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/
68 #include "qemu/osdep.h"
70 #include "qemu/atomic.h"
76 * We want to avoid false sharing of cache lines. Most systems have 64-byte
77 * cache lines so we go with it for simplicity.
79 * Note that systems with smaller cache lines will be fine (the struct is
80 * almost 64-bytes); systems with larger cache lines might suffer from
83 #define QHT_BUCKET_ALIGN 64
85 /* define these to keep sizeof(qht_bucket) within QHT_BUCKET_ALIGN */
86 #if HOST_LONG_BITS == 32
87 #define QHT_BUCKET_ENTRIES 6
89 #define QHT_BUCKET_ENTRIES 4
93 * Note: reading partially-updated pointers in @pointers could lead to
94 * segfaults. We thus access them with atomic_read/set; this guarantees
95 * that the compiler makes all those accesses atomic. We also need the
96 * volatile-like behavior in atomic_read, since otherwise the compiler
97 * might refetch the pointer.
98 * atomic_read's are of course not necessary when the bucket lock is held.
100 * If both ht->lock and b->lock are grabbed, ht->lock should always
105 QemuSeqLock sequence
;
106 uint32_t hashes
[QHT_BUCKET_ENTRIES
];
107 void *pointers
[QHT_BUCKET_ENTRIES
];
108 struct qht_bucket
*next
;
109 } QEMU_ALIGNED(QHT_BUCKET_ALIGN
);
111 QEMU_BUILD_BUG_ON(sizeof(struct qht_bucket
) > QHT_BUCKET_ALIGN
);
114 * struct qht_map - structure to track an array of buckets
115 * @rcu: used by RCU. Keep it as the top field in the struct to help valgrind
116 * find the whole struct.
117 * @buckets: array of head buckets. It is constant once the map is created.
118 * @n_buckets: number of head buckets. It is constant once the map is created.
119 * @n_added_buckets: number of added (i.e. "non-head") buckets
120 * @n_added_buckets_threshold: threshold to trigger an upward resize once the
121 * number of added buckets surpasses it.
123 * Buckets are tracked in what we call a "map", i.e. this structure.
127 struct qht_bucket
*buckets
;
129 size_t n_added_buckets
;
130 size_t n_added_buckets_threshold
;
133 /* trigger a resize when n_added_buckets > n_buckets / div */
134 #define QHT_NR_ADDED_BUCKETS_THRESHOLD_DIV 8
136 static void qht_do_resize(struct qht
*ht
, struct qht_map
*new);
137 static void qht_grow_maybe(struct qht
*ht
);
141 #define qht_debug_assert(X) do { assert(X); } while (0)
143 static void qht_bucket_debug__locked(struct qht_bucket
*b
)
145 bool seen_empty
= false;
146 bool corrupt
= false;
150 for (i
= 0; i
< QHT_BUCKET_ENTRIES
; i
++) {
151 if (b
->pointers
[i
] == NULL
) {
156 fprintf(stderr
, "%s: b: %p, pos: %i, hash: 0x%x, p: %p\n",
157 __func__
, b
, i
, b
->hashes
[i
], b
->pointers
[i
]);
163 qht_debug_assert(!corrupt
);
166 static void qht_map_debug__all_locked(struct qht_map
*map
)
170 for (i
= 0; i
< map
->n_buckets
; i
++) {
171 qht_bucket_debug__locked(&map
->buckets
[i
]);
176 #define qht_debug_assert(X) do { (void)(X); } while (0)
178 static inline void qht_bucket_debug__locked(struct qht_bucket
*b
)
181 static inline void qht_map_debug__all_locked(struct qht_map
*map
)
183 #endif /* QHT_DEBUG */
185 static inline size_t qht_elems_to_buckets(size_t n_elems
)
187 return pow2ceil(n_elems
/ QHT_BUCKET_ENTRIES
);
190 static inline void qht_head_init(struct qht_bucket
*b
)
192 memset(b
, 0, sizeof(*b
));
193 qemu_spin_init(&b
->lock
);
194 seqlock_init(&b
->sequence
);
198 struct qht_bucket
*qht_map_to_bucket(struct qht_map
*map
, uint32_t hash
)
200 return &map
->buckets
[hash
& (map
->n_buckets
- 1)];
203 /* acquire all bucket locks from a map */
204 static void qht_map_lock_buckets(struct qht_map
*map
)
208 for (i
= 0; i
< map
->n_buckets
; i
++) {
209 struct qht_bucket
*b
= &map
->buckets
[i
];
211 qemu_spin_lock(&b
->lock
);
215 static void qht_map_unlock_buckets(struct qht_map
*map
)
219 for (i
= 0; i
< map
->n_buckets
; i
++) {
220 struct qht_bucket
*b
= &map
->buckets
[i
];
222 qemu_spin_unlock(&b
->lock
);
227 * Call with at least a bucket lock held.
228 * @map should be the value read before acquiring the lock (or locks).
230 static inline bool qht_map_is_stale__locked(struct qht
*ht
, struct qht_map
*map
)
232 return map
!= ht
->map
;
236 * Grab all bucket locks, and set @pmap after making sure the map isn't stale.
238 * Pairs with qht_map_unlock_buckets(), hence the pass-by-reference.
240 * Note: callers cannot have ht->lock held.
243 void qht_map_lock_buckets__no_stale(struct qht
*ht
, struct qht_map
**pmap
)
247 map
= atomic_rcu_read(&ht
->map
);
248 qht_map_lock_buckets(map
);
249 if (likely(!qht_map_is_stale__locked(ht
, map
))) {
253 qht_map_unlock_buckets(map
);
255 /* we raced with a resize; acquire ht->lock to see the updated ht->map */
256 qemu_mutex_lock(&ht
->lock
);
258 qht_map_lock_buckets(map
);
259 qemu_mutex_unlock(&ht
->lock
);
265 * Get a head bucket and lock it, making sure its parent map is not stale.
266 * @pmap is filled with a pointer to the bucket's parent map.
268 * Unlock with qemu_spin_unlock(&b->lock).
270 * Note: callers cannot have ht->lock held.
273 struct qht_bucket
*qht_bucket_lock__no_stale(struct qht
*ht
, uint32_t hash
,
274 struct qht_map
**pmap
)
276 struct qht_bucket
*b
;
279 map
= atomic_rcu_read(&ht
->map
);
280 b
= qht_map_to_bucket(map
, hash
);
282 qemu_spin_lock(&b
->lock
);
283 if (likely(!qht_map_is_stale__locked(ht
, map
))) {
287 qemu_spin_unlock(&b
->lock
);
289 /* we raced with a resize; acquire ht->lock to see the updated ht->map */
290 qemu_mutex_lock(&ht
->lock
);
292 b
= qht_map_to_bucket(map
, hash
);
293 qemu_spin_lock(&b
->lock
);
294 qemu_mutex_unlock(&ht
->lock
);
299 static inline bool qht_map_needs_resize(struct qht_map
*map
)
301 return atomic_read(&map
->n_added_buckets
) > map
->n_added_buckets_threshold
;
304 static inline void qht_chain_destroy(struct qht_bucket
*head
)
306 struct qht_bucket
*curr
= head
->next
;
307 struct qht_bucket
*prev
;
316 /* pass only an orphan map */
317 static void qht_map_destroy(struct qht_map
*map
)
321 for (i
= 0; i
< map
->n_buckets
; i
++) {
322 qht_chain_destroy(&map
->buckets
[i
]);
324 qemu_vfree(map
->buckets
);
328 static struct qht_map
*qht_map_create(size_t n_buckets
)
333 map
= g_malloc(sizeof(*map
));
334 map
->n_buckets
= n_buckets
;
336 map
->n_added_buckets
= 0;
337 map
->n_added_buckets_threshold
= n_buckets
/
338 QHT_NR_ADDED_BUCKETS_THRESHOLD_DIV
;
340 /* let tiny hash tables to at least add one non-head bucket */
341 if (unlikely(map
->n_added_buckets_threshold
== 0)) {
342 map
->n_added_buckets_threshold
= 1;
345 map
->buckets
= qemu_memalign(QHT_BUCKET_ALIGN
,
346 sizeof(*map
->buckets
) * n_buckets
);
347 for (i
= 0; i
< n_buckets
; i
++) {
348 qht_head_init(&map
->buckets
[i
]);
353 void qht_init(struct qht
*ht
, size_t n_elems
, unsigned int mode
)
356 size_t n_buckets
= qht_elems_to_buckets(n_elems
);
359 qemu_mutex_init(&ht
->lock
);
360 map
= qht_map_create(n_buckets
);
361 atomic_rcu_set(&ht
->map
, map
);
364 /* call only when there are no readers/writers left */
365 void qht_destroy(struct qht
*ht
)
367 qht_map_destroy(ht
->map
);
368 memset(ht
, 0, sizeof(*ht
));
371 static void qht_bucket_reset__locked(struct qht_bucket
*head
)
373 struct qht_bucket
*b
= head
;
376 seqlock_write_begin(&head
->sequence
);
378 for (i
= 0; i
< QHT_BUCKET_ENTRIES
; i
++) {
379 if (b
->pointers
[i
] == NULL
) {
383 atomic_set(&b
->pointers
[i
], NULL
);
388 seqlock_write_end(&head
->sequence
);
391 /* call with all bucket locks held */
392 static void qht_map_reset__all_locked(struct qht_map
*map
)
396 for (i
= 0; i
< map
->n_buckets
; i
++) {
397 qht_bucket_reset__locked(&map
->buckets
[i
]);
399 qht_map_debug__all_locked(map
);
402 void qht_reset(struct qht
*ht
)
406 qht_map_lock_buckets__no_stale(ht
, &map
);
407 qht_map_reset__all_locked(map
);
408 qht_map_unlock_buckets(map
);
411 bool qht_reset_size(struct qht
*ht
, size_t n_elems
)
418 n_buckets
= qht_elems_to_buckets(n_elems
);
420 qemu_mutex_lock(&ht
->lock
);
422 if (n_buckets
!= map
->n_buckets
) {
423 new = qht_map_create(n_buckets
);
427 qht_map_lock_buckets(map
);
428 qht_map_reset__all_locked(map
);
430 qht_do_resize(ht
, new);
432 qht_map_unlock_buckets(map
);
433 qemu_mutex_unlock(&ht
->lock
);
439 void *qht_do_lookup(struct qht_bucket
*head
, qht_lookup_func_t func
,
440 const void *userp
, uint32_t hash
)
442 struct qht_bucket
*b
= head
;
446 for (i
= 0; i
< QHT_BUCKET_ENTRIES
; i
++) {
447 if (b
->hashes
[i
] == hash
) {
448 void *p
= atomic_read(&b
->pointers
[i
]);
450 if (likely(p
) && likely(func(p
, userp
))) {
455 b
= atomic_rcu_read(&b
->next
);
461 static __attribute__((noinline
))
462 void *qht_lookup__slowpath(struct qht_bucket
*b
, qht_lookup_func_t func
,
463 const void *userp
, uint32_t hash
)
465 unsigned int version
;
469 version
= seqlock_read_begin(&b
->sequence
);
470 ret
= qht_do_lookup(b
, func
, userp
, hash
);
471 } while (seqlock_read_retry(&b
->sequence
, version
));
475 void *qht_lookup(struct qht
*ht
, qht_lookup_func_t func
, const void *userp
,
478 struct qht_bucket
*b
;
480 unsigned int version
;
483 map
= atomic_rcu_read(&ht
->map
);
484 b
= qht_map_to_bucket(map
, hash
);
486 version
= seqlock_read_begin(&b
->sequence
);
487 ret
= qht_do_lookup(b
, func
, userp
, hash
);
488 if (likely(!seqlock_read_retry(&b
->sequence
, version
))) {
492 * Removing the do/while from the fastpath gives a 4% perf. increase when
493 * running a 100%-lookup microbenchmark.
495 return qht_lookup__slowpath(b
, func
, userp
, hash
);
498 /* call with head->lock held */
499 static bool qht_insert__locked(struct qht
*ht
, struct qht_map
*map
,
500 struct qht_bucket
*head
, void *p
, uint32_t hash
,
503 struct qht_bucket
*b
= head
;
504 struct qht_bucket
*prev
= NULL
;
505 struct qht_bucket
*new = NULL
;
509 for (i
= 0; i
< QHT_BUCKET_ENTRIES
; i
++) {
510 if (b
->pointers
[i
]) {
511 if (unlikely(b
->pointers
[i
] == p
)) {
522 b
= qemu_memalign(QHT_BUCKET_ALIGN
, sizeof(*b
));
523 memset(b
, 0, sizeof(*b
));
526 atomic_inc(&map
->n_added_buckets
);
527 if (unlikely(qht_map_needs_resize(map
)) && needs_resize
) {
528 *needs_resize
= true;
532 /* found an empty key: acquire the seqlock and write */
533 seqlock_write_begin(&head
->sequence
);
535 atomic_rcu_set(&prev
->next
, b
);
538 atomic_set(&b
->pointers
[i
], p
);
539 seqlock_write_end(&head
->sequence
);
543 static __attribute__((noinline
)) void qht_grow_maybe(struct qht
*ht
)
548 * If the lock is taken it probably means there's an ongoing resize,
551 if (qemu_mutex_trylock(&ht
->lock
)) {
555 /* another thread might have just performed the resize we were after */
556 if (qht_map_needs_resize(map
)) {
557 struct qht_map
*new = qht_map_create(map
->n_buckets
* 2);
559 qht_map_lock_buckets(map
);
560 qht_do_resize(ht
, new);
561 qht_map_unlock_buckets(map
);
563 qemu_mutex_unlock(&ht
->lock
);
566 bool qht_insert(struct qht
*ht
, void *p
, uint32_t hash
)
568 struct qht_bucket
*b
;
570 bool needs_resize
= false;
573 /* NULL pointers are not supported */
576 b
= qht_bucket_lock__no_stale(ht
, hash
, &map
);
577 ret
= qht_insert__locked(ht
, map
, b
, p
, hash
, &needs_resize
);
578 qht_bucket_debug__locked(b
);
579 qemu_spin_unlock(&b
->lock
);
581 if (unlikely(needs_resize
) && ht
->mode
& QHT_MODE_AUTO_RESIZE
) {
587 static inline bool qht_entry_is_last(struct qht_bucket
*b
, int pos
)
589 if (pos
== QHT_BUCKET_ENTRIES
- 1) {
590 if (b
->next
== NULL
) {
593 return b
->next
->pointers
[0] == NULL
;
595 return b
->pointers
[pos
+ 1] == NULL
;
599 qht_entry_move(struct qht_bucket
*to
, int i
, struct qht_bucket
*from
, int j
)
601 qht_debug_assert(!(to
== from
&& i
== j
));
602 qht_debug_assert(to
->pointers
[i
]);
603 qht_debug_assert(from
->pointers
[j
]);
605 to
->hashes
[i
] = from
->hashes
[j
];
606 atomic_set(&to
->pointers
[i
], from
->pointers
[j
]);
609 atomic_set(&from
->pointers
[j
], NULL
);
613 * Find the last valid entry in @head, and swap it with @orig[pos], which has
614 * just been invalidated.
616 static inline void qht_bucket_remove_entry(struct qht_bucket
*orig
, int pos
)
618 struct qht_bucket
*b
= orig
;
619 struct qht_bucket
*prev
= NULL
;
622 if (qht_entry_is_last(orig
, pos
)) {
623 orig
->hashes
[pos
] = 0;
624 atomic_set(&orig
->pointers
[pos
], NULL
);
628 for (i
= 0; i
< QHT_BUCKET_ENTRIES
; i
++) {
629 if (b
->pointers
[i
]) {
633 return qht_entry_move(orig
, pos
, b
, i
- 1);
635 qht_debug_assert(prev
);
636 return qht_entry_move(orig
, pos
, prev
, QHT_BUCKET_ENTRIES
- 1);
641 /* no free entries other than orig[pos], so swap it with the last one */
642 qht_entry_move(orig
, pos
, prev
, QHT_BUCKET_ENTRIES
- 1);
645 /* call with b->lock held */
647 bool qht_remove__locked(struct qht_map
*map
, struct qht_bucket
*head
,
648 const void *p
, uint32_t hash
)
650 struct qht_bucket
*b
= head
;
654 for (i
= 0; i
< QHT_BUCKET_ENTRIES
; i
++) {
655 void *q
= b
->pointers
[i
];
657 if (unlikely(q
== NULL
)) {
661 qht_debug_assert(b
->hashes
[i
] == hash
);
662 seqlock_write_begin(&head
->sequence
);
663 qht_bucket_remove_entry(b
, i
);
664 seqlock_write_end(&head
->sequence
);
673 bool qht_remove(struct qht
*ht
, const void *p
, uint32_t hash
)
675 struct qht_bucket
*b
;
679 /* NULL pointers are not supported */
682 b
= qht_bucket_lock__no_stale(ht
, hash
, &map
);
683 ret
= qht_remove__locked(map
, b
, p
, hash
);
684 qht_bucket_debug__locked(b
);
685 qemu_spin_unlock(&b
->lock
);
689 static inline void qht_bucket_iter(struct qht
*ht
, struct qht_bucket
*b
,
690 qht_iter_func_t func
, void *userp
)
695 for (i
= 0; i
< QHT_BUCKET_ENTRIES
; i
++) {
696 if (b
->pointers
[i
] == NULL
) {
699 func(ht
, b
->pointers
[i
], b
->hashes
[i
], userp
);
705 /* call with all of the map's locks held */
706 static inline void qht_map_iter__all_locked(struct qht
*ht
, struct qht_map
*map
,
707 qht_iter_func_t func
, void *userp
)
711 for (i
= 0; i
< map
->n_buckets
; i
++) {
712 qht_bucket_iter(ht
, &map
->buckets
[i
], func
, userp
);
716 void qht_iter(struct qht
*ht
, qht_iter_func_t func
, void *userp
)
720 map
= atomic_rcu_read(&ht
->map
);
721 qht_map_lock_buckets(map
);
722 /* Note: ht here is merely for carrying ht->mode; ht->map won't be read */
723 qht_map_iter__all_locked(ht
, map
, func
, userp
);
724 qht_map_unlock_buckets(map
);
727 static void qht_map_copy(struct qht
*ht
, void *p
, uint32_t hash
, void *userp
)
729 struct qht_map
*new = userp
;
730 struct qht_bucket
*b
= qht_map_to_bucket(new, hash
);
732 /* no need to acquire b->lock because no thread has seen this map yet */
733 qht_insert__locked(ht
, new, b
, p
, hash
, NULL
);
737 * Call with ht->lock and all bucket locks held.
739 * Creating the @new map here would add unnecessary delay while all the locks
740 * are held--holding up the bucket locks is particularly bad, since no writes
741 * can occur while these are held. Thus, we let callers create the new map,
742 * hopefully without the bucket locks held.
744 static void qht_do_resize(struct qht
*ht
, struct qht_map
*new)
749 g_assert_cmpuint(new->n_buckets
, !=, old
->n_buckets
);
751 qht_map_iter__all_locked(ht
, old
, qht_map_copy
, new);
752 qht_map_debug__all_locked(new);
754 atomic_rcu_set(&ht
->map
, new);
755 call_rcu(old
, qht_map_destroy
, rcu
);
758 bool qht_resize(struct qht
*ht
, size_t n_elems
)
760 size_t n_buckets
= qht_elems_to_buckets(n_elems
);
763 qemu_mutex_lock(&ht
->lock
);
764 if (n_buckets
!= ht
->map
->n_buckets
) {
766 struct qht_map
*old
= ht
->map
;
768 new = qht_map_create(n_buckets
);
769 qht_map_lock_buckets(old
);
770 qht_do_resize(ht
, new);
771 qht_map_unlock_buckets(old
);
774 qemu_mutex_unlock(&ht
->lock
);
779 /* pass @stats to qht_statistics_destroy() when done */
780 void qht_statistics_init(struct qht
*ht
, struct qht_stats
*stats
)
785 map
= atomic_rcu_read(&ht
->map
);
787 stats
->head_buckets
= map
->n_buckets
;
788 stats
->used_head_buckets
= 0;
790 qdist_init(&stats
->chain
);
791 qdist_init(&stats
->occupancy
);
793 for (i
= 0; i
< map
->n_buckets
; i
++) {
794 struct qht_bucket
*head
= &map
->buckets
[i
];
795 struct qht_bucket
*b
;
796 unsigned int version
;
802 version
= seqlock_read_begin(&head
->sequence
);
807 for (j
= 0; j
< QHT_BUCKET_ENTRIES
; j
++) {
808 if (atomic_read(&b
->pointers
[j
]) == NULL
) {
814 b
= atomic_rcu_read(&b
->next
);
816 } while (seqlock_read_retry(&head
->sequence
, version
));
819 qdist_inc(&stats
->chain
, buckets
);
820 qdist_inc(&stats
->occupancy
,
821 (double)entries
/ QHT_BUCKET_ENTRIES
/ buckets
);
822 stats
->used_head_buckets
++;
823 stats
->entries
+= entries
;
825 qdist_inc(&stats
->occupancy
, 0);
830 void qht_statistics_destroy(struct qht_stats
*stats
)
832 qdist_destroy(&stats
->occupancy
);
833 qdist_destroy(&stats
->chain
);