1 /* SPDX-License-Identifier: GPL-2.0 OR MIT */
2 /**************************************************************************
4 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
22 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
23 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
24 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
25 * USE OR OTHER DEALINGS IN THE SOFTWARE.
27 **************************************************************************/
29 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
32 #define pr_fmt(fmt) "[TTM] " fmt
34 #include <drm/ttm/ttm_module.h>
35 #include <drm/ttm/ttm_bo_driver.h>
36 #include <drm/ttm/ttm_placement.h>
37 #include <linux/jiffies.h>
38 #include <linux/slab.h>
39 #include <linux/sched.h>
41 #include <linux/file.h>
42 #include <linux/module.h>
43 #include <linux/atomic.h>
44 #include <linux/dma-resv.h>
46 static void ttm_bo_global_kobj_release(struct kobject *kobj);
49 * ttm_global_mutex - protecting the global BO state
51 DEFINE_MUTEX(ttm_global_mutex);
52 unsigned ttm_bo_glob_use_count;
53 struct ttm_bo_global ttm_bo_glob;
55 static struct attribute ttm_bo_count = {
60 /* default destructor */
61 static void ttm_bo_default_destroy(struct ttm_buffer_object *bo)
66 static inline int ttm_mem_type_from_place(const struct ttm_place *place,
71 pos = ffs(place->flags & TTM_PL_MASK_MEM);
79 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, struct drm_printer *p,
82 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
84 drm_printf(p, " has_type: %d\n", man->has_type);
85 drm_printf(p, " use_type: %d\n", man->use_type);
86 drm_printf(p, " flags: 0x%08X\n", man->flags);
87 drm_printf(p, " gpu_offset: 0x%08llX\n", man->gpu_offset);
88 drm_printf(p, " size: %llu\n", man->size);
89 drm_printf(p, " available_caching: 0x%08X\n", man->available_caching);
90 drm_printf(p, " default_caching: 0x%08X\n", man->default_caching);
91 if (mem_type != TTM_PL_SYSTEM)
92 (*man->func->debug)(man, p);
95 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
96 struct ttm_placement *placement)
98 struct drm_printer p = drm_debug_printer(TTM_PFX);
101 drm_printf(&p, "No space for %p (%lu pages, %luK, %luM)\n",
102 bo, bo->mem.num_pages, bo->mem.size >> 10,
104 for (i = 0; i < placement->num_placement; i++) {
105 ret = ttm_mem_type_from_place(&placement->placement[i],
109 drm_printf(&p, " placement[%d]=0x%08X (%d)\n",
110 i, placement->placement[i].flags, mem_type);
111 ttm_mem_type_debug(bo->bdev, &p, mem_type);
115 static ssize_t ttm_bo_global_show(struct kobject *kobj,
116 struct attribute *attr,
119 struct ttm_bo_global *glob =
120 container_of(kobj, struct ttm_bo_global, kobj);
122 return snprintf(buffer, PAGE_SIZE, "%d\n",
123 atomic_read(&glob->bo_count));
126 static struct attribute *ttm_bo_global_attrs[] = {
131 static const struct sysfs_ops ttm_bo_global_ops = {
132 .show = &ttm_bo_global_show
135 static struct kobj_type ttm_bo_glob_kobj_type = {
136 .release = &ttm_bo_global_kobj_release,
137 .sysfs_ops = &ttm_bo_global_ops,
138 .default_attrs = ttm_bo_global_attrs
142 static inline uint32_t ttm_bo_type_flags(unsigned type)
147 static void ttm_bo_release_list(struct kref *list_kref)
149 struct ttm_buffer_object *bo =
150 container_of(list_kref, struct ttm_buffer_object, list_kref);
151 struct ttm_bo_device *bdev = bo->bdev;
152 size_t acc_size = bo->acc_size;
154 BUG_ON(kref_read(&bo->list_kref));
155 BUG_ON(kref_read(&bo->kref));
156 BUG_ON(atomic_read(&bo->cpu_writers));
157 BUG_ON(bo->mem.mm_node != NULL);
158 BUG_ON(!list_empty(&bo->lru));
159 BUG_ON(!list_empty(&bo->ddestroy));
160 ttm_tt_destroy(bo->ttm);
161 atomic_dec(&bo->bdev->glob->bo_count);
162 dma_fence_put(bo->moving);
163 if (!ttm_bo_uses_embedded_gem_object(bo))
164 dma_resv_fini(&bo->base._resv);
165 mutex_destroy(&bo->wu_mutex);
167 ttm_mem_global_free(bdev->glob->mem_glob, acc_size);
170 static void ttm_bo_add_mem_to_lru(struct ttm_buffer_object *bo,
171 struct ttm_mem_reg *mem)
173 struct ttm_bo_device *bdev = bo->bdev;
174 struct ttm_mem_type_manager *man;
176 dma_resv_assert_held(bo->base.resv);
178 if (!list_empty(&bo->lru))
181 if (mem->placement & TTM_PL_FLAG_NO_EVICT)
184 man = &bdev->man[mem->mem_type];
185 list_add_tail(&bo->lru, &man->lru[bo->priority]);
186 kref_get(&bo->list_kref);
188 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm &&
189 !(bo->ttm->page_flags & (TTM_PAGE_FLAG_SG |
190 TTM_PAGE_FLAG_SWAPPED))) {
191 list_add_tail(&bo->swap, &bdev->glob->swap_lru[bo->priority]);
192 kref_get(&bo->list_kref);
196 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
198 ttm_bo_add_mem_to_lru(bo, &bo->mem);
200 EXPORT_SYMBOL(ttm_bo_add_to_lru);
202 static void ttm_bo_ref_bug(struct kref *list_kref)
207 void ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
209 struct ttm_bo_device *bdev = bo->bdev;
212 if (!list_empty(&bo->swap)) {
213 list_del_init(&bo->swap);
214 kref_put(&bo->list_kref, ttm_bo_ref_bug);
217 if (!list_empty(&bo->lru)) {
218 list_del_init(&bo->lru);
219 kref_put(&bo->list_kref, ttm_bo_ref_bug);
223 if (notify && bdev->driver->del_from_lru_notify)
224 bdev->driver->del_from_lru_notify(bo);
227 void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo)
229 struct ttm_bo_global *glob = bo->bdev->glob;
231 spin_lock(&glob->lru_lock);
232 ttm_bo_del_from_lru(bo);
233 spin_unlock(&glob->lru_lock);
235 EXPORT_SYMBOL(ttm_bo_del_sub_from_lru);
237 static void ttm_bo_bulk_move_set_pos(struct ttm_lru_bulk_move_pos *pos,
238 struct ttm_buffer_object *bo)
245 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo,
246 struct ttm_lru_bulk_move *bulk)
248 dma_resv_assert_held(bo->base.resv);
250 ttm_bo_del_from_lru(bo);
251 ttm_bo_add_to_lru(bo);
253 if (bulk && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
254 switch (bo->mem.mem_type) {
256 ttm_bo_bulk_move_set_pos(&bulk->tt[bo->priority], bo);
260 ttm_bo_bulk_move_set_pos(&bulk->vram[bo->priority], bo);
263 if (bo->ttm && !(bo->ttm->page_flags &
264 (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED)))
265 ttm_bo_bulk_move_set_pos(&bulk->swap[bo->priority], bo);
268 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
270 void ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move *bulk)
274 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
275 struct ttm_lru_bulk_move_pos *pos = &bulk->tt[i];
276 struct ttm_mem_type_manager *man;
281 dma_resv_assert_held(pos->first->base.resv);
282 dma_resv_assert_held(pos->last->base.resv);
284 man = &pos->first->bdev->man[TTM_PL_TT];
285 list_bulk_move_tail(&man->lru[i], &pos->first->lru,
289 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
290 struct ttm_lru_bulk_move_pos *pos = &bulk->vram[i];
291 struct ttm_mem_type_manager *man;
296 dma_resv_assert_held(pos->first->base.resv);
297 dma_resv_assert_held(pos->last->base.resv);
299 man = &pos->first->bdev->man[TTM_PL_VRAM];
300 list_bulk_move_tail(&man->lru[i], &pos->first->lru,
304 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
305 struct ttm_lru_bulk_move_pos *pos = &bulk->swap[i];
306 struct list_head *lru;
311 dma_resv_assert_held(pos->first->base.resv);
312 dma_resv_assert_held(pos->last->base.resv);
314 lru = &pos->first->bdev->glob->swap_lru[i];
315 list_bulk_move_tail(lru, &pos->first->swap, &pos->last->swap);
318 EXPORT_SYMBOL(ttm_bo_bulk_move_lru_tail);
320 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
321 struct ttm_mem_reg *mem, bool evict,
322 struct ttm_operation_ctx *ctx)
324 struct ttm_bo_device *bdev = bo->bdev;
325 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
326 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
327 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
328 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
331 if (old_is_pci || new_is_pci ||
332 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
333 ret = ttm_mem_io_lock(old_man, true);
334 if (unlikely(ret != 0))
336 ttm_bo_unmap_virtual_locked(bo);
337 ttm_mem_io_unlock(old_man);
341 * Create and bind a ttm if required.
344 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
345 if (bo->ttm == NULL) {
346 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
347 ret = ttm_tt_create(bo, zero);
352 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
356 if (mem->mem_type != TTM_PL_SYSTEM) {
357 ret = ttm_tt_bind(bo->ttm, mem, ctx);
362 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
363 if (bdev->driver->move_notify)
364 bdev->driver->move_notify(bo, evict, mem);
371 if (bdev->driver->move_notify)
372 bdev->driver->move_notify(bo, evict, mem);
374 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
375 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
376 ret = ttm_bo_move_ttm(bo, ctx, mem);
377 else if (bdev->driver->move)
378 ret = bdev->driver->move(bo, evict, ctx, mem);
380 ret = ttm_bo_move_memcpy(bo, ctx, mem);
383 if (bdev->driver->move_notify) {
385 bdev->driver->move_notify(bo, false, mem);
394 if (bdev->driver->invalidate_caches) {
395 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
397 pr_err("Can not flush read caches\n");
403 bo->offset = (bo->mem.start << PAGE_SHIFT) +
404 bdev->man[bo->mem.mem_type].gpu_offset;
408 ctx->bytes_moved += bo->num_pages << PAGE_SHIFT;
412 new_man = &bdev->man[bo->mem.mem_type];
413 if (new_man->flags & TTM_MEMTYPE_FLAG_FIXED) {
414 ttm_tt_destroy(bo->ttm);
423 * Will release GPU memory type usage on destruction.
424 * This is the place to put in driver specific hooks to release
425 * driver private resources.
426 * Will release the bo::reserved lock.
429 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
431 if (bo->bdev->driver->move_notify)
432 bo->bdev->driver->move_notify(bo, false, NULL);
434 ttm_tt_destroy(bo->ttm);
436 ttm_bo_mem_put(bo, &bo->mem);
439 static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
443 if (bo->base.resv == &bo->base._resv)
446 BUG_ON(!dma_resv_trylock(&bo->base._resv));
448 r = dma_resv_copy_fences(&bo->base._resv, bo->base.resv);
450 dma_resv_unlock(&bo->base._resv);
455 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
457 struct dma_resv_list *fobj;
458 struct dma_fence *fence;
461 fobj = dma_resv_get_list(&bo->base._resv);
462 fence = dma_resv_get_excl(&bo->base._resv);
463 if (fence && !fence->ops->signaled)
464 dma_fence_enable_sw_signaling(fence);
466 for (i = 0; fobj && i < fobj->shared_count; ++i) {
467 fence = rcu_dereference_protected(fobj->shared[i],
468 dma_resv_held(bo->base.resv));
470 if (!fence->ops->signaled)
471 dma_fence_enable_sw_signaling(fence);
475 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
477 struct ttm_bo_device *bdev = bo->bdev;
478 struct ttm_bo_global *glob = bdev->glob;
481 ret = ttm_bo_individualize_resv(bo);
483 /* Last resort, if we fail to allocate memory for the
484 * fences block for the BO to become idle
486 dma_resv_wait_timeout_rcu(bo->base.resv, true, false,
488 spin_lock(&glob->lru_lock);
492 spin_lock(&glob->lru_lock);
493 ret = dma_resv_trylock(bo->base.resv) ? 0 : -EBUSY;
495 if (dma_resv_test_signaled_rcu(&bo->base._resv, true)) {
496 ttm_bo_del_from_lru(bo);
497 spin_unlock(&glob->lru_lock);
498 if (bo->base.resv != &bo->base._resv)
499 dma_resv_unlock(&bo->base._resv);
501 ttm_bo_cleanup_memtype_use(bo);
502 dma_resv_unlock(bo->base.resv);
506 ttm_bo_flush_all_fences(bo);
509 * Make NO_EVICT bos immediately available to
510 * shrinkers, now that they are queued for
513 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) {
514 bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT;
515 ttm_bo_add_to_lru(bo);
518 dma_resv_unlock(bo->base.resv);
520 if (bo->base.resv != &bo->base._resv)
521 dma_resv_unlock(&bo->base._resv);
524 kref_get(&bo->list_kref);
525 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
526 spin_unlock(&glob->lru_lock);
528 schedule_delayed_work(&bdev->wq,
529 ((HZ / 100) < 1) ? 1 : HZ / 100);
533 * function ttm_bo_cleanup_refs
534 * If bo idle, remove from delayed- and lru lists, and unref.
535 * If not idle, do nothing.
537 * Must be called with lru_lock and reservation held, this function
538 * will drop the lru lock and optionally the reservation lock before returning.
540 * @interruptible Any sleeps should occur interruptibly.
541 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
542 * @unlock_resv Unlock the reservation lock as well.
545 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
546 bool interruptible, bool no_wait_gpu,
549 struct ttm_bo_global *glob = bo->bdev->glob;
550 struct dma_resv *resv;
553 if (unlikely(list_empty(&bo->ddestroy)))
554 resv = bo->base.resv;
556 resv = &bo->base._resv;
558 if (dma_resv_test_signaled_rcu(resv, true))
563 if (ret && !no_wait_gpu) {
567 dma_resv_unlock(bo->base.resv);
568 spin_unlock(&glob->lru_lock);
570 lret = dma_resv_wait_timeout_rcu(resv, true,
579 spin_lock(&glob->lru_lock);
580 if (unlock_resv && !dma_resv_trylock(bo->base.resv)) {
582 * We raced, and lost, someone else holds the reservation now,
583 * and is probably busy in ttm_bo_cleanup_memtype_use.
585 * Even if it's not the case, because we finished waiting any
586 * delayed destruction would succeed, so just return success
589 spin_unlock(&glob->lru_lock);
595 if (ret || unlikely(list_empty(&bo->ddestroy))) {
597 dma_resv_unlock(bo->base.resv);
598 spin_unlock(&glob->lru_lock);
602 ttm_bo_del_from_lru(bo);
603 list_del_init(&bo->ddestroy);
604 kref_put(&bo->list_kref, ttm_bo_ref_bug);
606 spin_unlock(&glob->lru_lock);
607 ttm_bo_cleanup_memtype_use(bo);
610 dma_resv_unlock(bo->base.resv);
616 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
617 * encountered buffers.
619 static bool ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
621 struct ttm_bo_global *glob = bdev->glob;
622 struct list_head removed;
625 INIT_LIST_HEAD(&removed);
627 spin_lock(&glob->lru_lock);
628 while (!list_empty(&bdev->ddestroy)) {
629 struct ttm_buffer_object *bo;
631 bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object,
633 kref_get(&bo->list_kref);
634 list_move_tail(&bo->ddestroy, &removed);
636 if (remove_all || bo->base.resv != &bo->base._resv) {
637 spin_unlock(&glob->lru_lock);
638 dma_resv_lock(bo->base.resv, NULL);
640 spin_lock(&glob->lru_lock);
641 ttm_bo_cleanup_refs(bo, false, !remove_all, true);
643 } else if (dma_resv_trylock(bo->base.resv)) {
644 ttm_bo_cleanup_refs(bo, false, !remove_all, true);
646 spin_unlock(&glob->lru_lock);
649 kref_put(&bo->list_kref, ttm_bo_release_list);
650 spin_lock(&glob->lru_lock);
652 list_splice_tail(&removed, &bdev->ddestroy);
653 empty = list_empty(&bdev->ddestroy);
654 spin_unlock(&glob->lru_lock);
659 static void ttm_bo_delayed_workqueue(struct work_struct *work)
661 struct ttm_bo_device *bdev =
662 container_of(work, struct ttm_bo_device, wq.work);
664 if (!ttm_bo_delayed_delete(bdev, false))
665 schedule_delayed_work(&bdev->wq,
666 ((HZ / 100) < 1) ? 1 : HZ / 100);
669 static void ttm_bo_release(struct kref *kref)
671 struct ttm_buffer_object *bo =
672 container_of(kref, struct ttm_buffer_object, kref);
673 struct ttm_bo_device *bdev = bo->bdev;
674 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
676 if (bo->bdev->driver->release_notify)
677 bo->bdev->driver->release_notify(bo);
679 drm_vma_offset_remove(bdev->vma_manager, &bo->base.vma_node);
680 ttm_mem_io_lock(man, false);
681 ttm_mem_io_free_vm(bo);
682 ttm_mem_io_unlock(man);
683 ttm_bo_cleanup_refs_or_queue(bo);
684 kref_put(&bo->list_kref, ttm_bo_release_list);
687 void ttm_bo_put(struct ttm_buffer_object *bo)
689 kref_put(&bo->kref, ttm_bo_release);
691 EXPORT_SYMBOL(ttm_bo_put);
693 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
695 return cancel_delayed_work_sync(&bdev->wq);
697 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
699 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
702 schedule_delayed_work(&bdev->wq,
703 ((HZ / 100) < 1) ? 1 : HZ / 100);
705 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
707 static int ttm_bo_evict(struct ttm_buffer_object *bo,
708 struct ttm_operation_ctx *ctx)
710 struct ttm_bo_device *bdev = bo->bdev;
711 struct ttm_mem_reg evict_mem;
712 struct ttm_placement placement;
715 dma_resv_assert_held(bo->base.resv);
717 placement.num_placement = 0;
718 placement.num_busy_placement = 0;
719 bdev->driver->evict_flags(bo, &placement);
721 if (!placement.num_placement && !placement.num_busy_placement) {
722 ret = ttm_bo_pipeline_gutting(bo);
726 return ttm_tt_create(bo, false);
730 evict_mem.mm_node = NULL;
731 evict_mem.bus.io_reserved_vm = false;
732 evict_mem.bus.io_reserved_count = 0;
734 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
736 if (ret != -ERESTARTSYS) {
737 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
739 ttm_bo_mem_space_debug(bo, &placement);
744 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, ctx);
746 if (ret != -ERESTARTSYS)
747 pr_err("Buffer eviction failed\n");
748 ttm_bo_mem_put(bo, &evict_mem);
756 bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
757 const struct ttm_place *place)
759 /* Don't evict this BO if it's outside of the
760 * requested placement range
762 if (place->fpfn >= (bo->mem.start + bo->mem.size) ||
763 (place->lpfn && place->lpfn <= bo->mem.start))
768 EXPORT_SYMBOL(ttm_bo_eviction_valuable);
771 * Check the target bo is allowable to be evicted or swapout, including cases:
773 * a. if share same reservation object with ctx->resv, have assumption
774 * reservation objects should already be locked, so not lock again and
775 * return true directly when either the opreation allow_reserved_eviction
776 * or the target bo already is in delayed free list;
778 * b. Otherwise, trylock it.
780 static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
781 struct ttm_operation_ctx *ctx, bool *locked, bool *busy)
785 if (bo->base.resv == ctx->resv) {
786 dma_resv_assert_held(bo->base.resv);
787 if (ctx->flags & TTM_OPT_FLAG_ALLOW_RES_EVICT
788 || !list_empty(&bo->ddestroy))
794 ret = dma_resv_trylock(bo->base.resv);
804 * ttm_mem_evict_wait_busy - wait for a busy BO to become available
806 * @busy_bo: BO which couldn't be locked with trylock
807 * @ctx: operation context
808 * @ticket: acquire ticket
810 * Try to lock a busy buffer object to avoid failing eviction.
812 static int ttm_mem_evict_wait_busy(struct ttm_buffer_object *busy_bo,
813 struct ttm_operation_ctx *ctx,
814 struct ww_acquire_ctx *ticket)
818 if (!busy_bo || !ticket)
821 if (ctx->interruptible)
822 r = dma_resv_lock_interruptible(busy_bo->base.resv,
825 r = dma_resv_lock(busy_bo->base.resv, ticket);
828 * TODO: It would be better to keep the BO locked until allocation is at
829 * least tried one more time, but that would mean a much larger rework
833 dma_resv_unlock(busy_bo->base.resv);
835 return r == -EDEADLK ? -EBUSY : r;
838 static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
840 const struct ttm_place *place,
841 struct ttm_operation_ctx *ctx,
842 struct ww_acquire_ctx *ticket)
844 struct ttm_buffer_object *bo = NULL, *busy_bo = NULL;
845 struct ttm_bo_global *glob = bdev->glob;
846 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
851 spin_lock(&glob->lru_lock);
852 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
853 list_for_each_entry(bo, &man->lru[i], lru) {
856 if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked,
858 if (busy && !busy_bo && ticket !=
859 dma_resv_locking_ctx(bo->base.resv))
864 if (place && !bdev->driver->eviction_valuable(bo,
867 dma_resv_unlock(bo->base.resv);
873 /* If the inner loop terminated early, we have our candidate */
874 if (&bo->lru != &man->lru[i])
882 kref_get(&busy_bo->list_kref);
883 spin_unlock(&glob->lru_lock);
884 ret = ttm_mem_evict_wait_busy(busy_bo, ctx, ticket);
886 kref_put(&busy_bo->list_kref, ttm_bo_release_list);
890 kref_get(&bo->list_kref);
892 if (!list_empty(&bo->ddestroy)) {
893 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
894 ctx->no_wait_gpu, locked);
895 kref_put(&bo->list_kref, ttm_bo_release_list);
899 ttm_bo_del_from_lru(bo);
900 spin_unlock(&glob->lru_lock);
902 ret = ttm_bo_evict(bo, ctx);
904 ttm_bo_unreserve(bo);
906 spin_lock(&glob->lru_lock);
907 ttm_bo_add_to_lru(bo);
908 spin_unlock(&glob->lru_lock);
911 kref_put(&bo->list_kref, ttm_bo_release_list);
915 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
917 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
920 (*man->func->put_node)(man, mem);
922 EXPORT_SYMBOL(ttm_bo_mem_put);
925 * Add the last move fence to the BO and reserve a new shared slot.
927 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
928 struct ttm_mem_type_manager *man,
929 struct ttm_mem_reg *mem,
932 struct dma_fence *fence;
935 spin_lock(&man->move_lock);
936 fence = dma_fence_get(man->move);
937 spin_unlock(&man->move_lock);
945 dma_resv_add_shared_fence(bo->base.resv, fence);
947 ret = dma_resv_reserve_shared(bo->base.resv, 1);
949 dma_fence_put(fence);
953 dma_fence_put(bo->moving);
959 * Repeatedly evict memory from the LRU for @mem_type until we create enough
960 * space, or we've evicted everything and there isn't enough space.
962 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
963 const struct ttm_place *place,
964 struct ttm_mem_reg *mem,
965 struct ttm_operation_ctx *ctx)
967 struct ttm_bo_device *bdev = bo->bdev;
968 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
969 struct ww_acquire_ctx *ticket;
972 ticket = dma_resv_locking_ctx(bo->base.resv);
974 ret = (*man->func->get_node)(man, bo, place, mem);
975 if (unlikely(ret != 0))
979 ret = ttm_mem_evict_first(bdev, mem->mem_type, place, ctx,
981 if (unlikely(ret != 0))
985 return ttm_bo_add_move_fence(bo, man, mem, ctx->no_wait_gpu);
988 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
989 uint32_t cur_placement,
990 uint32_t proposed_placement)
992 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
993 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
996 * Keep current caching if possible.
999 if ((cur_placement & caching) != 0)
1000 result |= (cur_placement & caching);
1001 else if ((man->default_caching & caching) != 0)
1002 result |= man->default_caching;
1003 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
1004 result |= TTM_PL_FLAG_CACHED;
1005 else if ((TTM_PL_FLAG_WC & caching) != 0)
1006 result |= TTM_PL_FLAG_WC;
1007 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
1008 result |= TTM_PL_FLAG_UNCACHED;
1013 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
1015 const struct ttm_place *place,
1016 uint32_t *masked_placement)
1018 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
1020 if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0)
1023 if ((place->flags & man->available_caching) == 0)
1026 cur_flags |= (place->flags & man->available_caching);
1028 *masked_placement = cur_flags;
1033 * ttm_bo_mem_placement - check if placement is compatible
1034 * @bo: BO to find memory for
1035 * @place: where to search
1036 * @mem: the memory object to fill in
1037 * @ctx: operation context
1039 * Check if placement is compatible and fill in mem structure.
1040 * Returns -EBUSY if placement won't work or negative error code.
1041 * 0 when placement can be used.
1043 static int ttm_bo_mem_placement(struct ttm_buffer_object *bo,
1044 const struct ttm_place *place,
1045 struct ttm_mem_reg *mem,
1046 struct ttm_operation_ctx *ctx)
1048 struct ttm_bo_device *bdev = bo->bdev;
1049 uint32_t mem_type = TTM_PL_SYSTEM;
1050 struct ttm_mem_type_manager *man;
1051 uint32_t cur_flags = 0;
1054 ret = ttm_mem_type_from_place(place, &mem_type);
1058 man = &bdev->man[mem_type];
1059 if (!man->has_type || !man->use_type)
1062 if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags))
1065 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, cur_flags);
1067 * Use the access and other non-mapping-related flag bits from
1068 * the memory placement flags to the current flags
1070 ttm_flag_masked(&cur_flags, place->flags, ~TTM_PL_MASK_MEMTYPE);
1072 mem->mem_type = mem_type;
1073 mem->placement = cur_flags;
1075 if (bo->mem.mem_type < mem_type && !list_empty(&bo->lru)) {
1076 spin_lock(&bo->bdev->glob->lru_lock);
1077 ttm_bo_del_from_lru(bo);
1078 ttm_bo_add_mem_to_lru(bo, mem);
1079 spin_unlock(&bo->bdev->glob->lru_lock);
1086 * Creates space for memory region @mem according to its type.
1088 * This function first searches for free space in compatible memory types in
1089 * the priority order defined by the driver. If free space isn't found, then
1090 * ttm_bo_mem_force_space is attempted in priority order to evict and find
1093 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
1094 struct ttm_placement *placement,
1095 struct ttm_mem_reg *mem,
1096 struct ttm_operation_ctx *ctx)
1098 struct ttm_bo_device *bdev = bo->bdev;
1099 bool type_found = false;
1102 ret = dma_resv_reserve_shared(bo->base.resv, 1);
1106 mem->mm_node = NULL;
1107 for (i = 0; i < placement->num_placement; ++i) {
1108 const struct ttm_place *place = &placement->placement[i];
1109 struct ttm_mem_type_manager *man;
1111 ret = ttm_bo_mem_placement(bo, place, mem, ctx);
1118 mem->mm_node = NULL;
1119 if (mem->mem_type == TTM_PL_SYSTEM)
1122 man = &bdev->man[mem->mem_type];
1123 ret = (*man->func->get_node)(man, bo, place, mem);
1130 ret = ttm_bo_add_move_fence(bo, man, mem, ctx->no_wait_gpu);
1131 if (unlikely(ret)) {
1132 (*man->func->put_node)(man, mem);
1141 for (i = 0; i < placement->num_busy_placement; ++i) {
1142 const struct ttm_place *place = &placement->busy_placement[i];
1144 ret = ttm_bo_mem_placement(bo, place, mem, ctx);
1151 mem->mm_node = NULL;
1152 if (mem->mem_type == TTM_PL_SYSTEM)
1155 ret = ttm_bo_mem_force_space(bo, place, mem, ctx);
1156 if (ret == 0 && mem->mm_node)
1159 if (ret && ret != -EBUSY)
1165 pr_err(TTM_PFX "No compatible memory type found\n");
1170 if (bo->mem.mem_type == TTM_PL_SYSTEM && !list_empty(&bo->lru)) {
1171 spin_lock(&bo->bdev->glob->lru_lock);
1172 ttm_bo_move_to_lru_tail(bo, NULL);
1173 spin_unlock(&bo->bdev->glob->lru_lock);
1178 EXPORT_SYMBOL(ttm_bo_mem_space);
1180 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1181 struct ttm_placement *placement,
1182 struct ttm_operation_ctx *ctx)
1185 struct ttm_mem_reg mem;
1187 dma_resv_assert_held(bo->base.resv);
1189 mem.num_pages = bo->num_pages;
1190 mem.size = mem.num_pages << PAGE_SHIFT;
1191 mem.page_alignment = bo->mem.page_alignment;
1192 mem.bus.io_reserved_vm = false;
1193 mem.bus.io_reserved_count = 0;
1195 * Determine where to move the buffer.
1197 ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
1200 ret = ttm_bo_handle_move_mem(bo, &mem, false, ctx);
1202 if (ret && mem.mm_node)
1203 ttm_bo_mem_put(bo, &mem);
1207 static bool ttm_bo_places_compat(const struct ttm_place *places,
1208 unsigned num_placement,
1209 struct ttm_mem_reg *mem,
1210 uint32_t *new_flags)
1214 for (i = 0; i < num_placement; i++) {
1215 const struct ttm_place *heap = &places[i];
1217 if (mem->mm_node && (mem->start < heap->fpfn ||
1218 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1221 *new_flags = heap->flags;
1222 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1223 (*new_flags & mem->placement & TTM_PL_MASK_MEM) &&
1224 (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) ||
1225 (mem->placement & TTM_PL_FLAG_CONTIGUOUS)))
1231 bool ttm_bo_mem_compat(struct ttm_placement *placement,
1232 struct ttm_mem_reg *mem,
1233 uint32_t *new_flags)
1235 if (ttm_bo_places_compat(placement->placement, placement->num_placement,
1239 if ((placement->busy_placement != placement->placement ||
1240 placement->num_busy_placement > placement->num_placement) &&
1241 ttm_bo_places_compat(placement->busy_placement,
1242 placement->num_busy_placement,
1248 EXPORT_SYMBOL(ttm_bo_mem_compat);
1250 int ttm_bo_validate(struct ttm_buffer_object *bo,
1251 struct ttm_placement *placement,
1252 struct ttm_operation_ctx *ctx)
1257 dma_resv_assert_held(bo->base.resv);
1259 * Check whether we need to move buffer.
1261 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1262 ret = ttm_bo_move_buffer(bo, placement, ctx);
1267 * Use the access and other non-mapping-related flag bits from
1268 * the compatible memory placement flags to the active flags
1270 ttm_flag_masked(&bo->mem.placement, new_flags,
1271 ~TTM_PL_MASK_MEMTYPE);
1274 * We might need to add a TTM.
1276 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1277 ret = ttm_tt_create(bo, true);
1283 EXPORT_SYMBOL(ttm_bo_validate);
1285 int ttm_bo_init_reserved(struct ttm_bo_device *bdev,
1286 struct ttm_buffer_object *bo,
1288 enum ttm_bo_type type,
1289 struct ttm_placement *placement,
1290 uint32_t page_alignment,
1291 struct ttm_operation_ctx *ctx,
1293 struct sg_table *sg,
1294 struct dma_resv *resv,
1295 void (*destroy) (struct ttm_buffer_object *))
1298 unsigned long num_pages;
1299 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1302 ret = ttm_mem_global_alloc(mem_glob, acc_size, ctx);
1304 pr_err("Out of kernel memory\n");
1312 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1313 if (num_pages == 0) {
1314 pr_err("Illegal buffer object size\n");
1319 ttm_mem_global_free(mem_glob, acc_size);
1322 bo->destroy = destroy ? destroy : ttm_bo_default_destroy;
1324 kref_init(&bo->kref);
1325 kref_init(&bo->list_kref);
1326 atomic_set(&bo->cpu_writers, 0);
1327 INIT_LIST_HEAD(&bo->lru);
1328 INIT_LIST_HEAD(&bo->ddestroy);
1329 INIT_LIST_HEAD(&bo->swap);
1330 INIT_LIST_HEAD(&bo->io_reserve_lru);
1331 mutex_init(&bo->wu_mutex);
1334 bo->num_pages = num_pages;
1335 bo->mem.size = num_pages << PAGE_SHIFT;
1336 bo->mem.mem_type = TTM_PL_SYSTEM;
1337 bo->mem.num_pages = bo->num_pages;
1338 bo->mem.mm_node = NULL;
1339 bo->mem.page_alignment = page_alignment;
1340 bo->mem.bus.io_reserved_vm = false;
1341 bo->mem.bus.io_reserved_count = 0;
1343 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1344 bo->acc_size = acc_size;
1347 bo->base.resv = resv;
1348 dma_resv_assert_held(bo->base.resv);
1350 bo->base.resv = &bo->base._resv;
1352 if (!ttm_bo_uses_embedded_gem_object(bo)) {
1354 * bo.gem is not initialized, so we have to setup the
1355 * struct elements we want use regardless.
1357 dma_resv_init(&bo->base._resv);
1358 drm_vma_node_reset(&bo->base.vma_node);
1360 atomic_inc(&bo->bdev->glob->bo_count);
1363 * For ttm_bo_type_device buffers, allocate
1364 * address space from the device.
1366 if (bo->type == ttm_bo_type_device ||
1367 bo->type == ttm_bo_type_sg)
1368 ret = drm_vma_offset_add(bdev->vma_manager, &bo->base.vma_node,
1371 /* passed reservation objects should already be locked,
1372 * since otherwise lockdep will be angered in radeon.
1375 locked = dma_resv_trylock(bo->base.resv);
1380 ret = ttm_bo_validate(bo, placement, ctx);
1382 if (unlikely(ret)) {
1384 ttm_bo_unreserve(bo);
1390 if (resv && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
1391 spin_lock(&bdev->glob->lru_lock);
1392 ttm_bo_add_to_lru(bo);
1393 spin_unlock(&bdev->glob->lru_lock);
1398 EXPORT_SYMBOL(ttm_bo_init_reserved);
1400 int ttm_bo_init(struct ttm_bo_device *bdev,
1401 struct ttm_buffer_object *bo,
1403 enum ttm_bo_type type,
1404 struct ttm_placement *placement,
1405 uint32_t page_alignment,
1408 struct sg_table *sg,
1409 struct dma_resv *resv,
1410 void (*destroy) (struct ttm_buffer_object *))
1412 struct ttm_operation_ctx ctx = { interruptible, false };
1415 ret = ttm_bo_init_reserved(bdev, bo, size, type, placement,
1416 page_alignment, &ctx, acc_size,
1422 ttm_bo_unreserve(bo);
1426 EXPORT_SYMBOL(ttm_bo_init);
1428 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1429 unsigned long bo_size,
1430 unsigned struct_size)
1432 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1435 size += ttm_round_pot(struct_size);
1436 size += ttm_round_pot(npages * sizeof(void *));
1437 size += ttm_round_pot(sizeof(struct ttm_tt));
1440 EXPORT_SYMBOL(ttm_bo_acc_size);
1442 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1443 unsigned long bo_size,
1444 unsigned struct_size)
1446 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1449 size += ttm_round_pot(struct_size);
1450 size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t)));
1451 size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1454 EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1456 int ttm_bo_create(struct ttm_bo_device *bdev,
1458 enum ttm_bo_type type,
1459 struct ttm_placement *placement,
1460 uint32_t page_alignment,
1462 struct ttm_buffer_object **p_bo)
1464 struct ttm_buffer_object *bo;
1468 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1469 if (unlikely(bo == NULL))
1472 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1473 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1474 interruptible, acc_size,
1476 if (likely(ret == 0))
1481 EXPORT_SYMBOL(ttm_bo_create);
1483 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1486 struct ttm_operation_ctx ctx = {
1487 .interruptible = false,
1488 .no_wait_gpu = false,
1489 .flags = TTM_OPT_FLAG_FORCE_ALLOC
1491 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1492 struct ttm_bo_global *glob = bdev->glob;
1493 struct dma_fence *fence;
1498 * Can't use standard list traversal since we're unlocking.
1501 spin_lock(&glob->lru_lock);
1502 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
1503 while (!list_empty(&man->lru[i])) {
1504 spin_unlock(&glob->lru_lock);
1505 ret = ttm_mem_evict_first(bdev, mem_type, NULL, &ctx,
1509 spin_lock(&glob->lru_lock);
1512 spin_unlock(&glob->lru_lock);
1514 spin_lock(&man->move_lock);
1515 fence = dma_fence_get(man->move);
1516 spin_unlock(&man->move_lock);
1519 ret = dma_fence_wait(fence, false);
1520 dma_fence_put(fence);
1528 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1530 struct ttm_mem_type_manager *man;
1533 if (mem_type >= TTM_NUM_MEM_TYPES) {
1534 pr_err("Illegal memory type %d\n", mem_type);
1537 man = &bdev->man[mem_type];
1539 if (!man->has_type) {
1540 pr_err("Trying to take down uninitialized memory manager type %u\n",
1545 man->use_type = false;
1546 man->has_type = false;
1550 ret = ttm_bo_force_list_clean(bdev, mem_type);
1552 pr_err("Cleanup eviction failed\n");
1556 ret = (*man->func->takedown)(man);
1559 dma_fence_put(man->move);
1564 EXPORT_SYMBOL(ttm_bo_clean_mm);
1566 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1568 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1570 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1571 pr_err("Illegal memory manager memory type %u\n", mem_type);
1575 if (!man->has_type) {
1576 pr_err("Memory type %u has not been initialized\n", mem_type);
1580 return ttm_bo_force_list_clean(bdev, mem_type);
1582 EXPORT_SYMBOL(ttm_bo_evict_mm);
1584 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1585 unsigned long p_size)
1588 struct ttm_mem_type_manager *man;
1591 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1592 man = &bdev->man[type];
1593 BUG_ON(man->has_type);
1594 man->io_reserve_fastpath = true;
1595 man->use_io_reserve_lru = false;
1596 mutex_init(&man->io_reserve_mutex);
1597 spin_lock_init(&man->move_lock);
1598 INIT_LIST_HEAD(&man->io_reserve_lru);
1600 ret = bdev->driver->init_mem_type(bdev, type, man);
1605 if (type != TTM_PL_SYSTEM) {
1606 ret = (*man->func->init)(man, p_size);
1610 man->has_type = true;
1611 man->use_type = true;
1614 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1615 INIT_LIST_HEAD(&man->lru[i]);
1620 EXPORT_SYMBOL(ttm_bo_init_mm);
1622 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1624 struct ttm_bo_global *glob =
1625 container_of(kobj, struct ttm_bo_global, kobj);
1627 __free_page(glob->dummy_read_page);
1630 static void ttm_bo_global_release(void)
1632 struct ttm_bo_global *glob = &ttm_bo_glob;
1634 mutex_lock(&ttm_global_mutex);
1635 if (--ttm_bo_glob_use_count > 0)
1638 kobject_del(&glob->kobj);
1639 kobject_put(&glob->kobj);
1640 ttm_mem_global_release(&ttm_mem_glob);
1641 memset(glob, 0, sizeof(*glob));
1643 mutex_unlock(&ttm_global_mutex);
1646 static int ttm_bo_global_init(void)
1648 struct ttm_bo_global *glob = &ttm_bo_glob;
1652 mutex_lock(&ttm_global_mutex);
1653 if (++ttm_bo_glob_use_count > 1)
1656 ret = ttm_mem_global_init(&ttm_mem_glob);
1660 spin_lock_init(&glob->lru_lock);
1661 glob->mem_glob = &ttm_mem_glob;
1662 glob->mem_glob->bo_glob = glob;
1663 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1665 if (unlikely(glob->dummy_read_page == NULL)) {
1670 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1671 INIT_LIST_HEAD(&glob->swap_lru[i]);
1672 INIT_LIST_HEAD(&glob->device_list);
1673 atomic_set(&glob->bo_count, 0);
1675 ret = kobject_init_and_add(
1676 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1677 if (unlikely(ret != 0))
1678 kobject_put(&glob->kobj);
1680 mutex_unlock(&ttm_global_mutex);
1684 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1687 unsigned i = TTM_NUM_MEM_TYPES;
1688 struct ttm_mem_type_manager *man;
1689 struct ttm_bo_global *glob = bdev->glob;
1692 man = &bdev->man[i];
1693 if (man->has_type) {
1694 man->use_type = false;
1695 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1697 pr_err("DRM memory manager type %d is not clean\n",
1700 man->has_type = false;
1704 mutex_lock(&ttm_global_mutex);
1705 list_del(&bdev->device_list);
1706 mutex_unlock(&ttm_global_mutex);
1708 cancel_delayed_work_sync(&bdev->wq);
1710 if (ttm_bo_delayed_delete(bdev, true))
1711 pr_debug("Delayed destroy list was clean\n");
1713 spin_lock(&glob->lru_lock);
1714 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1715 if (list_empty(&bdev->man[0].lru[0]))
1716 pr_debug("Swap list %d was clean\n", i);
1717 spin_unlock(&glob->lru_lock);
1720 ttm_bo_global_release();
1724 EXPORT_SYMBOL(ttm_bo_device_release);
1726 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1727 struct ttm_bo_driver *driver,
1728 struct address_space *mapping,
1729 struct drm_vma_offset_manager *vma_manager,
1732 struct ttm_bo_global *glob = &ttm_bo_glob;
1735 if (WARN_ON(vma_manager == NULL))
1738 ret = ttm_bo_global_init();
1742 bdev->driver = driver;
1744 memset(bdev->man, 0, sizeof(bdev->man));
1747 * Initialize the system memory buffer type.
1748 * Other types need to be driver / IOCTL initialized.
1750 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1751 if (unlikely(ret != 0))
1754 bdev->vma_manager = vma_manager;
1755 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1756 INIT_LIST_HEAD(&bdev->ddestroy);
1757 bdev->dev_mapping = mapping;
1759 bdev->need_dma32 = need_dma32;
1760 mutex_lock(&ttm_global_mutex);
1761 list_add_tail(&bdev->device_list, &glob->device_list);
1762 mutex_unlock(&ttm_global_mutex);
1766 ttm_bo_global_release();
1769 EXPORT_SYMBOL(ttm_bo_device_init);
1772 * buffer object vm functions.
1775 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1777 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1779 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1780 if (mem->mem_type == TTM_PL_SYSTEM)
1783 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1786 if (mem->placement & TTM_PL_FLAG_CACHED)
1792 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1794 struct ttm_bo_device *bdev = bo->bdev;
1796 drm_vma_node_unmap(&bo->base.vma_node, bdev->dev_mapping);
1797 ttm_mem_io_free_vm(bo);
1800 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1802 struct ttm_bo_device *bdev = bo->bdev;
1803 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1805 ttm_mem_io_lock(man, false);
1806 ttm_bo_unmap_virtual_locked(bo);
1807 ttm_mem_io_unlock(man);
1811 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1813 int ttm_bo_wait(struct ttm_buffer_object *bo,
1814 bool interruptible, bool no_wait)
1816 long timeout = 15 * HZ;
1819 if (dma_resv_test_signaled_rcu(bo->base.resv, true))
1825 timeout = dma_resv_wait_timeout_rcu(bo->base.resv, true,
1826 interruptible, timeout);
1833 dma_resv_add_excl_fence(bo->base.resv, NULL);
1836 EXPORT_SYMBOL(ttm_bo_wait);
1838 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1843 * Using ttm_bo_reserve makes sure the lru lists are updated.
1846 ret = ttm_bo_reserve(bo, true, no_wait, NULL);
1847 if (unlikely(ret != 0))
1849 ret = ttm_bo_wait(bo, true, no_wait);
1850 if (likely(ret == 0))
1851 atomic_inc(&bo->cpu_writers);
1852 ttm_bo_unreserve(bo);
1855 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1857 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1859 atomic_dec(&bo->cpu_writers);
1861 EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1864 * A buffer object shrink method that tries to swap out the first
1865 * buffer object on the bo_global::swap_lru list.
1867 int ttm_bo_swapout(struct ttm_bo_global *glob, struct ttm_operation_ctx *ctx)
1869 struct ttm_buffer_object *bo;
1874 spin_lock(&glob->lru_lock);
1875 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
1876 list_for_each_entry(bo, &glob->swap_lru[i], swap) {
1877 if (ttm_bo_evict_swapout_allowable(bo, ctx, &locked,
1888 spin_unlock(&glob->lru_lock);
1892 kref_get(&bo->list_kref);
1894 if (!list_empty(&bo->ddestroy)) {
1895 ret = ttm_bo_cleanup_refs(bo, false, false, locked);
1896 kref_put(&bo->list_kref, ttm_bo_release_list);
1900 ttm_bo_del_from_lru(bo);
1901 spin_unlock(&glob->lru_lock);
1904 * Move to system cached
1907 if (bo->mem.mem_type != TTM_PL_SYSTEM ||
1908 bo->ttm->caching_state != tt_cached) {
1909 struct ttm_operation_ctx ctx = { false, false };
1910 struct ttm_mem_reg evict_mem;
1912 evict_mem = bo->mem;
1913 evict_mem.mm_node = NULL;
1914 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1915 evict_mem.mem_type = TTM_PL_SYSTEM;
1917 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, &ctx);
1918 if (unlikely(ret != 0))
1923 * Make sure BO is idle.
1926 ret = ttm_bo_wait(bo, false, false);
1927 if (unlikely(ret != 0))
1930 ttm_bo_unmap_virtual(bo);
1933 * Swap out. Buffer will be swapped in again as soon as
1934 * anyone tries to access a ttm page.
1937 if (bo->bdev->driver->swap_notify)
1938 bo->bdev->driver->swap_notify(bo);
1940 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1945 * Unreserve without putting on LRU to avoid swapping out an
1946 * already swapped buffer.
1949 dma_resv_unlock(bo->base.resv);
1950 kref_put(&bo->list_kref, ttm_bo_release_list);
1953 EXPORT_SYMBOL(ttm_bo_swapout);
1955 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1957 struct ttm_operation_ctx ctx = {
1958 .interruptible = false,
1959 .no_wait_gpu = false
1962 while (ttm_bo_swapout(bdev->glob, &ctx) == 0)
1965 EXPORT_SYMBOL(ttm_bo_swapout_all);
1968 * ttm_bo_wait_unreserved - interruptible wait for a buffer object to become
1971 * @bo: Pointer to buffer
1973 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo)
1978 * In the absense of a wait_unlocked API,
1979 * Use the bo::wu_mutex to avoid triggering livelocks due to
1980 * concurrent use of this function. Note that this use of
1981 * bo::wu_mutex can go away if we change locking order to
1982 * mmap_sem -> bo::reserve.
1984 ret = mutex_lock_interruptible(&bo->wu_mutex);
1985 if (unlikely(ret != 0))
1986 return -ERESTARTSYS;
1987 if (!dma_resv_is_locked(bo->base.resv))
1989 ret = dma_resv_lock_interruptible(bo->base.resv, NULL);
1992 if (unlikely(ret != 0))
1994 dma_resv_unlock(bo->base.resv);
1997 mutex_unlock(&bo->wu_mutex);