1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
9 #include "alloc-util.h"
13 #include "memory-util.h"
15 #include "missing_syscall.h"
16 #include "process-util.h"
17 #include "random-util.h"
19 #include "siphash24.h"
20 #include "string-util.h"
23 #if ENABLE_DEBUG_HASHMAP
28 * Implementation of hashmaps.
30 * - uses less RAM compared to closed addressing (chaining), because
31 * our entries are small (especially in Sets, which tend to contain
32 * the majority of entries in systemd).
33 * Collision resolution: Robin Hood
34 * - tends to equalize displacement of entries from their optimal buckets.
35 * Probe sequence: linear
36 * - though theoretically worse than random probing/uniform hashing/double
37 * hashing, it is good for cache locality.
40 * Celis, P. 1986. Robin Hood Hashing.
41 * Ph.D. Dissertation. University of Waterloo, Waterloo, Ont., Canada, Canada.
42 * https://cs.uwaterloo.ca/research/tr/1986/CS-86-14.pdf
43 * - The results are derived for random probing. Suggests deletion with
44 * tombstones and two mean-centered search methods. None of that works
45 * well for linear probing.
47 * Janson, S. 2005. Individual displacements for linear probing hashing with different insertion policies.
48 * ACM Trans. Algorithms 1, 2 (October 2005), 177-213.
49 * DOI=10.1145/1103963.1103964 http://doi.acm.org/10.1145/1103963.1103964
50 * http://www.math.uu.se/~svante/papers/sj157.pdf
51 * - Applies to Robin Hood with linear probing. Contains remarks on
52 * the unsuitability of mean-centered search with linear probing.
54 * Viola, A. 2005. Exact distribution of individual displacements in linear probing hashing.
55 * ACM Trans. Algorithms 1, 2 (October 2005), 214-242.
56 * DOI=10.1145/1103963.1103965 http://doi.acm.org/10.1145/1103963.1103965
57 * - Similar to Janson. Note that Viola writes about C_{m,n} (number of probes
58 * in a successful search), and Janson writes about displacement. C = d + 1.
60 * Goossaert, E. 2013. Robin Hood hashing: backward shift deletion.
61 * http://codecapsule.com/2013/11/17/robin-hood-hashing-backward-shift-deletion/
62 * - Explanation of backward shift deletion with pictures.
64 * Khuong, P. 2013. The Other Robin Hood Hashing.
65 * http://www.pvk.ca/Blog/2013/11/26/the-other-robin-hood-hashing/
66 * - Short summary of random vs. linear probing, and tombstones vs. backward shift.
70 * XXX Ideas for improvement:
71 * For unordered hashmaps, randomize iteration order, similarly to Perl:
72 * http://blog.booking.com/hardening-perls-hash-function.html
75 /* INV_KEEP_FREE = 1 / (1 - max_load_factor)
76 * e.g. 1 / (1 - 0.8) = 5 ... keep one fifth of the buckets free. */
77 #define INV_KEEP_FREE 5U
79 /* Fields common to entries of all hashmap/set types */
80 struct hashmap_base_entry
{
84 /* Entry types for specific hashmap/set types
85 * hashmap_base_entry must be at the beginning of each entry struct. */
87 struct plain_hashmap_entry
{
88 struct hashmap_base_entry b
;
92 struct ordered_hashmap_entry
{
93 struct plain_hashmap_entry p
;
94 unsigned iterate_next
, iterate_previous
;
98 struct hashmap_base_entry b
;
101 /* In several functions it is advantageous to have the hash table extended
102 * virtually by a couple of additional buckets. We reserve special index values
103 * for these "swap" buckets. */
104 #define _IDX_SWAP_BEGIN (UINT_MAX - 3)
105 #define IDX_PUT (_IDX_SWAP_BEGIN + 0)
106 #define IDX_TMP (_IDX_SWAP_BEGIN + 1)
107 #define _IDX_SWAP_END (_IDX_SWAP_BEGIN + 2)
109 #define IDX_FIRST (UINT_MAX - 1) /* special index for freshly initialized iterators */
110 #define IDX_NIL UINT_MAX /* special index value meaning "none" or "end" */
112 assert_cc(IDX_FIRST
== _IDX_SWAP_END
);
113 assert_cc(IDX_FIRST
== _IDX_ITERATOR_FIRST
);
115 /* Storage space for the "swap" buckets.
116 * All entry types can fit into an ordered_hashmap_entry. */
117 struct swap_entries
{
118 struct ordered_hashmap_entry e
[_IDX_SWAP_END
- _IDX_SWAP_BEGIN
];
121 /* Distance from Initial Bucket */
122 typedef uint8_t dib_raw_t
;
123 #define DIB_RAW_OVERFLOW ((dib_raw_t)0xfdU) /* indicates DIB value is greater than representable */
124 #define DIB_RAW_REHASH ((dib_raw_t)0xfeU) /* entry yet to be rehashed during in-place resize */
125 #define DIB_RAW_FREE ((dib_raw_t)0xffU) /* a free bucket */
126 #define DIB_RAW_INIT ((char)DIB_RAW_FREE) /* a byte to memset a DIB store with when initializing */
128 #define DIB_FREE UINT_MAX
130 #if ENABLE_DEBUG_HASHMAP
131 struct hashmap_debug_info
{
132 LIST_FIELDS(struct hashmap_debug_info
, debug_list
);
133 unsigned max_entries
; /* high watermark of n_entries */
135 /* who allocated this hashmap */
140 /* fields to detect modification while iterating */
141 unsigned put_count
; /* counts puts into the hashmap */
142 unsigned rem_count
; /* counts removals from hashmap */
143 unsigned last_rem_idx
; /* remembers last removal index */
146 /* Tracks all existing hashmaps. Get at it from gdb. See sd_dump_hashmaps.py */
147 static LIST_HEAD(struct hashmap_debug_info
, hashmap_debug_list
);
148 static pthread_mutex_t hashmap_debug_list_mutex
= PTHREAD_MUTEX_INITIALIZER
;
153 HASHMAP_TYPE_ORDERED
,
158 struct _packed_ indirect_storage
{
159 void *storage
; /* where buckets and DIBs are stored */
160 uint8_t hash_key
[HASH_KEY_SIZE
]; /* hash key; changes during resize */
162 unsigned n_entries
; /* number of stored entries */
163 unsigned n_buckets
; /* number of buckets */
165 unsigned idx_lowest_entry
; /* Index below which all buckets are free.
166 Makes "while (hashmap_steal_first())" loops
167 O(n) instead of O(n^2) for unordered hashmaps. */
168 uint8_t _pad
[3]; /* padding for the whole HashmapBase */
169 /* The bitfields in HashmapBase complete the alignment of the whole thing. */
172 struct direct_storage
{
173 /* This gives us 39 bytes on 64bit, or 35 bytes on 32bit.
174 * That's room for 4 set_entries + 4 DIB bytes + 3 unused bytes on 64bit,
175 * or 7 set_entries + 7 DIB bytes + 0 unused bytes on 32bit. */
176 uint8_t storage
[sizeof(struct indirect_storage
)];
179 #define DIRECT_BUCKETS(entry_t) \
180 (sizeof(struct direct_storage) / (sizeof(entry_t) + sizeof(dib_raw_t)))
182 /* We should be able to store at least one entry directly. */
183 assert_cc(DIRECT_BUCKETS(struct ordered_hashmap_entry
) >= 1);
185 /* We have 3 bits for n_direct_entries. */
186 assert_cc(DIRECT_BUCKETS(struct set_entry
) < (1 << 3));
188 /* Hashmaps with directly stored entries all use this shared hash key.
189 * It's no big deal if the key is guessed, because there can be only
190 * a handful of directly stored entries in a hashmap. When a hashmap
191 * outgrows direct storage, it gets its own key for indirect storage. */
192 static uint8_t shared_hash_key
[HASH_KEY_SIZE
];
194 /* Fields that all hashmap/set types must have */
196 const struct hash_ops
*hash_ops
; /* hash and compare ops to use */
199 struct indirect_storage indirect
; /* if has_indirect */
200 struct direct_storage direct
; /* if !has_indirect */
203 enum HashmapType type
:2; /* HASHMAP_TYPE_* */
204 bool has_indirect
:1; /* whether indirect storage is used */
205 unsigned n_direct_entries
:3; /* Number of entries in direct storage.
206 * Only valid if !has_indirect. */
207 bool from_pool
:1; /* whether was allocated from mempool */
208 bool dirty
:1; /* whether dirtied since last iterated_cache_get() */
209 bool cached
:1; /* whether this hashmap is being cached */
211 #if ENABLE_DEBUG_HASHMAP
212 struct hashmap_debug_info debug
;
216 /* Specific hash types
217 * HashmapBase must be at the beginning of each hashmap struct. */
220 struct HashmapBase b
;
223 struct OrderedHashmap
{
224 struct HashmapBase b
;
225 unsigned iterate_list_head
, iterate_list_tail
;
229 struct HashmapBase b
;
232 typedef struct CacheMem
{
238 struct IteratedCache
{
239 HashmapBase
*hashmap
;
240 CacheMem keys
, values
;
243 DEFINE_MEMPOOL(hashmap_pool
, Hashmap
, 8);
244 DEFINE_MEMPOOL(ordered_hashmap_pool
, OrderedHashmap
, 8);
245 /* No need for a separate Set pool */
246 assert_cc(sizeof(Hashmap
) == sizeof(Set
));
248 struct hashmap_type_info
{
251 struct mempool
*mempool
;
252 unsigned n_direct_buckets
;
255 static _used_
const struct hashmap_type_info hashmap_type_info
[_HASHMAP_TYPE_MAX
] = {
256 [HASHMAP_TYPE_PLAIN
] = {
257 .head_size
= sizeof(Hashmap
),
258 .entry_size
= sizeof(struct plain_hashmap_entry
),
259 .mempool
= &hashmap_pool
,
260 .n_direct_buckets
= DIRECT_BUCKETS(struct plain_hashmap_entry
),
262 [HASHMAP_TYPE_ORDERED
] = {
263 .head_size
= sizeof(OrderedHashmap
),
264 .entry_size
= sizeof(struct ordered_hashmap_entry
),
265 .mempool
= &ordered_hashmap_pool
,
266 .n_direct_buckets
= DIRECT_BUCKETS(struct ordered_hashmap_entry
),
268 [HASHMAP_TYPE_SET
] = {
269 .head_size
= sizeof(Set
),
270 .entry_size
= sizeof(struct set_entry
),
271 .mempool
= &hashmap_pool
,
272 .n_direct_buckets
= DIRECT_BUCKETS(struct set_entry
),
277 _destructor_
static void cleanup_pools(void) {
278 _cleanup_free_
char *t
= NULL
;
281 /* Be nice to valgrind */
283 /* The pool is only allocated by the main thread, but the memory can
284 * be passed to other threads. Let's clean up if we are the main thread
285 * and no other threads are live. */
286 /* We build our own is_main_thread() here, which doesn't use C11
287 * TLS based caching of the result. That's because valgrind apparently
288 * doesn't like malloc() (which C11 TLS internally uses) to be called
289 * from a GCC destructors. */
290 if (getpid() != gettid())
293 r
= get_proc_field("/proc/self/status", "Threads", WHITESPACE
, &t
);
294 if (r
< 0 || !streq(t
, "1"))
297 mempool_drop(&hashmap_pool
);
298 mempool_drop(&ordered_hashmap_pool
);
302 static unsigned n_buckets(HashmapBase
*h
) {
303 return h
->has_indirect
? h
->indirect
.n_buckets
304 : hashmap_type_info
[h
->type
].n_direct_buckets
;
307 static unsigned n_entries(HashmapBase
*h
) {
308 return h
->has_indirect
? h
->indirect
.n_entries
309 : h
->n_direct_entries
;
312 static void n_entries_inc(HashmapBase
*h
) {
314 h
->indirect
.n_entries
++;
316 h
->n_direct_entries
++;
319 static void n_entries_dec(HashmapBase
*h
) {
321 h
->indirect
.n_entries
--;
323 h
->n_direct_entries
--;
326 static void* storage_ptr(HashmapBase
*h
) {
327 return h
->has_indirect
? h
->indirect
.storage
331 static uint8_t* hash_key(HashmapBase
*h
) {
332 return h
->has_indirect
? h
->indirect
.hash_key
336 static unsigned base_bucket_hash(HashmapBase
*h
, const void *p
) {
337 struct siphash state
;
340 siphash24_init(&state
, hash_key(h
));
342 h
->hash_ops
->hash(p
, &state
);
344 hash
= siphash24_finalize(&state
);
346 return (unsigned) (hash
% n_buckets(h
));
348 #define bucket_hash(h, p) base_bucket_hash(HASHMAP_BASE(h), p)
350 static void base_set_dirty(HashmapBase
*h
) {
353 #define hashmap_set_dirty(h) base_set_dirty(HASHMAP_BASE(h))
355 static void get_hash_key(uint8_t hash_key
[HASH_KEY_SIZE
], bool reuse_is_ok
) {
356 static uint8_t current
[HASH_KEY_SIZE
];
357 static bool current_initialized
= false;
359 /* Returns a hash function key to use. In order to keep things
360 * fast we will not generate a new key each time we allocate a
361 * new hash table. Instead, we'll just reuse the most recently
362 * generated one, except if we never generated one or when we
363 * are rehashing an entire hash table because we reached a
366 if (!current_initialized
|| !reuse_is_ok
) {
367 random_bytes(current
, sizeof(current
));
368 current_initialized
= true;
371 memcpy(hash_key
, current
, sizeof(current
));
374 static struct hashmap_base_entry
* bucket_at(HashmapBase
*h
, unsigned idx
) {
375 return (struct hashmap_base_entry
*)
376 ((uint8_t*) storage_ptr(h
) + idx
* hashmap_type_info
[h
->type
].entry_size
);
379 static struct plain_hashmap_entry
* plain_bucket_at(Hashmap
*h
, unsigned idx
) {
380 return (struct plain_hashmap_entry
*) bucket_at(HASHMAP_BASE(h
), idx
);
383 static struct ordered_hashmap_entry
* ordered_bucket_at(OrderedHashmap
*h
, unsigned idx
) {
384 return (struct ordered_hashmap_entry
*) bucket_at(HASHMAP_BASE(h
), idx
);
387 static struct set_entry
*set_bucket_at(Set
*h
, unsigned idx
) {
388 return (struct set_entry
*) bucket_at(HASHMAP_BASE(h
), idx
);
391 static struct ordered_hashmap_entry
* bucket_at_swap(struct swap_entries
*swap
, unsigned idx
) {
392 return &swap
->e
[idx
- _IDX_SWAP_BEGIN
];
395 /* Returns a pointer to the bucket at index idx.
396 * Understands real indexes and swap indexes, hence "_virtual". */
397 static struct hashmap_base_entry
* bucket_at_virtual(HashmapBase
*h
, struct swap_entries
*swap
,
399 if (idx
< _IDX_SWAP_BEGIN
)
400 return bucket_at(h
, idx
);
402 if (idx
< _IDX_SWAP_END
)
403 return &bucket_at_swap(swap
, idx
)->p
.b
;
405 assert_not_reached();
408 static dib_raw_t
* dib_raw_ptr(HashmapBase
*h
) {
410 ((uint8_t*) storage_ptr(h
) + hashmap_type_info
[h
->type
].entry_size
* n_buckets(h
));
413 static unsigned bucket_distance(HashmapBase
*h
, unsigned idx
, unsigned from
) {
414 return idx
>= from
? idx
- from
415 : n_buckets(h
) + idx
- from
;
418 static unsigned bucket_calculate_dib(HashmapBase
*h
, unsigned idx
, dib_raw_t raw_dib
) {
419 unsigned initial_bucket
;
421 if (raw_dib
== DIB_RAW_FREE
)
424 if (_likely_(raw_dib
< DIB_RAW_OVERFLOW
))
428 * Having an overflow DIB value is very unlikely. The hash function
429 * would have to be bad. For example, in a table of size 2^24 filled
430 * to load factor 0.9 the maximum observed DIB is only about 60.
431 * In theory (assuming I used Maxima correctly), for an infinite size
432 * hash table with load factor 0.8 the probability of a given entry
433 * having DIB > 40 is 1.9e-8.
434 * This returns the correct DIB value by recomputing the hash value in
435 * the unlikely case. XXX Hitting this case could be a hint to rehash.
437 initial_bucket
= bucket_hash(h
, bucket_at(h
, idx
)->key
);
438 return bucket_distance(h
, idx
, initial_bucket
);
441 static void bucket_set_dib(HashmapBase
*h
, unsigned idx
, unsigned dib
) {
442 dib_raw_ptr(h
)[idx
] = dib
!= DIB_FREE
? MIN(dib
, DIB_RAW_OVERFLOW
) : DIB_RAW_FREE
;
445 static unsigned skip_free_buckets(HashmapBase
*h
, unsigned idx
) {
448 dibs
= dib_raw_ptr(h
);
450 for ( ; idx
< n_buckets(h
); idx
++)
451 if (dibs
[idx
] != DIB_RAW_FREE
)
457 static void bucket_mark_free(HashmapBase
*h
, unsigned idx
) {
458 memzero(bucket_at(h
, idx
), hashmap_type_info
[h
->type
].entry_size
);
459 bucket_set_dib(h
, idx
, DIB_FREE
);
462 static void bucket_move_entry(HashmapBase
*h
, struct swap_entries
*swap
,
463 unsigned from
, unsigned to
) {
464 struct hashmap_base_entry
*e_from
, *e_to
;
468 e_from
= bucket_at_virtual(h
, swap
, from
);
469 e_to
= bucket_at_virtual(h
, swap
, to
);
471 memcpy(e_to
, e_from
, hashmap_type_info
[h
->type
].entry_size
);
473 if (h
->type
== HASHMAP_TYPE_ORDERED
) {
474 OrderedHashmap
*lh
= (OrderedHashmap
*) h
;
475 struct ordered_hashmap_entry
*le
, *le_to
;
477 le_to
= (struct ordered_hashmap_entry
*) e_to
;
479 if (le_to
->iterate_next
!= IDX_NIL
) {
480 le
= (struct ordered_hashmap_entry
*)
481 bucket_at_virtual(h
, swap
, le_to
->iterate_next
);
482 le
->iterate_previous
= to
;
485 if (le_to
->iterate_previous
!= IDX_NIL
) {
486 le
= (struct ordered_hashmap_entry
*)
487 bucket_at_virtual(h
, swap
, le_to
->iterate_previous
);
488 le
->iterate_next
= to
;
491 if (lh
->iterate_list_head
== from
)
492 lh
->iterate_list_head
= to
;
493 if (lh
->iterate_list_tail
== from
)
494 lh
->iterate_list_tail
= to
;
498 static unsigned next_idx(HashmapBase
*h
, unsigned idx
) {
499 return (idx
+ 1U) % n_buckets(h
);
502 static unsigned prev_idx(HashmapBase
*h
, unsigned idx
) {
503 return (n_buckets(h
) + idx
- 1U) % n_buckets(h
);
506 static void* entry_value(HashmapBase
*h
, struct hashmap_base_entry
*e
) {
509 case HASHMAP_TYPE_PLAIN
:
510 case HASHMAP_TYPE_ORDERED
:
511 return ((struct plain_hashmap_entry
*)e
)->value
;
513 case HASHMAP_TYPE_SET
:
514 return (void*) e
->key
;
517 assert_not_reached();
521 static void base_remove_entry(HashmapBase
*h
, unsigned idx
) {
522 unsigned left
, right
, prev
, dib
;
523 dib_raw_t raw_dib
, *dibs
;
525 dibs
= dib_raw_ptr(h
);
526 assert(dibs
[idx
] != DIB_RAW_FREE
);
528 #if ENABLE_DEBUG_HASHMAP
529 h
->debug
.rem_count
++;
530 h
->debug
.last_rem_idx
= idx
;
534 /* Find the stop bucket ("right"). It is either free or has DIB == 0. */
535 for (right
= next_idx(h
, left
); ; right
= next_idx(h
, right
)) {
536 raw_dib
= dibs
[right
];
537 if (IN_SET(raw_dib
, 0, DIB_RAW_FREE
))
540 /* The buckets are not supposed to be all occupied and with DIB > 0.
541 * That would mean we could make everyone better off by shifting them
542 * backward. This scenario is impossible. */
543 assert(left
!= right
);
546 if (h
->type
== HASHMAP_TYPE_ORDERED
) {
547 OrderedHashmap
*lh
= (OrderedHashmap
*) h
;
548 struct ordered_hashmap_entry
*le
= ordered_bucket_at(lh
, idx
);
550 if (le
->iterate_next
!= IDX_NIL
)
551 ordered_bucket_at(lh
, le
->iterate_next
)->iterate_previous
= le
->iterate_previous
;
553 lh
->iterate_list_tail
= le
->iterate_previous
;
555 if (le
->iterate_previous
!= IDX_NIL
)
556 ordered_bucket_at(lh
, le
->iterate_previous
)->iterate_next
= le
->iterate_next
;
558 lh
->iterate_list_head
= le
->iterate_next
;
561 /* Now shift all buckets in the interval (left, right) one step backwards */
562 for (prev
= left
, left
= next_idx(h
, left
); left
!= right
;
563 prev
= left
, left
= next_idx(h
, left
)) {
564 dib
= bucket_calculate_dib(h
, left
, dibs
[left
]);
566 bucket_move_entry(h
, NULL
, left
, prev
);
567 bucket_set_dib(h
, prev
, dib
- 1);
570 bucket_mark_free(h
, prev
);
574 #define remove_entry(h, idx) base_remove_entry(HASHMAP_BASE(h), idx)
576 static unsigned hashmap_iterate_in_insertion_order(OrderedHashmap
*h
, Iterator
*i
) {
577 struct ordered_hashmap_entry
*e
;
583 if (i
->idx
== IDX_NIL
)
586 if (i
->idx
== IDX_FIRST
&& h
->iterate_list_head
== IDX_NIL
)
589 if (i
->idx
== IDX_FIRST
) {
590 idx
= h
->iterate_list_head
;
591 e
= ordered_bucket_at(h
, idx
);
594 e
= ordered_bucket_at(h
, idx
);
596 * We allow removing the current entry while iterating, but removal may cause
597 * a backward shift. The next entry may thus move one bucket to the left.
598 * To detect when it happens, we remember the key pointer of the entry we were
599 * going to iterate next. If it does not match, there was a backward shift.
601 if (e
->p
.b
.key
!= i
->next_key
) {
602 idx
= prev_idx(HASHMAP_BASE(h
), idx
);
603 e
= ordered_bucket_at(h
, idx
);
605 assert(e
->p
.b
.key
== i
->next_key
);
608 #if ENABLE_DEBUG_HASHMAP
612 if (e
->iterate_next
!= IDX_NIL
) {
613 struct ordered_hashmap_entry
*n
;
614 i
->idx
= e
->iterate_next
;
615 n
= ordered_bucket_at(h
, i
->idx
);
616 i
->next_key
= n
->p
.b
.key
;
627 static unsigned hashmap_iterate_in_internal_order(HashmapBase
*h
, Iterator
*i
) {
633 if (i
->idx
== IDX_NIL
)
636 if (i
->idx
== IDX_FIRST
) {
637 /* fast forward to the first occupied bucket */
638 if (h
->has_indirect
) {
639 i
->idx
= skip_free_buckets(h
, h
->indirect
.idx_lowest_entry
);
640 h
->indirect
.idx_lowest_entry
= i
->idx
;
642 i
->idx
= skip_free_buckets(h
, 0);
644 if (i
->idx
== IDX_NIL
)
647 struct hashmap_base_entry
*e
;
651 e
= bucket_at(h
, i
->idx
);
653 * We allow removing the current entry while iterating, but removal may cause
654 * a backward shift. The next entry may thus move one bucket to the left.
655 * To detect when it happens, we remember the key pointer of the entry we were
656 * going to iterate next. If it does not match, there was a backward shift.
658 if (e
->key
!= i
->next_key
)
659 e
= bucket_at(h
, --i
->idx
);
661 assert(e
->key
== i
->next_key
);
665 #if ENABLE_DEBUG_HASHMAP
669 i
->idx
= skip_free_buckets(h
, i
->idx
+ 1);
670 if (i
->idx
!= IDX_NIL
)
671 i
->next_key
= bucket_at(h
, i
->idx
)->key
;
682 static unsigned hashmap_iterate_entry(HashmapBase
*h
, Iterator
*i
) {
688 #if ENABLE_DEBUG_HASHMAP
689 if (i
->idx
== IDX_FIRST
) {
690 i
->put_count
= h
->debug
.put_count
;
691 i
->rem_count
= h
->debug
.rem_count
;
693 /* While iterating, must not add any new entries */
694 assert(i
->put_count
== h
->debug
.put_count
);
695 /* ... or remove entries other than the current one */
696 assert(i
->rem_count
== h
->debug
.rem_count
||
697 (i
->rem_count
== h
->debug
.rem_count
- 1 &&
698 i
->prev_idx
== h
->debug
.last_rem_idx
));
699 /* Reset our removals counter */
700 i
->rem_count
= h
->debug
.rem_count
;
704 return h
->type
== HASHMAP_TYPE_ORDERED
? hashmap_iterate_in_insertion_order((OrderedHashmap
*) h
, i
)
705 : hashmap_iterate_in_internal_order(h
, i
);
708 bool _hashmap_iterate(HashmapBase
*h
, Iterator
*i
, void **value
, const void **key
) {
709 struct hashmap_base_entry
*e
;
713 idx
= hashmap_iterate_entry(h
, i
);
714 if (idx
== IDX_NIL
) {
723 e
= bucket_at(h
, idx
);
724 data
= entry_value(h
, e
);
733 #define HASHMAP_FOREACH_IDX(idx, h, i) \
734 for ((i) = ITERATOR_FIRST, (idx) = hashmap_iterate_entry((h), &(i)); \
736 (idx) = hashmap_iterate_entry((h), &(i)))
738 IteratedCache
* _hashmap_iterated_cache_new(HashmapBase
*h
) {
739 IteratedCache
*cache
;
747 cache
= new0(IteratedCache
, 1);
757 static void reset_direct_storage(HashmapBase
*h
) {
758 const struct hashmap_type_info
*hi
= &hashmap_type_info
[h
->type
];
761 assert(!h
->has_indirect
);
763 p
= mempset(h
->direct
.storage
, 0, hi
->entry_size
* hi
->n_direct_buckets
);
764 memset(p
, DIB_RAW_INIT
, sizeof(dib_raw_t
) * hi
->n_direct_buckets
);
767 static void shared_hash_key_initialize(void) {
768 random_bytes(shared_hash_key
, sizeof(shared_hash_key
));
771 static struct HashmapBase
* hashmap_base_new(const struct hash_ops
*hash_ops
, enum HashmapType type HASHMAP_DEBUG_PARAMS
) {
773 const struct hashmap_type_info
*hi
= &hashmap_type_info
[type
];
775 bool use_pool
= mempool_enabled
&& mempool_enabled();
777 h
= use_pool
? mempool_alloc0_tile(hi
->mempool
) : malloc0(hi
->head_size
);
782 h
->from_pool
= use_pool
;
783 h
->hash_ops
= hash_ops
?: &trivial_hash_ops
;
785 if (type
== HASHMAP_TYPE_ORDERED
) {
786 OrderedHashmap
*lh
= (OrderedHashmap
*)h
;
787 lh
->iterate_list_head
= lh
->iterate_list_tail
= IDX_NIL
;
790 reset_direct_storage(h
);
792 static pthread_once_t once
= PTHREAD_ONCE_INIT
;
793 assert_se(pthread_once(&once
, shared_hash_key_initialize
) == 0);
795 #if ENABLE_DEBUG_HASHMAP
796 h
->debug
.func
= func
;
797 h
->debug
.file
= file
;
798 h
->debug
.line
= line
;
799 assert_se(pthread_mutex_lock(&hashmap_debug_list_mutex
) == 0);
800 LIST_PREPEND(debug_list
, hashmap_debug_list
, &h
->debug
);
801 assert_se(pthread_mutex_unlock(&hashmap_debug_list_mutex
) == 0);
807 Hashmap
*_hashmap_new(const struct hash_ops
*hash_ops HASHMAP_DEBUG_PARAMS
) {
808 return (Hashmap
*) hashmap_base_new(hash_ops
, HASHMAP_TYPE_PLAIN HASHMAP_DEBUG_PASS_ARGS
);
811 OrderedHashmap
*_ordered_hashmap_new(const struct hash_ops
*hash_ops HASHMAP_DEBUG_PARAMS
) {
812 return (OrderedHashmap
*) hashmap_base_new(hash_ops
, HASHMAP_TYPE_ORDERED HASHMAP_DEBUG_PASS_ARGS
);
815 Set
*_set_new(const struct hash_ops
*hash_ops HASHMAP_DEBUG_PARAMS
) {
816 return (Set
*) hashmap_base_new(hash_ops
, HASHMAP_TYPE_SET HASHMAP_DEBUG_PASS_ARGS
);
819 static int hashmap_base_ensure_allocated(HashmapBase
**h
, const struct hash_ops
*hash_ops
,
820 enum HashmapType type HASHMAP_DEBUG_PARAMS
) {
828 q
= hashmap_base_new(hash_ops
, type HASHMAP_DEBUG_PASS_ARGS
);
836 int _hashmap_ensure_allocated(Hashmap
**h
, const struct hash_ops
*hash_ops HASHMAP_DEBUG_PARAMS
) {
837 return hashmap_base_ensure_allocated((HashmapBase
**)h
, hash_ops
, HASHMAP_TYPE_PLAIN HASHMAP_DEBUG_PASS_ARGS
);
840 int _ordered_hashmap_ensure_allocated(OrderedHashmap
**h
, const struct hash_ops
*hash_ops HASHMAP_DEBUG_PARAMS
) {
841 return hashmap_base_ensure_allocated((HashmapBase
**)h
, hash_ops
, HASHMAP_TYPE_ORDERED HASHMAP_DEBUG_PASS_ARGS
);
844 int _set_ensure_allocated(Set
**s
, const struct hash_ops
*hash_ops HASHMAP_DEBUG_PARAMS
) {
845 return hashmap_base_ensure_allocated((HashmapBase
**)s
, hash_ops
, HASHMAP_TYPE_SET HASHMAP_DEBUG_PASS_ARGS
);
848 int _hashmap_ensure_put(Hashmap
**h
, const struct hash_ops
*hash_ops
, const void *key
, void *value HASHMAP_DEBUG_PARAMS
) {
851 r
= _hashmap_ensure_allocated(h
, hash_ops HASHMAP_DEBUG_PASS_ARGS
);
855 return hashmap_put(*h
, key
, value
);
858 int _ordered_hashmap_ensure_put(OrderedHashmap
**h
, const struct hash_ops
*hash_ops
, const void *key
, void *value HASHMAP_DEBUG_PARAMS
) {
861 r
= _ordered_hashmap_ensure_allocated(h
, hash_ops HASHMAP_DEBUG_PASS_ARGS
);
865 return ordered_hashmap_put(*h
, key
, value
);
868 static void hashmap_free_no_clear(HashmapBase
*h
) {
869 assert(!h
->has_indirect
);
870 assert(h
->n_direct_entries
== 0);
872 #if ENABLE_DEBUG_HASHMAP
873 assert_se(pthread_mutex_lock(&hashmap_debug_list_mutex
) == 0);
874 LIST_REMOVE(debug_list
, hashmap_debug_list
, &h
->debug
);
875 assert_se(pthread_mutex_unlock(&hashmap_debug_list_mutex
) == 0);
879 /* Ensure that the object didn't get migrated between threads. */
880 assert_se(is_main_thread());
881 mempool_free_tile(hashmap_type_info
[h
->type
].mempool
, h
);
886 HashmapBase
* _hashmap_free(HashmapBase
*h
, free_func_t default_free_key
, free_func_t default_free_value
) {
888 _hashmap_clear(h
, default_free_key
, default_free_value
);
889 hashmap_free_no_clear(h
);
895 void _hashmap_clear(HashmapBase
*h
, free_func_t default_free_key
, free_func_t default_free_value
) {
896 free_func_t free_key
, free_value
;
900 free_key
= h
->hash_ops
->free_key
?: default_free_key
;
901 free_value
= h
->hash_ops
->free_value
?: default_free_value
;
903 if (free_key
|| free_value
) {
905 /* If destructor calls are defined, let's destroy things defensively: let's take the item out of the
906 * hash table, and only then call the destructor functions. If these destructors then try to unregister
907 * themselves from our hash table a second time, the entry is already gone. */
909 while (_hashmap_size(h
) > 0) {
913 v
= _hashmap_first_key_and_value(h
, true, &k
);
923 if (h
->has_indirect
) {
924 free(h
->indirect
.storage
);
925 h
->has_indirect
= false;
928 h
->n_direct_entries
= 0;
929 reset_direct_storage(h
);
931 if (h
->type
== HASHMAP_TYPE_ORDERED
) {
932 OrderedHashmap
*lh
= (OrderedHashmap
*) h
;
933 lh
->iterate_list_head
= lh
->iterate_list_tail
= IDX_NIL
;
939 static int resize_buckets(HashmapBase
*h
, unsigned entries_add
);
942 * Finds an empty bucket to put an entry into, starting the scan at 'idx'.
943 * Performs Robin Hood swaps as it goes. The entry to put must be placed
944 * by the caller into swap slot IDX_PUT.
945 * If used for in-place resizing, may leave a displaced entry in swap slot
946 * IDX_PUT. Caller must rehash it next.
947 * Returns: true if it left a displaced entry to rehash next in IDX_PUT,
950 static bool hashmap_put_robin_hood(HashmapBase
*h
, unsigned idx
,
951 struct swap_entries
*swap
) {
952 dib_raw_t raw_dib
, *dibs
;
953 unsigned dib
, distance
;
955 #if ENABLE_DEBUG_HASHMAP
956 h
->debug
.put_count
++;
959 dibs
= dib_raw_ptr(h
);
961 for (distance
= 0; ; distance
++) {
963 if (IN_SET(raw_dib
, DIB_RAW_FREE
, DIB_RAW_REHASH
)) {
964 if (raw_dib
== DIB_RAW_REHASH
)
965 bucket_move_entry(h
, swap
, idx
, IDX_TMP
);
967 if (h
->has_indirect
&& h
->indirect
.idx_lowest_entry
> idx
)
968 h
->indirect
.idx_lowest_entry
= idx
;
970 bucket_set_dib(h
, idx
, distance
);
971 bucket_move_entry(h
, swap
, IDX_PUT
, idx
);
972 if (raw_dib
== DIB_RAW_REHASH
) {
973 bucket_move_entry(h
, swap
, IDX_TMP
, IDX_PUT
);
980 dib
= bucket_calculate_dib(h
, idx
, raw_dib
);
982 if (dib
< distance
) {
983 /* Found a wealthier entry. Go Robin Hood! */
984 bucket_set_dib(h
, idx
, distance
);
986 /* swap the entries */
987 bucket_move_entry(h
, swap
, idx
, IDX_TMP
);
988 bucket_move_entry(h
, swap
, IDX_PUT
, idx
);
989 bucket_move_entry(h
, swap
, IDX_TMP
, IDX_PUT
);
994 idx
= next_idx(h
, idx
);
999 * Puts an entry into a hashmap, boldly - no check whether key already exists.
1000 * The caller must place the entry (only its key and value, not link indexes)
1001 * in swap slot IDX_PUT.
1002 * Caller must ensure: the key does not exist yet in the hashmap.
1003 * that resize is not needed if !may_resize.
1004 * Returns: 1 if entry was put successfully.
1005 * -ENOMEM if may_resize==true and resize failed with -ENOMEM.
1006 * Cannot return -ENOMEM if !may_resize.
1008 static int hashmap_base_put_boldly(HashmapBase
*h
, unsigned idx
,
1009 struct swap_entries
*swap
, bool may_resize
) {
1010 struct ordered_hashmap_entry
*new_entry
;
1013 assert(idx
< n_buckets(h
));
1015 new_entry
= bucket_at_swap(swap
, IDX_PUT
);
1018 r
= resize_buckets(h
, 1);
1022 idx
= bucket_hash(h
, new_entry
->p
.b
.key
);
1024 assert(n_entries(h
) < n_buckets(h
));
1026 if (h
->type
== HASHMAP_TYPE_ORDERED
) {
1027 OrderedHashmap
*lh
= (OrderedHashmap
*) h
;
1029 new_entry
->iterate_next
= IDX_NIL
;
1030 new_entry
->iterate_previous
= lh
->iterate_list_tail
;
1032 if (lh
->iterate_list_tail
!= IDX_NIL
) {
1033 struct ordered_hashmap_entry
*old_tail
;
1035 old_tail
= ordered_bucket_at(lh
, lh
->iterate_list_tail
);
1036 assert(old_tail
->iterate_next
== IDX_NIL
);
1037 old_tail
->iterate_next
= IDX_PUT
;
1040 lh
->iterate_list_tail
= IDX_PUT
;
1041 if (lh
->iterate_list_head
== IDX_NIL
)
1042 lh
->iterate_list_head
= IDX_PUT
;
1045 assert_se(hashmap_put_robin_hood(h
, idx
, swap
) == false);
1048 #if ENABLE_DEBUG_HASHMAP
1049 h
->debug
.max_entries
= MAX(h
->debug
.max_entries
, n_entries(h
));
1056 #define hashmap_put_boldly(h, idx, swap, may_resize) \
1057 hashmap_base_put_boldly(HASHMAP_BASE(h), idx, swap, may_resize)
1060 * Returns 0 if resize is not needed.
1061 * 1 if successfully resized.
1062 * -ENOMEM on allocation failure.
1064 static int resize_buckets(HashmapBase
*h
, unsigned entries_add
) {
1065 struct swap_entries swap
;
1067 dib_raw_t
*old_dibs
, *new_dibs
;
1068 const struct hashmap_type_info
*hi
;
1069 unsigned idx
, optimal_idx
;
1070 unsigned old_n_buckets
, new_n_buckets
, n_rehashed
, new_n_entries
;
1076 hi
= &hashmap_type_info
[h
->type
];
1077 new_n_entries
= n_entries(h
) + entries_add
;
1080 if (_unlikely_(new_n_entries
< entries_add
))
1083 /* For direct storage we allow 100% load, because it's tiny. */
1084 if (!h
->has_indirect
&& new_n_entries
<= hi
->n_direct_buckets
)
1088 * Load factor = n/m = 1 - (1/INV_KEEP_FREE).
1089 * From it follows: m = n + n/(INV_KEEP_FREE - 1)
1091 new_n_buckets
= new_n_entries
+ new_n_entries
/ (INV_KEEP_FREE
- 1);
1093 if (_unlikely_(new_n_buckets
< new_n_entries
))
1096 if (_unlikely_(new_n_buckets
> UINT_MAX
/ (hi
->entry_size
+ sizeof(dib_raw_t
))))
1099 old_n_buckets
= n_buckets(h
);
1101 if (_likely_(new_n_buckets
<= old_n_buckets
))
1104 new_shift
= log2u_round_up(MAX(
1105 new_n_buckets
* (hi
->entry_size
+ sizeof(dib_raw_t
)),
1106 2 * sizeof(struct direct_storage
)));
1108 /* Realloc storage (buckets and DIB array). */
1109 new_storage
= realloc(h
->has_indirect
? h
->indirect
.storage
: NULL
,
1114 /* Must upgrade direct to indirect storage. */
1115 if (!h
->has_indirect
) {
1116 memcpy(new_storage
, h
->direct
.storage
,
1117 old_n_buckets
* (hi
->entry_size
+ sizeof(dib_raw_t
)));
1118 h
->indirect
.n_entries
= h
->n_direct_entries
;
1119 h
->indirect
.idx_lowest_entry
= 0;
1120 h
->n_direct_entries
= 0;
1123 /* Get a new hash key. If we've just upgraded to indirect storage,
1124 * allow reusing a previously generated key. It's still a different key
1125 * from the shared one that we used for direct storage. */
1126 get_hash_key(h
->indirect
.hash_key
, !h
->has_indirect
);
1128 h
->has_indirect
= true;
1129 h
->indirect
.storage
= new_storage
;
1130 h
->indirect
.n_buckets
= (1U << new_shift
) /
1131 (hi
->entry_size
+ sizeof(dib_raw_t
));
1133 old_dibs
= (dib_raw_t
*)((uint8_t*) new_storage
+ hi
->entry_size
* old_n_buckets
);
1134 new_dibs
= dib_raw_ptr(h
);
1137 * Move the DIB array to the new place, replacing valid DIB values with
1138 * DIB_RAW_REHASH to indicate all of the used buckets need rehashing.
1139 * Note: Overlap is not possible, because we have at least doubled the
1140 * number of buckets and dib_raw_t is smaller than any entry type.
1142 for (idx
= 0; idx
< old_n_buckets
; idx
++) {
1143 assert(old_dibs
[idx
] != DIB_RAW_REHASH
);
1144 new_dibs
[idx
] = old_dibs
[idx
] == DIB_RAW_FREE
? DIB_RAW_FREE
1148 /* Zero the area of newly added entries (including the old DIB area) */
1149 memzero(bucket_at(h
, old_n_buckets
),
1150 (n_buckets(h
) - old_n_buckets
) * hi
->entry_size
);
1152 /* The upper half of the new DIB array needs initialization */
1153 memset(&new_dibs
[old_n_buckets
], DIB_RAW_INIT
,
1154 (n_buckets(h
) - old_n_buckets
) * sizeof(dib_raw_t
));
1156 /* Rehash entries that need it */
1158 for (idx
= 0; idx
< old_n_buckets
; idx
++) {
1159 if (new_dibs
[idx
] != DIB_RAW_REHASH
)
1162 optimal_idx
= bucket_hash(h
, bucket_at(h
, idx
)->key
);
1165 * Not much to do if by luck the entry hashes to its current
1166 * location. Just set its DIB.
1168 if (optimal_idx
== idx
) {
1174 new_dibs
[idx
] = DIB_RAW_FREE
;
1175 bucket_move_entry(h
, &swap
, idx
, IDX_PUT
);
1176 /* bucket_move_entry does not clear the source */
1177 memzero(bucket_at(h
, idx
), hi
->entry_size
);
1181 * Find the new bucket for the current entry. This may make
1182 * another entry homeless and load it into IDX_PUT.
1184 rehash_next
= hashmap_put_robin_hood(h
, optimal_idx
, &swap
);
1187 /* Did the current entry displace another one? */
1189 optimal_idx
= bucket_hash(h
, bucket_at_swap(&swap
, IDX_PUT
)->p
.b
.key
);
1190 } while (rehash_next
);
1193 assert(n_rehashed
== n_entries(h
));
1199 * Finds an entry with a matching key
1200 * Returns: index of the found entry, or IDX_NIL if not found.
1202 static unsigned base_bucket_scan(HashmapBase
*h
, unsigned idx
, const void *key
) {
1203 struct hashmap_base_entry
*e
;
1204 unsigned dib
, distance
;
1205 dib_raw_t
*dibs
= dib_raw_ptr(h
);
1207 assert(idx
< n_buckets(h
));
1209 for (distance
= 0; ; distance
++) {
1210 if (dibs
[idx
] == DIB_RAW_FREE
)
1213 dib
= bucket_calculate_dib(h
, idx
, dibs
[idx
]);
1217 if (dib
== distance
) {
1218 e
= bucket_at(h
, idx
);
1219 if (h
->hash_ops
->compare(e
->key
, key
) == 0)
1223 idx
= next_idx(h
, idx
);
1226 #define bucket_scan(h, idx, key) base_bucket_scan(HASHMAP_BASE(h), idx, key)
1228 int hashmap_put(Hashmap
*h
, const void *key
, void *value
) {
1229 struct swap_entries swap
;
1230 struct plain_hashmap_entry
*e
;
1235 hash
= bucket_hash(h
, key
);
1236 idx
= bucket_scan(h
, hash
, key
);
1237 if (idx
!= IDX_NIL
) {
1238 e
= plain_bucket_at(h
, idx
);
1239 if (e
->value
== value
)
1244 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
;
1247 return hashmap_put_boldly(h
, hash
, &swap
, true);
1250 int set_put(Set
*s
, const void *key
) {
1251 struct swap_entries swap
;
1252 struct hashmap_base_entry
*e
;
1257 hash
= bucket_hash(s
, key
);
1258 idx
= bucket_scan(s
, hash
, key
);
1262 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
.b
;
1264 return hashmap_put_boldly(s
, hash
, &swap
, true);
1267 int _set_ensure_put(Set
**s
, const struct hash_ops
*hash_ops
, const void *key HASHMAP_DEBUG_PARAMS
) {
1270 r
= _set_ensure_allocated(s
, hash_ops HASHMAP_DEBUG_PASS_ARGS
);
1274 return set_put(*s
, key
);
1277 int _set_ensure_consume(Set
**s
, const struct hash_ops
*hash_ops
, void *key HASHMAP_DEBUG_PARAMS
) {
1280 r
= _set_ensure_put(s
, hash_ops
, key HASHMAP_DEBUG_PASS_ARGS
);
1282 if (hash_ops
&& hash_ops
->free_key
)
1283 hash_ops
->free_key(key
);
1291 int hashmap_replace(Hashmap
*h
, const void *key
, void *value
) {
1292 struct swap_entries swap
;
1293 struct plain_hashmap_entry
*e
;
1298 hash
= bucket_hash(h
, key
);
1299 idx
= bucket_scan(h
, hash
, key
);
1300 if (idx
!= IDX_NIL
) {
1301 e
= plain_bucket_at(h
, idx
);
1302 #if ENABLE_DEBUG_HASHMAP
1303 /* Although the key is equal, the key pointer may have changed,
1304 * and this would break our assumption for iterating. So count
1305 * this operation as incompatible with iteration. */
1306 if (e
->b
.key
!= key
) {
1307 h
->b
.debug
.put_count
++;
1308 h
->b
.debug
.rem_count
++;
1309 h
->b
.debug
.last_rem_idx
= idx
;
1314 hashmap_set_dirty(h
);
1319 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
;
1322 return hashmap_put_boldly(h
, hash
, &swap
, true);
1325 int hashmap_update(Hashmap
*h
, const void *key
, void *value
) {
1326 struct plain_hashmap_entry
*e
;
1331 hash
= bucket_hash(h
, key
);
1332 idx
= bucket_scan(h
, hash
, key
);
1336 e
= plain_bucket_at(h
, idx
);
1338 hashmap_set_dirty(h
);
1343 void* _hashmap_get(HashmapBase
*h
, const void *key
) {
1344 struct hashmap_base_entry
*e
;
1350 hash
= bucket_hash(h
, key
);
1351 idx
= bucket_scan(h
, hash
, key
);
1355 e
= bucket_at(h
, idx
);
1356 return entry_value(h
, e
);
1359 void* hashmap_get2(Hashmap
*h
, const void *key
, void **key2
) {
1360 struct plain_hashmap_entry
*e
;
1366 hash
= bucket_hash(h
, key
);
1367 idx
= bucket_scan(h
, hash
, key
);
1371 e
= plain_bucket_at(h
, idx
);
1373 *key2
= (void*) e
->b
.key
;
1378 bool _hashmap_contains(HashmapBase
*h
, const void *key
) {
1384 hash
= bucket_hash(h
, key
);
1385 return bucket_scan(h
, hash
, key
) != IDX_NIL
;
1388 void* _hashmap_remove(HashmapBase
*h
, const void *key
) {
1389 struct hashmap_base_entry
*e
;
1396 hash
= bucket_hash(h
, key
);
1397 idx
= bucket_scan(h
, hash
, key
);
1401 e
= bucket_at(h
, idx
);
1402 data
= entry_value(h
, e
);
1403 remove_entry(h
, idx
);
1408 void* hashmap_remove2(Hashmap
*h
, const void *key
, void **rkey
) {
1409 struct plain_hashmap_entry
*e
;
1419 hash
= bucket_hash(h
, key
);
1420 idx
= bucket_scan(h
, hash
, key
);
1421 if (idx
== IDX_NIL
) {
1427 e
= plain_bucket_at(h
, idx
);
1430 *rkey
= (void*) e
->b
.key
;
1432 remove_entry(h
, idx
);
1437 int hashmap_remove_and_put(Hashmap
*h
, const void *old_key
, const void *new_key
, void *value
) {
1438 struct swap_entries swap
;
1439 struct plain_hashmap_entry
*e
;
1440 unsigned old_hash
, new_hash
, idx
;
1445 old_hash
= bucket_hash(h
, old_key
);
1446 idx
= bucket_scan(h
, old_hash
, old_key
);
1450 new_hash
= bucket_hash(h
, new_key
);
1451 if (bucket_scan(h
, new_hash
, new_key
) != IDX_NIL
)
1454 remove_entry(h
, idx
);
1456 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
;
1459 assert_se(hashmap_put_boldly(h
, new_hash
, &swap
, false) == 1);
1464 int set_remove_and_put(Set
*s
, const void *old_key
, const void *new_key
) {
1465 struct swap_entries swap
;
1466 struct hashmap_base_entry
*e
;
1467 unsigned old_hash
, new_hash
, idx
;
1472 old_hash
= bucket_hash(s
, old_key
);
1473 idx
= bucket_scan(s
, old_hash
, old_key
);
1477 new_hash
= bucket_hash(s
, new_key
);
1478 if (bucket_scan(s
, new_hash
, new_key
) != IDX_NIL
)
1481 remove_entry(s
, idx
);
1483 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
.b
;
1485 assert_se(hashmap_put_boldly(s
, new_hash
, &swap
, false) == 1);
1490 int hashmap_remove_and_replace(Hashmap
*h
, const void *old_key
, const void *new_key
, void *value
) {
1491 struct swap_entries swap
;
1492 struct plain_hashmap_entry
*e
;
1493 unsigned old_hash
, new_hash
, idx_old
, idx_new
;
1498 old_hash
= bucket_hash(h
, old_key
);
1499 idx_old
= bucket_scan(h
, old_hash
, old_key
);
1500 if (idx_old
== IDX_NIL
)
1503 old_key
= bucket_at(HASHMAP_BASE(h
), idx_old
)->key
;
1505 new_hash
= bucket_hash(h
, new_key
);
1506 idx_new
= bucket_scan(h
, new_hash
, new_key
);
1507 if (idx_new
!= IDX_NIL
)
1508 if (idx_old
!= idx_new
) {
1509 remove_entry(h
, idx_new
);
1510 /* Compensate for a possible backward shift. */
1511 if (old_key
!= bucket_at(HASHMAP_BASE(h
), idx_old
)->key
)
1512 idx_old
= prev_idx(HASHMAP_BASE(h
), idx_old
);
1513 assert(old_key
== bucket_at(HASHMAP_BASE(h
), idx_old
)->key
);
1516 remove_entry(h
, idx_old
);
1518 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
;
1521 assert_se(hashmap_put_boldly(h
, new_hash
, &swap
, false) == 1);
1526 void* _hashmap_remove_value(HashmapBase
*h
, const void *key
, void *value
) {
1527 struct hashmap_base_entry
*e
;
1533 hash
= bucket_hash(h
, key
);
1534 idx
= bucket_scan(h
, hash
, key
);
1538 e
= bucket_at(h
, idx
);
1539 if (entry_value(h
, e
) != value
)
1542 remove_entry(h
, idx
);
1547 static unsigned find_first_entry(HashmapBase
*h
) {
1548 Iterator i
= ITERATOR_FIRST
;
1550 if (!h
|| !n_entries(h
))
1553 return hashmap_iterate_entry(h
, &i
);
1556 void* _hashmap_first_key_and_value(HashmapBase
*h
, bool remove
, void **ret_key
) {
1557 struct hashmap_base_entry
*e
;
1561 idx
= find_first_entry(h
);
1562 if (idx
== IDX_NIL
) {
1568 e
= bucket_at(h
, idx
);
1569 key
= (void*) e
->key
;
1570 data
= entry_value(h
, e
);
1573 remove_entry(h
, idx
);
1581 unsigned _hashmap_size(HashmapBase
*h
) {
1585 return n_entries(h
);
1588 unsigned _hashmap_buckets(HashmapBase
*h
) {
1592 return n_buckets(h
);
1595 int _hashmap_merge(Hashmap
*h
, Hashmap
*other
) {
1601 HASHMAP_FOREACH_IDX(idx
, HASHMAP_BASE(other
), i
) {
1602 struct plain_hashmap_entry
*pe
= plain_bucket_at(other
, idx
);
1605 r
= hashmap_put(h
, pe
->b
.key
, pe
->value
);
1606 if (r
< 0 && r
!= -EEXIST
)
1613 int set_merge(Set
*s
, Set
*other
) {
1619 HASHMAP_FOREACH_IDX(idx
, HASHMAP_BASE(other
), i
) {
1620 struct set_entry
*se
= set_bucket_at(other
, idx
);
1623 r
= set_put(s
, se
->b
.key
);
1631 int _hashmap_reserve(HashmapBase
*h
, unsigned entries_add
) {
1636 r
= resize_buckets(h
, entries_add
);
1644 * The same as hashmap_merge(), but every new item from other is moved to h.
1645 * Keys already in h are skipped and stay in other.
1646 * Returns: 0 on success.
1647 * -ENOMEM on alloc failure, in which case no move has been done.
1649 int _hashmap_move(HashmapBase
*h
, HashmapBase
*other
) {
1650 struct swap_entries swap
;
1651 struct hashmap_base_entry
*e
, *n
;
1661 assert(other
->type
== h
->type
);
1664 * This reserves buckets for the worst case, where none of other's
1665 * entries are yet present in h. This is preferable to risking
1666 * an allocation failure in the middle of the moving and having to
1667 * rollback or return a partial result.
1669 r
= resize_buckets(h
, n_entries(other
));
1673 HASHMAP_FOREACH_IDX(idx
, other
, i
) {
1676 e
= bucket_at(other
, idx
);
1677 h_hash
= bucket_hash(h
, e
->key
);
1678 if (bucket_scan(h
, h_hash
, e
->key
) != IDX_NIL
)
1681 n
= &bucket_at_swap(&swap
, IDX_PUT
)->p
.b
;
1683 if (h
->type
!= HASHMAP_TYPE_SET
)
1684 ((struct plain_hashmap_entry
*) n
)->value
=
1685 ((struct plain_hashmap_entry
*) e
)->value
;
1686 assert_se(hashmap_put_boldly(h
, h_hash
, &swap
, false) == 1);
1688 remove_entry(other
, idx
);
1694 int _hashmap_move_one(HashmapBase
*h
, HashmapBase
*other
, const void *key
) {
1695 struct swap_entries swap
;
1696 unsigned h_hash
, other_hash
, idx
;
1697 struct hashmap_base_entry
*e
, *n
;
1702 h_hash
= bucket_hash(h
, key
);
1703 if (bucket_scan(h
, h_hash
, key
) != IDX_NIL
)
1709 assert(other
->type
== h
->type
);
1711 other_hash
= bucket_hash(other
, key
);
1712 idx
= bucket_scan(other
, other_hash
, key
);
1716 e
= bucket_at(other
, idx
);
1718 n
= &bucket_at_swap(&swap
, IDX_PUT
)->p
.b
;
1720 if (h
->type
!= HASHMAP_TYPE_SET
)
1721 ((struct plain_hashmap_entry
*) n
)->value
=
1722 ((struct plain_hashmap_entry
*) e
)->value
;
1723 r
= hashmap_put_boldly(h
, h_hash
, &swap
, true);
1727 remove_entry(other
, idx
);
1731 HashmapBase
* _hashmap_copy(HashmapBase
*h HASHMAP_DEBUG_PARAMS
) {
1737 copy
= hashmap_base_new(h
->hash_ops
, h
->type HASHMAP_DEBUG_PASS_ARGS
);
1742 case HASHMAP_TYPE_PLAIN
:
1743 case HASHMAP_TYPE_ORDERED
:
1744 r
= hashmap_merge((Hashmap
*)copy
, (Hashmap
*)h
);
1746 case HASHMAP_TYPE_SET
:
1747 r
= set_merge((Set
*)copy
, (Set
*)h
);
1750 assert_not_reached();
1754 return _hashmap_free(copy
, false, false);
1759 char** _hashmap_get_strv(HashmapBase
*h
) {
1765 return new0(char*, 1);
1767 sv
= new(char*, n_entries(h
)+1);
1772 HASHMAP_FOREACH_IDX(idx
, h
, i
)
1773 sv
[n
++] = entry_value(h
, bucket_at(h
, idx
));
1779 void* ordered_hashmap_next(OrderedHashmap
*h
, const void *key
) {
1780 struct ordered_hashmap_entry
*e
;
1786 hash
= bucket_hash(h
, key
);
1787 idx
= bucket_scan(h
, hash
, key
);
1791 e
= ordered_bucket_at(h
, idx
);
1792 if (e
->iterate_next
== IDX_NIL
)
1794 return ordered_bucket_at(h
, e
->iterate_next
)->p
.value
;
1797 int set_consume(Set
*s
, void *value
) {
1803 r
= set_put(s
, value
);
1810 int _hashmap_put_strdup_full(Hashmap
**h
, const struct hash_ops
*hash_ops
, const char *k
, const char *v HASHMAP_DEBUG_PARAMS
) {
1813 r
= _hashmap_ensure_allocated(h
, hash_ops HASHMAP_DEBUG_PASS_ARGS
);
1817 _cleanup_free_
char *kdup
= NULL
, *vdup
= NULL
;
1829 r
= hashmap_put(*h
, kdup
, vdup
);
1831 if (r
== -EEXIST
&& streq_ptr(v
, hashmap_get(*h
, kdup
)))
1836 /* 0 with non-null vdup would mean vdup is already in the hashmap, which cannot be */
1837 assert(vdup
== NULL
|| r
> 0);
1844 int _set_put_strndup_full(Set
**s
, const struct hash_ops
*hash_ops
, const char *p
, size_t n HASHMAP_DEBUG_PARAMS
) {
1851 r
= _set_ensure_allocated(s
, hash_ops HASHMAP_DEBUG_PASS_ARGS
);
1855 if (n
== SIZE_MAX
) {
1856 if (set_contains(*s
, (char*) p
))
1865 return set_consume(*s
, c
);
1868 int _set_put_strdupv_full(Set
**s
, const struct hash_ops
*hash_ops
, char **l HASHMAP_DEBUG_PARAMS
) {
1873 STRV_FOREACH(i
, l
) {
1874 r
= _set_put_strndup_full(s
, hash_ops
, *i
, SIZE_MAX HASHMAP_DEBUG_PASS_ARGS
);
1884 int set_put_strsplit(Set
*s
, const char *v
, const char *separators
, ExtractFlags flags
) {
1894 r
= extract_first_word(&p
, &word
, separators
, flags
);
1898 r
= set_consume(s
, word
);
1904 /* expand the cachemem if needed, return true if newly (re)activated. */
1905 static int cachemem_maintain(CacheMem
*mem
, size_t size
) {
1908 if (!GREEDY_REALLOC(mem
->ptr
, size
)) {
1921 int iterated_cache_get(IteratedCache
*cache
, const void ***res_keys
, const void ***res_values
, unsigned *res_n_entries
) {
1922 bool sync_keys
= false, sync_values
= false;
1927 assert(cache
->hashmap
);
1929 size
= n_entries(cache
->hashmap
);
1932 r
= cachemem_maintain(&cache
->keys
, size
);
1938 cache
->keys
.active
= false;
1941 r
= cachemem_maintain(&cache
->values
, size
);
1947 cache
->values
.active
= false;
1949 if (cache
->hashmap
->dirty
) {
1950 if (cache
->keys
.active
)
1952 if (cache
->values
.active
)
1955 cache
->hashmap
->dirty
= false;
1958 if (sync_keys
|| sync_values
) {
1963 HASHMAP_FOREACH_IDX(idx
, cache
->hashmap
, iter
) {
1964 struct hashmap_base_entry
*e
;
1966 e
= bucket_at(cache
->hashmap
, idx
);
1969 cache
->keys
.ptr
[i
] = e
->key
;
1971 cache
->values
.ptr
[i
] = entry_value(cache
->hashmap
, e
);
1977 *res_keys
= cache
->keys
.ptr
;
1979 *res_values
= cache
->values
.ptr
;
1981 *res_n_entries
= size
;
1986 IteratedCache
* iterated_cache_free(IteratedCache
*cache
) {
1988 free(cache
->keys
.ptr
);
1989 free(cache
->values
.ptr
);
1992 return mfree(cache
);
1995 int set_strjoin(Set
*s
, const char *separator
, bool wrap_with_separator
, char **ret
) {
1996 _cleanup_free_
char *str
= NULL
;
1997 size_t separator_len
, len
= 0;
2003 if (set_isempty(s
)) {
2008 separator_len
= strlen_ptr(separator
);
2010 if (separator_len
== 0)
2011 wrap_with_separator
= false;
2013 first
= !wrap_with_separator
;
2015 SET_FOREACH(value
, s
) {
2016 size_t l
= strlen_ptr(value
);
2021 if (!GREEDY_REALLOC(str
, len
+ l
+ (first
? 0 : separator_len
) + (wrap_with_separator
? separator_len
: 0) + 1))
2024 if (separator_len
> 0 && !first
) {
2025 memcpy(str
+ len
, separator
, separator_len
);
2026 len
+= separator_len
;
2029 memcpy(str
+ len
, value
, l
);
2034 if (wrap_with_separator
) {
2035 memcpy(str
+ len
, separator
, separator_len
);
2036 len
+= separator_len
;
2041 *ret
= TAKE_PTR(str
);
2045 bool set_equal(Set
*a
, Set
*b
) {
2048 /* Checks whether each entry of 'a' is also in 'b' and vice versa, i.e. the two sets contain the same
2054 if (set_isempty(a
) && set_isempty(b
))
2057 if (set_size(a
) != set_size(b
)) /* Cheap check that hopefully catches a lot of inequality cases
2062 if (!set_contains(b
, p
))
2065 /* If we have the same hashops, then we don't need to check things backwards given we compared the
2066 * size and that all of a is in b. */
2067 if (a
->b
.hash_ops
== b
->b
.hash_ops
)
2071 if (!set_contains(a
, p
))
2077 static bool set_fnmatch_one(Set
*patterns
, const char *needle
) {
2082 SET_FOREACH(p
, patterns
)
2083 if (fnmatch(p
, needle
, 0) == 0)
2089 bool set_fnmatch(Set
*include_patterns
, Set
*exclude_patterns
, const char *needle
) {
2092 if (set_fnmatch_one(exclude_patterns
, needle
))
2095 if (set_isempty(include_patterns
))
2098 return set_fnmatch_one(include_patterns
, needle
);