2 * qht.c - QEMU Hash Table, designed to scale for read-mostly workloads.
4 * Copyright (C) 2016, Emilio G. Cota <cota@braap.org>
6 * License: GNU GPL, version 2 or later.
7 * See the COPYING file in the top-level directory.
10 * - NULL cannot be inserted/removed as a pointer value.
11 * - Trying to insert an already-existing hash-pointer pair is OK. However,
12 * it is not OK to insert into the same hash table different hash-pointer
13 * pairs that have the same pointer value, but not the hashes.
14 * - Lookups are performed under an RCU read-critical section; removals
15 * must wait for a grace period to elapse before freeing removed objects.
18 * - Reads (i.e. lookups and iterators) can be concurrent with other reads.
19 * Lookups that are concurrent with writes to the same bucket will retry
20 * via a seqlock; iterators acquire all bucket locks and therefore can be
21 * concurrent with lookups and are serialized wrt writers.
22 * - Writes (i.e. insertions/removals) can be concurrent with writes to
23 * different buckets; writes to the same bucket are serialized through a lock.
24 * - Optional auto-resizing: the hash table resizes up if the load surpasses
25 * a certain threshold. Resizing is done concurrently with readers; writes
26 * are serialized with the resize operation.
28 * The key structure is the bucket, which is cacheline-sized. Buckets
29 * contain a few hash values and pointers; the u32 hash values are stored in
30 * full so that resizing is fast. Having this structure instead of directly
31 * chaining items has two advantages:
32 * - Failed lookups fail fast, and touch a minimum number of cache lines.
33 * - Resizing the hash table with concurrent lookups is easy.
35 * There are two types of buckets:
36 * 1. "head" buckets are the ones allocated in the array of buckets in qht_map.
37 * 2. all "non-head" buckets (i.e. all others) are members of a chain that
38 * starts from a head bucket.
39 * Note that the seqlock and spinlock of a head bucket applies to all buckets
40 * chained to it; these two fields are unused in non-head buckets.
42 * On removals, we move the last valid item in the chain to the position of the
43 * just-removed entry. This makes lookups slightly faster, since the moment an
44 * invalid entry is found, the (failed) lookup is over.
46 * Resizing is done by taking all bucket spinlocks (so that no other writers can
47 * race with us) and then copying all entries into a new hash map. Then, the
48 * ht->map pointer is set, and the old map is freed once no RCU readers can see
51 * Writers check for concurrent resizes by comparing ht->map before and after
52 * acquiring their bucket lock. If they don't match, a resize has occured
53 * while the bucket spinlock was being acquired.
56 * - Idea of cacheline-sized buckets with full hashes taken from:
57 * David, Guerraoui & Trigonakis, "Asynchronized Concurrency:
58 * The Secret to Scaling Concurrent Search Data Structures", ASPLOS'15.
59 * - Why not RCU-based hash tables? They would allow us to get rid of the
60 * seqlock, but resizing would take forever since RCU read critical
61 * sections in QEMU take quite a long time.
62 * More info on relativistic hash tables:
63 * + Triplett, McKenney & Walpole, "Resizable, Scalable, Concurrent Hash
64 * Tables via Relativistic Programming", USENIX ATC'11.
65 * + Corbet, "Relativistic hash tables, part 1: Algorithms", @ lwn.net, 2014.
66 * https://lwn.net/Articles/612021/
69 #include "qemu/atomic.h"
75 * We want to avoid false sharing of cache lines. Most systems have 64-byte
76 * cache lines so we go with it for simplicity.
78 * Note that systems with smaller cache lines will be fine (the struct is
79 * almost 64-bytes); systems with larger cache lines might suffer from
82 #define QHT_BUCKET_ALIGN 64
84 /* define these to keep sizeof(qht_bucket) within QHT_BUCKET_ALIGN */
85 #if HOST_LONG_BITS == 32
86 #define QHT_BUCKET_ENTRIES 6
88 #define QHT_BUCKET_ENTRIES 4
92 * Note: reading partially-updated pointers in @pointers could lead to
93 * segfaults. We thus access them with atomic_read/set; this guarantees
94 * that the compiler makes all those accesses atomic. We also need the
95 * volatile-like behavior in atomic_read, since otherwise the compiler
96 * might refetch the pointer.
97 * atomic_read's are of course not necessary when the bucket lock is held.
99 * If both ht->lock and b->lock are grabbed, ht->lock should always
104 QemuSeqLock sequence
;
105 uint32_t hashes
[QHT_BUCKET_ENTRIES
];
106 void *pointers
[QHT_BUCKET_ENTRIES
];
107 struct qht_bucket
*next
;
108 } QEMU_ALIGNED(QHT_BUCKET_ALIGN
);
110 QEMU_BUILD_BUG_ON(sizeof(struct qht_bucket
) > QHT_BUCKET_ALIGN
);
113 * struct qht_map - structure to track an array of buckets
114 * @rcu: used by RCU. Keep it as the top field in the struct to help valgrind
115 * find the whole struct.
116 * @buckets: array of head buckets. It is constant once the map is created.
117 * @n_buckets: number of head buckets. It is constant once the map is created.
118 * @n_added_buckets: number of added (i.e. "non-head") buckets
119 * @n_added_buckets_threshold: threshold to trigger an upward resize once the
120 * number of added buckets surpasses it.
122 * Buckets are tracked in what we call a "map", i.e. this structure.
126 struct qht_bucket
*buckets
;
128 size_t n_added_buckets
;
129 size_t n_added_buckets_threshold
;
132 /* trigger a resize when n_added_buckets > n_buckets / div */
133 #define QHT_NR_ADDED_BUCKETS_THRESHOLD_DIV 8
135 static void qht_do_resize(struct qht
*ht
, struct qht_map
*new);
136 static void qht_grow_maybe(struct qht
*ht
);
140 #define qht_debug_assert(X) do { assert(X); } while (0)
142 static void qht_bucket_debug__locked(struct qht_bucket
*b
)
144 bool seen_empty
= false;
145 bool corrupt
= false;
149 for (i
= 0; i
< QHT_BUCKET_ENTRIES
; i
++) {
150 if (b
->pointers
[i
] == NULL
) {
155 fprintf(stderr
, "%s: b: %p, pos: %i, hash: 0x%x, p: %p\n",
156 __func__
, b
, i
, b
->hashes
[i
], b
->pointers
[i
]);
162 qht_debug_assert(!corrupt
);
165 static void qht_map_debug__all_locked(struct qht_map
*map
)
169 for (i
= 0; i
< map
->n_buckets
; i
++) {
170 qht_bucket_debug__locked(&map
->buckets
[i
]);
175 #define qht_debug_assert(X) do { (void)(X); } while (0)
177 static inline void qht_bucket_debug__locked(struct qht_bucket
*b
)
180 static inline void qht_map_debug__all_locked(struct qht_map
*map
)
182 #endif /* QHT_DEBUG */
184 static inline size_t qht_elems_to_buckets(size_t n_elems
)
186 return pow2ceil(n_elems
/ QHT_BUCKET_ENTRIES
);
189 static inline void qht_head_init(struct qht_bucket
*b
)
191 memset(b
, 0, sizeof(*b
));
192 qemu_spin_init(&b
->lock
);
193 seqlock_init(&b
->sequence
);
197 struct qht_bucket
*qht_map_to_bucket(struct qht_map
*map
, uint32_t hash
)
199 return &map
->buckets
[hash
& (map
->n_buckets
- 1)];
202 /* acquire all bucket locks from a map */
203 static void qht_map_lock_buckets(struct qht_map
*map
)
207 for (i
= 0; i
< map
->n_buckets
; i
++) {
208 struct qht_bucket
*b
= &map
->buckets
[i
];
210 qemu_spin_lock(&b
->lock
);
214 static void qht_map_unlock_buckets(struct qht_map
*map
)
218 for (i
= 0; i
< map
->n_buckets
; i
++) {
219 struct qht_bucket
*b
= &map
->buckets
[i
];
221 qemu_spin_unlock(&b
->lock
);
226 * Call with at least a bucket lock held.
227 * @map should be the value read before acquiring the lock (or locks).
229 static inline bool qht_map_is_stale__locked(struct qht
*ht
, struct qht_map
*map
)
231 return map
!= ht
->map
;
235 * Grab all bucket locks, and set @pmap after making sure the map isn't stale.
237 * Pairs with qht_map_unlock_buckets(), hence the pass-by-reference.
239 * Note: callers cannot have ht->lock held.
242 void qht_map_lock_buckets__no_stale(struct qht
*ht
, struct qht_map
**pmap
)
246 map
= atomic_rcu_read(&ht
->map
);
247 qht_map_lock_buckets(map
);
248 if (likely(!qht_map_is_stale__locked(ht
, map
))) {
252 qht_map_unlock_buckets(map
);
254 /* we raced with a resize; acquire ht->lock to see the updated ht->map */
255 qemu_mutex_lock(&ht
->lock
);
257 qht_map_lock_buckets(map
);
258 qemu_mutex_unlock(&ht
->lock
);
264 * Get a head bucket and lock it, making sure its parent map is not stale.
265 * @pmap is filled with a pointer to the bucket's parent map.
267 * Unlock with qemu_spin_unlock(&b->lock).
269 * Note: callers cannot have ht->lock held.
272 struct qht_bucket
*qht_bucket_lock__no_stale(struct qht
*ht
, uint32_t hash
,
273 struct qht_map
**pmap
)
275 struct qht_bucket
*b
;
278 map
= atomic_rcu_read(&ht
->map
);
279 b
= qht_map_to_bucket(map
, hash
);
281 qemu_spin_lock(&b
->lock
);
282 if (likely(!qht_map_is_stale__locked(ht
, map
))) {
286 qemu_spin_unlock(&b
->lock
);
288 /* we raced with a resize; acquire ht->lock to see the updated ht->map */
289 qemu_mutex_lock(&ht
->lock
);
291 b
= qht_map_to_bucket(map
, hash
);
292 qemu_spin_lock(&b
->lock
);
293 qemu_mutex_unlock(&ht
->lock
);
298 static inline bool qht_map_needs_resize(struct qht_map
*map
)
300 return atomic_read(&map
->n_added_buckets
) > map
->n_added_buckets_threshold
;
303 static inline void qht_chain_destroy(struct qht_bucket
*head
)
305 struct qht_bucket
*curr
= head
->next
;
306 struct qht_bucket
*prev
;
315 /* pass only an orphan map */
316 static void qht_map_destroy(struct qht_map
*map
)
320 for (i
= 0; i
< map
->n_buckets
; i
++) {
321 qht_chain_destroy(&map
->buckets
[i
]);
323 qemu_vfree(map
->buckets
);
327 static struct qht_map
*qht_map_create(size_t n_buckets
)
332 map
= g_malloc(sizeof(*map
));
333 map
->n_buckets
= n_buckets
;
335 map
->n_added_buckets
= 0;
336 map
->n_added_buckets_threshold
= n_buckets
/
337 QHT_NR_ADDED_BUCKETS_THRESHOLD_DIV
;
339 /* let tiny hash tables to at least add one non-head bucket */
340 if (unlikely(map
->n_added_buckets_threshold
== 0)) {
341 map
->n_added_buckets_threshold
= 1;
344 map
->buckets
= qemu_memalign(QHT_BUCKET_ALIGN
,
345 sizeof(*map
->buckets
) * n_buckets
);
346 for (i
= 0; i
< n_buckets
; i
++) {
347 qht_head_init(&map
->buckets
[i
]);
352 void qht_init(struct qht
*ht
, size_t n_elems
, unsigned int mode
)
355 size_t n_buckets
= qht_elems_to_buckets(n_elems
);
358 qemu_mutex_init(&ht
->lock
);
359 map
= qht_map_create(n_buckets
);
360 atomic_rcu_set(&ht
->map
, map
);
363 /* call only when there are no readers/writers left */
364 void qht_destroy(struct qht
*ht
)
366 qht_map_destroy(ht
->map
);
367 memset(ht
, 0, sizeof(*ht
));
370 static void qht_bucket_reset__locked(struct qht_bucket
*head
)
372 struct qht_bucket
*b
= head
;
375 seqlock_write_begin(&head
->sequence
);
377 for (i
= 0; i
< QHT_BUCKET_ENTRIES
; i
++) {
378 if (b
->pointers
[i
] == NULL
) {
382 atomic_set(&b
->pointers
[i
], NULL
);
387 seqlock_write_end(&head
->sequence
);
390 /* call with all bucket locks held */
391 static void qht_map_reset__all_locked(struct qht_map
*map
)
395 for (i
= 0; i
< map
->n_buckets
; i
++) {
396 qht_bucket_reset__locked(&map
->buckets
[i
]);
398 qht_map_debug__all_locked(map
);
401 void qht_reset(struct qht
*ht
)
405 qht_map_lock_buckets__no_stale(ht
, &map
);
406 qht_map_reset__all_locked(map
);
407 qht_map_unlock_buckets(map
);
410 bool qht_reset_size(struct qht
*ht
, size_t n_elems
)
417 n_buckets
= qht_elems_to_buckets(n_elems
);
419 qemu_mutex_lock(&ht
->lock
);
421 if (n_buckets
!= map
->n_buckets
) {
422 new = qht_map_create(n_buckets
);
426 qht_map_lock_buckets(map
);
427 qht_map_reset__all_locked(map
);
429 qht_do_resize(ht
, new);
431 qht_map_unlock_buckets(map
);
432 qemu_mutex_unlock(&ht
->lock
);
438 void *qht_do_lookup(struct qht_bucket
*head
, qht_lookup_func_t func
,
439 const void *userp
, uint32_t hash
)
441 struct qht_bucket
*b
= head
;
445 for (i
= 0; i
< QHT_BUCKET_ENTRIES
; i
++) {
446 if (b
->hashes
[i
] == hash
) {
447 void *p
= atomic_read(&b
->pointers
[i
]);
449 if (likely(p
) && likely(func(p
, userp
))) {
454 b
= atomic_rcu_read(&b
->next
);
460 static __attribute__((noinline
))
461 void *qht_lookup__slowpath(struct qht_bucket
*b
, qht_lookup_func_t func
,
462 const void *userp
, uint32_t hash
)
464 unsigned int version
;
468 version
= seqlock_read_begin(&b
->sequence
);
469 ret
= qht_do_lookup(b
, func
, userp
, hash
);
470 } while (seqlock_read_retry(&b
->sequence
, version
));
474 void *qht_lookup(struct qht
*ht
, qht_lookup_func_t func
, const void *userp
,
477 struct qht_bucket
*b
;
479 unsigned int version
;
482 map
= atomic_rcu_read(&ht
->map
);
483 b
= qht_map_to_bucket(map
, hash
);
485 version
= seqlock_read_begin(&b
->sequence
);
486 ret
= qht_do_lookup(b
, func
, userp
, hash
);
487 if (likely(!seqlock_read_retry(&b
->sequence
, version
))) {
491 * Removing the do/while from the fastpath gives a 4% perf. increase when
492 * running a 100%-lookup microbenchmark.
494 return qht_lookup__slowpath(b
, func
, userp
, hash
);
497 /* call with head->lock held */
498 static bool qht_insert__locked(struct qht
*ht
, struct qht_map
*map
,
499 struct qht_bucket
*head
, void *p
, uint32_t hash
,
502 struct qht_bucket
*b
= head
;
503 struct qht_bucket
*prev
= NULL
;
504 struct qht_bucket
*new = NULL
;
508 for (i
= 0; i
< QHT_BUCKET_ENTRIES
; i
++) {
509 if (b
->pointers
[i
]) {
510 if (unlikely(b
->pointers
[i
] == p
)) {
521 b
= qemu_memalign(QHT_BUCKET_ALIGN
, sizeof(*b
));
522 memset(b
, 0, sizeof(*b
));
525 atomic_inc(&map
->n_added_buckets
);
526 if (unlikely(qht_map_needs_resize(map
)) && needs_resize
) {
527 *needs_resize
= true;
531 /* found an empty key: acquire the seqlock and write */
532 seqlock_write_begin(&head
->sequence
);
534 atomic_rcu_set(&prev
->next
, b
);
537 atomic_set(&b
->pointers
[i
], p
);
538 seqlock_write_end(&head
->sequence
);
542 static __attribute__((noinline
)) void qht_grow_maybe(struct qht
*ht
)
547 * If the lock is taken it probably means there's an ongoing resize,
550 if (qemu_mutex_trylock(&ht
->lock
)) {
554 /* another thread might have just performed the resize we were after */
555 if (qht_map_needs_resize(map
)) {
556 struct qht_map
*new = qht_map_create(map
->n_buckets
* 2);
558 qht_map_lock_buckets(map
);
559 qht_do_resize(ht
, new);
560 qht_map_unlock_buckets(map
);
562 qemu_mutex_unlock(&ht
->lock
);
565 bool qht_insert(struct qht
*ht
, void *p
, uint32_t hash
)
567 struct qht_bucket
*b
;
569 bool needs_resize
= false;
572 /* NULL pointers are not supported */
575 b
= qht_bucket_lock__no_stale(ht
, hash
, &map
);
576 ret
= qht_insert__locked(ht
, map
, b
, p
, hash
, &needs_resize
);
577 qht_bucket_debug__locked(b
);
578 qemu_spin_unlock(&b
->lock
);
580 if (unlikely(needs_resize
) && ht
->mode
& QHT_MODE_AUTO_RESIZE
) {
586 static inline bool qht_entry_is_last(struct qht_bucket
*b
, int pos
)
588 if (pos
== QHT_BUCKET_ENTRIES
- 1) {
589 if (b
->next
== NULL
) {
592 return b
->next
->pointers
[0] == NULL
;
594 return b
->pointers
[pos
+ 1] == NULL
;
598 qht_entry_move(struct qht_bucket
*to
, int i
, struct qht_bucket
*from
, int j
)
600 qht_debug_assert(!(to
== from
&& i
== j
));
601 qht_debug_assert(to
->pointers
[i
]);
602 qht_debug_assert(from
->pointers
[j
]);
604 to
->hashes
[i
] = from
->hashes
[j
];
605 atomic_set(&to
->pointers
[i
], from
->pointers
[j
]);
608 atomic_set(&from
->pointers
[j
], NULL
);
612 * Find the last valid entry in @head, and swap it with @orig[pos], which has
613 * just been invalidated.
615 static inline void qht_bucket_remove_entry(struct qht_bucket
*orig
, int pos
)
617 struct qht_bucket
*b
= orig
;
618 struct qht_bucket
*prev
= NULL
;
621 if (qht_entry_is_last(orig
, pos
)) {
622 orig
->hashes
[pos
] = 0;
623 atomic_set(&orig
->pointers
[pos
], NULL
);
627 for (i
= 0; i
< QHT_BUCKET_ENTRIES
; i
++) {
628 if (b
->pointers
[i
]) {
632 return qht_entry_move(orig
, pos
, b
, i
- 1);
634 qht_debug_assert(prev
);
635 return qht_entry_move(orig
, pos
, prev
, QHT_BUCKET_ENTRIES
- 1);
640 /* no free entries other than orig[pos], so swap it with the last one */
641 qht_entry_move(orig
, pos
, prev
, QHT_BUCKET_ENTRIES
- 1);
644 /* call with b->lock held */
646 bool qht_remove__locked(struct qht_map
*map
, struct qht_bucket
*head
,
647 const void *p
, uint32_t hash
)
649 struct qht_bucket
*b
= head
;
653 for (i
= 0; i
< QHT_BUCKET_ENTRIES
; i
++) {
654 void *q
= b
->pointers
[i
];
656 if (unlikely(q
== NULL
)) {
660 qht_debug_assert(b
->hashes
[i
] == hash
);
661 seqlock_write_begin(&head
->sequence
);
662 qht_bucket_remove_entry(b
, i
);
663 seqlock_write_end(&head
->sequence
);
672 bool qht_remove(struct qht
*ht
, const void *p
, uint32_t hash
)
674 struct qht_bucket
*b
;
678 /* NULL pointers are not supported */
681 b
= qht_bucket_lock__no_stale(ht
, hash
, &map
);
682 ret
= qht_remove__locked(map
, b
, p
, hash
);
683 qht_bucket_debug__locked(b
);
684 qemu_spin_unlock(&b
->lock
);
688 static inline void qht_bucket_iter(struct qht
*ht
, struct qht_bucket
*b
,
689 qht_iter_func_t func
, void *userp
)
694 for (i
= 0; i
< QHT_BUCKET_ENTRIES
; i
++) {
695 if (b
->pointers
[i
] == NULL
) {
698 func(ht
, b
->pointers
[i
], b
->hashes
[i
], userp
);
704 /* call with all of the map's locks held */
705 static inline void qht_map_iter__all_locked(struct qht
*ht
, struct qht_map
*map
,
706 qht_iter_func_t func
, void *userp
)
710 for (i
= 0; i
< map
->n_buckets
; i
++) {
711 qht_bucket_iter(ht
, &map
->buckets
[i
], func
, userp
);
715 void qht_iter(struct qht
*ht
, qht_iter_func_t func
, void *userp
)
719 map
= atomic_rcu_read(&ht
->map
);
720 qht_map_lock_buckets(map
);
721 /* Note: ht here is merely for carrying ht->mode; ht->map won't be read */
722 qht_map_iter__all_locked(ht
, map
, func
, userp
);
723 qht_map_unlock_buckets(map
);
726 static void qht_map_copy(struct qht
*ht
, void *p
, uint32_t hash
, void *userp
)
728 struct qht_map
*new = userp
;
729 struct qht_bucket
*b
= qht_map_to_bucket(new, hash
);
731 /* no need to acquire b->lock because no thread has seen this map yet */
732 qht_insert__locked(ht
, new, b
, p
, hash
, NULL
);
736 * Call with ht->lock and all bucket locks held.
738 * Creating the @new map here would add unnecessary delay while all the locks
739 * are held--holding up the bucket locks is particularly bad, since no writes
740 * can occur while these are held. Thus, we let callers create the new map,
741 * hopefully without the bucket locks held.
743 static void qht_do_resize(struct qht
*ht
, struct qht_map
*new)
748 g_assert_cmpuint(new->n_buckets
, !=, old
->n_buckets
);
750 qht_map_iter__all_locked(ht
, old
, qht_map_copy
, new);
751 qht_map_debug__all_locked(new);
753 atomic_rcu_set(&ht
->map
, new);
754 call_rcu(old
, qht_map_destroy
, rcu
);
757 bool qht_resize(struct qht
*ht
, size_t n_elems
)
759 size_t n_buckets
= qht_elems_to_buckets(n_elems
);
762 qemu_mutex_lock(&ht
->lock
);
763 if (n_buckets
!= ht
->map
->n_buckets
) {
765 struct qht_map
*old
= ht
->map
;
767 new = qht_map_create(n_buckets
);
768 qht_map_lock_buckets(old
);
769 qht_do_resize(ht
, new);
770 qht_map_unlock_buckets(old
);
773 qemu_mutex_unlock(&ht
->lock
);
778 /* pass @stats to qht_statistics_destroy() when done */
779 void qht_statistics_init(struct qht
*ht
, struct qht_stats
*stats
)
784 map
= atomic_rcu_read(&ht
->map
);
786 stats
->head_buckets
= map
->n_buckets
;
787 stats
->used_head_buckets
= 0;
789 qdist_init(&stats
->chain
);
790 qdist_init(&stats
->occupancy
);
792 for (i
= 0; i
< map
->n_buckets
; i
++) {
793 struct qht_bucket
*head
= &map
->buckets
[i
];
794 struct qht_bucket
*b
;
795 unsigned int version
;
801 version
= seqlock_read_begin(&head
->sequence
);
806 for (j
= 0; j
< QHT_BUCKET_ENTRIES
; j
++) {
807 if (atomic_read(&b
->pointers
[j
]) == NULL
) {
813 b
= atomic_rcu_read(&b
->next
);
815 } while (seqlock_read_retry(&head
->sequence
, version
));
818 qdist_inc(&stats
->chain
, buckets
);
819 qdist_inc(&stats
->occupancy
,
820 (double)entries
/ QHT_BUCKET_ENTRIES
/ buckets
);
821 stats
->used_head_buckets
++;
822 stats
->entries
+= entries
;
824 qdist_inc(&stats
->occupancy
, 0);
829 void qht_statistics_destroy(struct qht_stats
*stats
)
831 qdist_destroy(&stats
->occupancy
);
832 qdist_destroy(&stats
->chain
);