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 #define pr_fmt(fmt) "[TTM] " fmt
33 #include <drm/ttm/ttm_module.h>
34 #include <drm/ttm/ttm_bo_driver.h>
35 #include <drm/ttm/ttm_placement.h>
36 #include <linux/jiffies.h>
37 #include <linux/slab.h>
38 #include <linux/sched.h>
40 #include <linux/file.h>
41 #include <linux/module.h>
42 #include <linux/atomic.h>
43 #include <linux/reservation.h>
45 #define TTM_ASSERT_LOCKED(param)
46 #define TTM_DEBUG(fmt, arg...)
47 #define TTM_BO_HASH_ORDER 13
49 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink);
50 static void ttm_bo_global_kobj_release(struct kobject *kobj);
52 static struct attribute ttm_bo_count = {
57 static inline int ttm_mem_type_from_place(const struct ttm_place *place,
62 for (i = 0; i <= TTM_PL_PRIV5; i++)
63 if (place->flags & (1 << i)) {
70 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type)
72 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
74 pr_err(" has_type: %d\n", man->has_type);
75 pr_err(" use_type: %d\n", man->use_type);
76 pr_err(" flags: 0x%08X\n", man->flags);
77 pr_err(" gpu_offset: 0x%08lX\n", man->gpu_offset);
78 pr_err(" size: %llu\n", man->size);
79 pr_err(" available_caching: 0x%08X\n", man->available_caching);
80 pr_err(" default_caching: 0x%08X\n", man->default_caching);
81 if (mem_type != TTM_PL_SYSTEM)
82 (*man->func->debug)(man, TTM_PFX);
85 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
86 struct ttm_placement *placement)
90 pr_err("No space for %p (%lu pages, %luK, %luM)\n",
91 bo, bo->mem.num_pages, bo->mem.size >> 10,
93 for (i = 0; i < placement->num_placement; i++) {
94 ret = ttm_mem_type_from_place(&placement->placement[i],
98 pr_err(" placement[%d]=0x%08X (%d)\n",
99 i, placement->placement[i].flags, mem_type);
100 ttm_mem_type_debug(bo->bdev, mem_type);
104 static ssize_t ttm_bo_global_show(struct kobject *kobj,
105 struct attribute *attr,
108 struct ttm_bo_global *glob =
109 container_of(kobj, struct ttm_bo_global, kobj);
111 return snprintf(buffer, PAGE_SIZE, "%lu\n",
112 (unsigned long) atomic_read(&glob->bo_count));
115 static struct attribute *ttm_bo_global_attrs[] = {
120 static const struct sysfs_ops ttm_bo_global_ops = {
121 .show = &ttm_bo_global_show
124 static struct kobj_type ttm_bo_glob_kobj_type = {
125 .release = &ttm_bo_global_kobj_release,
126 .sysfs_ops = &ttm_bo_global_ops,
127 .default_attrs = ttm_bo_global_attrs
131 static inline uint32_t ttm_bo_type_flags(unsigned type)
136 static void ttm_bo_release_list(struct kref *list_kref)
138 struct ttm_buffer_object *bo =
139 container_of(list_kref, struct ttm_buffer_object, list_kref);
140 struct ttm_bo_device *bdev = bo->bdev;
141 size_t acc_size = bo->acc_size;
143 BUG_ON(atomic_read(&bo->list_kref.refcount));
144 BUG_ON(atomic_read(&bo->kref.refcount));
145 BUG_ON(atomic_read(&bo->cpu_writers));
146 BUG_ON(bo->mem.mm_node != NULL);
147 BUG_ON(!list_empty(&bo->lru));
148 BUG_ON(!list_empty(&bo->ddestroy));
151 ttm_tt_destroy(bo->ttm);
152 atomic_dec(&bo->glob->bo_count);
153 if (bo->resv == &bo->ttm_resv)
154 reservation_object_fini(&bo->ttm_resv);
155 mutex_destroy(&bo->wu_mutex);
161 ttm_mem_global_free(bdev->glob->mem_glob, acc_size);
164 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
166 struct ttm_bo_device *bdev = bo->bdev;
167 struct ttm_mem_type_manager *man;
169 lockdep_assert_held(&bo->resv->lock.base);
171 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
173 BUG_ON(!list_empty(&bo->lru));
175 man = &bdev->man[bo->mem.mem_type];
176 list_add_tail(&bo->lru, &man->lru);
177 kref_get(&bo->list_kref);
179 if (bo->ttm != NULL) {
180 list_add_tail(&bo->swap, &bo->glob->swap_lru);
181 kref_get(&bo->list_kref);
185 EXPORT_SYMBOL(ttm_bo_add_to_lru);
187 int ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
191 if (!list_empty(&bo->swap)) {
192 list_del_init(&bo->swap);
195 if (!list_empty(&bo->lru)) {
196 list_del_init(&bo->lru);
201 * TODO: Add a driver hook to delete from
202 * driver-specific LRU's here.
208 static void ttm_bo_ref_bug(struct kref *list_kref)
213 void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count,
216 kref_sub(&bo->list_kref, count,
217 (never_free) ? ttm_bo_ref_bug : ttm_bo_release_list);
220 void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo)
224 spin_lock(&bo->glob->lru_lock);
225 put_count = ttm_bo_del_from_lru(bo);
226 spin_unlock(&bo->glob->lru_lock);
227 ttm_bo_list_ref_sub(bo, put_count, true);
229 EXPORT_SYMBOL(ttm_bo_del_sub_from_lru);
232 * Call bo->mutex locked.
234 static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
236 struct ttm_bo_device *bdev = bo->bdev;
237 struct ttm_bo_global *glob = bo->glob;
239 uint32_t page_flags = 0;
241 TTM_ASSERT_LOCKED(&bo->mutex);
244 if (bdev->need_dma32)
245 page_flags |= TTM_PAGE_FLAG_DMA32;
248 case ttm_bo_type_device:
250 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
251 case ttm_bo_type_kernel:
252 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
253 page_flags, glob->dummy_read_page);
254 if (unlikely(bo->ttm == NULL))
258 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
259 page_flags | TTM_PAGE_FLAG_SG,
260 glob->dummy_read_page);
261 if (unlikely(bo->ttm == NULL)) {
265 bo->ttm->sg = bo->sg;
268 pr_err("Illegal buffer object type\n");
276 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
277 struct ttm_mem_reg *mem,
278 bool evict, bool interruptible,
281 struct ttm_bo_device *bdev = bo->bdev;
282 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
283 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
284 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
285 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
288 if (old_is_pci || new_is_pci ||
289 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
290 ret = ttm_mem_io_lock(old_man, true);
291 if (unlikely(ret != 0))
293 ttm_bo_unmap_virtual_locked(bo);
294 ttm_mem_io_unlock(old_man);
298 * Create and bind a ttm if required.
301 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
302 if (bo->ttm == NULL) {
303 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
304 ret = ttm_bo_add_ttm(bo, zero);
309 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
313 if (mem->mem_type != TTM_PL_SYSTEM) {
314 ret = ttm_tt_bind(bo->ttm, mem);
319 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
320 if (bdev->driver->move_notify)
321 bdev->driver->move_notify(bo, mem);
328 if (bdev->driver->move_notify)
329 bdev->driver->move_notify(bo, mem);
331 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
332 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
333 ret = ttm_bo_move_ttm(bo, evict, no_wait_gpu, mem);
334 else if (bdev->driver->move)
335 ret = bdev->driver->move(bo, evict, interruptible,
338 ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, mem);
341 if (bdev->driver->move_notify) {
342 struct ttm_mem_reg tmp_mem = *mem;
345 bdev->driver->move_notify(bo, mem);
355 if (bdev->driver->invalidate_caches) {
356 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
358 pr_err("Can not flush read caches\n");
363 if (bo->mem.mm_node) {
364 bo->offset = (bo->mem.start << PAGE_SHIFT) +
365 bdev->man[bo->mem.mem_type].gpu_offset;
366 bo->cur_placement = bo->mem.placement;
373 new_man = &bdev->man[bo->mem.mem_type];
374 if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) {
375 ttm_tt_unbind(bo->ttm);
376 ttm_tt_destroy(bo->ttm);
385 * Will release GPU memory type usage on destruction.
386 * This is the place to put in driver specific hooks to release
387 * driver private resources.
388 * Will release the bo::reserved lock.
391 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
393 if (bo->bdev->driver->move_notify)
394 bo->bdev->driver->move_notify(bo, NULL);
397 ttm_tt_unbind(bo->ttm);
398 ttm_tt_destroy(bo->ttm);
401 ttm_bo_mem_put(bo, &bo->mem);
403 ww_mutex_unlock (&bo->resv->lock);
406 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
408 struct reservation_object_list *fobj;
412 fobj = reservation_object_get_list(bo->resv);
413 fence = reservation_object_get_excl(bo->resv);
414 if (fence && !fence->ops->signaled)
415 fence_enable_sw_signaling(fence);
417 for (i = 0; fobj && i < fobj->shared_count; ++i) {
418 fence = rcu_dereference_protected(fobj->shared[i],
419 reservation_object_held(bo->resv));
421 if (!fence->ops->signaled)
422 fence_enable_sw_signaling(fence);
426 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
428 struct ttm_bo_device *bdev = bo->bdev;
429 struct ttm_bo_global *glob = bo->glob;
433 spin_lock(&glob->lru_lock);
434 ret = __ttm_bo_reserve(bo, false, true, false, NULL);
437 if (!ttm_bo_wait(bo, false, false, true)) {
438 put_count = ttm_bo_del_from_lru(bo);
440 spin_unlock(&glob->lru_lock);
441 ttm_bo_cleanup_memtype_use(bo);
443 ttm_bo_list_ref_sub(bo, put_count, true);
447 ttm_bo_flush_all_fences(bo);
450 * Make NO_EVICT bos immediately available to
451 * shrinkers, now that they are queued for
454 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) {
455 bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT;
456 ttm_bo_add_to_lru(bo);
459 __ttm_bo_unreserve(bo);
462 kref_get(&bo->list_kref);
463 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
464 spin_unlock(&glob->lru_lock);
466 schedule_delayed_work(&bdev->wq,
467 ((HZ / 100) < 1) ? 1 : HZ / 100);
471 * function ttm_bo_cleanup_refs_and_unlock
472 * If bo idle, remove from delayed- and lru lists, and unref.
473 * If not idle, do nothing.
475 * Must be called with lru_lock and reservation held, this function
476 * will drop both before returning.
478 * @interruptible Any sleeps should occur interruptibly.
479 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
482 static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
486 struct ttm_bo_global *glob = bo->glob;
490 ret = ttm_bo_wait(bo, false, false, true);
492 if (ret && !no_wait_gpu) {
494 ww_mutex_unlock(&bo->resv->lock);
495 spin_unlock(&glob->lru_lock);
497 lret = reservation_object_wait_timeout_rcu(bo->resv,
507 spin_lock(&glob->lru_lock);
508 ret = __ttm_bo_reserve(bo, false, true, false, NULL);
511 * We raced, and lost, someone else holds the reservation now,
512 * and is probably busy in ttm_bo_cleanup_memtype_use.
514 * Even if it's not the case, because we finished waiting any
515 * delayed destruction would succeed, so just return success
519 spin_unlock(&glob->lru_lock);
524 * remove sync_obj with ttm_bo_wait, the wait should be
525 * finished, and no new wait object should have been added.
527 ret = ttm_bo_wait(bo, false, false, true);
531 if (ret || unlikely(list_empty(&bo->ddestroy))) {
532 __ttm_bo_unreserve(bo);
533 spin_unlock(&glob->lru_lock);
537 put_count = ttm_bo_del_from_lru(bo);
538 list_del_init(&bo->ddestroy);
541 spin_unlock(&glob->lru_lock);
542 ttm_bo_cleanup_memtype_use(bo);
544 ttm_bo_list_ref_sub(bo, put_count, true);
550 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
551 * encountered buffers.
554 static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
556 struct ttm_bo_global *glob = bdev->glob;
557 struct ttm_buffer_object *entry = NULL;
560 spin_lock(&glob->lru_lock);
561 if (list_empty(&bdev->ddestroy))
564 entry = list_first_entry(&bdev->ddestroy,
565 struct ttm_buffer_object, ddestroy);
566 kref_get(&entry->list_kref);
569 struct ttm_buffer_object *nentry = NULL;
571 if (entry->ddestroy.next != &bdev->ddestroy) {
572 nentry = list_first_entry(&entry->ddestroy,
573 struct ttm_buffer_object, ddestroy);
574 kref_get(&nentry->list_kref);
577 ret = __ttm_bo_reserve(entry, false, true, false, NULL);
578 if (remove_all && ret) {
579 spin_unlock(&glob->lru_lock);
580 ret = __ttm_bo_reserve(entry, false, false,
582 spin_lock(&glob->lru_lock);
586 ret = ttm_bo_cleanup_refs_and_unlock(entry, false,
589 spin_unlock(&glob->lru_lock);
591 kref_put(&entry->list_kref, ttm_bo_release_list);
597 spin_lock(&glob->lru_lock);
598 if (list_empty(&entry->ddestroy))
603 spin_unlock(&glob->lru_lock);
606 kref_put(&entry->list_kref, ttm_bo_release_list);
610 static void ttm_bo_delayed_workqueue(struct work_struct *work)
612 struct ttm_bo_device *bdev =
613 container_of(work, struct ttm_bo_device, wq.work);
615 if (ttm_bo_delayed_delete(bdev, false)) {
616 schedule_delayed_work(&bdev->wq,
617 ((HZ / 100) < 1) ? 1 : HZ / 100);
621 static void ttm_bo_release(struct kref *kref)
623 struct ttm_buffer_object *bo =
624 container_of(kref, struct ttm_buffer_object, kref);
625 struct ttm_bo_device *bdev = bo->bdev;
626 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
628 drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node);
629 ttm_mem_io_lock(man, false);
630 ttm_mem_io_free_vm(bo);
631 ttm_mem_io_unlock(man);
632 ttm_bo_cleanup_refs_or_queue(bo);
633 kref_put(&bo->list_kref, ttm_bo_release_list);
636 void ttm_bo_unref(struct ttm_buffer_object **p_bo)
638 struct ttm_buffer_object *bo = *p_bo;
641 kref_put(&bo->kref, ttm_bo_release);
643 EXPORT_SYMBOL(ttm_bo_unref);
645 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
647 return cancel_delayed_work_sync(&bdev->wq);
649 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
651 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
654 schedule_delayed_work(&bdev->wq,
655 ((HZ / 100) < 1) ? 1 : HZ / 100);
657 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
659 static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
662 struct ttm_bo_device *bdev = bo->bdev;
663 struct ttm_mem_reg evict_mem;
664 struct ttm_placement placement;
667 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
669 if (unlikely(ret != 0)) {
670 if (ret != -ERESTARTSYS) {
671 pr_err("Failed to expire sync object before buffer eviction\n");
676 lockdep_assert_held(&bo->resv->lock.base);
679 evict_mem.mm_node = NULL;
680 evict_mem.bus.io_reserved_vm = false;
681 evict_mem.bus.io_reserved_count = 0;
683 placement.num_placement = 0;
684 placement.num_busy_placement = 0;
685 bdev->driver->evict_flags(bo, &placement);
686 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
689 if (ret != -ERESTARTSYS) {
690 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
692 ttm_bo_mem_space_debug(bo, &placement);
697 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
700 if (ret != -ERESTARTSYS)
701 pr_err("Buffer eviction failed\n");
702 ttm_bo_mem_put(bo, &evict_mem);
710 static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
715 struct ttm_bo_global *glob = bdev->glob;
716 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
717 struct ttm_buffer_object *bo;
718 int ret = -EBUSY, put_count;
720 spin_lock(&glob->lru_lock);
721 list_for_each_entry(bo, &man->lru, lru) {
722 ret = __ttm_bo_reserve(bo, false, true, false, NULL);
728 spin_unlock(&glob->lru_lock);
732 kref_get(&bo->list_kref);
734 if (!list_empty(&bo->ddestroy)) {
735 ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible,
737 kref_put(&bo->list_kref, ttm_bo_release_list);
741 put_count = ttm_bo_del_from_lru(bo);
742 spin_unlock(&glob->lru_lock);
746 ttm_bo_list_ref_sub(bo, put_count, true);
748 ret = ttm_bo_evict(bo, interruptible, no_wait_gpu);
749 ttm_bo_unreserve(bo);
751 kref_put(&bo->list_kref, ttm_bo_release_list);
755 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
757 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
760 (*man->func->put_node)(man, mem);
762 EXPORT_SYMBOL(ttm_bo_mem_put);
765 * Repeatedly evict memory from the LRU for @mem_type until we create enough
766 * space, or we've evicted everything and there isn't enough space.
768 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
770 const struct ttm_place *place,
771 struct ttm_mem_reg *mem,
775 struct ttm_bo_device *bdev = bo->bdev;
776 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
780 ret = (*man->func->get_node)(man, bo, place, mem);
781 if (unlikely(ret != 0))
785 ret = ttm_mem_evict_first(bdev, mem_type,
786 interruptible, no_wait_gpu);
787 if (unlikely(ret != 0))
790 if (mem->mm_node == NULL)
792 mem->mem_type = mem_type;
796 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
797 uint32_t cur_placement,
798 uint32_t proposed_placement)
800 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
801 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
804 * Keep current caching if possible.
807 if ((cur_placement & caching) != 0)
808 result |= (cur_placement & caching);
809 else if ((man->default_caching & caching) != 0)
810 result |= man->default_caching;
811 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
812 result |= TTM_PL_FLAG_CACHED;
813 else if ((TTM_PL_FLAG_WC & caching) != 0)
814 result |= TTM_PL_FLAG_WC;
815 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
816 result |= TTM_PL_FLAG_UNCACHED;
821 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
823 const struct ttm_place *place,
824 uint32_t *masked_placement)
826 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
828 if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0)
831 if ((place->flags & man->available_caching) == 0)
834 cur_flags |= (place->flags & man->available_caching);
836 *masked_placement = cur_flags;
841 * Creates space for memory region @mem according to its type.
843 * This function first searches for free space in compatible memory types in
844 * the priority order defined by the driver. If free space isn't found, then
845 * ttm_bo_mem_force_space is attempted in priority order to evict and find
848 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
849 struct ttm_placement *placement,
850 struct ttm_mem_reg *mem,
854 struct ttm_bo_device *bdev = bo->bdev;
855 struct ttm_mem_type_manager *man;
856 uint32_t mem_type = TTM_PL_SYSTEM;
857 uint32_t cur_flags = 0;
858 bool type_found = false;
859 bool type_ok = false;
860 bool has_erestartsys = false;
864 for (i = 0; i < placement->num_placement; ++i) {
865 const struct ttm_place *place = &placement->placement[i];
867 ret = ttm_mem_type_from_place(place, &mem_type);
870 man = &bdev->man[mem_type];
872 type_ok = ttm_bo_mt_compatible(man, mem_type, place,
878 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
881 * Use the access and other non-mapping-related flag bits from
882 * the memory placement flags to the current flags
884 ttm_flag_masked(&cur_flags, place->flags,
885 ~TTM_PL_MASK_MEMTYPE);
887 if (mem_type == TTM_PL_SYSTEM)
890 if (man->has_type && man->use_type) {
892 ret = (*man->func->get_node)(man, bo, place, mem);
900 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
901 mem->mem_type = mem_type;
902 mem->placement = cur_flags;
909 for (i = 0; i < placement->num_busy_placement; ++i) {
910 const struct ttm_place *place = &placement->busy_placement[i];
912 ret = ttm_mem_type_from_place(place, &mem_type);
915 man = &bdev->man[mem_type];
918 if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags))
921 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
924 * Use the access and other non-mapping-related flag bits from
925 * the memory placement flags to the current flags
927 ttm_flag_masked(&cur_flags, place->flags,
928 ~TTM_PL_MASK_MEMTYPE);
930 if (mem_type == TTM_PL_SYSTEM) {
931 mem->mem_type = mem_type;
932 mem->placement = cur_flags;
937 ret = ttm_bo_mem_force_space(bo, mem_type, place, mem,
938 interruptible, no_wait_gpu);
939 if (ret == 0 && mem->mm_node) {
940 mem->placement = cur_flags;
943 if (ret == -ERESTARTSYS)
944 has_erestartsys = true;
946 ret = (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
949 EXPORT_SYMBOL(ttm_bo_mem_space);
951 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
952 struct ttm_placement *placement,
957 struct ttm_mem_reg mem;
959 lockdep_assert_held(&bo->resv->lock.base);
962 * FIXME: It's possible to pipeline buffer moves.
963 * Have the driver move function wait for idle when necessary,
964 * instead of doing it here.
966 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
969 mem.num_pages = bo->num_pages;
970 mem.size = mem.num_pages << PAGE_SHIFT;
971 mem.page_alignment = bo->mem.page_alignment;
972 mem.bus.io_reserved_vm = false;
973 mem.bus.io_reserved_count = 0;
975 * Determine where to move the buffer.
977 ret = ttm_bo_mem_space(bo, placement, &mem,
978 interruptible, no_wait_gpu);
981 ret = ttm_bo_handle_move_mem(bo, &mem, false,
982 interruptible, no_wait_gpu);
984 if (ret && mem.mm_node)
985 ttm_bo_mem_put(bo, &mem);
989 static bool ttm_bo_mem_compat(struct ttm_placement *placement,
990 struct ttm_mem_reg *mem,
995 for (i = 0; i < placement->num_placement; i++) {
996 const struct ttm_place *heap = &placement->placement[i];
997 if (mem->mm_node && heap->lpfn != 0 &&
998 (mem->start < heap->fpfn ||
999 mem->start + mem->num_pages > heap->lpfn))
1002 *new_flags = heap->flags;
1003 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1004 (*new_flags & mem->placement & TTM_PL_MASK_MEM))
1008 for (i = 0; i < placement->num_busy_placement; i++) {
1009 const struct ttm_place *heap = &placement->busy_placement[i];
1010 if (mem->mm_node && heap->lpfn != 0 &&
1011 (mem->start < heap->fpfn ||
1012 mem->start + mem->num_pages > heap->lpfn))
1015 *new_flags = heap->flags;
1016 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1017 (*new_flags & mem->placement & TTM_PL_MASK_MEM))
1024 int ttm_bo_validate(struct ttm_buffer_object *bo,
1025 struct ttm_placement *placement,
1032 lockdep_assert_held(&bo->resv->lock.base);
1034 * Check whether we need to move buffer.
1036 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1037 ret = ttm_bo_move_buffer(bo, placement, interruptible,
1043 * Use the access and other non-mapping-related flag bits from
1044 * the compatible memory placement flags to the active flags
1046 ttm_flag_masked(&bo->mem.placement, new_flags,
1047 ~TTM_PL_MASK_MEMTYPE);
1050 * We might need to add a TTM.
1052 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1053 ret = ttm_bo_add_ttm(bo, true);
1059 EXPORT_SYMBOL(ttm_bo_validate);
1061 int ttm_bo_init(struct ttm_bo_device *bdev,
1062 struct ttm_buffer_object *bo,
1064 enum ttm_bo_type type,
1065 struct ttm_placement *placement,
1066 uint32_t page_alignment,
1068 struct file *persistent_swap_storage,
1070 struct sg_table *sg,
1071 struct reservation_object *resv,
1072 void (*destroy) (struct ttm_buffer_object *))
1075 unsigned long num_pages;
1076 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1079 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1081 pr_err("Out of kernel memory\n");
1089 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1090 if (num_pages == 0) {
1091 pr_err("Illegal buffer object size\n");
1096 ttm_mem_global_free(mem_glob, acc_size);
1099 bo->destroy = destroy;
1101 kref_init(&bo->kref);
1102 kref_init(&bo->list_kref);
1103 atomic_set(&bo->cpu_writers, 0);
1104 INIT_LIST_HEAD(&bo->lru);
1105 INIT_LIST_HEAD(&bo->ddestroy);
1106 INIT_LIST_HEAD(&bo->swap);
1107 INIT_LIST_HEAD(&bo->io_reserve_lru);
1108 mutex_init(&bo->wu_mutex);
1110 bo->glob = bdev->glob;
1112 bo->num_pages = num_pages;
1113 bo->mem.size = num_pages << PAGE_SHIFT;
1114 bo->mem.mem_type = TTM_PL_SYSTEM;
1115 bo->mem.num_pages = bo->num_pages;
1116 bo->mem.mm_node = NULL;
1117 bo->mem.page_alignment = page_alignment;
1118 bo->mem.bus.io_reserved_vm = false;
1119 bo->mem.bus.io_reserved_count = 0;
1121 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1122 bo->persistent_swap_storage = persistent_swap_storage;
1123 bo->acc_size = acc_size;
1127 lockdep_assert_held(&bo->resv->lock.base);
1129 bo->resv = &bo->ttm_resv;
1130 reservation_object_init(&bo->ttm_resv);
1132 atomic_inc(&bo->glob->bo_count);
1133 drm_vma_node_reset(&bo->vma_node);
1136 * For ttm_bo_type_device buffers, allocate
1137 * address space from the device.
1139 if (bo->type == ttm_bo_type_device ||
1140 bo->type == ttm_bo_type_sg)
1141 ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node,
1144 /* passed reservation objects should already be locked,
1145 * since otherwise lockdep will be angered in radeon.
1148 locked = ww_mutex_trylock(&bo->resv->lock);
1153 ret = ttm_bo_validate(bo, placement, interruptible, false);
1156 ttm_bo_unreserve(bo);
1163 EXPORT_SYMBOL(ttm_bo_init);
1165 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1166 unsigned long bo_size,
1167 unsigned struct_size)
1169 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1172 size += ttm_round_pot(struct_size);
1173 size += PAGE_ALIGN(npages * sizeof(void *));
1174 size += ttm_round_pot(sizeof(struct ttm_tt));
1177 EXPORT_SYMBOL(ttm_bo_acc_size);
1179 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1180 unsigned long bo_size,
1181 unsigned struct_size)
1183 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1186 size += ttm_round_pot(struct_size);
1187 size += PAGE_ALIGN(npages * sizeof(void *));
1188 size += PAGE_ALIGN(npages * sizeof(dma_addr_t));
1189 size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1192 EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1194 int ttm_bo_create(struct ttm_bo_device *bdev,
1196 enum ttm_bo_type type,
1197 struct ttm_placement *placement,
1198 uint32_t page_alignment,
1200 struct file *persistent_swap_storage,
1201 struct ttm_buffer_object **p_bo)
1203 struct ttm_buffer_object *bo;
1207 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1208 if (unlikely(bo == NULL))
1211 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1212 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1213 interruptible, persistent_swap_storage, acc_size,
1215 if (likely(ret == 0))
1220 EXPORT_SYMBOL(ttm_bo_create);
1222 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1223 unsigned mem_type, bool allow_errors)
1225 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1226 struct ttm_bo_global *glob = bdev->glob;
1230 * Can't use standard list traversal since we're unlocking.
1233 spin_lock(&glob->lru_lock);
1234 while (!list_empty(&man->lru)) {
1235 spin_unlock(&glob->lru_lock);
1236 ret = ttm_mem_evict_first(bdev, mem_type, false, false);
1241 pr_err("Cleanup eviction failed\n");
1244 spin_lock(&glob->lru_lock);
1246 spin_unlock(&glob->lru_lock);
1250 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1252 struct ttm_mem_type_manager *man;
1255 if (mem_type >= TTM_NUM_MEM_TYPES) {
1256 pr_err("Illegal memory type %d\n", mem_type);
1259 man = &bdev->man[mem_type];
1261 if (!man->has_type) {
1262 pr_err("Trying to take down uninitialized memory manager type %u\n",
1267 man->use_type = false;
1268 man->has_type = false;
1272 ttm_bo_force_list_clean(bdev, mem_type, false);
1274 ret = (*man->func->takedown)(man);
1279 EXPORT_SYMBOL(ttm_bo_clean_mm);
1281 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1283 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1285 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1286 pr_err("Illegal memory manager memory type %u\n", mem_type);
1290 if (!man->has_type) {
1291 pr_err("Memory type %u has not been initialized\n", mem_type);
1295 return ttm_bo_force_list_clean(bdev, mem_type, true);
1297 EXPORT_SYMBOL(ttm_bo_evict_mm);
1299 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1300 unsigned long p_size)
1303 struct ttm_mem_type_manager *man;
1305 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1306 man = &bdev->man[type];
1307 BUG_ON(man->has_type);
1308 man->io_reserve_fastpath = true;
1309 man->use_io_reserve_lru = false;
1310 mutex_init(&man->io_reserve_mutex);
1311 INIT_LIST_HEAD(&man->io_reserve_lru);
1313 ret = bdev->driver->init_mem_type(bdev, type, man);
1319 if (type != TTM_PL_SYSTEM) {
1320 ret = (*man->func->init)(man, p_size);
1324 man->has_type = true;
1325 man->use_type = true;
1328 INIT_LIST_HEAD(&man->lru);
1332 EXPORT_SYMBOL(ttm_bo_init_mm);
1334 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1336 struct ttm_bo_global *glob =
1337 container_of(kobj, struct ttm_bo_global, kobj);
1339 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
1340 __free_page(glob->dummy_read_page);
1344 void ttm_bo_global_release(struct drm_global_reference *ref)
1346 struct ttm_bo_global *glob = ref->object;
1348 kobject_del(&glob->kobj);
1349 kobject_put(&glob->kobj);
1351 EXPORT_SYMBOL(ttm_bo_global_release);
1353 int ttm_bo_global_init(struct drm_global_reference *ref)
1355 struct ttm_bo_global_ref *bo_ref =
1356 container_of(ref, struct ttm_bo_global_ref, ref);
1357 struct ttm_bo_global *glob = ref->object;
1360 mutex_init(&glob->device_list_mutex);
1361 spin_lock_init(&glob->lru_lock);
1362 glob->mem_glob = bo_ref->mem_glob;
1363 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1365 if (unlikely(glob->dummy_read_page == NULL)) {
1370 INIT_LIST_HEAD(&glob->swap_lru);
1371 INIT_LIST_HEAD(&glob->device_list);
1373 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
1374 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
1375 if (unlikely(ret != 0)) {
1376 pr_err("Could not register buffer object swapout\n");
1380 atomic_set(&glob->bo_count, 0);
1382 ret = kobject_init_and_add(
1383 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1384 if (unlikely(ret != 0))
1385 kobject_put(&glob->kobj);
1388 __free_page(glob->dummy_read_page);
1393 EXPORT_SYMBOL(ttm_bo_global_init);
1396 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1399 unsigned i = TTM_NUM_MEM_TYPES;
1400 struct ttm_mem_type_manager *man;
1401 struct ttm_bo_global *glob = bdev->glob;
1404 man = &bdev->man[i];
1405 if (man->has_type) {
1406 man->use_type = false;
1407 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1409 pr_err("DRM memory manager type %d is not clean\n",
1412 man->has_type = false;
1416 mutex_lock(&glob->device_list_mutex);
1417 list_del(&bdev->device_list);
1418 mutex_unlock(&glob->device_list_mutex);
1420 cancel_delayed_work_sync(&bdev->wq);
1422 while (ttm_bo_delayed_delete(bdev, true))
1425 spin_lock(&glob->lru_lock);
1426 if (list_empty(&bdev->ddestroy))
1427 TTM_DEBUG("Delayed destroy list was clean\n");
1429 if (list_empty(&bdev->man[0].lru))
1430 TTM_DEBUG("Swap list was clean\n");
1431 spin_unlock(&glob->lru_lock);
1433 drm_vma_offset_manager_destroy(&bdev->vma_manager);
1437 EXPORT_SYMBOL(ttm_bo_device_release);
1439 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1440 struct ttm_bo_global *glob,
1441 struct ttm_bo_driver *driver,
1442 struct address_space *mapping,
1443 uint64_t file_page_offset,
1448 bdev->driver = driver;
1450 memset(bdev->man, 0, sizeof(bdev->man));
1453 * Initialize the system memory buffer type.
1454 * Other types need to be driver / IOCTL initialized.
1456 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1457 if (unlikely(ret != 0))
1460 drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset,
1462 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1463 INIT_LIST_HEAD(&bdev->ddestroy);
1464 bdev->dev_mapping = mapping;
1466 bdev->need_dma32 = need_dma32;
1468 mutex_lock(&glob->device_list_mutex);
1469 list_add_tail(&bdev->device_list, &glob->device_list);
1470 mutex_unlock(&glob->device_list_mutex);
1476 EXPORT_SYMBOL(ttm_bo_device_init);
1479 * buffer object vm functions.
1482 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1484 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1486 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1487 if (mem->mem_type == TTM_PL_SYSTEM)
1490 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1493 if (mem->placement & TTM_PL_FLAG_CACHED)
1499 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1501 struct ttm_bo_device *bdev = bo->bdev;
1503 drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping);
1504 ttm_mem_io_free_vm(bo);
1507 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1509 struct ttm_bo_device *bdev = bo->bdev;
1510 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1512 ttm_mem_io_lock(man, false);
1513 ttm_bo_unmap_virtual_locked(bo);
1514 ttm_mem_io_unlock(man);
1518 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1520 int ttm_bo_wait(struct ttm_buffer_object *bo,
1521 bool lazy, bool interruptible, bool no_wait)
1523 struct reservation_object_list *fobj;
1524 struct reservation_object *resv;
1526 long timeout = 15 * HZ;
1530 fobj = reservation_object_get_list(resv);
1531 excl = reservation_object_get_excl(resv);
1533 if (!fence_is_signaled(excl)) {
1537 timeout = fence_wait_timeout(excl,
1538 interruptible, timeout);
1542 for (i = 0; fobj && timeout > 0 && i < fobj->shared_count; ++i) {
1543 struct fence *fence;
1544 fence = rcu_dereference_protected(fobj->shared[i],
1545 reservation_object_held(resv));
1547 if (!fence_is_signaled(fence)) {
1551 timeout = fence_wait_timeout(fence,
1552 interruptible, timeout);
1562 reservation_object_add_excl_fence(resv, NULL);
1563 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
1566 EXPORT_SYMBOL(ttm_bo_wait);
1568 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1573 * Using ttm_bo_reserve makes sure the lru lists are updated.
1576 ret = ttm_bo_reserve(bo, true, no_wait, false, NULL);
1577 if (unlikely(ret != 0))
1579 ret = ttm_bo_wait(bo, false, true, no_wait);
1580 if (likely(ret == 0))
1581 atomic_inc(&bo->cpu_writers);
1582 ttm_bo_unreserve(bo);
1585 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1587 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1589 atomic_dec(&bo->cpu_writers);
1591 EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1594 * A buffer object shrink method that tries to swap out the first
1595 * buffer object on the bo_global::swap_lru list.
1598 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1600 struct ttm_bo_global *glob =
1601 container_of(shrink, struct ttm_bo_global, shrink);
1602 struct ttm_buffer_object *bo;
1605 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM);
1607 spin_lock(&glob->lru_lock);
1608 list_for_each_entry(bo, &glob->swap_lru, swap) {
1609 ret = __ttm_bo_reserve(bo, false, true, false, NULL);
1615 spin_unlock(&glob->lru_lock);
1619 kref_get(&bo->list_kref);
1621 if (!list_empty(&bo->ddestroy)) {
1622 ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false);
1623 kref_put(&bo->list_kref, ttm_bo_release_list);
1627 put_count = ttm_bo_del_from_lru(bo);
1628 spin_unlock(&glob->lru_lock);
1630 ttm_bo_list_ref_sub(bo, put_count, true);
1633 * Wait for GPU, then move to system cached.
1636 ret = ttm_bo_wait(bo, false, false, false);
1638 if (unlikely(ret != 0))
1641 if ((bo->mem.placement & swap_placement) != swap_placement) {
1642 struct ttm_mem_reg evict_mem;
1644 evict_mem = bo->mem;
1645 evict_mem.mm_node = NULL;
1646 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1647 evict_mem.mem_type = TTM_PL_SYSTEM;
1649 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1651 if (unlikely(ret != 0))
1655 ttm_bo_unmap_virtual(bo);
1658 * Swap out. Buffer will be swapped in again as soon as
1659 * anyone tries to access a ttm page.
1662 if (bo->bdev->driver->swap_notify)
1663 bo->bdev->driver->swap_notify(bo);
1665 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1670 * Unreserve without putting on LRU to avoid swapping out an
1671 * already swapped buffer.
1674 __ttm_bo_unreserve(bo);
1675 kref_put(&bo->list_kref, ttm_bo_release_list);
1679 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1681 while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1684 EXPORT_SYMBOL(ttm_bo_swapout_all);
1687 * ttm_bo_wait_unreserved - interruptible wait for a buffer object to become
1690 * @bo: Pointer to buffer
1692 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo)
1697 * In the absense of a wait_unlocked API,
1698 * Use the bo::wu_mutex to avoid triggering livelocks due to
1699 * concurrent use of this function. Note that this use of
1700 * bo::wu_mutex can go away if we change locking order to
1701 * mmap_sem -> bo::reserve.
1703 ret = mutex_lock_interruptible(&bo->wu_mutex);
1704 if (unlikely(ret != 0))
1705 return -ERESTARTSYS;
1706 if (!ww_mutex_is_locked(&bo->resv->lock))
1708 ret = __ttm_bo_reserve(bo, true, false, false, NULL);
1709 if (unlikely(ret != 0))
1711 __ttm_bo_unreserve(bo);
1714 mutex_unlock(&bo->wu_mutex);