need
1796 * to re-check that nobody removed us from the swap-list while
1800 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
1801 if (unlikely(ret == -EBUSY)) {
1802 spin_unlock(&glob->lru_lock);
1803 ttm_bo_wait_unreserved(bo, false);
1804 kref_put(&bo->list_kref, ttm_bo_release_list);
1805 spin_lock(&glob->lru_lock);
1810 put_count = ttm_bo_del_from_lru(bo);
1811 spin_unlock(&glob->lru_lock);
1813 ttm_bo_list_ref_sub(bo, put_count, true);
1816 * Wait for GPU, then move to system cached.
1819 spin_lock(&bo->bdev->fence_lock);
1820 ret = ttm_bo_wait(bo, false, false, false);
1821 spin_unlock(&bo->bdev->fence_lock);
1823 if (unlikely(ret != 0))
1826 if ((bo->mem.placement & swap_placement) != swap_placement) {
1827 struct ttm_mem_reg evict_mem;
1829 evict_mem = bo->mem;
1830 evict_mem.mm_node = NULL;
1831 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1832 evict_mem.mem_type = TTM_PL_SYSTEM;
1834 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1835 false, false, false);
1836 if (unlikely(ret != 0))
1840 ttm_bo_unmap_virtual(bo);
1843 * Swap out. Buffer will be swapped in again as soon as
1844 * anyone tries to access a ttm page.
1847 if (bo->bdev->driver->swap_notify)
1848 bo->bdev->driver->swap_notify(bo);
1850 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1855 * Unreserve without putting on LRU to avoid swapping out an
1856 * already swapped buffer.
1859 atomic_set(&bo->reserved, 0);
1860 wake_up_all(&bo->event_queue);
1861 kref_put(&bo->list_kref, ttm_bo_release_list);
1865 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1867 while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1870 EXPORT_SYMBOL(ttm_bo_swapout_all);