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 */
24 static struct kmem_cache *slab_cache;
25
26 struct 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
35 static inline struct active_node *
node_from_active(struct i915_active_fence * active)36 node_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
is_barrier(const struct i915_active_fence * active)43 static inline bool is_barrier(const struct i915_active_fence *active)
44 {
45 return IS_ERR(rcu_access_pointer(active->fence));
46 }
47
barrier_to_ll(struct active_node * node)48 static 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
54 static inline struct intel_engine_cs *
__barrier_to_engine(struct active_node * node)55 __barrier_to_engine(struct active_node *node)
56 {
57 return (struct intel_engine_cs *)READ_ONCE(node->base.cb.node.prev);
58 }
59
60 static inline struct intel_engine_cs *
barrier_to_engine(struct active_node * node)61 barrier_to_engine(struct active_node *node)
62 {
63 GEM_BUG_ON(!is_barrier(&node->base));
64 return __barrier_to_engine(node);
65 }
66
barrier_from_ll(struct llist_node * x)67 static 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
active_debug_hint(void * addr)75 static 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
82 static const struct debug_obj_descr active_debug_desc = {
83 .name = "i915_active",
84 .debug_hint = active_debug_hint,
85 };
86
debug_active_init(struct i915_active * ref)87 static void debug_active_init(struct i915_active *ref)
88 {
89 debug_object_init(ref, &active_debug_desc);
90 }
91
debug_active_activate(struct i915_active * ref)92 static void debug_active_activate(struct i915_active *ref)
93 {
94 lockdep_assert_held(&ref->tree_lock);
95 debug_object_activate(ref, &active_debug_desc);
96 }
97
debug_active_deactivate(struct i915_active * ref)98 static void debug_active_deactivate(struct i915_active *ref)
99 {
100 lockdep_assert_held(&ref->tree_lock);
101 if (!atomic_read(&ref->count)) /* after the last dec */
102 debug_object_deactivate(ref, &active_debug_desc);
103 }
104
debug_active_fini(struct i915_active * ref)105 static void debug_active_fini(struct i915_active *ref)
106 {
107 debug_object_free(ref, &active_debug_desc);
108 }
109
debug_active_assert(struct i915_active * ref)110 static void debug_active_assert(struct i915_active *ref)
111 {
112 debug_object_assert_init(ref, &active_debug_desc);
113 }
114
115 #else
116
debug_active_init(struct i915_active * ref)117 static inline void debug_active_init(struct i915_active *ref) { }
debug_active_activate(struct i915_active * ref)118 static inline void debug_active_activate(struct i915_active *ref) { }
debug_active_deactivate(struct i915_active * ref)119 static inline void debug_active_deactivate(struct i915_active *ref) { }
debug_active_fini(struct i915_active * ref)120 static inline void debug_active_fini(struct i915_active *ref) { }
debug_active_assert(struct i915_active * ref)121 static inline void debug_active_assert(struct i915_active *ref) { }
122
123 #endif
124
125 static void
__active_retire(struct i915_active * ref)126 __active_retire(struct i915_active *ref)
127 {
128 struct rb_root root = RB_ROOT;
129 struct active_node *it, *n;
130 unsigned long flags;
131
132 GEM_BUG_ON(i915_active_is_idle(ref));
133
134 /* return the unused nodes to our slabcache -- flushing the allocator */
135 if (!atomic_dec_and_lock_irqsave(&ref->count, &ref->tree_lock, flags))
136 return;
137
138 GEM_BUG_ON(rcu_access_pointer(ref->excl.fence));
139 debug_active_deactivate(ref);
140
141 /* Even if we have not used the cache, we may still have a barrier */
142 if (!ref->cache)
143 ref->cache = fetch_node(ref->tree.rb_node);
144
145 /* Keep the MRU cached node for reuse */
146 if (ref->cache) {
147 /* Discard all other nodes in the tree */
148 rb_erase(&ref->cache->node, &ref->tree);
149 root = ref->tree;
150
151 /* Rebuild the tree with only the cached node */
152 rb_link_node(&ref->cache->node, NULL, &ref->tree.rb_node);
153 rb_insert_color(&ref->cache->node, &ref->tree);
154 GEM_BUG_ON(ref->tree.rb_node != &ref->cache->node);
155
156 /* Make the cached node available for reuse with any timeline */
157 ref->cache->timeline = 0; /* needs cmpxchg(u64) */
158 }
159
160 spin_unlock_irqrestore(&ref->tree_lock, flags);
161
162 /* After the final retire, the entire struct may be freed */
163 if (ref->retire)
164 ref->retire(ref);
165
166 /* ... except if you wait on it, you must manage your own references! */
167 wake_up_var(ref);
168
169 /* Finally free the discarded timeline tree */
170 rbtree_postorder_for_each_entry_safe(it, n, &root, node) {
171 GEM_BUG_ON(i915_active_fence_isset(&it->base));
172 kmem_cache_free(slab_cache, it);
173 }
174 }
175
176 static void
active_work(struct work_struct * wrk)177 active_work(struct work_struct *wrk)
178 {
179 struct i915_active *ref = container_of(wrk, typeof(*ref), work);
180
181 GEM_BUG_ON(!atomic_read(&ref->count));
182 if (atomic_add_unless(&ref->count, -1, 1))
183 return;
184
185 __active_retire(ref);
186 }
187
188 static void
active_retire(struct i915_active * ref)189 active_retire(struct i915_active *ref)
190 {
191 GEM_BUG_ON(!atomic_read(&ref->count));
192 if (atomic_add_unless(&ref->count, -1, 1))
193 return;
194
195 if (ref->flags & I915_ACTIVE_RETIRE_SLEEPS) {
196 queue_work(system_unbound_wq, &ref->work);
197 return;
198 }
199
200 __active_retire(ref);
201 }
202
203 static inline struct dma_fence **
__active_fence_slot(struct i915_active_fence * active)204 __active_fence_slot(struct i915_active_fence *active)
205 {
206 return (struct dma_fence ** __force)&active->fence;
207 }
208
209 static inline bool
active_fence_cb(struct dma_fence * fence,struct dma_fence_cb * cb)210 active_fence_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
211 {
212 struct i915_active_fence *active =
213 container_of(cb, typeof(*active), cb);
214
215 return cmpxchg(__active_fence_slot(active), fence, NULL) == fence;
216 }
217
218 static void
node_retire(struct dma_fence * fence,struct dma_fence_cb * cb)219 node_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
220 {
221 if (active_fence_cb(fence, cb))
222 active_retire(container_of(cb, struct active_node, base.cb)->ref);
223 }
224
225 static void
excl_retire(struct dma_fence * fence,struct dma_fence_cb * cb)226 excl_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
227 {
228 if (active_fence_cb(fence, cb))
229 active_retire(container_of(cb, struct i915_active, excl.cb));
230 }
231
__active_lookup(struct i915_active * ref,u64 idx)232 static struct active_node *__active_lookup(struct i915_active *ref, u64 idx)
233 {
234 struct active_node *it;
235
236 GEM_BUG_ON(idx == 0); /* 0 is the unordered timeline, rsvd for cache */
237
238 /*
239 * We track the most recently used timeline to skip a rbtree search
240 * for the common case, under typical loads we never need the rbtree
241 * at all. We can reuse the last slot if it is empty, that is
242 * after the previous activity has been retired, or if it matches the
243 * current timeline.
244 */
245 it = READ_ONCE(ref->cache);
246 if (it) {
247 u64 cached = READ_ONCE(it->timeline);
248
249 /* Once claimed, this slot will only belong to this idx */
250 if (cached == idx)
251 return it;
252
253 /*
254 * An unclaimed cache [.timeline=0] can only be claimed once.
255 *
256 * If the value is already non-zero, some other thread has
257 * claimed the cache and we know that is does not match our
258 * idx. If, and only if, the timeline is currently zero is it
259 * worth competing to claim it atomically for ourselves (for
260 * only the winner of that race will cmpxchg return the old
261 * value of 0).
262 */
263 if (!cached && !cmpxchg64(&it->timeline, 0, idx))
264 return it;
265 }
266
267 BUILD_BUG_ON(offsetof(typeof(*it), node));
268
269 /* While active, the tree can only be built; not destroyed */
270 GEM_BUG_ON(i915_active_is_idle(ref));
271
272 it = fetch_node(ref->tree.rb_node);
273 while (it) {
274 if (it->timeline < idx) {
275 it = fetch_node(it->node.rb_right);
276 } else if (it->timeline > idx) {
277 it = fetch_node(it->node.rb_left);
278 } else {
279 WRITE_ONCE(ref->cache, it);
280 break;
281 }
282 }
283
284 /* NB: If the tree rotated beneath us, we may miss our target. */
285 return it;
286 }
287
288 static struct i915_active_fence *
active_instance(struct i915_active * ref,u64 idx)289 active_instance(struct i915_active *ref, u64 idx)
290 {
291 struct active_node *node;
292 struct rb_node **p, *parent;
293
294 node = __active_lookup(ref, idx);
295 if (likely(node))
296 return &node->base;
297
298 spin_lock_irq(&ref->tree_lock);
299 GEM_BUG_ON(i915_active_is_idle(ref));
300
301 parent = NULL;
302 p = &ref->tree.rb_node;
303 while (*p) {
304 parent = *p;
305
306 node = rb_entry(parent, struct active_node, node);
307 if (node->timeline == idx)
308 goto out;
309
310 if (node->timeline < idx)
311 p = &parent->rb_right;
312 else
313 p = &parent->rb_left;
314 }
315
316 /*
317 * XXX: We should preallocate this before i915_active_ref() is ever
318 * called, but we cannot call into fs_reclaim() anyway, so use GFP_ATOMIC.
319 */
320 node = kmem_cache_alloc(slab_cache, GFP_ATOMIC);
321 if (!node)
322 goto out;
323
324 __i915_active_fence_init(&node->base, NULL, node_retire);
325 node->ref = ref;
326 node->timeline = idx;
327
328 rb_link_node(&node->node, parent, p);
329 rb_insert_color(&node->node, &ref->tree);
330
331 out:
332 WRITE_ONCE(ref->cache, node);
333 spin_unlock_irq(&ref->tree_lock);
334
335 return &node->base;
336 }
337
__i915_active_init(struct i915_active * ref,int (* active)(struct i915_active * ref),void (* retire)(struct i915_active * ref),unsigned long flags,struct lock_class_key * mkey,struct lock_class_key * wkey)338 void __i915_active_init(struct i915_active *ref,
339 int (*active)(struct i915_active *ref),
340 void (*retire)(struct i915_active *ref),
341 unsigned long flags,
342 struct lock_class_key *mkey,
343 struct lock_class_key *wkey)
344 {
345 debug_active_init(ref);
346
347 ref->flags = flags;
348 ref->active = active;
349 ref->retire = retire;
350
351 spin_lock_init(&ref->tree_lock);
352 ref->tree = RB_ROOT;
353 ref->cache = NULL;
354
355 init_llist_head(&ref->preallocated_barriers);
356 atomic_set(&ref->count, 0);
357 __mutex_init(&ref->mutex, "i915_active", mkey);
358 __i915_active_fence_init(&ref->excl, NULL, excl_retire);
359 INIT_WORK(&ref->work, active_work);
360 #if IS_ENABLED(CONFIG_LOCKDEP)
361 lockdep_init_map(&ref->work.lockdep_map, "i915_active.work", wkey, 0);
362 #endif
363 }
364
____active_del_barrier(struct i915_active * ref,struct active_node * node,struct intel_engine_cs * engine)365 static bool ____active_del_barrier(struct i915_active *ref,
366 struct active_node *node,
367 struct intel_engine_cs *engine)
368
369 {
370 struct llist_node *head = NULL, *tail = NULL;
371 struct llist_node *pos, *next;
372
373 GEM_BUG_ON(node->timeline != engine->kernel_context->timeline->fence_context);
374
375 /*
376 * Rebuild the llist excluding our node. We may perform this
377 * outside of the kernel_context timeline mutex and so someone
378 * else may be manipulating the engine->barrier_tasks, in
379 * which case either we or they will be upset :)
380 *
381 * A second __active_del_barrier() will report failure to claim
382 * the active_node and the caller will just shrug and know not to
383 * claim ownership of its node.
384 *
385 * A concurrent i915_request_add_active_barriers() will miss adding
386 * any of the tasks, but we will try again on the next -- and since
387 * we are actively using the barrier, we know that there will be
388 * at least another opportunity when we idle.
389 */
390 llist_for_each_safe(pos, next, llist_del_all(&engine->barrier_tasks)) {
391 if (node == barrier_from_ll(pos)) {
392 node = NULL;
393 continue;
394 }
395
396 pos->next = head;
397 head = pos;
398 if (!tail)
399 tail = pos;
400 }
401 if (head)
402 llist_add_batch(head, tail, &engine->barrier_tasks);
403
404 return !node;
405 }
406
407 static bool
__active_del_barrier(struct i915_active * ref,struct active_node * node)408 __active_del_barrier(struct i915_active *ref, struct active_node *node)
409 {
410 return ____active_del_barrier(ref, node, barrier_to_engine(node));
411 }
412
413 static bool
replace_barrier(struct i915_active * ref,struct i915_active_fence * active)414 replace_barrier(struct i915_active *ref, struct i915_active_fence *active)
415 {
416 if (!is_barrier(active)) /* proto-node used by our idle barrier? */
417 return false;
418
419 /*
420 * This request is on the kernel_context timeline, and so
421 * we can use it to substitute for the pending idle-barrer
422 * request that we want to emit on the kernel_context.
423 */
424 return __active_del_barrier(ref, node_from_active(active));
425 }
426
i915_active_ref(struct i915_active * ref,u64 idx,struct dma_fence * fence)427 int i915_active_ref(struct i915_active *ref, u64 idx, struct dma_fence *fence)
428 {
429 struct i915_active_fence *active;
430 int err;
431
432 /* Prevent reaping in case we malloc/wait while building the tree */
433 err = i915_active_acquire(ref);
434 if (err)
435 return err;
436
437 do {
438 active = active_instance(ref, idx);
439 if (!active) {
440 err = -ENOMEM;
441 goto out;
442 }
443
444 if (replace_barrier(ref, active)) {
445 RCU_INIT_POINTER(active->fence, NULL);
446 atomic_dec(&ref->count);
447 }
448 } while (unlikely(is_barrier(active)));
449
450 fence = __i915_active_fence_set(active, fence);
451 if (!fence)
452 __i915_active_acquire(ref);
453 else
454 dma_fence_put(fence);
455
456 out:
457 i915_active_release(ref);
458 return err;
459 }
460
461 static struct dma_fence *
__i915_active_set_fence(struct i915_active * ref,struct i915_active_fence * active,struct dma_fence * 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 prev = __i915_active_fence_set(active, fence);
474 if (!prev)
475 __i915_active_acquire(ref);
476
477 return prev;
478 }
479
480 static struct i915_active_fence *
__active_fence(struct i915_active * ref,u64 idx)481 __active_fence(struct i915_active *ref, u64 idx)
482 {
483 struct active_node *it;
484
485 it = __active_lookup(ref, idx);
486 if (unlikely(!it)) { /* Contention with parallel tree builders! */
487 spin_lock_irq(&ref->tree_lock);
488 it = __active_lookup(ref, idx);
489 spin_unlock_irq(&ref->tree_lock);
490 }
491 GEM_BUG_ON(!it); /* slot must be preallocated */
492
493 return &it->base;
494 }
495
496 struct dma_fence *
__i915_active_ref(struct i915_active * ref,u64 idx,struct dma_fence * fence)497 __i915_active_ref(struct i915_active *ref, u64 idx, struct dma_fence *fence)
498 {
499 /* Only valid while active, see i915_active_acquire_for_context() */
500 return __i915_active_set_fence(ref, __active_fence(ref, idx), fence);
501 }
502
503 struct dma_fence *
i915_active_set_exclusive(struct i915_active * ref,struct dma_fence * f)504 i915_active_set_exclusive(struct i915_active *ref, struct dma_fence *f)
505 {
506 /* We expect the caller to manage the exclusive timeline ordering */
507 return __i915_active_set_fence(ref, &ref->excl, f);
508 }
509
i915_active_acquire_if_busy(struct i915_active * ref)510 bool i915_active_acquire_if_busy(struct i915_active *ref)
511 {
512 debug_active_assert(ref);
513 return atomic_add_unless(&ref->count, 1, 0);
514 }
515
__i915_active_activate(struct i915_active * ref)516 static void __i915_active_activate(struct i915_active *ref)
517 {
518 spin_lock_irq(&ref->tree_lock); /* __active_retire() */
519 if (!atomic_fetch_inc(&ref->count))
520 debug_active_activate(ref);
521 spin_unlock_irq(&ref->tree_lock);
522 }
523
i915_active_acquire(struct i915_active * ref)524 int i915_active_acquire(struct i915_active *ref)
525 {
526 int err;
527
528 if (i915_active_acquire_if_busy(ref))
529 return 0;
530
531 if (!ref->active) {
532 __i915_active_activate(ref);
533 return 0;
534 }
535
536 err = mutex_lock_interruptible(&ref->mutex);
537 if (err)
538 return err;
539
540 if (likely(!i915_active_acquire_if_busy(ref))) {
541 err = ref->active(ref);
542 if (!err)
543 __i915_active_activate(ref);
544 }
545
546 mutex_unlock(&ref->mutex);
547
548 return err;
549 }
550
i915_active_acquire_for_context(struct i915_active * ref,u64 idx)551 int i915_active_acquire_for_context(struct i915_active *ref, u64 idx)
552 {
553 struct i915_active_fence *active;
554 int err;
555
556 err = i915_active_acquire(ref);
557 if (err)
558 return err;
559
560 active = active_instance(ref, idx);
561 if (!active) {
562 i915_active_release(ref);
563 return -ENOMEM;
564 }
565
566 return 0; /* return with active ref */
567 }
568
i915_active_release(struct i915_active * ref)569 void i915_active_release(struct i915_active *ref)
570 {
571 debug_active_assert(ref);
572 active_retire(ref);
573 }
574
enable_signaling(struct i915_active_fence * active)575 static void enable_signaling(struct i915_active_fence *active)
576 {
577 struct dma_fence *fence;
578
579 if (unlikely(is_barrier(active)))
580 return;
581
582 fence = i915_active_fence_get(active);
583 if (!fence)
584 return;
585
586 dma_fence_enable_sw_signaling(fence);
587 dma_fence_put(fence);
588 }
589
flush_barrier(struct active_node * it)590 static int flush_barrier(struct active_node *it)
591 {
592 struct intel_engine_cs *engine;
593
594 if (likely(!is_barrier(&it->base)))
595 return 0;
596
597 engine = __barrier_to_engine(it);
598 smp_rmb(); /* serialise with add_active_barriers */
599 if (!is_barrier(&it->base))
600 return 0;
601
602 return intel_engine_flush_barriers(engine);
603 }
604
flush_lazy_signals(struct i915_active * ref)605 static int flush_lazy_signals(struct i915_active *ref)
606 {
607 struct active_node *it, *n;
608 int err = 0;
609
610 enable_signaling(&ref->excl);
611 rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
612 err = flush_barrier(it); /* unconnected idle barrier? */
613 if (err)
614 break;
615
616 enable_signaling(&it->base);
617 }
618
619 return err;
620 }
621
__i915_active_wait(struct i915_active * ref,int state)622 int __i915_active_wait(struct i915_active *ref, int state)
623 {
624 might_sleep();
625
626 /* Any fence added after the wait begins will not be auto-signaled */
627 if (i915_active_acquire_if_busy(ref)) {
628 int err;
629
630 err = flush_lazy_signals(ref);
631 i915_active_release(ref);
632 if (err)
633 return err;
634
635 if (___wait_var_event(ref, i915_active_is_idle(ref),
636 state, 0, 0, schedule()))
637 return -EINTR;
638 }
639
640 /*
641 * After the wait is complete, the caller may free the active.
642 * We have to flush any concurrent retirement before returning.
643 */
644 flush_work(&ref->work);
645 return 0;
646 }
647
__await_active(struct i915_active_fence * active,int (* fn)(void * arg,struct dma_fence * fence),void * arg)648 static int __await_active(struct i915_active_fence *active,
649 int (*fn)(void *arg, struct dma_fence *fence),
650 void *arg)
651 {
652 struct dma_fence *fence;
653
654 if (is_barrier(active)) /* XXX flush the barrier? */
655 return 0;
656
657 fence = i915_active_fence_get(active);
658 if (fence) {
659 int err;
660
661 err = fn(arg, fence);
662 dma_fence_put(fence);
663 if (err < 0)
664 return err;
665 }
666
667 return 0;
668 }
669
670 struct wait_barrier {
671 struct wait_queue_entry base;
672 struct i915_active *ref;
673 };
674
675 static int
barrier_wake(wait_queue_entry_t * wq,unsigned int mode,int flags,void * key)676 barrier_wake(wait_queue_entry_t *wq, unsigned int mode, int flags, void *key)
677 {
678 struct wait_barrier *wb = container_of(wq, typeof(*wb), base);
679
680 if (i915_active_is_idle(wb->ref)) {
681 list_del(&wq->entry);
682 i915_sw_fence_complete(wq->private);
683 kfree(wq);
684 }
685
686 return 0;
687 }
688
__await_barrier(struct i915_active * ref,struct i915_sw_fence * fence)689 static int __await_barrier(struct i915_active *ref, struct i915_sw_fence *fence)
690 {
691 struct wait_barrier *wb;
692
693 wb = kmalloc(sizeof(*wb), GFP_KERNEL);
694 if (unlikely(!wb))
695 return -ENOMEM;
696
697 GEM_BUG_ON(i915_active_is_idle(ref));
698 if (!i915_sw_fence_await(fence)) {
699 kfree(wb);
700 return -EINVAL;
701 }
702
703 wb->base.flags = 0;
704 wb->base.func = barrier_wake;
705 wb->base.private = fence;
706 wb->ref = ref;
707
708 add_wait_queue(__var_waitqueue(ref), &wb->base);
709 return 0;
710 }
711
await_active(struct i915_active * ref,unsigned int flags,int (* fn)(void * arg,struct dma_fence * fence),void * arg,struct i915_sw_fence * barrier)712 static int await_active(struct i915_active *ref,
713 unsigned int flags,
714 int (*fn)(void *arg, struct dma_fence *fence),
715 void *arg, struct i915_sw_fence *barrier)
716 {
717 int err = 0;
718
719 if (!i915_active_acquire_if_busy(ref))
720 return 0;
721
722 if (flags & I915_ACTIVE_AWAIT_EXCL &&
723 rcu_access_pointer(ref->excl.fence)) {
724 err = __await_active(&ref->excl, fn, arg);
725 if (err)
726 goto out;
727 }
728
729 if (flags & I915_ACTIVE_AWAIT_ACTIVE) {
730 struct active_node *it, *n;
731
732 rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
733 err = __await_active(&it->base, fn, arg);
734 if (err)
735 goto out;
736 }
737 }
738
739 if (flags & I915_ACTIVE_AWAIT_BARRIER) {
740 err = flush_lazy_signals(ref);
741 if (err)
742 goto out;
743
744 err = __await_barrier(ref, barrier);
745 if (err)
746 goto out;
747 }
748
749 out:
750 i915_active_release(ref);
751 return err;
752 }
753
rq_await_fence(void * arg,struct dma_fence * fence)754 static int rq_await_fence(void *arg, struct dma_fence *fence)
755 {
756 return i915_request_await_dma_fence(arg, fence);
757 }
758
i915_request_await_active(struct i915_request * rq,struct i915_active * ref,unsigned int flags)759 int i915_request_await_active(struct i915_request *rq,
760 struct i915_active *ref,
761 unsigned int flags)
762 {
763 return await_active(ref, flags, rq_await_fence, rq, &rq->submit);
764 }
765
sw_await_fence(void * arg,struct dma_fence * fence)766 static int sw_await_fence(void *arg, struct dma_fence *fence)
767 {
768 return i915_sw_fence_await_dma_fence(arg, fence, 0,
769 GFP_NOWAIT | __GFP_NOWARN);
770 }
771
i915_sw_fence_await_active(struct i915_sw_fence * fence,struct i915_active * ref,unsigned int flags)772 int i915_sw_fence_await_active(struct i915_sw_fence *fence,
773 struct i915_active *ref,
774 unsigned int flags)
775 {
776 return await_active(ref, flags, sw_await_fence, fence, fence);
777 }
778
i915_active_fini(struct i915_active * ref)779 void i915_active_fini(struct i915_active *ref)
780 {
781 debug_active_fini(ref);
782 GEM_BUG_ON(atomic_read(&ref->count));
783 GEM_BUG_ON(work_pending(&ref->work));
784 mutex_destroy(&ref->mutex);
785
786 if (ref->cache)
787 kmem_cache_free(slab_cache, ref->cache);
788 }
789
is_idle_barrier(struct active_node * node,u64 idx)790 static inline bool is_idle_barrier(struct active_node *node, u64 idx)
791 {
792 return node->timeline == idx && !i915_active_fence_isset(&node->base);
793 }
794
reuse_idle_barrier(struct i915_active * ref,u64 idx)795 static struct active_node *reuse_idle_barrier(struct i915_active *ref, u64 idx)
796 {
797 struct rb_node *prev, *p;
798
799 if (RB_EMPTY_ROOT(&ref->tree))
800 return NULL;
801
802 GEM_BUG_ON(i915_active_is_idle(ref));
803
804 /*
805 * Try to reuse any existing barrier nodes already allocated for this
806 * i915_active, due to overlapping active phases there is likely a
807 * node kept alive (as we reuse before parking). We prefer to reuse
808 * completely idle barriers (less hassle in manipulating the llists),
809 * but otherwise any will do.
810 */
811 if (ref->cache && is_idle_barrier(ref->cache, idx)) {
812 p = &ref->cache->node;
813 goto match;
814 }
815
816 prev = NULL;
817 p = ref->tree.rb_node;
818 while (p) {
819 struct active_node *node =
820 rb_entry(p, struct active_node, node);
821
822 if (is_idle_barrier(node, idx))
823 goto match;
824
825 prev = p;
826 if (node->timeline < idx)
827 p = READ_ONCE(p->rb_right);
828 else
829 p = READ_ONCE(p->rb_left);
830 }
831
832 /*
833 * No quick match, but we did find the leftmost rb_node for the
834 * kernel_context. Walk the rb_tree in-order to see if there were
835 * any idle-barriers on this timeline that we missed, or just use
836 * the first pending barrier.
837 */
838 for (p = prev; p; p = rb_next(p)) {
839 struct active_node *node =
840 rb_entry(p, struct active_node, node);
841 struct intel_engine_cs *engine;
842
843 if (node->timeline > idx)
844 break;
845
846 if (node->timeline < idx)
847 continue;
848
849 if (is_idle_barrier(node, idx))
850 goto match;
851
852 /*
853 * The list of pending barriers is protected by the
854 * kernel_context timeline, which notably we do not hold
855 * here. i915_request_add_active_barriers() may consume
856 * the barrier before we claim it, so we have to check
857 * for success.
858 */
859 engine = __barrier_to_engine(node);
860 smp_rmb(); /* serialise with add_active_barriers */
861 if (is_barrier(&node->base) &&
862 ____active_del_barrier(ref, node, engine))
863 goto match;
864 }
865
866 return NULL;
867
868 match:
869 spin_lock_irq(&ref->tree_lock);
870 rb_erase(p, &ref->tree); /* Hide from waits and sibling allocations */
871 if (p == &ref->cache->node)
872 WRITE_ONCE(ref->cache, NULL);
873 spin_unlock_irq(&ref->tree_lock);
874
875 return rb_entry(p, struct active_node, node);
876 }
877
i915_active_acquire_preallocate_barrier(struct i915_active * ref,struct intel_engine_cs * engine)878 int i915_active_acquire_preallocate_barrier(struct i915_active *ref,
879 struct intel_engine_cs *engine)
880 {
881 intel_engine_mask_t tmp, mask = engine->mask;
882 struct llist_node *first = NULL, *last = NULL;
883 struct intel_gt *gt = engine->gt;
884
885 GEM_BUG_ON(i915_active_is_idle(ref));
886
887 /* Wait until the previous preallocation is completed */
888 while (!llist_empty(&ref->preallocated_barriers))
889 cond_resched();
890
891 /*
892 * Preallocate a node for each physical engine supporting the target
893 * engine (remember virtual engines have more than one sibling).
894 * We can then use the preallocated nodes in
895 * i915_active_acquire_barrier()
896 */
897 GEM_BUG_ON(!mask);
898 for_each_engine_masked(engine, gt, mask, tmp) {
899 u64 idx = engine->kernel_context->timeline->fence_context;
900 struct llist_node *prev = first;
901 struct active_node *node;
902
903 rcu_read_lock();
904 node = reuse_idle_barrier(ref, idx);
905 rcu_read_unlock();
906 if (!node) {
907 node = kmem_cache_alloc(slab_cache, GFP_KERNEL);
908 if (!node)
909 goto unwind;
910
911 RCU_INIT_POINTER(node->base.fence, NULL);
912 node->base.cb.func = node_retire;
913 node->timeline = idx;
914 node->ref = ref;
915 }
916
917 if (!i915_active_fence_isset(&node->base)) {
918 /*
919 * Mark this as being *our* unconnected proto-node.
920 *
921 * Since this node is not in any list, and we have
922 * decoupled it from the rbtree, we can reuse the
923 * request to indicate this is an idle-barrier node
924 * and then we can use the rb_node and list pointers
925 * for our tracking of the pending barrier.
926 */
927 RCU_INIT_POINTER(node->base.fence, ERR_PTR(-EAGAIN));
928 node->base.cb.node.prev = (void *)engine;
929 __i915_active_acquire(ref);
930 }
931 GEM_BUG_ON(rcu_access_pointer(node->base.fence) != ERR_PTR(-EAGAIN));
932
933 GEM_BUG_ON(barrier_to_engine(node) != engine);
934 first = barrier_to_ll(node);
935 first->next = prev;
936 if (!last)
937 last = first;
938 intel_engine_pm_get(engine);
939 }
940
941 GEM_BUG_ON(!llist_empty(&ref->preallocated_barriers));
942 llist_add_batch(first, last, &ref->preallocated_barriers);
943
944 return 0;
945
946 unwind:
947 while (first) {
948 struct active_node *node = barrier_from_ll(first);
949
950 first = first->next;
951
952 atomic_dec(&ref->count);
953 intel_engine_pm_put(barrier_to_engine(node));
954
955 kmem_cache_free(slab_cache, node);
956 }
957 return -ENOMEM;
958 }
959
i915_active_acquire_barrier(struct i915_active * ref)960 void i915_active_acquire_barrier(struct i915_active *ref)
961 {
962 struct llist_node *pos, *next;
963 unsigned long flags;
964
965 GEM_BUG_ON(i915_active_is_idle(ref));
966
967 /*
968 * Transfer the list of preallocated barriers into the
969 * i915_active rbtree, but only as proto-nodes. They will be
970 * populated by i915_request_add_active_barriers() to point to the
971 * request that will eventually release them.
972 */
973 llist_for_each_safe(pos, next, take_preallocated_barriers(ref)) {
974 struct active_node *node = barrier_from_ll(pos);
975 struct intel_engine_cs *engine = barrier_to_engine(node);
976 struct rb_node **p, *parent;
977
978 spin_lock_irqsave_nested(&ref->tree_lock, flags,
979 SINGLE_DEPTH_NESTING);
980 parent = NULL;
981 p = &ref->tree.rb_node;
982 while (*p) {
983 struct active_node *it;
984
985 parent = *p;
986
987 it = rb_entry(parent, struct active_node, node);
988 if (it->timeline < node->timeline)
989 p = &parent->rb_right;
990 else
991 p = &parent->rb_left;
992 }
993 rb_link_node(&node->node, parent, p);
994 rb_insert_color(&node->node, &ref->tree);
995 spin_unlock_irqrestore(&ref->tree_lock, flags);
996
997 GEM_BUG_ON(!intel_engine_pm_is_awake(engine));
998 llist_add(barrier_to_ll(node), &engine->barrier_tasks);
999 intel_engine_pm_put_delay(engine, 1);
1000 }
1001 }
1002
ll_to_fence_slot(struct llist_node * node)1003 static struct dma_fence **ll_to_fence_slot(struct llist_node *node)
1004 {
1005 return __active_fence_slot(&barrier_from_ll(node)->base);
1006 }
1007
i915_request_add_active_barriers(struct i915_request * rq)1008 void i915_request_add_active_barriers(struct i915_request *rq)
1009 {
1010 struct intel_engine_cs *engine = rq->engine;
1011 struct llist_node *node, *next;
1012 unsigned long flags;
1013
1014 GEM_BUG_ON(!intel_context_is_barrier(rq->context));
1015 GEM_BUG_ON(intel_engine_is_virtual(engine));
1016 GEM_BUG_ON(i915_request_timeline(rq) != engine->kernel_context->timeline);
1017
1018 node = llist_del_all(&engine->barrier_tasks);
1019 if (!node)
1020 return;
1021 /*
1022 * Attach the list of proto-fences to the in-flight request such
1023 * that the parent i915_active will be released when this request
1024 * is retired.
1025 */
1026 spin_lock_irqsave(&rq->lock, flags);
1027 llist_for_each_safe(node, next, node) {
1028 /* serialise with reuse_idle_barrier */
1029 smp_store_mb(*ll_to_fence_slot(node), &rq->fence);
1030 list_add_tail((struct list_head *)node, &rq->fence.cb_list);
1031 }
1032 spin_unlock_irqrestore(&rq->lock, flags);
1033 }
1034
1035 /*
1036 * __i915_active_fence_set: Update the last active fence along its timeline
1037 * @active: the active tracker
1038 * @fence: the new fence (under construction)
1039 *
1040 * Records the new @fence as the last active fence along its timeline in
1041 * this active tracker, moving the tracking callbacks from the previous
1042 * fence onto this one. Gets and returns a reference to the previous fence
1043 * (if not already completed), which the caller must put after making sure
1044 * that it is executed before the new fence. To ensure that the order of
1045 * fences within the timeline of the i915_active_fence is understood, it
1046 * should be locked by the caller.
1047 */
1048 struct dma_fence *
__i915_active_fence_set(struct i915_active_fence * active,struct dma_fence * fence)1049 __i915_active_fence_set(struct i915_active_fence *active,
1050 struct dma_fence *fence)
1051 {
1052 struct dma_fence *prev;
1053 unsigned long flags;
1054
1055 /*
1056 * In case of fences embedded in i915_requests, their memory is
1057 * SLAB_FAILSAFE_BY_RCU, then it can be reused right after release
1058 * by new requests. Then, there is a risk of passing back a pointer
1059 * to a new, completely unrelated fence that reuses the same memory
1060 * while tracked under a different active tracker. Combined with i915
1061 * perf open/close operations that build await dependencies between
1062 * engine kernel context requests and user requests from different
1063 * timelines, this can lead to dependency loops and infinite waits.
1064 *
1065 * As a countermeasure, we try to get a reference to the active->fence
1066 * first, so if we succeed and pass it back to our user then it is not
1067 * released and potentially reused by an unrelated request before the
1068 * user has a chance to set up an await dependency on it.
1069 */
1070 prev = i915_active_fence_get(active);
1071 if (fence == prev)
1072 return fence;
1073
1074 GEM_BUG_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags));
1075
1076 /*
1077 * Consider that we have two threads arriving (A and B), with
1078 * C already resident as the active->fence.
1079 *
1080 * Both A and B have got a reference to C or NULL, depending on the
1081 * timing of the interrupt handler. Let's assume that if A has got C
1082 * then it has locked C first (before B).
1083 *
1084 * Note the strong ordering of the timeline also provides consistent
1085 * nesting rules for the fence->lock; the inner lock is always the
1086 * older lock.
1087 */
1088 spin_lock_irqsave(fence->lock, flags);
1089 if (prev)
1090 spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING);
1091
1092 /*
1093 * A does the cmpxchg first, and so it sees C or NULL, as before, or
1094 * something else, depending on the timing of other threads and/or
1095 * interrupt handler. If not the same as before then A unlocks C if
1096 * applicable and retries, starting from an attempt to get a new
1097 * active->fence. Meanwhile, B follows the same path as A.
1098 * Once A succeeds with cmpxch, B fails again, retires, gets A from
1099 * active->fence, locks it as soon as A completes, and possibly
1100 * succeeds with cmpxchg.
1101 */
1102 while (cmpxchg(__active_fence_slot(active), prev, fence) != prev) {
1103 if (prev) {
1104 spin_unlock(prev->lock);
1105 dma_fence_put(prev);
1106 }
1107 spin_unlock_irqrestore(fence->lock, flags);
1108
1109 prev = i915_active_fence_get(active);
1110 GEM_BUG_ON(prev == fence);
1111
1112 spin_lock_irqsave(fence->lock, flags);
1113 if (prev)
1114 spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING);
1115 }
1116
1117 /*
1118 * If prev is NULL then the previous fence must have been signaled
1119 * and we know that we are first on the timeline. If it is still
1120 * present then, having the lock on that fence already acquired, we
1121 * serialise with the interrupt handler, in the process of removing it
1122 * from any future interrupt callback. A will then wait on C before
1123 * executing (if present).
1124 *
1125 * As B is second, it sees A as the previous fence and so waits for
1126 * it to complete its transition and takes over the occupancy for
1127 * itself -- remembering that it needs to wait on A before executing.
1128 */
1129 if (prev) {
1130 __list_del_entry(&active->cb.node);
1131 spin_unlock(prev->lock); /* serialise with prev->cb_list */
1132 }
1133 list_add_tail(&active->cb.node, &fence->cb_list);
1134 spin_unlock_irqrestore(fence->lock, flags);
1135
1136 return prev;
1137 }
1138
i915_active_fence_set(struct i915_active_fence * active,struct i915_request * rq)1139 int i915_active_fence_set(struct i915_active_fence *active,
1140 struct i915_request *rq)
1141 {
1142 struct dma_fence *fence;
1143 int err = 0;
1144
1145 /* Must maintain timeline ordering wrt previous active requests */
1146 fence = __i915_active_fence_set(active, &rq->fence);
1147 if (fence) {
1148 err = i915_request_await_dma_fence(rq, fence);
1149 dma_fence_put(fence);
1150 }
1151
1152 return err;
1153 }
1154
i915_active_noop(struct dma_fence * fence,struct dma_fence_cb * cb)1155 void i915_active_noop(struct dma_fence *fence, struct dma_fence_cb *cb)
1156 {
1157 active_fence_cb(fence, cb);
1158 }
1159
1160 struct auto_active {
1161 struct i915_active base;
1162 struct kref ref;
1163 };
1164
i915_active_get(struct i915_active * ref)1165 struct i915_active *i915_active_get(struct i915_active *ref)
1166 {
1167 struct auto_active *aa = container_of(ref, typeof(*aa), base);
1168
1169 kref_get(&aa->ref);
1170 return &aa->base;
1171 }
1172
auto_release(struct kref * ref)1173 static void auto_release(struct kref *ref)
1174 {
1175 struct auto_active *aa = container_of(ref, typeof(*aa), ref);
1176
1177 i915_active_fini(&aa->base);
1178 kfree(aa);
1179 }
1180
i915_active_put(struct i915_active * ref)1181 void i915_active_put(struct i915_active *ref)
1182 {
1183 struct auto_active *aa = container_of(ref, typeof(*aa), base);
1184
1185 kref_put(&aa->ref, auto_release);
1186 }
1187
auto_active(struct i915_active * ref)1188 static int auto_active(struct i915_active *ref)
1189 {
1190 i915_active_get(ref);
1191 return 0;
1192 }
1193
auto_retire(struct i915_active * ref)1194 static void auto_retire(struct i915_active *ref)
1195 {
1196 i915_active_put(ref);
1197 }
1198
i915_active_create(void)1199 struct i915_active *i915_active_create(void)
1200 {
1201 struct auto_active *aa;
1202
1203 aa = kmalloc(sizeof(*aa), GFP_KERNEL);
1204 if (!aa)
1205 return NULL;
1206
1207 kref_init(&aa->ref);
1208 i915_active_init(&aa->base, auto_active, auto_retire, 0);
1209
1210 return &aa->base;
1211 }
1212
1213 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1214 #include "selftests/i915_active.c"
1215 #endif
1216
i915_active_module_exit(void)1217 void i915_active_module_exit(void)
1218 {
1219 kmem_cache_destroy(slab_cache);
1220 }
1221
i915_active_module_init(void)1222 int __init i915_active_module_init(void)
1223 {
1224 slab_cache = KMEM_CACHE(active_node, SLAB_HWCACHE_ALIGN);
1225 if (!slab_cache)
1226 return -ENOMEM;
1227
1228 return 0;
1229 }
1230