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>
41 #define TTM_ASSERT_LOCKED(param)
42 #define TTM_DEBUG(fmt, arg...)
43 #define TTM_BO_HASH_ORDER 13
45 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo);
46 static void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo);
47 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink);
49 static inline uint32_t ttm_bo_type_flags(unsigned type)
54 static void ttm_bo_release_list(struct kref *list_kref)
56 struct ttm_buffer_object *bo =
57 container_of(list_kref, struct ttm_buffer_object, list_kref);
58 struct ttm_bo_device *bdev = bo->bdev;
60 BUG_ON(atomic_read(&bo->list_kref.refcount));
61 BUG_ON(atomic_read(&bo->kref.refcount));
62 BUG_ON(atomic_read(&bo->cpu_writers));
63 BUG_ON(bo->sync_obj != NULL);
64 BUG_ON(bo->mem.mm_node != NULL);
65 BUG_ON(!list_empty(&bo->lru));
66 BUG_ON(!list_empty(&bo->ddestroy));
69 ttm_tt_destroy(bo->ttm);
73 ttm_mem_global_free(bdev->mem_glob, bo->acc_size, false);
78 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo, bool interruptible)
84 ret = wait_event_interruptible(bo->event_queue,
85 atomic_read(&bo->reserved) == 0);
86 if (unlikely(ret != 0))
89 wait_event(bo->event_queue, atomic_read(&bo->reserved) == 0);
94 static void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
96 struct ttm_bo_device *bdev = bo->bdev;
97 struct ttm_mem_type_manager *man;
99 BUG_ON(!atomic_read(&bo->reserved));
101 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
103 BUG_ON(!list_empty(&bo->lru));
105 man = &bdev->man[bo->mem.mem_type];
106 list_add_tail(&bo->lru, &man->lru);
107 kref_get(&bo->list_kref);
109 if (bo->ttm != NULL) {
110 list_add_tail(&bo->swap, &bdev->swap_lru);
111 kref_get(&bo->list_kref);
117 * Call with the lru_lock held.
120 static int ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
124 if (!list_empty(&bo->swap)) {
125 list_del_init(&bo->swap);
128 if (!list_empty(&bo->lru)) {
129 list_del_init(&bo->lru);
134 * TODO: Add a driver hook to delete from
135 * driver-specific LRU's here.
141 int ttm_bo_reserve_locked(struct ttm_buffer_object *bo,
143 bool no_wait, bool use_sequence, uint32_t sequence)
145 struct ttm_bo_device *bdev = bo->bdev;
148 while (unlikely(atomic_cmpxchg(&bo->reserved, 0, 1) != 0)) {
149 if (use_sequence && bo->seq_valid &&
150 (sequence - bo->val_seq < (1 << 31))) {
157 spin_unlock(&bdev->lru_lock);
158 ret = ttm_bo_wait_unreserved(bo, interruptible);
159 spin_lock(&bdev->lru_lock);
166 bo->val_seq = sequence;
167 bo->seq_valid = true;
169 bo->seq_valid = false;
174 EXPORT_SYMBOL(ttm_bo_reserve);
176 static void ttm_bo_ref_bug(struct kref *list_kref)
181 int ttm_bo_reserve(struct ttm_buffer_object *bo,
183 bool no_wait, bool use_sequence, uint32_t sequence)
185 struct ttm_bo_device *bdev = bo->bdev;
189 spin_lock(&bdev->lru_lock);
190 ret = ttm_bo_reserve_locked(bo, interruptible, no_wait, use_sequence,
192 if (likely(ret == 0))
193 put_count = ttm_bo_del_from_lru(bo);
194 spin_unlock(&bdev->lru_lock);
197 kref_put(&bo->list_kref, ttm_bo_ref_bug);
202 void ttm_bo_unreserve(struct ttm_buffer_object *bo)
204 struct ttm_bo_device *bdev = bo->bdev;
206 spin_lock(&bdev->lru_lock);
207 ttm_bo_add_to_lru(bo);
208 atomic_set(&bo->reserved, 0);
209 wake_up_all(&bo->event_queue);
210 spin_unlock(&bdev->lru_lock);
212 EXPORT_SYMBOL(ttm_bo_unreserve);
215 * Call bo->mutex locked.
218 static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
220 struct ttm_bo_device *bdev = bo->bdev;
222 uint32_t page_flags = 0;
224 TTM_ASSERT_LOCKED(&bo->mutex);
228 case ttm_bo_type_device:
230 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
231 case ttm_bo_type_kernel:
232 bo->ttm = ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
233 page_flags, bdev->dummy_read_page);
234 if (unlikely(bo->ttm == NULL))
237 case ttm_bo_type_user:
238 bo->ttm = ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
239 page_flags | TTM_PAGE_FLAG_USER,
240 bdev->dummy_read_page);
241 if (unlikely(bo->ttm == NULL))
245 ret = ttm_tt_set_user(bo->ttm, current,
246 bo->buffer_start, bo->num_pages);
247 if (unlikely(ret != 0))
248 ttm_tt_destroy(bo->ttm);
251 printk(KERN_ERR TTM_PFX "Illegal buffer object type\n");
259 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
260 struct ttm_mem_reg *mem,
261 bool evict, bool interruptible, bool no_wait)
263 struct ttm_bo_device *bdev = bo->bdev;
264 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
265 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
266 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
267 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
270 if (old_is_pci || new_is_pci ||
271 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0))
272 ttm_bo_unmap_virtual(bo);
275 * Create and bind a ttm if required.
278 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && (bo->ttm == NULL)) {
279 ret = ttm_bo_add_ttm(bo, false);
283 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
287 if (mem->mem_type != TTM_PL_SYSTEM) {
288 ret = ttm_tt_bind(bo->ttm, mem);
293 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
295 struct ttm_mem_reg *old_mem = &bo->mem;
296 uint32_t save_flags = old_mem->placement;
300 ttm_flag_masked(&save_flags, mem->placement,
301 TTM_PL_MASK_MEMTYPE);
307 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
308 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
309 ret = ttm_bo_move_ttm(bo, evict, no_wait, mem);
310 else if (bdev->driver->move)
311 ret = bdev->driver->move(bo, evict, interruptible,
314 ret = ttm_bo_move_memcpy(bo, evict, no_wait, mem);
321 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
323 printk(KERN_ERR TTM_PFX "Can not flush read caches\n");
327 if (bo->mem.mm_node) {
328 spin_lock(&bo->lock);
329 bo->offset = (bo->mem.mm_node->start << PAGE_SHIFT) +
330 bdev->man[bo->mem.mem_type].gpu_offset;
331 bo->cur_placement = bo->mem.placement;
332 spin_unlock(&bo->lock);
338 new_man = &bdev->man[bo->mem.mem_type];
339 if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) {
340 ttm_tt_unbind(bo->ttm);
341 ttm_tt_destroy(bo->ttm);
349 * If bo idle, remove from delayed- and lru lists, and unref.
350 * If not idle, and already on delayed list, do nothing.
351 * If not idle, and not on delayed list, put on delayed list,
352 * up the list_kref and schedule a delayed list check.
355 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo, bool remove_all)
357 struct ttm_bo_device *bdev = bo->bdev;
358 struct ttm_bo_driver *driver = bdev->driver;
361 spin_lock(&bo->lock);
362 (void) ttm_bo_wait(bo, false, false, !remove_all);
367 spin_unlock(&bo->lock);
369 spin_lock(&bdev->lru_lock);
370 ret = ttm_bo_reserve_locked(bo, false, false, false, 0);
373 ttm_tt_unbind(bo->ttm);
375 if (!list_empty(&bo->ddestroy)) {
376 list_del_init(&bo->ddestroy);
377 kref_put(&bo->list_kref, ttm_bo_ref_bug);
379 if (bo->mem.mm_node) {
380 drm_mm_put_block(bo->mem.mm_node);
381 bo->mem.mm_node = NULL;
383 put_count = ttm_bo_del_from_lru(bo);
384 spin_unlock(&bdev->lru_lock);
386 atomic_set(&bo->reserved, 0);
389 kref_put(&bo->list_kref, ttm_bo_release_list);
394 spin_lock(&bdev->lru_lock);
395 if (list_empty(&bo->ddestroy)) {
396 void *sync_obj = bo->sync_obj;
397 void *sync_obj_arg = bo->sync_obj_arg;
399 kref_get(&bo->list_kref);
400 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
401 spin_unlock(&bdev->lru_lock);
402 spin_unlock(&bo->lock);
405 driver->sync_obj_flush(sync_obj, sync_obj_arg);
406 schedule_delayed_work(&bdev->wq,
407 ((HZ / 100) < 1) ? 1 : HZ / 100);
411 spin_unlock(&bdev->lru_lock);
412 spin_unlock(&bo->lock);
420 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
421 * encountered buffers.
424 static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
426 struct ttm_buffer_object *entry, *nentry;
427 struct list_head *list, *next;
430 spin_lock(&bdev->lru_lock);
431 list_for_each_safe(list, next, &bdev->ddestroy) {
432 entry = list_entry(list, struct ttm_buffer_object, ddestroy);
436 * Protect the next list entry from destruction while we
437 * unlock the lru_lock.
440 if (next != &bdev->ddestroy) {
441 nentry = list_entry(next, struct ttm_buffer_object,
443 kref_get(&nentry->list_kref);
445 kref_get(&entry->list_kref);
447 spin_unlock(&bdev->lru_lock);
448 ret = ttm_bo_cleanup_refs(entry, remove_all);
449 kref_put(&entry->list_kref, ttm_bo_release_list);
451 spin_lock(&bdev->lru_lock);
453 bool next_onlist = !list_empty(next);
454 spin_unlock(&bdev->lru_lock);
455 kref_put(&nentry->list_kref, ttm_bo_release_list);
456 spin_lock(&bdev->lru_lock);
458 * Someone might have raced us and removed the
459 * next entry from the list. We don't bother restarting
469 ret = !list_empty(&bdev->ddestroy);
470 spin_unlock(&bdev->lru_lock);
475 static void ttm_bo_delayed_workqueue(struct work_struct *work)
477 struct ttm_bo_device *bdev =
478 container_of(work, struct ttm_bo_device, wq.work);
480 if (ttm_bo_delayed_delete(bdev, false)) {
481 schedule_delayed_work(&bdev->wq,
482 ((HZ / 100) < 1) ? 1 : HZ / 100);
486 static void ttm_bo_release(struct kref *kref)
488 struct ttm_buffer_object *bo =
489 container_of(kref, struct ttm_buffer_object, kref);
490 struct ttm_bo_device *bdev = bo->bdev;
492 if (likely(bo->vm_node != NULL)) {
493 rb_erase(&bo->vm_rb, &bdev->addr_space_rb);
494 drm_mm_put_block(bo->vm_node);
497 write_unlock(&bdev->vm_lock);
498 ttm_bo_cleanup_refs(bo, false);
499 kref_put(&bo->list_kref, ttm_bo_release_list);
500 write_lock(&bdev->vm_lock);
503 void ttm_bo_unref(struct ttm_buffer_object **p_bo)
505 struct ttm_buffer_object *bo = *p_bo;
506 struct ttm_bo_device *bdev = bo->bdev;
509 write_lock(&bdev->vm_lock);
510 kref_put(&bo->kref, ttm_bo_release);
511 write_unlock(&bdev->vm_lock);
513 EXPORT_SYMBOL(ttm_bo_unref);
515 static int ttm_bo_evict(struct ttm_buffer_object *bo, unsigned mem_type,
516 bool interruptible, bool no_wait)
519 struct ttm_bo_device *bdev = bo->bdev;
520 struct ttm_mem_reg evict_mem;
521 uint32_t proposed_placement;
523 if (bo->mem.mem_type != mem_type)
526 spin_lock(&bo->lock);
527 ret = ttm_bo_wait(bo, false, interruptible, no_wait);
528 spin_unlock(&bo->lock);
530 if (ret && ret != -ERESTART) {
531 printk(KERN_ERR TTM_PFX "Failed to expire sync object before "
532 "buffer eviction.\n");
536 BUG_ON(!atomic_read(&bo->reserved));
539 evict_mem.mm_node = NULL;
541 proposed_placement = bdev->driver->evict_flags(bo);
543 ret = ttm_bo_mem_space(bo, proposed_placement,
544 &evict_mem, interruptible, no_wait);
545 if (unlikely(ret != 0 && ret != -ERESTART))
546 ret = ttm_bo_mem_space(bo, TTM_PL_FLAG_SYSTEM,
547 &evict_mem, interruptible, no_wait);
550 if (ret != -ERESTART)
551 printk(KERN_ERR TTM_PFX
552 "Failed to find memory space for "
553 "buffer 0x%p eviction.\n", bo);
557 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
560 if (ret != -ERESTART)
561 printk(KERN_ERR TTM_PFX "Buffer eviction failed\n");
565 spin_lock(&bdev->lru_lock);
566 if (evict_mem.mm_node) {
567 drm_mm_put_block(evict_mem.mm_node);
568 evict_mem.mm_node = NULL;
570 spin_unlock(&bdev->lru_lock);
577 * Repeatedly evict memory from the LRU for @mem_type until we create enough
578 * space, or we've evicted everything and there isn't enough space.
580 static int ttm_bo_mem_force_space(struct ttm_bo_device *bdev,
581 struct ttm_mem_reg *mem,
583 bool interruptible, bool no_wait)
585 struct drm_mm_node *node;
586 struct ttm_buffer_object *entry;
587 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
588 struct list_head *lru;
589 unsigned long num_pages = mem->num_pages;
594 ret = drm_mm_pre_get(&man->manager);
595 if (unlikely(ret != 0))
598 spin_lock(&bdev->lru_lock);
600 node = drm_mm_search_free(&man->manager, num_pages,
601 mem->page_alignment, 1);
609 entry = list_first_entry(lru, struct ttm_buffer_object, lru);
610 kref_get(&entry->list_kref);
613 ttm_bo_reserve_locked(entry, interruptible, no_wait,
616 if (likely(ret == 0))
617 put_count = ttm_bo_del_from_lru(entry);
619 spin_unlock(&bdev->lru_lock);
621 if (unlikely(ret != 0))
625 kref_put(&entry->list_kref, ttm_bo_ref_bug);
627 ret = ttm_bo_evict(entry, mem_type, interruptible, no_wait);
629 ttm_bo_unreserve(entry);
631 kref_put(&entry->list_kref, ttm_bo_release_list);
635 spin_lock(&bdev->lru_lock);
639 spin_unlock(&bdev->lru_lock);
643 node = drm_mm_get_block_atomic(node, num_pages, mem->page_alignment);
644 if (unlikely(!node)) {
645 spin_unlock(&bdev->lru_lock);
649 spin_unlock(&bdev->lru_lock);
651 mem->mem_type = mem_type;
655 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
658 uint32_t mask, uint32_t *res_mask)
660 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
662 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) && disallow_fixed)
665 if ((cur_flags & mask & TTM_PL_MASK_MEM) == 0)
668 if ((mask & man->available_caching) == 0)
670 if (mask & man->default_caching)
671 cur_flags |= man->default_caching;
672 else if (mask & TTM_PL_FLAG_CACHED)
673 cur_flags |= TTM_PL_FLAG_CACHED;
674 else if (mask & TTM_PL_FLAG_WC)
675 cur_flags |= TTM_PL_FLAG_WC;
677 cur_flags |= TTM_PL_FLAG_UNCACHED;
679 *res_mask = cur_flags;
684 * Creates space for memory region @mem according to its type.
686 * This function first searches for free space in compatible memory types in
687 * the priority order defined by the driver. If free space isn't found, then
688 * ttm_bo_mem_force_space is attempted in priority order to evict and find
691 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
692 uint32_t proposed_placement,
693 struct ttm_mem_reg *mem,
694 bool interruptible, bool no_wait)
696 struct ttm_bo_device *bdev = bo->bdev;
697 struct ttm_mem_type_manager *man;
699 uint32_t num_prios = bdev->driver->num_mem_type_prio;
700 const uint32_t *prios = bdev->driver->mem_type_prio;
702 uint32_t mem_type = TTM_PL_SYSTEM;
703 uint32_t cur_flags = 0;
704 bool type_found = false;
705 bool type_ok = false;
706 bool has_eagain = false;
707 struct drm_mm_node *node = NULL;
711 for (i = 0; i < num_prios; ++i) {
713 man = &bdev->man[mem_type];
715 type_ok = ttm_bo_mt_compatible(man,
716 bo->type == ttm_bo_type_user,
717 mem_type, proposed_placement,
723 if (mem_type == TTM_PL_SYSTEM)
726 if (man->has_type && man->use_type) {
729 ret = drm_mm_pre_get(&man->manager);
733 spin_lock(&bdev->lru_lock);
734 node = drm_mm_search_free(&man->manager,
738 if (unlikely(!node)) {
739 spin_unlock(&bdev->lru_lock);
742 node = drm_mm_get_block_atomic(node,
746 spin_unlock(&bdev->lru_lock);
753 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || node) {
755 mem->mem_type = mem_type;
756 mem->placement = cur_flags;
763 num_prios = bdev->driver->num_mem_busy_prio;
764 prios = bdev->driver->mem_busy_prio;
766 for (i = 0; i < num_prios; ++i) {
768 man = &bdev->man[mem_type];
773 if (!ttm_bo_mt_compatible(man,
774 bo->type == ttm_bo_type_user,
776 proposed_placement, &cur_flags))
779 ret = ttm_bo_mem_force_space(bdev, mem, mem_type,
780 interruptible, no_wait);
782 if (ret == 0 && mem->mm_node) {
783 mem->placement = cur_flags;
787 if (ret == -ERESTART)
791 ret = (has_eagain) ? -ERESTART : -ENOMEM;
794 EXPORT_SYMBOL(ttm_bo_mem_space);
796 int ttm_bo_wait_cpu(struct ttm_buffer_object *bo, bool no_wait)
800 if ((atomic_read(&bo->cpu_writers) > 0) && no_wait)
803 ret = wait_event_interruptible(bo->event_queue,
804 atomic_read(&bo->cpu_writers) == 0);
806 if (ret == -ERESTARTSYS)
812 int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
813 uint32_t proposed_placement,
814 bool interruptible, bool no_wait)
816 struct ttm_bo_device *bdev = bo->bdev;
818 struct ttm_mem_reg mem;
820 BUG_ON(!atomic_read(&bo->reserved));
823 * FIXME: It's possible to pipeline buffer moves.
824 * Have the driver move function wait for idle when necessary,
825 * instead of doing it here.
828 spin_lock(&bo->lock);
829 ret = ttm_bo_wait(bo, false, interruptible, no_wait);
830 spin_unlock(&bo->lock);
835 mem.num_pages = bo->num_pages;
836 mem.size = mem.num_pages << PAGE_SHIFT;
837 mem.page_alignment = bo->mem.page_alignment;
840 * Determine where to move the buffer.
843 ret = ttm_bo_mem_space(bo, proposed_placement, &mem,
844 interruptible, no_wait);
848 ret = ttm_bo_handle_move_mem(bo, &mem, false, interruptible, no_wait);
851 if (ret && mem.mm_node) {
852 spin_lock(&bdev->lru_lock);
853 drm_mm_put_block(mem.mm_node);
854 spin_unlock(&bdev->lru_lock);
859 static int ttm_bo_mem_compat(uint32_t proposed_placement,
860 struct ttm_mem_reg *mem)
862 if ((proposed_placement & mem->placement & TTM_PL_MASK_MEM) == 0)
864 if ((proposed_placement & mem->placement & TTM_PL_MASK_CACHING) == 0)
870 int ttm_buffer_object_validate(struct ttm_buffer_object *bo,
871 uint32_t proposed_placement,
872 bool interruptible, bool no_wait)
876 BUG_ON(!atomic_read(&bo->reserved));
877 bo->proposed_placement = proposed_placement;
879 TTM_DEBUG("Proposed placement 0x%08lx, Old flags 0x%08lx\n",
880 (unsigned long)proposed_placement,
881 (unsigned long)bo->mem.placement);
884 * Check whether we need to move buffer.
887 if (!ttm_bo_mem_compat(bo->proposed_placement, &bo->mem)) {
888 ret = ttm_bo_move_buffer(bo, bo->proposed_placement,
889 interruptible, no_wait);
891 if (ret != -ERESTART)
892 printk(KERN_ERR TTM_PFX
893 "Failed moving buffer. "
894 "Proposed placement 0x%08x\n",
895 bo->proposed_placement);
897 printk(KERN_ERR TTM_PFX
898 "Out of aperture space or "
899 "DRM memory quota.\n");
905 * We might need to add a TTM.
908 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
909 ret = ttm_bo_add_ttm(bo, true);
914 * Validation has succeeded, move the access and other
915 * non-mapping-related flag bits from the proposed flags to
919 ttm_flag_masked(&bo->mem.placement, bo->proposed_placement,
920 ~TTM_PL_MASK_MEMTYPE);
924 EXPORT_SYMBOL(ttm_buffer_object_validate);
927 ttm_bo_check_placement(struct ttm_buffer_object *bo,
928 uint32_t set_flags, uint32_t clr_flags)
930 uint32_t new_mask = set_flags | clr_flags;
932 if ((bo->type == ttm_bo_type_user) &&
933 (clr_flags & TTM_PL_FLAG_CACHED)) {
934 printk(KERN_ERR TTM_PFX
935 "User buffers require cache-coherent memory.\n");
939 if (!capable(CAP_SYS_ADMIN)) {
940 if (new_mask & TTM_PL_FLAG_NO_EVICT) {
941 printk(KERN_ERR TTM_PFX "Need to be root to modify"
942 " NO_EVICT status.\n");
946 if ((clr_flags & bo->mem.placement & TTM_PL_MASK_MEMTYPE) &&
947 (bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
948 printk(KERN_ERR TTM_PFX
949 "Incompatible memory specification"
950 " for NO_EVICT buffer.\n");
957 int ttm_buffer_object_init(struct ttm_bo_device *bdev,
958 struct ttm_buffer_object *bo,
960 enum ttm_bo_type type,
962 uint32_t page_alignment,
963 unsigned long buffer_start,
965 struct file *persistant_swap_storage,
967 void (*destroy) (struct ttm_buffer_object *))
970 unsigned long num_pages;
972 size += buffer_start & ~PAGE_MASK;
973 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
974 if (num_pages == 0) {
975 printk(KERN_ERR TTM_PFX "Illegal buffer object size.\n");
978 bo->destroy = destroy;
980 spin_lock_init(&bo->lock);
981 kref_init(&bo->kref);
982 kref_init(&bo->list_kref);
983 atomic_set(&bo->cpu_writers, 0);
984 atomic_set(&bo->reserved, 1);
985 init_waitqueue_head(&bo->event_queue);
986 INIT_LIST_HEAD(&bo->lru);
987 INIT_LIST_HEAD(&bo->ddestroy);
988 INIT_LIST_HEAD(&bo->swap);
991 bo->num_pages = num_pages;
992 bo->mem.mem_type = TTM_PL_SYSTEM;
993 bo->mem.num_pages = bo->num_pages;
994 bo->mem.mm_node = NULL;
995 bo->mem.page_alignment = page_alignment;
996 bo->buffer_start = buffer_start & PAGE_MASK;
998 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
999 bo->seq_valid = false;
1000 bo->persistant_swap_storage = persistant_swap_storage;
1001 bo->acc_size = acc_size;
1003 ret = ttm_bo_check_placement(bo, flags, 0ULL);
1004 if (unlikely(ret != 0))
1008 * If no caching attributes are set, accept any form of caching.
1011 if ((flags & TTM_PL_MASK_CACHING) == 0)
1012 flags |= TTM_PL_MASK_CACHING;
1015 * For ttm_bo_type_device buffers, allocate
1016 * address space from the device.
1019 if (bo->type == ttm_bo_type_device) {
1020 ret = ttm_bo_setup_vm(bo);
1025 ret = ttm_buffer_object_validate(bo, flags, interruptible, false);
1029 ttm_bo_unreserve(bo);
1033 ttm_bo_unreserve(bo);
1038 EXPORT_SYMBOL(ttm_buffer_object_init);
1040 static inline size_t ttm_bo_size(struct ttm_bo_device *bdev,
1041 unsigned long num_pages)
1043 size_t page_array_size = (num_pages * sizeof(void *) + PAGE_SIZE - 1) &
1046 return bdev->ttm_bo_size + 2 * page_array_size;
1049 int ttm_buffer_object_create(struct ttm_bo_device *bdev,
1051 enum ttm_bo_type type,
1053 uint32_t page_alignment,
1054 unsigned long buffer_start,
1056 struct file *persistant_swap_storage,
1057 struct ttm_buffer_object **p_bo)
1059 struct ttm_buffer_object *bo;
1061 struct ttm_mem_global *mem_glob = bdev->mem_glob;
1064 ttm_bo_size(bdev, (size + PAGE_SIZE - 1) >> PAGE_SHIFT);
1065 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false, false);
1066 if (unlikely(ret != 0))
1069 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1071 if (unlikely(bo == NULL)) {
1072 ttm_mem_global_free(mem_glob, acc_size, false);
1076 ret = ttm_buffer_object_init(bdev, bo, size, type, flags,
1077 page_alignment, buffer_start,
1079 persistant_swap_storage, acc_size, NULL);
1080 if (likely(ret == 0))
1086 static int ttm_bo_leave_list(struct ttm_buffer_object *bo,
1087 uint32_t mem_type, bool allow_errors)
1091 spin_lock(&bo->lock);
1092 ret = ttm_bo_wait(bo, false, false, false);
1093 spin_unlock(&bo->lock);
1095 if (ret && allow_errors)
1098 if (bo->mem.mem_type == mem_type)
1099 ret = ttm_bo_evict(bo, mem_type, false, false);
1106 printk(KERN_ERR TTM_PFX "Cleanup eviction failed\n");
1114 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1115 struct list_head *head,
1116 unsigned mem_type, bool allow_errors)
1118 struct ttm_buffer_object *entry;
1123 * Can't use standard list traversal since we're unlocking.
1126 spin_lock(&bdev->lru_lock);
1128 while (!list_empty(head)) {
1129 entry = list_first_entry(head, struct ttm_buffer_object, lru);
1130 kref_get(&entry->list_kref);
1131 ret = ttm_bo_reserve_locked(entry, false, false, false, 0);
1132 put_count = ttm_bo_del_from_lru(entry);
1133 spin_unlock(&bdev->lru_lock);
1135 kref_put(&entry->list_kref, ttm_bo_ref_bug);
1137 ret = ttm_bo_leave_list(entry, mem_type, allow_errors);
1138 ttm_bo_unreserve(entry);
1139 kref_put(&entry->list_kref, ttm_bo_release_list);
1140 spin_lock(&bdev->lru_lock);
1143 spin_unlock(&bdev->lru_lock);
1148 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1150 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1153 if (mem_type >= TTM_NUM_MEM_TYPES) {
1154 printk(KERN_ERR TTM_PFX "Illegal memory type %d\n", mem_type);
1158 if (!man->has_type) {
1159 printk(KERN_ERR TTM_PFX "Trying to take down uninitialized "
1160 "memory manager type %u\n", mem_type);
1164 man->use_type = false;
1165 man->has_type = false;
1169 ttm_bo_force_list_clean(bdev, &man->lru, mem_type, false);
1171 spin_lock(&bdev->lru_lock);
1172 if (drm_mm_clean(&man->manager))
1173 drm_mm_takedown(&man->manager);
1177 spin_unlock(&bdev->lru_lock);
1182 EXPORT_SYMBOL(ttm_bo_clean_mm);
1184 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1186 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1188 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1189 printk(KERN_ERR TTM_PFX
1190 "Illegal memory manager memory type %u.\n",
1195 if (!man->has_type) {
1196 printk(KERN_ERR TTM_PFX
1197 "Memory type %u has not been initialized.\n",
1202 return ttm_bo_force_list_clean(bdev, &man->lru, mem_type, true);
1204 EXPORT_SYMBOL(ttm_bo_evict_mm);
1206 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1207 unsigned long p_offset, unsigned long p_size)
1210 struct ttm_mem_type_manager *man;
1212 if (type >= TTM_NUM_MEM_TYPES) {
1213 printk(KERN_ERR TTM_PFX "Illegal memory type %d\n", type);
1217 man = &bdev->man[type];
1218 if (man->has_type) {
1219 printk(KERN_ERR TTM_PFX
1220 "Memory manager already initialized for type %d\n",
1225 ret = bdev->driver->init_mem_type(bdev, type, man);
1230 if (type != TTM_PL_SYSTEM) {
1232 printk(KERN_ERR TTM_PFX
1233 "Zero size memory manager type %d\n",
1237 ret = drm_mm_init(&man->manager, p_offset, p_size);
1241 man->has_type = true;
1242 man->use_type = true;
1245 INIT_LIST_HEAD(&man->lru);
1249 EXPORT_SYMBOL(ttm_bo_init_mm);
1251 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1254 unsigned i = TTM_NUM_MEM_TYPES;
1255 struct ttm_mem_type_manager *man;
1258 man = &bdev->man[i];
1259 if (man->has_type) {
1260 man->use_type = false;
1261 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1263 printk(KERN_ERR TTM_PFX
1264 "DRM memory manager type %d "
1265 "is not clean.\n", i);
1267 man->has_type = false;
1271 if (!cancel_delayed_work(&bdev->wq))
1272 flush_scheduled_work();
1274 while (ttm_bo_delayed_delete(bdev, true))
1277 spin_lock(&bdev->lru_lock);
1278 if (list_empty(&bdev->ddestroy))
1279 TTM_DEBUG("Delayed destroy list was clean\n");
1281 if (list_empty(&bdev->man[0].lru))
1282 TTM_DEBUG("Swap list was clean\n");
1283 spin_unlock(&bdev->lru_lock);
1285 ttm_mem_unregister_shrink(bdev->mem_glob, &bdev->shrink);
1286 BUG_ON(!drm_mm_clean(&bdev->addr_space_mm));
1287 write_lock(&bdev->vm_lock);
1288 drm_mm_takedown(&bdev->addr_space_mm);
1289 write_unlock(&bdev->vm_lock);
1291 __free_page(bdev->dummy_read_page);
1294 EXPORT_SYMBOL(ttm_bo_device_release);
1297 * This function is intended to be called on drm driver load.
1298 * If you decide to call it from firstopen, you must protect the call
1299 * from a potentially racing ttm_bo_driver_finish in lastclose.
1300 * (This may happen on X server restart).
1303 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1304 struct ttm_mem_global *mem_glob,
1305 struct ttm_bo_driver *driver, uint64_t file_page_offset)
1309 bdev->dummy_read_page = NULL;
1310 rwlock_init(&bdev->vm_lock);
1311 spin_lock_init(&bdev->lru_lock);
1313 bdev->driver = driver;
1314 bdev->mem_glob = mem_glob;
1316 memset(bdev->man, 0, sizeof(bdev->man));
1318 bdev->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1319 if (unlikely(bdev->dummy_read_page == NULL)) {
1325 * Initialize the system memory buffer type.
1326 * Other types need to be driver / IOCTL initialized.
1328 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0, 0);
1329 if (unlikely(ret != 0))
1332 bdev->addr_space_rb = RB_ROOT;
1333 ret = drm_mm_init(&bdev->addr_space_mm, file_page_offset, 0x10000000);
1334 if (unlikely(ret != 0))
1337 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1338 bdev->nice_mode = true;
1339 INIT_LIST_HEAD(&bdev->ddestroy);
1340 INIT_LIST_HEAD(&bdev->swap_lru);
1341 bdev->dev_mapping = NULL;
1342 ttm_mem_init_shrink(&bdev->shrink, ttm_bo_swapout);
1343 ret = ttm_mem_register_shrink(mem_glob, &bdev->shrink);
1344 if (unlikely(ret != 0)) {
1345 printk(KERN_ERR TTM_PFX
1346 "Could not register buffer object swapout.\n");
1350 bdev->ttm_bo_extra_size =
1351 ttm_round_pot(sizeof(struct ttm_tt)) +
1352 ttm_round_pot(sizeof(struct ttm_backend));
1354 bdev->ttm_bo_size = bdev->ttm_bo_extra_size +
1355 ttm_round_pot(sizeof(struct ttm_buffer_object));
1359 ttm_bo_clean_mm(bdev, 0);
1361 __free_page(bdev->dummy_read_page);
1365 EXPORT_SYMBOL(ttm_bo_device_init);
1368 * buffer object vm functions.
1371 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1373 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1375 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1376 if (mem->mem_type == TTM_PL_SYSTEM)
1379 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1382 if (mem->placement & TTM_PL_FLAG_CACHED)
1388 int ttm_bo_pci_offset(struct ttm_bo_device *bdev,
1389 struct ttm_mem_reg *mem,
1390 unsigned long *bus_base,
1391 unsigned long *bus_offset, unsigned long *bus_size)
1393 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1396 if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
1399 if (ttm_mem_reg_is_pci(bdev, mem)) {
1400 *bus_offset = mem->mm_node->start << PAGE_SHIFT;
1401 *bus_size = mem->num_pages << PAGE_SHIFT;
1402 *bus_base = man->io_offset;
1408 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1410 struct ttm_bo_device *bdev = bo->bdev;
1411 loff_t offset = (loff_t) bo->addr_space_offset;
1412 loff_t holelen = ((loff_t) bo->mem.num_pages) << PAGE_SHIFT;
1414 if (!bdev->dev_mapping)
1417 unmap_mapping_range(bdev->dev_mapping, offset, holelen, 1);
1420 static void ttm_bo_vm_insert_rb(struct ttm_buffer_object *bo)
1422 struct ttm_bo_device *bdev = bo->bdev;
1423 struct rb_node **cur = &bdev->addr_space_rb.rb_node;
1424 struct rb_node *parent = NULL;
1425 struct ttm_buffer_object *cur_bo;
1426 unsigned long offset = bo->vm_node->start;
1427 unsigned long cur_offset;
1431 cur_bo = rb_entry(parent, struct ttm_buffer_object, vm_rb);
1432 cur_offset = cur_bo->vm_node->start;
1433 if (offset < cur_offset)
1434 cur = &parent->rb_left;
1435 else if (offset > cur_offset)
1436 cur = &parent->rb_right;
1441 rb_link_node(&bo->vm_rb, parent, cur);
1442 rb_insert_color(&bo->vm_rb, &bdev->addr_space_rb);
1448 * @bo: the buffer to allocate address space for
1450 * Allocate address space in the drm device so that applications
1451 * can mmap the buffer and access the contents. This only
1452 * applies to ttm_bo_type_device objects as others are not
1453 * placed in the drm device address space.
1456 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo)
1458 struct ttm_bo_device *bdev = bo->bdev;
1462 ret = drm_mm_pre_get(&bdev->addr_space_mm);
1463 if (unlikely(ret != 0))
1466 write_lock(&bdev->vm_lock);
1467 bo->vm_node = drm_mm_search_free(&bdev->addr_space_mm,
1468 bo->mem.num_pages, 0, 0);
1470 if (unlikely(bo->vm_node == NULL)) {
1475 bo->vm_node = drm_mm_get_block_atomic(bo->vm_node,
1476 bo->mem.num_pages, 0);
1478 if (unlikely(bo->vm_node == NULL)) {
1479 write_unlock(&bdev->vm_lock);
1483 ttm_bo_vm_insert_rb(bo);
1484 write_unlock(&bdev->vm_lock);
1485 bo->addr_space_offset = ((uint64_t) bo->vm_node->start) << PAGE_SHIFT;
1489 write_unlock(&bdev->vm_lock);
1493 int ttm_bo_wait(struct ttm_buffer_object *bo,
1494 bool lazy, bool interruptible, bool no_wait)
1496 struct ttm_bo_driver *driver = bo->bdev->driver;
1501 if (likely(bo->sync_obj == NULL))
1504 while (bo->sync_obj) {
1506 if (driver->sync_obj_signaled(bo->sync_obj, bo->sync_obj_arg)) {
1507 void *tmp_obj = bo->sync_obj;
1508 bo->sync_obj = NULL;
1509 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
1510 spin_unlock(&bo->lock);
1511 driver->sync_obj_unref(&tmp_obj);
1512 spin_lock(&bo->lock);
1519 sync_obj = driver->sync_obj_ref(bo->sync_obj);
1520 sync_obj_arg = bo->sync_obj_arg;
1521 spin_unlock(&bo->lock);
1522 ret = driver->sync_obj_wait(sync_obj, sync_obj_arg,
1523 lazy, interruptible);
1524 if (unlikely(ret != 0)) {
1525 driver->sync_obj_unref(&sync_obj);
1526 spin_lock(&bo->lock);
1529 spin_lock(&bo->lock);
1530 if (likely(bo->sync_obj == sync_obj &&
1531 bo->sync_obj_arg == sync_obj_arg)) {
1532 void *tmp_obj = bo->sync_obj;
1533 bo->sync_obj = NULL;
1534 clear_bit(TTM_BO_PRIV_FLAG_MOVING,
1536 spin_unlock(&bo->lock);
1537 driver->sync_obj_unref(&sync_obj);
1538 driver->sync_obj_unref(&tmp_obj);
1539 spin_lock(&bo->lock);
1544 EXPORT_SYMBOL(ttm_bo_wait);
1546 void ttm_bo_unblock_reservation(struct ttm_buffer_object *bo)
1548 atomic_set(&bo->reserved, 0);
1549 wake_up_all(&bo->event_queue);
1552 int ttm_bo_block_reservation(struct ttm_buffer_object *bo, bool interruptible,
1557 while (unlikely(atomic_cmpxchg(&bo->reserved, 0, 1) != 0)) {
1560 else if (interruptible) {
1561 ret = wait_event_interruptible
1562 (bo->event_queue, atomic_read(&bo->reserved) == 0);
1563 if (unlikely(ret != 0))
1566 wait_event(bo->event_queue,
1567 atomic_read(&bo->reserved) == 0);
1573 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1578 * Using ttm_bo_reserve instead of ttm_bo_block_reservation
1579 * makes sure the lru lists are updated.
1582 ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
1583 if (unlikely(ret != 0))
1585 spin_lock(&bo->lock);
1586 ret = ttm_bo_wait(bo, false, true, no_wait);
1587 spin_unlock(&bo->lock);
1588 if (likely(ret == 0))
1589 atomic_inc(&bo->cpu_writers);
1590 ttm_bo_unreserve(bo);
1594 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1596 if (atomic_dec_and_test(&bo->cpu_writers))
1597 wake_up_all(&bo->event_queue);
1601 * A buffer object shrink method that tries to swap out the first
1602 * buffer object on the bo_global::swap_lru list.
1605 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1607 struct ttm_bo_device *bdev =
1608 container_of(shrink, struct ttm_bo_device, shrink);
1609 struct ttm_buffer_object *bo;
1612 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM);
1614 spin_lock(&bdev->lru_lock);
1615 while (ret == -EBUSY) {
1616 if (unlikely(list_empty(&bdev->swap_lru))) {
1617 spin_unlock(&bdev->lru_lock);
1621 bo = list_first_entry(&bdev->swap_lru,
1622 struct ttm_buffer_object, swap);
1623 kref_get(&bo->list_kref);
1626 * Reserve buffer. Since we unlock while sleeping, we need
1627 * to re-check that nobody removed us from the swap-list while
1631 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
1632 if (unlikely(ret == -EBUSY)) {
1633 spin_unlock(&bdev->lru_lock);
1634 ttm_bo_wait_unreserved(bo, false);
1635 kref_put(&bo->list_kref, ttm_bo_release_list);
1636 spin_lock(&bdev->lru_lock);
1641 put_count = ttm_bo_del_from_lru(bo);
1642 spin_unlock(&bdev->lru_lock);
1645 kref_put(&bo->list_kref, ttm_bo_ref_bug);
1648 * Wait for GPU, then move to system cached.
1651 spin_lock(&bo->lock);
1652 ret = ttm_bo_wait(bo, false, false, false);
1653 spin_unlock(&bo->lock);
1655 if (unlikely(ret != 0))
1658 if ((bo->mem.placement & swap_placement) != swap_placement) {
1659 struct ttm_mem_reg evict_mem;
1661 evict_mem = bo->mem;
1662 evict_mem.mm_node = NULL;
1663 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1664 evict_mem.mem_type = TTM_PL_SYSTEM;
1666 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1668 if (unlikely(ret != 0))
1672 ttm_bo_unmap_virtual(bo);
1675 * Swap out. Buffer will be swapped in again as soon as
1676 * anyone tries to access a ttm page.
1679 ret = ttm_tt_swapout(bo->ttm, bo->persistant_swap_storage);
1684 * Unreserve without putting on LRU to avoid swapping out an
1685 * already swapped buffer.
1688 atomic_set(&bo->reserved, 0);
1689 wake_up_all(&bo->event_queue);
1690 kref_put(&bo->list_kref, ttm_bo_release_list);
1694 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1696 while (ttm_bo_swapout(&bdev->shrink) == 0)