1 /* SPDX-License-Identifier: LGPL-2.1+ */
8 #include "alloc-util.h"
12 #include "memory-util.h"
14 #include "missing_syscall.h"
15 #include "process-util.h"
16 #include "random-util.h"
18 #include "siphash24.h"
19 #include "string-util.h"
22 #if ENABLE_DEBUG_HASHMAP
27 * Implementation of hashmaps.
29 * - uses less RAM compared to closed addressing (chaining), because
30 * our entries are small (especially in Sets, which tend to contain
31 * the majority of entries in systemd).
32 * Collision resolution: Robin Hood
33 * - tends to equalize displacement of entries from their optimal buckets.
34 * Probe sequence: linear
35 * - though theoretically worse than random probing/uniform hashing/double
36 * hashing, it is good for cache locality.
39 * Celis, P. 1986. Robin Hood Hashing.
40 * Ph.D. Dissertation. University of Waterloo, Waterloo, Ont., Canada, Canada.
41 * https://cs.uwaterloo.ca/research/tr/1986/CS-86-14.pdf
42 * - The results are derived for random probing. Suggests deletion with
43 * tombstones and two mean-centered search methods. None of that works
44 * well for linear probing.
46 * Janson, S. 2005. Individual displacements for linear probing hashing with different insertion policies.
47 * ACM Trans. Algorithms 1, 2 (October 2005), 177-213.
48 * DOI=10.1145/1103963.1103964 http://doi.acm.org/10.1145/1103963.1103964
49 * http://www.math.uu.se/~svante/papers/sj157.pdf
50 * - Applies to Robin Hood with linear probing. Contains remarks on
51 * the unsuitability of mean-centered search with linear probing.
53 * Viola, A. 2005. Exact distribution of individual displacements in linear probing hashing.
54 * ACM Trans. Algorithms 1, 2 (October 2005), 214-242.
55 * DOI=10.1145/1103963.1103965 http://doi.acm.org/10.1145/1103963.1103965
56 * - Similar to Janson. Note that Viola writes about C_{m,n} (number of probes
57 * in a successful search), and Janson writes about displacement. C = d + 1.
59 * Goossaert, E. 2013. Robin Hood hashing: backward shift deletion.
60 * http://codecapsule.com/2013/11/17/robin-hood-hashing-backward-shift-deletion/
61 * - Explanation of backward shift deletion with pictures.
63 * Khuong, P. 2013. The Other Robin Hood Hashing.
64 * http://www.pvk.ca/Blog/2013/11/26/the-other-robin-hood-hashing/
65 * - Short summary of random vs. linear probing, and tombstones vs. backward shift.
69 * XXX Ideas for improvement:
70 * For unordered hashmaps, randomize iteration order, similarly to Perl:
71 * http://blog.booking.com/hardening-perls-hash-function.html
74 /* INV_KEEP_FREE = 1 / (1 - max_load_factor)
75 * e.g. 1 / (1 - 0.8) = 5 ... keep one fifth of the buckets free. */
76 #define INV_KEEP_FREE 5U
78 /* Fields common to entries of all hashmap/set types */
79 struct hashmap_base_entry
{
83 /* Entry types for specific hashmap/set types
84 * hashmap_base_entry must be at the beginning of each entry struct. */
86 struct plain_hashmap_entry
{
87 struct hashmap_base_entry b
;
91 struct ordered_hashmap_entry
{
92 struct plain_hashmap_entry p
;
93 unsigned iterate_next
, iterate_previous
;
97 struct hashmap_base_entry b
;
100 /* In several functions it is advantageous to have the hash table extended
101 * virtually by a couple of additional buckets. We reserve special index values
102 * for these "swap" buckets. */
103 #define _IDX_SWAP_BEGIN (UINT_MAX - 3)
104 #define IDX_PUT (_IDX_SWAP_BEGIN + 0)
105 #define IDX_TMP (_IDX_SWAP_BEGIN + 1)
106 #define _IDX_SWAP_END (_IDX_SWAP_BEGIN + 2)
108 #define IDX_FIRST (UINT_MAX - 1) /* special index for freshly initialized iterators */
109 #define IDX_NIL UINT_MAX /* special index value meaning "none" or "end" */
111 assert_cc(IDX_FIRST
== _IDX_SWAP_END
);
112 assert_cc(IDX_FIRST
== _IDX_ITERATOR_FIRST
);
114 /* Storage space for the "swap" buckets.
115 * All entry types can fit into a ordered_hashmap_entry. */
116 struct swap_entries
{
117 struct ordered_hashmap_entry e
[_IDX_SWAP_END
- _IDX_SWAP_BEGIN
];
120 /* Distance from Initial Bucket */
121 typedef uint8_t dib_raw_t
;
122 #define DIB_RAW_OVERFLOW ((dib_raw_t)0xfdU) /* indicates DIB value is greater than representable */
123 #define DIB_RAW_REHASH ((dib_raw_t)0xfeU) /* entry yet to be rehashed during in-place resize */
124 #define DIB_RAW_FREE ((dib_raw_t)0xffU) /* a free bucket */
125 #define DIB_RAW_INIT ((char)DIB_RAW_FREE) /* a byte to memset a DIB store with when initializing */
127 #define DIB_FREE UINT_MAX
129 #if ENABLE_DEBUG_HASHMAP
130 struct hashmap_debug_info
{
131 LIST_FIELDS(struct hashmap_debug_info
, debug_list
);
132 unsigned max_entries
; /* high watermark of n_entries */
134 /* who allocated this hashmap */
139 /* fields to detect modification while iterating */
140 unsigned put_count
; /* counts puts into the hashmap */
141 unsigned rem_count
; /* counts removals from hashmap */
142 unsigned last_rem_idx
; /* remembers last removal index */
145 /* Tracks all existing hashmaps. Get at it from gdb. See sd_dump_hashmaps.py */
146 static LIST_HEAD(struct hashmap_debug_info
, hashmap_debug_list
);
147 static pthread_mutex_t hashmap_debug_list_mutex
= PTHREAD_MUTEX_INITIALIZER
;
152 HASHMAP_TYPE_ORDERED
,
157 struct _packed_ indirect_storage
{
158 void *storage
; /* where buckets and DIBs are stored */
159 uint8_t hash_key
[HASH_KEY_SIZE
]; /* hash key; changes during resize */
161 unsigned n_entries
; /* number of stored entries */
162 unsigned n_buckets
; /* number of buckets */
164 unsigned idx_lowest_entry
; /* Index below which all buckets are free.
165 Makes "while(hashmap_steal_first())" loops
166 O(n) instead of O(n^2) for unordered hashmaps. */
167 uint8_t _pad
[3]; /* padding for the whole HashmapBase */
168 /* The bitfields in HashmapBase complete the alignment of the whole thing. */
171 struct direct_storage
{
172 /* This gives us 39 bytes on 64bit, or 35 bytes on 32bit.
173 * That's room for 4 set_entries + 4 DIB bytes + 3 unused bytes on 64bit,
174 * or 7 set_entries + 7 DIB bytes + 0 unused bytes on 32bit. */
175 uint8_t storage
[sizeof(struct indirect_storage
)];
178 #define DIRECT_BUCKETS(entry_t) \
179 (sizeof(struct direct_storage) / (sizeof(entry_t) + sizeof(dib_raw_t)))
181 /* We should be able to store at least one entry directly. */
182 assert_cc(DIRECT_BUCKETS(struct ordered_hashmap_entry
) >= 1);
184 /* We have 3 bits for n_direct_entries. */
185 assert_cc(DIRECT_BUCKETS(struct set_entry
) < (1 << 3));
187 /* Hashmaps with directly stored entries all use this shared hash key.
188 * It's no big deal if the key is guessed, because there can be only
189 * a handful of directly stored entries in a hashmap. When a hashmap
190 * outgrows direct storage, it gets its own key for indirect storage. */
191 static uint8_t shared_hash_key
[HASH_KEY_SIZE
];
193 /* Fields that all hashmap/set types must have */
195 const struct hash_ops
*hash_ops
; /* hash and compare ops to use */
198 struct indirect_storage indirect
; /* if has_indirect */
199 struct direct_storage direct
; /* if !has_indirect */
202 enum HashmapType type
:2; /* HASHMAP_TYPE_* */
203 bool has_indirect
:1; /* whether indirect storage is used */
204 unsigned n_direct_entries
:3; /* Number of entries in direct storage.
205 * Only valid if !has_indirect. */
206 bool from_pool
:1; /* whether was allocated from mempool */
207 bool dirty
:1; /* whether dirtied since last iterated_cache_get() */
208 bool cached
:1; /* whether this hashmap is being cached */
210 #if ENABLE_DEBUG_HASHMAP
211 struct hashmap_debug_info debug
;
215 /* Specific hash types
216 * HashmapBase must be at the beginning of each hashmap struct. */
219 struct HashmapBase b
;
222 struct OrderedHashmap
{
223 struct HashmapBase b
;
224 unsigned iterate_list_head
, iterate_list_tail
;
228 struct HashmapBase b
;
231 typedef struct CacheMem
{
233 size_t n_populated
, n_allocated
;
237 struct IteratedCache
{
238 HashmapBase
*hashmap
;
239 CacheMem keys
, values
;
242 DEFINE_MEMPOOL(hashmap_pool
, Hashmap
, 8);
243 DEFINE_MEMPOOL(ordered_hashmap_pool
, OrderedHashmap
, 8);
244 /* No need for a separate Set pool */
245 assert_cc(sizeof(Hashmap
) == sizeof(Set
));
247 struct hashmap_type_info
{
250 struct mempool
*mempool
;
251 unsigned n_direct_buckets
;
254 static _used_
const struct hashmap_type_info hashmap_type_info
[_HASHMAP_TYPE_MAX
] = {
255 [HASHMAP_TYPE_PLAIN
] = {
256 .head_size
= sizeof(Hashmap
),
257 .entry_size
= sizeof(struct plain_hashmap_entry
),
258 .mempool
= &hashmap_pool
,
259 .n_direct_buckets
= DIRECT_BUCKETS(struct plain_hashmap_entry
),
261 [HASHMAP_TYPE_ORDERED
] = {
262 .head_size
= sizeof(OrderedHashmap
),
263 .entry_size
= sizeof(struct ordered_hashmap_entry
),
264 .mempool
= &ordered_hashmap_pool
,
265 .n_direct_buckets
= DIRECT_BUCKETS(struct ordered_hashmap_entry
),
267 [HASHMAP_TYPE_SET
] = {
268 .head_size
= sizeof(Set
),
269 .entry_size
= sizeof(struct set_entry
),
270 .mempool
= &hashmap_pool
,
271 .n_direct_buckets
= DIRECT_BUCKETS(struct set_entry
),
276 _destructor_
static void cleanup_pools(void) {
277 _cleanup_free_
char *t
= NULL
;
280 /* Be nice to valgrind */
282 /* The pool is only allocated by the main thread, but the memory can
283 * be passed to other threads. Let's clean up if we are the main thread
284 * and no other threads are live. */
285 /* We build our own is_main_thread() here, which doesn't use C11
286 * TLS based caching of the result. That's because valgrind apparently
287 * doesn't like malloc() (which C11 TLS internally uses) to be called
288 * from a GCC destructors. */
289 if (getpid() != gettid())
292 r
= get_proc_field("/proc/self/status", "Threads", WHITESPACE
, &t
);
293 if (r
< 0 || !streq(t
, "1"))
296 mempool_drop(&hashmap_pool
);
297 mempool_drop(&ordered_hashmap_pool
);
301 static unsigned n_buckets(HashmapBase
*h
) {
302 return h
->has_indirect
? h
->indirect
.n_buckets
303 : hashmap_type_info
[h
->type
].n_direct_buckets
;
306 static unsigned n_entries(HashmapBase
*h
) {
307 return h
->has_indirect
? h
->indirect
.n_entries
308 : h
->n_direct_entries
;
311 static void n_entries_inc(HashmapBase
*h
) {
313 h
->indirect
.n_entries
++;
315 h
->n_direct_entries
++;
318 static void n_entries_dec(HashmapBase
*h
) {
320 h
->indirect
.n_entries
--;
322 h
->n_direct_entries
--;
325 static void* storage_ptr(HashmapBase
*h
) {
326 return h
->has_indirect
? h
->indirect
.storage
330 static uint8_t* hash_key(HashmapBase
*h
) {
331 return h
->has_indirect
? h
->indirect
.hash_key
335 static unsigned base_bucket_hash(HashmapBase
*h
, const void *p
) {
336 struct siphash state
;
339 siphash24_init(&state
, hash_key(h
));
341 h
->hash_ops
->hash(p
, &state
);
343 hash
= siphash24_finalize(&state
);
345 return (unsigned) (hash
% n_buckets(h
));
347 #define bucket_hash(h, p) base_bucket_hash(HASHMAP_BASE(h), p)
349 static void base_set_dirty(HashmapBase
*h
) {
352 #define hashmap_set_dirty(h) base_set_dirty(HASHMAP_BASE(h))
354 static void get_hash_key(uint8_t hash_key
[HASH_KEY_SIZE
], bool reuse_is_ok
) {
355 static uint8_t current
[HASH_KEY_SIZE
];
356 static bool current_initialized
= false;
358 /* Returns a hash function key to use. In order to keep things
359 * fast we will not generate a new key each time we allocate a
360 * new hash table. Instead, we'll just reuse the most recently
361 * generated one, except if we never generated one or when we
362 * are rehashing an entire hash table because we reached a
365 if (!current_initialized
|| !reuse_is_ok
) {
366 random_bytes(current
, sizeof(current
));
367 current_initialized
= true;
370 memcpy(hash_key
, current
, sizeof(current
));
373 static struct hashmap_base_entry
* bucket_at(HashmapBase
*h
, unsigned idx
) {
374 return (struct hashmap_base_entry
*)
375 ((uint8_t*) storage_ptr(h
) + idx
* hashmap_type_info
[h
->type
].entry_size
);
378 static struct plain_hashmap_entry
* plain_bucket_at(Hashmap
*h
, unsigned idx
) {
379 return (struct plain_hashmap_entry
*) bucket_at(HASHMAP_BASE(h
), idx
);
382 static struct ordered_hashmap_entry
* ordered_bucket_at(OrderedHashmap
*h
, unsigned idx
) {
383 return (struct ordered_hashmap_entry
*) bucket_at(HASHMAP_BASE(h
), idx
);
386 static struct set_entry
*set_bucket_at(Set
*h
, unsigned idx
) {
387 return (struct set_entry
*) bucket_at(HASHMAP_BASE(h
), idx
);
390 static struct ordered_hashmap_entry
* bucket_at_swap(struct swap_entries
*swap
, unsigned idx
) {
391 return &swap
->e
[idx
- _IDX_SWAP_BEGIN
];
394 /* Returns a pointer to the bucket at index idx.
395 * Understands real indexes and swap indexes, hence "_virtual". */
396 static struct hashmap_base_entry
* bucket_at_virtual(HashmapBase
*h
, struct swap_entries
*swap
,
398 if (idx
< _IDX_SWAP_BEGIN
)
399 return bucket_at(h
, idx
);
401 if (idx
< _IDX_SWAP_END
)
402 return &bucket_at_swap(swap
, idx
)->p
.b
;
404 assert_not_reached("Invalid index");
407 static dib_raw_t
* dib_raw_ptr(HashmapBase
*h
) {
409 ((uint8_t*) storage_ptr(h
) + hashmap_type_info
[h
->type
].entry_size
* n_buckets(h
));
412 static unsigned bucket_distance(HashmapBase
*h
, unsigned idx
, unsigned from
) {
413 return idx
>= from
? idx
- from
414 : n_buckets(h
) + idx
- from
;
417 static unsigned bucket_calculate_dib(HashmapBase
*h
, unsigned idx
, dib_raw_t raw_dib
) {
418 unsigned initial_bucket
;
420 if (raw_dib
== DIB_RAW_FREE
)
423 if (_likely_(raw_dib
< DIB_RAW_OVERFLOW
))
427 * Having an overflow DIB value is very unlikely. The hash function
428 * would have to be bad. For example, in a table of size 2^24 filled
429 * to load factor 0.9 the maximum observed DIB is only about 60.
430 * In theory (assuming I used Maxima correctly), for an infinite size
431 * hash table with load factor 0.8 the probability of a given entry
432 * having DIB > 40 is 1.9e-8.
433 * This returns the correct DIB value by recomputing the hash value in
434 * the unlikely case. XXX Hitting this case could be a hint to rehash.
436 initial_bucket
= bucket_hash(h
, bucket_at(h
, idx
)->key
);
437 return bucket_distance(h
, idx
, initial_bucket
);
440 static void bucket_set_dib(HashmapBase
*h
, unsigned idx
, unsigned dib
) {
441 dib_raw_ptr(h
)[idx
] = dib
!= DIB_FREE
? MIN(dib
, DIB_RAW_OVERFLOW
) : DIB_RAW_FREE
;
444 static unsigned skip_free_buckets(HashmapBase
*h
, unsigned idx
) {
447 dibs
= dib_raw_ptr(h
);
449 for ( ; idx
< n_buckets(h
); idx
++)
450 if (dibs
[idx
] != DIB_RAW_FREE
)
456 static void bucket_mark_free(HashmapBase
*h
, unsigned idx
) {
457 memzero(bucket_at(h
, idx
), hashmap_type_info
[h
->type
].entry_size
);
458 bucket_set_dib(h
, idx
, DIB_FREE
);
461 static void bucket_move_entry(HashmapBase
*h
, struct swap_entries
*swap
,
462 unsigned from
, unsigned to
) {
463 struct hashmap_base_entry
*e_from
, *e_to
;
467 e_from
= bucket_at_virtual(h
, swap
, from
);
468 e_to
= bucket_at_virtual(h
, swap
, to
);
470 memcpy(e_to
, e_from
, hashmap_type_info
[h
->type
].entry_size
);
472 if (h
->type
== HASHMAP_TYPE_ORDERED
) {
473 OrderedHashmap
*lh
= (OrderedHashmap
*) h
;
474 struct ordered_hashmap_entry
*le
, *le_to
;
476 le_to
= (struct ordered_hashmap_entry
*) e_to
;
478 if (le_to
->iterate_next
!= IDX_NIL
) {
479 le
= (struct ordered_hashmap_entry
*)
480 bucket_at_virtual(h
, swap
, le_to
->iterate_next
);
481 le
->iterate_previous
= to
;
484 if (le_to
->iterate_previous
!= IDX_NIL
) {
485 le
= (struct ordered_hashmap_entry
*)
486 bucket_at_virtual(h
, swap
, le_to
->iterate_previous
);
487 le
->iterate_next
= to
;
490 if (lh
->iterate_list_head
== from
)
491 lh
->iterate_list_head
= to
;
492 if (lh
->iterate_list_tail
== from
)
493 lh
->iterate_list_tail
= to
;
497 static unsigned next_idx(HashmapBase
*h
, unsigned idx
) {
498 return (idx
+ 1U) % n_buckets(h
);
501 static unsigned prev_idx(HashmapBase
*h
, unsigned idx
) {
502 return (n_buckets(h
) + idx
- 1U) % n_buckets(h
);
505 static void* entry_value(HashmapBase
*h
, struct hashmap_base_entry
*e
) {
508 case HASHMAP_TYPE_PLAIN
:
509 case HASHMAP_TYPE_ORDERED
:
510 return ((struct plain_hashmap_entry
*)e
)->value
;
512 case HASHMAP_TYPE_SET
:
513 return (void*) e
->key
;
516 assert_not_reached("Unknown hashmap type");
520 static void base_remove_entry(HashmapBase
*h
, unsigned idx
) {
521 unsigned left
, right
, prev
, dib
;
522 dib_raw_t raw_dib
, *dibs
;
524 dibs
= dib_raw_ptr(h
);
525 assert(dibs
[idx
] != DIB_RAW_FREE
);
527 #if ENABLE_DEBUG_HASHMAP
528 h
->debug
.rem_count
++;
529 h
->debug
.last_rem_idx
= idx
;
533 /* Find the stop bucket ("right"). It is either free or has DIB == 0. */
534 for (right
= next_idx(h
, left
); ; right
= next_idx(h
, right
)) {
535 raw_dib
= dibs
[right
];
536 if (IN_SET(raw_dib
, 0, DIB_RAW_FREE
))
539 /* The buckets are not supposed to be all occupied and with DIB > 0.
540 * That would mean we could make everyone better off by shifting them
541 * backward. This scenario is impossible. */
542 assert(left
!= right
);
545 if (h
->type
== HASHMAP_TYPE_ORDERED
) {
546 OrderedHashmap
*lh
= (OrderedHashmap
*) h
;
547 struct ordered_hashmap_entry
*le
= ordered_bucket_at(lh
, idx
);
549 if (le
->iterate_next
!= IDX_NIL
)
550 ordered_bucket_at(lh
, le
->iterate_next
)->iterate_previous
= le
->iterate_previous
;
552 lh
->iterate_list_tail
= le
->iterate_previous
;
554 if (le
->iterate_previous
!= IDX_NIL
)
555 ordered_bucket_at(lh
, le
->iterate_previous
)->iterate_next
= le
->iterate_next
;
557 lh
->iterate_list_head
= le
->iterate_next
;
560 /* Now shift all buckets in the interval (left, right) one step backwards */
561 for (prev
= left
, left
= next_idx(h
, left
); left
!= right
;
562 prev
= left
, left
= next_idx(h
, left
)) {
563 dib
= bucket_calculate_dib(h
, left
, dibs
[left
]);
565 bucket_move_entry(h
, NULL
, left
, prev
);
566 bucket_set_dib(h
, prev
, dib
- 1);
569 bucket_mark_free(h
, prev
);
573 #define remove_entry(h, idx) base_remove_entry(HASHMAP_BASE(h), idx)
575 static unsigned hashmap_iterate_in_insertion_order(OrderedHashmap
*h
, Iterator
*i
) {
576 struct ordered_hashmap_entry
*e
;
582 if (i
->idx
== IDX_NIL
)
585 if (i
->idx
== IDX_FIRST
&& h
->iterate_list_head
== IDX_NIL
)
588 if (i
->idx
== IDX_FIRST
) {
589 idx
= h
->iterate_list_head
;
590 e
= ordered_bucket_at(h
, idx
);
593 e
= ordered_bucket_at(h
, idx
);
595 * We allow removing the current entry while iterating, but removal may cause
596 * a backward shift. The next entry may thus move one bucket to the left.
597 * To detect when it happens, we remember the key pointer of the entry we were
598 * going to iterate next. If it does not match, there was a backward shift.
600 if (e
->p
.b
.key
!= i
->next_key
) {
601 idx
= prev_idx(HASHMAP_BASE(h
), idx
);
602 e
= ordered_bucket_at(h
, idx
);
604 assert(e
->p
.b
.key
== i
->next_key
);
607 #if ENABLE_DEBUG_HASHMAP
611 if (e
->iterate_next
!= IDX_NIL
) {
612 struct ordered_hashmap_entry
*n
;
613 i
->idx
= e
->iterate_next
;
614 n
= ordered_bucket_at(h
, i
->idx
);
615 i
->next_key
= n
->p
.b
.key
;
626 static unsigned hashmap_iterate_in_internal_order(HashmapBase
*h
, Iterator
*i
) {
632 if (i
->idx
== IDX_NIL
)
635 if (i
->idx
== IDX_FIRST
) {
636 /* fast forward to the first occupied bucket */
637 if (h
->has_indirect
) {
638 i
->idx
= skip_free_buckets(h
, h
->indirect
.idx_lowest_entry
);
639 h
->indirect
.idx_lowest_entry
= i
->idx
;
641 i
->idx
= skip_free_buckets(h
, 0);
643 if (i
->idx
== IDX_NIL
)
646 struct hashmap_base_entry
*e
;
650 e
= bucket_at(h
, i
->idx
);
652 * We allow removing the current entry while iterating, but removal may cause
653 * a backward shift. The next entry may thus move one bucket to the left.
654 * To detect when it happens, we remember the key pointer of the entry we were
655 * going to iterate next. If it does not match, there was a backward shift.
657 if (e
->key
!= i
->next_key
)
658 e
= bucket_at(h
, --i
->idx
);
660 assert(e
->key
== i
->next_key
);
664 #if ENABLE_DEBUG_HASHMAP
668 i
->idx
= skip_free_buckets(h
, i
->idx
+ 1);
669 if (i
->idx
!= IDX_NIL
)
670 i
->next_key
= bucket_at(h
, i
->idx
)->key
;
681 static unsigned hashmap_iterate_entry(HashmapBase
*h
, Iterator
*i
) {
687 #if ENABLE_DEBUG_HASHMAP
688 if (i
->idx
== IDX_FIRST
) {
689 i
->put_count
= h
->debug
.put_count
;
690 i
->rem_count
= h
->debug
.rem_count
;
692 /* While iterating, must not add any new entries */
693 assert(i
->put_count
== h
->debug
.put_count
);
694 /* ... or remove entries other than the current one */
695 assert(i
->rem_count
== h
->debug
.rem_count
||
696 (i
->rem_count
== h
->debug
.rem_count
- 1 &&
697 i
->prev_idx
== h
->debug
.last_rem_idx
));
698 /* Reset our removals counter */
699 i
->rem_count
= h
->debug
.rem_count
;
703 return h
->type
== HASHMAP_TYPE_ORDERED
? hashmap_iterate_in_insertion_order((OrderedHashmap
*) h
, i
)
704 : hashmap_iterate_in_internal_order(h
, i
);
707 bool _hashmap_iterate(HashmapBase
*h
, Iterator
*i
, void **value
, const void **key
) {
708 struct hashmap_base_entry
*e
;
712 idx
= hashmap_iterate_entry(h
, i
);
713 if (idx
== IDX_NIL
) {
722 e
= bucket_at(h
, idx
);
723 data
= entry_value(h
, e
);
732 #define HASHMAP_FOREACH_IDX(idx, h, i) \
733 for ((i) = ITERATOR_FIRST, (idx) = hashmap_iterate_entry((h), &(i)); \
735 (idx) = hashmap_iterate_entry((h), &(i)))
737 IteratedCache
* _hashmap_iterated_cache_new(HashmapBase
*h
) {
738 IteratedCache
*cache
;
746 cache
= new0(IteratedCache
, 1);
756 static void reset_direct_storage(HashmapBase
*h
) {
757 const struct hashmap_type_info
*hi
= &hashmap_type_info
[h
->type
];
760 assert(!h
->has_indirect
);
762 p
= mempset(h
->direct
.storage
, 0, hi
->entry_size
* hi
->n_direct_buckets
);
763 memset(p
, DIB_RAW_INIT
, sizeof(dib_raw_t
) * hi
->n_direct_buckets
);
766 static void shared_hash_key_initialize(void) {
767 random_bytes(shared_hash_key
, sizeof(shared_hash_key
));
770 static struct HashmapBase
* hashmap_base_new(const struct hash_ops
*hash_ops
, enum HashmapType type HASHMAP_DEBUG_PARAMS
) {
772 const struct hashmap_type_info
*hi
= &hashmap_type_info
[type
];
775 up
= mempool_enabled();
777 h
= up
? mempool_alloc0_tile(hi
->mempool
) : malloc0(hi
->head_size
);
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 _ordered_hashmap_ensure_put(OrderedHashmap
**h
, const struct hash_ops
*hash_ops
, const void *key
, void *value HASHMAP_DEBUG_PARAMS
) {
851 r
= _ordered_hashmap_ensure_allocated(h
, hash_ops HASHMAP_DEBUG_PASS_ARGS
);
855 return ordered_hashmap_put(*h
, key
, value
);
858 static void hashmap_free_no_clear(HashmapBase
*h
) {
859 assert(!h
->has_indirect
);
860 assert(h
->n_direct_entries
== 0);
862 #if ENABLE_DEBUG_HASHMAP
863 assert_se(pthread_mutex_lock(&hashmap_debug_list_mutex
) == 0);
864 LIST_REMOVE(debug_list
, hashmap_debug_list
, &h
->debug
);
865 assert_se(pthread_mutex_unlock(&hashmap_debug_list_mutex
) == 0);
869 /* Ensure that the object didn't get migrated between threads. */
870 assert_se(is_main_thread());
871 mempool_free_tile(hashmap_type_info
[h
->type
].mempool
, h
);
876 HashmapBase
* _hashmap_free(HashmapBase
*h
, free_func_t default_free_key
, free_func_t default_free_value
) {
878 _hashmap_clear(h
, default_free_key
, default_free_value
);
879 hashmap_free_no_clear(h
);
885 void _hashmap_clear(HashmapBase
*h
, free_func_t default_free_key
, free_func_t default_free_value
) {
886 free_func_t free_key
, free_value
;
890 free_key
= h
->hash_ops
->free_key
?: default_free_key
;
891 free_value
= h
->hash_ops
->free_value
?: default_free_value
;
893 if (free_key
|| free_value
) {
895 /* If destructor calls are defined, let's destroy things defensively: let's take the item out of the
896 * hash table, and only then call the destructor functions. If these destructors then try to unregister
897 * themselves from our hash table a second time, the entry is already gone. */
899 while (_hashmap_size(h
) > 0) {
903 v
= _hashmap_first_key_and_value(h
, true, &k
);
913 if (h
->has_indirect
) {
914 free(h
->indirect
.storage
);
915 h
->has_indirect
= false;
918 h
->n_direct_entries
= 0;
919 reset_direct_storage(h
);
921 if (h
->type
== HASHMAP_TYPE_ORDERED
) {
922 OrderedHashmap
*lh
= (OrderedHashmap
*) h
;
923 lh
->iterate_list_head
= lh
->iterate_list_tail
= IDX_NIL
;
929 static int resize_buckets(HashmapBase
*h
, unsigned entries_add
);
932 * Finds an empty bucket to put an entry into, starting the scan at 'idx'.
933 * Performs Robin Hood swaps as it goes. The entry to put must be placed
934 * by the caller into swap slot IDX_PUT.
935 * If used for in-place resizing, may leave a displaced entry in swap slot
936 * IDX_PUT. Caller must rehash it next.
937 * Returns: true if it left a displaced entry to rehash next in IDX_PUT,
940 static bool hashmap_put_robin_hood(HashmapBase
*h
, unsigned idx
,
941 struct swap_entries
*swap
) {
942 dib_raw_t raw_dib
, *dibs
;
943 unsigned dib
, distance
;
945 #if ENABLE_DEBUG_HASHMAP
946 h
->debug
.put_count
++;
949 dibs
= dib_raw_ptr(h
);
951 for (distance
= 0; ; distance
++) {
953 if (IN_SET(raw_dib
, DIB_RAW_FREE
, DIB_RAW_REHASH
)) {
954 if (raw_dib
== DIB_RAW_REHASH
)
955 bucket_move_entry(h
, swap
, idx
, IDX_TMP
);
957 if (h
->has_indirect
&& h
->indirect
.idx_lowest_entry
> idx
)
958 h
->indirect
.idx_lowest_entry
= idx
;
960 bucket_set_dib(h
, idx
, distance
);
961 bucket_move_entry(h
, swap
, IDX_PUT
, idx
);
962 if (raw_dib
== DIB_RAW_REHASH
) {
963 bucket_move_entry(h
, swap
, IDX_TMP
, IDX_PUT
);
970 dib
= bucket_calculate_dib(h
, idx
, raw_dib
);
972 if (dib
< distance
) {
973 /* Found a wealthier entry. Go Robin Hood! */
974 bucket_set_dib(h
, idx
, distance
);
976 /* swap the entries */
977 bucket_move_entry(h
, swap
, idx
, IDX_TMP
);
978 bucket_move_entry(h
, swap
, IDX_PUT
, idx
);
979 bucket_move_entry(h
, swap
, IDX_TMP
, IDX_PUT
);
984 idx
= next_idx(h
, idx
);
989 * Puts an entry into a hashmap, boldly - no check whether key already exists.
990 * The caller must place the entry (only its key and value, not link indexes)
991 * in swap slot IDX_PUT.
992 * Caller must ensure: the key does not exist yet in the hashmap.
993 * that resize is not needed if !may_resize.
994 * Returns: 1 if entry was put successfully.
995 * -ENOMEM if may_resize==true and resize failed with -ENOMEM.
996 * Cannot return -ENOMEM if !may_resize.
998 static int hashmap_base_put_boldly(HashmapBase
*h
, unsigned idx
,
999 struct swap_entries
*swap
, bool may_resize
) {
1000 struct ordered_hashmap_entry
*new_entry
;
1003 assert(idx
< n_buckets(h
));
1005 new_entry
= bucket_at_swap(swap
, IDX_PUT
);
1008 r
= resize_buckets(h
, 1);
1012 idx
= bucket_hash(h
, new_entry
->p
.b
.key
);
1014 assert(n_entries(h
) < n_buckets(h
));
1016 if (h
->type
== HASHMAP_TYPE_ORDERED
) {
1017 OrderedHashmap
*lh
= (OrderedHashmap
*) h
;
1019 new_entry
->iterate_next
= IDX_NIL
;
1020 new_entry
->iterate_previous
= lh
->iterate_list_tail
;
1022 if (lh
->iterate_list_tail
!= IDX_NIL
) {
1023 struct ordered_hashmap_entry
*old_tail
;
1025 old_tail
= ordered_bucket_at(lh
, lh
->iterate_list_tail
);
1026 assert(old_tail
->iterate_next
== IDX_NIL
);
1027 old_tail
->iterate_next
= IDX_PUT
;
1030 lh
->iterate_list_tail
= IDX_PUT
;
1031 if (lh
->iterate_list_head
== IDX_NIL
)
1032 lh
->iterate_list_head
= IDX_PUT
;
1035 assert_se(hashmap_put_robin_hood(h
, idx
, swap
) == false);
1038 #if ENABLE_DEBUG_HASHMAP
1039 h
->debug
.max_entries
= MAX(h
->debug
.max_entries
, n_entries(h
));
1046 #define hashmap_put_boldly(h, idx, swap, may_resize) \
1047 hashmap_base_put_boldly(HASHMAP_BASE(h), idx, swap, may_resize)
1050 * Returns 0 if resize is not needed.
1051 * 1 if successfully resized.
1052 * -ENOMEM on allocation failure.
1054 static int resize_buckets(HashmapBase
*h
, unsigned entries_add
) {
1055 struct swap_entries swap
;
1057 dib_raw_t
*old_dibs
, *new_dibs
;
1058 const struct hashmap_type_info
*hi
;
1059 unsigned idx
, optimal_idx
;
1060 unsigned old_n_buckets
, new_n_buckets
, n_rehashed
, new_n_entries
;
1066 hi
= &hashmap_type_info
[h
->type
];
1067 new_n_entries
= n_entries(h
) + entries_add
;
1070 if (_unlikely_(new_n_entries
< entries_add
))
1073 /* For direct storage we allow 100% load, because it's tiny. */
1074 if (!h
->has_indirect
&& new_n_entries
<= hi
->n_direct_buckets
)
1078 * Load factor = n/m = 1 - (1/INV_KEEP_FREE).
1079 * From it follows: m = n + n/(INV_KEEP_FREE - 1)
1081 new_n_buckets
= new_n_entries
+ new_n_entries
/ (INV_KEEP_FREE
- 1);
1083 if (_unlikely_(new_n_buckets
< new_n_entries
))
1086 if (_unlikely_(new_n_buckets
> UINT_MAX
/ (hi
->entry_size
+ sizeof(dib_raw_t
))))
1089 old_n_buckets
= n_buckets(h
);
1091 if (_likely_(new_n_buckets
<= old_n_buckets
))
1094 new_shift
= log2u_round_up(MAX(
1095 new_n_buckets
* (hi
->entry_size
+ sizeof(dib_raw_t
)),
1096 2 * sizeof(struct direct_storage
)));
1098 /* Realloc storage (buckets and DIB array). */
1099 new_storage
= realloc(h
->has_indirect
? h
->indirect
.storage
: NULL
,
1104 /* Must upgrade direct to indirect storage. */
1105 if (!h
->has_indirect
) {
1106 memcpy(new_storage
, h
->direct
.storage
,
1107 old_n_buckets
* (hi
->entry_size
+ sizeof(dib_raw_t
)));
1108 h
->indirect
.n_entries
= h
->n_direct_entries
;
1109 h
->indirect
.idx_lowest_entry
= 0;
1110 h
->n_direct_entries
= 0;
1113 /* Get a new hash key. If we've just upgraded to indirect storage,
1114 * allow reusing a previously generated key. It's still a different key
1115 * from the shared one that we used for direct storage. */
1116 get_hash_key(h
->indirect
.hash_key
, !h
->has_indirect
);
1118 h
->has_indirect
= true;
1119 h
->indirect
.storage
= new_storage
;
1120 h
->indirect
.n_buckets
= (1U << new_shift
) /
1121 (hi
->entry_size
+ sizeof(dib_raw_t
));
1123 old_dibs
= (dib_raw_t
*)((uint8_t*) new_storage
+ hi
->entry_size
* old_n_buckets
);
1124 new_dibs
= dib_raw_ptr(h
);
1127 * Move the DIB array to the new place, replacing valid DIB values with
1128 * DIB_RAW_REHASH to indicate all of the used buckets need rehashing.
1129 * Note: Overlap is not possible, because we have at least doubled the
1130 * number of buckets and dib_raw_t is smaller than any entry type.
1132 for (idx
= 0; idx
< old_n_buckets
; idx
++) {
1133 assert(old_dibs
[idx
] != DIB_RAW_REHASH
);
1134 new_dibs
[idx
] = old_dibs
[idx
] == DIB_RAW_FREE
? DIB_RAW_FREE
1138 /* Zero the area of newly added entries (including the old DIB area) */
1139 memzero(bucket_at(h
, old_n_buckets
),
1140 (n_buckets(h
) - old_n_buckets
) * hi
->entry_size
);
1142 /* The upper half of the new DIB array needs initialization */
1143 memset(&new_dibs
[old_n_buckets
], DIB_RAW_INIT
,
1144 (n_buckets(h
) - old_n_buckets
) * sizeof(dib_raw_t
));
1146 /* Rehash entries that need it */
1148 for (idx
= 0; idx
< old_n_buckets
; idx
++) {
1149 if (new_dibs
[idx
] != DIB_RAW_REHASH
)
1152 optimal_idx
= bucket_hash(h
, bucket_at(h
, idx
)->key
);
1155 * Not much to do if by luck the entry hashes to its current
1156 * location. Just set its DIB.
1158 if (optimal_idx
== idx
) {
1164 new_dibs
[idx
] = DIB_RAW_FREE
;
1165 bucket_move_entry(h
, &swap
, idx
, IDX_PUT
);
1166 /* bucket_move_entry does not clear the source */
1167 memzero(bucket_at(h
, idx
), hi
->entry_size
);
1171 * Find the new bucket for the current entry. This may make
1172 * another entry homeless and load it into IDX_PUT.
1174 rehash_next
= hashmap_put_robin_hood(h
, optimal_idx
, &swap
);
1177 /* Did the current entry displace another one? */
1179 optimal_idx
= bucket_hash(h
, bucket_at_swap(&swap
, IDX_PUT
)->p
.b
.key
);
1180 } while (rehash_next
);
1183 assert(n_rehashed
== n_entries(h
));
1189 * Finds an entry with a matching key
1190 * Returns: index of the found entry, or IDX_NIL if not found.
1192 static unsigned base_bucket_scan(HashmapBase
*h
, unsigned idx
, const void *key
) {
1193 struct hashmap_base_entry
*e
;
1194 unsigned dib
, distance
;
1195 dib_raw_t
*dibs
= dib_raw_ptr(h
);
1197 assert(idx
< n_buckets(h
));
1199 for (distance
= 0; ; distance
++) {
1200 if (dibs
[idx
] == DIB_RAW_FREE
)
1203 dib
= bucket_calculate_dib(h
, idx
, dibs
[idx
]);
1207 if (dib
== distance
) {
1208 e
= bucket_at(h
, idx
);
1209 if (h
->hash_ops
->compare(e
->key
, key
) == 0)
1213 idx
= next_idx(h
, idx
);
1216 #define bucket_scan(h, idx, key) base_bucket_scan(HASHMAP_BASE(h), idx, key)
1218 int hashmap_put(Hashmap
*h
, const void *key
, void *value
) {
1219 struct swap_entries swap
;
1220 struct plain_hashmap_entry
*e
;
1225 hash
= bucket_hash(h
, key
);
1226 idx
= bucket_scan(h
, hash
, key
);
1227 if (idx
!= IDX_NIL
) {
1228 e
= plain_bucket_at(h
, idx
);
1229 if (e
->value
== value
)
1234 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
;
1237 return hashmap_put_boldly(h
, hash
, &swap
, true);
1240 int set_put(Set
*s
, const void *key
) {
1241 struct swap_entries swap
;
1242 struct hashmap_base_entry
*e
;
1247 hash
= bucket_hash(s
, key
);
1248 idx
= bucket_scan(s
, hash
, key
);
1252 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
.b
;
1254 return hashmap_put_boldly(s
, hash
, &swap
, true);
1257 int _set_ensure_put(Set
**s
, const struct hash_ops
*hash_ops
, const void *key HASHMAP_DEBUG_PARAMS
) {
1260 r
= _set_ensure_allocated(s
, hash_ops HASHMAP_DEBUG_PASS_ARGS
);
1264 return set_put(*s
, key
);
1267 int _set_ensure_consume(Set
**s
, const struct hash_ops
*hash_ops
, void *key HASHMAP_DEBUG_PARAMS
) {
1270 r
= _set_ensure_put(s
, hash_ops
, key HASHMAP_DEBUG_PASS_ARGS
);
1272 if (hash_ops
&& hash_ops
->free_key
)
1273 hash_ops
->free_key(key
);
1281 int hashmap_replace(Hashmap
*h
, const void *key
, void *value
) {
1282 struct swap_entries swap
;
1283 struct plain_hashmap_entry
*e
;
1288 hash
= bucket_hash(h
, key
);
1289 idx
= bucket_scan(h
, hash
, key
);
1290 if (idx
!= IDX_NIL
) {
1291 e
= plain_bucket_at(h
, idx
);
1292 #if ENABLE_DEBUG_HASHMAP
1293 /* Although the key is equal, the key pointer may have changed,
1294 * and this would break our assumption for iterating. So count
1295 * this operation as incompatible with iteration. */
1296 if (e
->b
.key
!= key
) {
1297 h
->b
.debug
.put_count
++;
1298 h
->b
.debug
.rem_count
++;
1299 h
->b
.debug
.last_rem_idx
= idx
;
1304 hashmap_set_dirty(h
);
1309 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
;
1312 return hashmap_put_boldly(h
, hash
, &swap
, true);
1315 int hashmap_update(Hashmap
*h
, const void *key
, void *value
) {
1316 struct plain_hashmap_entry
*e
;
1321 hash
= bucket_hash(h
, key
);
1322 idx
= bucket_scan(h
, hash
, key
);
1326 e
= plain_bucket_at(h
, idx
);
1328 hashmap_set_dirty(h
);
1333 void* _hashmap_get(HashmapBase
*h
, const void *key
) {
1334 struct hashmap_base_entry
*e
;
1340 hash
= bucket_hash(h
, key
);
1341 idx
= bucket_scan(h
, hash
, key
);
1345 e
= bucket_at(h
, idx
);
1346 return entry_value(h
, e
);
1349 void* hashmap_get2(Hashmap
*h
, const void *key
, void **key2
) {
1350 struct plain_hashmap_entry
*e
;
1356 hash
= bucket_hash(h
, key
);
1357 idx
= bucket_scan(h
, hash
, key
);
1361 e
= plain_bucket_at(h
, idx
);
1363 *key2
= (void*) e
->b
.key
;
1368 bool _hashmap_contains(HashmapBase
*h
, const void *key
) {
1374 hash
= bucket_hash(h
, key
);
1375 return bucket_scan(h
, hash
, key
) != IDX_NIL
;
1378 void* _hashmap_remove(HashmapBase
*h
, const void *key
) {
1379 struct hashmap_base_entry
*e
;
1386 hash
= bucket_hash(h
, key
);
1387 idx
= bucket_scan(h
, hash
, key
);
1391 e
= bucket_at(h
, idx
);
1392 data
= entry_value(h
, e
);
1393 remove_entry(h
, idx
);
1398 void* hashmap_remove2(Hashmap
*h
, const void *key
, void **rkey
) {
1399 struct plain_hashmap_entry
*e
;
1409 hash
= bucket_hash(h
, key
);
1410 idx
= bucket_scan(h
, hash
, key
);
1411 if (idx
== IDX_NIL
) {
1417 e
= plain_bucket_at(h
, idx
);
1420 *rkey
= (void*) e
->b
.key
;
1422 remove_entry(h
, idx
);
1427 int hashmap_remove_and_put(Hashmap
*h
, const void *old_key
, const void *new_key
, void *value
) {
1428 struct swap_entries swap
;
1429 struct plain_hashmap_entry
*e
;
1430 unsigned old_hash
, new_hash
, idx
;
1435 old_hash
= bucket_hash(h
, old_key
);
1436 idx
= bucket_scan(h
, old_hash
, old_key
);
1440 new_hash
= bucket_hash(h
, new_key
);
1441 if (bucket_scan(h
, new_hash
, new_key
) != IDX_NIL
)
1444 remove_entry(h
, idx
);
1446 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
;
1449 assert_se(hashmap_put_boldly(h
, new_hash
, &swap
, false) == 1);
1454 int set_remove_and_put(Set
*s
, const void *old_key
, const void *new_key
) {
1455 struct swap_entries swap
;
1456 struct hashmap_base_entry
*e
;
1457 unsigned old_hash
, new_hash
, idx
;
1462 old_hash
= bucket_hash(s
, old_key
);
1463 idx
= bucket_scan(s
, old_hash
, old_key
);
1467 new_hash
= bucket_hash(s
, new_key
);
1468 if (bucket_scan(s
, new_hash
, new_key
) != IDX_NIL
)
1471 remove_entry(s
, idx
);
1473 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
.b
;
1475 assert_se(hashmap_put_boldly(s
, new_hash
, &swap
, false) == 1);
1480 int hashmap_remove_and_replace(Hashmap
*h
, const void *old_key
, const void *new_key
, void *value
) {
1481 struct swap_entries swap
;
1482 struct plain_hashmap_entry
*e
;
1483 unsigned old_hash
, new_hash
, idx_old
, idx_new
;
1488 old_hash
= bucket_hash(h
, old_key
);
1489 idx_old
= bucket_scan(h
, old_hash
, old_key
);
1490 if (idx_old
== IDX_NIL
)
1493 old_key
= bucket_at(HASHMAP_BASE(h
), idx_old
)->key
;
1495 new_hash
= bucket_hash(h
, new_key
);
1496 idx_new
= bucket_scan(h
, new_hash
, new_key
);
1497 if (idx_new
!= IDX_NIL
)
1498 if (idx_old
!= idx_new
) {
1499 remove_entry(h
, idx_new
);
1500 /* Compensate for a possible backward shift. */
1501 if (old_key
!= bucket_at(HASHMAP_BASE(h
), idx_old
)->key
)
1502 idx_old
= prev_idx(HASHMAP_BASE(h
), idx_old
);
1503 assert(old_key
== bucket_at(HASHMAP_BASE(h
), idx_old
)->key
);
1506 remove_entry(h
, idx_old
);
1508 e
= &bucket_at_swap(&swap
, IDX_PUT
)->p
;
1511 assert_se(hashmap_put_boldly(h
, new_hash
, &swap
, false) == 1);
1516 void* _hashmap_remove_value(HashmapBase
*h
, const void *key
, void *value
) {
1517 struct hashmap_base_entry
*e
;
1523 hash
= bucket_hash(h
, key
);
1524 idx
= bucket_scan(h
, hash
, key
);
1528 e
= bucket_at(h
, idx
);
1529 if (entry_value(h
, e
) != value
)
1532 remove_entry(h
, idx
);
1537 static unsigned find_first_entry(HashmapBase
*h
) {
1538 Iterator i
= ITERATOR_FIRST
;
1540 if (!h
|| !n_entries(h
))
1543 return hashmap_iterate_entry(h
, &i
);
1546 void* _hashmap_first_key_and_value(HashmapBase
*h
, bool remove
, void **ret_key
) {
1547 struct hashmap_base_entry
*e
;
1551 idx
= find_first_entry(h
);
1552 if (idx
== IDX_NIL
) {
1558 e
= bucket_at(h
, idx
);
1559 key
= (void*) e
->key
;
1560 data
= entry_value(h
, e
);
1563 remove_entry(h
, idx
);
1571 unsigned _hashmap_size(HashmapBase
*h
) {
1575 return n_entries(h
);
1578 unsigned _hashmap_buckets(HashmapBase
*h
) {
1582 return n_buckets(h
);
1585 int _hashmap_merge(Hashmap
*h
, Hashmap
*other
) {
1591 HASHMAP_FOREACH_IDX(idx
, HASHMAP_BASE(other
), i
) {
1592 struct plain_hashmap_entry
*pe
= plain_bucket_at(other
, idx
);
1595 r
= hashmap_put(h
, pe
->b
.key
, pe
->value
);
1596 if (r
< 0 && r
!= -EEXIST
)
1603 int set_merge(Set
*s
, Set
*other
) {
1609 HASHMAP_FOREACH_IDX(idx
, HASHMAP_BASE(other
), i
) {
1610 struct set_entry
*se
= set_bucket_at(other
, idx
);
1613 r
= set_put(s
, se
->b
.key
);
1621 int _hashmap_reserve(HashmapBase
*h
, unsigned entries_add
) {
1626 r
= resize_buckets(h
, entries_add
);
1634 * The same as hashmap_merge(), but every new item from other is moved to h.
1635 * Keys already in h are skipped and stay in other.
1636 * Returns: 0 on success.
1637 * -ENOMEM on alloc failure, in which case no move has been done.
1639 int _hashmap_move(HashmapBase
*h
, HashmapBase
*other
) {
1640 struct swap_entries swap
;
1641 struct hashmap_base_entry
*e
, *n
;
1651 assert(other
->type
== h
->type
);
1654 * This reserves buckets for the worst case, where none of other's
1655 * entries are yet present in h. This is preferable to risking
1656 * an allocation failure in the middle of the moving and having to
1657 * rollback or return a partial result.
1659 r
= resize_buckets(h
, n_entries(other
));
1663 HASHMAP_FOREACH_IDX(idx
, other
, i
) {
1666 e
= bucket_at(other
, idx
);
1667 h_hash
= bucket_hash(h
, e
->key
);
1668 if (bucket_scan(h
, h_hash
, e
->key
) != IDX_NIL
)
1671 n
= &bucket_at_swap(&swap
, IDX_PUT
)->p
.b
;
1673 if (h
->type
!= HASHMAP_TYPE_SET
)
1674 ((struct plain_hashmap_entry
*) n
)->value
=
1675 ((struct plain_hashmap_entry
*) e
)->value
;
1676 assert_se(hashmap_put_boldly(h
, h_hash
, &swap
, false) == 1);
1678 remove_entry(other
, idx
);
1684 int _hashmap_move_one(HashmapBase
*h
, HashmapBase
*other
, const void *key
) {
1685 struct swap_entries swap
;
1686 unsigned h_hash
, other_hash
, idx
;
1687 struct hashmap_base_entry
*e
, *n
;
1692 h_hash
= bucket_hash(h
, key
);
1693 if (bucket_scan(h
, h_hash
, key
) != IDX_NIL
)
1699 assert(other
->type
== h
->type
);
1701 other_hash
= bucket_hash(other
, key
);
1702 idx
= bucket_scan(other
, other_hash
, key
);
1706 e
= bucket_at(other
, idx
);
1708 n
= &bucket_at_swap(&swap
, IDX_PUT
)->p
.b
;
1710 if (h
->type
!= HASHMAP_TYPE_SET
)
1711 ((struct plain_hashmap_entry
*) n
)->value
=
1712 ((struct plain_hashmap_entry
*) e
)->value
;
1713 r
= hashmap_put_boldly(h
, h_hash
, &swap
, true);
1717 remove_entry(other
, idx
);
1721 HashmapBase
* _hashmap_copy(HashmapBase
*h HASHMAP_DEBUG_PARAMS
) {
1727 copy
= hashmap_base_new(h
->hash_ops
, h
->type HASHMAP_DEBUG_PASS_ARGS
);
1732 case HASHMAP_TYPE_PLAIN
:
1733 case HASHMAP_TYPE_ORDERED
:
1734 r
= hashmap_merge((Hashmap
*)copy
, (Hashmap
*)h
);
1736 case HASHMAP_TYPE_SET
:
1737 r
= set_merge((Set
*)copy
, (Set
*)h
);
1740 assert_not_reached("Unknown hashmap type");
1744 return _hashmap_free(copy
, false, false);
1749 char** _hashmap_get_strv(HashmapBase
*h
) {
1754 sv
= new(char*, n_entries(h
)+1);
1759 HASHMAP_FOREACH_IDX(idx
, h
, i
)
1760 sv
[n
++] = entry_value(h
, bucket_at(h
, idx
));
1766 void* ordered_hashmap_next(OrderedHashmap
*h
, const void *key
) {
1767 struct ordered_hashmap_entry
*e
;
1773 hash
= bucket_hash(h
, key
);
1774 idx
= bucket_scan(h
, hash
, key
);
1778 e
= ordered_bucket_at(h
, idx
);
1779 if (e
->iterate_next
== IDX_NIL
)
1781 return ordered_bucket_at(h
, e
->iterate_next
)->p
.value
;
1784 int set_consume(Set
*s
, void *value
) {
1790 r
= set_put(s
, value
);
1797 int _hashmap_put_strdup(Hashmap
**h
, const char *k
, const char *v HASHMAP_DEBUG_PARAMS
) {
1800 r
= _hashmap_ensure_allocated(h
, &string_hash_ops_free_free HASHMAP_DEBUG_PASS_ARGS
);
1804 _cleanup_free_
char *kdup
= NULL
, *vdup
= NULL
;
1816 r
= hashmap_put(*h
, kdup
, vdup
);
1818 if (r
== -EEXIST
&& streq_ptr(v
, hashmap_get(*h
, kdup
)))
1823 /* 0 with non-null vdup would mean vdup is already in the hashmap, which cannot be */
1824 assert(vdup
== NULL
|| r
> 0);
1831 int _set_put_strdup(Set
**s
, const char *p HASHMAP_DEBUG_PARAMS
) {
1838 r
= _set_ensure_allocated(s
, &string_hash_ops_free HASHMAP_DEBUG_PASS_ARGS
);
1842 if (set_contains(*s
, (char*) p
))
1849 return set_consume(*s
, c
);
1852 int _set_put_strdupv(Set
**s
, char **l HASHMAP_DEBUG_PARAMS
) {
1858 STRV_FOREACH(i
, l
) {
1859 r
= _set_put_strdup(s
, *i HASHMAP_DEBUG_PASS_ARGS
);
1869 int set_put_strsplit(Set
*s
, const char *v
, const char *separators
, ExtractFlags flags
) {
1879 r
= extract_first_word(&p
, &word
, separators
, flags
);
1883 r
= set_consume(s
, word
);
1889 /* expand the cachemem if needed, return true if newly (re)activated. */
1890 static int cachemem_maintain(CacheMem
*mem
, unsigned size
) {
1893 if (!GREEDY_REALLOC(mem
->ptr
, mem
->n_allocated
, size
)) {
1906 int iterated_cache_get(IteratedCache
*cache
, const void ***res_keys
, const void ***res_values
, unsigned *res_n_entries
) {
1907 bool sync_keys
= false, sync_values
= false;
1912 assert(cache
->hashmap
);
1914 size
= n_entries(cache
->hashmap
);
1917 r
= cachemem_maintain(&cache
->keys
, size
);
1923 cache
->keys
.active
= false;
1926 r
= cachemem_maintain(&cache
->values
, size
);
1932 cache
->values
.active
= false;
1934 if (cache
->hashmap
->dirty
) {
1935 if (cache
->keys
.active
)
1937 if (cache
->values
.active
)
1940 cache
->hashmap
->dirty
= false;
1943 if (sync_keys
|| sync_values
) {
1948 HASHMAP_FOREACH_IDX(idx
, cache
->hashmap
, iter
) {
1949 struct hashmap_base_entry
*e
;
1951 e
= bucket_at(cache
->hashmap
, idx
);
1954 cache
->keys
.ptr
[i
] = e
->key
;
1956 cache
->values
.ptr
[i
] = entry_value(cache
->hashmap
, e
);
1962 *res_keys
= cache
->keys
.ptr
;
1964 *res_values
= cache
->values
.ptr
;
1966 *res_n_entries
= size
;
1971 IteratedCache
* iterated_cache_free(IteratedCache
*cache
) {
1973 free(cache
->keys
.ptr
);
1974 free(cache
->values
.ptr
);
1977 return mfree(cache
);