1 /* SPDX-License-Identifier: LGPL-2.1+ */
3 Copyright © 2014 Michal Schmidt
11 #include "alloc-util.h"
16 #include "process-util.h"
17 #include "random-util.h"
19 #include "siphash24.h"
20 #include "string-util.h"
24 #if ENABLE_DEBUG_HASHMAP
30 * Implementation of hashmaps.
32 * - uses less RAM compared to closed addressing (chaining), because
33 * our entries are small (especially in Sets, which tend to contain
34 * the majority of entries in systemd).
35 * Collision resolution: Robin Hood
36 * - tends to equalize displacement of entries from their optimal buckets.
37 * Probe sequence: linear
38 * - though theoretically worse than random probing/uniform hashing/double
39 * hashing, it is good for cache locality.
42 * Celis, P. 1986. Robin Hood Hashing.
43 * Ph.D. Dissertation. University of Waterloo, Waterloo, Ont., Canada, Canada.
44 * https://cs.uwaterloo.ca/research/tr/1986/CS-86-14.pdf
45 * - The results are derived for random probing. Suggests deletion with
46 * tombstones and two mean-centered search methods. None of that works
47 * well for linear probing.
49 * Janson, S. 2005. Individual displacements for linear probing hashing with different insertion policies.
50 * ACM Trans. Algorithms 1, 2 (October 2005), 177-213.
51 * DOI=10.1145/1103963.1103964 http://doi.acm.org/10.1145/1103963.1103964
52 * http://www.math.uu.se/~svante/papers/sj157.pdf
53 * - Applies to Robin Hood with linear probing. Contains remarks on
54 * the unsuitability of mean-centered search with linear probing.
56 * Viola, A. 2005. Exact distribution of individual displacements in linear probing hashing.
57 * ACM Trans. Algorithms 1, 2 (October 2005), 214-242.
58 * DOI=10.1145/1103963.1103965 http://doi.acm.org/10.1145/1103963.1103965
59 * - Similar to Janson. Note that Viola writes about C_{m,n} (number of probes
60 * in a successful search), and Janson writes about displacement. C = d + 1.
62 * Goossaert, E. 2013. Robin Hood hashing: backward shift deletion.
63 * http://codecapsule.com/2013/11/17/robin-hood-hashing-backward-shift-deletion/
64 * - Explanation of backward shift deletion with pictures.
66 * Khuong, P. 2013. The Other Robin Hood Hashing.
67 * http://www.pvk.ca/Blog/2013/11/26/the-other-robin-hood-hashing/
68 * - Short summary of random vs. linear probing, and tombstones vs. backward shift.
72 * XXX Ideas for improvement:
73 * For unordered hashmaps, randomize iteration order, similarly to Perl:
74 * http://blog.booking.com/hardening-perls-hash-function.html
77 /* INV_KEEP_FREE = 1 / (1 - max_load_factor)
78 * e.g. 1 / (1 - 0.8) = 5 ... keep one fifth of the buckets free. */
79 #define INV_KEEP_FREE 5U
81 /* Fields common to entries of all hashmap/set types */
82 struct hashmap_base_entry
{
86 /* Entry types for specific hashmap/set types
87 * hashmap_base_entry must be at the beginning of each entry struct. */
89 struct plain_hashmap_entry
{
90 struct hashmap_base_entry b
;
94 struct ordered_hashmap_entry
{
95 struct plain_hashmap_entry p
;
96 unsigned iterate_next
, iterate_previous
;
100 struct hashmap_base_entry b
;
103 /* In several functions it is advantageous to have the hash table extended
104 * virtually by a couple of additional buckets. We reserve special index values
105 * for these "swap" buckets. */
106 #define _IDX_SWAP_BEGIN (UINT_MAX - 3)
107 #define IDX_PUT (_IDX_SWAP_BEGIN + 0)
108 #define IDX_TMP (_IDX_SWAP_BEGIN + 1)
109 #define _IDX_SWAP_END (_IDX_SWAP_BEGIN + 2)
111 #define IDX_FIRST (UINT_MAX - 1) /* special index for freshly initialized iterators */
112 #define IDX_NIL UINT_MAX /* special index value meaning "none" or "end" */
114 assert_cc(IDX_FIRST
== _IDX_SWAP_END
);
115 assert_cc(IDX_FIRST
== _IDX_ITERATOR_FIRST
);
117 /* Storage space for the "swap" buckets.
118 * All entry types can fit into a ordered_hashmap_entry. */
119 struct swap_entries
{
120 struct ordered_hashmap_entry e
[_IDX_SWAP_END
- _IDX_SWAP_BEGIN
];
123 /* Distance from Initial Bucket */
124 typedef uint8_t dib_raw_t
;
125 #define DIB_RAW_OVERFLOW ((dib_raw_t)0xfdU) /* indicates DIB value is greater than representable */
126 #define DIB_RAW_REHASH ((dib_raw_t)0xfeU) /* entry yet to be rehashed during in-place resize */
127 #define DIB_RAW_FREE ((dib_raw_t)0xffU) /* a free bucket */
128 #define DIB_RAW_INIT ((char)DIB_RAW_FREE) /* a byte to memset a DIB store with when initializing */
130 #define DIB_FREE UINT_MAX
132 #if ENABLE_DEBUG_HASHMAP
133 struct hashmap_debug_info
{
134 LIST_FIELDS(struct hashmap_debug_info
, debug_list
);
135 unsigned max_entries
; /* high watermark of n_entries */
137 /* who allocated this hashmap */
142 /* fields to detect modification while iterating */
143 unsigned put_count
; /* counts puts into the hashmap */
144 unsigned rem_count
; /* counts removals from hashmap */
145 unsigned last_rem_idx
; /* remembers last removal index */
148 /* Tracks all existing hashmaps. Get at it from gdb. See sd_dump_hashmaps.py */
149 static LIST_HEAD(struct hashmap_debug_info
, hashmap_debug_list
);
150 static pthread_mutex_t hashmap_debug_list_mutex
= PTHREAD_MUTEX_INITIALIZER
;
152 #define HASHMAP_DEBUG_FIELDS struct hashmap_debug_info debug;
154 #else /* !ENABLE_DEBUG_HASHMAP */
155 #define HASHMAP_DEBUG_FIELDS
156 #endif /* ENABLE_DEBUG_HASHMAP */
160 HASHMAP_TYPE_ORDERED
,
165 struct _packed_ indirect_storage
{
166 void *storage
; /* where buckets and DIBs are stored */
167 uint8_t hash_key
[HASH_KEY_SIZE
]; /* hash key; changes during resize */
169 unsigned n_entries
; /* number of stored entries */
170 unsigned n_buckets
; /* number of buckets */
172 unsigned idx_lowest_entry
; /* Index below which all buckets are free.
173 Makes "while(hashmap_steal_first())" loops
174 O(n) instead of O(n^2) for unordered hashmaps. */
175 uint8_t _pad
[3]; /* padding for the whole HashmapBase */
176 /* The bitfields in HashmapBase complete the alignment of the whole thing. */
179 struct direct_storage
{
180 /* This gives us 39 bytes on 64bit, or 35 bytes on 32bit.
181 * That's room for 4 set_entries + 4 DIB bytes + 3 unused bytes on 64bit,
182 * or 7 set_entries + 7 DIB bytes + 0 unused bytes on 32bit. */
183 uint8_t storage
[sizeof(struct indirect_storage
)];
186 #define DIRECT_BUCKETS(entry_t) \
187 (sizeof(struct direct_storage) / (sizeof(entry_t) + sizeof(dib_raw_t)))
189 /* We should be able to store at least one entry directly. */
190 assert_cc(DIRECT_BUCKETS(struct ordered_hashmap_entry
) >= 1);
192 /* We have 3 bits for n_direct_entries. */
193 assert_cc(DIRECT_BUCKETS(struct set_entry
) < (1 << 3));
195 /* Hashmaps with directly stored entries all use this shared hash key.
196 * It's no big deal if the key is guessed, because there can be only
197 * a handful of directly stored entries in a hashmap. When a hashmap
198 * outgrows direct storage, it gets its own key for indirect storage. */
199 static uint8_t shared_hash_key
[HASH_KEY_SIZE
];
200 static bool shared_hash_key_initialized
;
202 /* Fields that all hashmap/set types must have */
204 const struct hash_ops
*hash_ops
; /* hash and compare ops to use */
207 struct indirect_storage indirect
; /* if has_indirect */
208 struct direct_storage direct
; /* if !has_indirect */
211 enum HashmapType type
:2; /* HASHMAP_TYPE_* */
212 bool has_indirect
:1; /* whether indirect storage is used */
213 unsigned n_direct_entries
:3; /* Number of entries in direct storage.
214 * Only valid if !has_indirect. */
215 bool from_pool
:1; /* whether was allocated from mempool */
216 bool dirty
:1; /* whether dirtied since last iterated_cache_get() */
217 bool cached
:1; /* whether this hashmap is being cached */
218 HASHMAP_DEBUG_FIELDS
/* optional hashmap_debug_info */
221 /* Specific hash types
222 * HashmapBase must be at the beginning of each hashmap struct. */
225 struct HashmapBase b
;
228 struct OrderedHashmap
{
229 struct HashmapBase b
;
230 unsigned iterate_list_head
, iterate_list_tail
;
234 struct HashmapBase b
;
237 typedef struct CacheMem
{
239 size_t n_populated
, n_allocated
;
243 struct IteratedCache
{
244 HashmapBase
*hashmap
;
245 CacheMem keys
, values
;
248 DEFINE_MEMPOOL(hashmap_pool
, Hashmap
, 8);
249 DEFINE_MEMPOOL(ordered_hashmap_pool
, OrderedHashmap
, 8);
250 /* No need for a separate Set pool */
251 assert_cc(sizeof(Hashmap
) == sizeof(Set
));
253 struct hashmap_type_info
{
256 struct mempool
*mempool
;
257 unsigned n_direct_buckets
;
260 static const struct hashmap_type_info hashmap_type_info
[_HASHMAP_TYPE_MAX
] = {
261 [HASHMAP_TYPE_PLAIN
] = {
262 .head_size
= sizeof(Hashmap
),
263 .entry_size
= sizeof(struct plain_hashmap_entry
),
264 .mempool
= &hashmap_pool
,
265 .n_direct_buckets
= DIRECT_BUCKETS(struct plain_hashmap_entry
),
267 [HASHMAP_TYPE_ORDERED
] = {
268 .head_size
= sizeof(OrderedHashmap
),
269 .entry_size
= sizeof(struct ordered_hashmap_entry
),
270 .mempool
= &ordered_hashmap_pool
,
271 .n_direct_buckets
= DIRECT_BUCKETS(struct ordered_hashmap_entry
),
273 [HASHMAP_TYPE_SET
] = {
274 .head_size
= sizeof(Set
),
275 .entry_size
= sizeof(struct set_entry
),
276 .mempool
= &hashmap_pool
,
277 .n_direct_buckets
= DIRECT_BUCKETS(struct set_entry
),
282 __attribute__((destructor
)) static void cleanup_pools(void) {
283 _cleanup_free_
char *t
= NULL
;
286 /* Be nice to valgrind */
288 /* The pool is only allocated by the main thread, but the memory can
289 * be passed to other threads. Let's clean up if we are the main thread
290 * and no other threads are live. */
291 if (!is_main_thread())
294 r
= get_proc_field("/proc/self/status", "Threads", WHITESPACE
, &t
);
295 if (r
< 0 || !streq(t
, "1"))
298 mempool_drop(&hashmap_pool
);
299 mempool_drop(&ordered_hashmap_pool
);
303 static unsigned n_buckets(HashmapBase
*h
) {
304 return h
->has_indirect
? h
->indirect
.n_buckets
305 : hashmap_type_info
[h
->type
].n_direct_buckets
;
308 static unsigned n_entries(HashmapBase
*h
) {
309 return h
->has_indirect
? h
->indirect
.n_entries
310 : h
->n_direct_entries
;
313 static void n_entries_inc(HashmapBase
*h
) {
315 h
->indirect
.n_entries
++;
317 h
->n_direct_entries
++;
320 static void n_entries_dec(HashmapBase
*h
) {
322 h
->indirect
.n_entries
--;
324 h
->n_direct_entries
--;
327 static void *storage_ptr(HashmapBase
*h
) {
328 return h
->has_indirect
? h
->indirect
.storage
332 static uint8_t *hash_key(HashmapBase
*h
) {
333 return h
->has_indirect
? h
->indirect
.hash_key
337 static unsigned base_bucket_hash(HashmapBase
*h
, const void *p
) {
338 struct siphash state
;
341 siphash24_init(&state
, hash_key(h
));
343 h
->hash_ops
->hash(p
, &state
);
345 hash
= siphash24_finalize(&state
);
347 return (unsigned) (hash
% n_buckets(h
));
349 #define bucket_hash(h, p) base_bucket_hash(HASHMAP_BASE(h), p)
351 static inline void base_set_dirty(HashmapBase
*h
) {
354 #define hashmap_set_dirty(h) base_set_dirty(HASHMAP_BASE(h))
356 static void get_hash_key(uint8_t hash_key
[HASH_KEY_SIZE
], bool reuse_is_ok
) {
357 static uint8_t current
[HASH_KEY_SIZE
];
358 static bool current_initialized
= false;
360 /* Returns a hash function key to use. In order to keep things
361 * fast we will not generate a new key each time we allocate a
362 * new hash table. Instead, we'll just reuse the most recently
363 * generated one, except if we never generated one or when we
364 * are rehashing an entire hash table because we reached a
367 if (!current_initialized
|| !reuse_is_ok
) {
368 random_bytes(current
, sizeof(current
));
369 current_initialized
= true;
372 memcpy(hash_key
, current
, sizeof(current
));
375 static struct hashmap_base_entry
*bucket_at(HashmapBase
*h
, unsigned idx
) {
376 return (struct hashmap_base_entry
*)
377 ((uint8_t*) storage_ptr(h
) + idx
* hashmap_type_info
[h
->type
].entry_size
);
380 static struct plain_hashmap_entry
*plain_bucket_at(Hashmap
*h
, unsigned idx
) {
381 return (struct plain_hashmap_entry
*) bucket_at(HASHMAP_BASE(h
), idx
);
384 static struct ordered_hashmap_entry
*ordered_bucket_at(OrderedHashmap
*h
, unsigned idx
) {
385 return (struct ordered_hashmap_entry
*) bucket_at(HASHMAP_BASE(h
), idx
);
388 static struct set_entry
*set_bucket_at(Set
*h
, unsigned idx
) {
389 return (struct set_entry
*) bucket_at(HASHMAP_BASE(h
), idx
);
392 static struct ordered_hashmap_entry
*bucket_at_swap(struct swap_entries
*swap
, unsigned idx
) {
393 return &swap
->e
[idx
- _IDX_SWAP_BEGIN
];
396 /* Returns a pointer to the bucket at index idx.
397 * Understands real indexes and swap indexes, hence "_virtual". */
398 static struct hashmap_base_entry
*bucket_at_virtual(HashmapBase
*h
, struct swap_entries
*swap
,
400 if (idx
< _IDX_SWAP_BEGIN
)
401 return bucket_at(h
, idx
);
403 if (idx
< _IDX_SWAP_END
)
404 return &bucket_at_swap(swap
, idx
)->p
.b
;
406 assert_not_reached("Invalid index");
409 static dib_raw_t
*dib_raw_ptr(HashmapBase
*h
) {
411 ((uint8_t*) storage_ptr(h
) + hashmap_type_info
[h
->type
].entry_size
* n_buckets(h
));
414 static unsigned bucket_distance(HashmapBase
*h
, unsigned idx
, unsigned from
) {
415 return idx
>= from
? idx
- from
416 : n_buckets(h
) + idx
- from
;
419 static unsigned bucket_calculate_dib(HashmapBase
*h
, unsigned idx
, dib_raw_t raw_dib
) {
420 unsigned initial_bucket
;
422 if (raw_dib
== DIB_RAW_FREE
)
425 if (_likely_(raw_dib
< DIB_RAW_OVERFLOW
))
429 * Having an overflow DIB value is very unlikely. The hash function
430 * would have to be bad. For example, in a table of size 2^24 filled
431 * to load factor 0.9 the maximum observed DIB is only about 60.
432 * In theory (assuming I used Maxima correctly), for an infinite size
433 * hash table with load factor 0.8 the probability of a given entry
434 * having DIB > 40 is 1.9e-8.
435 * This returns the correct DIB value by recomputing the hash value in
436 * the unlikely case. XXX Hitting this case could be a hint to rehash.
438 initial_bucket
= bucket_hash(h
, bucket_at(h
, idx
)->key
);
439 return bucket_distance(h
, idx
, initial_bucket
);
442 static void bucket_set_dib(HashmapBase
*h
, unsigned idx
, unsigned dib
) {
443 dib_raw_ptr(h
)[idx
] = dib
!= DIB_FREE
? MIN(dib
, DIB_RAW_OVERFLOW
) : DIB_RAW_FREE
;
446 static unsigned skip_free_buckets(HashmapBase
*h
, unsigned idx
) {
449 dibs
= dib_raw_ptr(h
);
451 for ( ; idx
< n_buckets(h
); idx
++)
452 if (dibs
[idx
] != DIB_RAW_FREE
)
458 static void bucket_mark_free(HashmapBase
*h
, unsigned idx
) {
459 memzero(bucket_at(h
, idx
), hashmap_type_info
[h
->type
].entry_size
);
460 bucket_set_dib(h
, idx
, DIB_FREE
);
463 static void bucket_move_entry(HashmapBase
*h
, struct swap_entries
*swap
,
464 unsigned from
, unsigned to
) {
465 struct hashmap_base_entry
*e_from
, *e_to
;
469 e_from
= bucket_at_virtual(h
, swap
, from
);
470 e_to
= bucket_at_virtual(h
, swap
, to
);
472 memcpy(e_to
, e_from
, hashmap_type_info
[h
->type
].entry_size
);
474 if (h
->type
== HASHMAP_TYPE_ORDERED
) {
475 OrderedHashmap
*lh
= (OrderedHashmap
*) h
;
476 struct ordered_hashmap_entry
*le
, *le_to
;
478 le_to
= (struct ordered_hashmap_entry
*) e_to
;
480 if (le_to
->iterate_next
!= IDX_NIL
) {
481 le
= (struct ordered_hashmap_entry
*)
482 bucket_at_virtual(h
, swap
, le_to
->iterate_next
);
483 le
->iterate_previous
= to
;
486 if (le_to
->iterate_previous
!= IDX_NIL
) {
487 le
= (struct ordered_hashmap_entry
*)
488 bucket_at_virtual(h
, swap
, le_to
->iterate_previous
);
489 le
->iterate_next
= to
;
492 if (lh
->iterate_list_head
== from
)
493 lh
->iterate_list_head
= to
;
494 if (lh
->iterate_list_tail
== from
)
495 lh
->iterate_list_tail
= to
;
499 static unsigned next_idx(HashmapBase
*h
, unsigned idx
) {
500 return (idx
+ 1U) % n_buckets(h
);
503 static unsigned prev_idx(HashmapBase
*h
, unsigned idx
) {
504 return (n_buckets(h
) + idx
- 1U) % n_buckets(h
);
507 static void *entry_value(HashmapBase
*h
, struct hashmap_base_entry
*e
) {
510 case HASHMAP_TYPE_PLAIN
:
511 case HASHMAP_TYPE_ORDERED
:
512 return ((struct plain_hashmap_entry
*)e
)->value
;
514 case HASHMAP_TYPE_SET
:
515 return (void*) e
->key
;
518 assert_not_reached("Unknown hashmap type");
522 static void base_remove_entry(HashmapBase
*h
, unsigned idx
) {
523 unsigned left
, right
, prev
, dib
;
524 dib_raw_t raw_dib
, *dibs
;
526 dibs
= dib_raw_ptr(h
);
527 assert(dibs
[idx
] != DIB_RAW_FREE
);
529 #if ENABLE_DEBUG_HASHMAP
530 h
->debug
.rem_count
++;
531 h
->debug
.last_rem_idx
= idx
;
535 /* Find the stop bucket ("right"). It is either free or has DIB == 0. */
536 for (right
= next_idx(h
, left
); ; right
= next_idx(h
, right
)) {
537 raw_dib
= dibs
[right
];
538 if (IN_SET(raw_dib
, 0, DIB_RAW_FREE
))
541 /* The buckets are not supposed to be all occupied and with DIB > 0.
542 * That would mean we could make everyone better off by shifting them
543 * backward. This scenario is impossible. */
544 assert(left
!= right
);
547 if (h
->type
== HASHMAP_TYPE_ORDERED
) {
548 OrderedHashmap
*lh
= (OrderedHashmap
*) h
;
549 struct ordered_hashmap_entry
*le
= ordered_bucket_at(lh
, idx
);
551 if (le
->iterate_next
!= IDX_NIL
)
552 ordered_bucket_at(lh
, le
->iterate_next
)->iterate_previous
= le
->iterate_previous
;
554 lh
->iterate_list_tail
= le
->iterate_previous
;
556 if (le
->iterate_previous
!= IDX_NIL
)
557 ordered_bucket_at(lh
, le
->iterate_previous
)->iterate_next
= le
->iterate_next
;
559 lh
->iterate_list_head
= le
->iterate_next
;
562 /* Now shift all buckets in the interval (left, right) one step backwards */
563 for (prev
= left
, left
= next_idx(h
, left
); left
!= right
;
564 prev
= left
, left
= next_idx(h
, left
)) {
565 dib
= bucket_calculate_dib(h
, left
, dibs
[left
]);
567 bucket_move_entry(h
, NULL
, left
, prev
);
568 bucket_set_dib(h
, prev
, dib
- 1);
571 bucket_mark_free(h
, prev
);
575 #define remove_entry(h, idx) base_remove_entry(HASHMAP_BASE(h), idx)
577 static unsigned hashmap_iterate_in_insertion_order(OrderedHashmap
*h
, Iterator
*i
) {
578 struct ordered_hashmap_entry
*e
;
584 if (i
->idx
== IDX_NIL
)
587 if (i
->idx
== IDX_FIRST
&& h
->iterate_list_head
== IDX_NIL
)
590 if (i
->idx
== IDX_FIRST
) {
591 idx
= h
->iterate_list_head
;
592 e
= ordered_bucket_at(h
, idx
);
595 e
= ordered_bucket_at(h
, idx
);
597 * We allow removing the current entry while iterating, but removal may cause
598 * a backward shift. The next entry may thus move one bucket to the left.
599 * To detect when it happens, we remember the key pointer of the entry we were
600 * going to iterate next. If it does not match, there was a backward shift.
602 if (e
->p
.b
.key
!= i
->next_key
) {
603 idx
= prev_idx(HASHMAP_BASE(h
), idx
);
604 e
= ordered_bucket_at(h
, idx
);
606 assert(e
->p
.b
.key
== i
->next_key
);
609 #if ENABLE_DEBUG_HASHMAP
613 if (e
->iterate_next
!= IDX_NIL
) {
614 struct ordered_hashmap_entry
*n
;
615 i
->idx
= e
->iterate_next
;
616 n
= ordered_bucket_at(h
, i
->idx
);
617 i
->next_key
= n
->p
.b
.key
;
628 static unsigned hashmap_iterate_in_internal_order(HashmapBase
*h
, Iterator
*i
) {
634 if (i
->idx
== IDX_NIL
)
637 if (i
->idx
== IDX_FIRST
) {
638 /* fast forward to the first occupied bucket */
639 if (h
->has_indirect
) {
640 i
->idx
= skip_free_buckets(h
, h
->indirect
.idx_lowest_entry
);
641 h
->indirect
.idx_lowest_entry
= i
->idx
;
643 i
->idx
= skip_free_buckets(h
, 0);
645 if (i
->idx
== IDX_NIL
)
648 struct hashmap_base_entry
*e
;
652 e
= bucket_at(h
, i
->idx
);
654 * We allow removing the current entry while iterating, but removal may cause
655 * a backward shift. The next entry may thus move one bucket to the left.
656 * To detect when it happens, we remember the key pointer of the entry we were
657 * going to iterate next. If it does not match, there was a backward shift.
659 if (e
->key
!= i
->next_key
)
660 e
= bucket_at(h
, --i
->idx
);
662 assert(e
->key
== i
->next_key
);
666 #if ENABLE_DEBUG_HASHMAP
670 i
->idx
= skip_free_buckets(h
, i
->idx
+ 1);
671 if (i
->idx
!= IDX_NIL
)
672 i
->next_key
= bucket_at(h
, i
->idx
)->key
;
683 static unsigned hashmap_iterate_entry(HashmapBase
*h
, Iterator
*i
) {
689 #if ENABLE_DEBUG_HASHMAP
690 if (i
->idx
== IDX_FIRST
) {
691 i
->put_count
= h
->debug
.put_count
;
692 i
->rem_count
= h
->debug
.rem_count
;
694 /* While iterating, must not add any new entries */
695 assert(i
->put_count
== h
->debug
.put_count
);
696 /* ... or remove entries other than the current one */
697 assert(i
->rem_count
== h
->debug
.rem_count
||
698 (i
->rem_count
== h
->debug
.rem_count
- 1 &&
699 i
->prev_idx
== h
->debug
.last_rem_idx
));
700 /* Reset our removals counter */
701 i
->rem_count
= h
->debug
.rem_count
;
705 return h
->type
== HASHMAP_TYPE_ORDERED
? hashmap_iterate_in_insertion_order((OrderedHashmap
*) h
, i
)
706 : hashmap_iterate_in_internal_order(h
, i
);
709 bool internal_hashmap_iterate(HashmapBase
*h
, Iterator
*i
, void **value
, const void **key
) {
710 struct hashmap_base_entry
*e
;
714 idx
= hashmap_iterate_entry(h
, i
);
715 if (idx
== IDX_NIL
) {
724 e
= bucket_at(h
, idx
);
725 data
= entry_value(h
, e
);
734 bool set_iterate(Set
*s
, Iterator
*i
, void **value
) {
735 return internal_hashmap_iterate(HASHMAP_BASE(s
), i
, value
, NULL
);
738 #define HASHMAP_FOREACH_IDX(idx, h, i) \
739 for ((i) = ITERATOR_FIRST, (idx) = hashmap_iterate_entry((h), &(i)); \
741 (idx) = hashmap_iterate_entry((h), &(i)))
743 IteratedCache
*internal_hashmap_iterated_cache_new(HashmapBase
*h
) {
744 IteratedCache
*cache
;
752 cache
= new0(IteratedCache
, 1);
762 static void reset_direct_storage(HashmapBase
*h
) {
763 const struct hashmap_type_info
*hi
= &hashmap_type_info
[h
->type
];
766 assert(!h
->has_indirect
);
768 p
= mempset(h
->direct
.storage
, 0, hi
->entry_size
* hi
->n_direct_buckets
);
769 memset(p
, DIB_RAW_INIT
, sizeof(dib_raw_t
) * hi
->n_direct_buckets
);
772 static struct HashmapBase
*hashmap_base_new(const struct hash_ops
*hash_ops
, enum HashmapType type HASHMAP_DEBUG_PARAMS
) {
774 const struct hashmap_type_info
*hi
= &hashmap_type_info
[type
];
777 use_pool
= is_main_thread();
779 h
= use_pool
? mempool_alloc0_tile(hi
->mempool
) : malloc0(hi
->head_size
);
785 h
->from_pool
= use_pool
;
786 h
->hash_ops
= hash_ops
? hash_ops
: &trivial_hash_ops
;
788 if (type
== HASHMAP_TYPE_ORDERED
) {
789 OrderedHashmap
*lh
= (OrderedHashmap
*)h
;
790 lh
->iterate_list_head
= lh
->iterate_list_tail
= IDX_NIL
;
793 reset_direct_storage(h
);
795 if (!shared_hash_key_initialized
) {
796 random_bytes(shared_hash_key
, sizeof(shared_hash_key
));
797 shared_hash_key_initialized
= true;
800 #if ENABLE_DEBUG_HASHMAP
801 h
->debug
.func
= func
;
802 h
->debug
.file
= file
;
803 h
->debug
.line
= line
;
804 assert_se(pthread_mutex_lock(&hashmap_debug_list_mutex
) == 0);
805 LIST_PREPEND(debug_list
, hashmap_debug_list
, &h
->debug
);
806 assert_se(pthread_mutex_unlock(&hashmap_debug_list_mutex
) == 0);
812 Hashmap
*internal_hashmap_new(const struct hash_ops
*hash_ops HASHMAP_DEBUG_PARAMS
) {
813 return (Hashmap
*) hashmap_base_new(hash_ops
, HASHMAP_TYPE_PLAIN HASHMAP_DEBUG_PASS_ARGS
);
816 OrderedHashmap
*internal_ordered_hashmap_new(const struct hash_ops
*hash_ops HASHMAP_DEBUG_PARAMS
) {
817 return (OrderedHashmap
*) hashmap_base_new(hash_ops
, HASHMAP_TYPE_ORDERED HASHMAP_DEBUG_PASS_ARGS
);
820 Set
*internal_set_new(const struct hash_ops
*hash_ops HASHMAP_DEBUG_PARAMS
) {
821 return (Set
*) hashmap_base_new(hash_ops
, HASHMAP_TYPE_SET HASHMAP_DEBUG_PASS_ARGS
);
824 static int hashmap_base_ensure_allocated(HashmapBase
**h
, const struct hash_ops
*hash_ops
,
825 enum HashmapType type HASHMAP_DEBUG_PARAMS
) {
833 q
= hashmap_base_new(hash_ops
, type HASHMAP_DEBUG_PASS_ARGS
);
841 int internal_hashmap_ensure_allocated(Hashmap
**h
, const struct hash_ops
*hash_ops HASHMAP_DEBUG_PARAMS
) {
842 return hashmap_base_ensure_allocated((HashmapBase
**)h
, hash_ops
, HASHMAP_TYPE_PLAIN HASHMAP_DEBUG_PASS_ARGS
);
845 int internal_ordered_hashmap_ensure_allocated(OrderedHashmap
**h
, const struct hash_ops
*hash_ops HASHMAP_DEBUG_PARAMS
) {
846 return hashmap_base_ensure_allocated((HashmapBase
**)h
, hash_ops
, HASHMAP_TYPE_ORDERED HASHMAP_DEBUG_PASS_ARGS
);
849 int internal_set_ensure_allocated(Set
**s
, const struct hash_ops
*hash_ops HASHMAP_DEBUG_PARAMS
) {
850 return hashmap_base_ensure_allocated((HashmapBase
**)s
, hash_ops
, HASHMAP_TYPE_SET HASHMAP_DEBUG_PASS_ARGS
);
853 static void hashmap_free_no_clear(HashmapBase
*h
) {
854 assert(!h
->has_indirect
);
855 assert(!h
->n_direct_entries
);
857 #if ENABLE_DEBUG_HASHMAP
858 assert_se(pthread_mutex_lock(&hashmap_debug_list_mutex
) == 0);
859 LIST_REMOVE(debug_list
, hashmap_debug_list
, &h
->debug
);
860 assert_se(pthread_mutex_unlock(&hashmap_debug_list_mutex
) == 0);
864 mempool_free_tile(hashmap_type_info
[h
->type
].mempool
, h
);
869 HashmapBase
*internal_hashmap_free(HashmapBase
*h
) {
871 /* Free the hashmap, but nothing in it */
874 internal_hashmap_clear(h
);
875 hashmap_free_no_clear(h
);
881 HashmapBase
*internal_hashmap_free_free(HashmapBase
*h
) {
883 /* Free the hashmap and all data objects in it, but not the
887 internal_hashmap_clear_free(h
);
888 hashmap_free_no_clear(h
);
894 Hashmap
*hashmap_free_free_free(Hashmap
*h
) {
896 /* Free the hashmap and all data and key objects in it */
899 hashmap_clear_free_free(h
);
900 hashmap_free_no_clear(HASHMAP_BASE(h
));
906 void internal_hashmap_clear(HashmapBase
*h
) {
910 if (h
->has_indirect
) {
911 free(h
->indirect
.storage
);
912 h
->has_indirect
= false;
915 h
->n_direct_entries
= 0;
916 reset_direct_storage(h
);
918 if (h
->type
== HASHMAP_TYPE_ORDERED
) {
919 OrderedHashmap
*lh
= (OrderedHashmap
*) h
;
920 lh
->iterate_list_head
= lh
->iterate_list_tail
= IDX_NIL
;
926 void internal_hashmap_clear_free(HashmapBase
*h
) {
932 for (idx
= skip_free_buckets(h
, 0); idx
!= IDX_NIL
;
933 idx
= skip_free_buckets(h
, idx
+ 1))
934 free(entry_value(h
, bucket_at(h
, idx
)));
936 internal_hashmap_clear(h
);
939 void hashmap_clear_free_free(Hashmap
*h
) {
945 for (idx
= skip_free_buckets(HASHMAP_BASE(h
), 0); idx
!= IDX_NIL
;
946 idx
= skip_free_buckets(HASHMAP_BASE(h
), idx
+ 1)) {
947 struct plain_hashmap_entry
*e
= plain_bucket_at(h
, idx
);
948 free((void*)e
->b
.key
);
952 internal_hashmap_clear(HASHMAP_BASE(h
));
955 static int resize_buckets(HashmapBase
*h
, unsigned entries_add
);
958 * Finds an empty bucket to put an entry into, starting the scan at 'idx'.
959 * Performs Robin Hood swaps as it goes. The entry to put must be placed
960 * by the caller into swap slot IDX_PUT.
961 * If used for in-place resizing, may leave a displaced entry in swap slot
962 * IDX_PUT. Caller must rehash it next.
963 * Returns: true if it left a displaced entry to rehash next in IDX_PUT,
966 static bool hashmap_put_robin_hood(HashmapBase
*h
, unsigned idx
,
967 struct swap_entries
*swap
) {
968 dib_raw_t raw_dib
, *dibs
;
969 unsigned dib
, distance
;
971 #if ENABLE_DEBUG_HASHMAP
972 h
->debug
.put_count
++;
975 dibs
= dib_raw_ptr(h
);
977 for (distance
= 0; ; distance
++) {
979 if (IN_SET(raw_dib
, DIB_RAW_FREE
, DIB_RAW_REHASH
)) {
980 if (raw_dib
== DIB_RAW_REHASH
)
981 bucket_move_entry(h
, swap
, idx
, IDX_TMP
);
983 if (h
->has_indirect
&& h
->indirect
.idx_lowest_entry
> idx
)
984 h
->indirect
.idx_lowest_entry
= idx
;
986 bucket_set_dib(h
, idx
, distance
);
987 bucket_move_entry(h
, swap
, IDX_PUT
, idx
);
988 if (raw_dib
== DIB_RAW_REHASH
) {
989 bucket_move_entry(h
, swap
, IDX_TMP
, IDX_PUT
);
996 dib
= bucket_calculate_dib(h
, idx
, raw_dib
);
998 if (dib
< distance
) {
999 /* Found a wealthier entry. Go Robin Hood! */
1000 bucket_set_dib(h
, idx
, distance
);
1002 /* swap the entries */
1003 bucket_move_entry(h
, swap
, idx
, IDX_TMP
);
1004 bucket_move_entry(h
, swap
, IDX_PUT
, idx
);
1005 bucket_move_entry(h
, swap
, IDX_TMP
, IDX_PUT
);
1010 idx
= next_idx(h
, idx
);
1015 * Puts an entry into a hashmap, boldly - no check whether key already exists.
1016 * The caller must place the entry (only its key and value, not link indexes)
1017 * in swap slot IDX_PUT.
1018 * Caller must ensure: the key does not exist yet in the hashmap.
1019 * that resize is not needed if !may_resize.
1020 * Returns: 1 if entry was put successfully.
1021 * -ENOMEM if may_resize==true and resize failed with -ENOMEM.
1022 * Cannot return -ENOMEM if !may_resize.
1024 static int hashmap_base_put_boldly(HashmapBase
*h
, unsigned idx
,
1025 struct swap_entries
*swap
, bool may_resize
) {
1026 struct ordered_hashmap_entry
*new_entry
;
1029 assert(idx
< n_buckets(h
));
1031 new_entry
= bucket_at_swap(swap
, IDX_PUT
);
1034 r
= resize_buckets(h
, 1);
1038 idx
= bucket_hash(h
, new_entry
->p
.b
.key
);
1040 assert(n_entries(h
) < n_buckets(h
));
1042 if (h
->type
== HASHMAP_TYPE_ORDERED
) {
1043 OrderedHashmap
*lh
= (OrderedHashmap
*) h
;
1045 new_entry
->iterate_next
= IDX_NIL
;
1046 new_entry
->iterate_previous
= lh
->iterate_list_tail
;
1048 if (lh
->iterate_list_tail
!= IDX_NIL
) {
1049 struct ordered_hashmap_entry
*old_tail
;
1051 old_tail
= ordered_bucket_at(lh
, lh
->iterate_list_tail
);
1052 assert(old_tail
->iterate_next
== IDX_NIL
);
1053 old_tail
->iterate_next
= IDX_PUT
;
1056 lh
->iterate_list_tail
= IDX_PUT
;
1057 if (lh
->iterate_list_head
== IDX_NIL
)
1058 lh
->iterate_list_head
= IDX_PUT
;
1061 assert_se(hashmap_put_robin_hood(h
, idx
, swap
) == false);
1064 #if ENABLE_DEBUG_HASHMAP
1065 h
->debug
.max_entries
= MAX(h
->debug
.max_entries
, n_entries(h
));
1072 #define hashmap_put_boldly(h, idx, swap, may_resize) \
1073 hashmap_base_put_boldly(HASHMAP_BASE(h), idx, swap, may_resize)
1076 * Returns 0 if resize is not needed.
1077 * 1 if successfully resized.
1078 * -ENOMEM on allocation failure.
1080 static int resize_buckets(HashmapBase
*h
, unsigned entries_add
) {
1081 struct swap_entries swap
;
1083 dib_raw_t
*old_dibs
, *new_dibs
;
1084 const struct hashmap_type_info
*hi
;
1085 unsigned idx
, optimal_idx
;
1086 unsigned old_n_buckets
, new_n_buckets
, n_rehashed
, new_n_entries
;
1092 hi
= &hashmap_type_info
[h
->type
];
1093 new_n_entries
= n_entries(h
) + entries_add
;
1096 if (_unlikely_(new_n_entries
< entries_add
))
1099 /* For direct storage we allow 100% load, because it's tiny. */
1100 if (!h
->has_indirect
&& new_n_entries
<= hi
->n_direct_buckets
)
1104 * Load factor = n/m = 1 - (1/INV_KEEP_FREE).
1105 * From it follows: m = n + n/(INV_KEEP_FREE - 1)
1107 new_n_buckets
= new_n_entries
+ new_n_entries
/ (INV_KEEP_FREE
- 1);
1109 if (_unlikely_(new_n_buckets
< new_n_entries
))
1112 if (_unlikely_(new_n_buckets
> UINT_MAX
/ (hi
->entry_size
+ sizeof(dib_raw_t
))))
1115 old_n_buckets
= n_buckets(h
);
1117 if (_likely_(new_n_buckets
<= old_n_buckets
))
1120 new_shift
= log2u_round_up(MAX(
1121 new_n_buckets
* (hi
->entry_size
+ sizeof(dib_raw_t
)),
1122 2 * sizeof(struct direct_storage
)));
1124 /* Realloc storage (buckets and DIB array). */
1125 new_storage
= realloc(h
->has_indirect
? h
->indirect
.storage
: NULL
,
1130 /* Must upgrade direct to indirect storage. */
1131 if (!h
->has_indirect
) {
1132 memcpy(new_storage
, h
->direct
.storage
,
1133 old_n_buckets
* (hi
->entry_size
+ sizeof(dib_raw_t
)));
1134 h
->indirect
.n_entries
= h
->n_direct_entries
;
1135 h
->indirect
.idx_lowest_entry
= 0;
1136 h
->n_direct_entries
= 0;
1139 /* Get a new hash key. If we've just upgraded to indirect storage,
1140 * allow reusing a previously generated key. It's still a different key
1141 * from the shared one that we used for direct storage. */
1142 get_hash_key(h
->indirect
.hash_key
, !h
->has_indirect
);
1144 h
->has_indirect
= true;
1145 h
->indirect
.storage
= new_storage
;
1146 h
->indirect
.n_buckets
= (1U << new_shift
) /
1147 (hi
->entry_size
+ sizeof(dib_raw_t
));
1149 old_dibs
= (dib_raw_t
*)((uint8_t*) new_storage
+ hi
->entry_size
* old_n_buckets
);
1150 new_dibs
= dib_raw_ptr(h
);
1153 * Move the DIB array to the new place, replacing valid DIB values with
1154 * DIB_RAW_REHASH to indicate all of the used buckets need rehashing.
1155 * Note: Overlap is not possible, because we have at least doubled the
1156 * number of buckets and dib_raw_t is smaller than any entry type.
1158 for (idx
= 0; idx
< old_n_buckets
; idx
++) {
1159 assert(old_dibs
[idx
] != DIB_RAW_REHASH
);
1160 new_dibs
[idx
] = old_dibs
[idx
] == DIB_RAW_FREE
? DIB_RAW_FREE
1164 /* Zero the area of newly added entries (including the old DIB area) */
1165 memzero(bucket_at(h
, old_n_buckets
),
1166 (n_buckets(h
) - old_n_buckets
) * hi
->entry_size
);
1168 /* The upper half of the new DIB array needs initialization */
1169 memset(&new_dibs
[old_n_buckets
], DIB_RAW_INIT
,
1170 (n_buckets(h
) - old_n_buckets
) * sizeof(dib_raw_t
));
1172 /* Rehash entries that need it */
1174 for (idx
= 0; idx
< old_n_buckets
; idx
++) {
1175 if (new_dibs
[idx
] != DIB_RAW_REHASH
)
1178 optimal_idx
= bucket_hash(h
, bucket_at(h
, idx
)->key
);
1181 * Not much to do if by luck the entry hashes to its current
1182 * location. Just set its DIB.
1184 if (optimal_idx
== idx
) {
1190 new_dibs
[idx
] = DIB_RAW_FREE
;
1191 bucket_move_entry(h
, &swap
, idx
, IDX_PUT
);
1192 /* bucket_move_entry does not clear the source */
1193 memzero(bucket_at(h
, idx
), hi
->entry_size
);
1197 * Find the new bucket for the current entry. This may make
1198 * another entry homeless and load it into IDX_PUT.
1200 rehash_next
= hashmap_put_robin_hood(h
, optimal_idx
, &swap
);
1203 /* Did the current entry displace another one? */
1205 optimal_idx
= bucket_hash(h
, bucket_at_swap(&swap
, IDX_PUT
)->p
.b
.key
);
1206 } while (rehash_next
);
1209 assert(n_rehashed
== n_entries(h
));
1215 * Finds an entry with a matching key
1216 * Returns: index of the found entry, or IDX_NIL if not found.
1218 static unsigned base_bucket_scan(HashmapBase
*h
, unsigned idx
, const void *key
) {
1219 struct hashmap_base_entry
*e
;
1220 unsigned dib
, distance
;
1221 dib_raw_t
*dibs
= dib_raw_ptr(h
);
1223 assert(idx
< n_buckets(h
));
1225 for (distance
= 0; ; distance
++) {
1226 if (dibs
[idx
] == DIB_RAW_FREE
)
1229 dib
= bucket_calculate_dib(h
, idx
, dibs
[idx
]);
1233 if (dib
== distance
) {
1234 e
= bucket_at(h
, idx
);
1235 if (h
->hash_ops
->compare(e
->key
, key
) == 0)
1239 idx
= next_idx(h
, idx
);
1242 #define bucket_scan(h, idx, key) base_bucket_scan(HASHMAP_BASE(h), idx, key)
1244 int hashmap_put(Hashmap
*h
, const void *key
, void *value
) {
1245 struct swap_entries swap
;
1246 struct plain_hashmap_entry
*e
;
1251 hash
= bucket_hash(h
, key
);
1252 idx
= bucket_scan(h
, hash
, key
);
1253 if (idx
!= IDX_NIL
) {
1254 e
= plain_bucket_at(h
, idx
);
1255 if (e
->value
== value
)
1260 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
;
1263 return hashmap_put_boldly(h
, hash
, &swap
, true);
1266 int set_put(Set
*s
, const void *key
) {
1267 struct swap_entries swap
;
1268 struct hashmap_base_entry
*e
;
1273 hash
= bucket_hash(s
, key
);
1274 idx
= bucket_scan(s
, hash
, key
);
1278 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
.b
;
1280 return hashmap_put_boldly(s
, hash
, &swap
, true);
1283 int hashmap_replace(Hashmap
*h
, const void *key
, void *value
) {
1284 struct swap_entries swap
;
1285 struct plain_hashmap_entry
*e
;
1290 hash
= bucket_hash(h
, key
);
1291 idx
= bucket_scan(h
, hash
, key
);
1292 if (idx
!= IDX_NIL
) {
1293 e
= plain_bucket_at(h
, idx
);
1294 #if ENABLE_DEBUG_HASHMAP
1295 /* Although the key is equal, the key pointer may have changed,
1296 * and this would break our assumption for iterating. So count
1297 * this operation as incompatible with iteration. */
1298 if (e
->b
.key
!= key
) {
1299 h
->b
.debug
.put_count
++;
1300 h
->b
.debug
.rem_count
++;
1301 h
->b
.debug
.last_rem_idx
= idx
;
1306 hashmap_set_dirty(h
);
1311 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
;
1314 return hashmap_put_boldly(h
, hash
, &swap
, true);
1317 int hashmap_update(Hashmap
*h
, const void *key
, void *value
) {
1318 struct plain_hashmap_entry
*e
;
1323 hash
= bucket_hash(h
, key
);
1324 idx
= bucket_scan(h
, hash
, key
);
1328 e
= plain_bucket_at(h
, idx
);
1330 hashmap_set_dirty(h
);
1335 void *internal_hashmap_get(HashmapBase
*h
, const void *key
) {
1336 struct hashmap_base_entry
*e
;
1342 hash
= bucket_hash(h
, key
);
1343 idx
= bucket_scan(h
, hash
, key
);
1347 e
= bucket_at(h
, idx
);
1348 return entry_value(h
, e
);
1351 void *hashmap_get2(Hashmap
*h
, const void *key
, void **key2
) {
1352 struct plain_hashmap_entry
*e
;
1358 hash
= bucket_hash(h
, key
);
1359 idx
= bucket_scan(h
, hash
, key
);
1363 e
= plain_bucket_at(h
, idx
);
1365 *key2
= (void*) e
->b
.key
;
1370 bool internal_hashmap_contains(HashmapBase
*h
, const void *key
) {
1376 hash
= bucket_hash(h
, key
);
1377 return bucket_scan(h
, hash
, key
) != IDX_NIL
;
1380 void *internal_hashmap_remove(HashmapBase
*h
, const void *key
) {
1381 struct hashmap_base_entry
*e
;
1388 hash
= bucket_hash(h
, key
);
1389 idx
= bucket_scan(h
, hash
, key
);
1393 e
= bucket_at(h
, idx
);
1394 data
= entry_value(h
, e
);
1395 remove_entry(h
, idx
);
1400 void *hashmap_remove2(Hashmap
*h
, const void *key
, void **rkey
) {
1401 struct plain_hashmap_entry
*e
;
1411 hash
= bucket_hash(h
, key
);
1412 idx
= bucket_scan(h
, hash
, key
);
1413 if (idx
== IDX_NIL
) {
1419 e
= plain_bucket_at(h
, idx
);
1422 *rkey
= (void*) e
->b
.key
;
1424 remove_entry(h
, idx
);
1429 int hashmap_remove_and_put(Hashmap
*h
, const void *old_key
, const void *new_key
, void *value
) {
1430 struct swap_entries swap
;
1431 struct plain_hashmap_entry
*e
;
1432 unsigned old_hash
, new_hash
, idx
;
1437 old_hash
= bucket_hash(h
, old_key
);
1438 idx
= bucket_scan(h
, old_hash
, old_key
);
1442 new_hash
= bucket_hash(h
, new_key
);
1443 if (bucket_scan(h
, new_hash
, new_key
) != IDX_NIL
)
1446 remove_entry(h
, idx
);
1448 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
;
1451 assert_se(hashmap_put_boldly(h
, new_hash
, &swap
, false) == 1);
1456 int set_remove_and_put(Set
*s
, const void *old_key
, const void *new_key
) {
1457 struct swap_entries swap
;
1458 struct hashmap_base_entry
*e
;
1459 unsigned old_hash
, new_hash
, idx
;
1464 old_hash
= bucket_hash(s
, old_key
);
1465 idx
= bucket_scan(s
, old_hash
, old_key
);
1469 new_hash
= bucket_hash(s
, new_key
);
1470 if (bucket_scan(s
, new_hash
, new_key
) != IDX_NIL
)
1473 remove_entry(s
, idx
);
1475 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
.b
;
1477 assert_se(hashmap_put_boldly(s
, new_hash
, &swap
, false) == 1);
1482 int hashmap_remove_and_replace(Hashmap
*h
, const void *old_key
, const void *new_key
, void *value
) {
1483 struct swap_entries swap
;
1484 struct plain_hashmap_entry
*e
;
1485 unsigned old_hash
, new_hash
, idx_old
, idx_new
;
1490 old_hash
= bucket_hash(h
, old_key
);
1491 idx_old
= bucket_scan(h
, old_hash
, old_key
);
1492 if (idx_old
== IDX_NIL
)
1495 old_key
= bucket_at(HASHMAP_BASE(h
), idx_old
)->key
;
1497 new_hash
= bucket_hash(h
, new_key
);
1498 idx_new
= bucket_scan(h
, new_hash
, new_key
);
1499 if (idx_new
!= IDX_NIL
)
1500 if (idx_old
!= idx_new
) {
1501 remove_entry(h
, idx_new
);
1502 /* Compensate for a possible backward shift. */
1503 if (old_key
!= bucket_at(HASHMAP_BASE(h
), idx_old
)->key
)
1504 idx_old
= prev_idx(HASHMAP_BASE(h
), idx_old
);
1505 assert(old_key
== bucket_at(HASHMAP_BASE(h
), idx_old
)->key
);
1508 remove_entry(h
, idx_old
);
1510 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
;
1513 assert_se(hashmap_put_boldly(h
, new_hash
, &swap
, false) == 1);
1518 void *hashmap_remove_value(Hashmap
*h
, const void *key
, void *value
) {
1519 struct plain_hashmap_entry
*e
;
1525 hash
= bucket_hash(h
, key
);
1526 idx
= bucket_scan(h
, hash
, key
);
1530 e
= plain_bucket_at(h
, idx
);
1531 if (e
->value
!= value
)
1534 remove_entry(h
, idx
);
1539 static unsigned find_first_entry(HashmapBase
*h
) {
1540 Iterator i
= ITERATOR_FIRST
;
1542 if (!h
|| !n_entries(h
))
1545 return hashmap_iterate_entry(h
, &i
);
1548 void *internal_hashmap_first(HashmapBase
*h
) {
1551 idx
= find_first_entry(h
);
1555 return entry_value(h
, bucket_at(h
, idx
));
1558 void *internal_hashmap_first_key(HashmapBase
*h
) {
1559 struct hashmap_base_entry
*e
;
1562 idx
= find_first_entry(h
);
1566 e
= bucket_at(h
, idx
);
1567 return (void*) e
->key
;
1570 void *internal_hashmap_steal_first(HashmapBase
*h
) {
1571 struct hashmap_base_entry
*e
;
1575 idx
= find_first_entry(h
);
1579 e
= bucket_at(h
, idx
);
1580 data
= entry_value(h
, e
);
1581 remove_entry(h
, idx
);
1586 void *internal_hashmap_steal_first_key(HashmapBase
*h
) {
1587 struct hashmap_base_entry
*e
;
1591 idx
= find_first_entry(h
);
1595 e
= bucket_at(h
, idx
);
1596 key
= (void*) e
->key
;
1597 remove_entry(h
, idx
);
1602 unsigned internal_hashmap_size(HashmapBase
*h
) {
1607 return n_entries(h
);
1610 unsigned internal_hashmap_buckets(HashmapBase
*h
) {
1615 return n_buckets(h
);
1618 int internal_hashmap_merge(Hashmap
*h
, Hashmap
*other
) {
1624 HASHMAP_FOREACH_IDX(idx
, HASHMAP_BASE(other
), i
) {
1625 struct plain_hashmap_entry
*pe
= plain_bucket_at(other
, idx
);
1628 r
= hashmap_put(h
, pe
->b
.key
, pe
->value
);
1629 if (r
< 0 && r
!= -EEXIST
)
1636 int set_merge(Set
*s
, Set
*other
) {
1642 HASHMAP_FOREACH_IDX(idx
, HASHMAP_BASE(other
), i
) {
1643 struct set_entry
*se
= set_bucket_at(other
, idx
);
1646 r
= set_put(s
, se
->b
.key
);
1654 int internal_hashmap_reserve(HashmapBase
*h
, unsigned entries_add
) {
1659 r
= resize_buckets(h
, entries_add
);
1667 * The same as hashmap_merge(), but every new item from other is moved to h.
1668 * Keys already in h are skipped and stay in other.
1669 * Returns: 0 on success.
1670 * -ENOMEM on alloc failure, in which case no move has been done.
1672 int internal_hashmap_move(HashmapBase
*h
, HashmapBase
*other
) {
1673 struct swap_entries swap
;
1674 struct hashmap_base_entry
*e
, *n
;
1684 assert(other
->type
== h
->type
);
1687 * This reserves buckets for the worst case, where none of other's
1688 * entries are yet present in h. This is preferable to risking
1689 * an allocation failure in the middle of the moving and having to
1690 * rollback or return a partial result.
1692 r
= resize_buckets(h
, n_entries(other
));
1696 HASHMAP_FOREACH_IDX(idx
, other
, i
) {
1699 e
= bucket_at(other
, idx
);
1700 h_hash
= bucket_hash(h
, e
->key
);
1701 if (bucket_scan(h
, h_hash
, e
->key
) != IDX_NIL
)
1704 n
= &bucket_at_swap(&swap
, IDX_PUT
)->p
.b
;
1706 if (h
->type
!= HASHMAP_TYPE_SET
)
1707 ((struct plain_hashmap_entry
*) n
)->value
=
1708 ((struct plain_hashmap_entry
*) e
)->value
;
1709 assert_se(hashmap_put_boldly(h
, h_hash
, &swap
, false) == 1);
1711 remove_entry(other
, idx
);
1717 int internal_hashmap_move_one(HashmapBase
*h
, HashmapBase
*other
, const void *key
) {
1718 struct swap_entries swap
;
1719 unsigned h_hash
, other_hash
, idx
;
1720 struct hashmap_base_entry
*e
, *n
;
1725 h_hash
= bucket_hash(h
, key
);
1726 if (bucket_scan(h
, h_hash
, key
) != IDX_NIL
)
1732 assert(other
->type
== h
->type
);
1734 other_hash
= bucket_hash(other
, key
);
1735 idx
= bucket_scan(other
, other_hash
, key
);
1739 e
= bucket_at(other
, idx
);
1741 n
= &bucket_at_swap(&swap
, IDX_PUT
)->p
.b
;
1743 if (h
->type
!= HASHMAP_TYPE_SET
)
1744 ((struct plain_hashmap_entry
*) n
)->value
=
1745 ((struct plain_hashmap_entry
*) e
)->value
;
1746 r
= hashmap_put_boldly(h
, h_hash
, &swap
, true);
1750 remove_entry(other
, idx
);
1754 HashmapBase
*internal_hashmap_copy(HashmapBase
*h
) {
1760 copy
= hashmap_base_new(h
->hash_ops
, h
->type HASHMAP_DEBUG_SRC_ARGS
);
1765 case HASHMAP_TYPE_PLAIN
:
1766 case HASHMAP_TYPE_ORDERED
:
1767 r
= hashmap_merge((Hashmap
*)copy
, (Hashmap
*)h
);
1769 case HASHMAP_TYPE_SET
:
1770 r
= set_merge((Set
*)copy
, (Set
*)h
);
1773 assert_not_reached("Unknown hashmap type");
1777 internal_hashmap_free(copy
);
1784 char **internal_hashmap_get_strv(HashmapBase
*h
) {
1789 sv
= new(char*, n_entries(h
)+1);
1794 HASHMAP_FOREACH_IDX(idx
, h
, i
)
1795 sv
[n
++] = entry_value(h
, bucket_at(h
, idx
));
1801 void *ordered_hashmap_next(OrderedHashmap
*h
, const void *key
) {
1802 struct ordered_hashmap_entry
*e
;
1808 hash
= bucket_hash(h
, key
);
1809 idx
= bucket_scan(h
, hash
, key
);
1813 e
= ordered_bucket_at(h
, idx
);
1814 if (e
->iterate_next
== IDX_NIL
)
1816 return ordered_bucket_at(h
, e
->iterate_next
)->p
.value
;
1819 int set_consume(Set
*s
, void *value
) {
1825 r
= set_put(s
, value
);
1832 int set_put_strdup(Set
*s
, const char *p
) {
1838 if (set_contains(s
, (char*) p
))
1845 return set_consume(s
, c
);
1848 int set_put_strdupv(Set
*s
, char **l
) {
1854 STRV_FOREACH(i
, l
) {
1855 r
= set_put_strdup(s
, *i
);
1865 int set_put_strsplit(Set
*s
, const char *v
, const char *separators
, ExtractFlags flags
) {
1875 r
= extract_first_word(&p
, &word
, separators
, flags
);
1879 r
= set_consume(s
, word
);
1885 /* expand the cachemem if needed, return true if newly (re)activated. */
1886 static int cachemem_maintain(CacheMem
*mem
, unsigned size
) {
1889 if (!GREEDY_REALLOC(mem
->ptr
, mem
->n_allocated
, size
)) {
1902 int iterated_cache_get(IteratedCache
*cache
, const void ***res_keys
, const void ***res_values
, unsigned *res_n_entries
) {
1903 bool sync_keys
= false, sync_values
= false;
1908 assert(cache
->hashmap
);
1910 size
= n_entries(cache
->hashmap
);
1913 r
= cachemem_maintain(&cache
->keys
, size
);
1919 cache
->keys
.active
= false;
1922 r
= cachemem_maintain(&cache
->values
, size
);
1928 cache
->values
.active
= false;
1930 if (cache
->hashmap
->dirty
) {
1931 if (cache
->keys
.active
)
1933 if (cache
->values
.active
)
1936 cache
->hashmap
->dirty
= false;
1939 if (sync_keys
|| sync_values
) {
1944 HASHMAP_FOREACH_IDX(idx
, cache
->hashmap
, iter
) {
1945 struct hashmap_base_entry
*e
;
1947 e
= bucket_at(cache
->hashmap
, idx
);
1950 cache
->keys
.ptr
[i
] = e
->key
;
1952 cache
->values
.ptr
[i
] = entry_value(cache
->hashmap
, e
);
1958 *res_keys
= cache
->keys
.ptr
;
1960 *res_values
= cache
->values
.ptr
;
1962 *res_n_entries
= size
;
1967 IteratedCache
*iterated_cache_free(IteratedCache
*cache
) {
1969 free(cache
->keys
.ptr
);
1970 free(cache
->values
.ptr
);