2 * SPDX-License-Identifier: MIT
4 * Copyright © 2019 Intel Corporation
7 #include <linux/debugobjects.h>
9 #include "gt/intel_context.h"
10 #include "gt/intel_engine_heartbeat.h"
11 #include "gt/intel_engine_pm.h"
12 #include "gt/intel_ring.h"
15 #include "i915_active.h"
18 * Active refs memory management
20 * To be more economical with memory, we reap all the i915_active trees as
21 * they idle (when we know the active requests are inactive) and allocate the
22 * nodes from a local slab cache to hopefully reduce the fragmentation.
24 static struct kmem_cache *slab_cache;
28 struct i915_active_fence base;
29 struct i915_active *ref;
33 #define fetch_node(x) rb_entry(READ_ONCE(x), typeof(struct active_node), node)
35 static inline struct active_node *
36 node_from_active(struct i915_active_fence *active)
38 return container_of(active, struct active_node, base);
41 #define take_preallocated_barriers(x) llist_del_all(&(x)->preallocated_barriers)
43 static inline bool is_barrier(const struct i915_active_fence *active)
45 return IS_ERR(rcu_access_pointer(active->fence));
48 static inline struct llist_node *barrier_to_ll(struct active_node *node)
50 GEM_BUG_ON(!is_barrier(&node->base));
51 return (struct llist_node *)&node->base.cb.node;
54 static inline struct intel_engine_cs *
55 __barrier_to_engine(struct active_node *node)
57 return (struct intel_engine_cs *)READ_ONCE(node->base.cb.node.prev);
60 static inline struct intel_engine_cs *
61 barrier_to_engine(struct active_node *node)
63 GEM_BUG_ON(!is_barrier(&node->base));
64 return __barrier_to_engine(node);
67 static inline struct active_node *barrier_from_ll(struct llist_node *x)
69 return container_of((struct list_head *)x,
70 struct active_node, base.cb.node);
73 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM) && IS_ENABLED(CONFIG_DEBUG_OBJECTS)
75 static void *active_debug_hint(void *addr)
77 struct i915_active *ref = addr;
79 return (void *)ref->active ?: (void *)ref->retire ?: (void *)ref;
82 static const struct debug_obj_descr active_debug_desc = {
83 .name = "i915_active",
84 .debug_hint = active_debug_hint,
87 static void debug_active_init(struct i915_active *ref)
89 debug_object_init(ref, &active_debug_desc);
92 static void debug_active_activate(struct i915_active *ref)
94 lockdep_assert_held(&ref->tree_lock);
95 debug_object_activate(ref, &active_debug_desc);
98 static void debug_active_deactivate(struct i915_active *ref)
100 lockdep_assert_held(&ref->tree_lock);
101 if (!atomic_read(&ref->count)) /* after the last dec */
102 debug_object_deactivate(ref, &active_debug_desc);
105 static void debug_active_fini(struct i915_active *ref)
107 debug_object_free(ref, &active_debug_desc);
110 static void debug_active_assert(struct i915_active *ref)
112 debug_object_assert_init(ref, &active_debug_desc);
117 static inline void debug_active_init(struct i915_active *ref) { }
118 static inline void debug_active_activate(struct i915_active *ref) { }
119 static inline void debug_active_deactivate(struct i915_active *ref) { }
120 static inline void debug_active_fini(struct i915_active *ref) { }
121 static inline void debug_active_assert(struct i915_active *ref) { }
126 __active_retire(struct i915_active *ref)
128 struct rb_root root = RB_ROOT;
129 struct active_node *it, *n;
132 GEM_BUG_ON(i915_active_is_idle(ref));
134 /* return the unused nodes to our slabcache -- flushing the allocator */
135 if (!atomic_dec_and_lock_irqsave(&ref->count, &ref->tree_lock, flags))
138 GEM_BUG_ON(rcu_access_pointer(ref->excl.fence));
139 debug_active_deactivate(ref);
141 /* Even if we have not used the cache, we may still have a barrier */
143 ref->cache = fetch_node(ref->tree.rb_node);
145 /* Keep the MRU cached node for reuse */
147 /* Discard all other nodes in the tree */
148 rb_erase(&ref->cache->node, &ref->tree);
151 /* Rebuild the tree with only the cached node */
152 rb_link_node(&ref->cache->node, NULL, &ref->tree.rb_node);
153 rb_insert_color(&ref->cache->node, &ref->tree);
154 GEM_BUG_ON(ref->tree.rb_node != &ref->cache->node);
156 /* Make the cached node available for reuse with any timeline */
157 ref->cache->timeline = 0; /* needs cmpxchg(u64) */
160 spin_unlock_irqrestore(&ref->tree_lock, flags);
162 /* After the final retire, the entire struct may be freed */
166 /* ... except if you wait on it, you must manage your own references! */
169 /* Finally free the discarded timeline tree */
170 rbtree_postorder_for_each_entry_safe(it, n, &root, node) {
171 GEM_BUG_ON(i915_active_fence_isset(&it->base));
172 kmem_cache_free(slab_cache, it);
177 active_work(struct work_struct *wrk)
179 struct i915_active *ref = container_of(wrk, typeof(*ref), work);
181 GEM_BUG_ON(!atomic_read(&ref->count));
182 if (atomic_add_unless(&ref->count, -1, 1))
185 __active_retire(ref);
189 active_retire(struct i915_active *ref)
191 GEM_BUG_ON(!atomic_read(&ref->count));
192 if (atomic_add_unless(&ref->count, -1, 1))
195 if (ref->flags & I915_ACTIVE_RETIRE_SLEEPS) {
196 queue_work(system_unbound_wq, &ref->work);
200 __active_retire(ref);
203 static inline struct dma_fence **
204 __active_fence_slot(struct i915_active_fence *active)
206 return (struct dma_fence ** __force)&active->fence;
210 active_fence_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
212 struct i915_active_fence *active =
213 container_of(cb, typeof(*active), cb);
215 return cmpxchg(__active_fence_slot(active), fence, NULL) == fence;
219 node_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
221 if (active_fence_cb(fence, cb))
222 active_retire(container_of(cb, struct active_node, base.cb)->ref);
226 excl_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
228 if (active_fence_cb(fence, cb))
229 active_retire(container_of(cb, struct i915_active, excl.cb));
232 static struct active_node *__active_lookup(struct i915_active *ref, u64 idx)
234 struct active_node *it;
236 GEM_BUG_ON(idx == 0); /* 0 is the unordered timeline, rsvd for cache */
239 * We track the most recently used timeline to skip a rbtree search
240 * for the common case, under typical loads we never need the rbtree
241 * at all. We can reuse the last slot if it is empty, that is
242 * after the previous activity has been retired, or if it matches the
245 it = READ_ONCE(ref->cache);
247 u64 cached = READ_ONCE(it->timeline);
249 /* Once claimed, this slot will only belong to this idx */
254 * An unclaimed cache [.timeline=0] can only be claimed once.
256 * If the value is already non-zero, some other thread has
257 * claimed the cache and we know that is does not match our
258 * idx. If, and only if, the timeline is currently zero is it
259 * worth competing to claim it atomically for ourselves (for
260 * only the winner of that race will cmpxchg return the old
263 if (!cached && !cmpxchg64(&it->timeline, 0, idx))
267 BUILD_BUG_ON(offsetof(typeof(*it), node));
269 /* While active, the tree can only be built; not destroyed */
270 GEM_BUG_ON(i915_active_is_idle(ref));
272 it = fetch_node(ref->tree.rb_node);
274 if (it->timeline < idx) {
275 it = fetch_node(it->node.rb_right);
276 } else if (it->timeline > idx) {
277 it = fetch_node(it->node.rb_left);
279 WRITE_ONCE(ref->cache, it);
284 /* NB: If the tree rotated beneath us, we may miss our target. */
288 static struct i915_active_fence *
289 active_instance(struct i915_active *ref, u64 idx)
291 struct active_node *node;
292 struct rb_node **p, *parent;
294 node = __active_lookup(ref, idx);
298 spin_lock_irq(&ref->tree_lock);
299 GEM_BUG_ON(i915_active_is_idle(ref));
302 p = &ref->tree.rb_node;
306 node = rb_entry(parent, struct active_node, node);
307 if (node->timeline == idx)
310 if (node->timeline < idx)
311 p = &parent->rb_right;
313 p = &parent->rb_left;
317 * XXX: We should preallocate this before i915_active_ref() is ever
318 * called, but we cannot call into fs_reclaim() anyway, so use GFP_ATOMIC.
320 node = kmem_cache_alloc(slab_cache, GFP_ATOMIC);
324 __i915_active_fence_init(&node->base, NULL, node_retire);
326 node->timeline = idx;
328 rb_link_node(&node->node, parent, p);
329 rb_insert_color(&node->node, &ref->tree);
332 WRITE_ONCE(ref->cache, node);
333 spin_unlock_irq(&ref->tree_lock);
338 void __i915_active_init(struct i915_active *ref,
339 int (*active)(struct i915_active *ref),
340 void (*retire)(struct i915_active *ref),
342 struct lock_class_key *mkey,
343 struct lock_class_key *wkey)
345 debug_active_init(ref);
348 ref->active = active;
349 ref->retire = retire;
351 spin_lock_init(&ref->tree_lock);
355 init_llist_head(&ref->preallocated_barriers);
356 atomic_set(&ref->count, 0);
357 __mutex_init(&ref->mutex, "i915_active", mkey);
358 __i915_active_fence_init(&ref->excl, NULL, excl_retire);
359 INIT_WORK(&ref->work, active_work);
360 #if IS_ENABLED(CONFIG_LOCKDEP)
361 lockdep_init_map(&ref->work.lockdep_map, "i915_active.work", wkey, 0);
365 static bool ____active_del_barrier(struct i915_active *ref,
366 struct active_node *node,
367 struct intel_engine_cs *engine)
370 struct llist_node *head = NULL, *tail = NULL;
371 struct llist_node *pos, *next;
373 GEM_BUG_ON(node->timeline != engine->kernel_context->timeline->fence_context);
376 * Rebuild the llist excluding our node. We may perform this
377 * outside of the kernel_context timeline mutex and so someone
378 * else may be manipulating the engine->barrier_tasks, in
379 * which case either we or they will be upset :)
381 * A second __active_del_barrier() will report failure to claim
382 * the active_node and the caller will just shrug and know not to
383 * claim ownership of its node.
385 * A concurrent i915_request_add_active_barriers() will miss adding
386 * any of the tasks, but we will try again on the next -- and since
387 * we are actively using the barrier, we know that there will be
388 * at least another opportunity when we idle.
390 llist_for_each_safe(pos, next, llist_del_all(&engine->barrier_tasks)) {
391 if (node == barrier_from_ll(pos)) {
402 llist_add_batch(head, tail, &engine->barrier_tasks);
408 __active_del_barrier(struct i915_active *ref, struct active_node *node)
410 return ____active_del_barrier(ref, node, barrier_to_engine(node));
414 replace_barrier(struct i915_active *ref, struct i915_active_fence *active)
416 if (!is_barrier(active)) /* proto-node used by our idle barrier? */
420 * This request is on the kernel_context timeline, and so
421 * we can use it to substitute for the pending idle-barrer
422 * request that we want to emit on the kernel_context.
424 return __active_del_barrier(ref, node_from_active(active));
427 int i915_active_add_request(struct i915_active *ref, struct i915_request *rq)
429 u64 idx = i915_request_timeline(rq)->fence_context;
430 struct dma_fence *fence = &rq->fence;
431 struct i915_active_fence *active;
434 /* Prevent reaping in case we malloc/wait while building the tree */
435 err = i915_active_acquire(ref);
440 active = active_instance(ref, idx);
446 if (replace_barrier(ref, active)) {
447 RCU_INIT_POINTER(active->fence, NULL);
448 atomic_dec(&ref->count);
450 } while (unlikely(is_barrier(active)));
452 fence = __i915_active_fence_set(active, fence);
454 __i915_active_acquire(ref);
456 dma_fence_put(fence);
459 i915_active_release(ref);
463 static struct dma_fence *
464 __i915_active_set_fence(struct i915_active *ref,
465 struct i915_active_fence *active,
466 struct dma_fence *fence)
468 struct dma_fence *prev;
470 if (replace_barrier(ref, active)) {
471 RCU_INIT_POINTER(active->fence, fence);
475 prev = __i915_active_fence_set(active, fence);
477 __i915_active_acquire(ref);
483 i915_active_set_exclusive(struct i915_active *ref, struct dma_fence *f)
485 /* We expect the caller to manage the exclusive timeline ordering */
486 return __i915_active_set_fence(ref, &ref->excl, f);
489 bool i915_active_acquire_if_busy(struct i915_active *ref)
491 debug_active_assert(ref);
492 return atomic_add_unless(&ref->count, 1, 0);
495 static void __i915_active_activate(struct i915_active *ref)
497 spin_lock_irq(&ref->tree_lock); /* __active_retire() */
498 if (!atomic_fetch_inc(&ref->count))
499 debug_active_activate(ref);
500 spin_unlock_irq(&ref->tree_lock);
503 int i915_active_acquire(struct i915_active *ref)
507 if (i915_active_acquire_if_busy(ref))
511 __i915_active_activate(ref);
515 err = mutex_lock_interruptible(&ref->mutex);
519 if (likely(!i915_active_acquire_if_busy(ref))) {
520 err = ref->active(ref);
522 __i915_active_activate(ref);
525 mutex_unlock(&ref->mutex);
530 int i915_active_acquire_for_context(struct i915_active *ref, u64 idx)
532 struct i915_active_fence *active;
535 err = i915_active_acquire(ref);
539 active = active_instance(ref, idx);
541 i915_active_release(ref);
545 return 0; /* return with active ref */
548 void i915_active_release(struct i915_active *ref)
550 debug_active_assert(ref);
554 static void enable_signaling(struct i915_active_fence *active)
556 struct dma_fence *fence;
558 if (unlikely(is_barrier(active)))
561 fence = i915_active_fence_get(active);
565 dma_fence_enable_sw_signaling(fence);
566 dma_fence_put(fence);
569 static int flush_barrier(struct active_node *it)
571 struct intel_engine_cs *engine;
573 if (likely(!is_barrier(&it->base)))
576 engine = __barrier_to_engine(it);
577 smp_rmb(); /* serialise with add_active_barriers */
578 if (!is_barrier(&it->base))
581 return intel_engine_flush_barriers(engine);
584 static int flush_lazy_signals(struct i915_active *ref)
586 struct active_node *it, *n;
589 enable_signaling(&ref->excl);
590 rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
591 err = flush_barrier(it); /* unconnected idle barrier? */
595 enable_signaling(&it->base);
601 int __i915_active_wait(struct i915_active *ref, int state)
605 /* Any fence added after the wait begins will not be auto-signaled */
606 if (i915_active_acquire_if_busy(ref)) {
609 err = flush_lazy_signals(ref);
610 i915_active_release(ref);
614 if (___wait_var_event(ref, i915_active_is_idle(ref),
615 state, 0, 0, schedule()))
620 * After the wait is complete, the caller may free the active.
621 * We have to flush any concurrent retirement before returning.
623 flush_work(&ref->work);
627 static int __await_active(struct i915_active_fence *active,
628 int (*fn)(void *arg, struct dma_fence *fence),
631 struct dma_fence *fence;
633 if (is_barrier(active)) /* XXX flush the barrier? */
636 fence = i915_active_fence_get(active);
640 err = fn(arg, fence);
641 dma_fence_put(fence);
649 struct wait_barrier {
650 struct wait_queue_entry base;
651 struct i915_active *ref;
655 barrier_wake(wait_queue_entry_t *wq, unsigned int mode, int flags, void *key)
657 struct wait_barrier *wb = container_of(wq, typeof(*wb), base);
659 if (i915_active_is_idle(wb->ref)) {
660 list_del(&wq->entry);
661 i915_sw_fence_complete(wq->private);
668 static int __await_barrier(struct i915_active *ref, struct i915_sw_fence *fence)
670 struct wait_barrier *wb;
672 wb = kmalloc(sizeof(*wb), GFP_KERNEL);
676 GEM_BUG_ON(i915_active_is_idle(ref));
677 if (!i915_sw_fence_await(fence)) {
683 wb->base.func = barrier_wake;
684 wb->base.private = fence;
687 add_wait_queue(__var_waitqueue(ref), &wb->base);
691 static int await_active(struct i915_active *ref,
693 int (*fn)(void *arg, struct dma_fence *fence),
694 void *arg, struct i915_sw_fence *barrier)
698 if (!i915_active_acquire_if_busy(ref))
701 if (flags & I915_ACTIVE_AWAIT_EXCL &&
702 rcu_access_pointer(ref->excl.fence)) {
703 err = __await_active(&ref->excl, fn, arg);
708 if (flags & I915_ACTIVE_AWAIT_ACTIVE) {
709 struct active_node *it, *n;
711 rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
712 err = __await_active(&it->base, fn, arg);
718 if (flags & I915_ACTIVE_AWAIT_BARRIER) {
719 err = flush_lazy_signals(ref);
723 err = __await_barrier(ref, barrier);
729 i915_active_release(ref);
733 static int rq_await_fence(void *arg, struct dma_fence *fence)
735 return i915_request_await_dma_fence(arg, fence);
738 int i915_request_await_active(struct i915_request *rq,
739 struct i915_active *ref,
742 return await_active(ref, flags, rq_await_fence, rq, &rq->submit);
745 static int sw_await_fence(void *arg, struct dma_fence *fence)
747 return i915_sw_fence_await_dma_fence(arg, fence, 0,
748 GFP_NOWAIT | __GFP_NOWARN);
751 int i915_sw_fence_await_active(struct i915_sw_fence *fence,
752 struct i915_active *ref,
755 return await_active(ref, flags, sw_await_fence, fence, fence);
758 void i915_active_fini(struct i915_active *ref)
760 debug_active_fini(ref);
761 GEM_BUG_ON(atomic_read(&ref->count));
762 GEM_BUG_ON(work_pending(&ref->work));
763 mutex_destroy(&ref->mutex);
766 kmem_cache_free(slab_cache, ref->cache);
769 static inline bool is_idle_barrier(struct active_node *node, u64 idx)
771 return node->timeline == idx && !i915_active_fence_isset(&node->base);
774 static struct active_node *reuse_idle_barrier(struct i915_active *ref, u64 idx)
776 struct rb_node *prev, *p;
778 if (RB_EMPTY_ROOT(&ref->tree))
781 GEM_BUG_ON(i915_active_is_idle(ref));
784 * Try to reuse any existing barrier nodes already allocated for this
785 * i915_active, due to overlapping active phases there is likely a
786 * node kept alive (as we reuse before parking). We prefer to reuse
787 * completely idle barriers (less hassle in manipulating the llists),
788 * but otherwise any will do.
790 if (ref->cache && is_idle_barrier(ref->cache, idx)) {
791 p = &ref->cache->node;
796 p = ref->tree.rb_node;
798 struct active_node *node =
799 rb_entry(p, struct active_node, node);
801 if (is_idle_barrier(node, idx))
805 if (node->timeline < idx)
806 p = READ_ONCE(p->rb_right);
808 p = READ_ONCE(p->rb_left);
812 * No quick match, but we did find the leftmost rb_node for the
813 * kernel_context. Walk the rb_tree in-order to see if there were
814 * any idle-barriers on this timeline that we missed, or just use
815 * the first pending barrier.
817 for (p = prev; p; p = rb_next(p)) {
818 struct active_node *node =
819 rb_entry(p, struct active_node, node);
820 struct intel_engine_cs *engine;
822 if (node->timeline > idx)
825 if (node->timeline < idx)
828 if (is_idle_barrier(node, idx))
832 * The list of pending barriers is protected by the
833 * kernel_context timeline, which notably we do not hold
834 * here. i915_request_add_active_barriers() may consume
835 * the barrier before we claim it, so we have to check
838 engine = __barrier_to_engine(node);
839 smp_rmb(); /* serialise with add_active_barriers */
840 if (is_barrier(&node->base) &&
841 ____active_del_barrier(ref, node, engine))
848 spin_lock_irq(&ref->tree_lock);
849 rb_erase(p, &ref->tree); /* Hide from waits and sibling allocations */
850 if (p == &ref->cache->node)
851 WRITE_ONCE(ref->cache, NULL);
852 spin_unlock_irq(&ref->tree_lock);
854 return rb_entry(p, struct active_node, node);
857 int i915_active_acquire_preallocate_barrier(struct i915_active *ref,
858 struct intel_engine_cs *engine)
860 intel_engine_mask_t tmp, mask = engine->mask;
861 struct llist_node *first = NULL, *last = NULL;
862 struct intel_gt *gt = engine->gt;
864 GEM_BUG_ON(i915_active_is_idle(ref));
866 /* Wait until the previous preallocation is completed */
867 while (!llist_empty(&ref->preallocated_barriers))
871 * Preallocate a node for each physical engine supporting the target
872 * engine (remember virtual engines have more than one sibling).
873 * We can then use the preallocated nodes in
874 * i915_active_acquire_barrier()
877 for_each_engine_masked(engine, gt, mask, tmp) {
878 u64 idx = engine->kernel_context->timeline->fence_context;
879 struct llist_node *prev = first;
880 struct active_node *node;
883 node = reuse_idle_barrier(ref, idx);
886 node = kmem_cache_alloc(slab_cache, GFP_KERNEL);
890 RCU_INIT_POINTER(node->base.fence, NULL);
891 node->base.cb.func = node_retire;
892 node->timeline = idx;
896 if (!i915_active_fence_isset(&node->base)) {
898 * Mark this as being *our* unconnected proto-node.
900 * Since this node is not in any list, and we have
901 * decoupled it from the rbtree, we can reuse the
902 * request to indicate this is an idle-barrier node
903 * and then we can use the rb_node and list pointers
904 * for our tracking of the pending barrier.
906 RCU_INIT_POINTER(node->base.fence, ERR_PTR(-EAGAIN));
907 node->base.cb.node.prev = (void *)engine;
908 __i915_active_acquire(ref);
910 GEM_BUG_ON(rcu_access_pointer(node->base.fence) != ERR_PTR(-EAGAIN));
912 GEM_BUG_ON(barrier_to_engine(node) != engine);
913 first = barrier_to_ll(node);
917 intel_engine_pm_get(engine);
920 GEM_BUG_ON(!llist_empty(&ref->preallocated_barriers));
921 llist_add_batch(first, last, &ref->preallocated_barriers);
927 struct active_node *node = barrier_from_ll(first);
931 atomic_dec(&ref->count);
932 intel_engine_pm_put(barrier_to_engine(node));
934 kmem_cache_free(slab_cache, node);
939 void i915_active_acquire_barrier(struct i915_active *ref)
941 struct llist_node *pos, *next;
944 GEM_BUG_ON(i915_active_is_idle(ref));
947 * Transfer the list of preallocated barriers into the
948 * i915_active rbtree, but only as proto-nodes. They will be
949 * populated by i915_request_add_active_barriers() to point to the
950 * request that will eventually release them.
952 llist_for_each_safe(pos, next, take_preallocated_barriers(ref)) {
953 struct active_node *node = barrier_from_ll(pos);
954 struct intel_engine_cs *engine = barrier_to_engine(node);
955 struct rb_node **p, *parent;
957 spin_lock_irqsave_nested(&ref->tree_lock, flags,
958 SINGLE_DEPTH_NESTING);
960 p = &ref->tree.rb_node;
962 struct active_node *it;
966 it = rb_entry(parent, struct active_node, node);
967 if (it->timeline < node->timeline)
968 p = &parent->rb_right;
970 p = &parent->rb_left;
972 rb_link_node(&node->node, parent, p);
973 rb_insert_color(&node->node, &ref->tree);
974 spin_unlock_irqrestore(&ref->tree_lock, flags);
976 GEM_BUG_ON(!intel_engine_pm_is_awake(engine));
977 llist_add(barrier_to_ll(node), &engine->barrier_tasks);
978 intel_engine_pm_put_delay(engine, 2);
982 static struct dma_fence **ll_to_fence_slot(struct llist_node *node)
984 return __active_fence_slot(&barrier_from_ll(node)->base);
987 void i915_request_add_active_barriers(struct i915_request *rq)
989 struct intel_engine_cs *engine = rq->engine;
990 struct llist_node *node, *next;
993 GEM_BUG_ON(!intel_context_is_barrier(rq->context));
994 GEM_BUG_ON(intel_engine_is_virtual(engine));
995 GEM_BUG_ON(i915_request_timeline(rq) != engine->kernel_context->timeline);
997 node = llist_del_all(&engine->barrier_tasks);
1001 * Attach the list of proto-fences to the in-flight request such
1002 * that the parent i915_active will be released when this request
1005 spin_lock_irqsave(&rq->lock, flags);
1006 llist_for_each_safe(node, next, node) {
1007 /* serialise with reuse_idle_barrier */
1008 smp_store_mb(*ll_to_fence_slot(node), &rq->fence);
1009 list_add_tail((struct list_head *)node, &rq->fence.cb_list);
1011 spin_unlock_irqrestore(&rq->lock, flags);
1015 * __i915_active_fence_set: Update the last active fence along its timeline
1016 * @active: the active tracker
1017 * @fence: the new fence (under construction)
1019 * Records the new @fence as the last active fence along its timeline in
1020 * this active tracker, moving the tracking callbacks from the previous
1021 * fence onto this one. Gets and returns a reference to the previous fence
1022 * (if not already completed), which the caller must put after making sure
1023 * that it is executed before the new fence. To ensure that the order of
1024 * fences within the timeline of the i915_active_fence is understood, it
1025 * should be locked by the caller.
1028 __i915_active_fence_set(struct i915_active_fence *active,
1029 struct dma_fence *fence)
1031 struct dma_fence *prev;
1032 unsigned long flags;
1035 * In case of fences embedded in i915_requests, their memory is
1036 * SLAB_FAILSAFE_BY_RCU, then it can be reused right after release
1037 * by new requests. Then, there is a risk of passing back a pointer
1038 * to a new, completely unrelated fence that reuses the same memory
1039 * while tracked under a different active tracker. Combined with i915
1040 * perf open/close operations that build await dependencies between
1041 * engine kernel context requests and user requests from different
1042 * timelines, this can lead to dependency loops and infinite waits.
1044 * As a countermeasure, we try to get a reference to the active->fence
1045 * first, so if we succeed and pass it back to our user then it is not
1046 * released and potentially reused by an unrelated request before the
1047 * user has a chance to set up an await dependency on it.
1049 prev = i915_active_fence_get(active);
1053 GEM_BUG_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags));
1056 * Consider that we have two threads arriving (A and B), with
1057 * C already resident as the active->fence.
1059 * Both A and B have got a reference to C or NULL, depending on the
1060 * timing of the interrupt handler. Let's assume that if A has got C
1061 * then it has locked C first (before B).
1063 * Note the strong ordering of the timeline also provides consistent
1064 * nesting rules for the fence->lock; the inner lock is always the
1067 spin_lock_irqsave(fence->lock, flags);
1069 spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING);
1072 * A does the cmpxchg first, and so it sees C or NULL, as before, or
1073 * something else, depending on the timing of other threads and/or
1074 * interrupt handler. If not the same as before then A unlocks C if
1075 * applicable and retries, starting from an attempt to get a new
1076 * active->fence. Meanwhile, B follows the same path as A.
1077 * Once A succeeds with cmpxch, B fails again, retires, gets A from
1078 * active->fence, locks it as soon as A completes, and possibly
1079 * succeeds with cmpxchg.
1081 while (cmpxchg(__active_fence_slot(active), prev, fence) != prev) {
1083 spin_unlock(prev->lock);
1084 dma_fence_put(prev);
1086 spin_unlock_irqrestore(fence->lock, flags);
1088 prev = i915_active_fence_get(active);
1089 GEM_BUG_ON(prev == fence);
1091 spin_lock_irqsave(fence->lock, flags);
1093 spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING);
1097 * If prev is NULL then the previous fence must have been signaled
1098 * and we know that we are first on the timeline. If it is still
1099 * present then, having the lock on that fence already acquired, we
1100 * serialise with the interrupt handler, in the process of removing it
1101 * from any future interrupt callback. A will then wait on C before
1102 * executing (if present).
1104 * As B is second, it sees A as the previous fence and so waits for
1105 * it to complete its transition and takes over the occupancy for
1106 * itself -- remembering that it needs to wait on A before executing.
1109 __list_del_entry(&active->cb.node);
1110 spin_unlock(prev->lock); /* serialise with prev->cb_list */
1112 list_add_tail(&active->cb.node, &fence->cb_list);
1113 spin_unlock_irqrestore(fence->lock, flags);
1118 int i915_active_fence_set(struct i915_active_fence *active,
1119 struct i915_request *rq)
1121 struct dma_fence *fence;
1124 /* Must maintain timeline ordering wrt previous active requests */
1125 fence = __i915_active_fence_set(active, &rq->fence);
1127 err = i915_request_await_dma_fence(rq, fence);
1128 dma_fence_put(fence);
1134 void i915_active_noop(struct dma_fence *fence, struct dma_fence_cb *cb)
1136 active_fence_cb(fence, cb);
1139 struct auto_active {
1140 struct i915_active base;
1144 struct i915_active *i915_active_get(struct i915_active *ref)
1146 struct auto_active *aa = container_of(ref, typeof(*aa), base);
1152 static void auto_release(struct kref *ref)
1154 struct auto_active *aa = container_of(ref, typeof(*aa), ref);
1156 i915_active_fini(&aa->base);
1160 void i915_active_put(struct i915_active *ref)
1162 struct auto_active *aa = container_of(ref, typeof(*aa), base);
1164 kref_put(&aa->ref, auto_release);
1167 static int auto_active(struct i915_active *ref)
1169 i915_active_get(ref);
1173 static void auto_retire(struct i915_active *ref)
1175 i915_active_put(ref);
1178 struct i915_active *i915_active_create(void)
1180 struct auto_active *aa;
1182 aa = kmalloc(sizeof(*aa), GFP_KERNEL);
1186 kref_init(&aa->ref);
1187 i915_active_init(&aa->base, auto_active, auto_retire, 0);
1192 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1193 #include "selftests/i915_active.c"
1196 void i915_active_module_exit(void)
1198 kmem_cache_destroy(slab_cache);
1201 int __init i915_active_module_init(void)
1203 slab_cache = KMEM_CACHE(active_node, SLAB_HWCACHE_ALIGN);