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_pm.h"
11 #include "gt/intel_ring.h"
14 #include "i915_active.h"
15 #include "i915_globals.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 i915_global_active {
25 struct i915_global base;
26 struct kmem_cache *slab_cache;
30 struct i915_active_fence base;
31 struct i915_active *ref;
36 static inline struct active_node *
37 node_from_active(struct i915_active_fence *active)
39 return container_of(active, struct active_node, base);
42 #define take_preallocated_barriers(x) llist_del_all(&(x)->preallocated_barriers)
44 static inline bool is_barrier(const struct i915_active_fence *active)
46 return IS_ERR(rcu_access_pointer(active->fence));
49 static inline struct llist_node *barrier_to_ll(struct active_node *node)
51 GEM_BUG_ON(!is_barrier(&node->base));
52 return (struct llist_node *)&node->base.cb.node;
55 static inline struct intel_engine_cs *
56 __barrier_to_engine(struct active_node *node)
58 return (struct intel_engine_cs *)READ_ONCE(node->base.cb.node.prev);
61 static inline struct intel_engine_cs *
62 barrier_to_engine(struct active_node *node)
64 GEM_BUG_ON(!is_barrier(&node->base));
65 return __barrier_to_engine(node);
68 static inline struct active_node *barrier_from_ll(struct llist_node *x)
70 return container_of((struct list_head *)x,
71 struct active_node, base.cb.node);
74 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM) && IS_ENABLED(CONFIG_DEBUG_OBJECTS)
76 static void *active_debug_hint(void *addr)
78 struct i915_active *ref = addr;
80 return (void *)ref->active ?: (void *)ref->retire ?: (void *)ref;
83 static struct debug_obj_descr active_debug_desc = {
84 .name = "i915_active",
85 .debug_hint = active_debug_hint,
88 static void debug_active_init(struct i915_active *ref)
90 debug_object_init(ref, &active_debug_desc);
93 static void debug_active_activate(struct i915_active *ref)
95 lockdep_assert_held(&ref->tree_lock);
96 if (!atomic_read(&ref->count)) /* before the first inc */
97 debug_object_activate(ref, &active_debug_desc);
100 static void debug_active_deactivate(struct i915_active *ref)
102 lockdep_assert_held(&ref->tree_lock);
103 if (!atomic_read(&ref->count)) /* after the last dec */
104 debug_object_deactivate(ref, &active_debug_desc);
107 static void debug_active_fini(struct i915_active *ref)
109 debug_object_free(ref, &active_debug_desc);
112 static void debug_active_assert(struct i915_active *ref)
114 debug_object_assert_init(ref, &active_debug_desc);
119 static inline void debug_active_init(struct i915_active *ref) { }
120 static inline void debug_active_activate(struct i915_active *ref) { }
121 static inline void debug_active_deactivate(struct i915_active *ref) { }
122 static inline void debug_active_fini(struct i915_active *ref) { }
123 static inline void debug_active_assert(struct i915_active *ref) { }
128 __active_retire(struct i915_active *ref)
130 struct active_node *it, *n;
134 GEM_BUG_ON(i915_active_is_idle(ref));
136 /* return the unused nodes to our slabcache -- flushing the allocator */
137 if (!atomic_dec_and_lock_irqsave(&ref->count, &ref->tree_lock, flags))
140 GEM_BUG_ON(rcu_access_pointer(ref->excl.fence));
141 debug_active_deactivate(ref);
147 spin_unlock_irqrestore(&ref->tree_lock, flags);
149 /* After the final retire, the entire struct may be freed */
153 /* ... except if you wait on it, you must manage your own references! */
156 rbtree_postorder_for_each_entry_safe(it, n, &root, node) {
157 GEM_BUG_ON(i915_active_fence_isset(&it->base));
158 kmem_cache_free(global.slab_cache, it);
163 active_work(struct work_struct *wrk)
165 struct i915_active *ref = container_of(wrk, typeof(*ref), work);
167 GEM_BUG_ON(!atomic_read(&ref->count));
168 if (atomic_add_unless(&ref->count, -1, 1))
171 __active_retire(ref);
175 active_retire(struct i915_active *ref)
177 GEM_BUG_ON(!atomic_read(&ref->count));
178 if (atomic_add_unless(&ref->count, -1, 1))
181 if (ref->flags & I915_ACTIVE_RETIRE_SLEEPS) {
182 queue_work(system_unbound_wq, &ref->work);
186 __active_retire(ref);
189 static inline struct dma_fence **
190 __active_fence_slot(struct i915_active_fence *active)
192 return (struct dma_fence ** __force)&active->fence;
196 active_fence_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
198 struct i915_active_fence *active =
199 container_of(cb, typeof(*active), cb);
201 return cmpxchg(__active_fence_slot(active), fence, NULL) == fence;
205 node_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
207 if (active_fence_cb(fence, cb))
208 active_retire(container_of(cb, struct active_node, base.cb)->ref);
212 excl_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
214 if (active_fence_cb(fence, cb))
215 active_retire(container_of(cb, struct i915_active, excl.cb));
218 static struct i915_active_fence *
219 active_instance(struct i915_active *ref, struct intel_timeline *tl)
221 struct active_node *node, *prealloc;
222 struct rb_node **p, *parent;
223 u64 idx = tl->fence_context;
226 * We track the most recently used timeline to skip a rbtree search
227 * for the common case, under typical loads we never need the rbtree
228 * at all. We can reuse the last slot if it is empty, that is
229 * after the previous activity has been retired, or if it matches the
232 node = READ_ONCE(ref->cache);
233 if (node && node->timeline == idx)
236 /* Preallocate a replacement, just in case */
237 prealloc = kmem_cache_alloc(global.slab_cache, GFP_KERNEL);
241 spin_lock_irq(&ref->tree_lock);
242 GEM_BUG_ON(i915_active_is_idle(ref));
245 p = &ref->tree.rb_node;
249 node = rb_entry(parent, struct active_node, node);
250 if (node->timeline == idx) {
251 kmem_cache_free(global.slab_cache, prealloc);
255 if (node->timeline < idx)
256 p = &parent->rb_right;
258 p = &parent->rb_left;
262 __i915_active_fence_init(&node->base, NULL, node_retire);
264 node->timeline = idx;
266 rb_link_node(&node->node, parent, p);
267 rb_insert_color(&node->node, &ref->tree);
271 spin_unlock_irq(&ref->tree_lock);
273 BUILD_BUG_ON(offsetof(typeof(*node), base));
277 void __i915_active_init(struct i915_active *ref,
278 int (*active)(struct i915_active *ref),
279 void (*retire)(struct i915_active *ref),
280 struct lock_class_key *mkey,
281 struct lock_class_key *wkey)
285 debug_active_init(ref);
288 ref->active = active;
289 ref->retire = ptr_unpack_bits(retire, &bits, 2);
290 if (bits & I915_ACTIVE_MAY_SLEEP)
291 ref->flags |= I915_ACTIVE_RETIRE_SLEEPS;
293 spin_lock_init(&ref->tree_lock);
297 init_llist_head(&ref->preallocated_barriers);
298 atomic_set(&ref->count, 0);
299 __mutex_init(&ref->mutex, "i915_active", mkey);
300 __i915_active_fence_init(&ref->excl, NULL, excl_retire);
301 INIT_WORK(&ref->work, active_work);
302 #if IS_ENABLED(CONFIG_LOCKDEP)
303 lockdep_init_map(&ref->work.lockdep_map, "i915_active.work", wkey, 0);
307 static bool ____active_del_barrier(struct i915_active *ref,
308 struct active_node *node,
309 struct intel_engine_cs *engine)
312 struct llist_node *head = NULL, *tail = NULL;
313 struct llist_node *pos, *next;
315 GEM_BUG_ON(node->timeline != engine->kernel_context->timeline->fence_context);
318 * Rebuild the llist excluding our node. We may perform this
319 * outside of the kernel_context timeline mutex and so someone
320 * else may be manipulating the engine->barrier_tasks, in
321 * which case either we or they will be upset :)
323 * A second __active_del_barrier() will report failure to claim
324 * the active_node and the caller will just shrug and know not to
325 * claim ownership of its node.
327 * A concurrent i915_request_add_active_barriers() will miss adding
328 * any of the tasks, but we will try again on the next -- and since
329 * we are actively using the barrier, we know that there will be
330 * at least another opportunity when we idle.
332 llist_for_each_safe(pos, next, llist_del_all(&engine->barrier_tasks)) {
333 if (node == barrier_from_ll(pos)) {
344 llist_add_batch(head, tail, &engine->barrier_tasks);
350 __active_del_barrier(struct i915_active *ref, struct active_node *node)
352 return ____active_del_barrier(ref, node, barrier_to_engine(node));
355 int i915_active_ref(struct i915_active *ref,
356 struct intel_timeline *tl,
357 struct dma_fence *fence)
359 struct i915_active_fence *active;
362 lockdep_assert_held(&tl->mutex);
364 /* Prevent reaping in case we malloc/wait while building the tree */
365 err = i915_active_acquire(ref);
369 active = active_instance(ref, tl);
375 if (is_barrier(active)) { /* proto-node used by our idle barrier */
377 * This request is on the kernel_context timeline, and so
378 * we can use it to substitute for the pending idle-barrer
379 * request that we want to emit on the kernel_context.
381 __active_del_barrier(ref, node_from_active(active));
382 RCU_INIT_POINTER(active->fence, NULL);
383 atomic_dec(&ref->count);
385 if (!__i915_active_fence_set(active, fence))
386 atomic_inc(&ref->count);
389 i915_active_release(ref);
394 i915_active_set_exclusive(struct i915_active *ref, struct dma_fence *f)
396 struct dma_fence *prev;
398 /* We expect the caller to manage the exclusive timeline ordering */
399 GEM_BUG_ON(i915_active_is_idle(ref));
402 prev = __i915_active_fence_set(&ref->excl, f);
404 prev = dma_fence_get_rcu(prev);
406 atomic_inc(&ref->count);
412 bool i915_active_acquire_if_busy(struct i915_active *ref)
414 debug_active_assert(ref);
415 return atomic_add_unless(&ref->count, 1, 0);
418 int i915_active_acquire(struct i915_active *ref)
422 if (i915_active_acquire_if_busy(ref))
425 err = mutex_lock_interruptible(&ref->mutex);
429 if (likely(!i915_active_acquire_if_busy(ref))) {
431 err = ref->active(ref);
433 spin_lock_irq(&ref->tree_lock); /* __active_retire() */
434 debug_active_activate(ref);
435 atomic_inc(&ref->count);
436 spin_unlock_irq(&ref->tree_lock);
440 mutex_unlock(&ref->mutex);
445 void i915_active_release(struct i915_active *ref)
447 debug_active_assert(ref);
451 static void enable_signaling(struct i915_active_fence *active)
453 struct dma_fence *fence;
455 fence = i915_active_fence_get(active);
459 dma_fence_enable_sw_signaling(fence);
460 dma_fence_put(fence);
463 int i915_active_wait(struct i915_active *ref)
465 struct active_node *it, *n;
470 if (!i915_active_acquire_if_busy(ref))
473 /* Flush lazy signals */
474 enable_signaling(&ref->excl);
475 rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
476 if (is_barrier(&it->base)) /* unconnected idle barrier */
479 enable_signaling(&it->base);
481 /* Any fence added after the wait begins will not be auto-signaled */
483 i915_active_release(ref);
487 if (wait_var_event_interruptible(ref, i915_active_is_idle(ref)))
490 flush_work(&ref->work);
494 int i915_request_await_active(struct i915_request *rq, struct i915_active *ref)
498 if (rcu_access_pointer(ref->excl.fence)) {
499 struct dma_fence *fence;
502 fence = dma_fence_get_rcu_safe(&ref->excl.fence);
505 err = i915_request_await_dma_fence(rq, fence);
506 dma_fence_put(fence);
510 /* In the future we may choose to await on all fences */
515 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)
516 void i915_active_fini(struct i915_active *ref)
518 debug_active_fini(ref);
519 GEM_BUG_ON(atomic_read(&ref->count));
520 GEM_BUG_ON(work_pending(&ref->work));
521 GEM_BUG_ON(!RB_EMPTY_ROOT(&ref->tree));
522 mutex_destroy(&ref->mutex);
526 static inline bool is_idle_barrier(struct active_node *node, u64 idx)
528 return node->timeline == idx && !i915_active_fence_isset(&node->base);
531 static struct active_node *reuse_idle_barrier(struct i915_active *ref, u64 idx)
533 struct rb_node *prev, *p;
535 if (RB_EMPTY_ROOT(&ref->tree))
538 spin_lock_irq(&ref->tree_lock);
539 GEM_BUG_ON(i915_active_is_idle(ref));
542 * Try to reuse any existing barrier nodes already allocated for this
543 * i915_active, due to overlapping active phases there is likely a
544 * node kept alive (as we reuse before parking). We prefer to reuse
545 * completely idle barriers (less hassle in manipulating the llists),
546 * but otherwise any will do.
548 if (ref->cache && is_idle_barrier(ref->cache, idx)) {
549 p = &ref->cache->node;
554 p = ref->tree.rb_node;
556 struct active_node *node =
557 rb_entry(p, struct active_node, node);
559 if (is_idle_barrier(node, idx))
563 if (node->timeline < idx)
570 * No quick match, but we did find the leftmost rb_node for the
571 * kernel_context. Walk the rb_tree in-order to see if there were
572 * any idle-barriers on this timeline that we missed, or just use
573 * the first pending barrier.
575 for (p = prev; p; p = rb_next(p)) {
576 struct active_node *node =
577 rb_entry(p, struct active_node, node);
578 struct intel_engine_cs *engine;
580 if (node->timeline > idx)
583 if (node->timeline < idx)
586 if (is_idle_barrier(node, idx))
590 * The list of pending barriers is protected by the
591 * kernel_context timeline, which notably we do not hold
592 * here. i915_request_add_active_barriers() may consume
593 * the barrier before we claim it, so we have to check
596 engine = __barrier_to_engine(node);
597 smp_rmb(); /* serialise with add_active_barriers */
598 if (is_barrier(&node->base) &&
599 ____active_del_barrier(ref, node, engine))
603 spin_unlock_irq(&ref->tree_lock);
608 rb_erase(p, &ref->tree); /* Hide from waits and sibling allocations */
609 if (p == &ref->cache->node)
611 spin_unlock_irq(&ref->tree_lock);
613 return rb_entry(p, struct active_node, node);
616 int i915_active_acquire_preallocate_barrier(struct i915_active *ref,
617 struct intel_engine_cs *engine)
619 intel_engine_mask_t tmp, mask = engine->mask;
620 struct llist_node *first = NULL, *last = NULL;
621 struct intel_gt *gt = engine->gt;
624 GEM_BUG_ON(i915_active_is_idle(ref));
626 /* Wait until the previous preallocation is completed */
627 while (!llist_empty(&ref->preallocated_barriers))
631 * Preallocate a node for each physical engine supporting the target
632 * engine (remember virtual engines have more than one sibling).
633 * We can then use the preallocated nodes in
634 * i915_active_acquire_barrier()
637 for_each_engine_masked(engine, gt, mask, tmp) {
638 u64 idx = engine->kernel_context->timeline->fence_context;
639 struct llist_node *prev = first;
640 struct active_node *node;
642 node = reuse_idle_barrier(ref, idx);
644 node = kmem_cache_alloc(global.slab_cache, GFP_KERNEL);
650 RCU_INIT_POINTER(node->base.fence, NULL);
651 node->base.cb.func = node_retire;
652 node->timeline = idx;
656 if (!i915_active_fence_isset(&node->base)) {
658 * Mark this as being *our* unconnected proto-node.
660 * Since this node is not in any list, and we have
661 * decoupled it from the rbtree, we can reuse the
662 * request to indicate this is an idle-barrier node
663 * and then we can use the rb_node and list pointers
664 * for our tracking of the pending barrier.
666 RCU_INIT_POINTER(node->base.fence, ERR_PTR(-EAGAIN));
667 node->base.cb.node.prev = (void *)engine;
668 atomic_inc(&ref->count);
670 GEM_BUG_ON(rcu_access_pointer(node->base.fence) != ERR_PTR(-EAGAIN));
672 GEM_BUG_ON(barrier_to_engine(node) != engine);
673 first = barrier_to_ll(node);
677 intel_engine_pm_get(engine);
680 GEM_BUG_ON(!llist_empty(&ref->preallocated_barriers));
681 llist_add_batch(first, last, &ref->preallocated_barriers);
687 struct active_node *node = barrier_from_ll(first);
691 atomic_dec(&ref->count);
692 intel_engine_pm_put(barrier_to_engine(node));
694 kmem_cache_free(global.slab_cache, node);
699 void i915_active_acquire_barrier(struct i915_active *ref)
701 struct llist_node *pos, *next;
704 GEM_BUG_ON(i915_active_is_idle(ref));
707 * Transfer the list of preallocated barriers into the
708 * i915_active rbtree, but only as proto-nodes. They will be
709 * populated by i915_request_add_active_barriers() to point to the
710 * request that will eventually release them.
712 llist_for_each_safe(pos, next, take_preallocated_barriers(ref)) {
713 struct active_node *node = barrier_from_ll(pos);
714 struct intel_engine_cs *engine = barrier_to_engine(node);
715 struct rb_node **p, *parent;
717 spin_lock_irqsave_nested(&ref->tree_lock, flags,
718 SINGLE_DEPTH_NESTING);
720 p = &ref->tree.rb_node;
722 struct active_node *it;
726 it = rb_entry(parent, struct active_node, node);
727 if (it->timeline < node->timeline)
728 p = &parent->rb_right;
730 p = &parent->rb_left;
732 rb_link_node(&node->node, parent, p);
733 rb_insert_color(&node->node, &ref->tree);
734 spin_unlock_irqrestore(&ref->tree_lock, flags);
736 GEM_BUG_ON(!intel_engine_pm_is_awake(engine));
737 llist_add(barrier_to_ll(node), &engine->barrier_tasks);
738 intel_engine_pm_put(engine);
742 static struct dma_fence **ll_to_fence_slot(struct llist_node *node)
744 return __active_fence_slot(&barrier_from_ll(node)->base);
747 void i915_request_add_active_barriers(struct i915_request *rq)
749 struct intel_engine_cs *engine = rq->engine;
750 struct llist_node *node, *next;
753 GEM_BUG_ON(!intel_context_is_barrier(rq->context));
754 GEM_BUG_ON(intel_engine_is_virtual(engine));
755 GEM_BUG_ON(i915_request_timeline(rq) != engine->kernel_context->timeline);
757 node = llist_del_all(&engine->barrier_tasks);
761 * Attach the list of proto-fences to the in-flight request such
762 * that the parent i915_active will be released when this request
765 spin_lock_irqsave(&rq->lock, flags);
766 llist_for_each_safe(node, next, node) {
767 /* serialise with reuse_idle_barrier */
768 smp_store_mb(*ll_to_fence_slot(node), &rq->fence);
769 list_add_tail((struct list_head *)node, &rq->fence.cb_list);
771 spin_unlock_irqrestore(&rq->lock, flags);
775 * __i915_active_fence_set: Update the last active fence along its timeline
776 * @active: the active tracker
777 * @fence: the new fence (under construction)
779 * Records the new @fence as the last active fence along its timeline in
780 * this active tracker, moving the tracking callbacks from the previous
781 * fence onto this one. Returns the previous fence (if not already completed),
782 * which the caller must ensure is executed before the new fence. To ensure
783 * that the order of fences within the timeline of the i915_active_fence is
784 * understood, it should be locked by the caller.
787 __i915_active_fence_set(struct i915_active_fence *active,
788 struct dma_fence *fence)
790 struct dma_fence *prev;
793 if (fence == rcu_access_pointer(active->fence))
796 GEM_BUG_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags));
799 * Consider that we have two threads arriving (A and B), with
800 * C already resident as the active->fence.
802 * A does the xchg first, and so it sees C or NULL depending
803 * on the timing of the interrupt handler. If it is NULL, the
804 * previous fence must have been signaled and we know that
805 * we are first on the timeline. If it is still present,
806 * we acquire the lock on that fence and serialise with the interrupt
807 * handler, in the process removing it from any future interrupt
808 * callback. A will then wait on C before executing (if present).
810 * As B is second, it sees A as the previous fence and so waits for
811 * it to complete its transition and takes over the occupancy for
812 * itself -- remembering that it needs to wait on A before executing.
814 * Note the strong ordering of the timeline also provides consistent
815 * nesting rules for the fence->lock; the inner lock is always the
818 spin_lock_irqsave(fence->lock, flags);
819 prev = xchg(__active_fence_slot(active), fence);
821 GEM_BUG_ON(prev == fence);
822 spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING);
823 __list_del_entry(&active->cb.node);
824 spin_unlock(prev->lock); /* serialise with prev->cb_list */
826 list_add_tail(&active->cb.node, &fence->cb_list);
827 spin_unlock_irqrestore(fence->lock, flags);
832 int i915_active_fence_set(struct i915_active_fence *active,
833 struct i915_request *rq)
835 struct dma_fence *fence;
838 /* Must maintain timeline ordering wrt previous active requests */
840 fence = __i915_active_fence_set(active, &rq->fence);
841 if (fence) /* but the previous fence may not belong to that timeline! */
842 fence = dma_fence_get_rcu(fence);
845 err = i915_request_await_dma_fence(rq, fence);
846 dma_fence_put(fence);
852 void i915_active_noop(struct dma_fence *fence, struct dma_fence_cb *cb)
854 active_fence_cb(fence, cb);
857 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
858 #include "selftests/i915_active.c"
861 static void i915_global_active_shrink(void)
863 kmem_cache_shrink(global.slab_cache);
866 static void i915_global_active_exit(void)
868 kmem_cache_destroy(global.slab_cache);
871 static struct i915_global_active global = { {
872 .shrink = i915_global_active_shrink,
873 .exit = i915_global_active_exit,
876 int __init i915_global_active_init(void)
878 global.slab_cache = KMEM_CACHE(active_node, SLAB_HWCACHE_ALIGN);
879 if (!global.slab_cache)
882 i915_global_register(&global.base);