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"
16 #include "i915_globals.h"
19 * Active refs memory management
21 * To be more economical with memory, we reap all the i915_active trees as
22 * they idle (when we know the active requests are inactive) and allocate the
23 * nodes from a local slab cache to hopefully reduce the fragmentation.
25 static struct i915_global_active {
26 struct i915_global base;
27 struct kmem_cache *slab_cache;
32 struct i915_active_fence base;
33 struct i915_active *ref;
37 #define fetch_node(x) rb_entry(READ_ONCE(x), typeof(struct active_node), node)
39 static inline struct active_node *
40 node_from_active(struct i915_active_fence *active)
42 return container_of(active, struct active_node, base);
45 #define take_preallocated_barriers(x) llist_del_all(&(x)->preallocated_barriers)
47 static inline bool is_barrier(const struct i915_active_fence *active)
49 return IS_ERR(rcu_access_pointer(active->fence));
52 static inline struct llist_node *barrier_to_ll(struct active_node *node)
54 GEM_BUG_ON(!is_barrier(&node->base));
55 return (struct llist_node *)&node->base.cb.node;
58 static inline struct intel_engine_cs *
59 __barrier_to_engine(struct active_node *node)
61 return (struct intel_engine_cs *)READ_ONCE(node->base.cb.node.prev);
64 static inline struct intel_engine_cs *
65 barrier_to_engine(struct active_node *node)
67 GEM_BUG_ON(!is_barrier(&node->base));
68 return __barrier_to_engine(node);
71 static inline struct active_node *barrier_from_ll(struct llist_node *x)
73 return container_of((struct list_head *)x,
74 struct active_node, base.cb.node);
77 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM) && IS_ENABLED(CONFIG_DEBUG_OBJECTS)
79 static void *active_debug_hint(void *addr)
81 struct i915_active *ref = addr;
83 return (void *)ref->active ?: (void *)ref->retire ?: (void *)ref;
86 static const struct debug_obj_descr active_debug_desc = {
87 .name = "i915_active",
88 .debug_hint = active_debug_hint,
91 static void debug_active_init(struct i915_active *ref)
93 debug_object_init(ref, &active_debug_desc);
96 static void debug_active_activate(struct i915_active *ref)
98 lockdep_assert_held(&ref->tree_lock);
99 debug_object_activate(ref, &active_debug_desc);
102 static void debug_active_deactivate(struct i915_active *ref)
104 lockdep_assert_held(&ref->tree_lock);
105 if (!atomic_read(&ref->count)) /* after the last dec */
106 debug_object_deactivate(ref, &active_debug_desc);
109 static void debug_active_fini(struct i915_active *ref)
111 debug_object_free(ref, &active_debug_desc);
114 static void debug_active_assert(struct i915_active *ref)
116 debug_object_assert_init(ref, &active_debug_desc);
121 static inline void debug_active_init(struct i915_active *ref) { }
122 static inline void debug_active_activate(struct i915_active *ref) { }
123 static inline void debug_active_deactivate(struct i915_active *ref) { }
124 static inline void debug_active_fini(struct i915_active *ref) { }
125 static inline void debug_active_assert(struct i915_active *ref) { }
130 __active_retire(struct i915_active *ref)
132 struct rb_root root = RB_ROOT;
133 struct active_node *it, *n;
136 GEM_BUG_ON(i915_active_is_idle(ref));
138 /* return the unused nodes to our slabcache -- flushing the allocator */
139 if (!atomic_dec_and_lock_irqsave(&ref->count, &ref->tree_lock, flags))
142 GEM_BUG_ON(rcu_access_pointer(ref->excl.fence));
143 debug_active_deactivate(ref);
145 /* Even if we have not used the cache, we may still have a barrier */
147 ref->cache = fetch_node(ref->tree.rb_node);
149 /* Keep the MRU cached node for reuse */
151 /* Discard all other nodes in the tree */
152 rb_erase(&ref->cache->node, &ref->tree);
155 /* Rebuild the tree with only the cached node */
156 rb_link_node(&ref->cache->node, NULL, &ref->tree.rb_node);
157 rb_insert_color(&ref->cache->node, &ref->tree);
158 GEM_BUG_ON(ref->tree.rb_node != &ref->cache->node);
160 /* Make the cached node available for reuse with any timeline */
161 if (IS_ENABLED(CONFIG_64BIT))
162 ref->cache->timeline = 0; /* needs cmpxchg(u64) */
165 spin_unlock_irqrestore(&ref->tree_lock, flags);
167 /* After the final retire, the entire struct may be freed */
171 /* ... except if you wait on it, you must manage your own references! */
174 /* Finally free the discarded timeline tree */
175 rbtree_postorder_for_each_entry_safe(it, n, &root, node) {
176 GEM_BUG_ON(i915_active_fence_isset(&it->base));
177 kmem_cache_free(global.slab_cache, it);
182 active_work(struct work_struct *wrk)
184 struct i915_active *ref = container_of(wrk, typeof(*ref), work);
186 GEM_BUG_ON(!atomic_read(&ref->count));
187 if (atomic_add_unless(&ref->count, -1, 1))
190 __active_retire(ref);
194 active_retire(struct i915_active *ref)
196 GEM_BUG_ON(!atomic_read(&ref->count));
197 if (atomic_add_unless(&ref->count, -1, 1))
200 if (ref->flags & I915_ACTIVE_RETIRE_SLEEPS) {
201 queue_work(system_unbound_wq, &ref->work);
205 __active_retire(ref);
208 static inline struct dma_fence **
209 __active_fence_slot(struct i915_active_fence *active)
211 return (struct dma_fence ** __force)&active->fence;
215 active_fence_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
217 struct i915_active_fence *active =
218 container_of(cb, typeof(*active), cb);
220 return cmpxchg(__active_fence_slot(active), fence, NULL) == fence;
224 node_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
226 if (active_fence_cb(fence, cb))
227 active_retire(container_of(cb, struct active_node, base.cb)->ref);
231 excl_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
233 if (active_fence_cb(fence, cb))
234 active_retire(container_of(cb, struct i915_active, excl.cb));
237 static struct active_node *__active_lookup(struct i915_active *ref, u64 idx)
239 struct active_node *it;
241 GEM_BUG_ON(idx == 0); /* 0 is the unordered timeline, rsvd for cache */
244 * We track the most recently used timeline to skip a rbtree search
245 * for the common case, under typical loads we never need the rbtree
246 * at all. We can reuse the last slot if it is empty, that is
247 * after the previous activity has been retired, or if it matches the
250 it = READ_ONCE(ref->cache);
252 u64 cached = READ_ONCE(it->timeline);
254 /* Once claimed, this slot will only belong to this idx */
258 #ifdef CONFIG_64BIT /* for cmpxchg(u64) */
260 * An unclaimed cache [.timeline=0] can only be claimed once.
262 * If the value is already non-zero, some other thread has
263 * claimed the cache and we know that is does not match our
264 * idx. If, and only if, the timeline is currently zero is it
265 * worth competing to claim it atomically for ourselves (for
266 * only the winner of that race will cmpxchg return the old
269 if (!cached && !cmpxchg(&it->timeline, 0, idx))
274 BUILD_BUG_ON(offsetof(typeof(*it), node));
276 /* While active, the tree can only be built; not destroyed */
277 GEM_BUG_ON(i915_active_is_idle(ref));
279 it = fetch_node(ref->tree.rb_node);
281 if (it->timeline < idx) {
282 it = fetch_node(it->node.rb_right);
283 } else if (it->timeline > idx) {
284 it = fetch_node(it->node.rb_left);
286 WRITE_ONCE(ref->cache, it);
291 /* NB: If the tree rotated beneath us, we may miss our target. */
295 static struct i915_active_fence *
296 active_instance(struct i915_active *ref, u64 idx)
298 struct active_node *node, *prealloc;
299 struct rb_node **p, *parent;
301 node = __active_lookup(ref, idx);
305 /* Preallocate a replacement, just in case */
306 prealloc = kmem_cache_alloc(global.slab_cache, GFP_KERNEL);
310 spin_lock_irq(&ref->tree_lock);
311 GEM_BUG_ON(i915_active_is_idle(ref));
314 p = &ref->tree.rb_node;
318 node = rb_entry(parent, struct active_node, node);
319 if (node->timeline == idx) {
320 kmem_cache_free(global.slab_cache, prealloc);
324 if (node->timeline < idx)
325 p = &parent->rb_right;
327 p = &parent->rb_left;
331 __i915_active_fence_init(&node->base, NULL, node_retire);
333 node->timeline = idx;
335 rb_link_node(&node->node, parent, p);
336 rb_insert_color(&node->node, &ref->tree);
339 WRITE_ONCE(ref->cache, node);
340 spin_unlock_irq(&ref->tree_lock);
345 void __i915_active_init(struct i915_active *ref,
346 int (*active)(struct i915_active *ref),
347 void (*retire)(struct i915_active *ref),
348 struct lock_class_key *mkey,
349 struct lock_class_key *wkey)
353 debug_active_init(ref);
356 ref->active = active;
357 ref->retire = ptr_unpack_bits(retire, &bits, 2);
358 if (bits & I915_ACTIVE_MAY_SLEEP)
359 ref->flags |= I915_ACTIVE_RETIRE_SLEEPS;
361 spin_lock_init(&ref->tree_lock);
365 init_llist_head(&ref->preallocated_barriers);
366 atomic_set(&ref->count, 0);
367 __mutex_init(&ref->mutex, "i915_active", mkey);
368 __i915_active_fence_init(&ref->excl, NULL, excl_retire);
369 INIT_WORK(&ref->work, active_work);
370 #if IS_ENABLED(CONFIG_LOCKDEP)
371 lockdep_init_map(&ref->work.lockdep_map, "i915_active.work", wkey, 0);
375 static bool ____active_del_barrier(struct i915_active *ref,
376 struct active_node *node,
377 struct intel_engine_cs *engine)
380 struct llist_node *head = NULL, *tail = NULL;
381 struct llist_node *pos, *next;
383 GEM_BUG_ON(node->timeline != engine->kernel_context->timeline->fence_context);
386 * Rebuild the llist excluding our node. We may perform this
387 * outside of the kernel_context timeline mutex and so someone
388 * else may be manipulating the engine->barrier_tasks, in
389 * which case either we or they will be upset :)
391 * A second __active_del_barrier() will report failure to claim
392 * the active_node and the caller will just shrug and know not to
393 * claim ownership of its node.
395 * A concurrent i915_request_add_active_barriers() will miss adding
396 * any of the tasks, but we will try again on the next -- and since
397 * we are actively using the barrier, we know that there will be
398 * at least another opportunity when we idle.
400 llist_for_each_safe(pos, next, llist_del_all(&engine->barrier_tasks)) {
401 if (node == barrier_from_ll(pos)) {
412 llist_add_batch(head, tail, &engine->barrier_tasks);
418 __active_del_barrier(struct i915_active *ref, struct active_node *node)
420 return ____active_del_barrier(ref, node, barrier_to_engine(node));
424 replace_barrier(struct i915_active *ref, struct i915_active_fence *active)
426 if (!is_barrier(active)) /* proto-node used by our idle barrier? */
430 * This request is on the kernel_context timeline, and so
431 * we can use it to substitute for the pending idle-barrer
432 * request that we want to emit on the kernel_context.
434 return __active_del_barrier(ref, node_from_active(active));
437 int i915_active_ref(struct i915_active *ref, u64 idx, struct dma_fence *fence)
439 struct i915_active_fence *active;
442 /* Prevent reaping in case we malloc/wait while building the tree */
443 err = i915_active_acquire(ref);
448 active = active_instance(ref, idx);
454 if (replace_barrier(ref, active)) {
455 RCU_INIT_POINTER(active->fence, NULL);
456 atomic_dec(&ref->count);
458 } while (unlikely(is_barrier(active)));
460 fence = __i915_active_fence_set(active, fence);
462 __i915_active_acquire(ref);
464 dma_fence_put(fence);
467 i915_active_release(ref);
471 static struct dma_fence *
472 __i915_active_set_fence(struct i915_active *ref,
473 struct i915_active_fence *active,
474 struct dma_fence *fence)
476 struct dma_fence *prev;
478 if (replace_barrier(ref, active)) {
479 RCU_INIT_POINTER(active->fence, fence);
483 prev = __i915_active_fence_set(active, fence);
485 __i915_active_acquire(ref);
490 static struct i915_active_fence *
491 __active_fence(struct i915_active *ref, u64 idx)
493 struct active_node *it;
495 it = __active_lookup(ref, idx);
496 if (unlikely(!it)) { /* Contention with parallel tree builders! */
497 spin_lock_irq(&ref->tree_lock);
498 it = __active_lookup(ref, idx);
499 spin_unlock_irq(&ref->tree_lock);
501 GEM_BUG_ON(!it); /* slot must be preallocated */
507 __i915_active_ref(struct i915_active *ref, u64 idx, struct dma_fence *fence)
509 /* Only valid while active, see i915_active_acquire_for_context() */
510 return __i915_active_set_fence(ref, __active_fence(ref, idx), fence);
514 i915_active_set_exclusive(struct i915_active *ref, struct dma_fence *f)
516 /* We expect the caller to manage the exclusive timeline ordering */
517 return __i915_active_set_fence(ref, &ref->excl, f);
520 bool i915_active_acquire_if_busy(struct i915_active *ref)
522 debug_active_assert(ref);
523 return atomic_add_unless(&ref->count, 1, 0);
526 static void __i915_active_activate(struct i915_active *ref)
528 spin_lock_irq(&ref->tree_lock); /* __active_retire() */
529 if (!atomic_fetch_inc(&ref->count))
530 debug_active_activate(ref);
531 spin_unlock_irq(&ref->tree_lock);
534 int i915_active_acquire(struct i915_active *ref)
538 if (i915_active_acquire_if_busy(ref))
542 __i915_active_activate(ref);
546 err = mutex_lock_interruptible(&ref->mutex);
550 if (likely(!i915_active_acquire_if_busy(ref))) {
551 err = ref->active(ref);
553 __i915_active_activate(ref);
556 mutex_unlock(&ref->mutex);
561 int i915_active_acquire_for_context(struct i915_active *ref, u64 idx)
563 struct i915_active_fence *active;
566 err = i915_active_acquire(ref);
570 active = active_instance(ref, idx);
572 i915_active_release(ref);
576 return 0; /* return with active ref */
579 void i915_active_release(struct i915_active *ref)
581 debug_active_assert(ref);
585 static void enable_signaling(struct i915_active_fence *active)
587 struct dma_fence *fence;
589 if (unlikely(is_barrier(active)))
592 fence = i915_active_fence_get(active);
596 dma_fence_enable_sw_signaling(fence);
597 dma_fence_put(fence);
600 static int flush_barrier(struct active_node *it)
602 struct intel_engine_cs *engine;
604 if (likely(!is_barrier(&it->base)))
607 engine = __barrier_to_engine(it);
608 smp_rmb(); /* serialise with add_active_barriers */
609 if (!is_barrier(&it->base))
612 return intel_engine_flush_barriers(engine);
615 static int flush_lazy_signals(struct i915_active *ref)
617 struct active_node *it, *n;
620 enable_signaling(&ref->excl);
621 rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
622 err = flush_barrier(it); /* unconnected idle barrier? */
626 enable_signaling(&it->base);
632 int __i915_active_wait(struct i915_active *ref, int state)
636 /* Any fence added after the wait begins will not be auto-signaled */
637 if (i915_active_acquire_if_busy(ref)) {
640 err = flush_lazy_signals(ref);
641 i915_active_release(ref);
645 if (___wait_var_event(ref, i915_active_is_idle(ref),
646 state, 0, 0, schedule()))
651 * After the wait is complete, the caller may free the active.
652 * We have to flush any concurrent retirement before returning.
654 flush_work(&ref->work);
658 static int __await_active(struct i915_active_fence *active,
659 int (*fn)(void *arg, struct dma_fence *fence),
662 struct dma_fence *fence;
664 if (is_barrier(active)) /* XXX flush the barrier? */
667 fence = i915_active_fence_get(active);
671 err = fn(arg, fence);
672 dma_fence_put(fence);
680 struct wait_barrier {
681 struct wait_queue_entry base;
682 struct i915_active *ref;
686 barrier_wake(wait_queue_entry_t *wq, unsigned int mode, int flags, void *key)
688 struct wait_barrier *wb = container_of(wq, typeof(*wb), base);
690 if (i915_active_is_idle(wb->ref)) {
691 list_del(&wq->entry);
692 i915_sw_fence_complete(wq->private);
699 static int __await_barrier(struct i915_active *ref, struct i915_sw_fence *fence)
701 struct wait_barrier *wb;
703 wb = kmalloc(sizeof(*wb), GFP_KERNEL);
707 GEM_BUG_ON(i915_active_is_idle(ref));
708 if (!i915_sw_fence_await(fence)) {
714 wb->base.func = barrier_wake;
715 wb->base.private = fence;
718 add_wait_queue(__var_waitqueue(ref), &wb->base);
722 static int await_active(struct i915_active *ref,
724 int (*fn)(void *arg, struct dma_fence *fence),
725 void *arg, struct i915_sw_fence *barrier)
729 if (!i915_active_acquire_if_busy(ref))
732 if (flags & I915_ACTIVE_AWAIT_EXCL &&
733 rcu_access_pointer(ref->excl.fence)) {
734 err = __await_active(&ref->excl, fn, arg);
739 if (flags & I915_ACTIVE_AWAIT_ACTIVE) {
740 struct active_node *it, *n;
742 rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
743 err = __await_active(&it->base, fn, arg);
749 if (flags & I915_ACTIVE_AWAIT_BARRIER) {
750 err = flush_lazy_signals(ref);
754 err = __await_barrier(ref, barrier);
760 i915_active_release(ref);
764 static int rq_await_fence(void *arg, struct dma_fence *fence)
766 return i915_request_await_dma_fence(arg, fence);
769 int i915_request_await_active(struct i915_request *rq,
770 struct i915_active *ref,
773 return await_active(ref, flags, rq_await_fence, rq, &rq->submit);
776 static int sw_await_fence(void *arg, struct dma_fence *fence)
778 return i915_sw_fence_await_dma_fence(arg, fence, 0,
779 GFP_NOWAIT | __GFP_NOWARN);
782 int i915_sw_fence_await_active(struct i915_sw_fence *fence,
783 struct i915_active *ref,
786 return await_active(ref, flags, sw_await_fence, fence, fence);
789 void i915_active_fini(struct i915_active *ref)
791 debug_active_fini(ref);
792 GEM_BUG_ON(atomic_read(&ref->count));
793 GEM_BUG_ON(work_pending(&ref->work));
794 mutex_destroy(&ref->mutex);
797 kmem_cache_free(global.slab_cache, ref->cache);
800 static inline bool is_idle_barrier(struct active_node *node, u64 idx)
802 return node->timeline == idx && !i915_active_fence_isset(&node->base);
805 static struct active_node *reuse_idle_barrier(struct i915_active *ref, u64 idx)
807 struct rb_node *prev, *p;
809 if (RB_EMPTY_ROOT(&ref->tree))
812 GEM_BUG_ON(i915_active_is_idle(ref));
815 * Try to reuse any existing barrier nodes already allocated for this
816 * i915_active, due to overlapping active phases there is likely a
817 * node kept alive (as we reuse before parking). We prefer to reuse
818 * completely idle barriers (less hassle in manipulating the llists),
819 * but otherwise any will do.
821 if (ref->cache && is_idle_barrier(ref->cache, idx)) {
822 p = &ref->cache->node;
827 p = ref->tree.rb_node;
829 struct active_node *node =
830 rb_entry(p, struct active_node, node);
832 if (is_idle_barrier(node, idx))
836 if (node->timeline < idx)
837 p = READ_ONCE(p->rb_right);
839 p = READ_ONCE(p->rb_left);
843 * No quick match, but we did find the leftmost rb_node for the
844 * kernel_context. Walk the rb_tree in-order to see if there were
845 * any idle-barriers on this timeline that we missed, or just use
846 * the first pending barrier.
848 for (p = prev; p; p = rb_next(p)) {
849 struct active_node *node =
850 rb_entry(p, struct active_node, node);
851 struct intel_engine_cs *engine;
853 if (node->timeline > idx)
856 if (node->timeline < idx)
859 if (is_idle_barrier(node, idx))
863 * The list of pending barriers is protected by the
864 * kernel_context timeline, which notably we do not hold
865 * here. i915_request_add_active_barriers() may consume
866 * the barrier before we claim it, so we have to check
869 engine = __barrier_to_engine(node);
870 smp_rmb(); /* serialise with add_active_barriers */
871 if (is_barrier(&node->base) &&
872 ____active_del_barrier(ref, node, engine))
879 spin_lock_irq(&ref->tree_lock);
880 rb_erase(p, &ref->tree); /* Hide from waits and sibling allocations */
881 if (p == &ref->cache->node)
882 WRITE_ONCE(ref->cache, NULL);
883 spin_unlock_irq(&ref->tree_lock);
885 return rb_entry(p, struct active_node, node);
888 int i915_active_acquire_preallocate_barrier(struct i915_active *ref,
889 struct intel_engine_cs *engine)
891 intel_engine_mask_t tmp, mask = engine->mask;
892 struct llist_node *first = NULL, *last = NULL;
893 struct intel_gt *gt = engine->gt;
895 GEM_BUG_ON(i915_active_is_idle(ref));
897 /* Wait until the previous preallocation is completed */
898 while (!llist_empty(&ref->preallocated_barriers))
902 * Preallocate a node for each physical engine supporting the target
903 * engine (remember virtual engines have more than one sibling).
904 * We can then use the preallocated nodes in
905 * i915_active_acquire_barrier()
908 for_each_engine_masked(engine, gt, mask, tmp) {
909 u64 idx = engine->kernel_context->timeline->fence_context;
910 struct llist_node *prev = first;
911 struct active_node *node;
914 node = reuse_idle_barrier(ref, idx);
917 node = kmem_cache_alloc(global.slab_cache, GFP_KERNEL);
921 RCU_INIT_POINTER(node->base.fence, NULL);
922 node->base.cb.func = node_retire;
923 node->timeline = idx;
927 if (!i915_active_fence_isset(&node->base)) {
929 * Mark this as being *our* unconnected proto-node.
931 * Since this node is not in any list, and we have
932 * decoupled it from the rbtree, we can reuse the
933 * request to indicate this is an idle-barrier node
934 * and then we can use the rb_node and list pointers
935 * for our tracking of the pending barrier.
937 RCU_INIT_POINTER(node->base.fence, ERR_PTR(-EAGAIN));
938 node->base.cb.node.prev = (void *)engine;
939 __i915_active_acquire(ref);
941 GEM_BUG_ON(rcu_access_pointer(node->base.fence) != ERR_PTR(-EAGAIN));
943 GEM_BUG_ON(barrier_to_engine(node) != engine);
944 first = barrier_to_ll(node);
948 intel_engine_pm_get(engine);
951 GEM_BUG_ON(!llist_empty(&ref->preallocated_barriers));
952 llist_add_batch(first, last, &ref->preallocated_barriers);
958 struct active_node *node = barrier_from_ll(first);
962 atomic_dec(&ref->count);
963 intel_engine_pm_put(barrier_to_engine(node));
965 kmem_cache_free(global.slab_cache, node);
970 void i915_active_acquire_barrier(struct i915_active *ref)
972 struct llist_node *pos, *next;
975 GEM_BUG_ON(i915_active_is_idle(ref));
978 * Transfer the list of preallocated barriers into the
979 * i915_active rbtree, but only as proto-nodes. They will be
980 * populated by i915_request_add_active_barriers() to point to the
981 * request that will eventually release them.
983 llist_for_each_safe(pos, next, take_preallocated_barriers(ref)) {
984 struct active_node *node = barrier_from_ll(pos);
985 struct intel_engine_cs *engine = barrier_to_engine(node);
986 struct rb_node **p, *parent;
988 spin_lock_irqsave_nested(&ref->tree_lock, flags,
989 SINGLE_DEPTH_NESTING);
991 p = &ref->tree.rb_node;
993 struct active_node *it;
997 it = rb_entry(parent, struct active_node, node);
998 if (it->timeline < node->timeline)
999 p = &parent->rb_right;
1001 p = &parent->rb_left;
1003 rb_link_node(&node->node, parent, p);
1004 rb_insert_color(&node->node, &ref->tree);
1005 spin_unlock_irqrestore(&ref->tree_lock, flags);
1007 GEM_BUG_ON(!intel_engine_pm_is_awake(engine));
1008 llist_add(barrier_to_ll(node), &engine->barrier_tasks);
1009 intel_engine_pm_put_delay(engine, 1);
1013 static struct dma_fence **ll_to_fence_slot(struct llist_node *node)
1015 return __active_fence_slot(&barrier_from_ll(node)->base);
1018 void i915_request_add_active_barriers(struct i915_request *rq)
1020 struct intel_engine_cs *engine = rq->engine;
1021 struct llist_node *node, *next;
1022 unsigned long flags;
1024 GEM_BUG_ON(!intel_context_is_barrier(rq->context));
1025 GEM_BUG_ON(intel_engine_is_virtual(engine));
1026 GEM_BUG_ON(i915_request_timeline(rq) != engine->kernel_context->timeline);
1028 node = llist_del_all(&engine->barrier_tasks);
1032 * Attach the list of proto-fences to the in-flight request such
1033 * that the parent i915_active will be released when this request
1036 spin_lock_irqsave(&rq->lock, flags);
1037 llist_for_each_safe(node, next, node) {
1038 /* serialise with reuse_idle_barrier */
1039 smp_store_mb(*ll_to_fence_slot(node), &rq->fence);
1040 list_add_tail((struct list_head *)node, &rq->fence.cb_list);
1042 spin_unlock_irqrestore(&rq->lock, flags);
1046 * __i915_active_fence_set: Update the last active fence along its timeline
1047 * @active: the active tracker
1048 * @fence: the new fence (under construction)
1050 * Records the new @fence as the last active fence along its timeline in
1051 * this active tracker, moving the tracking callbacks from the previous
1052 * fence onto this one. Gets and returns a reference to the previous fence
1053 * (if not already completed), which the caller must put after making sure
1054 * that it is executed before the new fence. To ensure that the order of
1055 * fences within the timeline of the i915_active_fence is understood, it
1056 * should be locked by the caller.
1059 __i915_active_fence_set(struct i915_active_fence *active,
1060 struct dma_fence *fence)
1062 struct dma_fence *prev;
1063 unsigned long flags;
1066 * In case of fences embedded in i915_requests, their memory is
1067 * SLAB_FAILSAFE_BY_RCU, then it can be reused right after release
1068 * by new requests. Then, there is a risk of passing back a pointer
1069 * to a new, completely unrelated fence that reuses the same memory
1070 * while tracked under a different active tracker. Combined with i915
1071 * perf open/close operations that build await dependencies between
1072 * engine kernel context requests and user requests from different
1073 * timelines, this can lead to dependency loops and infinite waits.
1075 * As a countermeasure, we try to get a reference to the active->fence
1076 * first, so if we succeed and pass it back to our user then it is not
1077 * released and potentially reused by an unrelated request before the
1078 * user has a chance to set up an await dependency on it.
1080 prev = i915_active_fence_get(active);
1084 GEM_BUG_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags));
1087 * Consider that we have two threads arriving (A and B), with
1088 * C already resident as the active->fence.
1090 * Both A and B have got a reference to C or NULL, depending on the
1091 * timing of the interrupt handler. Let's assume that if A has got C
1092 * then it has locked C first (before B).
1094 * Note the strong ordering of the timeline also provides consistent
1095 * nesting rules for the fence->lock; the inner lock is always the
1098 spin_lock_irqsave(fence->lock, flags);
1100 spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING);
1103 * A does the cmpxchg first, and so it sees C or NULL, as before, or
1104 * something else, depending on the timing of other threads and/or
1105 * interrupt handler. If not the same as before then A unlocks C if
1106 * applicable and retries, starting from an attempt to get a new
1107 * active->fence. Meanwhile, B follows the same path as A.
1108 * Once A succeeds with cmpxch, B fails again, retires, gets A from
1109 * active->fence, locks it as soon as A completes, and possibly
1110 * succeeds with cmpxchg.
1112 while (cmpxchg(__active_fence_slot(active), prev, fence) != prev) {
1114 spin_unlock(prev->lock);
1115 dma_fence_put(prev);
1117 spin_unlock_irqrestore(fence->lock, flags);
1119 prev = i915_active_fence_get(active);
1120 GEM_BUG_ON(prev == fence);
1122 spin_lock_irqsave(fence->lock, flags);
1124 spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING);
1128 * If prev is NULL then the previous fence must have been signaled
1129 * and we know that we are first on the timeline. If it is still
1130 * present then, having the lock on that fence already acquired, we
1131 * serialise with the interrupt handler, in the process of removing it
1132 * from any future interrupt callback. A will then wait on C before
1133 * executing (if present).
1135 * As B is second, it sees A as the previous fence and so waits for
1136 * it to complete its transition and takes over the occupancy for
1137 * itself -- remembering that it needs to wait on A before executing.
1140 __list_del_entry(&active->cb.node);
1141 spin_unlock(prev->lock); /* serialise with prev->cb_list */
1143 list_add_tail(&active->cb.node, &fence->cb_list);
1144 spin_unlock_irqrestore(fence->lock, flags);
1149 int i915_active_fence_set(struct i915_active_fence *active,
1150 struct i915_request *rq)
1152 struct dma_fence *fence;
1155 /* Must maintain timeline ordering wrt previous active requests */
1156 fence = __i915_active_fence_set(active, &rq->fence);
1158 err = i915_request_await_dma_fence(rq, fence);
1159 dma_fence_put(fence);
1165 void i915_active_noop(struct dma_fence *fence, struct dma_fence_cb *cb)
1167 active_fence_cb(fence, cb);
1170 struct auto_active {
1171 struct i915_active base;
1175 struct i915_active *i915_active_get(struct i915_active *ref)
1177 struct auto_active *aa = container_of(ref, typeof(*aa), base);
1183 static void auto_release(struct kref *ref)
1185 struct auto_active *aa = container_of(ref, typeof(*aa), ref);
1187 i915_active_fini(&aa->base);
1191 void i915_active_put(struct i915_active *ref)
1193 struct auto_active *aa = container_of(ref, typeof(*aa), base);
1195 kref_put(&aa->ref, auto_release);
1198 static int auto_active(struct i915_active *ref)
1200 i915_active_get(ref);
1204 __i915_active_call static void
1205 auto_retire(struct i915_active *ref)
1207 i915_active_put(ref);
1210 struct i915_active *i915_active_create(void)
1212 struct auto_active *aa;
1214 aa = kmalloc(sizeof(*aa), GFP_KERNEL);
1218 kref_init(&aa->ref);
1219 i915_active_init(&aa->base, auto_active, auto_retire);
1224 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1225 #include "selftests/i915_active.c"
1228 static void i915_global_active_shrink(void)
1230 kmem_cache_shrink(global.slab_cache);
1233 static void i915_global_active_exit(void)
1235 kmem_cache_destroy(global.slab_cache);
1238 static struct i915_global_active global = { {
1239 .shrink = i915_global_active_shrink,
1240 .exit = i915_global_active_exit,
1243 int __init i915_global_active_init(void)
1245 global.slab_cache = KMEM_CACHE(active_node, SLAB_HWCACHE_ALIGN);
1246 if (!global.slab_cache)
1249 i915_global_register(&global.base);