1 /*
2 * qht.c - QEMU Hash Table, designed to scale for read-mostly workloads.
3 *
4 * Copyright (C) 2016, Emilio G. Cota <cota@braap.org>
5 *
6 * License: GNU GPL, version 2 or later.
7 * See the COPYING file in the top-level directory.
8 *
9 * Assumptions:
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.
16 *
17 * Features:
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.
27 *
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.
34 *
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.
41 *
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.
45 *
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
49 * it anymore.
50 *
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 occurred
53 * while the bucket spinlock was being acquired.
54 *
55 * Related Work:
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/
67 */
68 #include "qemu/osdep.h"
69 #include "qemu/qht.h"
70 #include "qemu/atomic.h"
71 #include "qemu/rcu.h"
72 #include "qemu/memalign.h"
73
74 //#define QHT_DEBUG
75
76 /*
77 * We want to avoid false sharing of cache lines. Most systems have 64-byte
78 * cache lines so we go with it for simplicity.
79 *
80 * Note that systems with smaller cache lines will be fine (the struct is
81 * almost 64-bytes); systems with larger cache lines might suffer from
82 * some false sharing.
83 */
84 #define QHT_BUCKET_ALIGN 64
85
86 /* define these to keep sizeof(qht_bucket) within QHT_BUCKET_ALIGN */
87 #if HOST_LONG_BITS == 32
88 #define QHT_BUCKET_ENTRIES 6
89 #else /* 64-bit */
90 #define QHT_BUCKET_ENTRIES 4
91 #endif
92
93 enum qht_iter_type {
94 QHT_ITER_VOID, /* do nothing; use retvoid */
95 QHT_ITER_RM, /* remove element if retbool returns true */
96 };
97
98 struct qht_iter {
99 union {
100 qht_iter_func_t retvoid;
101 qht_iter_bool_func_t retbool;
102 } f;
103 enum qht_iter_type type;
104 };
105
106 /*
107 * Do _not_ use qemu_mutex_[try]lock directly! Use these macros, otherwise
108 * the profiler (QSP) will deadlock.
109 */
qht_lock(struct qht * ht)110 static inline void qht_lock(struct qht *ht)
111 {
112 if (ht->mode & QHT_MODE_RAW_MUTEXES) {
113 qemu_mutex_lock__raw(&ht->lock);
114 } else {
115 qemu_mutex_lock(&ht->lock);
116 }
117 }
118
qht_trylock(struct qht * ht)119 static inline int qht_trylock(struct qht *ht)
120 {
121 if (ht->mode & QHT_MODE_RAW_MUTEXES) {
122 return qemu_mutex_trylock__raw(&(ht)->lock);
123 }
124 return qemu_mutex_trylock(&(ht)->lock);
125 }
126
127 /* this inline is not really necessary, but it helps keep code consistent */
qht_unlock(struct qht * ht)128 static inline void qht_unlock(struct qht *ht)
129 {
130 qemu_mutex_unlock(&ht->lock);
131 }
132
133 /*
134 * Note: reading partially-updated pointers in @pointers could lead to
135 * segfaults. We thus access them with qatomic_read/set; this guarantees
136 * that the compiler makes all those accesses atomic. We also need the
137 * volatile-like behavior in qatomic_read, since otherwise the compiler
138 * might refetch the pointer.
139 * qatomic_read's are of course not necessary when the bucket lock is held.
140 *
141 * If both ht->lock and b->lock are grabbed, ht->lock should always
142 * be grabbed first.
143 */
144 struct qht_bucket {
145 QemuSpin lock;
146 QemuSeqLock sequence;
147 uint32_t hashes[QHT_BUCKET_ENTRIES];
148 void *pointers[QHT_BUCKET_ENTRIES];
149 struct qht_bucket *next;
150 } QEMU_ALIGNED(QHT_BUCKET_ALIGN);
151
152 QEMU_BUILD_BUG_ON(sizeof(struct qht_bucket) > QHT_BUCKET_ALIGN);
153
154 /*
155 * Under TSAN, we use striped locks instead of one lock per bucket chain.
156 * This avoids crashing under TSAN, since TSAN aborts the program if more than
157 * 64 locks are held (this is a hardcoded limit in TSAN).
158 * When resizing a QHT we grab all the buckets' locks, which can easily
159 * go over TSAN's limit. By using striped locks, we avoid this problem.
160 *
161 * Note: this number must be a power of two for easy index computation.
162 */
163 #define QHT_TSAN_BUCKET_LOCKS_BITS 4
164 #define QHT_TSAN_BUCKET_LOCKS (1 << QHT_TSAN_BUCKET_LOCKS_BITS)
165
166 struct qht_tsan_lock {
167 QemuSpin lock;
168 } QEMU_ALIGNED(QHT_BUCKET_ALIGN);
169
170 /**
171 * struct qht_map - structure to track an array of buckets
172 * @rcu: used by RCU. Keep it as the top field in the struct to help valgrind
173 * find the whole struct.
174 * @buckets: array of head buckets. It is constant once the map is created.
175 * @n_buckets: number of head buckets. It is constant once the map is created.
176 * @n_added_buckets: number of added (i.e. "non-head") buckets
177 * @n_added_buckets_threshold: threshold to trigger an upward resize once the
178 * number of added buckets surpasses it.
179 * @tsan_bucket_locks: Array of striped locks to be used only under TSAN.
180 *
181 * Buckets are tracked in what we call a "map", i.e. this structure.
182 */
183 struct qht_map {
184 struct rcu_head rcu;
185 struct qht_bucket *buckets;
186 size_t n_buckets;
187 size_t n_added_buckets;
188 size_t n_added_buckets_threshold;
189 #ifdef CONFIG_TSAN
190 struct qht_tsan_lock tsan_bucket_locks[QHT_TSAN_BUCKET_LOCKS];
191 #endif
192 };
193
194 /* trigger a resize when n_added_buckets > n_buckets / div */
195 #define QHT_NR_ADDED_BUCKETS_THRESHOLD_DIV 8
196
197 static void qht_do_resize_reset(struct qht *ht, struct qht_map *new,
198 bool reset);
199 static void qht_grow_maybe(struct qht *ht);
200
201 #ifdef QHT_DEBUG
202
203 #define qht_debug_assert(X) do { assert(X); } while (0)
204
qht_bucket_debug__locked(struct qht_bucket * b)205 static void qht_bucket_debug__locked(struct qht_bucket *b)
206 {
207 bool seen_empty = false;
208 bool corrupt = false;
209 int i;
210
211 do {
212 for (i = 0; i < QHT_BUCKET_ENTRIES; i++) {
213 if (b->pointers[i] == NULL) {
214 seen_empty = true;
215 continue;
216 }
217 if (seen_empty) {
218 fprintf(stderr, "%s: b: %p, pos: %i, hash: 0x%x, p: %p\n",
219 __func__, b, i, b->hashes[i], b->pointers[i]);
220 corrupt = true;
221 }
222 }
223 b = b->next;
224 } while (b);
225 qht_debug_assert(!corrupt);
226 }
227
qht_map_debug__all_locked(struct qht_map * map)228 static void qht_map_debug__all_locked(struct qht_map *map)
229 {
230 int i;
231
232 for (i = 0; i < map->n_buckets; i++) {
233 qht_bucket_debug__locked(&map->buckets[i]);
234 }
235 }
236 #else
237
238 #define qht_debug_assert(X) do { (void)(X); } while (0)
239
qht_bucket_debug__locked(struct qht_bucket * b)240 static inline void qht_bucket_debug__locked(struct qht_bucket *b)
241 { }
242
qht_map_debug__all_locked(struct qht_map * map)243 static inline void qht_map_debug__all_locked(struct qht_map *map)
244 { }
245 #endif /* QHT_DEBUG */
246
qht_elems_to_buckets(size_t n_elems)247 static inline size_t qht_elems_to_buckets(size_t n_elems)
248 {
249 return pow2ceil(n_elems / QHT_BUCKET_ENTRIES);
250 }
251
252 /*
253 * When using striped locks (i.e. under TSAN), we have to be careful not
254 * to operate on the same lock twice (e.g. when iterating through all buckets).
255 * We achieve this by operating only on each stripe's first matching lock.
256 */
qht_do_if_first_in_stripe(struct qht_map * map,struct qht_bucket * b,void (* func)(QemuSpin * spin))257 static inline void qht_do_if_first_in_stripe(struct qht_map *map,
258 struct qht_bucket *b,
259 void (*func)(QemuSpin *spin))
260 {
261 #ifdef CONFIG_TSAN
262 unsigned long bucket_idx = b - map->buckets;
263 bool is_first_in_stripe = (bucket_idx >> QHT_TSAN_BUCKET_LOCKS_BITS) == 0;
264 if (is_first_in_stripe) {
265 unsigned long lock_idx = bucket_idx & (QHT_TSAN_BUCKET_LOCKS - 1);
266 func(&map->tsan_bucket_locks[lock_idx].lock);
267 }
268 #else
269 func(&b->lock);
270 #endif
271 }
272
qht_bucket_lock_do(struct qht_map * map,struct qht_bucket * b,void (* func)(QemuSpin * lock))273 static inline void qht_bucket_lock_do(struct qht_map *map,
274 struct qht_bucket *b,
275 void (*func)(QemuSpin *lock))
276 {
277 #ifdef CONFIG_TSAN
278 unsigned long bucket_idx = b - map->buckets;
279 unsigned long lock_idx = bucket_idx & (QHT_TSAN_BUCKET_LOCKS - 1);
280 func(&map->tsan_bucket_locks[lock_idx].lock);
281 #else
282 func(&b->lock);
283 #endif
284 }
285
qht_bucket_lock(struct qht_map * map,struct qht_bucket * b)286 static inline void qht_bucket_lock(struct qht_map *map,
287 struct qht_bucket *b)
288 {
289 qht_bucket_lock_do(map, b, qemu_spin_lock);
290 }
291
qht_bucket_unlock(struct qht_map * map,struct qht_bucket * b)292 static inline void qht_bucket_unlock(struct qht_map *map,
293 struct qht_bucket *b)
294 {
295 qht_bucket_lock_do(map, b, qemu_spin_unlock);
296 }
297
qht_head_init(struct qht_map * map,struct qht_bucket * b)298 static inline void qht_head_init(struct qht_map *map, struct qht_bucket *b)
299 {
300 memset(b, 0, sizeof(*b));
301 qht_do_if_first_in_stripe(map, b, qemu_spin_init);
302 seqlock_init(&b->sequence);
303 }
304
305 static inline
qht_map_to_bucket(const struct qht_map * map,uint32_t hash)306 struct qht_bucket *qht_map_to_bucket(const struct qht_map *map, uint32_t hash)
307 {
308 return &map->buckets[hash & (map->n_buckets - 1)];
309 }
310
311 /* acquire all bucket locks from a map */
qht_map_lock_buckets(struct qht_map * map)312 static void qht_map_lock_buckets(struct qht_map *map)
313 {
314 size_t i;
315
316 for (i = 0; i < map->n_buckets; i++) {
317 struct qht_bucket *b = &map->buckets[i];
318
319 qht_do_if_first_in_stripe(map, b, qemu_spin_lock);
320 }
321 }
322
qht_map_unlock_buckets(struct qht_map * map)323 static void qht_map_unlock_buckets(struct qht_map *map)
324 {
325 size_t i;
326
327 for (i = 0; i < map->n_buckets; i++) {
328 struct qht_bucket *b = &map->buckets[i];
329
330 qht_do_if_first_in_stripe(map, b, qemu_spin_unlock);
331 }
332 }
333
334 /*
335 * Call with at least a bucket lock held.
336 * @map should be the value read before acquiring the lock (or locks).
337 */
qht_map_is_stale__locked(const struct qht * ht,const struct qht_map * map)338 static inline bool qht_map_is_stale__locked(const struct qht *ht,
339 const struct qht_map *map)
340 {
341 return map != ht->map;
342 }
343
344 /*
345 * Grab all bucket locks, and set @pmap after making sure the map isn't stale.
346 *
347 * Pairs with qht_map_unlock_buckets(), hence the pass-by-reference.
348 *
349 * Note: callers cannot have ht->lock held.
350 */
351 static inline
qht_map_lock_buckets__no_stale(struct qht * ht,struct qht_map ** pmap)352 void qht_map_lock_buckets__no_stale(struct qht *ht, struct qht_map **pmap)
353 {
354 struct qht_map *map;
355
356 map = qatomic_rcu_read(&ht->map);
357 qht_map_lock_buckets(map);
358 if (likely(!qht_map_is_stale__locked(ht, map))) {
359 *pmap = map;
360 return;
361 }
362 qht_map_unlock_buckets(map);
363
364 /* we raced with a resize; acquire ht->lock to see the updated ht->map */
365 qht_lock(ht);
366 map = ht->map;
367 qht_map_lock_buckets(map);
368 qht_unlock(ht);
369 *pmap = map;
370 }
371
372 /*
373 * Get a head bucket and lock it, making sure its parent map is not stale.
374 * @pmap is filled with a pointer to the bucket's parent map.
375 *
376 * Unlock with qht_bucket_unlock.
377 *
378 * Note: callers cannot have ht->lock held.
379 */
380 static inline
qht_bucket_lock__no_stale(struct qht * ht,uint32_t hash,struct qht_map ** pmap)381 struct qht_bucket *qht_bucket_lock__no_stale(struct qht *ht, uint32_t hash,
382 struct qht_map **pmap)
383 {
384 struct qht_bucket *b;
385 struct qht_map *map;
386
387 map = qatomic_rcu_read(&ht->map);
388 b = qht_map_to_bucket(map, hash);
389
390 qht_bucket_lock(map, b);
391 if (likely(!qht_map_is_stale__locked(ht, map))) {
392 *pmap = map;
393 return b;
394 }
395 qht_bucket_unlock(map, b);
396
397 /* we raced with a resize; acquire ht->lock to see the updated ht->map */
398 qht_lock(ht);
399 map = ht->map;
400 b = qht_map_to_bucket(map, hash);
401 qht_bucket_lock(map, b);
402 qht_unlock(ht);
403 *pmap = map;
404 return b;
405 }
406
qht_map_needs_resize(const struct qht_map * map)407 static inline bool qht_map_needs_resize(const struct qht_map *map)
408 {
409 return qatomic_read(&map->n_added_buckets) >
410 map->n_added_buckets_threshold;
411 }
412
qht_chain_destroy(struct qht_map * map,struct qht_bucket * head)413 static inline void qht_chain_destroy(struct qht_map *map,
414 struct qht_bucket *head)
415 {
416 struct qht_bucket *curr = head->next;
417 struct qht_bucket *prev;
418
419 qht_do_if_first_in_stripe(map, head, qemu_spin_destroy);
420 while (curr) {
421 prev = curr;
422 curr = curr->next;
423 qemu_vfree(prev);
424 }
425 }
426
427 /* pass only an orphan map */
qht_map_destroy(struct qht_map * map)428 static void qht_map_destroy(struct qht_map *map)
429 {
430 size_t i;
431
432 for (i = 0; i < map->n_buckets; i++) {
433 qht_chain_destroy(map, &map->buckets[i]);
434 }
435 qemu_vfree(map->buckets);
436 g_free(map);
437 }
438
qht_map_create(size_t n_buckets)439 static struct qht_map *qht_map_create(size_t n_buckets)
440 {
441 struct qht_map *map;
442 size_t i;
443
444 map = g_malloc(sizeof(*map));
445 map->n_buckets = n_buckets;
446
447 map->n_added_buckets = 0;
448 map->n_added_buckets_threshold = n_buckets /
449 QHT_NR_ADDED_BUCKETS_THRESHOLD_DIV;
450
451 /* let tiny hash tables to at least add one non-head bucket */
452 if (unlikely(map->n_added_buckets_threshold == 0)) {
453 map->n_added_buckets_threshold = 1;
454 }
455
456 map->buckets = qemu_memalign(QHT_BUCKET_ALIGN,
457 sizeof(*map->buckets) * n_buckets);
458 for (i = 0; i < n_buckets; i++) {
459 qht_head_init(map, &map->buckets[i]);
460 }
461 return map;
462 }
463
qht_init(struct qht * ht,qht_cmp_func_t cmp,size_t n_elems,unsigned int mode)464 void qht_init(struct qht *ht, qht_cmp_func_t cmp, size_t n_elems,
465 unsigned int mode)
466 {
467 struct qht_map *map;
468 size_t n_buckets = qht_elems_to_buckets(n_elems);
469
470 g_assert(cmp);
471 ht->cmp = cmp;
472 ht->mode = mode;
473 qemu_mutex_init(&ht->lock);
474 map = qht_map_create(n_buckets);
475 qatomic_rcu_set(&ht->map, map);
476 }
477
478 /* call only when there are no readers/writers left */
qht_destroy(struct qht * ht)479 void qht_destroy(struct qht *ht)
480 {
481 qht_map_destroy(ht->map);
482 memset(ht, 0, sizeof(*ht));
483 }
484
qht_bucket_reset__locked(struct qht_bucket * head)485 static void qht_bucket_reset__locked(struct qht_bucket *head)
486 {
487 struct qht_bucket *b = head;
488 int i;
489
490 seqlock_write_begin(&head->sequence);
491 do {
492 for (i = 0; i < QHT_BUCKET_ENTRIES; i++) {
493 if (b->pointers[i] == NULL) {
494 goto done;
495 }
496 qatomic_set(&b->hashes[i], 0);
497 qatomic_set(&b->pointers[i], NULL);
498 }
499 b = b->next;
500 } while (b);
501 done:
502 seqlock_write_end(&head->sequence);
503 }
504
505 /* call with all bucket locks held */
qht_map_reset__all_locked(struct qht_map * map)506 static void qht_map_reset__all_locked(struct qht_map *map)
507 {
508 size_t i;
509
510 for (i = 0; i < map->n_buckets; i++) {
511 qht_bucket_reset__locked(&map->buckets[i]);
512 }
513 qht_map_debug__all_locked(map);
514 }
515
qht_reset(struct qht * ht)516 void qht_reset(struct qht *ht)
517 {
518 struct qht_map *map;
519
520 qht_map_lock_buckets__no_stale(ht, &map);
521 qht_map_reset__all_locked(map);
522 qht_map_unlock_buckets(map);
523 }
524
qht_do_resize(struct qht * ht,struct qht_map * new)525 static inline void qht_do_resize(struct qht *ht, struct qht_map *new)
526 {
527 qht_do_resize_reset(ht, new, false);
528 }
529
qht_do_resize_and_reset(struct qht * ht,struct qht_map * new)530 static inline void qht_do_resize_and_reset(struct qht *ht, struct qht_map *new)
531 {
532 qht_do_resize_reset(ht, new, true);
533 }
534
qht_reset_size(struct qht * ht,size_t n_elems)535 bool qht_reset_size(struct qht *ht, size_t n_elems)
536 {
537 struct qht_map *new = NULL;
538 struct qht_map *map;
539 size_t n_buckets;
540
541 n_buckets = qht_elems_to_buckets(n_elems);
542
543 qht_lock(ht);
544 map = ht->map;
545 if (n_buckets != map->n_buckets) {
546 new = qht_map_create(n_buckets);
547 }
548 qht_do_resize_and_reset(ht, new);
549 qht_unlock(ht);
550
551 return !!new;
552 }
553
554 static inline
qht_do_lookup(const struct qht_bucket * head,qht_lookup_func_t func,const void * userp,uint32_t hash)555 void *qht_do_lookup(const struct qht_bucket *head, qht_lookup_func_t func,
556 const void *userp, uint32_t hash)
557 {
558 const struct qht_bucket *b = head;
559 int i;
560
561 do {
562 for (i = 0; i < QHT_BUCKET_ENTRIES; i++) {
563 if (qatomic_read(&b->hashes[i]) == hash) {
564 /* The pointer is dereferenced before seqlock_read_retry,
565 * so (unlike qht_insert__locked) we need to use
566 * qatomic_rcu_read here.
567 */
568 void *p = qatomic_rcu_read(&b->pointers[i]);
569
570 if (likely(p) && likely(func(p, userp))) {
571 return p;
572 }
573 }
574 }
575 b = qatomic_rcu_read(&b->next);
576 } while (b);
577
578 return NULL;
579 }
580
581 static __attribute__((noinline))
qht_lookup__slowpath(const struct qht_bucket * b,qht_lookup_func_t func,const void * userp,uint32_t hash)582 void *qht_lookup__slowpath(const struct qht_bucket *b, qht_lookup_func_t func,
583 const void *userp, uint32_t hash)
584 {
585 unsigned int version;
586 void *ret;
587
588 do {
589 version = seqlock_read_begin(&b->sequence);
590 ret = qht_do_lookup(b, func, userp, hash);
591 } while (seqlock_read_retry(&b->sequence, version));
592 return ret;
593 }
594
qht_lookup_custom(const struct qht * ht,const void * userp,uint32_t hash,qht_lookup_func_t func)595 void *qht_lookup_custom(const struct qht *ht, const void *userp, uint32_t hash,
596 qht_lookup_func_t func)
597 {
598 const struct qht_bucket *b;
599 const struct qht_map *map;
600 unsigned int version;
601 void *ret;
602
603 map = qatomic_rcu_read(&ht->map);
604 b = qht_map_to_bucket(map, hash);
605
606 version = seqlock_read_begin(&b->sequence);
607 ret = qht_do_lookup(b, func, userp, hash);
608 if (likely(!seqlock_read_retry(&b->sequence, version))) {
609 return ret;
610 }
611 /*
612 * Removing the do/while from the fastpath gives a 4% perf. increase when
613 * running a 100%-lookup microbenchmark.
614 */
615 return qht_lookup__slowpath(b, func, userp, hash);
616 }
617
qht_lookup(const struct qht * ht,const void * userp,uint32_t hash)618 void *qht_lookup(const struct qht *ht, const void *userp, uint32_t hash)
619 {
620 return qht_lookup_custom(ht, userp, hash, ht->cmp);
621 }
622
623 /*
624 * call with head->lock held
625 * @ht is const since it is only used for ht->cmp()
626 */
qht_insert__locked(const struct qht * ht,struct qht_map * map,struct qht_bucket * head,void * p,uint32_t hash,bool * needs_resize)627 static void *qht_insert__locked(const struct qht *ht, struct qht_map *map,
628 struct qht_bucket *head, void *p, uint32_t hash,
629 bool *needs_resize)
630 {
631 struct qht_bucket *b = head;
632 struct qht_bucket *prev = NULL;
633 struct qht_bucket *new = NULL;
634 int i;
635
636 do {
637 for (i = 0; i < QHT_BUCKET_ENTRIES; i++) {
638 if (b->pointers[i]) {
639 if (unlikely(b->hashes[i] == hash &&
640 ht->cmp(b->pointers[i], p))) {
641 return b->pointers[i];
642 }
643 } else {
644 goto found;
645 }
646 }
647 prev = b;
648 b = b->next;
649 } while (b);
650
651 b = qemu_memalign(QHT_BUCKET_ALIGN, sizeof(*b));
652 memset(b, 0, sizeof(*b));
653 new = b;
654 i = 0;
655 qatomic_inc(&map->n_added_buckets);
656 if (unlikely(qht_map_needs_resize(map)) && needs_resize) {
657 *needs_resize = true;
658 }
659
660 found:
661 /* found an empty key: acquire the seqlock and write */
662 seqlock_write_begin(&head->sequence);
663 if (new) {
664 qatomic_rcu_set(&prev->next, b);
665 }
666 /* smp_wmb() implicit in seqlock_write_begin. */
667 qatomic_set(&b->hashes[i], hash);
668 qatomic_set(&b->pointers[i], p);
669 seqlock_write_end(&head->sequence);
670 return NULL;
671 }
672
qht_grow_maybe(struct qht * ht)673 static __attribute__((noinline)) void qht_grow_maybe(struct qht *ht)
674 {
675 struct qht_map *map;
676
677 /*
678 * If the lock is taken it probably means there's an ongoing resize,
679 * so bail out.
680 */
681 if (qht_trylock(ht)) {
682 return;
683 }
684 map = ht->map;
685 /* another thread might have just performed the resize we were after */
686 if (qht_map_needs_resize(map)) {
687 struct qht_map *new = qht_map_create(map->n_buckets * 2);
688
689 qht_do_resize(ht, new);
690 }
691 qht_unlock(ht);
692 }
693
qht_insert(struct qht * ht,void * p,uint32_t hash,void ** existing)694 bool qht_insert(struct qht *ht, void *p, uint32_t hash, void **existing)
695 {
696 struct qht_bucket *b;
697 struct qht_map *map;
698 bool needs_resize = false;
699 void *prev;
700
701 /* NULL pointers are not supported */
702 qht_debug_assert(p);
703
704 b = qht_bucket_lock__no_stale(ht, hash, &map);
705 prev = qht_insert__locked(ht, map, b, p, hash, &needs_resize);
706 qht_bucket_debug__locked(b);
707 qht_bucket_unlock(map, b);
708
709 if (unlikely(needs_resize) && ht->mode & QHT_MODE_AUTO_RESIZE) {
710 qht_grow_maybe(ht);
711 }
712 if (likely(prev == NULL)) {
713 return true;
714 }
715 if (existing) {
716 *existing = prev;
717 }
718 return false;
719 }
720
qht_entry_is_last(const struct qht_bucket * b,int pos)721 static inline bool qht_entry_is_last(const struct qht_bucket *b, int pos)
722 {
723 if (pos == QHT_BUCKET_ENTRIES - 1) {
724 if (b->next == NULL) {
725 return true;
726 }
727 return b->next->pointers[0] == NULL;
728 }
729 return b->pointers[pos + 1] == NULL;
730 }
731
732 static void
qht_entry_move(struct qht_bucket * to,int i,struct qht_bucket * from,int j)733 qht_entry_move(struct qht_bucket *to, int i, struct qht_bucket *from, int j)
734 {
735 qht_debug_assert(!(to == from && i == j));
736 qht_debug_assert(to->pointers[i]);
737 qht_debug_assert(from->pointers[j]);
738
739 qatomic_set(&to->hashes[i], from->hashes[j]);
740 qatomic_set(&to->pointers[i], from->pointers[j]);
741
742 qatomic_set(&from->hashes[j], 0);
743 qatomic_set(&from->pointers[j], NULL);
744 }
745
746 /*
747 * Find the last valid entry in @orig, and swap it with @orig[pos], which has
748 * just been invalidated.
749 */
qht_bucket_remove_entry(struct qht_bucket * orig,int pos)750 static inline void qht_bucket_remove_entry(struct qht_bucket *orig, int pos)
751 {
752 struct qht_bucket *b = orig;
753 struct qht_bucket *prev = NULL;
754 int i;
755
756 if (qht_entry_is_last(orig, pos)) {
757 qatomic_set(&orig->hashes[pos], 0);
758 qatomic_set(&orig->pointers[pos], NULL);
759 return;
760 }
761 do {
762 for (i = 0; i < QHT_BUCKET_ENTRIES; i++) {
763 if (b->pointers[i]) {
764 continue;
765 }
766 if (i > 0) {
767 return qht_entry_move(orig, pos, b, i - 1);
768 }
769 qht_debug_assert(prev);
770 return qht_entry_move(orig, pos, prev, QHT_BUCKET_ENTRIES - 1);
771 }
772 prev = b;
773 b = b->next;
774 } while (b);
775 /* no free entries other than orig[pos], so swap it with the last one */
776 qht_entry_move(orig, pos, prev, QHT_BUCKET_ENTRIES - 1);
777 }
778
779 /* call with b->lock held */
780 static inline
qht_remove__locked(struct qht_bucket * head,const void * p,uint32_t hash)781 bool qht_remove__locked(struct qht_bucket *head, const void *p, uint32_t hash)
782 {
783 struct qht_bucket *b = head;
784 int i;
785
786 do {
787 for (i = 0; i < QHT_BUCKET_ENTRIES; i++) {
788 void *q = b->pointers[i];
789
790 if (unlikely(q == NULL)) {
791 return false;
792 }
793 if (q == p) {
794 qht_debug_assert(b->hashes[i] == hash);
795 seqlock_write_begin(&head->sequence);
796 qht_bucket_remove_entry(b, i);
797 seqlock_write_end(&head->sequence);
798 return true;
799 }
800 }
801 b = b->next;
802 } while (b);
803 return false;
804 }
805
qht_remove(struct qht * ht,const void * p,uint32_t hash)806 bool qht_remove(struct qht *ht, const void *p, uint32_t hash)
807 {
808 struct qht_bucket *b;
809 struct qht_map *map;
810 bool ret;
811
812 /* NULL pointers are not supported */
813 qht_debug_assert(p);
814
815 b = qht_bucket_lock__no_stale(ht, hash, &map);
816 ret = qht_remove__locked(b, p, hash);
817 qht_bucket_debug__locked(b);
818 qht_bucket_unlock(map, b);
819 return ret;
820 }
821
qht_bucket_iter(struct qht_bucket * head,const struct qht_iter * iter,void * userp)822 static inline void qht_bucket_iter(struct qht_bucket *head,
823 const struct qht_iter *iter, void *userp)
824 {
825 struct qht_bucket *b = head;
826 int i;
827
828 do {
829 for (i = 0; i < QHT_BUCKET_ENTRIES; i++) {
830 if (b->pointers[i] == NULL) {
831 return;
832 }
833 switch (iter->type) {
834 case QHT_ITER_VOID:
835 iter->f.retvoid(b->pointers[i], b->hashes[i], userp);
836 break;
837 case QHT_ITER_RM:
838 if (iter->f.retbool(b->pointers[i], b->hashes[i], userp)) {
839 /* replace i with the last valid element in the bucket */
840 seqlock_write_begin(&head->sequence);
841 qht_bucket_remove_entry(b, i);
842 seqlock_write_end(&head->sequence);
843 qht_bucket_debug__locked(b);
844 /* reevaluate i, since it just got replaced */
845 i--;
846 continue;
847 }
848 break;
849 default:
850 g_assert_not_reached();
851 }
852 }
853 b = b->next;
854 } while (b);
855 }
856
857 /* call with all of the map's locks held */
qht_map_iter__all_locked(struct qht_map * map,const struct qht_iter * iter,void * userp)858 static inline void qht_map_iter__all_locked(struct qht_map *map,
859 const struct qht_iter *iter,
860 void *userp)
861 {
862 size_t i;
863
864 for (i = 0; i < map->n_buckets; i++) {
865 qht_bucket_iter(&map->buckets[i], iter, userp);
866 }
867 }
868
869 static inline void
do_qht_iter(struct qht * ht,const struct qht_iter * iter,void * userp)870 do_qht_iter(struct qht *ht, const struct qht_iter *iter, void *userp)
871 {
872 struct qht_map *map;
873
874 map = qatomic_rcu_read(&ht->map);
875 qht_map_lock_buckets(map);
876 qht_map_iter__all_locked(map, iter, userp);
877 qht_map_unlock_buckets(map);
878 }
879
qht_iter(struct qht * ht,qht_iter_func_t func,void * userp)880 void qht_iter(struct qht *ht, qht_iter_func_t func, void *userp)
881 {
882 const struct qht_iter iter = {
883 .f.retvoid = func,
884 .type = QHT_ITER_VOID,
885 };
886
887 do_qht_iter(ht, &iter, userp);
888 }
889
qht_iter_remove(struct qht * ht,qht_iter_bool_func_t func,void * userp)890 void qht_iter_remove(struct qht *ht, qht_iter_bool_func_t func, void *userp)
891 {
892 const struct qht_iter iter = {
893 .f.retbool = func,
894 .type = QHT_ITER_RM,
895 };
896
897 do_qht_iter(ht, &iter, userp);
898 }
899
900 struct qht_map_copy_data {
901 struct qht *ht;
902 struct qht_map *new;
903 };
904
qht_map_copy(void * p,uint32_t hash,void * userp)905 static void qht_map_copy(void *p, uint32_t hash, void *userp)
906 {
907 struct qht_map_copy_data *data = userp;
908 struct qht *ht = data->ht;
909 struct qht_map *new = data->new;
910 struct qht_bucket *b = qht_map_to_bucket(new, hash);
911
912 /* no need to acquire b->lock because no thread has seen this map yet */
913 qht_insert__locked(ht, new, b, p, hash, NULL);
914 }
915
916 /*
917 * Atomically perform a resize and/or reset.
918 * Call with ht->lock held.
919 */
qht_do_resize_reset(struct qht * ht,struct qht_map * new,bool reset)920 static void qht_do_resize_reset(struct qht *ht, struct qht_map *new, bool reset)
921 {
922 struct qht_map *old;
923 const struct qht_iter iter = {
924 .f.retvoid = qht_map_copy,
925 .type = QHT_ITER_VOID,
926 };
927 struct qht_map_copy_data data;
928
929 old = ht->map;
930 qht_map_lock_buckets(old);
931
932 if (reset) {
933 qht_map_reset__all_locked(old);
934 }
935
936 if (new == NULL) {
937 qht_map_unlock_buckets(old);
938 return;
939 }
940
941 g_assert(new->n_buckets != old->n_buckets);
942 data.ht = ht;
943 data.new = new;
944 qht_map_iter__all_locked(old, &iter, &data);
945 qht_map_debug__all_locked(new);
946
947 qatomic_rcu_set(&ht->map, new);
948 qht_map_unlock_buckets(old);
949 call_rcu(old, qht_map_destroy, rcu);
950 }
951
qht_resize(struct qht * ht,size_t n_elems)952 bool qht_resize(struct qht *ht, size_t n_elems)
953 {
954 size_t n_buckets = qht_elems_to_buckets(n_elems);
955 size_t ret = false;
956
957 qht_lock(ht);
958 if (n_buckets != ht->map->n_buckets) {
959 struct qht_map *new;
960
961 new = qht_map_create(n_buckets);
962 qht_do_resize(ht, new);
963 ret = true;
964 }
965 qht_unlock(ht);
966
967 return ret;
968 }
969
970 /* pass @stats to qht_statistics_destroy() when done */
qht_statistics_init(const struct qht * ht,struct qht_stats * stats)971 void qht_statistics_init(const struct qht *ht, struct qht_stats *stats)
972 {
973 const struct qht_map *map;
974 int i;
975
976 map = qatomic_rcu_read(&ht->map);
977
978 stats->used_head_buckets = 0;
979 stats->entries = 0;
980 qdist_init(&stats->chain);
981 qdist_init(&stats->occupancy);
982 /* bail out if the qht has not yet been initialized */
983 if (unlikely(map == NULL)) {
984 stats->head_buckets = 0;
985 return;
986 }
987 stats->head_buckets = map->n_buckets;
988
989 for (i = 0; i < map->n_buckets; i++) {
990 const struct qht_bucket *head = &map->buckets[i];
991 const struct qht_bucket *b;
992 unsigned int version;
993 size_t buckets;
994 size_t entries;
995 int j;
996
997 do {
998 version = seqlock_read_begin(&head->sequence);
999 buckets = 0;
1000 entries = 0;
1001 b = head;
1002 do {
1003 for (j = 0; j < QHT_BUCKET_ENTRIES; j++) {
1004 if (qatomic_read(&b->pointers[j]) == NULL) {
1005 break;
1006 }
1007 entries++;
1008 }
1009 buckets++;
1010 b = qatomic_rcu_read(&b->next);
1011 } while (b);
1012 } while (seqlock_read_retry(&head->sequence, version));
1013
1014 if (entries) {
1015 qdist_inc(&stats->chain, buckets);
1016 qdist_inc(&stats->occupancy,
1017 (double)entries / QHT_BUCKET_ENTRIES / buckets);
1018 stats->used_head_buckets++;
1019 stats->entries += entries;
1020 } else {
1021 qdist_inc(&stats->occupancy, 0);
1022 }
1023 }
1024 }
1025
qht_statistics_destroy(struct qht_stats * stats)1026 void qht_statistics_destroy(struct qht_stats *stats)
1027 {
1028 qdist_destroy(&stats->occupancy);
1029 qdist_destroy(&stats->chain);
1030 }
1031