1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
9 #include "alloc-util.h"
12 #include "logarithm.h"
14 #include "memory-util.h"
16 #include "missing_syscall.h"
17 #include "process-util.h"
18 #include "random-util.h"
20 #include "siphash24.h"
21 #include "string-util.h"
24 #if ENABLE_DEBUG_HASHMAP
29 * Implementation of hashmaps.
31 * - uses less RAM compared to closed addressing (chaining), because
32 * our entries are small (especially in Sets, which tend to contain
33 * the majority of entries in systemd).
34 * Collision resolution: Robin Hood
35 * - tends to equalize displacement of entries from their optimal buckets.
36 * Probe sequence: linear
37 * - though theoretically worse than random probing/uniform hashing/double
38 * hashing, it is good for cache locality.
41 * Celis, P. 1986. Robin Hood Hashing.
42 * Ph.D. Dissertation. University of Waterloo, Waterloo, Ont., Canada, Canada.
43 * https://cs.uwaterloo.ca/research/tr/1986/CS-86-14.pdf
44 * - The results are derived for random probing. Suggests deletion with
45 * tombstones and two mean-centered search methods. None of that works
46 * well for linear probing.
48 * Janson, S. 2005. Individual displacements for linear probing hashing with different insertion policies.
49 * ACM Trans. Algorithms 1, 2 (October 2005), 177-213.
50 * DOI=10.1145/1103963.1103964 http://doi.acm.org/10.1145/1103963.1103964
51 * http://www.math.uu.se/~svante/papers/sj157.pdf
52 * - Applies to Robin Hood with linear probing. Contains remarks on
53 * the unsuitability of mean-centered search with linear probing.
55 * Viola, A. 2005. Exact distribution of individual displacements in linear probing hashing.
56 * ACM Trans. Algorithms 1, 2 (October 2005), 214-242.
57 * DOI=10.1145/1103963.1103965 http://doi.acm.org/10.1145/1103963.1103965
58 * - Similar to Janson. Note that Viola writes about C_{m,n} (number of probes
59 * in a successful search), and Janson writes about displacement. C = d + 1.
61 * Goossaert, E. 2013. Robin Hood hashing: backward shift deletion.
62 * http://codecapsule.com/2013/11/17/robin-hood-hashing-backward-shift-deletion/
63 * - Explanation of backward shift deletion with pictures.
65 * Khuong, P. 2013. The Other Robin Hood Hashing.
66 * http://www.pvk.ca/Blog/2013/11/26/the-other-robin-hood-hashing/
67 * - Short summary of random vs. linear probing, and tombstones vs. backward shift.
71 * XXX Ideas for improvement:
72 * For unordered hashmaps, randomize iteration order, similarly to Perl:
73 * http://blog.booking.com/hardening-perls-hash-function.html
76 /* INV_KEEP_FREE = 1 / (1 - max_load_factor)
77 * e.g. 1 / (1 - 0.8) = 5 ... keep one fifth of the buckets free. */
78 #define INV_KEEP_FREE 5U
80 /* Fields common to entries of all hashmap/set types */
81 struct hashmap_base_entry
{
85 /* Entry types for specific hashmap/set types
86 * hashmap_base_entry must be at the beginning of each entry struct. */
88 struct plain_hashmap_entry
{
89 struct hashmap_base_entry b
;
93 struct ordered_hashmap_entry
{
94 struct plain_hashmap_entry p
;
95 unsigned iterate_next
, iterate_previous
;
99 struct hashmap_base_entry b
;
102 /* In several functions it is advantageous to have the hash table extended
103 * virtually by a couple of additional buckets. We reserve special index values
104 * for these "swap" buckets. */
105 #define _IDX_SWAP_BEGIN (UINT_MAX - 3)
106 #define IDX_PUT (_IDX_SWAP_BEGIN + 0)
107 #define IDX_TMP (_IDX_SWAP_BEGIN + 1)
108 #define _IDX_SWAP_END (_IDX_SWAP_BEGIN + 2)
110 #define IDX_FIRST (UINT_MAX - 1) /* special index for freshly initialized iterators */
111 #define IDX_NIL UINT_MAX /* special index value meaning "none" or "end" */
113 assert_cc(IDX_FIRST
== _IDX_SWAP_END
);
114 assert_cc(IDX_FIRST
== _IDX_ITERATOR_FIRST
);
116 /* Storage space for the "swap" buckets.
117 * All entry types can fit into an ordered_hashmap_entry. */
118 struct swap_entries
{
119 struct ordered_hashmap_entry e
[_IDX_SWAP_END
- _IDX_SWAP_BEGIN
];
122 /* Distance from Initial Bucket */
123 typedef uint8_t dib_raw_t
;
124 #define DIB_RAW_OVERFLOW ((dib_raw_t)0xfdU) /* indicates DIB value is greater than representable */
125 #define DIB_RAW_REHASH ((dib_raw_t)0xfeU) /* entry yet to be rehashed during in-place resize */
126 #define DIB_RAW_FREE ((dib_raw_t)0xffU) /* a free bucket */
127 #define DIB_RAW_INIT ((char)DIB_RAW_FREE) /* a byte to memset a DIB store with when initializing */
129 #define DIB_FREE UINT_MAX
131 #if ENABLE_DEBUG_HASHMAP
132 struct hashmap_debug_info
{
133 LIST_FIELDS(struct hashmap_debug_info
, debug_list
);
134 unsigned max_entries
; /* high watermark of n_entries */
136 /* who allocated this hashmap */
141 /* fields to detect modification while iterating */
142 unsigned put_count
; /* counts puts into the hashmap */
143 unsigned rem_count
; /* counts removals from hashmap */
144 unsigned last_rem_idx
; /* remembers last removal index */
147 /* Tracks all existing hashmaps. Get at it from gdb. See sd_dump_hashmaps.py */
148 static LIST_HEAD(struct hashmap_debug_info
, hashmap_debug_list
);
149 static pthread_mutex_t hashmap_debug_list_mutex
= PTHREAD_MUTEX_INITIALIZER
;
154 HASHMAP_TYPE_ORDERED
,
159 struct _packed_ indirect_storage
{
160 void *storage
; /* where buckets and DIBs are stored */
161 uint8_t hash_key
[HASH_KEY_SIZE
]; /* hash key; changes during resize */
163 unsigned n_entries
; /* number of stored entries */
164 unsigned n_buckets
; /* number of buckets */
166 unsigned idx_lowest_entry
; /* Index below which all buckets are free.
167 Makes "while (hashmap_steal_first())" loops
168 O(n) instead of O(n^2) for unordered hashmaps. */
169 uint8_t _pad
[3]; /* padding for the whole HashmapBase */
170 /* The bitfields in HashmapBase complete the alignment of the whole thing. */
173 struct direct_storage
{
174 /* This gives us 39 bytes on 64bit, or 35 bytes on 32bit.
175 * That's room for 4 set_entries + 4 DIB bytes + 3 unused bytes on 64bit,
176 * or 7 set_entries + 7 DIB bytes + 0 unused bytes on 32bit. */
177 uint8_t storage
[sizeof(struct indirect_storage
)];
180 #define DIRECT_BUCKETS(entry_t) \
181 (sizeof(struct direct_storage) / (sizeof(entry_t) + sizeof(dib_raw_t)))
183 /* We should be able to store at least one entry directly. */
184 assert_cc(DIRECT_BUCKETS(struct ordered_hashmap_entry
) >= 1);
186 /* We have 3 bits for n_direct_entries. */
187 assert_cc(DIRECT_BUCKETS(struct set_entry
) < (1 << 3));
189 /* Hashmaps with directly stored entries all use this shared hash key.
190 * It's no big deal if the key is guessed, because there can be only
191 * a handful of directly stored entries in a hashmap. When a hashmap
192 * outgrows direct storage, it gets its own key for indirect storage. */
193 static uint8_t shared_hash_key
[HASH_KEY_SIZE
];
195 /* Fields that all hashmap/set types must have */
197 const struct hash_ops
*hash_ops
; /* hash and compare ops to use */
200 struct indirect_storage indirect
; /* if has_indirect */
201 struct direct_storage direct
; /* if !has_indirect */
204 enum HashmapType type
:2; /* HASHMAP_TYPE_* */
205 bool has_indirect
:1; /* whether indirect storage is used */
206 unsigned n_direct_entries
:3; /* Number of entries in direct storage.
207 * Only valid if !has_indirect. */
208 bool from_pool
:1; /* whether was allocated from mempool */
209 bool dirty
:1; /* whether dirtied since last iterated_cache_get() */
210 bool cached
:1; /* whether this hashmap is being cached */
212 #if ENABLE_DEBUG_HASHMAP
213 struct hashmap_debug_info debug
;
217 /* Specific hash types
218 * HashmapBase must be at the beginning of each hashmap struct. */
221 struct HashmapBase b
;
224 struct OrderedHashmap
{
225 struct HashmapBase b
;
226 unsigned iterate_list_head
, iterate_list_tail
;
230 struct HashmapBase b
;
233 typedef struct CacheMem
{
239 struct IteratedCache
{
240 HashmapBase
*hashmap
;
241 CacheMem keys
, values
;
244 DEFINE_MEMPOOL(hashmap_pool
, Hashmap
, 8);
245 DEFINE_MEMPOOL(ordered_hashmap_pool
, OrderedHashmap
, 8);
246 /* No need for a separate Set pool */
247 assert_cc(sizeof(Hashmap
) == sizeof(Set
));
249 struct hashmap_type_info
{
252 struct mempool
*mempool
;
253 unsigned n_direct_buckets
;
256 static _used_
const struct hashmap_type_info hashmap_type_info
[_HASHMAP_TYPE_MAX
] = {
257 [HASHMAP_TYPE_PLAIN
] = {
258 .head_size
= sizeof(Hashmap
),
259 .entry_size
= sizeof(struct plain_hashmap_entry
),
260 .mempool
= &hashmap_pool
,
261 .n_direct_buckets
= DIRECT_BUCKETS(struct plain_hashmap_entry
),
263 [HASHMAP_TYPE_ORDERED
] = {
264 .head_size
= sizeof(OrderedHashmap
),
265 .entry_size
= sizeof(struct ordered_hashmap_entry
),
266 .mempool
= &ordered_hashmap_pool
,
267 .n_direct_buckets
= DIRECT_BUCKETS(struct ordered_hashmap_entry
),
269 [HASHMAP_TYPE_SET
] = {
270 .head_size
= sizeof(Set
),
271 .entry_size
= sizeof(struct set_entry
),
272 .mempool
= &hashmap_pool
,
273 .n_direct_buckets
= DIRECT_BUCKETS(struct set_entry
),
277 void hashmap_cleanup_pools(void) {
280 /* The pool is only allocated by the main thread, but the memory can be passed to other
281 * threads. Let's clean up if we are the main thread and no other threads are live. */
283 /* We build our own is_main_thread() here, which doesn't use C11 TLS based caching of the
284 * result. That's because valgrind apparently doesn't like TLS to be used from a GCC destructor. */
285 if (getpid() != gettid())
286 return (void) log_debug("Not cleaning up memory pools, not in main thread.");
288 r
= get_process_threads(0);
290 return (void) log_debug_errno(r
, "Failed to determine number of threads, not cleaning up memory pools: %m");
292 return (void) log_debug("Not cleaning up memory pools, running in multi-threaded process.");
294 mempool_drop(&hashmap_pool
);
295 mempool_drop(&ordered_hashmap_pool
);
299 _destructor_
static void cleanup_pools(void) {
300 /* Be nice to valgrind */
301 hashmap_cleanup_pools();
305 static unsigned n_buckets(HashmapBase
*h
) {
306 return h
->has_indirect
? h
->indirect
.n_buckets
307 : hashmap_type_info
[h
->type
].n_direct_buckets
;
310 static unsigned n_entries(HashmapBase
*h
) {
311 return h
->has_indirect
? h
->indirect
.n_entries
312 : h
->n_direct_entries
;
315 static void n_entries_inc(HashmapBase
*h
) {
317 h
->indirect
.n_entries
++;
319 h
->n_direct_entries
++;
322 static void n_entries_dec(HashmapBase
*h
) {
324 h
->indirect
.n_entries
--;
326 h
->n_direct_entries
--;
329 static void* storage_ptr(HashmapBase
*h
) {
330 return h
->has_indirect
? h
->indirect
.storage
334 static uint8_t* hash_key(HashmapBase
*h
) {
335 return h
->has_indirect
? h
->indirect
.hash_key
339 static unsigned base_bucket_hash(HashmapBase
*h
, const void *p
) {
340 struct siphash state
;
343 siphash24_init(&state
, hash_key(h
));
345 h
->hash_ops
->hash(p
, &state
);
347 hash
= siphash24_finalize(&state
);
349 return (unsigned) (hash
% n_buckets(h
));
351 #define bucket_hash(h, p) base_bucket_hash(HASHMAP_BASE(h), p)
353 static void base_set_dirty(HashmapBase
*h
) {
356 #define hashmap_set_dirty(h) base_set_dirty(HASHMAP_BASE(h))
358 static void get_hash_key(uint8_t hash_key
[HASH_KEY_SIZE
], bool reuse_is_ok
) {
359 static uint8_t current
[HASH_KEY_SIZE
];
360 static bool current_initialized
= false;
362 /* Returns a hash function key to use. In order to keep things
363 * fast we will not generate a new key each time we allocate a
364 * new hash table. Instead, we'll just reuse the most recently
365 * generated one, except if we never generated one or when we
366 * are rehashing an entire hash table because we reached a
369 if (!current_initialized
|| !reuse_is_ok
) {
370 random_bytes(current
, sizeof(current
));
371 current_initialized
= true;
374 memcpy(hash_key
, current
, sizeof(current
));
377 static struct hashmap_base_entry
* bucket_at(HashmapBase
*h
, unsigned idx
) {
378 return CAST_ALIGN_PTR(
379 struct hashmap_base_entry
,
380 (uint8_t *) storage_ptr(h
) + idx
* hashmap_type_info
[h
->type
].entry_size
);
383 static struct plain_hashmap_entry
* plain_bucket_at(Hashmap
*h
, unsigned idx
) {
384 return (struct plain_hashmap_entry
*) bucket_at(HASHMAP_BASE(h
), idx
);
387 static struct ordered_hashmap_entry
* ordered_bucket_at(OrderedHashmap
*h
, unsigned idx
) {
388 return (struct ordered_hashmap_entry
*) bucket_at(HASHMAP_BASE(h
), idx
);
391 static struct set_entry
*set_bucket_at(Set
*h
, unsigned idx
) {
392 return (struct set_entry
*) bucket_at(HASHMAP_BASE(h
), idx
);
395 static struct ordered_hashmap_entry
* bucket_at_swap(struct swap_entries
*swap
, unsigned idx
) {
396 return &swap
->e
[idx
- _IDX_SWAP_BEGIN
];
399 /* Returns a pointer to the bucket at index idx.
400 * Understands real indexes and swap indexes, hence "_virtual". */
401 static struct hashmap_base_entry
* bucket_at_virtual(HashmapBase
*h
, struct swap_entries
*swap
,
403 if (idx
< _IDX_SWAP_BEGIN
)
404 return bucket_at(h
, idx
);
406 if (idx
< _IDX_SWAP_END
)
407 return &bucket_at_swap(swap
, idx
)->p
.b
;
409 assert_not_reached();
412 static dib_raw_t
* dib_raw_ptr(HashmapBase
*h
) {
414 ((uint8_t*) storage_ptr(h
) + hashmap_type_info
[h
->type
].entry_size
* n_buckets(h
));
417 static unsigned bucket_distance(HashmapBase
*h
, unsigned idx
, unsigned from
) {
418 return idx
>= from
? idx
- from
419 : n_buckets(h
) + idx
- from
;
422 static unsigned bucket_calculate_dib(HashmapBase
*h
, unsigned idx
, dib_raw_t raw_dib
) {
423 unsigned initial_bucket
;
425 if (raw_dib
== DIB_RAW_FREE
)
428 if (_likely_(raw_dib
< DIB_RAW_OVERFLOW
))
432 * Having an overflow DIB value is very unlikely. The hash function
433 * would have to be bad. For example, in a table of size 2^24 filled
434 * to load factor 0.9 the maximum observed DIB is only about 60.
435 * In theory (assuming I used Maxima correctly), for an infinite size
436 * hash table with load factor 0.8 the probability of a given entry
437 * having DIB > 40 is 1.9e-8.
438 * This returns the correct DIB value by recomputing the hash value in
439 * the unlikely case. XXX Hitting this case could be a hint to rehash.
441 initial_bucket
= bucket_hash(h
, bucket_at(h
, idx
)->key
);
442 return bucket_distance(h
, idx
, initial_bucket
);
445 static void bucket_set_dib(HashmapBase
*h
, unsigned idx
, unsigned dib
) {
446 dib_raw_ptr(h
)[idx
] = dib
!= DIB_FREE
? MIN(dib
, DIB_RAW_OVERFLOW
) : DIB_RAW_FREE
;
449 static unsigned skip_free_buckets(HashmapBase
*h
, unsigned idx
) {
452 dibs
= dib_raw_ptr(h
);
454 for ( ; idx
< n_buckets(h
); idx
++)
455 if (dibs
[idx
] != DIB_RAW_FREE
)
461 static void bucket_mark_free(HashmapBase
*h
, unsigned idx
) {
462 memzero(bucket_at(h
, idx
), hashmap_type_info
[h
->type
].entry_size
);
463 bucket_set_dib(h
, idx
, DIB_FREE
);
466 static void bucket_move_entry(HashmapBase
*h
, struct swap_entries
*swap
,
467 unsigned from
, unsigned to
) {
468 struct hashmap_base_entry
*e_from
, *e_to
;
472 e_from
= bucket_at_virtual(h
, swap
, from
);
473 e_to
= bucket_at_virtual(h
, swap
, to
);
475 memcpy(e_to
, e_from
, hashmap_type_info
[h
->type
].entry_size
);
477 if (h
->type
== HASHMAP_TYPE_ORDERED
) {
478 OrderedHashmap
*lh
= (OrderedHashmap
*) h
;
479 struct ordered_hashmap_entry
*le
, *le_to
;
481 le_to
= (struct ordered_hashmap_entry
*) e_to
;
483 if (le_to
->iterate_next
!= IDX_NIL
) {
484 le
= (struct ordered_hashmap_entry
*)
485 bucket_at_virtual(h
, swap
, le_to
->iterate_next
);
486 le
->iterate_previous
= to
;
489 if (le_to
->iterate_previous
!= IDX_NIL
) {
490 le
= (struct ordered_hashmap_entry
*)
491 bucket_at_virtual(h
, swap
, le_to
->iterate_previous
);
492 le
->iterate_next
= to
;
495 if (lh
->iterate_list_head
== from
)
496 lh
->iterate_list_head
= to
;
497 if (lh
->iterate_list_tail
== from
)
498 lh
->iterate_list_tail
= to
;
502 static unsigned next_idx(HashmapBase
*h
, unsigned idx
) {
503 return (idx
+ 1U) % n_buckets(h
);
506 static unsigned prev_idx(HashmapBase
*h
, unsigned idx
) {
507 return (n_buckets(h
) + idx
- 1U) % n_buckets(h
);
510 static void* entry_value(HashmapBase
*h
, struct hashmap_base_entry
*e
) {
513 case HASHMAP_TYPE_PLAIN
:
514 case HASHMAP_TYPE_ORDERED
:
515 return ((struct plain_hashmap_entry
*)e
)->value
;
517 case HASHMAP_TYPE_SET
:
518 return (void*) e
->key
;
521 assert_not_reached();
525 static void base_remove_entry(HashmapBase
*h
, unsigned idx
) {
526 unsigned left
, right
, prev
, dib
;
527 dib_raw_t raw_dib
, *dibs
;
529 dibs
= dib_raw_ptr(h
);
530 assert(dibs
[idx
] != DIB_RAW_FREE
);
532 #if ENABLE_DEBUG_HASHMAP
533 h
->debug
.rem_count
++;
534 h
->debug
.last_rem_idx
= idx
;
538 /* Find the stop bucket ("right"). It is either free or has DIB == 0. */
539 for (right
= next_idx(h
, left
); ; right
= next_idx(h
, right
)) {
540 raw_dib
= dibs
[right
];
541 if (IN_SET(raw_dib
, 0, DIB_RAW_FREE
))
544 /* The buckets are not supposed to be all occupied and with DIB > 0.
545 * That would mean we could make everyone better off by shifting them
546 * backward. This scenario is impossible. */
547 assert(left
!= right
);
550 if (h
->type
== HASHMAP_TYPE_ORDERED
) {
551 OrderedHashmap
*lh
= (OrderedHashmap
*) h
;
552 struct ordered_hashmap_entry
*le
= ordered_bucket_at(lh
, idx
);
554 if (le
->iterate_next
!= IDX_NIL
)
555 ordered_bucket_at(lh
, le
->iterate_next
)->iterate_previous
= le
->iterate_previous
;
557 lh
->iterate_list_tail
= le
->iterate_previous
;
559 if (le
->iterate_previous
!= IDX_NIL
)
560 ordered_bucket_at(lh
, le
->iterate_previous
)->iterate_next
= le
->iterate_next
;
562 lh
->iterate_list_head
= le
->iterate_next
;
565 /* Now shift all buckets in the interval (left, right) one step backwards */
566 for (prev
= left
, left
= next_idx(h
, left
); left
!= right
;
567 prev
= left
, left
= next_idx(h
, left
)) {
568 dib
= bucket_calculate_dib(h
, left
, dibs
[left
]);
570 bucket_move_entry(h
, NULL
, left
, prev
);
571 bucket_set_dib(h
, prev
, dib
- 1);
574 bucket_mark_free(h
, prev
);
578 #define remove_entry(h, idx) base_remove_entry(HASHMAP_BASE(h), idx)
580 static unsigned hashmap_iterate_in_insertion_order(OrderedHashmap
*h
, Iterator
*i
) {
581 struct ordered_hashmap_entry
*e
;
587 if (i
->idx
== IDX_NIL
)
590 if (i
->idx
== IDX_FIRST
&& h
->iterate_list_head
== IDX_NIL
)
593 if (i
->idx
== IDX_FIRST
) {
594 idx
= h
->iterate_list_head
;
595 e
= ordered_bucket_at(h
, idx
);
598 e
= ordered_bucket_at(h
, idx
);
600 * We allow removing the current entry while iterating, but removal may cause
601 * a backward shift. The next entry may thus move one bucket to the left.
602 * To detect when it happens, we remember the key pointer of the entry we were
603 * going to iterate next. If it does not match, there was a backward shift.
605 if (e
->p
.b
.key
!= i
->next_key
) {
606 idx
= prev_idx(HASHMAP_BASE(h
), idx
);
607 e
= ordered_bucket_at(h
, idx
);
609 assert(e
->p
.b
.key
== i
->next_key
);
612 #if ENABLE_DEBUG_HASHMAP
616 if (e
->iterate_next
!= IDX_NIL
) {
617 struct ordered_hashmap_entry
*n
;
618 i
->idx
= e
->iterate_next
;
619 n
= ordered_bucket_at(h
, i
->idx
);
620 i
->next_key
= n
->p
.b
.key
;
631 static unsigned hashmap_iterate_in_internal_order(HashmapBase
*h
, Iterator
*i
) {
637 if (i
->idx
== IDX_NIL
)
640 if (i
->idx
== IDX_FIRST
) {
641 /* fast forward to the first occupied bucket */
642 if (h
->has_indirect
) {
643 i
->idx
= skip_free_buckets(h
, h
->indirect
.idx_lowest_entry
);
644 h
->indirect
.idx_lowest_entry
= i
->idx
;
646 i
->idx
= skip_free_buckets(h
, 0);
648 if (i
->idx
== IDX_NIL
)
651 struct hashmap_base_entry
*e
;
655 e
= bucket_at(h
, i
->idx
);
657 * We allow removing the current entry while iterating, but removal may cause
658 * a backward shift. The next entry may thus move one bucket to the left.
659 * To detect when it happens, we remember the key pointer of the entry we were
660 * going to iterate next. If it does not match, there was a backward shift.
662 if (e
->key
!= i
->next_key
)
663 e
= bucket_at(h
, --i
->idx
);
665 assert(e
->key
== i
->next_key
);
669 #if ENABLE_DEBUG_HASHMAP
673 i
->idx
= skip_free_buckets(h
, i
->idx
+ 1);
674 if (i
->idx
!= IDX_NIL
)
675 i
->next_key
= bucket_at(h
, i
->idx
)->key
;
686 static unsigned hashmap_iterate_entry(HashmapBase
*h
, Iterator
*i
) {
692 #if ENABLE_DEBUG_HASHMAP
693 if (i
->idx
== IDX_FIRST
) {
694 i
->put_count
= h
->debug
.put_count
;
695 i
->rem_count
= h
->debug
.rem_count
;
697 /* While iterating, must not add any new entries */
698 assert(i
->put_count
== h
->debug
.put_count
);
699 /* ... or remove entries other than the current one */
700 assert(i
->rem_count
== h
->debug
.rem_count
||
701 (i
->rem_count
== h
->debug
.rem_count
- 1 &&
702 i
->prev_idx
== h
->debug
.last_rem_idx
));
703 /* Reset our removals counter */
704 i
->rem_count
= h
->debug
.rem_count
;
708 return h
->type
== HASHMAP_TYPE_ORDERED
? hashmap_iterate_in_insertion_order((OrderedHashmap
*) h
, i
)
709 : hashmap_iterate_in_internal_order(h
, i
);
712 bool _hashmap_iterate(HashmapBase
*h
, Iterator
*i
, void **value
, const void **key
) {
713 struct hashmap_base_entry
*e
;
717 idx
= hashmap_iterate_entry(h
, i
);
718 if (idx
== IDX_NIL
) {
727 e
= bucket_at(h
, idx
);
728 data
= entry_value(h
, e
);
737 #define HASHMAP_FOREACH_IDX(idx, h, i) \
738 for ((i) = ITERATOR_FIRST, (idx) = hashmap_iterate_entry((h), &(i)); \
740 (idx) = hashmap_iterate_entry((h), &(i)))
742 IteratedCache
* _hashmap_iterated_cache_new(HashmapBase
*h
) {
743 IteratedCache
*cache
;
751 cache
= new0(IteratedCache
, 1);
761 static void reset_direct_storage(HashmapBase
*h
) {
762 const struct hashmap_type_info
*hi
= &hashmap_type_info
[h
->type
];
765 assert(!h
->has_indirect
);
767 p
= mempset(h
->direct
.storage
, 0, hi
->entry_size
* hi
->n_direct_buckets
);
768 memset(p
, DIB_RAW_INIT
, sizeof(dib_raw_t
) * hi
->n_direct_buckets
);
771 static void shared_hash_key_initialize(void) {
772 random_bytes(shared_hash_key
, sizeof(shared_hash_key
));
775 static struct HashmapBase
* hashmap_base_new(const struct hash_ops
*hash_ops
, enum HashmapType type HASHMAP_DEBUG_PARAMS
) {
777 const struct hashmap_type_info
*hi
= &hashmap_type_info
[type
];
779 bool use_pool
= mempool_enabled
&& mempool_enabled(); /* mempool_enabled is a weak symbol */
781 h
= use_pool
? mempool_alloc0_tile(hi
->mempool
) : malloc0(hi
->head_size
);
786 h
->from_pool
= use_pool
;
787 h
->hash_ops
= hash_ops
?: &trivial_hash_ops
;
789 if (type
== HASHMAP_TYPE_ORDERED
) {
790 OrderedHashmap
*lh
= (OrderedHashmap
*)h
;
791 lh
->iterate_list_head
= lh
->iterate_list_tail
= IDX_NIL
;
794 reset_direct_storage(h
);
796 static pthread_once_t once
= PTHREAD_ONCE_INIT
;
797 assert_se(pthread_once(&once
, shared_hash_key_initialize
) == 0);
799 #if ENABLE_DEBUG_HASHMAP
800 h
->debug
.func
= func
;
801 h
->debug
.file
= file
;
802 h
->debug
.line
= line
;
803 assert_se(pthread_mutex_lock(&hashmap_debug_list_mutex
) == 0);
804 LIST_PREPEND(debug_list
, hashmap_debug_list
, &h
->debug
);
805 assert_se(pthread_mutex_unlock(&hashmap_debug_list_mutex
) == 0);
811 Hashmap
*_hashmap_new(const struct hash_ops
*hash_ops HASHMAP_DEBUG_PARAMS
) {
812 return (Hashmap
*) hashmap_base_new(hash_ops
, HASHMAP_TYPE_PLAIN HASHMAP_DEBUG_PASS_ARGS
);
815 OrderedHashmap
*_ordered_hashmap_new(const struct hash_ops
*hash_ops HASHMAP_DEBUG_PARAMS
) {
816 return (OrderedHashmap
*) hashmap_base_new(hash_ops
, HASHMAP_TYPE_ORDERED HASHMAP_DEBUG_PASS_ARGS
);
819 Set
*_set_new(const struct hash_ops
*hash_ops HASHMAP_DEBUG_PARAMS
) {
820 return (Set
*) hashmap_base_new(hash_ops
, HASHMAP_TYPE_SET HASHMAP_DEBUG_PASS_ARGS
);
823 static int hashmap_base_ensure_allocated(HashmapBase
**h
, const struct hash_ops
*hash_ops
,
824 enum HashmapType type HASHMAP_DEBUG_PARAMS
) {
832 q
= hashmap_base_new(hash_ops
, type HASHMAP_DEBUG_PASS_ARGS
);
840 int _hashmap_ensure_allocated(Hashmap
**h
, const struct hash_ops
*hash_ops HASHMAP_DEBUG_PARAMS
) {
841 return hashmap_base_ensure_allocated((HashmapBase
**)h
, hash_ops
, HASHMAP_TYPE_PLAIN HASHMAP_DEBUG_PASS_ARGS
);
844 int _ordered_hashmap_ensure_allocated(OrderedHashmap
**h
, const struct hash_ops
*hash_ops HASHMAP_DEBUG_PARAMS
) {
845 return hashmap_base_ensure_allocated((HashmapBase
**)h
, hash_ops
, HASHMAP_TYPE_ORDERED HASHMAP_DEBUG_PASS_ARGS
);
848 int _set_ensure_allocated(Set
**s
, const struct hash_ops
*hash_ops HASHMAP_DEBUG_PARAMS
) {
849 return hashmap_base_ensure_allocated((HashmapBase
**)s
, hash_ops
, HASHMAP_TYPE_SET HASHMAP_DEBUG_PASS_ARGS
);
852 int _hashmap_ensure_put(Hashmap
**h
, const struct hash_ops
*hash_ops
, const void *key
, void *value HASHMAP_DEBUG_PARAMS
) {
855 r
= _hashmap_ensure_allocated(h
, hash_ops HASHMAP_DEBUG_PASS_ARGS
);
859 return hashmap_put(*h
, key
, value
);
862 int _ordered_hashmap_ensure_put(OrderedHashmap
**h
, const struct hash_ops
*hash_ops
, const void *key
, void *value HASHMAP_DEBUG_PARAMS
) {
865 r
= _ordered_hashmap_ensure_allocated(h
, hash_ops HASHMAP_DEBUG_PASS_ARGS
);
869 return ordered_hashmap_put(*h
, key
, value
);
872 static void hashmap_free_no_clear(HashmapBase
*h
) {
873 assert(!h
->has_indirect
);
874 assert(h
->n_direct_entries
== 0);
876 #if ENABLE_DEBUG_HASHMAP
877 assert_se(pthread_mutex_lock(&hashmap_debug_list_mutex
) == 0);
878 LIST_REMOVE(debug_list
, hashmap_debug_list
, &h
->debug
);
879 assert_se(pthread_mutex_unlock(&hashmap_debug_list_mutex
) == 0);
883 /* Ensure that the object didn't get migrated between threads. */
884 assert_se(is_main_thread());
885 mempool_free_tile(hashmap_type_info
[h
->type
].mempool
, h
);
890 HashmapBase
* _hashmap_free(HashmapBase
*h
, free_func_t default_free_key
, free_func_t default_free_value
) {
892 _hashmap_clear(h
, default_free_key
, default_free_value
);
893 hashmap_free_no_clear(h
);
899 void _hashmap_clear(HashmapBase
*h
, free_func_t default_free_key
, free_func_t default_free_value
) {
900 free_func_t free_key
, free_value
;
904 free_key
= h
->hash_ops
->free_key
?: default_free_key
;
905 free_value
= h
->hash_ops
->free_value
?: default_free_value
;
907 if (free_key
|| free_value
) {
909 /* If destructor calls are defined, let's destroy things defensively: let's take the item out of the
910 * hash table, and only then call the destructor functions. If these destructors then try to unregister
911 * themselves from our hash table a second time, the entry is already gone. */
913 while (_hashmap_size(h
) > 0) {
917 v
= _hashmap_first_key_and_value(h
, true, &k
);
927 if (h
->has_indirect
) {
928 free(h
->indirect
.storage
);
929 h
->has_indirect
= false;
932 h
->n_direct_entries
= 0;
933 reset_direct_storage(h
);
935 if (h
->type
== HASHMAP_TYPE_ORDERED
) {
936 OrderedHashmap
*lh
= (OrderedHashmap
*) h
;
937 lh
->iterate_list_head
= lh
->iterate_list_tail
= IDX_NIL
;
943 static int resize_buckets(HashmapBase
*h
, unsigned entries_add
);
946 * Finds an empty bucket to put an entry into, starting the scan at 'idx'.
947 * Performs Robin Hood swaps as it goes. The entry to put must be placed
948 * by the caller into swap slot IDX_PUT.
949 * If used for in-place resizing, may leave a displaced entry in swap slot
950 * IDX_PUT. Caller must rehash it next.
951 * Returns: true if it left a displaced entry to rehash next in IDX_PUT,
954 static bool hashmap_put_robin_hood(HashmapBase
*h
, unsigned idx
,
955 struct swap_entries
*swap
) {
956 dib_raw_t raw_dib
, *dibs
;
957 unsigned dib
, distance
;
959 #if ENABLE_DEBUG_HASHMAP
960 h
->debug
.put_count
++;
963 dibs
= dib_raw_ptr(h
);
965 for (distance
= 0; ; distance
++) {
967 if (IN_SET(raw_dib
, DIB_RAW_FREE
, DIB_RAW_REHASH
)) {
968 if (raw_dib
== DIB_RAW_REHASH
)
969 bucket_move_entry(h
, swap
, idx
, IDX_TMP
);
971 if (h
->has_indirect
&& h
->indirect
.idx_lowest_entry
> idx
)
972 h
->indirect
.idx_lowest_entry
= idx
;
974 bucket_set_dib(h
, idx
, distance
);
975 bucket_move_entry(h
, swap
, IDX_PUT
, idx
);
976 if (raw_dib
== DIB_RAW_REHASH
) {
977 bucket_move_entry(h
, swap
, IDX_TMP
, IDX_PUT
);
984 dib
= bucket_calculate_dib(h
, idx
, raw_dib
);
986 if (dib
< distance
) {
987 /* Found a wealthier entry. Go Robin Hood! */
988 bucket_set_dib(h
, idx
, distance
);
990 /* swap the entries */
991 bucket_move_entry(h
, swap
, idx
, IDX_TMP
);
992 bucket_move_entry(h
, swap
, IDX_PUT
, idx
);
993 bucket_move_entry(h
, swap
, IDX_TMP
, IDX_PUT
);
998 idx
= next_idx(h
, idx
);
1003 * Puts an entry into a hashmap, boldly - no check whether key already exists.
1004 * The caller must place the entry (only its key and value, not link indexes)
1005 * in swap slot IDX_PUT.
1006 * Caller must ensure: the key does not exist yet in the hashmap.
1007 * that resize is not needed if !may_resize.
1008 * Returns: 1 if entry was put successfully.
1009 * -ENOMEM if may_resize==true and resize failed with -ENOMEM.
1010 * Cannot return -ENOMEM if !may_resize.
1012 static int hashmap_base_put_boldly(HashmapBase
*h
, unsigned idx
,
1013 struct swap_entries
*swap
, bool may_resize
) {
1014 struct ordered_hashmap_entry
*new_entry
;
1017 assert(idx
< n_buckets(h
));
1019 new_entry
= bucket_at_swap(swap
, IDX_PUT
);
1022 r
= resize_buckets(h
, 1);
1026 idx
= bucket_hash(h
, new_entry
->p
.b
.key
);
1028 assert(n_entries(h
) < n_buckets(h
));
1030 if (h
->type
== HASHMAP_TYPE_ORDERED
) {
1031 OrderedHashmap
*lh
= (OrderedHashmap
*) h
;
1033 new_entry
->iterate_next
= IDX_NIL
;
1034 new_entry
->iterate_previous
= lh
->iterate_list_tail
;
1036 if (lh
->iterate_list_tail
!= IDX_NIL
) {
1037 struct ordered_hashmap_entry
*old_tail
;
1039 old_tail
= ordered_bucket_at(lh
, lh
->iterate_list_tail
);
1040 assert(old_tail
->iterate_next
== IDX_NIL
);
1041 old_tail
->iterate_next
= IDX_PUT
;
1044 lh
->iterate_list_tail
= IDX_PUT
;
1045 if (lh
->iterate_list_head
== IDX_NIL
)
1046 lh
->iterate_list_head
= IDX_PUT
;
1049 assert_se(hashmap_put_robin_hood(h
, idx
, swap
) == false);
1052 #if ENABLE_DEBUG_HASHMAP
1053 h
->debug
.max_entries
= MAX(h
->debug
.max_entries
, n_entries(h
));
1060 #define hashmap_put_boldly(h, idx, swap, may_resize) \
1061 hashmap_base_put_boldly(HASHMAP_BASE(h), idx, swap, may_resize)
1064 * Returns 0 if resize is not needed.
1065 * 1 if successfully resized.
1066 * -ENOMEM on allocation failure.
1068 static int resize_buckets(HashmapBase
*h
, unsigned entries_add
) {
1069 struct swap_entries swap
;
1071 dib_raw_t
*old_dibs
, *new_dibs
;
1072 const struct hashmap_type_info
*hi
;
1073 unsigned idx
, optimal_idx
;
1074 unsigned old_n_buckets
, new_n_buckets
, n_rehashed
, new_n_entries
;
1080 hi
= &hashmap_type_info
[h
->type
];
1081 new_n_entries
= n_entries(h
) + entries_add
;
1084 if (_unlikely_(new_n_entries
< entries_add
))
1087 /* For direct storage we allow 100% load, because it's tiny. */
1088 if (!h
->has_indirect
&& new_n_entries
<= hi
->n_direct_buckets
)
1092 * Load factor = n/m = 1 - (1/INV_KEEP_FREE).
1093 * From it follows: m = n + n/(INV_KEEP_FREE - 1)
1095 new_n_buckets
= new_n_entries
+ new_n_entries
/ (INV_KEEP_FREE
- 1);
1097 if (_unlikely_(new_n_buckets
< new_n_entries
))
1100 if (_unlikely_(new_n_buckets
> UINT_MAX
/ (hi
->entry_size
+ sizeof(dib_raw_t
))))
1103 old_n_buckets
= n_buckets(h
);
1105 if (_likely_(new_n_buckets
<= old_n_buckets
))
1108 new_shift
= log2u_round_up(MAX(
1109 new_n_buckets
* (hi
->entry_size
+ sizeof(dib_raw_t
)),
1110 2 * sizeof(struct direct_storage
)));
1112 /* Realloc storage (buckets and DIB array). */
1113 new_storage
= realloc(h
->has_indirect
? h
->indirect
.storage
: NULL
,
1118 /* Must upgrade direct to indirect storage. */
1119 if (!h
->has_indirect
) {
1120 memcpy(new_storage
, h
->direct
.storage
,
1121 old_n_buckets
* (hi
->entry_size
+ sizeof(dib_raw_t
)));
1122 h
->indirect
.n_entries
= h
->n_direct_entries
;
1123 h
->indirect
.idx_lowest_entry
= 0;
1124 h
->n_direct_entries
= 0;
1127 /* Get a new hash key. If we've just upgraded to indirect storage,
1128 * allow reusing a previously generated key. It's still a different key
1129 * from the shared one that we used for direct storage. */
1130 get_hash_key(h
->indirect
.hash_key
, !h
->has_indirect
);
1132 h
->has_indirect
= true;
1133 h
->indirect
.storage
= new_storage
;
1134 h
->indirect
.n_buckets
= (1U << new_shift
) /
1135 (hi
->entry_size
+ sizeof(dib_raw_t
));
1137 old_dibs
= (dib_raw_t
*)((uint8_t*) new_storage
+ hi
->entry_size
* old_n_buckets
);
1138 new_dibs
= dib_raw_ptr(h
);
1141 * Move the DIB array to the new place, replacing valid DIB values with
1142 * DIB_RAW_REHASH to indicate all of the used buckets need rehashing.
1143 * Note: Overlap is not possible, because we have at least doubled the
1144 * number of buckets and dib_raw_t is smaller than any entry type.
1146 for (idx
= 0; idx
< old_n_buckets
; idx
++) {
1147 assert(old_dibs
[idx
] != DIB_RAW_REHASH
);
1148 new_dibs
[idx
] = old_dibs
[idx
] == DIB_RAW_FREE
? DIB_RAW_FREE
1152 /* Zero the area of newly added entries (including the old DIB area) */
1153 memzero(bucket_at(h
, old_n_buckets
),
1154 (n_buckets(h
) - old_n_buckets
) * hi
->entry_size
);
1156 /* The upper half of the new DIB array needs initialization */
1157 memset(&new_dibs
[old_n_buckets
], DIB_RAW_INIT
,
1158 (n_buckets(h
) - old_n_buckets
) * sizeof(dib_raw_t
));
1160 /* Rehash entries that need it */
1162 for (idx
= 0; idx
< old_n_buckets
; idx
++) {
1163 if (new_dibs
[idx
] != DIB_RAW_REHASH
)
1166 optimal_idx
= bucket_hash(h
, bucket_at(h
, idx
)->key
);
1169 * Not much to do if by luck the entry hashes to its current
1170 * location. Just set its DIB.
1172 if (optimal_idx
== idx
) {
1178 new_dibs
[idx
] = DIB_RAW_FREE
;
1179 bucket_move_entry(h
, &swap
, idx
, IDX_PUT
);
1180 /* bucket_move_entry does not clear the source */
1181 memzero(bucket_at(h
, idx
), hi
->entry_size
);
1185 * Find the new bucket for the current entry. This may make
1186 * another entry homeless and load it into IDX_PUT.
1188 rehash_next
= hashmap_put_robin_hood(h
, optimal_idx
, &swap
);
1191 /* Did the current entry displace another one? */
1193 optimal_idx
= bucket_hash(h
, bucket_at_swap(&swap
, IDX_PUT
)->p
.b
.key
);
1194 } while (rehash_next
);
1197 assert_se(n_rehashed
== n_entries(h
));
1203 * Finds an entry with a matching key
1204 * Returns: index of the found entry, or IDX_NIL if not found.
1206 static unsigned base_bucket_scan(HashmapBase
*h
, unsigned idx
, const void *key
) {
1207 struct hashmap_base_entry
*e
;
1208 unsigned dib
, distance
;
1209 dib_raw_t
*dibs
= dib_raw_ptr(h
);
1211 assert(idx
< n_buckets(h
));
1213 for (distance
= 0; ; distance
++) {
1214 if (dibs
[idx
] == DIB_RAW_FREE
)
1217 dib
= bucket_calculate_dib(h
, idx
, dibs
[idx
]);
1221 if (dib
== distance
) {
1222 e
= bucket_at(h
, idx
);
1223 if (h
->hash_ops
->compare(e
->key
, key
) == 0)
1227 idx
= next_idx(h
, idx
);
1230 #define bucket_scan(h, idx, key) base_bucket_scan(HASHMAP_BASE(h), idx, key)
1232 int hashmap_put(Hashmap
*h
, const void *key
, void *value
) {
1233 struct swap_entries swap
;
1234 struct plain_hashmap_entry
*e
;
1239 hash
= bucket_hash(h
, key
);
1240 idx
= bucket_scan(h
, hash
, key
);
1241 if (idx
!= IDX_NIL
) {
1242 e
= plain_bucket_at(h
, idx
);
1243 if (e
->value
== value
)
1248 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
;
1251 return hashmap_put_boldly(h
, hash
, &swap
, true);
1254 int set_put(Set
*s
, const void *key
) {
1255 struct swap_entries swap
;
1256 struct hashmap_base_entry
*e
;
1261 hash
= bucket_hash(s
, key
);
1262 idx
= bucket_scan(s
, hash
, key
);
1266 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
.b
;
1268 return hashmap_put_boldly(s
, hash
, &swap
, true);
1271 int _set_ensure_put(Set
**s
, const struct hash_ops
*hash_ops
, const void *key HASHMAP_DEBUG_PARAMS
) {
1274 r
= _set_ensure_allocated(s
, hash_ops HASHMAP_DEBUG_PASS_ARGS
);
1278 return set_put(*s
, key
);
1281 int _set_ensure_consume(Set
**s
, const struct hash_ops
*hash_ops
, void *key HASHMAP_DEBUG_PARAMS
) {
1284 r
= _set_ensure_put(s
, hash_ops
, key HASHMAP_DEBUG_PASS_ARGS
);
1286 if (hash_ops
&& hash_ops
->free_key
)
1287 hash_ops
->free_key(key
);
1295 int hashmap_replace(Hashmap
*h
, const void *key
, void *value
) {
1296 struct swap_entries swap
;
1297 struct plain_hashmap_entry
*e
;
1302 hash
= bucket_hash(h
, key
);
1303 idx
= bucket_scan(h
, hash
, key
);
1304 if (idx
!= IDX_NIL
) {
1305 e
= plain_bucket_at(h
, idx
);
1306 #if ENABLE_DEBUG_HASHMAP
1307 /* Although the key is equal, the key pointer may have changed,
1308 * and this would break our assumption for iterating. So count
1309 * this operation as incompatible with iteration. */
1310 if (e
->b
.key
!= key
) {
1311 h
->b
.debug
.put_count
++;
1312 h
->b
.debug
.rem_count
++;
1313 h
->b
.debug
.last_rem_idx
= idx
;
1318 hashmap_set_dirty(h
);
1323 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
;
1326 return hashmap_put_boldly(h
, hash
, &swap
, true);
1329 int hashmap_update(Hashmap
*h
, const void *key
, void *value
) {
1330 struct plain_hashmap_entry
*e
;
1335 hash
= bucket_hash(h
, key
);
1336 idx
= bucket_scan(h
, hash
, key
);
1340 e
= plain_bucket_at(h
, idx
);
1342 hashmap_set_dirty(h
);
1347 void* _hashmap_get(HashmapBase
*h
, const void *key
) {
1348 struct hashmap_base_entry
*e
;
1354 hash
= bucket_hash(h
, key
);
1355 idx
= bucket_scan(h
, hash
, key
);
1359 e
= bucket_at(h
, idx
);
1360 return entry_value(h
, e
);
1363 void* hashmap_get2(Hashmap
*h
, const void *key
, void **key2
) {
1364 struct plain_hashmap_entry
*e
;
1370 hash
= bucket_hash(h
, key
);
1371 idx
= bucket_scan(h
, hash
, key
);
1375 e
= plain_bucket_at(h
, idx
);
1377 *key2
= (void*) e
->b
.key
;
1382 bool _hashmap_contains(HashmapBase
*h
, const void *key
) {
1388 hash
= bucket_hash(h
, key
);
1389 return bucket_scan(h
, hash
, key
) != IDX_NIL
;
1392 void* _hashmap_remove(HashmapBase
*h
, const void *key
) {
1393 struct hashmap_base_entry
*e
;
1400 hash
= bucket_hash(h
, key
);
1401 idx
= bucket_scan(h
, hash
, key
);
1405 e
= bucket_at(h
, idx
);
1406 data
= entry_value(h
, e
);
1407 remove_entry(h
, idx
);
1412 void* hashmap_remove2(Hashmap
*h
, const void *key
, void **rkey
) {
1413 struct plain_hashmap_entry
*e
;
1423 hash
= bucket_hash(h
, key
);
1424 idx
= bucket_scan(h
, hash
, key
);
1425 if (idx
== IDX_NIL
) {
1431 e
= plain_bucket_at(h
, idx
);
1434 *rkey
= (void*) e
->b
.key
;
1436 remove_entry(h
, idx
);
1441 int hashmap_remove_and_put(Hashmap
*h
, const void *old_key
, const void *new_key
, void *value
) {
1442 struct swap_entries swap
;
1443 struct plain_hashmap_entry
*e
;
1444 unsigned old_hash
, new_hash
, idx
;
1449 old_hash
= bucket_hash(h
, old_key
);
1450 idx
= bucket_scan(h
, old_hash
, old_key
);
1454 new_hash
= bucket_hash(h
, new_key
);
1455 if (bucket_scan(h
, new_hash
, new_key
) != IDX_NIL
)
1458 remove_entry(h
, idx
);
1460 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
;
1463 assert_se(hashmap_put_boldly(h
, new_hash
, &swap
, false) == 1);
1468 int set_remove_and_put(Set
*s
, const void *old_key
, const void *new_key
) {
1469 struct swap_entries swap
;
1470 struct hashmap_base_entry
*e
;
1471 unsigned old_hash
, new_hash
, idx
;
1476 old_hash
= bucket_hash(s
, old_key
);
1477 idx
= bucket_scan(s
, old_hash
, old_key
);
1481 new_hash
= bucket_hash(s
, new_key
);
1482 if (bucket_scan(s
, new_hash
, new_key
) != IDX_NIL
)
1485 remove_entry(s
, idx
);
1487 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
.b
;
1489 assert_se(hashmap_put_boldly(s
, new_hash
, &swap
, false) == 1);
1494 int hashmap_remove_and_replace(Hashmap
*h
, const void *old_key
, const void *new_key
, void *value
) {
1495 struct swap_entries swap
;
1496 struct plain_hashmap_entry
*e
;
1497 unsigned old_hash
, new_hash
, idx_old
, idx_new
;
1502 old_hash
= bucket_hash(h
, old_key
);
1503 idx_old
= bucket_scan(h
, old_hash
, old_key
);
1504 if (idx_old
== IDX_NIL
)
1507 old_key
= bucket_at(HASHMAP_BASE(h
), idx_old
)->key
;
1509 new_hash
= bucket_hash(h
, new_key
);
1510 idx_new
= bucket_scan(h
, new_hash
, new_key
);
1511 if (idx_new
!= IDX_NIL
)
1512 if (idx_old
!= idx_new
) {
1513 remove_entry(h
, idx_new
);
1514 /* Compensate for a possible backward shift. */
1515 if (old_key
!= bucket_at(HASHMAP_BASE(h
), idx_old
)->key
)
1516 idx_old
= prev_idx(HASHMAP_BASE(h
), idx_old
);
1517 assert(old_key
== bucket_at(HASHMAP_BASE(h
), idx_old
)->key
);
1520 remove_entry(h
, idx_old
);
1522 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
;
1525 assert_se(hashmap_put_boldly(h
, new_hash
, &swap
, false) == 1);
1530 void* _hashmap_remove_value(HashmapBase
*h
, const void *key
, void *value
) {
1531 struct hashmap_base_entry
*e
;
1537 hash
= bucket_hash(h
, key
);
1538 idx
= bucket_scan(h
, hash
, key
);
1542 e
= bucket_at(h
, idx
);
1543 if (entry_value(h
, e
) != value
)
1546 remove_entry(h
, idx
);
1551 static unsigned find_first_entry(HashmapBase
*h
) {
1552 Iterator i
= ITERATOR_FIRST
;
1554 if (!h
|| !n_entries(h
))
1557 return hashmap_iterate_entry(h
, &i
);
1560 void* _hashmap_first_key_and_value(HashmapBase
*h
, bool remove
, void **ret_key
) {
1561 struct hashmap_base_entry
*e
;
1565 idx
= find_first_entry(h
);
1566 if (idx
== IDX_NIL
) {
1572 e
= bucket_at(h
, idx
);
1573 key
= (void*) e
->key
;
1574 data
= entry_value(h
, e
);
1577 remove_entry(h
, idx
);
1585 unsigned _hashmap_size(HashmapBase
*h
) {
1589 return n_entries(h
);
1592 unsigned _hashmap_buckets(HashmapBase
*h
) {
1596 return n_buckets(h
);
1599 int _hashmap_merge(Hashmap
*h
, Hashmap
*other
) {
1605 HASHMAP_FOREACH_IDX(idx
, HASHMAP_BASE(other
), i
) {
1606 struct plain_hashmap_entry
*pe
= plain_bucket_at(other
, idx
);
1609 r
= hashmap_put(h
, pe
->b
.key
, pe
->value
);
1610 if (r
< 0 && r
!= -EEXIST
)
1617 int set_merge(Set
*s
, Set
*other
) {
1623 HASHMAP_FOREACH_IDX(idx
, HASHMAP_BASE(other
), i
) {
1624 struct set_entry
*se
= set_bucket_at(other
, idx
);
1627 r
= set_put(s
, se
->b
.key
);
1635 int _hashmap_reserve(HashmapBase
*h
, unsigned entries_add
) {
1640 r
= resize_buckets(h
, entries_add
);
1648 * The same as hashmap_merge(), but every new item from other is moved to h.
1649 * Keys already in h are skipped and stay in other.
1650 * Returns: 0 on success.
1651 * -ENOMEM on alloc failure, in which case no move has been done.
1653 int _hashmap_move(HashmapBase
*h
, HashmapBase
*other
) {
1654 struct swap_entries swap
;
1655 struct hashmap_base_entry
*e
, *n
;
1665 assert(other
->type
== h
->type
);
1668 * This reserves buckets for the worst case, where none of other's
1669 * entries are yet present in h. This is preferable to risking
1670 * an allocation failure in the middle of the moving and having to
1671 * rollback or return a partial result.
1673 r
= resize_buckets(h
, n_entries(other
));
1677 HASHMAP_FOREACH_IDX(idx
, other
, i
) {
1680 e
= bucket_at(other
, idx
);
1681 h_hash
= bucket_hash(h
, e
->key
);
1682 if (bucket_scan(h
, h_hash
, e
->key
) != IDX_NIL
)
1685 n
= &bucket_at_swap(&swap
, IDX_PUT
)->p
.b
;
1687 if (h
->type
!= HASHMAP_TYPE_SET
)
1688 ((struct plain_hashmap_entry
*) n
)->value
=
1689 ((struct plain_hashmap_entry
*) e
)->value
;
1690 assert_se(hashmap_put_boldly(h
, h_hash
, &swap
, false) == 1);
1692 remove_entry(other
, idx
);
1698 int _hashmap_move_one(HashmapBase
*h
, HashmapBase
*other
, const void *key
) {
1699 struct swap_entries swap
;
1700 unsigned h_hash
, other_hash
, idx
;
1701 struct hashmap_base_entry
*e
, *n
;
1706 h_hash
= bucket_hash(h
, key
);
1707 if (bucket_scan(h
, h_hash
, key
) != IDX_NIL
)
1713 assert(other
->type
== h
->type
);
1715 other_hash
= bucket_hash(other
, key
);
1716 idx
= bucket_scan(other
, other_hash
, key
);
1720 e
= bucket_at(other
, idx
);
1722 n
= &bucket_at_swap(&swap
, IDX_PUT
)->p
.b
;
1724 if (h
->type
!= HASHMAP_TYPE_SET
)
1725 ((struct plain_hashmap_entry
*) n
)->value
=
1726 ((struct plain_hashmap_entry
*) e
)->value
;
1727 r
= hashmap_put_boldly(h
, h_hash
, &swap
, true);
1731 remove_entry(other
, idx
);
1735 HashmapBase
* _hashmap_copy(HashmapBase
*h HASHMAP_DEBUG_PARAMS
) {
1741 copy
= hashmap_base_new(h
->hash_ops
, h
->type HASHMAP_DEBUG_PASS_ARGS
);
1746 case HASHMAP_TYPE_PLAIN
:
1747 case HASHMAP_TYPE_ORDERED
:
1748 r
= hashmap_merge((Hashmap
*)copy
, (Hashmap
*)h
);
1750 case HASHMAP_TYPE_SET
:
1751 r
= set_merge((Set
*)copy
, (Set
*)h
);
1754 assert_not_reached();
1758 return _hashmap_free(copy
, NULL
, NULL
);
1763 char** _hashmap_get_strv(HashmapBase
*h
) {
1769 return new0(char*, 1);
1771 sv
= new(char*, n_entries(h
)+1);
1776 HASHMAP_FOREACH_IDX(idx
, h
, i
)
1777 sv
[n
++] = entry_value(h
, bucket_at(h
, idx
));
1783 void* ordered_hashmap_next(OrderedHashmap
*h
, const void *key
) {
1784 struct ordered_hashmap_entry
*e
;
1790 hash
= bucket_hash(h
, key
);
1791 idx
= bucket_scan(h
, hash
, key
);
1795 e
= ordered_bucket_at(h
, idx
);
1796 if (e
->iterate_next
== IDX_NIL
)
1798 return ordered_bucket_at(h
, e
->iterate_next
)->p
.value
;
1801 int set_consume(Set
*s
, void *value
) {
1807 r
= set_put(s
, value
);
1814 int _hashmap_put_strdup_full(Hashmap
**h
, const struct hash_ops
*hash_ops
, const char *k
, const char *v HASHMAP_DEBUG_PARAMS
) {
1817 r
= _hashmap_ensure_allocated(h
, hash_ops HASHMAP_DEBUG_PASS_ARGS
);
1821 _cleanup_free_
char *kdup
= NULL
, *vdup
= NULL
;
1833 r
= hashmap_put(*h
, kdup
, vdup
);
1835 if (r
== -EEXIST
&& streq_ptr(v
, hashmap_get(*h
, kdup
)))
1840 /* 0 with non-null vdup would mean vdup is already in the hashmap, which cannot be */
1841 assert(vdup
== NULL
|| r
> 0);
1848 int _set_put_strndup_full(Set
**s
, const struct hash_ops
*hash_ops
, const char *p
, size_t n HASHMAP_DEBUG_PARAMS
) {
1855 r
= _set_ensure_allocated(s
, hash_ops HASHMAP_DEBUG_PASS_ARGS
);
1859 if (n
== SIZE_MAX
) {
1860 if (set_contains(*s
, (char*) p
))
1869 return set_consume(*s
, c
);
1872 int _set_put_strdupv_full(Set
**s
, const struct hash_ops
*hash_ops
, char **l HASHMAP_DEBUG_PARAMS
) {
1877 STRV_FOREACH(i
, l
) {
1878 r
= _set_put_strndup_full(s
, hash_ops
, *i
, SIZE_MAX HASHMAP_DEBUG_PASS_ARGS
);
1888 int set_put_strsplit(Set
*s
, const char *v
, const char *separators
, ExtractFlags flags
) {
1889 const char *p
= ASSERT_PTR(v
);
1897 r
= extract_first_word(&p
, &word
, separators
, flags
);
1901 r
= set_consume(s
, word
);
1907 /* expand the cachemem if needed, return true if newly (re)activated. */
1908 static int cachemem_maintain(CacheMem
*mem
, size_t size
) {
1911 if (!GREEDY_REALLOC(mem
->ptr
, size
)) {
1924 int iterated_cache_get(IteratedCache
*cache
, const void ***res_keys
, const void ***res_values
, unsigned *res_n_entries
) {
1925 bool sync_keys
= false, sync_values
= false;
1930 assert(cache
->hashmap
);
1932 size
= n_entries(cache
->hashmap
);
1935 r
= cachemem_maintain(&cache
->keys
, size
);
1941 cache
->keys
.active
= false;
1944 r
= cachemem_maintain(&cache
->values
, size
);
1950 cache
->values
.active
= false;
1952 if (cache
->hashmap
->dirty
) {
1953 if (cache
->keys
.active
)
1955 if (cache
->values
.active
)
1958 cache
->hashmap
->dirty
= false;
1961 if (sync_keys
|| sync_values
) {
1966 HASHMAP_FOREACH_IDX(idx
, cache
->hashmap
, iter
) {
1967 struct hashmap_base_entry
*e
;
1969 e
= bucket_at(cache
->hashmap
, idx
);
1972 cache
->keys
.ptr
[i
] = e
->key
;
1974 cache
->values
.ptr
[i
] = entry_value(cache
->hashmap
, e
);
1980 *res_keys
= cache
->keys
.ptr
;
1982 *res_values
= cache
->values
.ptr
;
1984 *res_n_entries
= size
;
1989 IteratedCache
* iterated_cache_free(IteratedCache
*cache
) {
1991 free(cache
->keys
.ptr
);
1992 free(cache
->values
.ptr
);
1995 return mfree(cache
);
1998 int set_strjoin(Set
*s
, const char *separator
, bool wrap_with_separator
, char **ret
) {
1999 _cleanup_free_
char *str
= NULL
;
2000 size_t separator_len
, len
= 0;
2006 if (set_isempty(s
)) {
2011 separator_len
= strlen_ptr(separator
);
2013 if (separator_len
== 0)
2014 wrap_with_separator
= false;
2016 first
= !wrap_with_separator
;
2018 SET_FOREACH(value
, s
) {
2019 size_t l
= strlen_ptr(value
);
2024 if (!GREEDY_REALLOC(str
, len
+ l
+ (first
? 0 : separator_len
) + (wrap_with_separator
? separator_len
: 0) + 1))
2027 if (separator_len
> 0 && !first
) {
2028 memcpy(str
+ len
, separator
, separator_len
);
2029 len
+= separator_len
;
2032 memcpy(str
+ len
, value
, l
);
2037 if (wrap_with_separator
) {
2038 memcpy(str
+ len
, separator
, separator_len
);
2039 len
+= separator_len
;
2044 *ret
= TAKE_PTR(str
);
2048 bool set_equal(Set
*a
, Set
*b
) {
2051 /* Checks whether each entry of 'a' is also in 'b' and vice versa, i.e. the two sets contain the same
2057 if (set_isempty(a
) && set_isempty(b
))
2060 if (set_size(a
) != set_size(b
)) /* Cheap check that hopefully catches a lot of inequality cases
2065 if (!set_contains(b
, p
))
2068 /* If we have the same hashops, then we don't need to check things backwards given we compared the
2069 * size and that all of a is in b. */
2070 if (a
->b
.hash_ops
== b
->b
.hash_ops
)
2074 if (!set_contains(a
, p
))
2080 static bool set_fnmatch_one(Set
*patterns
, const char *needle
) {
2085 /* Any failure of fnmatch() is treated as equivalent to FNM_NOMATCH, i.e. as non-matching pattern */
2087 SET_FOREACH(p
, patterns
)
2088 if (fnmatch(p
, needle
, 0) == 0)
2094 bool set_fnmatch(Set
*include_patterns
, Set
*exclude_patterns
, const char *needle
) {
2097 if (set_fnmatch_one(exclude_patterns
, needle
))
2100 if (set_isempty(include_patterns
))
2103 return set_fnmatch_one(include_patterns
, needle
);