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))
345 case ttm_bo_type_user:
346 bo->ttm = ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
347 page_flags | TTM_PAGE_FLAG_USER,
348 glob->dummy_read_page);
349 if (unlikely(bo->ttm == NULL)) {
354 ret = ttm_tt_set_user(bo->ttm, current,
355 bo->buffer_start, bo->num_pages);
356 if (unlikely(ret != 0)) {
357 ttm_tt_destroy(bo->ttm);
362 printk(KERN_ERR TTM_PFX "Illegal buffer object type\n");
370 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
371 struct ttm_mem_reg *mem,
372 bool evict, bool interruptible,
373 bool no_wait_reserve, bool no_wait_gpu)
375 struct ttm_bo_device *bdev = bo->bdev;
376 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
377 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
378 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
379 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
382 if (old_is_pci || new_is_pci ||
383 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
384 ret = ttm_mem_io_lock(old_man, true);
385 if (unlikely(ret != 0))
387 ttm_bo_unmap_virtual_locked(bo);
388 ttm_mem_io_unlock(old_man);
392 * Create and bind a ttm if required.
395 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
396 if (bo->ttm == NULL) {
397 ret = ttm_bo_add_ttm(bo, false);
402 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
406 if (mem->mem_type != TTM_PL_SYSTEM) {
407 ret = ttm_tt_bind(bo->ttm, mem);
412 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
413 if (bdev->driver->move_notify)
414 bdev->driver->move_notify(bo, mem);
421 if (bdev->driver->move_notify)
422 bdev->driver->move_notify(bo, mem);
424 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
425 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
426 ret = ttm_bo_move_ttm(bo, evict, no_wait_reserve, no_wait_gpu, mem);
427 else if (bdev->driver->move)
428 ret = bdev->driver->move(bo, evict, interruptible,
429 no_wait_reserve, no_wait_gpu, mem);
431 ret = ttm_bo_move_memcpy(bo, evict, no_wait_reserve, no_wait_gpu, mem);
438 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
440 printk(KERN_ERR TTM_PFX "Can not flush read caches\n");
444 if (bo->mem.mm_node) {
445 bo->offset = (bo->mem.start << PAGE_SHIFT) +
446 bdev->man[bo->mem.mem_type].gpu_offset;
447 bo->cur_placement = bo->mem.placement;
454 new_man = &bdev->man[bo->mem.mem_type];
455 if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) {
456 ttm_tt_unbind(bo->ttm);
457 ttm_tt_destroy(bo->ttm);
466 * Will release GPU memory type usage on destruction.
467 * This is the place to put in driver specific hooks to release
468 * driver private resources.
469 * Will release the bo::reserved lock.
472 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
475 ttm_tt_unbind(bo->ttm);
476 ttm_tt_destroy(bo->ttm);
479 ttm_bo_mem_put(bo, &bo->mem);
481 atomic_set(&bo->reserved, 0);
484 * Make processes trying to reserve really pick it up.
486 smp_mb__after_atomic_dec();
487 wake_up_all(&bo->event_queue);
490 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
492 struct ttm_bo_device *bdev = bo->bdev;
493 struct ttm_bo_global *glob = bo->glob;
494 struct ttm_bo_driver *driver;
495 void *sync_obj = NULL;
500 spin_lock(&bdev->fence_lock);
501 (void) ttm_bo_wait(bo, false, false, true, TTM_USAGE_READWRITE);
504 spin_lock(&glob->lru_lock);
507 * Lock inversion between bo:reserve and bdev::fence_lock here,
508 * but that's OK, since we're only trylocking.
511 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
513 if (unlikely(ret == -EBUSY))
516 spin_unlock(&bdev->fence_lock);
517 put_count = ttm_bo_del_from_lru(bo);
519 spin_unlock(&glob->lru_lock);
520 ttm_bo_cleanup_memtype_use(bo);
522 ttm_bo_list_ref_sub(bo, put_count, true);
526 spin_lock(&glob->lru_lock);
529 driver = bdev->driver;
531 sync_obj = driver->sync_obj_ref(bo->sync_obj);
532 sync_obj_arg = bo->sync_obj_arg;
534 kref_get(&bo->list_kref);
535 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
536 spin_unlock(&glob->lru_lock);
537 spin_unlock(&bdev->fence_lock);
540 driver->sync_obj_flush(sync_obj, sync_obj_arg);
541 driver->sync_obj_unref(&sync_obj);
543 schedule_delayed_work(&bdev->wq,
544 ((HZ / 100) < 1) ? 1 : HZ / 100);
548 * function ttm_bo_cleanup_refs
549 * If bo idle, remove from delayed- and lru lists, and unref.
550 * If not idle, do nothing.
552 * @interruptible Any sleeps should occur interruptibly.
553 * @no_wait_reserve Never wait for reserve. Return -EBUSY instead.
554 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
557 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
559 bool no_wait_reserve,
562 struct ttm_bo_device *bdev = bo->bdev;
563 struct ttm_bo_global *glob = bo->glob;
568 spin_lock(&bdev->fence_lock);
569 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu,
570 TTM_USAGE_READWRITE);
571 spin_unlock(&bdev->fence_lock);
573 if (unlikely(ret != 0))
576 spin_lock(&glob->lru_lock);
577 ret = ttm_bo_reserve_locked(bo, interruptible,
578 no_wait_reserve, false, 0);
580 if (unlikely(ret != 0) || list_empty(&bo->ddestroy)) {
581 spin_unlock(&glob->lru_lock);
586 * We can re-check for sync object without taking
587 * the bo::lock since setting the sync object requires
588 * also bo::reserved. A busy object at this point may
589 * be caused by another thread recently starting an accelerated
593 if (unlikely(bo->sync_obj)) {
594 atomic_set(&bo->reserved, 0);
595 wake_up_all(&bo->event_queue);
596 spin_unlock(&glob->lru_lock);
600 put_count = ttm_bo_del_from_lru(bo);
601 list_del_init(&bo->ddestroy);
604 spin_unlock(&glob->lru_lock);
605 ttm_bo_cleanup_memtype_use(bo);
607 ttm_bo_list_ref_sub(bo, put_count, true);
613 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
614 * encountered buffers.
617 static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
619 struct ttm_bo_global *glob = bdev->glob;
620 struct ttm_buffer_object *entry = NULL;
623 spin_lock(&glob->lru_lock);
624 if (list_empty(&bdev->ddestroy))
627 entry = list_first_entry(&bdev->ddestroy,
628 struct ttm_buffer_object, ddestroy);
629 kref_get(&entry->list_kref);
632 struct ttm_buffer_object *nentry = NULL;
634 if (entry->ddestroy.next != &bdev->ddestroy) {
635 nentry = list_first_entry(&entry->ddestroy,
636 struct ttm_buffer_object, ddestroy);
637 kref_get(&nentry->list_kref);
640 spin_unlock(&glob->lru_lock);
641 ret = ttm_bo_cleanup_refs(entry, false, !remove_all,
643 kref_put(&entry->list_kref, ttm_bo_release_list);
649 spin_lock(&glob->lru_lock);
650 if (list_empty(&entry->ddestroy))
655 spin_unlock(&glob->lru_lock);
658 kref_put(&entry->list_kref, ttm_bo_release_list);
662 static void ttm_bo_delayed_workqueue(struct work_struct *work)
664 struct ttm_bo_device *bdev =
665 container_of(work, struct ttm_bo_device, wq.work);
667 if (ttm_bo_delayed_delete(bdev, false)) {
668 schedule_delayed_work(&bdev->wq,
669 ((HZ / 100) < 1) ? 1 : HZ / 100);
673 static void ttm_bo_release(struct kref *kref)
675 struct ttm_buffer_object *bo =
676 container_of(kref, struct ttm_buffer_object, kref);
677 struct ttm_bo_device *bdev = bo->bdev;
678 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
680 if (likely(bo->vm_node != NULL)) {
681 rb_erase(&bo->vm_rb, &bdev->addr_space_rb);
682 drm_mm_put_block(bo->vm_node);
685 write_unlock(&bdev->vm_lock);
686 ttm_mem_io_lock(man, false);
687 ttm_mem_io_free_vm(bo);
688 ttm_mem_io_unlock(man);
689 ttm_bo_cleanup_refs_or_queue(bo);
690 kref_put(&bo->list_kref, ttm_bo_release_list);
691 write_lock(&bdev->vm_lock);
694 void ttm_bo_unref(struct ttm_buffer_object **p_bo)
696 struct ttm_buffer_object *bo = *p_bo;
697 struct ttm_bo_device *bdev = bo->bdev;
700 write_lock(&bdev->vm_lock);
701 kref_put(&bo->kref, ttm_bo_release);
702 write_unlock(&bdev->vm_lock);
704 EXPORT_SYMBOL(ttm_bo_unref);
706 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
708 return cancel_delayed_work_sync(&bdev->wq);
710 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
712 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
715 schedule_delayed_work(&bdev->wq,
716 ((HZ / 100) < 1) ? 1 : HZ / 100);
718 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
720 static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
721 bool no_wait_reserve, bool no_wait_gpu)
723 struct ttm_bo_device *bdev = bo->bdev;
724 struct ttm_mem_reg evict_mem;
725 struct ttm_placement placement;
728 spin_lock(&bdev->fence_lock);
729 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu,
730 TTM_USAGE_READWRITE);
731 spin_unlock(&bdev->fence_lock);
733 if (unlikely(ret != 0)) {
734 if (ret != -ERESTARTSYS) {
735 printk(KERN_ERR TTM_PFX
736 "Failed to expire sync object before "
737 "buffer eviction.\n");
742 BUG_ON(!atomic_read(&bo->reserved));
745 evict_mem.mm_node = NULL;
746 evict_mem.bus.io_reserved_vm = false;
747 evict_mem.bus.io_reserved_count = 0;
751 placement.num_placement = 0;
752 placement.num_busy_placement = 0;
753 bdev->driver->evict_flags(bo, &placement);
754 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
755 no_wait_reserve, no_wait_gpu);
757 if (ret != -ERESTARTSYS) {
758 printk(KERN_ERR TTM_PFX
759 "Failed to find memory space for "
760 "buffer 0x%p eviction.\n", bo);
761 ttm_bo_mem_space_debug(bo, &placement);
766 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
767 no_wait_reserve, no_wait_gpu);
769 if (ret != -ERESTARTSYS)
770 printk(KERN_ERR TTM_PFX "Buffer eviction failed\n");
771 ttm_bo_mem_put(bo, &evict_mem);
779 static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
781 bool interruptible, bool no_wait_reserve,
784 struct ttm_bo_global *glob = bdev->glob;
785 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
786 struct ttm_buffer_object *bo;
787 int ret, put_count = 0;
790 spin_lock(&glob->lru_lock);
791 if (list_empty(&man->lru)) {
792 spin_unlock(&glob->lru_lock);
796 bo = list_first_entry(&man->lru, struct ttm_buffer_object, lru);
797 kref_get(&bo->list_kref);
799 if (!list_empty(&bo->ddestroy)) {
800 spin_unlock(&glob->lru_lock);
801 ret = ttm_bo_cleanup_refs(bo, interruptible,
802 no_wait_reserve, no_wait_gpu);
803 kref_put(&bo->list_kref, ttm_bo_release_list);
805 if (likely(ret == 0 || ret == -ERESTARTSYS))
811 ret = ttm_bo_reserve_locked(bo, false, no_wait_reserve, false, 0);
813 if (unlikely(ret == -EBUSY)) {
814 spin_unlock(&glob->lru_lock);
815 if (likely(!no_wait_gpu))
816 ret = ttm_bo_wait_unreserved(bo, interruptible);
818 kref_put(&bo->list_kref, ttm_bo_release_list);
821 * We *need* to retry after releasing the lru lock.
824 if (unlikely(ret != 0))
829 put_count = ttm_bo_del_from_lru(bo);
830 spin_unlock(&glob->lru_lock);
834 ttm_bo_list_ref_sub(bo, put_count, true);
836 ret = ttm_bo_evict(bo, interruptible, no_wait_reserve, no_wait_gpu);
837 ttm_bo_unreserve(bo);
839 kref_put(&bo->list_kref, ttm_bo_release_list);
843 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
845 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
848 (*man->func->put_node)(man, mem);
850 EXPORT_SYMBOL(ttm_bo_mem_put);
853 * Repeatedly evict memory from the LRU for @mem_type until we create enough
854 * space, or we've evicted everything and there isn't enough space.
856 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
858 struct ttm_placement *placement,
859 struct ttm_mem_reg *mem,
861 bool no_wait_reserve,
864 struct ttm_bo_device *bdev = bo->bdev;
865 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
869 ret = (*man->func->get_node)(man, bo, placement, mem);
870 if (unlikely(ret != 0))
874 ret = ttm_mem_evict_first(bdev, mem_type, interruptible,
875 no_wait_reserve, no_wait_gpu);
876 if (unlikely(ret != 0))
879 if (mem->mm_node == NULL)
881 mem->mem_type = mem_type;
885 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
886 uint32_t cur_placement,
887 uint32_t proposed_placement)
889 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
890 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
893 * Keep current caching if possible.
896 if ((cur_placement & caching) != 0)
897 result |= (cur_placement & caching);
898 else if ((man->default_caching & caching) != 0)
899 result |= man->default_caching;
900 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
901 result |= TTM_PL_FLAG_CACHED;
902 else if ((TTM_PL_FLAG_WC & caching) != 0)
903 result |= TTM_PL_FLAG_WC;
904 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
905 result |= TTM_PL_FLAG_UNCACHED;
910 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
913 uint32_t proposed_placement,
914 uint32_t *masked_placement)
916 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
918 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) && disallow_fixed)
921 if ((cur_flags & proposed_placement & TTM_PL_MASK_MEM) == 0)
924 if ((proposed_placement & man->available_caching) == 0)
927 cur_flags |= (proposed_placement & man->available_caching);
929 *masked_placement = cur_flags;
934 * Creates space for memory region @mem according to its type.
936 * This function first searches for free space in compatible memory types in
937 * the priority order defined by the driver. If free space isn't found, then
938 * ttm_bo_mem_force_space is attempted in priority order to evict and find
941 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
942 struct ttm_placement *placement,
943 struct ttm_mem_reg *mem,
944 bool interruptible, bool no_wait_reserve,
947 struct ttm_bo_device *bdev = bo->bdev;
948 struct ttm_mem_type_manager *man;
949 uint32_t mem_type = TTM_PL_SYSTEM;
950 uint32_t cur_flags = 0;
951 bool type_found = false;
952 bool type_ok = false;
953 bool has_erestartsys = false;
957 for (i = 0; i < placement->num_placement; ++i) {
958 ret = ttm_mem_type_from_flags(placement->placement[i],
962 man = &bdev->man[mem_type];
964 type_ok = ttm_bo_mt_compatible(man,
965 bo->type == ttm_bo_type_user,
967 placement->placement[i],
973 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
976 * Use the access and other non-mapping-related flag bits from
977 * the memory placement flags to the current flags
979 ttm_flag_masked(&cur_flags, placement->placement[i],
980 ~TTM_PL_MASK_MEMTYPE);
982 if (mem_type == TTM_PL_SYSTEM)
985 if (man->has_type && man->use_type) {
987 ret = (*man->func->get_node)(man, bo, placement, mem);
995 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
996 mem->mem_type = mem_type;
997 mem->placement = cur_flags;
1004 for (i = 0; i < placement->num_busy_placement; ++i) {
1005 ret = ttm_mem_type_from_flags(placement->busy_placement[i],
1009 man = &bdev->man[mem_type];
1012 if (!ttm_bo_mt_compatible(man,
1013 bo->type == ttm_bo_type_user,
1015 placement->busy_placement[i],
1019 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
1022 * Use the access and other non-mapping-related flag bits from
1023 * the memory placement flags to the current flags
1025 ttm_flag_masked(&cur_flags, placement->busy_placement[i],
1026 ~TTM_PL_MASK_MEMTYPE);
1029 if (mem_type == TTM_PL_SYSTEM) {
1030 mem->mem_type = mem_type;
1031 mem->placement = cur_flags;
1032 mem->mm_node = NULL;
1036 ret = ttm_bo_mem_force_space(bo, mem_type, placement, mem,
1037 interruptible, no_wait_reserve, no_wait_gpu);
1038 if (ret == 0 && mem->mm_node) {
1039 mem->placement = cur_flags;
1042 if (ret == -ERESTARTSYS)
1043 has_erestartsys = true;
1045 ret = (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
1048 EXPORT_SYMBOL(ttm_bo_mem_space);
1050 int ttm_bo_wait_cpu(struct ttm_buffer_object *bo, bool no_wait)
1052 if ((atomic_read(&bo->cpu_writers) > 0) && no_wait)
1055 return wait_event_interruptible(bo->event_queue,
1056 atomic_read(&bo->cpu_writers) == 0);
1058 EXPORT_SYMBOL(ttm_bo_wait_cpu);
1060 int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1061 struct ttm_placement *placement,
1062 bool interruptible, bool no_wait_reserve,
1066 struct ttm_mem_reg mem;
1067 struct ttm_bo_device *bdev = bo->bdev;
1069 BUG_ON(!atomic_read(&bo->reserved));
1072 * FIXME: It's possible to pipeline buffer moves.
1073 * Have the driver move function wait for idle when necessary,
1074 * instead of doing it here.
1076 spin_lock(&bdev->fence_lock);
1077 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu,
1078 TTM_USAGE_READWRITE);
1079 spin_unlock(&bdev->fence_lock);
1082 mem.num_pages = bo->num_pages;
1083 mem.size = mem.num_pages << PAGE_SHIFT;
1084 mem.page_alignment = bo->mem.page_alignment;
1085 mem.bus.io_reserved_vm = false;
1086 mem.bus.io_reserved_count = 0;
1088 * Determine where to move the buffer.
1090 ret = ttm_bo_mem_space(bo, placement, &mem, interruptible, no_wait_reserve, no_wait_gpu);
1093 ret = ttm_bo_handle_move_mem(bo, &mem, false, interruptible, no_wait_reserve, no_wait_gpu);
1095 if (ret && mem.mm_node)
1096 ttm_bo_mem_put(bo, &mem);
1100 static int ttm_bo_mem_compat(struct ttm_placement *placement,
1101 struct ttm_mem_reg *mem)
1105 if (mem->mm_node && placement->lpfn != 0 &&
1106 (mem->start < placement->fpfn ||
1107 mem->start + mem->num_pages > placement->lpfn))
1110 for (i = 0; i < placement->num_placement; i++) {
1111 if ((placement->placement[i] & mem->placement &
1112 TTM_PL_MASK_CACHING) &&
1113 (placement->placement[i] & mem->placement &
1120 int ttm_bo_validate(struct ttm_buffer_object *bo,
1121 struct ttm_placement *placement,
1122 bool interruptible, bool no_wait_reserve,
1127 BUG_ON(!atomic_read(&bo->reserved));
1128 /* Check that range is valid */
1129 if (placement->lpfn || placement->fpfn)
1130 if (placement->fpfn > placement->lpfn ||
1131 (placement->lpfn - placement->fpfn) < bo->num_pages)
1134 * Check whether we need to move buffer.
1136 ret = ttm_bo_mem_compat(placement, &bo->mem);
1138 ret = ttm_bo_move_buffer(bo, placement, interruptible, no_wait_reserve, no_wait_gpu);
1143 * Use the access and other non-mapping-related flag bits from
1144 * the compatible memory placement flags to the active flags
1146 ttm_flag_masked(&bo->mem.placement, placement->placement[ret],
1147 ~TTM_PL_MASK_MEMTYPE);
1150 * We might need to add a TTM.
1152 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1153 ret = ttm_bo_add_ttm(bo, true);
1159 EXPORT_SYMBOL(ttm_bo_validate);
1161 int ttm_bo_check_placement(struct ttm_buffer_object *bo,
1162 struct ttm_placement *placement)
1164 BUG_ON((placement->fpfn || placement->lpfn) &&
1165 (bo->mem.num_pages > (placement->lpfn - placement->fpfn)));
1170 int ttm_bo_init(struct ttm_bo_device *bdev,
1171 struct ttm_buffer_object *bo,
1173 enum ttm_bo_type type,
1174 struct ttm_placement *placement,
1175 uint32_t page_alignment,
1176 unsigned long buffer_start,
1178 struct file *persistent_swap_storage,
1180 void (*destroy) (struct ttm_buffer_object *))
1183 unsigned long num_pages;
1185 size += buffer_start & ~PAGE_MASK;
1186 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1187 if (num_pages == 0) {
1188 printk(KERN_ERR TTM_PFX "Illegal buffer object size.\n");
1195 bo->destroy = destroy;
1197 kref_init(&bo->kref);
1198 kref_init(&bo->list_kref);
1199 atomic_set(&bo->cpu_writers, 0);
1200 atomic_set(&bo->reserved, 1);
1201 init_waitqueue_head(&bo->event_queue);
1202 INIT_LIST_HEAD(&bo->lru);
1203 INIT_LIST_HEAD(&bo->ddestroy);
1204 INIT_LIST_HEAD(&bo->swap);
1205 INIT_LIST_HEAD(&bo->io_reserve_lru);
1207 bo->glob = bdev->glob;
1209 bo->num_pages = num_pages;
1210 bo->mem.size = num_pages << PAGE_SHIFT;
1211 bo->mem.mem_type = TTM_PL_SYSTEM;
1212 bo->mem.num_pages = bo->num_pages;
1213 bo->mem.mm_node = NULL;
1214 bo->mem.page_alignment = page_alignment;
1215 bo->mem.bus.io_reserved_vm = false;
1216 bo->mem.bus.io_reserved_count = 0;
1217 bo->buffer_start = buffer_start & PAGE_MASK;
1219 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1220 bo->seq_valid = false;
1221 bo->persistent_swap_storage = persistent_swap_storage;
1222 bo->acc_size = acc_size;
1223 atomic_inc(&bo->glob->bo_count);
1225 ret = ttm_bo_check_placement(bo, placement);
1226 if (unlikely(ret != 0))
1230 * For ttm_bo_type_device buffers, allocate
1231 * address space from the device.
1233 if (bo->type == ttm_bo_type_device) {
1234 ret = ttm_bo_setup_vm(bo);
1239 ret = ttm_bo_validate(bo, placement, interruptible, false, false);
1243 ttm_bo_unreserve(bo);
1247 ttm_bo_unreserve(bo);
1252 EXPORT_SYMBOL(ttm_bo_init);
1254 static inline size_t ttm_bo_size(struct ttm_bo_global *glob,
1255 unsigned long num_pages)
1257 size_t page_array_size = (num_pages * sizeof(void *) + PAGE_SIZE - 1) &
1260 return glob->ttm_bo_size + 2 * page_array_size;
1263 int ttm_bo_create(struct ttm_bo_device *bdev,
1265 enum ttm_bo_type type,
1266 struct ttm_placement *placement,
1267 uint32_t page_alignment,
1268 unsigned long buffer_start,
1270 struct file *persistent_swap_storage,
1271 struct ttm_buffer_object **p_bo)
1273 struct ttm_buffer_object *bo;
1274 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1278 ttm_bo_size(bdev->glob, (size + PAGE_SIZE - 1) >> PAGE_SHIFT);
1279 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1280 if (unlikely(ret != 0))
1283 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1285 if (unlikely(bo == NULL)) {
1286 ttm_mem_global_free(mem_glob, acc_size);
1290 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1291 buffer_start, interruptible,
1292 persistent_swap_storage, acc_size, NULL);
1293 if (likely(ret == 0))
1298 EXPORT_SYMBOL(ttm_bo_create);
1300 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1301 unsigned mem_type, bool allow_errors)
1303 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1304 struct ttm_bo_global *glob = bdev->glob;
1308 * Can't use standard list traversal since we're unlocking.
1311 spin_lock(&glob->lru_lock);
1312 while (!list_empty(&man->lru)) {
1313 spin_unlock(&glob->lru_lock);
1314 ret = ttm_mem_evict_first(bdev, mem_type, false, false, false);
1319 printk(KERN_ERR TTM_PFX
1320 "Cleanup eviction failed\n");
1323 spin_lock(&glob->lru_lock);
1325 spin_unlock(&glob->lru_lock);
1329 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1331 struct ttm_mem_type_manager *man;
1334 if (mem_type >= TTM_NUM_MEM_TYPES) {
1335 printk(KERN_ERR TTM_PFX "Illegal memory type %d\n", mem_type);
1338 man = &bdev->man[mem_type];
1340 if (!man->has_type) {
1341 printk(KERN_ERR TTM_PFX "Trying to take down uninitialized "
1342 "memory manager type %u\n", mem_type);
1346 man->use_type = false;
1347 man->has_type = false;
1351 ttm_bo_force_list_clean(bdev, mem_type, false);
1353 ret = (*man->func->takedown)(man);
1358 EXPORT_SYMBOL(ttm_bo_clean_mm);
1360 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1362 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1364 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1365 printk(KERN_ERR TTM_PFX
1366 "Illegal memory manager memory type %u.\n",
1371 if (!man->has_type) {
1372 printk(KERN_ERR TTM_PFX
1373 "Memory type %u has not been initialized.\n",
1378 return ttm_bo_force_list_clean(bdev, mem_type, true);
1380 EXPORT_SYMBOL(ttm_bo_evict_mm);
1382 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1383 unsigned long p_size)
1386 struct ttm_mem_type_manager *man;
1388 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1389 man = &bdev->man[type];
1390 BUG_ON(man->has_type);
1391 man->io_reserve_fastpath = true;
1392 man->use_io_reserve_lru = false;
1393 mutex_init(&man->io_reserve_mutex);
1394 INIT_LIST_HEAD(&man->io_reserve_lru);
1396 ret = bdev->driver->init_mem_type(bdev, type, man);
1402 if (type != TTM_PL_SYSTEM) {
1403 ret = (*man->func->init)(man, p_size);
1407 man->has_type = true;
1408 man->use_type = true;
1411 INIT_LIST_HEAD(&man->lru);
1415 EXPORT_SYMBOL(ttm_bo_init_mm);
1417 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1419 struct ttm_bo_global *glob =
1420 container_of(kobj, struct ttm_bo_global, kobj);
1422 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
1423 __free_page(glob->dummy_read_page);
1427 void ttm_bo_global_release(struct drm_global_reference *ref)
1429 struct ttm_bo_global *glob = ref->object;
1431 kobject_del(&glob->kobj);
1432 kobject_put(&glob->kobj);
1434 EXPORT_SYMBOL(ttm_bo_global_release);
1436 int ttm_bo_global_init(struct drm_global_reference *ref)
1438 struct ttm_bo_global_ref *bo_ref =
1439 container_of(ref, struct ttm_bo_global_ref, ref);
1440 struct ttm_bo_global *glob = ref->object;
1443 mutex_init(&glob->device_list_mutex);
1444 spin_lock_init(&glob->lru_lock);
1445 glob->mem_glob = bo_ref->mem_glob;
1446 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1448 if (unlikely(glob->dummy_read_page == NULL)) {
1453 INIT_LIST_HEAD(&glob->swap_lru);
1454 INIT_LIST_HEAD(&glob->device_list);
1456 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
1457 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
1458 if (unlikely(ret != 0)) {
1459 printk(KERN_ERR TTM_PFX
1460 "Could not register buffer object swapout.\n");
1464 glob->ttm_bo_extra_size =
1465 ttm_round_pot(sizeof(struct ttm_tt)) +
1466 ttm_round_pot(sizeof(struct ttm_backend));
1468 glob->ttm_bo_size = glob->ttm_bo_extra_size +
1469 ttm_round_pot(sizeof(struct ttm_buffer_object));
1471 atomic_set(&glob->bo_count, 0);
1473 ret = kobject_init_and_add(
1474 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1475 if (unlikely(ret != 0))
1476 kobject_put(&glob->kobj);
1479 __free_page(glob->dummy_read_page);
1484 EXPORT_SYMBOL(ttm_bo_global_init);
1487 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1490 unsigned i = TTM_NUM_MEM_TYPES;
1491 struct ttm_mem_type_manager *man;
1492 struct ttm_bo_global *glob = bdev->glob;
1495 man = &bdev->man[i];
1496 if (man->has_type) {
1497 man->use_type = false;
1498 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1500 printk(KERN_ERR TTM_PFX
1501 "DRM memory manager type %d "
1502 "is not clean.\n", i);
1504 man->has_type = false;
1508 mutex_lock(&glob->device_list_mutex);
1509 list_del(&bdev->device_list);
1510 mutex_unlock(&glob->device_list_mutex);
1512 cancel_delayed_work_sync(&bdev->wq);
1514 while (ttm_bo_delayed_delete(bdev, true))
1517 spin_lock(&glob->lru_lock);
1518 if (list_empty(&bdev->ddestroy))
1519 TTM_DEBUG("Delayed destroy list was clean\n");
1521 if (list_empty(&bdev->man[0].lru))
1522 TTM_DEBUG("Swap list was clean\n");
1523 spin_unlock(&glob->lru_lock);
1525 BUG_ON(!drm_mm_clean(&bdev->addr_space_mm));
1526 write_lock(&bdev->vm_lock);
1527 drm_mm_takedown(&bdev->addr_space_mm);
1528 write_unlock(&bdev->vm_lock);
1532 EXPORT_SYMBOL(ttm_bo_device_release);
1534 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1535 struct ttm_bo_global *glob,
1536 struct ttm_bo_driver *driver,
1537 uint64_t file_page_offset,
1542 rwlock_init(&bdev->vm_lock);
1543 bdev->driver = driver;
1545 memset(bdev->man, 0, sizeof(bdev->man));
1548 * Initialize the system memory buffer type.
1549 * Other types need to be driver / IOCTL initialized.
1551 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1552 if (unlikely(ret != 0))
1555 bdev->addr_space_rb = RB_ROOT;
1556 ret = drm_mm_init(&bdev->addr_space_mm, file_page_offset, 0x10000000);
1557 if (unlikely(ret != 0))
1558 goto out_no_addr_mm;
1560 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1561 bdev->nice_mode = true;
1562 INIT_LIST_HEAD(&bdev->ddestroy);
1563 bdev->dev_mapping = NULL;
1565 bdev->need_dma32 = need_dma32;
1567 spin_lock_init(&bdev->fence_lock);
1568 mutex_lock(&glob->device_list_mutex);
1569 list_add_tail(&bdev->device_list, &glob->device_list);
1570 mutex_unlock(&glob->device_list_mutex);
1574 ttm_bo_clean_mm(bdev, 0);
1578 EXPORT_SYMBOL(ttm_bo_device_init);
1581 * buffer object vm functions.
1584 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1586 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1588 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1589 if (mem->mem_type == TTM_PL_SYSTEM)
1592 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1595 if (mem->placement & TTM_PL_FLAG_CACHED)
1601 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1603 struct ttm_bo_device *bdev = bo->bdev;
1604 loff_t offset = (loff_t) bo->addr_space_offset;
1605 loff_t holelen = ((loff_t) bo->mem.num_pages) << PAGE_SHIFT;
1607 if (!bdev->dev_mapping)
1609 unmap_mapping_range(bdev->dev_mapping, offset, holelen, 1);
1610 ttm_mem_io_free_vm(bo);
1613 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1615 struct ttm_bo_device *bdev = bo->bdev;
1616 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1618 ttm_mem_io_lock(man, false);
1619 ttm_bo_unmap_virtual_locked(bo);
1620 ttm_mem_io_unlock(man);
1624 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1626 static void ttm_bo_vm_insert_rb(struct ttm_buffer_object *bo)
1628 struct ttm_bo_device *bdev = bo->bdev;
1629 struct rb_node **cur = &bdev->addr_space_rb.rb_node;
1630 struct rb_node *parent = NULL;
1631 struct ttm_buffer_object *cur_bo;
1632 unsigned long offset = bo->vm_node->start;
1633 unsigned long cur_offset;
1637 cur_bo = rb_entry(parent, struct ttm_buffer_object, vm_rb);
1638 cur_offset = cur_bo->vm_node->start;
1639 if (offset < cur_offset)
1640 cur = &parent->rb_left;
1641 else if (offset > cur_offset)
1642 cur = &parent->rb_right;
1647 rb_link_node(&bo->vm_rb, parent, cur);
1648 rb_insert_color(&bo->vm_rb, &bdev->addr_space_rb);
1654 * @bo: the buffer to allocate address space for
1656 * Allocate address space in the drm device so that applications
1657 * can mmap the buffer and access the contents. This only
1658 * applies to ttm_bo_type_device objects as others are not
1659 * placed in the drm device address space.
1662 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo)
1664 struct ttm_bo_device *bdev = bo->bdev;
1668 ret = drm_mm_pre_get(&bdev->addr_space_mm);
1669 if (unlikely(ret != 0))
1672 write_lock(&bdev->vm_lock);
1673 bo->vm_node = drm_mm_search_free(&bdev->addr_space_mm,
1674 bo->mem.num_pages, 0, 0);
1676 if (unlikely(bo->vm_node == NULL)) {
1681 bo->vm_node = drm_mm_get_block_atomic(bo->vm_node,
1682 bo->mem.num_pages, 0);
1684 if (unlikely(bo->vm_node == NULL)) {
1685 write_unlock(&bdev->vm_lock);
1689 ttm_bo_vm_insert_rb(bo);
1690 write_unlock(&bdev->vm_lock);
1691 bo->addr_space_offset = ((uint64_t) bo->vm_node->start) << PAGE_SHIFT;
1695 write_unlock(&bdev->vm_lock);
1699 static void ttm_bo_unref_sync_obj_locked(struct ttm_buffer_object *bo,
1701 void **extra_sync_obj)
1703 struct ttm_bo_device *bdev = bo->bdev;
1704 struct ttm_bo_driver *driver = bdev->driver;
1705 void *tmp_obj = NULL, *tmp_obj_read = NULL, *tmp_obj_write = NULL;
1707 /* We must unref the sync obj wherever it's ref'd.
1708 * Note that if we unref bo->sync_obj, we can unref both the read
1709 * and write sync objs too, because they can't be newer than
1710 * bo->sync_obj, so they are no longer relevant. */
1711 if (sync_obj == bo->sync_obj ||
1712 sync_obj == bo->sync_obj_read) {
1713 tmp_obj_read = bo->sync_obj_read;
1714 bo->sync_obj_read = NULL;
1716 if (sync_obj == bo->sync_obj ||
1717 sync_obj == bo->sync_obj_write) {
1718 tmp_obj_write = bo->sync_obj_write;
1719 bo->sync_obj_write = NULL;
1721 if (sync_obj == bo->sync_obj) {
1722 tmp_obj = bo->sync_obj;
1723 bo->sync_obj = NULL;
1726 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
1727 spin_unlock(&bdev->fence_lock);
1729 driver->sync_obj_unref(&tmp_obj);
1731 driver->sync_obj_unref(&tmp_obj_read);
1733 driver->sync_obj_unref(&tmp_obj_write);
1735 driver->sync_obj_unref(extra_sync_obj);
1736 spin_lock(&bdev->fence_lock);
1739 int ttm_bo_wait(struct ttm_buffer_object *bo,
1740 bool lazy, bool interruptible, bool no_wait,
1741 enum ttm_buffer_usage usage)
1743 struct ttm_bo_driver *driver = bo->bdev->driver;
1744 struct ttm_bo_device *bdev = bo->bdev;
1751 case TTM_USAGE_READ:
1752 bo_sync_obj = &bo->sync_obj_read;
1754 case TTM_USAGE_WRITE:
1755 bo_sync_obj = &bo->sync_obj_write;
1757 case TTM_USAGE_READWRITE:
1759 bo_sync_obj = &bo->sync_obj;
1762 if (likely(*bo_sync_obj == NULL))
1765 while (*bo_sync_obj) {
1767 if (driver->sync_obj_signaled(*bo_sync_obj, bo->sync_obj_arg)) {
1768 ttm_bo_unref_sync_obj_locked(bo, *bo_sync_obj, NULL);
1775 sync_obj = driver->sync_obj_ref(*bo_sync_obj);
1776 sync_obj_arg = bo->sync_obj_arg;
1777 spin_unlock(&bdev->fence_lock);
1778 ret = driver->sync_obj_wait(sync_obj, sync_obj_arg,
1779 lazy, interruptible);
1780 if (unlikely(ret != 0)) {
1781 driver->sync_obj_unref(&sync_obj);
1782 spin_lock(&bdev->fence_lock);
1785 spin_lock(&bdev->fence_lock);
1786 if (likely(*bo_sync_obj == sync_obj &&
1787 bo->sync_obj_arg == sync_obj_arg)) {
1788 ttm_bo_unref_sync_obj_locked(bo, *bo_sync_obj, &sync_obj);
1790 spin_unlock(&bdev->fence_lock);
1791 driver->sync_obj_unref(&sync_obj);
1792 spin_lock(&bdev->fence_lock);
1797 EXPORT_SYMBOL(ttm_bo_wait);
1799 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1801 struct ttm_bo_device *bdev = bo->bdev;
1805 * Using ttm_bo_reserve makes sure the lru lists are updated.
1808 ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
1809 if (unlikely(ret != 0))
1811 spin_lock(&bdev->fence_lock);
1812 ret = ttm_bo_wait(bo, false, true, no_wait, TTM_USAGE_READWRITE);
1813 spin_unlock(&bdev->fence_lock);
1814 if (likely(ret == 0))
1815 atomic_inc(&bo->cpu_writers);
1816 ttm_bo_unreserve(bo);
1819 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1821 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1823 if (atomic_dec_and_test(&bo->cpu_writers))
1824 wake_up_all(&bo->event_queue);
1826 EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1829 * A buffer object shrink method that tries to swap out the first
1830 * buffer object on the bo_global::swap_lru list.
1833 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1835 struct ttm_bo_global *glob =
1836 container_of(shrink, struct ttm_bo_global, shrink);
1837 struct ttm_buffer_object *bo;
1840 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM);
1842 spin_lock(&glob->lru_lock);
1843 while (ret == -EBUSY) {
1844 if (unlikely(list_empty(&glob->swap_lru))) {
1845 spin_unlock(&glob->lru_lock);
1849 bo = list_first_entry(&glob->swap_lru,
1850 struct ttm_buffer_object, swap);
1851 kref_get(&bo->list_kref);
1853 if (!list_empty(&bo->ddestroy)) {
1854 spin_unlock(&glob->lru_lock);
1855 (void) ttm_bo_cleanup_refs(bo, false, false, false);
1856 kref_put(&bo->list_kref, ttm_bo_release_list);
1861 * Reserve buffer. Since we unlock while sleeping, we need
1862 * to re-check that nobody removed us from the swap-list while
1866 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
1867 if (unlikely(ret == -EBUSY)) {
1868 spin_unlock(&glob->lru_lock);
1869 ttm_bo_wait_unreserved(bo, false);
1870 kref_put(&bo->list_kref, ttm_bo_release_list);
1871 spin_lock(&glob->lru_lock);
1876 put_count = ttm_bo_del_from_lru(bo);
1877 spin_unlock(&glob->lru_lock);
1879 ttm_bo_list_ref_sub(bo, put_count, true);
1882 * Wait for GPU, then move to system cached.
1885 spin_lock(&bo->bdev->fence_lock);
1886 ret = ttm_bo_wait(bo, false, false, false, TTM_USAGE_READWRITE);
1887 spin_unlock(&bo->bdev->fence_lock);
1889 if (unlikely(ret != 0))
1892 if ((bo->mem.placement & swap_placement) != swap_placement) {
1893 struct ttm_mem_reg evict_mem;
1895 evict_mem = bo->mem;
1896 evict_mem.mm_node = NULL;
1897 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1898 evict_mem.mem_type = TTM_PL_SYSTEM;
1900 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1901 false, false, false);
1902 if (unlikely(ret != 0))
1906 ttm_bo_unmap_virtual(bo);
1909 * Swap out. Buffer will be swapped in again as soon as
1910 * anyone tries to access a ttm page.
1913 if (bo->bdev->driver->swap_notify)
1914 bo->bdev->driver->swap_notify(bo);
1916 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1921 * Unreserve without putting on LRU to avoid swapping out an
1922 * already swapped buffer.
1925 atomic_set(&bo->reserved, 0);
1926 wake_up_all(&bo->event_queue);
1927 kref_put(&bo->list_kref, ttm_bo_release_list);
1931 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1933 while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1936 EXPORT_SYMBOL(ttm_bo_swapout_all);