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;
54 EXPORT_SYMBOL(ttm_bo_glob);
56 static struct attribute ttm_bo_count = {
61 /* default destructor */
62 static void ttm_bo_default_destroy(struct ttm_buffer_object *bo)
67 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
68 struct ttm_placement *placement)
70 struct drm_printer p = drm_debug_printer(TTM_PFX);
71 struct ttm_resource_manager *man;
74 drm_printf(&p, "No space for %p (%lu pages, %luK, %luM)\n",
75 bo, bo->mem.num_pages, bo->mem.size >> 10,
77 for (i = 0; i < placement->num_placement; i++) {
78 mem_type = placement->placement[i].mem_type;
79 drm_printf(&p, " placement[%d]=0x%08X (%d)\n",
80 i, placement->placement[i].flags, mem_type);
81 man = ttm_manager_type(bo->bdev, mem_type);
82 ttm_resource_manager_debug(man, &p);
86 static ssize_t ttm_bo_global_show(struct kobject *kobj,
87 struct attribute *attr,
90 struct ttm_bo_global *glob =
91 container_of(kobj, struct ttm_bo_global, kobj);
93 return snprintf(buffer, PAGE_SIZE, "%d\n",
94 atomic_read(&glob->bo_count));
97 static struct attribute *ttm_bo_global_attrs[] = {
102 static const struct sysfs_ops ttm_bo_global_ops = {
103 .show = &ttm_bo_global_show
106 static struct kobj_type ttm_bo_glob_kobj_type = {
107 .release = &ttm_bo_global_kobj_release,
108 .sysfs_ops = &ttm_bo_global_ops,
109 .default_attrs = ttm_bo_global_attrs
112 static void ttm_bo_add_mem_to_lru(struct ttm_buffer_object *bo,
113 struct ttm_resource *mem)
115 struct ttm_bo_device *bdev = bo->bdev;
116 struct ttm_resource_manager *man;
118 if (!list_empty(&bo->lru))
121 if (mem->placement & TTM_PL_FLAG_NO_EVICT)
124 man = ttm_manager_type(bdev, mem->mem_type);
125 list_add_tail(&bo->lru, &man->lru[bo->priority]);
127 if (man->use_tt && bo->ttm &&
128 !(bo->ttm->page_flags & (TTM_PAGE_FLAG_SG |
129 TTM_PAGE_FLAG_SWAPPED))) {
130 list_add_tail(&bo->swap, &ttm_bo_glob.swap_lru[bo->priority]);
134 static void ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
136 struct ttm_bo_device *bdev = bo->bdev;
139 if (!list_empty(&bo->swap)) {
140 list_del_init(&bo->swap);
143 if (!list_empty(&bo->lru)) {
144 list_del_init(&bo->lru);
148 if (notify && bdev->driver->del_from_lru_notify)
149 bdev->driver->del_from_lru_notify(bo);
152 static void ttm_bo_bulk_move_set_pos(struct ttm_lru_bulk_move_pos *pos,
153 struct ttm_buffer_object *bo)
160 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo,
161 struct ttm_lru_bulk_move *bulk)
163 dma_resv_assert_held(bo->base.resv);
165 ttm_bo_del_from_lru(bo);
166 ttm_bo_add_mem_to_lru(bo, &bo->mem);
168 if (bulk && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
169 switch (bo->mem.mem_type) {
171 ttm_bo_bulk_move_set_pos(&bulk->tt[bo->priority], bo);
175 ttm_bo_bulk_move_set_pos(&bulk->vram[bo->priority], bo);
178 if (bo->ttm && !(bo->ttm->page_flags &
179 (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED)))
180 ttm_bo_bulk_move_set_pos(&bulk->swap[bo->priority], bo);
183 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
185 void ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move *bulk)
189 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
190 struct ttm_lru_bulk_move_pos *pos = &bulk->tt[i];
191 struct ttm_resource_manager *man;
196 dma_resv_assert_held(pos->first->base.resv);
197 dma_resv_assert_held(pos->last->base.resv);
199 man = ttm_manager_type(pos->first->bdev, TTM_PL_TT);
200 list_bulk_move_tail(&man->lru[i], &pos->first->lru,
204 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
205 struct ttm_lru_bulk_move_pos *pos = &bulk->vram[i];
206 struct ttm_resource_manager *man;
211 dma_resv_assert_held(pos->first->base.resv);
212 dma_resv_assert_held(pos->last->base.resv);
214 man = ttm_manager_type(pos->first->bdev, TTM_PL_VRAM);
215 list_bulk_move_tail(&man->lru[i], &pos->first->lru,
219 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
220 struct ttm_lru_bulk_move_pos *pos = &bulk->swap[i];
221 struct list_head *lru;
226 dma_resv_assert_held(pos->first->base.resv);
227 dma_resv_assert_held(pos->last->base.resv);
229 lru = &ttm_bo_glob.swap_lru[i];
230 list_bulk_move_tail(lru, &pos->first->swap, &pos->last->swap);
233 EXPORT_SYMBOL(ttm_bo_bulk_move_lru_tail);
235 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
236 struct ttm_resource *mem, bool evict,
237 struct ttm_operation_ctx *ctx)
239 struct ttm_bo_device *bdev = bo->bdev;
240 struct ttm_resource_manager *old_man = ttm_manager_type(bdev, bo->mem.mem_type);
241 struct ttm_resource_manager *new_man = ttm_manager_type(bdev, mem->mem_type);
244 ttm_bo_unmap_virtual(bo);
247 * Create and bind a ttm if required.
250 if (new_man->use_tt) {
251 /* Zero init the new TTM structure if the old location should
252 * have used one as well.
254 ret = ttm_tt_create(bo, old_man->use_tt);
258 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
262 if (mem->mem_type != TTM_PL_SYSTEM) {
263 ret = ttm_tt_bind(bdev, bo->ttm, mem, ctx);
268 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
269 if (bdev->driver->move_notify)
270 bdev->driver->move_notify(bo, evict, mem);
276 if (bdev->driver->move_notify)
277 bdev->driver->move_notify(bo, evict, mem);
279 if (old_man->use_tt && new_man->use_tt)
280 ret = ttm_bo_move_ttm(bo, ctx, mem);
281 else if (bdev->driver->move)
282 ret = bdev->driver->move(bo, evict, ctx, mem);
284 ret = ttm_bo_move_memcpy(bo, ctx, mem);
287 if (bdev->driver->move_notify) {
289 bdev->driver->move_notify(bo, false, mem);
299 ctx->bytes_moved += bo->num_pages << PAGE_SHIFT;
303 new_man = ttm_manager_type(bdev, bo->mem.mem_type);
304 if (!new_man->use_tt) {
305 ttm_tt_destroy(bdev, bo->ttm);
314 * Will release GPU memory type usage on destruction.
315 * This is the place to put in driver specific hooks to release
316 * driver private resources.
317 * Will release the bo::reserved lock.
320 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
322 if (bo->bdev->driver->move_notify)
323 bo->bdev->driver->move_notify(bo, false, NULL);
325 ttm_tt_destroy(bo->bdev, bo->ttm);
327 ttm_resource_free(bo, &bo->mem);
330 static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
334 if (bo->base.resv == &bo->base._resv)
337 BUG_ON(!dma_resv_trylock(&bo->base._resv));
339 r = dma_resv_copy_fences(&bo->base._resv, bo->base.resv);
340 dma_resv_unlock(&bo->base._resv);
344 if (bo->type != ttm_bo_type_sg) {
345 /* This works because the BO is about to be destroyed and nobody
346 * reference it any more. The only tricky case is the trylock on
347 * the resv object while holding the lru_lock.
349 spin_lock(&ttm_bo_glob.lru_lock);
350 bo->base.resv = &bo->base._resv;
351 spin_unlock(&ttm_bo_glob.lru_lock);
357 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
359 struct dma_resv *resv = &bo->base._resv;
360 struct dma_resv_list *fobj;
361 struct dma_fence *fence;
365 fobj = rcu_dereference(resv->fence);
366 fence = rcu_dereference(resv->fence_excl);
367 if (fence && !fence->ops->signaled)
368 dma_fence_enable_sw_signaling(fence);
370 for (i = 0; fobj && i < fobj->shared_count; ++i) {
371 fence = rcu_dereference(fobj->shared[i]);
373 if (!fence->ops->signaled)
374 dma_fence_enable_sw_signaling(fence);
380 * function ttm_bo_cleanup_refs
381 * If bo idle, remove from lru lists, and unref.
382 * If not idle, block if possible.
384 * Must be called with lru_lock and reservation held, this function
385 * will drop the lru lock and optionally the reservation lock before returning.
387 * @interruptible Any sleeps should occur interruptibly.
388 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
389 * @unlock_resv Unlock the reservation lock as well.
392 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
393 bool interruptible, bool no_wait_gpu,
396 struct dma_resv *resv = &bo->base._resv;
399 if (dma_resv_test_signaled_rcu(resv, true))
404 if (ret && !no_wait_gpu) {
408 dma_resv_unlock(bo->base.resv);
409 spin_unlock(&ttm_bo_glob.lru_lock);
411 lret = dma_resv_wait_timeout_rcu(resv, true, interruptible,
419 spin_lock(&ttm_bo_glob.lru_lock);
420 if (unlock_resv && !dma_resv_trylock(bo->base.resv)) {
422 * We raced, and lost, someone else holds the reservation now,
423 * and is probably busy in ttm_bo_cleanup_memtype_use.
425 * Even if it's not the case, because we finished waiting any
426 * delayed destruction would succeed, so just return success
429 spin_unlock(&ttm_bo_glob.lru_lock);
435 if (ret || unlikely(list_empty(&bo->ddestroy))) {
437 dma_resv_unlock(bo->base.resv);
438 spin_unlock(&ttm_bo_glob.lru_lock);
442 ttm_bo_del_from_lru(bo);
443 list_del_init(&bo->ddestroy);
444 spin_unlock(&ttm_bo_glob.lru_lock);
445 ttm_bo_cleanup_memtype_use(bo);
448 dma_resv_unlock(bo->base.resv);
456 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
457 * encountered buffers.
459 static bool ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
461 struct ttm_bo_global *glob = &ttm_bo_glob;
462 struct list_head removed;
465 INIT_LIST_HEAD(&removed);
467 spin_lock(&glob->lru_lock);
468 while (!list_empty(&bdev->ddestroy)) {
469 struct ttm_buffer_object *bo;
471 bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object,
473 list_move_tail(&bo->ddestroy, &removed);
474 if (!ttm_bo_get_unless_zero(bo))
477 if (remove_all || bo->base.resv != &bo->base._resv) {
478 spin_unlock(&glob->lru_lock);
479 dma_resv_lock(bo->base.resv, NULL);
481 spin_lock(&glob->lru_lock);
482 ttm_bo_cleanup_refs(bo, false, !remove_all, true);
484 } else if (dma_resv_trylock(bo->base.resv)) {
485 ttm_bo_cleanup_refs(bo, false, !remove_all, true);
487 spin_unlock(&glob->lru_lock);
491 spin_lock(&glob->lru_lock);
493 list_splice_tail(&removed, &bdev->ddestroy);
494 empty = list_empty(&bdev->ddestroy);
495 spin_unlock(&glob->lru_lock);
500 static void ttm_bo_delayed_workqueue(struct work_struct *work)
502 struct ttm_bo_device *bdev =
503 container_of(work, struct ttm_bo_device, wq.work);
505 if (!ttm_bo_delayed_delete(bdev, false))
506 schedule_delayed_work(&bdev->wq,
507 ((HZ / 100) < 1) ? 1 : HZ / 100);
510 static void ttm_bo_release(struct kref *kref)
512 struct ttm_buffer_object *bo =
513 container_of(kref, struct ttm_buffer_object, kref);
514 struct ttm_bo_device *bdev = bo->bdev;
515 size_t acc_size = bo->acc_size;
519 ret = ttm_bo_individualize_resv(bo);
521 /* Last resort, if we fail to allocate memory for the
522 * fences block for the BO to become idle
524 dma_resv_wait_timeout_rcu(bo->base.resv, true, false,
528 if (bo->bdev->driver->release_notify)
529 bo->bdev->driver->release_notify(bo);
531 drm_vma_offset_remove(bdev->vma_manager, &bo->base.vma_node);
532 ttm_mem_io_free(bdev, &bo->mem);
535 if (!dma_resv_test_signaled_rcu(bo->base.resv, true) ||
536 !dma_resv_trylock(bo->base.resv)) {
537 /* The BO is not idle, resurrect it for delayed destroy */
538 ttm_bo_flush_all_fences(bo);
541 spin_lock(&ttm_bo_glob.lru_lock);
544 * Make NO_EVICT bos immediately available to
545 * shrinkers, now that they are queued for
548 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) {
549 bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT;
550 ttm_bo_del_from_lru(bo);
551 ttm_bo_add_mem_to_lru(bo, &bo->mem);
554 kref_init(&bo->kref);
555 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
556 spin_unlock(&ttm_bo_glob.lru_lock);
558 schedule_delayed_work(&bdev->wq,
559 ((HZ / 100) < 1) ? 1 : HZ / 100);
563 spin_lock(&ttm_bo_glob.lru_lock);
564 ttm_bo_del_from_lru(bo);
565 list_del(&bo->ddestroy);
566 spin_unlock(&ttm_bo_glob.lru_lock);
568 ttm_bo_cleanup_memtype_use(bo);
569 dma_resv_unlock(bo->base.resv);
571 atomic_dec(&ttm_bo_glob.bo_count);
572 dma_fence_put(bo->moving);
573 if (!ttm_bo_uses_embedded_gem_object(bo))
574 dma_resv_fini(&bo->base._resv);
576 ttm_mem_global_free(&ttm_mem_glob, acc_size);
579 void ttm_bo_put(struct ttm_buffer_object *bo)
581 kref_put(&bo->kref, ttm_bo_release);
583 EXPORT_SYMBOL(ttm_bo_put);
585 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
587 return cancel_delayed_work_sync(&bdev->wq);
589 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
591 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
594 schedule_delayed_work(&bdev->wq,
595 ((HZ / 100) < 1) ? 1 : HZ / 100);
597 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
599 static int ttm_bo_evict(struct ttm_buffer_object *bo,
600 struct ttm_operation_ctx *ctx)
602 struct ttm_bo_device *bdev = bo->bdev;
603 struct ttm_resource evict_mem;
604 struct ttm_placement placement;
607 dma_resv_assert_held(bo->base.resv);
609 placement.num_placement = 0;
610 placement.num_busy_placement = 0;
611 bdev->driver->evict_flags(bo, &placement);
613 if (!placement.num_placement && !placement.num_busy_placement) {
614 ttm_bo_wait(bo, false, false);
616 ttm_bo_cleanup_memtype_use(bo);
617 return ttm_tt_create(bo, false);
621 evict_mem.mm_node = NULL;
622 evict_mem.bus.offset = 0;
623 evict_mem.bus.addr = NULL;
625 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
627 if (ret != -ERESTARTSYS) {
628 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
630 ttm_bo_mem_space_debug(bo, &placement);
635 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, ctx);
637 if (ret != -ERESTARTSYS)
638 pr_err("Buffer eviction failed\n");
639 ttm_resource_free(bo, &evict_mem);
647 bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
648 const struct ttm_place *place)
650 /* Don't evict this BO if it's outside of the
651 * requested placement range
653 if (place->fpfn >= (bo->mem.start + bo->mem.size) ||
654 (place->lpfn && place->lpfn <= bo->mem.start))
659 EXPORT_SYMBOL(ttm_bo_eviction_valuable);
662 * Check the target bo is allowable to be evicted or swapout, including cases:
664 * a. if share same reservation object with ctx->resv, have assumption
665 * reservation objects should already be locked, so not lock again and
666 * return true directly when either the opreation allow_reserved_eviction
667 * or the target bo already is in delayed free list;
669 * b. Otherwise, trylock it.
671 static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
672 struct ttm_operation_ctx *ctx, bool *locked, bool *busy)
676 if (bo->base.resv == ctx->resv) {
677 dma_resv_assert_held(bo->base.resv);
678 if (ctx->flags & TTM_OPT_FLAG_ALLOW_RES_EVICT)
684 ret = dma_resv_trylock(bo->base.resv);
694 * ttm_mem_evict_wait_busy - wait for a busy BO to become available
696 * @busy_bo: BO which couldn't be locked with trylock
697 * @ctx: operation context
698 * @ticket: acquire ticket
700 * Try to lock a busy buffer object to avoid failing eviction.
702 static int ttm_mem_evict_wait_busy(struct ttm_buffer_object *busy_bo,
703 struct ttm_operation_ctx *ctx,
704 struct ww_acquire_ctx *ticket)
708 if (!busy_bo || !ticket)
711 if (ctx->interruptible)
712 r = dma_resv_lock_interruptible(busy_bo->base.resv,
715 r = dma_resv_lock(busy_bo->base.resv, ticket);
718 * TODO: It would be better to keep the BO locked until allocation is at
719 * least tried one more time, but that would mean a much larger rework
723 dma_resv_unlock(busy_bo->base.resv);
725 return r == -EDEADLK ? -EBUSY : r;
728 int ttm_mem_evict_first(struct ttm_bo_device *bdev,
729 struct ttm_resource_manager *man,
730 const struct ttm_place *place,
731 struct ttm_operation_ctx *ctx,
732 struct ww_acquire_ctx *ticket)
734 struct ttm_buffer_object *bo = NULL, *busy_bo = NULL;
739 spin_lock(&ttm_bo_glob.lru_lock);
740 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
741 list_for_each_entry(bo, &man->lru[i], lru) {
744 if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked,
746 if (busy && !busy_bo && ticket !=
747 dma_resv_locking_ctx(bo->base.resv))
752 if (place && !bdev->driver->eviction_valuable(bo,
755 dma_resv_unlock(bo->base.resv);
758 if (!ttm_bo_get_unless_zero(bo)) {
760 dma_resv_unlock(bo->base.resv);
766 /* If the inner loop terminated early, we have our candidate */
767 if (&bo->lru != &man->lru[i])
774 if (busy_bo && !ttm_bo_get_unless_zero(busy_bo))
776 spin_unlock(&ttm_bo_glob.lru_lock);
777 ret = ttm_mem_evict_wait_busy(busy_bo, ctx, ticket);
784 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
785 ctx->no_wait_gpu, locked);
790 spin_unlock(&ttm_bo_glob.lru_lock);
792 ret = ttm_bo_evict(bo, ctx);
794 ttm_bo_unreserve(bo);
801 * Add the last move fence to the BO and reserve a new shared slot.
803 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
804 struct ttm_resource_manager *man,
805 struct ttm_resource *mem,
808 struct dma_fence *fence;
811 spin_lock(&man->move_lock);
812 fence = dma_fence_get(man->move);
813 spin_unlock(&man->move_lock);
819 dma_fence_put(fence);
823 dma_resv_add_shared_fence(bo->base.resv, fence);
825 ret = dma_resv_reserve_shared(bo->base.resv, 1);
827 dma_fence_put(fence);
831 dma_fence_put(bo->moving);
837 * Repeatedly evict memory from the LRU for @mem_type until we create enough
838 * space, or we've evicted everything and there isn't enough space.
840 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
841 const struct ttm_place *place,
842 struct ttm_resource *mem,
843 struct ttm_operation_ctx *ctx)
845 struct ttm_bo_device *bdev = bo->bdev;
846 struct ttm_resource_manager *man = ttm_manager_type(bdev, mem->mem_type);
847 struct ww_acquire_ctx *ticket;
850 ticket = dma_resv_locking_ctx(bo->base.resv);
852 ret = ttm_resource_alloc(bo, place, mem);
855 if (unlikely(ret != -ENOSPC))
857 ret = ttm_mem_evict_first(bdev, man, place, ctx,
859 if (unlikely(ret != 0))
863 return ttm_bo_add_move_fence(bo, man, mem, ctx->no_wait_gpu);
866 static uint32_t ttm_bo_select_caching(struct ttm_resource_manager *man,
867 uint32_t cur_placement,
868 uint32_t proposed_placement)
870 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
871 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
874 * Keep current caching if possible.
877 if ((cur_placement & caching) != 0)
878 result |= (cur_placement & caching);
879 else if ((man->default_caching & caching) != 0)
880 result |= man->default_caching;
881 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
882 result |= TTM_PL_FLAG_CACHED;
883 else if ((TTM_PL_FLAG_WC & caching) != 0)
884 result |= TTM_PL_FLAG_WC;
885 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
886 result |= TTM_PL_FLAG_UNCACHED;
892 * ttm_bo_mem_placement - check if placement is compatible
893 * @bo: BO to find memory for
894 * @place: where to search
895 * @mem: the memory object to fill in
896 * @ctx: operation context
898 * Check if placement is compatible and fill in mem structure.
899 * Returns -EBUSY if placement won't work or negative error code.
900 * 0 when placement can be used.
902 static int ttm_bo_mem_placement(struct ttm_buffer_object *bo,
903 const struct ttm_place *place,
904 struct ttm_resource *mem,
905 struct ttm_operation_ctx *ctx)
907 struct ttm_bo_device *bdev = bo->bdev;
908 struct ttm_resource_manager *man;
909 uint32_t cur_flags = 0;
911 man = ttm_manager_type(bdev, place->mem_type);
912 if (!man || !ttm_resource_manager_used(man))
915 if ((place->flags & man->available_caching) == 0)
918 cur_flags = place->flags & man->available_caching;
919 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, cur_flags);
920 cur_flags |= place->flags & ~TTM_PL_MASK_CACHING;
922 mem->mem_type = place->mem_type;
923 mem->placement = cur_flags;
925 spin_lock(&ttm_bo_glob.lru_lock);
926 ttm_bo_del_from_lru(bo);
927 ttm_bo_add_mem_to_lru(bo, mem);
928 spin_unlock(&ttm_bo_glob.lru_lock);
934 * Creates space for memory region @mem according to its type.
936 * This function first searches for free space in compatible memory types in
937 * the priority order defined by the driver. If free space isn't found, then
938 * ttm_bo_mem_force_space is attempted in priority order to evict and find
941 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
942 struct ttm_placement *placement,
943 struct ttm_resource *mem,
944 struct ttm_operation_ctx *ctx)
946 struct ttm_bo_device *bdev = bo->bdev;
947 bool type_found = false;
950 ret = dma_resv_reserve_shared(bo->base.resv, 1);
954 for (i = 0; i < placement->num_placement; ++i) {
955 const struct ttm_place *place = &placement->placement[i];
956 struct ttm_resource_manager *man;
958 ret = ttm_bo_mem_placement(bo, place, mem, ctx);
963 ret = ttm_resource_alloc(bo, place, mem);
969 man = ttm_manager_type(bdev, mem->mem_type);
970 ret = ttm_bo_add_move_fence(bo, man, mem, ctx->no_wait_gpu);
972 ttm_resource_free(bo, mem);
981 for (i = 0; i < placement->num_busy_placement; ++i) {
982 const struct ttm_place *place = &placement->busy_placement[i];
984 ret = ttm_bo_mem_placement(bo, place, mem, ctx);
989 ret = ttm_bo_mem_force_space(bo, place, mem, ctx);
993 if (ret && ret != -EBUSY)
999 pr_err(TTM_PFX "No compatible memory type found\n");
1004 if (bo->mem.mem_type == TTM_PL_SYSTEM && !list_empty(&bo->lru)) {
1005 ttm_bo_move_to_lru_tail_unlocked(bo);
1010 EXPORT_SYMBOL(ttm_bo_mem_space);
1012 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1013 struct ttm_placement *placement,
1014 struct ttm_operation_ctx *ctx)
1017 struct ttm_resource mem;
1019 dma_resv_assert_held(bo->base.resv);
1021 mem.num_pages = bo->num_pages;
1022 mem.size = mem.num_pages << PAGE_SHIFT;
1023 mem.page_alignment = bo->mem.page_alignment;
1025 mem.bus.addr = NULL;
1029 * Determine where to move the buffer.
1031 ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
1034 ret = ttm_bo_handle_move_mem(bo, &mem, false, ctx);
1037 ttm_resource_free(bo, &mem);
1041 static bool ttm_bo_places_compat(const struct ttm_place *places,
1042 unsigned num_placement,
1043 struct ttm_resource *mem,
1044 uint32_t *new_flags)
1048 for (i = 0; i < num_placement; i++) {
1049 const struct ttm_place *heap = &places[i];
1051 if ((mem->start < heap->fpfn ||
1052 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1055 *new_flags = heap->flags;
1056 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1057 (mem->mem_type == heap->mem_type) &&
1058 (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) ||
1059 (mem->placement & TTM_PL_FLAG_CONTIGUOUS)))
1065 bool ttm_bo_mem_compat(struct ttm_placement *placement,
1066 struct ttm_resource *mem,
1067 uint32_t *new_flags)
1069 if (ttm_bo_places_compat(placement->placement, placement->num_placement,
1073 if ((placement->busy_placement != placement->placement ||
1074 placement->num_busy_placement > placement->num_placement) &&
1075 ttm_bo_places_compat(placement->busy_placement,
1076 placement->num_busy_placement,
1082 EXPORT_SYMBOL(ttm_bo_mem_compat);
1084 int ttm_bo_validate(struct ttm_buffer_object *bo,
1085 struct ttm_placement *placement,
1086 struct ttm_operation_ctx *ctx)
1091 dma_resv_assert_held(bo->base.resv);
1094 * Remove the backing store if no placement is given.
1096 if (!placement->num_placement && !placement->num_busy_placement) {
1097 ret = ttm_bo_pipeline_gutting(bo);
1101 return ttm_tt_create(bo, false);
1105 * Check whether we need to move buffer.
1107 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1108 ret = ttm_bo_move_buffer(bo, placement, ctx);
1112 bo->mem.placement &= TTM_PL_MASK_CACHING;
1113 bo->mem.placement |= new_flags & ~TTM_PL_MASK_CACHING;
1116 * We might need to add a TTM.
1118 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
1119 ret = ttm_tt_create(bo, true);
1125 EXPORT_SYMBOL(ttm_bo_validate);
1127 int ttm_bo_init_reserved(struct ttm_bo_device *bdev,
1128 struct ttm_buffer_object *bo,
1130 enum ttm_bo_type type,
1131 struct ttm_placement *placement,
1132 uint32_t page_alignment,
1133 struct ttm_operation_ctx *ctx,
1135 struct sg_table *sg,
1136 struct dma_resv *resv,
1137 void (*destroy) (struct ttm_buffer_object *))
1139 struct ttm_mem_global *mem_glob = &ttm_mem_glob;
1141 unsigned long num_pages;
1144 ret = ttm_mem_global_alloc(mem_glob, acc_size, ctx);
1146 pr_err("Out of kernel memory\n");
1154 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1155 if (num_pages == 0) {
1156 pr_err("Illegal buffer object size\n");
1161 ttm_mem_global_free(mem_glob, acc_size);
1164 bo->destroy = destroy ? destroy : ttm_bo_default_destroy;
1166 kref_init(&bo->kref);
1167 INIT_LIST_HEAD(&bo->lru);
1168 INIT_LIST_HEAD(&bo->ddestroy);
1169 INIT_LIST_HEAD(&bo->swap);
1172 bo->num_pages = num_pages;
1173 bo->mem.size = num_pages << PAGE_SHIFT;
1174 bo->mem.mem_type = TTM_PL_SYSTEM;
1175 bo->mem.num_pages = bo->num_pages;
1176 bo->mem.mm_node = NULL;
1177 bo->mem.page_alignment = page_alignment;
1178 bo->mem.bus.offset = 0;
1179 bo->mem.bus.addr = NULL;
1181 bo->mem.placement = TTM_PL_FLAG_CACHED;
1182 bo->acc_size = acc_size;
1185 bo->base.resv = resv;
1186 dma_resv_assert_held(bo->base.resv);
1188 bo->base.resv = &bo->base._resv;
1190 if (!ttm_bo_uses_embedded_gem_object(bo)) {
1192 * bo.gem is not initialized, so we have to setup the
1193 * struct elements we want use regardless.
1195 dma_resv_init(&bo->base._resv);
1196 drm_vma_node_reset(&bo->base.vma_node);
1198 atomic_inc(&ttm_bo_glob.bo_count);
1201 * For ttm_bo_type_device buffers, allocate
1202 * address space from the device.
1204 if (bo->type == ttm_bo_type_device ||
1205 bo->type == ttm_bo_type_sg)
1206 ret = drm_vma_offset_add(bdev->vma_manager, &bo->base.vma_node,
1209 /* passed reservation objects should already be locked,
1210 * since otherwise lockdep will be angered in radeon.
1213 locked = dma_resv_trylock(bo->base.resv);
1218 ret = ttm_bo_validate(bo, placement, ctx);
1220 if (unlikely(ret)) {
1222 ttm_bo_unreserve(bo);
1228 ttm_bo_move_to_lru_tail_unlocked(bo);
1232 EXPORT_SYMBOL(ttm_bo_init_reserved);
1234 int ttm_bo_init(struct ttm_bo_device *bdev,
1235 struct ttm_buffer_object *bo,
1237 enum ttm_bo_type type,
1238 struct ttm_placement *placement,
1239 uint32_t page_alignment,
1242 struct sg_table *sg,
1243 struct dma_resv *resv,
1244 void (*destroy) (struct ttm_buffer_object *))
1246 struct ttm_operation_ctx ctx = { interruptible, false };
1249 ret = ttm_bo_init_reserved(bdev, bo, size, type, placement,
1250 page_alignment, &ctx, acc_size,
1256 ttm_bo_unreserve(bo);
1260 EXPORT_SYMBOL(ttm_bo_init);
1262 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1263 unsigned long bo_size,
1264 unsigned struct_size)
1266 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1269 size += ttm_round_pot(struct_size);
1270 size += ttm_round_pot(npages * sizeof(void *));
1271 size += ttm_round_pot(sizeof(struct ttm_tt));
1274 EXPORT_SYMBOL(ttm_bo_acc_size);
1276 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1277 unsigned long bo_size,
1278 unsigned struct_size)
1280 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1283 size += ttm_round_pot(struct_size);
1284 size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t)));
1285 size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1288 EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1290 int ttm_bo_create(struct ttm_bo_device *bdev,
1292 enum ttm_bo_type type,
1293 struct ttm_placement *placement,
1294 uint32_t page_alignment,
1296 struct ttm_buffer_object **p_bo)
1298 struct ttm_buffer_object *bo;
1302 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1303 if (unlikely(bo == NULL))
1306 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1307 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1308 interruptible, acc_size,
1310 if (likely(ret == 0))
1315 EXPORT_SYMBOL(ttm_bo_create);
1317 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1319 struct ttm_resource_manager *man = ttm_manager_type(bdev, mem_type);
1321 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1322 pr_err("Illegal memory manager memory type %u\n", mem_type);
1327 pr_err("Memory type %u has not been initialized\n", mem_type);
1331 return ttm_resource_manager_force_list_clean(bdev, man);
1333 EXPORT_SYMBOL(ttm_bo_evict_mm);
1335 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1337 struct ttm_bo_global *glob =
1338 container_of(kobj, struct ttm_bo_global, kobj);
1340 __free_page(glob->dummy_read_page);
1343 static void ttm_bo_global_release(void)
1345 struct ttm_bo_global *glob = &ttm_bo_glob;
1347 mutex_lock(&ttm_global_mutex);
1348 if (--ttm_bo_glob_use_count > 0)
1351 kobject_del(&glob->kobj);
1352 kobject_put(&glob->kobj);
1353 ttm_mem_global_release(&ttm_mem_glob);
1354 memset(glob, 0, sizeof(*glob));
1356 mutex_unlock(&ttm_global_mutex);
1359 static int ttm_bo_global_init(void)
1361 struct ttm_bo_global *glob = &ttm_bo_glob;
1365 mutex_lock(&ttm_global_mutex);
1366 if (++ttm_bo_glob_use_count > 1)
1369 ret = ttm_mem_global_init(&ttm_mem_glob);
1373 spin_lock_init(&glob->lru_lock);
1374 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1376 if (unlikely(glob->dummy_read_page == NULL)) {
1381 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1382 INIT_LIST_HEAD(&glob->swap_lru[i]);
1383 INIT_LIST_HEAD(&glob->device_list);
1384 atomic_set(&glob->bo_count, 0);
1386 ret = kobject_init_and_add(
1387 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1388 if (unlikely(ret != 0))
1389 kobject_put(&glob->kobj);
1391 mutex_unlock(&ttm_global_mutex);
1395 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1397 struct ttm_bo_global *glob = &ttm_bo_glob;
1400 struct ttm_resource_manager *man;
1402 man = ttm_manager_type(bdev, TTM_PL_SYSTEM);
1403 ttm_resource_manager_set_used(man, false);
1404 ttm_set_driver_manager(bdev, TTM_PL_SYSTEM, NULL);
1406 mutex_lock(&ttm_global_mutex);
1407 list_del(&bdev->device_list);
1408 mutex_unlock(&ttm_global_mutex);
1410 cancel_delayed_work_sync(&bdev->wq);
1412 if (ttm_bo_delayed_delete(bdev, true))
1413 pr_debug("Delayed destroy list was clean\n");
1415 spin_lock(&glob->lru_lock);
1416 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1417 if (list_empty(&man->lru[0]))
1418 pr_debug("Swap list %d was clean\n", i);
1419 spin_unlock(&glob->lru_lock);
1422 ttm_bo_global_release();
1426 EXPORT_SYMBOL(ttm_bo_device_release);
1428 static void ttm_bo_init_sysman(struct ttm_bo_device *bdev)
1430 struct ttm_resource_manager *man = &bdev->sysman;
1433 * Initialize the system memory buffer type.
1434 * Other types need to be driver / IOCTL initialized.
1437 man->available_caching = TTM_PL_MASK_CACHING;
1438 man->default_caching = TTM_PL_FLAG_CACHED;
1440 ttm_resource_manager_init(man, 0);
1441 ttm_set_driver_manager(bdev, TTM_PL_SYSTEM, man);
1442 ttm_resource_manager_set_used(man, true);
1445 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1446 struct ttm_bo_driver *driver,
1447 struct address_space *mapping,
1448 struct drm_vma_offset_manager *vma_manager,
1451 struct ttm_bo_global *glob = &ttm_bo_glob;
1454 if (WARN_ON(vma_manager == NULL))
1457 ret = ttm_bo_global_init();
1461 bdev->driver = driver;
1463 ttm_bo_init_sysman(bdev);
1465 bdev->vma_manager = vma_manager;
1466 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1467 INIT_LIST_HEAD(&bdev->ddestroy);
1468 bdev->dev_mapping = mapping;
1469 bdev->need_dma32 = need_dma32;
1470 mutex_lock(&ttm_global_mutex);
1471 list_add_tail(&bdev->device_list, &glob->device_list);
1472 mutex_unlock(&ttm_global_mutex);
1476 EXPORT_SYMBOL(ttm_bo_device_init);
1479 * buffer object vm functions.
1482 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1484 struct ttm_bo_device *bdev = bo->bdev;
1486 drm_vma_node_unmap(&bo->base.vma_node, bdev->dev_mapping);
1487 ttm_mem_io_free(bdev, &bo->mem);
1489 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1491 int ttm_bo_wait(struct ttm_buffer_object *bo,
1492 bool interruptible, bool no_wait)
1494 long timeout = 15 * HZ;
1497 if (dma_resv_test_signaled_rcu(bo->base.resv, true))
1503 timeout = dma_resv_wait_timeout_rcu(bo->base.resv, true,
1504 interruptible, timeout);
1511 dma_resv_add_excl_fence(bo->base.resv, NULL);
1514 EXPORT_SYMBOL(ttm_bo_wait);
1517 * A buffer object shrink method that tries to swap out the first
1518 * buffer object on the bo_global::swap_lru list.
1520 int ttm_bo_swapout(struct ttm_bo_global *glob, struct ttm_operation_ctx *ctx)
1522 struct ttm_buffer_object *bo;
1527 spin_lock(&glob->lru_lock);
1528 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
1529 list_for_each_entry(bo, &glob->swap_lru[i], swap) {
1530 if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked,
1534 if (!ttm_bo_get_unless_zero(bo)) {
1536 dma_resv_unlock(bo->base.resv);
1548 spin_unlock(&glob->lru_lock);
1553 ret = ttm_bo_cleanup_refs(bo, false, false, locked);
1558 ttm_bo_del_from_lru(bo);
1559 spin_unlock(&glob->lru_lock);
1562 * Move to system cached
1565 if (bo->mem.mem_type != TTM_PL_SYSTEM ||
1566 bo->ttm->caching_state != tt_cached) {
1567 struct ttm_operation_ctx ctx = { false, false };
1568 struct ttm_resource evict_mem;
1570 evict_mem = bo->mem;
1571 evict_mem.mm_node = NULL;
1572 evict_mem.placement = TTM_PL_FLAG_CACHED;
1573 evict_mem.mem_type = TTM_PL_SYSTEM;
1575 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, &ctx);
1576 if (unlikely(ret != 0))
1581 * Make sure BO is idle.
1584 ret = ttm_bo_wait(bo, false, false);
1585 if (unlikely(ret != 0))
1588 ttm_bo_unmap_virtual(bo);
1591 * Swap out. Buffer will be swapped in again as soon as
1592 * anyone tries to access a ttm page.
1595 if (bo->bdev->driver->swap_notify)
1596 bo->bdev->driver->swap_notify(bo);
1598 ret = ttm_tt_swapout(bo->bdev, bo->ttm, bo->persistent_swap_storage);
1603 * Unreserve without putting on LRU to avoid swapping out an
1604 * already swapped buffer.
1607 dma_resv_unlock(bo->base.resv);
1611 EXPORT_SYMBOL(ttm_bo_swapout);
1613 void ttm_bo_swapout_all(void)
1615 struct ttm_operation_ctx ctx = {
1616 .interruptible = false,
1617 .no_wait_gpu = false
1620 while (ttm_bo_swapout(&ttm_bo_glob, &ctx) == 0);
1622 EXPORT_SYMBOL(ttm_bo_swapout_all);