Loading...
1/*
2 * SPDX-License-Identifier: MIT
3 *
4 * Copyright © 2019 Intel Corporation
5 */
6
7#include <linux/debugobjects.h>
8
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"
13
14#include "i915_drv.h"
15#include "i915_active.h"
16#include "i915_globals.h"
17
18/*
19 * Active refs memory management
20 *
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.
24 */
25static struct i915_global_active {
26 struct i915_global base;
27 struct kmem_cache *slab_cache;
28} global;
29
30struct active_node {
31 struct rb_node node;
32 struct i915_active_fence base;
33 struct i915_active *ref;
34 u64 timeline;
35};
36
37#define fetch_node(x) rb_entry(READ_ONCE(x), typeof(struct active_node), node)
38
39static inline struct active_node *
40node_from_active(struct i915_active_fence *active)
41{
42 return container_of(active, struct active_node, base);
43}
44
45#define take_preallocated_barriers(x) llist_del_all(&(x)->preallocated_barriers)
46
47static inline bool is_barrier(const struct i915_active_fence *active)
48{
49 return IS_ERR(rcu_access_pointer(active->fence));
50}
51
52static inline struct llist_node *barrier_to_ll(struct active_node *node)
53{
54 GEM_BUG_ON(!is_barrier(&node->base));
55 return (struct llist_node *)&node->base.cb.node;
56}
57
58static inline struct intel_engine_cs *
59__barrier_to_engine(struct active_node *node)
60{
61 return (struct intel_engine_cs *)READ_ONCE(node->base.cb.node.prev);
62}
63
64static inline struct intel_engine_cs *
65barrier_to_engine(struct active_node *node)
66{
67 GEM_BUG_ON(!is_barrier(&node->base));
68 return __barrier_to_engine(node);
69}
70
71static inline struct active_node *barrier_from_ll(struct llist_node *x)
72{
73 return container_of((struct list_head *)x,
74 struct active_node, base.cb.node);
75}
76
77#if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM) && IS_ENABLED(CONFIG_DEBUG_OBJECTS)
78
79static void *active_debug_hint(void *addr)
80{
81 struct i915_active *ref = addr;
82
83 return (void *)ref->active ?: (void *)ref->retire ?: (void *)ref;
84}
85
86static const struct debug_obj_descr active_debug_desc = {
87 .name = "i915_active",
88 .debug_hint = active_debug_hint,
89};
90
91static void debug_active_init(struct i915_active *ref)
92{
93 debug_object_init(ref, &active_debug_desc);
94}
95
96static void debug_active_activate(struct i915_active *ref)
97{
98 lockdep_assert_held(&ref->tree_lock);
99 if (!atomic_read(&ref->count)) /* before the first inc */
100 debug_object_activate(ref, &active_debug_desc);
101}
102
103static void debug_active_deactivate(struct i915_active *ref)
104{
105 lockdep_assert_held(&ref->tree_lock);
106 if (!atomic_read(&ref->count)) /* after the last dec */
107 debug_object_deactivate(ref, &active_debug_desc);
108}
109
110static void debug_active_fini(struct i915_active *ref)
111{
112 debug_object_free(ref, &active_debug_desc);
113}
114
115static void debug_active_assert(struct i915_active *ref)
116{
117 debug_object_assert_init(ref, &active_debug_desc);
118}
119
120#else
121
122static inline void debug_active_init(struct i915_active *ref) { }
123static inline void debug_active_activate(struct i915_active *ref) { }
124static inline void debug_active_deactivate(struct i915_active *ref) { }
125static inline void debug_active_fini(struct i915_active *ref) { }
126static inline void debug_active_assert(struct i915_active *ref) { }
127
128#endif
129
130static void
131__active_retire(struct i915_active *ref)
132{
133 struct rb_root root = RB_ROOT;
134 struct active_node *it, *n;
135 unsigned long flags;
136
137 GEM_BUG_ON(i915_active_is_idle(ref));
138
139 /* return the unused nodes to our slabcache -- flushing the allocator */
140 if (!atomic_dec_and_lock_irqsave(&ref->count, &ref->tree_lock, flags))
141 return;
142
143 GEM_BUG_ON(rcu_access_pointer(ref->excl.fence));
144 debug_active_deactivate(ref);
145
146 /* Even if we have not used the cache, we may still have a barrier */
147 if (!ref->cache)
148 ref->cache = fetch_node(ref->tree.rb_node);
149
150 /* Keep the MRU cached node for reuse */
151 if (ref->cache) {
152 /* Discard all other nodes in the tree */
153 rb_erase(&ref->cache->node, &ref->tree);
154 root = ref->tree;
155
156 /* Rebuild the tree with only the cached node */
157 rb_link_node(&ref->cache->node, NULL, &ref->tree.rb_node);
158 rb_insert_color(&ref->cache->node, &ref->tree);
159 GEM_BUG_ON(ref->tree.rb_node != &ref->cache->node);
160
161 /* Make the cached node available for reuse with any timeline */
162 ref->cache->timeline = 0; /* needs cmpxchg(u64) */
163 }
164
165 spin_unlock_irqrestore(&ref->tree_lock, flags);
166
167 /* After the final retire, the entire struct may be freed */
168 if (ref->retire)
169 ref->retire(ref);
170
171 /* ... except if you wait on it, you must manage your own references! */
172 wake_up_var(ref);
173
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);
178 }
179}
180
181static void
182active_work(struct work_struct *wrk)
183{
184 struct i915_active *ref = container_of(wrk, typeof(*ref), work);
185
186 GEM_BUG_ON(!atomic_read(&ref->count));
187 if (atomic_add_unless(&ref->count, -1, 1))
188 return;
189
190 __active_retire(ref);
191}
192
193static void
194active_retire(struct i915_active *ref)
195{
196 GEM_BUG_ON(!atomic_read(&ref->count));
197 if (atomic_add_unless(&ref->count, -1, 1))
198 return;
199
200 if (ref->flags & I915_ACTIVE_RETIRE_SLEEPS) {
201 queue_work(system_unbound_wq, &ref->work);
202 return;
203 }
204
205 __active_retire(ref);
206}
207
208static inline struct dma_fence **
209__active_fence_slot(struct i915_active_fence *active)
210{
211 return (struct dma_fence ** __force)&active->fence;
212}
213
214static inline bool
215active_fence_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
216{
217 struct i915_active_fence *active =
218 container_of(cb, typeof(*active), cb);
219
220 return cmpxchg(__active_fence_slot(active), fence, NULL) == fence;
221}
222
223static void
224node_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
225{
226 if (active_fence_cb(fence, cb))
227 active_retire(container_of(cb, struct active_node, base.cb)->ref);
228}
229
230static void
231excl_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
232{
233 if (active_fence_cb(fence, cb))
234 active_retire(container_of(cb, struct i915_active, excl.cb));
235}
236
237static struct active_node *__active_lookup(struct i915_active *ref, u64 idx)
238{
239 struct active_node *it;
240
241 GEM_BUG_ON(idx == 0); /* 0 is the unordered timeline, rsvd for cache */
242
243 /*
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
248 * current timeline.
249 */
250 it = READ_ONCE(ref->cache);
251 if (it) {
252 u64 cached = READ_ONCE(it->timeline);
253
254 /* Once claimed, this slot will only belong to this idx */
255 if (cached == idx)
256 return it;
257
258 /*
259 * An unclaimed cache [.timeline=0] can only be claimed once.
260 *
261 * If the value is already non-zero, some other thread has
262 * claimed the cache and we know that is does not match our
263 * idx. If, and only if, the timeline is currently zero is it
264 * worth competing to claim it atomically for ourselves (for
265 * only the winner of that race will cmpxchg return the old
266 * value of 0).
267 */
268 if (!cached && !cmpxchg64(&it->timeline, 0, idx))
269 return it;
270 }
271
272 BUILD_BUG_ON(offsetof(typeof(*it), node));
273
274 /* While active, the tree can only be built; not destroyed */
275 GEM_BUG_ON(i915_active_is_idle(ref));
276
277 it = fetch_node(ref->tree.rb_node);
278 while (it) {
279 if (it->timeline < idx) {
280 it = fetch_node(it->node.rb_right);
281 } else if (it->timeline > idx) {
282 it = fetch_node(it->node.rb_left);
283 } else {
284 WRITE_ONCE(ref->cache, it);
285 break;
286 }
287 }
288
289 /* NB: If the tree rotated beneath us, we may miss our target. */
290 return it;
291}
292
293static struct i915_active_fence *
294active_instance(struct i915_active *ref, u64 idx)
295{
296 struct active_node *node;
297 struct rb_node **p, *parent;
298
299 node = __active_lookup(ref, idx);
300 if (likely(node))
301 return &node->base;
302
303 spin_lock_irq(&ref->tree_lock);
304 GEM_BUG_ON(i915_active_is_idle(ref));
305
306 parent = NULL;
307 p = &ref->tree.rb_node;
308 while (*p) {
309 parent = *p;
310
311 node = rb_entry(parent, struct active_node, node);
312 if (node->timeline == idx)
313 goto out;
314
315 if (node->timeline < idx)
316 p = &parent->rb_right;
317 else
318 p = &parent->rb_left;
319 }
320
321 /*
322 * XXX: We should preallocate this before i915_active_ref() is ever
323 * called, but we cannot call into fs_reclaim() anyway, so use GFP_ATOMIC.
324 */
325 node = kmem_cache_alloc(global.slab_cache, GFP_ATOMIC);
326 if (!node)
327 goto out;
328
329 __i915_active_fence_init(&node->base, NULL, node_retire);
330 node->ref = ref;
331 node->timeline = idx;
332
333 rb_link_node(&node->node, parent, p);
334 rb_insert_color(&node->node, &ref->tree);
335
336out:
337 WRITE_ONCE(ref->cache, node);
338 spin_unlock_irq(&ref->tree_lock);
339
340 return &node->base;
341}
342
343void __i915_active_init(struct i915_active *ref,
344 int (*active)(struct i915_active *ref),
345 void (*retire)(struct i915_active *ref),
346 unsigned long flags,
347 struct lock_class_key *mkey,
348 struct lock_class_key *wkey)
349{
350 debug_active_init(ref);
351
352 ref->flags = flags;
353 ref->active = active;
354 ref->retire = retire;
355
356 spin_lock_init(&ref->tree_lock);
357 ref->tree = RB_ROOT;
358 ref->cache = NULL;
359
360 init_llist_head(&ref->preallocated_barriers);
361 atomic_set(&ref->count, 0);
362 __mutex_init(&ref->mutex, "i915_active", mkey);
363 __i915_active_fence_init(&ref->excl, NULL, excl_retire);
364 INIT_WORK(&ref->work, active_work);
365#if IS_ENABLED(CONFIG_LOCKDEP)
366 lockdep_init_map(&ref->work.lockdep_map, "i915_active.work", wkey, 0);
367#endif
368}
369
370static bool ____active_del_barrier(struct i915_active *ref,
371 struct active_node *node,
372 struct intel_engine_cs *engine)
373
374{
375 struct llist_node *head = NULL, *tail = NULL;
376 struct llist_node *pos, *next;
377
378 GEM_BUG_ON(node->timeline != engine->kernel_context->timeline->fence_context);
379
380 /*
381 * Rebuild the llist excluding our node. We may perform this
382 * outside of the kernel_context timeline mutex and so someone
383 * else may be manipulating the engine->barrier_tasks, in
384 * which case either we or they will be upset :)
385 *
386 * A second __active_del_barrier() will report failure to claim
387 * the active_node and the caller will just shrug and know not to
388 * claim ownership of its node.
389 *
390 * A concurrent i915_request_add_active_barriers() will miss adding
391 * any of the tasks, but we will try again on the next -- and since
392 * we are actively using the barrier, we know that there will be
393 * at least another opportunity when we idle.
394 */
395 llist_for_each_safe(pos, next, llist_del_all(&engine->barrier_tasks)) {
396 if (node == barrier_from_ll(pos)) {
397 node = NULL;
398 continue;
399 }
400
401 pos->next = head;
402 head = pos;
403 if (!tail)
404 tail = pos;
405 }
406 if (head)
407 llist_add_batch(head, tail, &engine->barrier_tasks);
408
409 return !node;
410}
411
412static bool
413__active_del_barrier(struct i915_active *ref, struct active_node *node)
414{
415 return ____active_del_barrier(ref, node, barrier_to_engine(node));
416}
417
418static bool
419replace_barrier(struct i915_active *ref, struct i915_active_fence *active)
420{
421 if (!is_barrier(active)) /* proto-node used by our idle barrier? */
422 return false;
423
424 /*
425 * This request is on the kernel_context timeline, and so
426 * we can use it to substitute for the pending idle-barrer
427 * request that we want to emit on the kernel_context.
428 */
429 __active_del_barrier(ref, node_from_active(active));
430 return true;
431}
432
433int i915_active_ref(struct i915_active *ref, u64 idx, struct dma_fence *fence)
434{
435 struct i915_active_fence *active;
436 int err;
437
438 /* Prevent reaping in case we malloc/wait while building the tree */
439 err = i915_active_acquire(ref);
440 if (err)
441 return err;
442
443 active = active_instance(ref, idx);
444 if (!active) {
445 err = -ENOMEM;
446 goto out;
447 }
448
449 if (replace_barrier(ref, active)) {
450 RCU_INIT_POINTER(active->fence, NULL);
451 atomic_dec(&ref->count);
452 }
453 if (!__i915_active_fence_set(active, fence))
454 __i915_active_acquire(ref);
455
456out:
457 i915_active_release(ref);
458 return err;
459}
460
461static struct dma_fence *
462__i915_active_set_fence(struct i915_active *ref,
463 struct i915_active_fence *active,
464 struct dma_fence *fence)
465{
466 struct dma_fence *prev;
467
468 if (replace_barrier(ref, active)) {
469 RCU_INIT_POINTER(active->fence, fence);
470 return NULL;
471 }
472
473 rcu_read_lock();
474 prev = __i915_active_fence_set(active, fence);
475 if (prev)
476 prev = dma_fence_get_rcu(prev);
477 else
478 __i915_active_acquire(ref);
479 rcu_read_unlock();
480
481 return prev;
482}
483
484static struct i915_active_fence *
485__active_fence(struct i915_active *ref, u64 idx)
486{
487 struct active_node *it;
488
489 it = __active_lookup(ref, idx);
490 if (unlikely(!it)) { /* Contention with parallel tree builders! */
491 spin_lock_irq(&ref->tree_lock);
492 it = __active_lookup(ref, idx);
493 spin_unlock_irq(&ref->tree_lock);
494 }
495 GEM_BUG_ON(!it); /* slot must be preallocated */
496
497 return &it->base;
498}
499
500struct dma_fence *
501__i915_active_ref(struct i915_active *ref, u64 idx, struct dma_fence *fence)
502{
503 /* Only valid while active, see i915_active_acquire_for_context() */
504 return __i915_active_set_fence(ref, __active_fence(ref, idx), fence);
505}
506
507struct dma_fence *
508i915_active_set_exclusive(struct i915_active *ref, struct dma_fence *f)
509{
510 /* We expect the caller to manage the exclusive timeline ordering */
511 return __i915_active_set_fence(ref, &ref->excl, f);
512}
513
514bool i915_active_acquire_if_busy(struct i915_active *ref)
515{
516 debug_active_assert(ref);
517 return atomic_add_unless(&ref->count, 1, 0);
518}
519
520static void __i915_active_activate(struct i915_active *ref)
521{
522 spin_lock_irq(&ref->tree_lock); /* __active_retire() */
523 if (!atomic_fetch_inc(&ref->count))
524 debug_active_activate(ref);
525 spin_unlock_irq(&ref->tree_lock);
526}
527
528int i915_active_acquire(struct i915_active *ref)
529{
530 int err;
531
532 if (i915_active_acquire_if_busy(ref))
533 return 0;
534
535 if (!ref->active) {
536 __i915_active_activate(ref);
537 return 0;
538 }
539
540 err = mutex_lock_interruptible(&ref->mutex);
541 if (err)
542 return err;
543
544 if (likely(!i915_active_acquire_if_busy(ref))) {
545 err = ref->active(ref);
546 if (!err)
547 __i915_active_activate(ref);
548 }
549
550 mutex_unlock(&ref->mutex);
551
552 return err;
553}
554
555int i915_active_acquire_for_context(struct i915_active *ref, u64 idx)
556{
557 struct i915_active_fence *active;
558 int err;
559
560 err = i915_active_acquire(ref);
561 if (err)
562 return err;
563
564 active = active_instance(ref, idx);
565 if (!active) {
566 i915_active_release(ref);
567 return -ENOMEM;
568 }
569
570 return 0; /* return with active ref */
571}
572
573void i915_active_release(struct i915_active *ref)
574{
575 debug_active_assert(ref);
576 active_retire(ref);
577}
578
579static void enable_signaling(struct i915_active_fence *active)
580{
581 struct dma_fence *fence;
582
583 if (unlikely(is_barrier(active)))
584 return;
585
586 fence = i915_active_fence_get(active);
587 if (!fence)
588 return;
589
590 dma_fence_enable_sw_signaling(fence);
591 dma_fence_put(fence);
592}
593
594static int flush_barrier(struct active_node *it)
595{
596 struct intel_engine_cs *engine;
597
598 if (likely(!is_barrier(&it->base)))
599 return 0;
600
601 engine = __barrier_to_engine(it);
602 smp_rmb(); /* serialise with add_active_barriers */
603 if (!is_barrier(&it->base))
604 return 0;
605
606 return intel_engine_flush_barriers(engine);
607}
608
609static int flush_lazy_signals(struct i915_active *ref)
610{
611 struct active_node *it, *n;
612 int err = 0;
613
614 enable_signaling(&ref->excl);
615 rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
616 err = flush_barrier(it); /* unconnected idle barrier? */
617 if (err)
618 break;
619
620 enable_signaling(&it->base);
621 }
622
623 return err;
624}
625
626int __i915_active_wait(struct i915_active *ref, int state)
627{
628 might_sleep();
629
630 /* Any fence added after the wait begins will not be auto-signaled */
631 if (i915_active_acquire_if_busy(ref)) {
632 int err;
633
634 err = flush_lazy_signals(ref);
635 i915_active_release(ref);
636 if (err)
637 return err;
638
639 if (___wait_var_event(ref, i915_active_is_idle(ref),
640 state, 0, 0, schedule()))
641 return -EINTR;
642 }
643
644 /*
645 * After the wait is complete, the caller may free the active.
646 * We have to flush any concurrent retirement before returning.
647 */
648 flush_work(&ref->work);
649 return 0;
650}
651
652static int __await_active(struct i915_active_fence *active,
653 int (*fn)(void *arg, struct dma_fence *fence),
654 void *arg)
655{
656 struct dma_fence *fence;
657
658 if (is_barrier(active)) /* XXX flush the barrier? */
659 return 0;
660
661 fence = i915_active_fence_get(active);
662 if (fence) {
663 int err;
664
665 err = fn(arg, fence);
666 dma_fence_put(fence);
667 if (err < 0)
668 return err;
669 }
670
671 return 0;
672}
673
674struct wait_barrier {
675 struct wait_queue_entry base;
676 struct i915_active *ref;
677};
678
679static int
680barrier_wake(wait_queue_entry_t *wq, unsigned int mode, int flags, void *key)
681{
682 struct wait_barrier *wb = container_of(wq, typeof(*wb), base);
683
684 if (i915_active_is_idle(wb->ref)) {
685 list_del(&wq->entry);
686 i915_sw_fence_complete(wq->private);
687 kfree(wq);
688 }
689
690 return 0;
691}
692
693static int __await_barrier(struct i915_active *ref, struct i915_sw_fence *fence)
694{
695 struct wait_barrier *wb;
696
697 wb = kmalloc(sizeof(*wb), GFP_KERNEL);
698 if (unlikely(!wb))
699 return -ENOMEM;
700
701 GEM_BUG_ON(i915_active_is_idle(ref));
702 if (!i915_sw_fence_await(fence)) {
703 kfree(wb);
704 return -EINVAL;
705 }
706
707 wb->base.flags = 0;
708 wb->base.func = barrier_wake;
709 wb->base.private = fence;
710 wb->ref = ref;
711
712 add_wait_queue(__var_waitqueue(ref), &wb->base);
713 return 0;
714}
715
716static int await_active(struct i915_active *ref,
717 unsigned int flags,
718 int (*fn)(void *arg, struct dma_fence *fence),
719 void *arg, struct i915_sw_fence *barrier)
720{
721 int err = 0;
722
723 if (!i915_active_acquire_if_busy(ref))
724 return 0;
725
726 if (flags & I915_ACTIVE_AWAIT_EXCL &&
727 rcu_access_pointer(ref->excl.fence)) {
728 err = __await_active(&ref->excl, fn, arg);
729 if (err)
730 goto out;
731 }
732
733 if (flags & I915_ACTIVE_AWAIT_ACTIVE) {
734 struct active_node *it, *n;
735
736 rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
737 err = __await_active(&it->base, fn, arg);
738 if (err)
739 goto out;
740 }
741 }
742
743 if (flags & I915_ACTIVE_AWAIT_BARRIER) {
744 err = flush_lazy_signals(ref);
745 if (err)
746 goto out;
747
748 err = __await_barrier(ref, barrier);
749 if (err)
750 goto out;
751 }
752
753out:
754 i915_active_release(ref);
755 return err;
756}
757
758static int rq_await_fence(void *arg, struct dma_fence *fence)
759{
760 return i915_request_await_dma_fence(arg, fence);
761}
762
763int i915_request_await_active(struct i915_request *rq,
764 struct i915_active *ref,
765 unsigned int flags)
766{
767 return await_active(ref, flags, rq_await_fence, rq, &rq->submit);
768}
769
770static int sw_await_fence(void *arg, struct dma_fence *fence)
771{
772 return i915_sw_fence_await_dma_fence(arg, fence, 0,
773 GFP_NOWAIT | __GFP_NOWARN);
774}
775
776int i915_sw_fence_await_active(struct i915_sw_fence *fence,
777 struct i915_active *ref,
778 unsigned int flags)
779{
780 return await_active(ref, flags, sw_await_fence, fence, fence);
781}
782
783void i915_active_fini(struct i915_active *ref)
784{
785 debug_active_fini(ref);
786 GEM_BUG_ON(atomic_read(&ref->count));
787 GEM_BUG_ON(work_pending(&ref->work));
788 mutex_destroy(&ref->mutex);
789
790 if (ref->cache)
791 kmem_cache_free(global.slab_cache, ref->cache);
792}
793
794static inline bool is_idle_barrier(struct active_node *node, u64 idx)
795{
796 return node->timeline == idx && !i915_active_fence_isset(&node->base);
797}
798
799static struct active_node *reuse_idle_barrier(struct i915_active *ref, u64 idx)
800{
801 struct rb_node *prev, *p;
802
803 if (RB_EMPTY_ROOT(&ref->tree))
804 return NULL;
805
806 GEM_BUG_ON(i915_active_is_idle(ref));
807
808 /*
809 * Try to reuse any existing barrier nodes already allocated for this
810 * i915_active, due to overlapping active phases there is likely a
811 * node kept alive (as we reuse before parking). We prefer to reuse
812 * completely idle barriers (less hassle in manipulating the llists),
813 * but otherwise any will do.
814 */
815 if (ref->cache && is_idle_barrier(ref->cache, idx)) {
816 p = &ref->cache->node;
817 goto match;
818 }
819
820 prev = NULL;
821 p = ref->tree.rb_node;
822 while (p) {
823 struct active_node *node =
824 rb_entry(p, struct active_node, node);
825
826 if (is_idle_barrier(node, idx))
827 goto match;
828
829 prev = p;
830 if (node->timeline < idx)
831 p = READ_ONCE(p->rb_right);
832 else
833 p = READ_ONCE(p->rb_left);
834 }
835
836 /*
837 * No quick match, but we did find the leftmost rb_node for the
838 * kernel_context. Walk the rb_tree in-order to see if there were
839 * any idle-barriers on this timeline that we missed, or just use
840 * the first pending barrier.
841 */
842 for (p = prev; p; p = rb_next(p)) {
843 struct active_node *node =
844 rb_entry(p, struct active_node, node);
845 struct intel_engine_cs *engine;
846
847 if (node->timeline > idx)
848 break;
849
850 if (node->timeline < idx)
851 continue;
852
853 if (is_idle_barrier(node, idx))
854 goto match;
855
856 /*
857 * The list of pending barriers is protected by the
858 * kernel_context timeline, which notably we do not hold
859 * here. i915_request_add_active_barriers() may consume
860 * the barrier before we claim it, so we have to check
861 * for success.
862 */
863 engine = __barrier_to_engine(node);
864 smp_rmb(); /* serialise with add_active_barriers */
865 if (is_barrier(&node->base) &&
866 ____active_del_barrier(ref, node, engine))
867 goto match;
868 }
869
870 return NULL;
871
872match:
873 spin_lock_irq(&ref->tree_lock);
874 rb_erase(p, &ref->tree); /* Hide from waits and sibling allocations */
875 if (p == &ref->cache->node)
876 WRITE_ONCE(ref->cache, NULL);
877 spin_unlock_irq(&ref->tree_lock);
878
879 return rb_entry(p, struct active_node, node);
880}
881
882int i915_active_acquire_preallocate_barrier(struct i915_active *ref,
883 struct intel_engine_cs *engine)
884{
885 intel_engine_mask_t tmp, mask = engine->mask;
886 struct llist_node *first = NULL, *last = NULL;
887 struct intel_gt *gt = engine->gt;
888
889 GEM_BUG_ON(i915_active_is_idle(ref));
890
891 /* Wait until the previous preallocation is completed */
892 while (!llist_empty(&ref->preallocated_barriers))
893 cond_resched();
894
895 /*
896 * Preallocate a node for each physical engine supporting the target
897 * engine (remember virtual engines have more than one sibling).
898 * We can then use the preallocated nodes in
899 * i915_active_acquire_barrier()
900 */
901 GEM_BUG_ON(!mask);
902 for_each_engine_masked(engine, gt, mask, tmp) {
903 u64 idx = engine->kernel_context->timeline->fence_context;
904 struct llist_node *prev = first;
905 struct active_node *node;
906
907 rcu_read_lock();
908 node = reuse_idle_barrier(ref, idx);
909 rcu_read_unlock();
910 if (!node) {
911 node = kmem_cache_alloc(global.slab_cache, GFP_KERNEL);
912 if (!node)
913 goto unwind;
914
915 RCU_INIT_POINTER(node->base.fence, NULL);
916 node->base.cb.func = node_retire;
917 node->timeline = idx;
918 node->ref = ref;
919 }
920
921 if (!i915_active_fence_isset(&node->base)) {
922 /*
923 * Mark this as being *our* unconnected proto-node.
924 *
925 * Since this node is not in any list, and we have
926 * decoupled it from the rbtree, we can reuse the
927 * request to indicate this is an idle-barrier node
928 * and then we can use the rb_node and list pointers
929 * for our tracking of the pending barrier.
930 */
931 RCU_INIT_POINTER(node->base.fence, ERR_PTR(-EAGAIN));
932 node->base.cb.node.prev = (void *)engine;
933 __i915_active_acquire(ref);
934 }
935 GEM_BUG_ON(rcu_access_pointer(node->base.fence) != ERR_PTR(-EAGAIN));
936
937 GEM_BUG_ON(barrier_to_engine(node) != engine);
938 first = barrier_to_ll(node);
939 first->next = prev;
940 if (!last)
941 last = first;
942 intel_engine_pm_get(engine);
943 }
944
945 GEM_BUG_ON(!llist_empty(&ref->preallocated_barriers));
946 llist_add_batch(first, last, &ref->preallocated_barriers);
947
948 return 0;
949
950unwind:
951 while (first) {
952 struct active_node *node = barrier_from_ll(first);
953
954 first = first->next;
955
956 atomic_dec(&ref->count);
957 intel_engine_pm_put(barrier_to_engine(node));
958
959 kmem_cache_free(global.slab_cache, node);
960 }
961 return -ENOMEM;
962}
963
964void i915_active_acquire_barrier(struct i915_active *ref)
965{
966 struct llist_node *pos, *next;
967 unsigned long flags;
968
969 GEM_BUG_ON(i915_active_is_idle(ref));
970
971 /*
972 * Transfer the list of preallocated barriers into the
973 * i915_active rbtree, but only as proto-nodes. They will be
974 * populated by i915_request_add_active_barriers() to point to the
975 * request that will eventually release them.
976 */
977 llist_for_each_safe(pos, next, take_preallocated_barriers(ref)) {
978 struct active_node *node = barrier_from_ll(pos);
979 struct intel_engine_cs *engine = barrier_to_engine(node);
980 struct rb_node **p, *parent;
981
982 spin_lock_irqsave_nested(&ref->tree_lock, flags,
983 SINGLE_DEPTH_NESTING);
984 parent = NULL;
985 p = &ref->tree.rb_node;
986 while (*p) {
987 struct active_node *it;
988
989 parent = *p;
990
991 it = rb_entry(parent, struct active_node, node);
992 if (it->timeline < node->timeline)
993 p = &parent->rb_right;
994 else
995 p = &parent->rb_left;
996 }
997 rb_link_node(&node->node, parent, p);
998 rb_insert_color(&node->node, &ref->tree);
999 spin_unlock_irqrestore(&ref->tree_lock, flags);
1000
1001 GEM_BUG_ON(!intel_engine_pm_is_awake(engine));
1002 llist_add(barrier_to_ll(node), &engine->barrier_tasks);
1003 intel_engine_pm_put_delay(engine, 1);
1004 }
1005}
1006
1007static struct dma_fence **ll_to_fence_slot(struct llist_node *node)
1008{
1009 return __active_fence_slot(&barrier_from_ll(node)->base);
1010}
1011
1012void i915_request_add_active_barriers(struct i915_request *rq)
1013{
1014 struct intel_engine_cs *engine = rq->engine;
1015 struct llist_node *node, *next;
1016 unsigned long flags;
1017
1018 GEM_BUG_ON(!intel_context_is_barrier(rq->context));
1019 GEM_BUG_ON(intel_engine_is_virtual(engine));
1020 GEM_BUG_ON(i915_request_timeline(rq) != engine->kernel_context->timeline);
1021
1022 node = llist_del_all(&engine->barrier_tasks);
1023 if (!node)
1024 return;
1025 /*
1026 * Attach the list of proto-fences to the in-flight request such
1027 * that the parent i915_active will be released when this request
1028 * is retired.
1029 */
1030 spin_lock_irqsave(&rq->lock, flags);
1031 llist_for_each_safe(node, next, node) {
1032 /* serialise with reuse_idle_barrier */
1033 smp_store_mb(*ll_to_fence_slot(node), &rq->fence);
1034 list_add_tail((struct list_head *)node, &rq->fence.cb_list);
1035 }
1036 spin_unlock_irqrestore(&rq->lock, flags);
1037}
1038
1039/*
1040 * __i915_active_fence_set: Update the last active fence along its timeline
1041 * @active: the active tracker
1042 * @fence: the new fence (under construction)
1043 *
1044 * Records the new @fence as the last active fence along its timeline in
1045 * this active tracker, moving the tracking callbacks from the previous
1046 * fence onto this one. Returns the previous fence (if not already completed),
1047 * which the caller must ensure is executed before the new fence. To ensure
1048 * that the order of fences within the timeline of the i915_active_fence is
1049 * understood, it should be locked by the caller.
1050 */
1051struct dma_fence *
1052__i915_active_fence_set(struct i915_active_fence *active,
1053 struct dma_fence *fence)
1054{
1055 struct dma_fence *prev;
1056 unsigned long flags;
1057
1058 if (fence == rcu_access_pointer(active->fence))
1059 return fence;
1060
1061 GEM_BUG_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags));
1062
1063 /*
1064 * Consider that we have two threads arriving (A and B), with
1065 * C already resident as the active->fence.
1066 *
1067 * A does the xchg first, and so it sees C or NULL depending
1068 * on the timing of the interrupt handler. If it is NULL, the
1069 * previous fence must have been signaled and we know that
1070 * we are first on the timeline. If it is still present,
1071 * we acquire the lock on that fence and serialise with the interrupt
1072 * handler, in the process removing it from any future interrupt
1073 * callback. A will then wait on C before executing (if present).
1074 *
1075 * As B is second, it sees A as the previous fence and so waits for
1076 * it to complete its transition and takes over the occupancy for
1077 * itself -- remembering that it needs to wait on A before executing.
1078 *
1079 * Note the strong ordering of the timeline also provides consistent
1080 * nesting rules for the fence->lock; the inner lock is always the
1081 * older lock.
1082 */
1083 spin_lock_irqsave(fence->lock, flags);
1084 prev = xchg(__active_fence_slot(active), fence);
1085 if (prev) {
1086 GEM_BUG_ON(prev == fence);
1087 spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING);
1088 __list_del_entry(&active->cb.node);
1089 spin_unlock(prev->lock); /* serialise with prev->cb_list */
1090 }
1091 list_add_tail(&active->cb.node, &fence->cb_list);
1092 spin_unlock_irqrestore(fence->lock, flags);
1093
1094 return prev;
1095}
1096
1097int i915_active_fence_set(struct i915_active_fence *active,
1098 struct i915_request *rq)
1099{
1100 struct dma_fence *fence;
1101 int err = 0;
1102
1103 /* Must maintain timeline ordering wrt previous active requests */
1104 rcu_read_lock();
1105 fence = __i915_active_fence_set(active, &rq->fence);
1106 if (fence) /* but the previous fence may not belong to that timeline! */
1107 fence = dma_fence_get_rcu(fence);
1108 rcu_read_unlock();
1109 if (fence) {
1110 err = i915_request_await_dma_fence(rq, fence);
1111 dma_fence_put(fence);
1112 }
1113
1114 return err;
1115}
1116
1117void i915_active_noop(struct dma_fence *fence, struct dma_fence_cb *cb)
1118{
1119 active_fence_cb(fence, cb);
1120}
1121
1122struct auto_active {
1123 struct i915_active base;
1124 struct kref ref;
1125};
1126
1127struct i915_active *i915_active_get(struct i915_active *ref)
1128{
1129 struct auto_active *aa = container_of(ref, typeof(*aa), base);
1130
1131 kref_get(&aa->ref);
1132 return &aa->base;
1133}
1134
1135static void auto_release(struct kref *ref)
1136{
1137 struct auto_active *aa = container_of(ref, typeof(*aa), ref);
1138
1139 i915_active_fini(&aa->base);
1140 kfree(aa);
1141}
1142
1143void i915_active_put(struct i915_active *ref)
1144{
1145 struct auto_active *aa = container_of(ref, typeof(*aa), base);
1146
1147 kref_put(&aa->ref, auto_release);
1148}
1149
1150static int auto_active(struct i915_active *ref)
1151{
1152 i915_active_get(ref);
1153 return 0;
1154}
1155
1156static void auto_retire(struct i915_active *ref)
1157{
1158 i915_active_put(ref);
1159}
1160
1161struct i915_active *i915_active_create(void)
1162{
1163 struct auto_active *aa;
1164
1165 aa = kmalloc(sizeof(*aa), GFP_KERNEL);
1166 if (!aa)
1167 return NULL;
1168
1169 kref_init(&aa->ref);
1170 i915_active_init(&aa->base, auto_active, auto_retire, 0);
1171
1172 return &aa->base;
1173}
1174
1175#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1176#include "selftests/i915_active.c"
1177#endif
1178
1179static void i915_global_active_shrink(void)
1180{
1181 kmem_cache_shrink(global.slab_cache);
1182}
1183
1184static void i915_global_active_exit(void)
1185{
1186 kmem_cache_destroy(global.slab_cache);
1187}
1188
1189static struct i915_global_active global = { {
1190 .shrink = i915_global_active_shrink,
1191 .exit = i915_global_active_exit,
1192} };
1193
1194int __init i915_global_active_init(void)
1195{
1196 global.slab_cache = KMEM_CACHE(active_node, SLAB_HWCACHE_ALIGN);
1197 if (!global.slab_cache)
1198 return -ENOMEM;
1199
1200 i915_global_register(&global.base);
1201 return 0;
1202}
1/*
2 * SPDX-License-Identifier: MIT
3 *
4 * Copyright © 2019 Intel Corporation
5 */
6
7#include <linux/debugobjects.h>
8
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"
13
14#include "i915_drv.h"
15#include "i915_active.h"
16
17/*
18 * Active refs memory management
19 *
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.
23 */
24static struct kmem_cache *slab_cache;
25
26struct active_node {
27 struct rb_node node;
28 struct i915_active_fence base;
29 struct i915_active *ref;
30 u64 timeline;
31};
32
33#define fetch_node(x) rb_entry(READ_ONCE(x), typeof(struct active_node), node)
34
35static inline struct active_node *
36node_from_active(struct i915_active_fence *active)
37{
38 return container_of(active, struct active_node, base);
39}
40
41#define take_preallocated_barriers(x) llist_del_all(&(x)->preallocated_barriers)
42
43static inline bool is_barrier(const struct i915_active_fence *active)
44{
45 return IS_ERR(rcu_access_pointer(active->fence));
46}
47
48static inline struct llist_node *barrier_to_ll(struct active_node *node)
49{
50 GEM_BUG_ON(!is_barrier(&node->base));
51 return (struct llist_node *)&node->base.cb.node;
52}
53
54static inline struct intel_engine_cs *
55__barrier_to_engine(struct active_node *node)
56{
57 return (struct intel_engine_cs *)READ_ONCE(node->base.cb.node.prev);
58}
59
60static inline struct intel_engine_cs *
61barrier_to_engine(struct active_node *node)
62{
63 GEM_BUG_ON(!is_barrier(&node->base));
64 return __barrier_to_engine(node);
65}
66
67static inline struct active_node *barrier_from_ll(struct llist_node *x)
68{
69 return container_of((struct list_head *)x,
70 struct active_node, base.cb.node);
71}
72
73#if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM) && IS_ENABLED(CONFIG_DEBUG_OBJECTS)
74
75static void *active_debug_hint(void *addr)
76{
77 struct i915_active *ref = addr;
78
79 return (void *)ref->active ?: (void *)ref->retire ?: (void *)ref;
80}
81
82static const struct debug_obj_descr active_debug_desc = {
83 .name = "i915_active",
84 .debug_hint = active_debug_hint,
85};
86
87static void debug_active_init(struct i915_active *ref)
88{
89 debug_object_init(ref, &active_debug_desc);
90}
91
92static void debug_active_activate(struct i915_active *ref)
93{
94 lockdep_assert_held(&ref->tree_lock);
95 if (!atomic_read(&ref->count)) /* before the first inc */
96 debug_object_activate(ref, &active_debug_desc);
97}
98
99static void debug_active_deactivate(struct i915_active *ref)
100{
101 lockdep_assert_held(&ref->tree_lock);
102 if (!atomic_read(&ref->count)) /* after the last dec */
103 debug_object_deactivate(ref, &active_debug_desc);
104}
105
106static void debug_active_fini(struct i915_active *ref)
107{
108 debug_object_free(ref, &active_debug_desc);
109}
110
111static void debug_active_assert(struct i915_active *ref)
112{
113 debug_object_assert_init(ref, &active_debug_desc);
114}
115
116#else
117
118static inline void debug_active_init(struct i915_active *ref) { }
119static inline void debug_active_activate(struct i915_active *ref) { }
120static inline void debug_active_deactivate(struct i915_active *ref) { }
121static inline void debug_active_fini(struct i915_active *ref) { }
122static inline void debug_active_assert(struct i915_active *ref) { }
123
124#endif
125
126static void
127__active_retire(struct i915_active *ref)
128{
129 struct rb_root root = RB_ROOT;
130 struct active_node *it, *n;
131 unsigned long flags;
132
133 GEM_BUG_ON(i915_active_is_idle(ref));
134
135 /* return the unused nodes to our slabcache -- flushing the allocator */
136 if (!atomic_dec_and_lock_irqsave(&ref->count, &ref->tree_lock, flags))
137 return;
138
139 GEM_BUG_ON(rcu_access_pointer(ref->excl.fence));
140 debug_active_deactivate(ref);
141
142 /* Even if we have not used the cache, we may still have a barrier */
143 if (!ref->cache)
144 ref->cache = fetch_node(ref->tree.rb_node);
145
146 /* Keep the MRU cached node for reuse */
147 if (ref->cache) {
148 /* Discard all other nodes in the tree */
149 rb_erase(&ref->cache->node, &ref->tree);
150 root = ref->tree;
151
152 /* Rebuild the tree with only the cached node */
153 rb_link_node(&ref->cache->node, NULL, &ref->tree.rb_node);
154 rb_insert_color(&ref->cache->node, &ref->tree);
155 GEM_BUG_ON(ref->tree.rb_node != &ref->cache->node);
156
157 /* Make the cached node available for reuse with any timeline */
158 ref->cache->timeline = 0; /* needs cmpxchg(u64) */
159 }
160
161 spin_unlock_irqrestore(&ref->tree_lock, flags);
162
163 /* After the final retire, the entire struct may be freed */
164 if (ref->retire)
165 ref->retire(ref);
166
167 /* ... except if you wait on it, you must manage your own references! */
168 wake_up_var(ref);
169
170 /* Finally free the discarded timeline tree */
171 rbtree_postorder_for_each_entry_safe(it, n, &root, node) {
172 GEM_BUG_ON(i915_active_fence_isset(&it->base));
173 kmem_cache_free(slab_cache, it);
174 }
175}
176
177static void
178active_work(struct work_struct *wrk)
179{
180 struct i915_active *ref = container_of(wrk, typeof(*ref), work);
181
182 GEM_BUG_ON(!atomic_read(&ref->count));
183 if (atomic_add_unless(&ref->count, -1, 1))
184 return;
185
186 __active_retire(ref);
187}
188
189static void
190active_retire(struct i915_active *ref)
191{
192 GEM_BUG_ON(!atomic_read(&ref->count));
193 if (atomic_add_unless(&ref->count, -1, 1))
194 return;
195
196 if (ref->flags & I915_ACTIVE_RETIRE_SLEEPS) {
197 queue_work(system_unbound_wq, &ref->work);
198 return;
199 }
200
201 __active_retire(ref);
202}
203
204static inline struct dma_fence **
205__active_fence_slot(struct i915_active_fence *active)
206{
207 return (struct dma_fence ** __force)&active->fence;
208}
209
210static inline bool
211active_fence_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
212{
213 struct i915_active_fence *active =
214 container_of(cb, typeof(*active), cb);
215
216 return cmpxchg(__active_fence_slot(active), fence, NULL) == fence;
217}
218
219static void
220node_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
221{
222 if (active_fence_cb(fence, cb))
223 active_retire(container_of(cb, struct active_node, base.cb)->ref);
224}
225
226static void
227excl_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
228{
229 if (active_fence_cb(fence, cb))
230 active_retire(container_of(cb, struct i915_active, excl.cb));
231}
232
233static struct active_node *__active_lookup(struct i915_active *ref, u64 idx)
234{
235 struct active_node *it;
236
237 GEM_BUG_ON(idx == 0); /* 0 is the unordered timeline, rsvd for cache */
238
239 /*
240 * We track the most recently used timeline to skip a rbtree search
241 * for the common case, under typical loads we never need the rbtree
242 * at all. We can reuse the last slot if it is empty, that is
243 * after the previous activity has been retired, or if it matches the
244 * current timeline.
245 */
246 it = READ_ONCE(ref->cache);
247 if (it) {
248 u64 cached = READ_ONCE(it->timeline);
249
250 /* Once claimed, this slot will only belong to this idx */
251 if (cached == idx)
252 return it;
253
254 /*
255 * An unclaimed cache [.timeline=0] can only be claimed once.
256 *
257 * If the value is already non-zero, some other thread has
258 * claimed the cache and we know that is does not match our
259 * idx. If, and only if, the timeline is currently zero is it
260 * worth competing to claim it atomically for ourselves (for
261 * only the winner of that race will cmpxchg return the old
262 * value of 0).
263 */
264 if (!cached && !cmpxchg64(&it->timeline, 0, idx))
265 return it;
266 }
267
268 BUILD_BUG_ON(offsetof(typeof(*it), node));
269
270 /* While active, the tree can only be built; not destroyed */
271 GEM_BUG_ON(i915_active_is_idle(ref));
272
273 it = fetch_node(ref->tree.rb_node);
274 while (it) {
275 if (it->timeline < idx) {
276 it = fetch_node(it->node.rb_right);
277 } else if (it->timeline > idx) {
278 it = fetch_node(it->node.rb_left);
279 } else {
280 WRITE_ONCE(ref->cache, it);
281 break;
282 }
283 }
284
285 /* NB: If the tree rotated beneath us, we may miss our target. */
286 return it;
287}
288
289static struct i915_active_fence *
290active_instance(struct i915_active *ref, u64 idx)
291{
292 struct active_node *node;
293 struct rb_node **p, *parent;
294
295 node = __active_lookup(ref, idx);
296 if (likely(node))
297 return &node->base;
298
299 spin_lock_irq(&ref->tree_lock);
300 GEM_BUG_ON(i915_active_is_idle(ref));
301
302 parent = NULL;
303 p = &ref->tree.rb_node;
304 while (*p) {
305 parent = *p;
306
307 node = rb_entry(parent, struct active_node, node);
308 if (node->timeline == idx)
309 goto out;
310
311 if (node->timeline < idx)
312 p = &parent->rb_right;
313 else
314 p = &parent->rb_left;
315 }
316
317 /*
318 * XXX: We should preallocate this before i915_active_ref() is ever
319 * called, but we cannot call into fs_reclaim() anyway, so use GFP_ATOMIC.
320 */
321 node = kmem_cache_alloc(slab_cache, GFP_ATOMIC);
322 if (!node)
323 goto out;
324
325 __i915_active_fence_init(&node->base, NULL, node_retire);
326 node->ref = ref;
327 node->timeline = idx;
328
329 rb_link_node(&node->node, parent, p);
330 rb_insert_color(&node->node, &ref->tree);
331
332out:
333 WRITE_ONCE(ref->cache, node);
334 spin_unlock_irq(&ref->tree_lock);
335
336 return &node->base;
337}
338
339void __i915_active_init(struct i915_active *ref,
340 int (*active)(struct i915_active *ref),
341 void (*retire)(struct i915_active *ref),
342 unsigned long flags,
343 struct lock_class_key *mkey,
344 struct lock_class_key *wkey)
345{
346 debug_active_init(ref);
347
348 ref->flags = flags;
349 ref->active = active;
350 ref->retire = retire;
351
352 spin_lock_init(&ref->tree_lock);
353 ref->tree = RB_ROOT;
354 ref->cache = NULL;
355
356 init_llist_head(&ref->preallocated_barriers);
357 atomic_set(&ref->count, 0);
358 __mutex_init(&ref->mutex, "i915_active", mkey);
359 __i915_active_fence_init(&ref->excl, NULL, excl_retire);
360 INIT_WORK(&ref->work, active_work);
361#if IS_ENABLED(CONFIG_LOCKDEP)
362 lockdep_init_map(&ref->work.lockdep_map, "i915_active.work", wkey, 0);
363#endif
364}
365
366static bool ____active_del_barrier(struct i915_active *ref,
367 struct active_node *node,
368 struct intel_engine_cs *engine)
369
370{
371 struct llist_node *head = NULL, *tail = NULL;
372 struct llist_node *pos, *next;
373
374 GEM_BUG_ON(node->timeline != engine->kernel_context->timeline->fence_context);
375
376 /*
377 * Rebuild the llist excluding our node. We may perform this
378 * outside of the kernel_context timeline mutex and so someone
379 * else may be manipulating the engine->barrier_tasks, in
380 * which case either we or they will be upset :)
381 *
382 * A second __active_del_barrier() will report failure to claim
383 * the active_node and the caller will just shrug and know not to
384 * claim ownership of its node.
385 *
386 * A concurrent i915_request_add_active_barriers() will miss adding
387 * any of the tasks, but we will try again on the next -- and since
388 * we are actively using the barrier, we know that there will be
389 * at least another opportunity when we idle.
390 */
391 llist_for_each_safe(pos, next, llist_del_all(&engine->barrier_tasks)) {
392 if (node == barrier_from_ll(pos)) {
393 node = NULL;
394 continue;
395 }
396
397 pos->next = head;
398 head = pos;
399 if (!tail)
400 tail = pos;
401 }
402 if (head)
403 llist_add_batch(head, tail, &engine->barrier_tasks);
404
405 return !node;
406}
407
408static bool
409__active_del_barrier(struct i915_active *ref, struct active_node *node)
410{
411 return ____active_del_barrier(ref, node, barrier_to_engine(node));
412}
413
414static bool
415replace_barrier(struct i915_active *ref, struct i915_active_fence *active)
416{
417 if (!is_barrier(active)) /* proto-node used by our idle barrier? */
418 return false;
419
420 /*
421 * This request is on the kernel_context timeline, and so
422 * we can use it to substitute for the pending idle-barrer
423 * request that we want to emit on the kernel_context.
424 */
425 __active_del_barrier(ref, node_from_active(active));
426 return true;
427}
428
429int i915_active_add_request(struct i915_active *ref, struct i915_request *rq)
430{
431 struct dma_fence *fence = &rq->fence;
432 struct i915_active_fence *active;
433 int err;
434
435 /* Prevent reaping in case we malloc/wait while building the tree */
436 err = i915_active_acquire(ref);
437 if (err)
438 return err;
439
440 active = active_instance(ref, i915_request_timeline(rq)->fence_context);
441 if (!active) {
442 err = -ENOMEM;
443 goto out;
444 }
445
446 if (replace_barrier(ref, active)) {
447 RCU_INIT_POINTER(active->fence, NULL);
448 atomic_dec(&ref->count);
449 }
450 if (!__i915_active_fence_set(active, fence))
451 __i915_active_acquire(ref);
452
453out:
454 i915_active_release(ref);
455 return err;
456}
457
458static struct dma_fence *
459__i915_active_set_fence(struct i915_active *ref,
460 struct i915_active_fence *active,
461 struct dma_fence *fence)
462{
463 struct dma_fence *prev;
464
465 if (replace_barrier(ref, active)) {
466 RCU_INIT_POINTER(active->fence, fence);
467 return NULL;
468 }
469
470 rcu_read_lock();
471 prev = __i915_active_fence_set(active, fence);
472 if (prev)
473 prev = dma_fence_get_rcu(prev);
474 else
475 __i915_active_acquire(ref);
476 rcu_read_unlock();
477
478 return prev;
479}
480
481struct dma_fence *
482i915_active_set_exclusive(struct i915_active *ref, struct dma_fence *f)
483{
484 /* We expect the caller to manage the exclusive timeline ordering */
485 return __i915_active_set_fence(ref, &ref->excl, f);
486}
487
488bool i915_active_acquire_if_busy(struct i915_active *ref)
489{
490 debug_active_assert(ref);
491 return atomic_add_unless(&ref->count, 1, 0);
492}
493
494static void __i915_active_activate(struct i915_active *ref)
495{
496 spin_lock_irq(&ref->tree_lock); /* __active_retire() */
497 if (!atomic_fetch_inc(&ref->count))
498 debug_active_activate(ref);
499 spin_unlock_irq(&ref->tree_lock);
500}
501
502int i915_active_acquire(struct i915_active *ref)
503{
504 int err;
505
506 if (i915_active_acquire_if_busy(ref))
507 return 0;
508
509 if (!ref->active) {
510 __i915_active_activate(ref);
511 return 0;
512 }
513
514 err = mutex_lock_interruptible(&ref->mutex);
515 if (err)
516 return err;
517
518 if (likely(!i915_active_acquire_if_busy(ref))) {
519 err = ref->active(ref);
520 if (!err)
521 __i915_active_activate(ref);
522 }
523
524 mutex_unlock(&ref->mutex);
525
526 return err;
527}
528
529int i915_active_acquire_for_context(struct i915_active *ref, u64 idx)
530{
531 struct i915_active_fence *active;
532 int err;
533
534 err = i915_active_acquire(ref);
535 if (err)
536 return err;
537
538 active = active_instance(ref, idx);
539 if (!active) {
540 i915_active_release(ref);
541 return -ENOMEM;
542 }
543
544 return 0; /* return with active ref */
545}
546
547void i915_active_release(struct i915_active *ref)
548{
549 debug_active_assert(ref);
550 active_retire(ref);
551}
552
553static void enable_signaling(struct i915_active_fence *active)
554{
555 struct dma_fence *fence;
556
557 if (unlikely(is_barrier(active)))
558 return;
559
560 fence = i915_active_fence_get(active);
561 if (!fence)
562 return;
563
564 dma_fence_enable_sw_signaling(fence);
565 dma_fence_put(fence);
566}
567
568static int flush_barrier(struct active_node *it)
569{
570 struct intel_engine_cs *engine;
571
572 if (likely(!is_barrier(&it->base)))
573 return 0;
574
575 engine = __barrier_to_engine(it);
576 smp_rmb(); /* serialise with add_active_barriers */
577 if (!is_barrier(&it->base))
578 return 0;
579
580 return intel_engine_flush_barriers(engine);
581}
582
583static int flush_lazy_signals(struct i915_active *ref)
584{
585 struct active_node *it, *n;
586 int err = 0;
587
588 enable_signaling(&ref->excl);
589 rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
590 err = flush_barrier(it); /* unconnected idle barrier? */
591 if (err)
592 break;
593
594 enable_signaling(&it->base);
595 }
596
597 return err;
598}
599
600int __i915_active_wait(struct i915_active *ref, int state)
601{
602 might_sleep();
603
604 /* Any fence added after the wait begins will not be auto-signaled */
605 if (i915_active_acquire_if_busy(ref)) {
606 int err;
607
608 err = flush_lazy_signals(ref);
609 i915_active_release(ref);
610 if (err)
611 return err;
612
613 if (___wait_var_event(ref, i915_active_is_idle(ref),
614 state, 0, 0, schedule()))
615 return -EINTR;
616 }
617
618 /*
619 * After the wait is complete, the caller may free the active.
620 * We have to flush any concurrent retirement before returning.
621 */
622 flush_work(&ref->work);
623 return 0;
624}
625
626static int __await_active(struct i915_active_fence *active,
627 int (*fn)(void *arg, struct dma_fence *fence),
628 void *arg)
629{
630 struct dma_fence *fence;
631
632 if (is_barrier(active)) /* XXX flush the barrier? */
633 return 0;
634
635 fence = i915_active_fence_get(active);
636 if (fence) {
637 int err;
638
639 err = fn(arg, fence);
640 dma_fence_put(fence);
641 if (err < 0)
642 return err;
643 }
644
645 return 0;
646}
647
648struct wait_barrier {
649 struct wait_queue_entry base;
650 struct i915_active *ref;
651};
652
653static int
654barrier_wake(wait_queue_entry_t *wq, unsigned int mode, int flags, void *key)
655{
656 struct wait_barrier *wb = container_of(wq, typeof(*wb), base);
657
658 if (i915_active_is_idle(wb->ref)) {
659 list_del(&wq->entry);
660 i915_sw_fence_complete(wq->private);
661 kfree(wq);
662 }
663
664 return 0;
665}
666
667static int __await_barrier(struct i915_active *ref, struct i915_sw_fence *fence)
668{
669 struct wait_barrier *wb;
670
671 wb = kmalloc(sizeof(*wb), GFP_KERNEL);
672 if (unlikely(!wb))
673 return -ENOMEM;
674
675 GEM_BUG_ON(i915_active_is_idle(ref));
676 if (!i915_sw_fence_await(fence)) {
677 kfree(wb);
678 return -EINVAL;
679 }
680
681 wb->base.flags = 0;
682 wb->base.func = barrier_wake;
683 wb->base.private = fence;
684 wb->ref = ref;
685
686 add_wait_queue(__var_waitqueue(ref), &wb->base);
687 return 0;
688}
689
690static int await_active(struct i915_active *ref,
691 unsigned int flags,
692 int (*fn)(void *arg, struct dma_fence *fence),
693 void *arg, struct i915_sw_fence *barrier)
694{
695 int err = 0;
696
697 if (!i915_active_acquire_if_busy(ref))
698 return 0;
699
700 if (flags & I915_ACTIVE_AWAIT_EXCL &&
701 rcu_access_pointer(ref->excl.fence)) {
702 err = __await_active(&ref->excl, fn, arg);
703 if (err)
704 goto out;
705 }
706
707 if (flags & I915_ACTIVE_AWAIT_ACTIVE) {
708 struct active_node *it, *n;
709
710 rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
711 err = __await_active(&it->base, fn, arg);
712 if (err)
713 goto out;
714 }
715 }
716
717 if (flags & I915_ACTIVE_AWAIT_BARRIER) {
718 err = flush_lazy_signals(ref);
719 if (err)
720 goto out;
721
722 err = __await_barrier(ref, barrier);
723 if (err)
724 goto out;
725 }
726
727out:
728 i915_active_release(ref);
729 return err;
730}
731
732static int rq_await_fence(void *arg, struct dma_fence *fence)
733{
734 return i915_request_await_dma_fence(arg, fence);
735}
736
737int i915_request_await_active(struct i915_request *rq,
738 struct i915_active *ref,
739 unsigned int flags)
740{
741 return await_active(ref, flags, rq_await_fence, rq, &rq->submit);
742}
743
744static int sw_await_fence(void *arg, struct dma_fence *fence)
745{
746 return i915_sw_fence_await_dma_fence(arg, fence, 0,
747 GFP_NOWAIT | __GFP_NOWARN);
748}
749
750int i915_sw_fence_await_active(struct i915_sw_fence *fence,
751 struct i915_active *ref,
752 unsigned int flags)
753{
754 return await_active(ref, flags, sw_await_fence, fence, fence);
755}
756
757void i915_active_fini(struct i915_active *ref)
758{
759 debug_active_fini(ref);
760 GEM_BUG_ON(atomic_read(&ref->count));
761 GEM_BUG_ON(work_pending(&ref->work));
762 mutex_destroy(&ref->mutex);
763
764 if (ref->cache)
765 kmem_cache_free(slab_cache, ref->cache);
766}
767
768static inline bool is_idle_barrier(struct active_node *node, u64 idx)
769{
770 return node->timeline == idx && !i915_active_fence_isset(&node->base);
771}
772
773static struct active_node *reuse_idle_barrier(struct i915_active *ref, u64 idx)
774{
775 struct rb_node *prev, *p;
776
777 if (RB_EMPTY_ROOT(&ref->tree))
778 return NULL;
779
780 GEM_BUG_ON(i915_active_is_idle(ref));
781
782 /*
783 * Try to reuse any existing barrier nodes already allocated for this
784 * i915_active, due to overlapping active phases there is likely a
785 * node kept alive (as we reuse before parking). We prefer to reuse
786 * completely idle barriers (less hassle in manipulating the llists),
787 * but otherwise any will do.
788 */
789 if (ref->cache && is_idle_barrier(ref->cache, idx)) {
790 p = &ref->cache->node;
791 goto match;
792 }
793
794 prev = NULL;
795 p = ref->tree.rb_node;
796 while (p) {
797 struct active_node *node =
798 rb_entry(p, struct active_node, node);
799
800 if (is_idle_barrier(node, idx))
801 goto match;
802
803 prev = p;
804 if (node->timeline < idx)
805 p = READ_ONCE(p->rb_right);
806 else
807 p = READ_ONCE(p->rb_left);
808 }
809
810 /*
811 * No quick match, but we did find the leftmost rb_node for the
812 * kernel_context. Walk the rb_tree in-order to see if there were
813 * any idle-barriers on this timeline that we missed, or just use
814 * the first pending barrier.
815 */
816 for (p = prev; p; p = rb_next(p)) {
817 struct active_node *node =
818 rb_entry(p, struct active_node, node);
819 struct intel_engine_cs *engine;
820
821 if (node->timeline > idx)
822 break;
823
824 if (node->timeline < idx)
825 continue;
826
827 if (is_idle_barrier(node, idx))
828 goto match;
829
830 /*
831 * The list of pending barriers is protected by the
832 * kernel_context timeline, which notably we do not hold
833 * here. i915_request_add_active_barriers() may consume
834 * the barrier before we claim it, so we have to check
835 * for success.
836 */
837 engine = __barrier_to_engine(node);
838 smp_rmb(); /* serialise with add_active_barriers */
839 if (is_barrier(&node->base) &&
840 ____active_del_barrier(ref, node, engine))
841 goto match;
842 }
843
844 return NULL;
845
846match:
847 spin_lock_irq(&ref->tree_lock);
848 rb_erase(p, &ref->tree); /* Hide from waits and sibling allocations */
849 if (p == &ref->cache->node)
850 WRITE_ONCE(ref->cache, NULL);
851 spin_unlock_irq(&ref->tree_lock);
852
853 return rb_entry(p, struct active_node, node);
854}
855
856int i915_active_acquire_preallocate_barrier(struct i915_active *ref,
857 struct intel_engine_cs *engine)
858{
859 intel_engine_mask_t tmp, mask = engine->mask;
860 struct llist_node *first = NULL, *last = NULL;
861 struct intel_gt *gt = engine->gt;
862
863 GEM_BUG_ON(i915_active_is_idle(ref));
864
865 /* Wait until the previous preallocation is completed */
866 while (!llist_empty(&ref->preallocated_barriers))
867 cond_resched();
868
869 /*
870 * Preallocate a node for each physical engine supporting the target
871 * engine (remember virtual engines have more than one sibling).
872 * We can then use the preallocated nodes in
873 * i915_active_acquire_barrier()
874 */
875 GEM_BUG_ON(!mask);
876 for_each_engine_masked(engine, gt, mask, tmp) {
877 u64 idx = engine->kernel_context->timeline->fence_context;
878 struct llist_node *prev = first;
879 struct active_node *node;
880
881 rcu_read_lock();
882 node = reuse_idle_barrier(ref, idx);
883 rcu_read_unlock();
884 if (!node) {
885 node = kmem_cache_alloc(slab_cache, GFP_KERNEL);
886 if (!node)
887 goto unwind;
888
889 RCU_INIT_POINTER(node->base.fence, NULL);
890 node->base.cb.func = node_retire;
891 node->timeline = idx;
892 node->ref = ref;
893 }
894
895 if (!i915_active_fence_isset(&node->base)) {
896 /*
897 * Mark this as being *our* unconnected proto-node.
898 *
899 * Since this node is not in any list, and we have
900 * decoupled it from the rbtree, we can reuse the
901 * request to indicate this is an idle-barrier node
902 * and then we can use the rb_node and list pointers
903 * for our tracking of the pending barrier.
904 */
905 RCU_INIT_POINTER(node->base.fence, ERR_PTR(-EAGAIN));
906 node->base.cb.node.prev = (void *)engine;
907 __i915_active_acquire(ref);
908 }
909 GEM_BUG_ON(rcu_access_pointer(node->base.fence) != ERR_PTR(-EAGAIN));
910
911 GEM_BUG_ON(barrier_to_engine(node) != engine);
912 first = barrier_to_ll(node);
913 first->next = prev;
914 if (!last)
915 last = first;
916 intel_engine_pm_get(engine);
917 }
918
919 GEM_BUG_ON(!llist_empty(&ref->preallocated_barriers));
920 llist_add_batch(first, last, &ref->preallocated_barriers);
921
922 return 0;
923
924unwind:
925 while (first) {
926 struct active_node *node = barrier_from_ll(first);
927
928 first = first->next;
929
930 atomic_dec(&ref->count);
931 intel_engine_pm_put(barrier_to_engine(node));
932
933 kmem_cache_free(slab_cache, node);
934 }
935 return -ENOMEM;
936}
937
938void i915_active_acquire_barrier(struct i915_active *ref)
939{
940 struct llist_node *pos, *next;
941 unsigned long flags;
942
943 GEM_BUG_ON(i915_active_is_idle(ref));
944
945 /*
946 * Transfer the list of preallocated barriers into the
947 * i915_active rbtree, but only as proto-nodes. They will be
948 * populated by i915_request_add_active_barriers() to point to the
949 * request that will eventually release them.
950 */
951 llist_for_each_safe(pos, next, take_preallocated_barriers(ref)) {
952 struct active_node *node = barrier_from_ll(pos);
953 struct intel_engine_cs *engine = barrier_to_engine(node);
954 struct rb_node **p, *parent;
955
956 spin_lock_irqsave_nested(&ref->tree_lock, flags,
957 SINGLE_DEPTH_NESTING);
958 parent = NULL;
959 p = &ref->tree.rb_node;
960 while (*p) {
961 struct active_node *it;
962
963 parent = *p;
964
965 it = rb_entry(parent, struct active_node, node);
966 if (it->timeline < node->timeline)
967 p = &parent->rb_right;
968 else
969 p = &parent->rb_left;
970 }
971 rb_link_node(&node->node, parent, p);
972 rb_insert_color(&node->node, &ref->tree);
973 spin_unlock_irqrestore(&ref->tree_lock, flags);
974
975 GEM_BUG_ON(!intel_engine_pm_is_awake(engine));
976 llist_add(barrier_to_ll(node), &engine->barrier_tasks);
977 intel_engine_pm_put_delay(engine, 2);
978 }
979}
980
981static struct dma_fence **ll_to_fence_slot(struct llist_node *node)
982{
983 return __active_fence_slot(&barrier_from_ll(node)->base);
984}
985
986void i915_request_add_active_barriers(struct i915_request *rq)
987{
988 struct intel_engine_cs *engine = rq->engine;
989 struct llist_node *node, *next;
990 unsigned long flags;
991
992 GEM_BUG_ON(!intel_context_is_barrier(rq->context));
993 GEM_BUG_ON(intel_engine_is_virtual(engine));
994 GEM_BUG_ON(i915_request_timeline(rq) != engine->kernel_context->timeline);
995
996 node = llist_del_all(&engine->barrier_tasks);
997 if (!node)
998 return;
999 /*
1000 * Attach the list of proto-fences to the in-flight request such
1001 * that the parent i915_active will be released when this request
1002 * is retired.
1003 */
1004 spin_lock_irqsave(&rq->lock, flags);
1005 llist_for_each_safe(node, next, node) {
1006 /* serialise with reuse_idle_barrier */
1007 smp_store_mb(*ll_to_fence_slot(node), &rq->fence);
1008 list_add_tail((struct list_head *)node, &rq->fence.cb_list);
1009 }
1010 spin_unlock_irqrestore(&rq->lock, flags);
1011}
1012
1013/*
1014 * __i915_active_fence_set: Update the last active fence along its timeline
1015 * @active: the active tracker
1016 * @fence: the new fence (under construction)
1017 *
1018 * Records the new @fence as the last active fence along its timeline in
1019 * this active tracker, moving the tracking callbacks from the previous
1020 * fence onto this one. Returns the previous fence (if not already completed),
1021 * which the caller must ensure is executed before the new fence. To ensure
1022 * that the order of fences within the timeline of the i915_active_fence is
1023 * understood, it should be locked by the caller.
1024 */
1025struct dma_fence *
1026__i915_active_fence_set(struct i915_active_fence *active,
1027 struct dma_fence *fence)
1028{
1029 struct dma_fence *prev;
1030 unsigned long flags;
1031
1032 if (fence == rcu_access_pointer(active->fence))
1033 return fence;
1034
1035 GEM_BUG_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags));
1036
1037 /*
1038 * Consider that we have two threads arriving (A and B), with
1039 * C already resident as the active->fence.
1040 *
1041 * A does the xchg first, and so it sees C or NULL depending
1042 * on the timing of the interrupt handler. If it is NULL, the
1043 * previous fence must have been signaled and we know that
1044 * we are first on the timeline. If it is still present,
1045 * we acquire the lock on that fence and serialise with the interrupt
1046 * handler, in the process removing it from any future interrupt
1047 * callback. A will then wait on C before executing (if present).
1048 *
1049 * As B is second, it sees A as the previous fence and so waits for
1050 * it to complete its transition and takes over the occupancy for
1051 * itself -- remembering that it needs to wait on A before executing.
1052 *
1053 * Note the strong ordering of the timeline also provides consistent
1054 * nesting rules for the fence->lock; the inner lock is always the
1055 * older lock.
1056 */
1057 spin_lock_irqsave(fence->lock, flags);
1058 prev = xchg(__active_fence_slot(active), fence);
1059 if (prev) {
1060 GEM_BUG_ON(prev == fence);
1061 spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING);
1062 __list_del_entry(&active->cb.node);
1063 spin_unlock(prev->lock); /* serialise with prev->cb_list */
1064 }
1065 list_add_tail(&active->cb.node, &fence->cb_list);
1066 spin_unlock_irqrestore(fence->lock, flags);
1067
1068 return prev;
1069}
1070
1071int i915_active_fence_set(struct i915_active_fence *active,
1072 struct i915_request *rq)
1073{
1074 struct dma_fence *fence;
1075 int err = 0;
1076
1077 /* Must maintain timeline ordering wrt previous active requests */
1078 rcu_read_lock();
1079 fence = __i915_active_fence_set(active, &rq->fence);
1080 if (fence) /* but the previous fence may not belong to that timeline! */
1081 fence = dma_fence_get_rcu(fence);
1082 rcu_read_unlock();
1083 if (fence) {
1084 err = i915_request_await_dma_fence(rq, fence);
1085 dma_fence_put(fence);
1086 }
1087
1088 return err;
1089}
1090
1091void i915_active_noop(struct dma_fence *fence, struct dma_fence_cb *cb)
1092{
1093 active_fence_cb(fence, cb);
1094}
1095
1096struct auto_active {
1097 struct i915_active base;
1098 struct kref ref;
1099};
1100
1101struct i915_active *i915_active_get(struct i915_active *ref)
1102{
1103 struct auto_active *aa = container_of(ref, typeof(*aa), base);
1104
1105 kref_get(&aa->ref);
1106 return &aa->base;
1107}
1108
1109static void auto_release(struct kref *ref)
1110{
1111 struct auto_active *aa = container_of(ref, typeof(*aa), ref);
1112
1113 i915_active_fini(&aa->base);
1114 kfree(aa);
1115}
1116
1117void i915_active_put(struct i915_active *ref)
1118{
1119 struct auto_active *aa = container_of(ref, typeof(*aa), base);
1120
1121 kref_put(&aa->ref, auto_release);
1122}
1123
1124static int auto_active(struct i915_active *ref)
1125{
1126 i915_active_get(ref);
1127 return 0;
1128}
1129
1130static void auto_retire(struct i915_active *ref)
1131{
1132 i915_active_put(ref);
1133}
1134
1135struct i915_active *i915_active_create(void)
1136{
1137 struct auto_active *aa;
1138
1139 aa = kmalloc(sizeof(*aa), GFP_KERNEL);
1140 if (!aa)
1141 return NULL;
1142
1143 kref_init(&aa->ref);
1144 i915_active_init(&aa->base, auto_active, auto_retire, 0);
1145
1146 return &aa->base;
1147}
1148
1149#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1150#include "selftests/i915_active.c"
1151#endif
1152
1153void i915_active_module_exit(void)
1154{
1155 kmem_cache_destroy(slab_cache);
1156}
1157
1158int __init i915_active_module_init(void)
1159{
1160 slab_cache = KMEM_CACHE(active_node, SLAB_HWCACHE_ALIGN);
1161 if (!slab_cache)
1162 return -ENOMEM;
1163
1164 return 0;
1165}