Debian Codegraph @

src/backend/utils/hash/dynahash.c

raw
1 /*-------------------------------------------------------------------------
2 *
3 * dynahash.c
4 * dynamic chained hash tables
5 *
6 * dynahash.c supports both local-to-a-backend hash tables and hash tables in
7 * shared memory. For shared hash tables, it is the caller's responsibility
8 * to provide appropriate access interlocking. The simplest convention is
9 * that a single LWLock protects the whole hash table. Searches (HASH_FIND or
10 * hash_seq_search) need only shared lock, but any update requires exclusive
11 * lock. For heavily-used shared tables, the single-lock approach creates a
12 * concurrency bottleneck, so we also support "partitioned" locking wherein
13 * there are multiple LWLocks guarding distinct subsets of the table. To use
14 * a hash table in partitioned mode, the HASH_PARTITION flag must be given
15 * to hash_create. This prevents any attempt to split buckets on-the-fly.
16 * Therefore, each hash bucket chain operates independently, and no fields
17 * of the hash header change after init except nentries and freeList.
18 * (A partitioned table uses multiple copies of those fields, guarded by
19 * spinlocks, for additional concurrency.)
20 * This lets any subset of the hash buckets be treated as a separately
21 * lockable partition. We expect callers to use the low-order bits of a
22 * lookup key's hash value as a partition number --- this will work because
23 * of the way calc_bucket() maps hash values to bucket numbers.
24 *
25 * For hash tables in shared memory, the memory allocator function should
26 * match malloc's semantics of returning NULL on failure. For hash tables
27 * in local memory, we typically use palloc() which will throw error on
28 * failure. The code in this file has to cope with both cases.
29 *
30 * dynahash.c provides support for these types of lookup keys:
31 *
32 * 1. Null-terminated C strings (truncated if necessary to fit in keysize),
33 * compared as though by strcmp(). This is selected by specifying the
34 * HASH_STRINGS flag to hash_create.
35 *
36 * 2. Arbitrary binary data of size keysize, compared as though by memcmp().
37 * (Caller must ensure there are no undefined padding bits in the keys!)
38 * This is selected by specifying the HASH_BLOBS flag to hash_create.
39 *
40 * 3. More complex key behavior can be selected by specifying user-supplied
41 * hashing, comparison, and/or key-copying functions. At least a hashing
42 * function must be supplied; comparison defaults to memcmp() and key copying
43 * to memcpy() when a user-defined hashing function is selected.
44 *
45 * Compared to simplehash, dynahash has the following benefits:
46 *
47 * - It supports partitioning, which is useful for shared memory access using
48 * locks.
49 * - Shared memory hashes are allocated in a fixed size area at startup and
50 * are discoverable by name from other processes.
51 * - Because entries don't need to be moved in the case of hash conflicts,
52 * dynahash has better performance for large entries.
53 * - Guarantees stable pointers to entries.
54 *
55 * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
56 * Portions Copyright (c) 1994, Regents of the University of California
57 *
58 *
59 * IDENTIFICATION
60 * src/backend/utils/hash/dynahash.c
61 *
62 *-------------------------------------------------------------------------
63 */
64
65 /*
66 * Original comments:
67 *
68 * Dynamic hashing, after CACM April 1988 pp 446-457, by Per-Ake Larson.
69 * Coded into C, with minor code improvements, and with hsearch(3) interface,
70 * by [email protected], Jul 26, 1988: 13:16;
71 * also, hcreate/hdestroy routines added to simulate hsearch(3).
72 *
73 * These routines simulate hsearch(3) and family, with the important
74 * difference that the hash table is dynamic - can grow indefinitely
75 * beyond its original size (as supplied to hcreate()).
76 *
77 * Performance appears to be comparable to that of hsearch(3).
78 * The 'source-code' options referred to in hsearch(3)'s 'man' page
79 * are not implemented; otherwise functionality is identical.
80 *
81 * Compilation controls:
82 * HASH_DEBUG controls some informative traces, mainly for debugging.
83 * HASH_STATISTICS causes HashAccesses and HashCollisions to be maintained;
84 * when combined with HASH_DEBUG, these are displayed by hdestroy().
85 *
86 * Problems & fixes to [email protected]. WARNING: relies on pre-processor
87 * concatenation property, in probably unnecessary code 'optimization'.
88 *
89 * Modified [email protected] February 1990
90 * added multiple table interface
91 * Modified by [email protected] April 1990
92 * changed ctl structure for shared memory
93 */
94
95 #include "postgres.h"
96
97 #include <limits.h>
98
99 #include "access/xact.h"
100 #include "common/hashfn.h"
101 #include "port/pg_bitutils.h"
102 #include "storage/shmem.h"
103 #include "storage/spin.h"
104 #include "utils/dynahash.h"
105 #include "utils/memutils.h"
106
107
108 /*
109 * Constants
110 *
111 * A hash table has a top-level "directory", each of whose entries points
112 * to a "segment" of ssize bucket headers. The maximum number of hash
113 * buckets is thus dsize * ssize (but dsize may be expansible). Of course,
114 * the number of records in the table can be larger, but we don't want a
115 * whole lot of records per bucket or performance goes down.
116 *
117 * In a hash table allocated in shared memory, the directory cannot be
118 * expanded because it must stay at a fixed address. The directory size
119 * should be selected using hash_select_dirsize (and you'd better have
120 * a good idea of the maximum number of entries!). For non-shared hash
121 * tables, the initial directory size can be left at the default.
122 */
123 #define DEF_SEGSIZE 256
124 #define DEF_SEGSIZE_SHIFT 8 /* must be log2(DEF_SEGSIZE) */
125 #define DEF_DIRSIZE 256
126
127 /* Number of freelists to be used for a partitioned hash table. */
128 #define NUM_FREELISTS 32
129
130 /* A hash bucket is a linked list of HASHELEMENTs */
131 typedef HASHELEMENT *HASHBUCKET;
132
133 /* A hash segment is an array of bucket headers */
134 typedef HASHBUCKET *HASHSEGMENT;
135
136 /*
137 * Per-freelist data.
138 *
139 * In a partitioned hash table, each freelist is associated with a specific
140 * set of hashcodes, as determined by the FREELIST_IDX() macro below.
141 * nentries tracks the number of live hashtable entries having those hashcodes
142 * (NOT the number of entries in the freelist, as you might expect).
143 *
144 * The coverage of a freelist might be more or less than one partition, so it
145 * needs its own lock rather than relying on caller locking. Relying on that
146 * wouldn't work even if the coverage was the same, because of the occasional
147 * need to "borrow" entries from another freelist; see get_hash_entry().
148 *
149 * Using an array of FreeListData instead of separate arrays of mutexes,
150 * nentries and freeLists helps to reduce sharing of cache lines between
151 * different mutexes.
152 */
153 typedef struct
154 {
155 slock_t mutex; /* spinlock for this freelist */
156 long nentries; /* number of entries in associated buckets */
157 HASHELEMENT *freeList; /* chain of free elements */
158 } FreeListData;
159
160 /*
161 * Header structure for a hash table --- contains all changeable info
162 *
163 * In a shared-memory hash table, the HASHHDR is in shared memory, while
164 * each backend has a local HTAB struct. For a non-shared table, there isn't
165 * any functional difference between HASHHDR and HTAB, but we separate them
166 * anyway to share code between shared and non-shared tables.
167 */
168 struct HASHHDR
169 {
170 /*
171 * The freelist can become a point of contention in high-concurrency hash
172 * tables, so we use an array of freelists, each with its own mutex and
173 * nentries count, instead of just a single one. Although the freelists
174 * normally operate independently, we will scavenge entries from freelists
175 * other than a hashcode's default freelist when necessary.
176 *
177 * If the hash table is not partitioned, only freeList[0] is used and its
178 * spinlock is not used at all; callers' locking is assumed sufficient.
179 */
180 FreeListData freeList[NUM_FREELISTS];
181
182 /* These fields can change, but not in a partitioned table */
183 /* Also, dsize can't change in a shared table, even if unpartitioned */
184 long dsize; /* directory size */
185 long nsegs; /* number of allocated segments (<= dsize) */
186 uint32 max_bucket; /* ID of maximum bucket in use */
187 uint32 high_mask; /* mask to modulo into entire table */
188 uint32 low_mask; /* mask to modulo into lower half of table */
189
190 /* These fields are fixed at hashtable creation */
191 Size keysize; /* hash key length in bytes */
192 Size entrysize; /* total user element size in bytes */
193 long num_partitions; /* # partitions (must be power of 2), or 0 */
194 long max_dsize; /* 'dsize' limit if directory is fixed size */
195 long ssize; /* segment size --- must be power of 2 */
196 int sshift; /* segment shift = log2(ssize) */
197 int nelem_alloc; /* number of entries to allocate at once */
198
199 #ifdef HASH_STATISTICS
200
201 /*
202 * Count statistics here. NB: stats code doesn't bother with mutex, so
203 * counts could be corrupted a bit in a partitioned table.
204 */
205 long accesses;
206 long collisions;
207 #endif
208 };
209
210 #define IS_PARTITIONED(hctl) ((hctl)->num_partitions != 0)
211
212 #define FREELIST_IDX(hctl, hashcode) \
213 (IS_PARTITIONED(hctl) ? (hashcode) % NUM_FREELISTS : 0)
214
215 /*
216 * Top control structure for a hashtable --- in a shared table, each backend
217 * has its own copy (OK since no fields change at runtime)
218 */
219 struct HTAB
220 {
221 HASHHDR *hctl; /* => shared control information */
222 HASHSEGMENT *dir; /* directory of segment starts */
223 HashValueFunc hash; /* hash function */
224 HashCompareFunc match; /* key comparison function */
225 HashCopyFunc keycopy; /* key copying function */
226 HashAllocFunc alloc; /* memory allocator */
227 MemoryContext hcxt; /* memory context if default allocator used */
228 char *tabname; /* table name (for error messages) */
229 bool isshared; /* true if table is in shared memory */
230 bool isfixed; /* if true, don't enlarge */
231
232 /* freezing a shared table isn't allowed, so we can keep state here */
233 bool frozen; /* true = no more inserts allowed */
234
235 /* We keep local copies of these fixed values to reduce contention */
236 Size keysize; /* hash key length in bytes */
237 long ssize; /* segment size --- must be power of 2 */
238 int sshift; /* segment shift = log2(ssize) */
239 };
240
241 /*
242 * Key (also entry) part of a HASHELEMENT
243 */
244 #define ELEMENTKEY(helem) (((char *)(helem)) + MAXALIGN(sizeof(HASHELEMENT)))
245
246 /*
247 * Obtain element pointer given pointer to key
248 */
249 #define ELEMENT_FROM_KEY(key) \
250 ((HASHELEMENT *) (((char *) (key)) - MAXALIGN(sizeof(HASHELEMENT))))
251
252 /*
253 * Fast MOD arithmetic, assuming that y is a power of 2 !
254 */
255 #define MOD(x,y) ((x) & ((y)-1))
256
257 #ifdef HASH_STATISTICS
258 static long hash_accesses,
259 hash_collisions,
260 hash_expansions;
261 #endif
262
263 /*
264 * Private function prototypes
265 */
266 static void *DynaHashAlloc(Size size);
267 static HASHSEGMENT seg_alloc(HTAB *hashp);
268 static bool element_alloc(HTAB *hashp, int nelem, int freelist_idx);
269 static bool dir_realloc(HTAB *hashp);
270 static bool expand_table(HTAB *hashp);
271 static HASHBUCKET get_hash_entry(HTAB *hashp, int freelist_idx);
272 static void hdefault(HTAB *hashp);
273 static int choose_nelem_alloc(Size entrysize);
274 static bool init_htab(HTAB *hashp, long nelem);
275 pg_noreturn static void hash_corrupted(HTAB *hashp);
276 static uint32 hash_initial_lookup(HTAB *hashp, uint32 hashvalue,
277 HASHBUCKET **bucketptr);
278 static long next_pow2_long(long num);
279 static int next_pow2_int(long num);
280 static void register_seq_scan(HTAB *hashp);
281 static void deregister_seq_scan(HTAB *hashp);
282 static bool has_seq_scans(HTAB *hashp);
283
284
285 /*
286 * memory allocation support
287 */
288 static MemoryContext CurrentDynaHashCxt = NULL;
289
290 static void *
291 DynaHashAlloc(Size size)
292 {
293 Assert(MemoryContextIsValid(CurrentDynaHashCxt));
294 return MemoryContextAllocExtended(CurrentDynaHashCxt, size,
295 MCXT_ALLOC_NO_OOM);
296 }
297
298
299 /*
300 * HashCompareFunc for string keys
301 *
302 * Because we copy keys with strlcpy(), they will be truncated at keysize-1
303 * bytes, so we can only compare that many ... hence strncmp is almost but
304 * not quite the right thing.
305 */
306 static int
307 string_compare(const char *key1, const char *key2, Size keysize)
308 {
309 return strncmp(key1, key2, keysize - 1);
310 }
311
312
313 /************************** CREATE ROUTINES **********************/
314
315 /*
316 * hash_create -- create a new dynamic hash table
317 *
318 * tabname: a name for the table (for debugging purposes)
319 * nelem: maximum number of elements expected
320 * *info: additional table parameters, as indicated by flags
321 * flags: bitmask indicating which parameters to take from *info
322 *
323 * The flags value *must* include HASH_ELEM. (Formerly, this was nominally
324 * optional, but the default keysize and entrysize values were useless.)
325 * The flags value must also include exactly one of HASH_STRINGS, HASH_BLOBS,
326 * or HASH_FUNCTION, to define the key hashing semantics (C strings,
327 * binary blobs, or custom, respectively). Callers specifying a custom
328 * hash function will likely also want to use HASH_COMPARE, and perhaps
329 * also HASH_KEYCOPY, to control key comparison and copying.
330 * Another often-used flag is HASH_CONTEXT, to allocate the hash table
331 * under info->hcxt rather than under TopMemoryContext; the default
332 * behavior is only suitable for session-lifespan hash tables.
333 * Other flags bits are special-purpose and seldom used, except for those
334 * associated with shared-memory hash tables, for which see ShmemInitHash().
335 *
336 * Fields in *info are read only when the associated flags bit is set.
337 * It is not necessary to initialize other fields of *info.
338 * Neither tabname nor *info need persist after the hash_create() call.
339 *
340 * Note: It is deprecated for callers of hash_create() to explicitly specify
341 * string_hash, tag_hash, uint32_hash, or oid_hash. Just set HASH_STRINGS or
342 * HASH_BLOBS. Use HASH_FUNCTION only when you want something other than
343 * one of these.
344 *
345 * Note: for a shared-memory hashtable, nelem needs to be a pretty good
346 * estimate, since we can't expand the table on the fly. But an unshared
347 * hashtable can be expanded on-the-fly, so it's better for nelem to be
348 * on the small side and let the table grow if it's exceeded. An overly
349 * large nelem will penalize hash_seq_search speed without buying much.
350 */
351 HTAB *
352 hash_create(const char *tabname, long nelem, const HASHCTL *info, int flags)
353 {
354 HTAB *hashp;
355 HASHHDR *hctl;
356
357 /*
358 * Hash tables now allocate space for key and data, but you have to say
359 * how much space to allocate.
360 */
361 Assert(flags & HASH_ELEM);
362 Assert(info->keysize > 0);
363 Assert(info->entrysize >= info->keysize);
364
365 /*
366 * For shared hash tables, we have a local hash header (HTAB struct) that
367 * we allocate in TopMemoryContext; all else is in shared memory.
368 *
369 * For non-shared hash tables, everything including the hash header is in
370 * a memory context created specially for the hash table --- this makes
371 * hash_destroy very simple. The memory context is made a child of either
372 * a context specified by the caller, or TopMemoryContext if nothing is
373 * specified.
374 */
375 if (flags & HASH_SHARED_MEM)
376 {
377 /* Set up to allocate the hash header */
378 CurrentDynaHashCxt = TopMemoryContext;
379 }
380 else
381 {
382 /* Create the hash table's private memory context */
383 if (flags & HASH_CONTEXT)
384 CurrentDynaHashCxt = info->hcxt;
385 else
386 CurrentDynaHashCxt = TopMemoryContext;
387 CurrentDynaHashCxt = AllocSetContextCreate(CurrentDynaHashCxt,
388 "dynahash",
389 ALLOCSET_DEFAULT_SIZES);
390 }
391
392 /* Initialize the hash header, plus a copy of the table name */
393 hashp = (HTAB *) MemoryContextAlloc(CurrentDynaHashCxt,
394 sizeof(HTAB) + strlen(tabname) + 1);
395 MemSet(hashp, 0, sizeof(HTAB));
396
397 hashp->tabname = (char *) (hashp + 1);
398 strcpy(hashp->tabname, tabname);
399
400 /* If we have a private context, label it with hashtable's name */
401 if (!(flags & HASH_SHARED_MEM))
402 MemoryContextSetIdentifier(CurrentDynaHashCxt, hashp->tabname);
403
404 /*
405 * Select the appropriate hash function (see comments at head of file).
406 */
407 if (flags & HASH_FUNCTION)
408 {
409 Assert(!(flags & (HASH_BLOBS | HASH_STRINGS)));
410 hashp->hash = info->hash;
411 }
412 else if (flags & HASH_BLOBS)
413 {
414 Assert(!(flags & HASH_STRINGS));
415 /* We can optimize hashing for common key sizes */
416 if (info->keysize == sizeof(uint32))
417 hashp->hash = uint32_hash;
418 else
419 hashp->hash = tag_hash;
420 }
421 else
422 {
423 /*
424 * string_hash used to be considered the default hash method, and in a
425 * non-assert build it effectively still is. But we now consider it
426 * an assertion error to not say HASH_STRINGS explicitly. To help
427 * catch mistaken usage of HASH_STRINGS, we also insist on a
428 * reasonably long string length: if the keysize is only 4 or 8 bytes,
429 * it's almost certainly an integer or pointer not a string.
430 */
431 Assert(flags & HASH_STRINGS);
432 Assert(info->keysize > 8);
433
434 hashp->hash = string_hash;
435 }
436
437 /*
438 * If you don't specify a match function, it defaults to string_compare if
439 * you used string_hash, and to memcmp otherwise.
440 *
441 * Note: explicitly specifying string_hash is deprecated, because this
442 * might not work for callers in loadable modules on some platforms due to
443 * referencing a trampoline instead of the string_hash function proper.
444 * Specify HASH_STRINGS instead.
445 */
446 if (flags & HASH_COMPARE)
447 hashp->match = info->match;
448 else if (hashp->hash == string_hash)
449 hashp->match = (HashCompareFunc) string_compare;
450 else
451 hashp->match = memcmp;
452
453 /*
454 * Similarly, the key-copying function defaults to strlcpy or memcpy.
455 */
456 if (flags & HASH_KEYCOPY)
457 hashp->keycopy = info->keycopy;
458 else if (hashp->hash == string_hash)
459 {
460 /*
461 * The signature of keycopy is meant for memcpy(), which returns
462 * void*, but strlcpy() returns size_t. Since we never use the return
463 * value of keycopy, and size_t is pretty much always the same size as
464 * void *, this should be safe. The extra cast in the middle is to
465 * avoid warnings from -Wcast-function-type.
466 */
467 hashp->keycopy = (HashCopyFunc) (pg_funcptr_t) strlcpy;
468 }
469 else
470 hashp->keycopy = memcpy;
471
472 /* And select the entry allocation function, too. */
473 if (flags & HASH_ALLOC)
474 hashp->alloc = info->alloc;
475 else
476 hashp->alloc = DynaHashAlloc;
477
478 if (flags & HASH_SHARED_MEM)
479 {
480 /*
481 * ctl structure and directory are preallocated for shared memory
482 * tables. Note that HASH_DIRSIZE and HASH_ALLOC had better be set as
483 * well.
484 */
485 hashp->hctl = info->hctl;
486 hashp->dir = (HASHSEGMENT *) (((char *) info->hctl) + sizeof(HASHHDR));
487 hashp->hcxt = NULL;
488 hashp->isshared = true;
489
490 /* hash table already exists, we're just attaching to it */
491 if (flags & HASH_ATTACH)
492 {
493 /* make local copies of some heavily-used values */
494 hctl = hashp->hctl;
495 hashp->keysize = hctl->keysize;
496 hashp->ssize = hctl->ssize;
497 hashp->sshift = hctl->sshift;
498
499 return hashp;
500 }
501 }
502 else
503 {
504 /* setup hash table defaults */
505 hashp->hctl = NULL;
506 hashp->dir = NULL;
507 hashp->hcxt = CurrentDynaHashCxt;
508 hashp->isshared = false;
509 }
510
511 if (!hashp->hctl)
512 {
513 hashp->hctl = (HASHHDR *) hashp->alloc(sizeof(HASHHDR));
514 if (!hashp->hctl)
515 ereport(ERROR,
516 (errcode(ERRCODE_OUT_OF_MEMORY),
517 errmsg("out of memory")));
518 }
519
520 hashp->frozen = false;
521
522 hdefault(hashp);
523
524 hctl = hashp->hctl;
525
526 if (flags & HASH_PARTITION)
527 {
528 /* Doesn't make sense to partition a local hash table */
529 Assert(flags & HASH_SHARED_MEM);
530
531 /*
532 * The number of partitions had better be a power of 2. Also, it must
533 * be less than INT_MAX (see init_htab()), so call the int version of
534 * next_pow2.
535 */
536 Assert(info->num_partitions == next_pow2_int(info->num_partitions));
537
538 hctl->num_partitions = info->num_partitions;
539 }
540
541 if (flags & HASH_SEGMENT)
542 {
543 hctl->ssize = info->ssize;
544 hctl->sshift = my_log2(info->ssize);
545 /* ssize had better be a power of 2 */
546 Assert(hctl->ssize == (1L << hctl->sshift));
547 }
548
549 /*
550 * SHM hash tables have fixed directory size passed by the caller.
551 */
552 if (flags & HASH_DIRSIZE)
553 {
554 hctl->max_dsize = info->max_dsize;
555 hctl->dsize = info->dsize;
556 }
557
558 /* remember the entry sizes, too */
559 hctl->keysize = info->keysize;
560 hctl->entrysize = info->entrysize;
561
562 /* make local copies of heavily-used constant fields */
563 hashp->keysize = hctl->keysize;
564 hashp->ssize = hctl->ssize;
565 hashp->sshift = hctl->sshift;
566
567 /* Build the hash directory structure */
568 if (!init_htab(hashp, nelem))
569 elog(ERROR, "failed to initialize hash table \"%s\"", hashp->tabname);
570
571 /*
572 * For a shared hash table, preallocate the requested number of elements.
573 * This reduces problems with run-time out-of-shared-memory conditions.
574 *
575 * For a non-shared hash table, preallocate the requested number of
576 * elements if it's less than our chosen nelem_alloc. This avoids wasting
577 * space if the caller correctly estimates a small table size.
578 */
579 if ((flags & HASH_SHARED_MEM) ||
580 nelem < hctl->nelem_alloc)
581 {
582 int i,
583 freelist_partitions,
584 nelem_alloc,
585 nelem_alloc_first;
586
587 /*
588 * If hash table is partitioned, give each freelist an equal share of
589 * the initial allocation. Otherwise only freeList[0] is used.
590 */
591 if (IS_PARTITIONED(hashp->hctl))
592 freelist_partitions = NUM_FREELISTS;
593 else
594 freelist_partitions = 1;
595
596 nelem_alloc = nelem / freelist_partitions;
597 if (nelem_alloc <= 0)
598 nelem_alloc = 1;
599
600 /*
601 * Make sure we'll allocate all the requested elements; freeList[0]
602 * gets the excess if the request isn't divisible by NUM_FREELISTS.
603 */
604 if (nelem_alloc * freelist_partitions < nelem)
605 nelem_alloc_first =
606 nelem - nelem_alloc * (freelist_partitions - 1);
607 else
608 nelem_alloc_first = nelem_alloc;
609
610 for (i = 0; i < freelist_partitions; i++)
611 {
612 int temp = (i == 0) ? nelem_alloc_first : nelem_alloc;
613
614 if (!element_alloc(hashp, temp, i))
615 ereport(ERROR,
616 (errcode(ERRCODE_OUT_OF_MEMORY),
617 errmsg("out of memory")));
618 }
619 }
620
621 if (flags & HASH_FIXED_SIZE)
622 hashp->isfixed = true;
623 return hashp;
624 }
625
626 /*
627 * Set default HASHHDR parameters.
628 */
629 static void
630 hdefault(HTAB *hashp)
631 {
632 HASHHDR *hctl = hashp->hctl;
633
634 MemSet(hctl, 0, sizeof(HASHHDR));
635
636 hctl->dsize = DEF_DIRSIZE;
637 hctl->nsegs = 0;
638
639 hctl->num_partitions = 0; /* not partitioned */
640
641 /* table has no fixed maximum size */
642 hctl->max_dsize = NO_MAX_DSIZE;
643
644 hctl->ssize = DEF_SEGSIZE;
645 hctl->sshift = DEF_SEGSIZE_SHIFT;
646
647 #ifdef HASH_STATISTICS
648 hctl->accesses = hctl->collisions = 0;
649 #endif
650 }
651
652 /*
653 * Given the user-specified entry size, choose nelem_alloc, ie, how many
654 * elements to add to the hash table when we need more.
655 */
656 static int
657 choose_nelem_alloc(Size entrysize)
658 {
659 int nelem_alloc;
660 Size elementSize;
661 Size allocSize;
662
663 /* Each element has a HASHELEMENT header plus user data. */
664 /* NB: this had better match element_alloc() */
665 elementSize = MAXALIGN(sizeof(HASHELEMENT)) + MAXALIGN(entrysize);
666
667 /*
668 * The idea here is to choose nelem_alloc at least 32, but round up so
669 * that the allocation request will be a power of 2 or just less. This
670 * makes little difference for hash tables in shared memory, but for hash
671 * tables managed by palloc, the allocation request will be rounded up to
672 * a power of 2 anyway. If we fail to take this into account, we'll waste
673 * as much as half the allocated space.
674 */
675 allocSize = 32 * 4; /* assume elementSize at least 8 */
676 do
677 {
678 allocSize <<= 1;
679 nelem_alloc = allocSize / elementSize;
680 } while (nelem_alloc < 32);
681
682 return nelem_alloc;
683 }
684
685 /*
686 * Compute derived fields of hctl and build the initial directory/segment
687 * arrays
688 */
689 static bool
690 init_htab(HTAB *hashp, long nelem)
691 {
692 HASHHDR *hctl = hashp->hctl;
693 HASHSEGMENT *segp;
694 int nbuckets;
695 int nsegs;
696 int i;
697
698 /*
699 * initialize mutexes if it's a partitioned table
700 */
701 if (IS_PARTITIONED(hctl))
702 for (i = 0; i < NUM_FREELISTS; i++)
703 SpinLockInit(&(hctl->freeList[i].mutex));
704
705 /*
706 * Allocate space for the next greater power of two number of buckets,
707 * assuming a desired maximum load factor of 1.
708 */
709 nbuckets = next_pow2_int(nelem);
710
711 /*
712 * In a partitioned table, nbuckets must be at least equal to
713 * num_partitions; were it less, keys with apparently different partition
714 * numbers would map to the same bucket, breaking partition independence.
715 * (Normally nbuckets will be much bigger; this is just a safety check.)
716 */
717 while (nbuckets < hctl->num_partitions)
718 nbuckets <<= 1;
719
720 hctl->max_bucket = hctl->low_mask = nbuckets - 1;
721 hctl->high_mask = (nbuckets << 1) - 1;
722
723 /*
724 * Figure number of directory segments needed, round up to a power of 2
725 */
726 nsegs = (nbuckets - 1) / hctl->ssize + 1;
727 nsegs = next_pow2_int(nsegs);
728
729 /*
730 * Make sure directory is big enough. If pre-allocated directory is too
731 * small, choke (caller screwed up).
732 */
733 if (nsegs > hctl->dsize)
734 {
735 if (!(hashp->dir))
736 hctl->dsize = nsegs;
737 else
738 return false;
739 }
740
741 /* Allocate a directory */
742 if (!(hashp->dir))
743 {
744 CurrentDynaHashCxt = hashp->hcxt;
745 hashp->dir = (HASHSEGMENT *)
746 hashp->alloc(hctl->dsize * sizeof(HASHSEGMENT));
747 if (!hashp->dir)
748 return false;
749 }
750
751 /* Allocate initial segments */
752 for (segp = hashp->dir; hctl->nsegs < nsegs; hctl->nsegs++, segp++)
753 {
754 *segp = seg_alloc(hashp);
755 if (*segp == NULL)
756 return false;
757 }
758
759 /* Choose number of entries to allocate at a time */
760 hctl->nelem_alloc = choose_nelem_alloc(hctl->entrysize);
761
762 #ifdef HASH_DEBUG
763 fprintf(stderr, "init_htab:\n%s%p\n%s%ld\n%s%ld\n%s%d\n%s%u\n%s%x\n%s%x\n%s%ld\n",
764 "TABLE POINTER ", hashp,
765 "DIRECTORY SIZE ", hctl->dsize,
766 "SEGMENT SIZE ", hctl->ssize,
767 "SEGMENT SHIFT ", hctl->sshift,
768 "MAX BUCKET ", hctl->max_bucket,
769 "HIGH MASK ", hctl->high_mask,
770 "LOW MASK ", hctl->low_mask,
771 "NSEGS ", hctl->nsegs);
772 #endif
773 return true;
774 }
775
776 /*
777 * Estimate the space needed for a hashtable containing the given number
778 * of entries of given size.
779 * NOTE: this is used to estimate the footprint of hashtables in shared
780 * memory; therefore it does not count HTAB which is in local memory.
781 * NB: assumes that all hash structure parameters have default values!
782 */
783 Size
784 hash_estimate_size(long num_entries, Size entrysize)
785 {
786 Size size;
787 long nBuckets,
788 nSegments,
789 nDirEntries,
790 nElementAllocs,
791 elementSize,
792 elementAllocCnt;
793
794 /* estimate number of buckets wanted */
795 nBuckets = next_pow2_long(num_entries);
796 /* # of segments needed for nBuckets */
797 nSegments = next_pow2_long((nBuckets - 1) / DEF_SEGSIZE + 1);
798 /* directory entries */
799 nDirEntries = DEF_DIRSIZE;
800 while (nDirEntries < nSegments)
801 nDirEntries <<= 1; /* dir_alloc doubles dsize at each call */
802
803 /* fixed control info */
804 size = MAXALIGN(sizeof(HASHHDR)); /* but not HTAB, per above */
805 /* directory */
806 size = add_size(size, mul_size(nDirEntries, sizeof(HASHSEGMENT)));
807 /* segments */
808 size = add_size(size, mul_size(nSegments,
809 MAXALIGN(DEF_SEGSIZE * sizeof(HASHBUCKET))));
810 /* elements --- allocated in groups of choose_nelem_alloc() entries */
811 elementAllocCnt = choose_nelem_alloc(entrysize);
812 nElementAllocs = (num_entries - 1) / elementAllocCnt + 1;
813 elementSize = MAXALIGN(sizeof(HASHELEMENT)) + MAXALIGN(entrysize);
814 size = add_size(size,
815 mul_size(nElementAllocs,
816 mul_size(elementAllocCnt, elementSize)));
817
818 return size;
819 }
820
821 /*
822 * Select an appropriate directory size for a hashtable with the given
823 * maximum number of entries.
824 * This is only needed for hashtables in shared memory, whose directories
825 * cannot be expanded dynamically.
826 * NB: assumes that all hash structure parameters have default values!
827 *
828 * XXX this had better agree with the behavior of init_htab()...
829 */
830 long
831 hash_select_dirsize(long num_entries)
832 {
833 long nBuckets,
834 nSegments,
835 nDirEntries;
836
837 /* estimate number of buckets wanted */
838 nBuckets = next_pow2_long(num_entries);
839 /* # of segments needed for nBuckets */
840 nSegments = next_pow2_long((nBuckets - 1) / DEF_SEGSIZE + 1);
841 /* directory entries */
842 nDirEntries = DEF_DIRSIZE;
843 while (nDirEntries < nSegments)
844 nDirEntries <<= 1; /* dir_alloc doubles dsize at each call */
845
846 return nDirEntries;
847 }
848
849 /*
850 * Compute the required initial memory allocation for a shared-memory
851 * hashtable with the given parameters. We need space for the HASHHDR
852 * and for the (non expansible) directory.
853 */
854 Size
855 hash_get_shared_size(HASHCTL *info, int flags)
856 {
857 Assert(flags & HASH_DIRSIZE);
858 Assert(info->dsize == info->max_dsize);
859 return sizeof(HASHHDR) + info->dsize * sizeof(HASHSEGMENT);
860 }
861
862
863 /********************** DESTROY ROUTINES ************************/
864
865 void
866 hash_destroy(HTAB *hashp)
867 {
868 if (hashp != NULL)
869 {
870 /* allocation method must be one we know how to free, too */
871 Assert(hashp->alloc == DynaHashAlloc);
872 /* so this hashtable must have its own context */
873 Assert(hashp->hcxt != NULL);
874
875 hash_stats("destroy", hashp);
876
877 /*
878 * Free everything by destroying the hash table's memory context.
879 */
880 MemoryContextDelete(hashp->hcxt);
881 }
882 }
883
884 void
885 hash_stats(const char *where, HTAB *hashp)
886 {
887 #ifdef HASH_STATISTICS
888 fprintf(stderr, "%s: this HTAB -- accesses %ld collisions %ld\n",
889 where, hashp->hctl->accesses, hashp->hctl->collisions);
890
891 fprintf(stderr, "hash_stats: entries %ld keysize %ld maxp %u segmentcount %ld\n",
892 hash_get_num_entries(hashp), (long) hashp->hctl->keysize,
893 hashp->hctl->max_bucket, hashp->hctl->nsegs);
894 fprintf(stderr, "%s: total accesses %ld total collisions %ld\n",
895 where, hash_accesses, hash_collisions);
896 fprintf(stderr, "hash_stats: total expansions %ld\n",
897 hash_expansions);
898 #endif
899 }
900
901 /*******************************SEARCH ROUTINES *****************************/
902
903
904 /*
905 * get_hash_value -- exported routine to calculate a key's hash value
906 *
907 * We export this because for partitioned tables, callers need to compute
908 * the partition number (from the low-order bits of the hash value) before
909 * searching.
910 */
911 uint32
912 get_hash_value(HTAB *hashp, const void *keyPtr)
913 {
914 return hashp->hash(keyPtr, hashp->keysize);
915 }
916
917 /* Convert a hash value to a bucket number */
918 static inline uint32
919 calc_bucket(HASHHDR *hctl, uint32 hash_val)
920 {
921 uint32 bucket;
922
923 bucket = hash_val & hctl->high_mask;
924 if (bucket > hctl->max_bucket)
925 bucket = bucket & hctl->low_mask;
926
927 return bucket;
928 }
929
930 /*
931 * hash_search -- look up key in table and perform action
932 * hash_search_with_hash_value -- same, with key's hash value already computed
933 *
934 * action is one of:
935 * HASH_FIND: look up key in table
936 * HASH_ENTER: look up key in table, creating entry if not present
937 * HASH_ENTER_NULL: same, but return NULL if out of memory
938 * HASH_REMOVE: look up key in table, remove entry if present
939 *
940 * Return value is a pointer to the element found/entered/removed if any,
941 * or NULL if no match was found. (NB: in the case of the REMOVE action,
942 * the result is a dangling pointer that shouldn't be dereferenced!)
943 *
944 * HASH_ENTER will normally ereport a generic "out of memory" error if
945 * it is unable to create a new entry. The HASH_ENTER_NULL operation is
946 * the same except it will return NULL if out of memory.
947 *
948 * If foundPtr isn't NULL, then *foundPtr is set true if we found an
949 * existing entry in the table, false otherwise. This is needed in the
950 * HASH_ENTER case, but is redundant with the return value otherwise.
951 *
952 * For hash_search_with_hash_value, the hashvalue parameter must have been
953 * calculated with get_hash_value().
954 */
955 void *
956 hash_search(HTAB *hashp,
957 const void *keyPtr,
958 HASHACTION action,
959 bool *foundPtr)
960 {
961 return hash_search_with_hash_value(hashp,
962 keyPtr,
963 hashp->hash(keyPtr, hashp->keysize),
964 action,
965 foundPtr);
966 }
967
968 void *
969 hash_search_with_hash_value(HTAB *hashp,
970 const void *keyPtr,
971 uint32 hashvalue,
972 HASHACTION action,
973 bool *foundPtr)
974 {
975 HASHHDR *hctl = hashp->hctl;
976 int freelist_idx = FREELIST_IDX(hctl, hashvalue);
977 Size keysize;
978 HASHBUCKET currBucket;
979 HASHBUCKET *prevBucketPtr;
980 HashCompareFunc match;
981
982 #ifdef HASH_STATISTICS
983 hash_accesses++;
984 hctl->accesses++;
985 #endif
986
987 /*
988 * If inserting, check if it is time to split a bucket.
989 *
990 * NOTE: failure to expand table is not a fatal error, it just means we
991 * have to run at higher fill factor than we wanted. However, if we're
992 * using the palloc allocator then it will throw error anyway on
993 * out-of-memory, so we must do this before modifying the table.
994 */
995 if (action == HASH_ENTER || action == HASH_ENTER_NULL)
996 {
997 /*
998 * Can't split if running in partitioned mode, nor if frozen, nor if
999 * table is the subject of any active hash_seq_search scans.
1000 */
1001 if (hctl->freeList[0].nentries > (long) hctl->max_bucket &&
1002 !IS_PARTITIONED(hctl) && !hashp->frozen &&
1003 !has_seq_scans(hashp))
1004 (void) expand_table(hashp);
1005 }
1006
1007 /*
1008 * Do the initial lookup
1009 */
1010 (void) hash_initial_lookup(hashp, hashvalue, &prevBucketPtr);
1011 currBucket = *prevBucketPtr;
1012
1013 /*
1014 * Follow collision chain looking for matching key
1015 */
1016 match = hashp->match; /* save one fetch in inner loop */
1017 keysize = hashp->keysize; /* ditto */
1018
1019 while (currBucket != NULL)
1020 {
1021 if (currBucket->hashvalue == hashvalue &&
1022 match(ELEMENTKEY(currBucket), keyPtr, keysize) == 0)
1023 break;
1024 prevBucketPtr = &(currBucket->link);
1025 currBucket = *prevBucketPtr;
1026 #ifdef HASH_STATISTICS
1027 hash_collisions++;
1028 hctl->collisions++;
1029 #endif
1030 }
1031
1032 if (foundPtr)
1033 *foundPtr = (bool) (currBucket != NULL);
1034
1035 /*
1036 * OK, now what?
1037 */
1038 switch (action)
1039 {
1040 case HASH_FIND:
1041 if (currBucket != NULL)
1042 return ELEMENTKEY(currBucket);
1043 return NULL;
1044
1045 case HASH_REMOVE:
1046 if (currBucket != NULL)
1047 {
1048 /* if partitioned, must lock to touch nentries and freeList */
1049 if (IS_PARTITIONED(hctl))
1050 SpinLockAcquire(&(hctl->freeList[freelist_idx].mutex));
1051
1052 /* delete the record from the appropriate nentries counter. */
1053 Assert(hctl->freeList[freelist_idx].nentries > 0);
1054 hctl->freeList[freelist_idx].nentries--;
1055
1056 /* remove record from hash bucket's chain. */
1057 *prevBucketPtr = currBucket->link;
1058
1059 /* add the record to the appropriate freelist. */
1060 currBucket->link = hctl->freeList[freelist_idx].freeList;
1061 hctl->freeList[freelist_idx].freeList = currBucket;
1062
1063 if (IS_PARTITIONED(hctl))
1064 SpinLockRelease(&hctl->freeList[freelist_idx].mutex);
1065
1066 /*
1067 * better hope the caller is synchronizing access to this
1068 * element, because someone else is going to reuse it the next
1069 * time something is added to the table
1070 */
1071 return ELEMENTKEY(currBucket);
1072 }
1073 return NULL;
1074
1075 case HASH_ENTER:
1076 case HASH_ENTER_NULL:
1077 /* Return existing element if found, else create one */
1078 if (currBucket != NULL)
1079 return ELEMENTKEY(currBucket);
1080
1081 /* disallow inserts if frozen */
1082 if (hashp->frozen)
1083 elog(ERROR, "cannot insert into frozen hashtable \"%s\"",
1084 hashp->tabname);
1085
1086 currBucket = get_hash_entry(hashp, freelist_idx);
1087 if (currBucket == NULL)
1088 {
1089 /* out of memory */
1090 if (action == HASH_ENTER_NULL)
1091 return NULL;
1092 /* report a generic message */
1093 if (hashp->isshared)
1094 ereport(ERROR,
1095 (errcode(ERRCODE_OUT_OF_MEMORY),
1096 errmsg("out of shared memory")));
1097 else
1098 ereport(ERROR,
1099 (errcode(ERRCODE_OUT_OF_MEMORY),
1100 errmsg("out of memory")));
1101 }
1102
1103 /* link into hashbucket chain */
1104 *prevBucketPtr = currBucket;
1105 currBucket->link = NULL;
1106
1107 /* copy key into record */
1108 currBucket->hashvalue = hashvalue;
1109 hashp->keycopy(ELEMENTKEY(currBucket), keyPtr, keysize);
1110
1111 /*
1112 * Caller is expected to fill the data field on return. DO NOT
1113 * insert any code that could possibly throw error here, as doing
1114 * so would leave the table entry incomplete and hence corrupt the
1115 * caller's data structure.
1116 */
1117
1118 return ELEMENTKEY(currBucket);
1119 }
1120
1121 elog(ERROR, "unrecognized hash action code: %d", (int) action);
1122
1123 return NULL; /* keep compiler quiet */
1124 }
1125
1126 /*
1127 * hash_update_hash_key -- change the hash key of an existing table entry
1128 *
1129 * This is equivalent to removing the entry, making a new entry, and copying
1130 * over its data, except that the entry never goes to the table's freelist.
1131 * Therefore this cannot suffer an out-of-memory failure, even if there are
1132 * other processes operating in other partitions of the hashtable.
1133 *
1134 * Returns true if successful, false if the requested new hash key is already
1135 * present. Throws error if the specified entry pointer isn't actually a
1136 * table member.
1137 *
1138 * NB: currently, there is no special case for old and new hash keys being
1139 * identical, which means we'll report false for that situation. This is
1140 * preferable for existing uses.
1141 *
1142 * NB: for a partitioned hashtable, caller must hold lock on both relevant
1143 * partitions, if the new hash key would belong to a different partition.
1144 */
1145 bool
1146 hash_update_hash_key(HTAB *hashp,
1147 void *existingEntry,
1148 const void *newKeyPtr)
1149 {
1150 HASHELEMENT *existingElement = ELEMENT_FROM_KEY(existingEntry);
1151 uint32 newhashvalue;
1152 Size keysize;
1153 uint32 bucket;
1154 uint32 newbucket;
1155 HASHBUCKET currBucket;
1156 HASHBUCKET *prevBucketPtr;
1157 HASHBUCKET *oldPrevPtr;
1158 HashCompareFunc match;
1159
1160 #ifdef HASH_STATISTICS
1161 HASHHDR *hctl = hashp->hctl;
1162
1163 hash_accesses++;
1164 hctl->accesses++;
1165 #endif
1166
1167 /* disallow updates if frozen */
1168 if (hashp->frozen)
1169 elog(ERROR, "cannot update in frozen hashtable \"%s\"",
1170 hashp->tabname);
1171
1172 /*
1173 * Lookup the existing element using its saved hash value. We need to do
1174 * this to be able to unlink it from its hash chain, but as a side benefit
1175 * we can verify the validity of the passed existingEntry pointer.
1176 */
1177 bucket = hash_initial_lookup(hashp, existingElement->hashvalue,
1178 &prevBucketPtr);
1179 currBucket = *prevBucketPtr;
1180
1181 while (currBucket != NULL)
1182 {
1183 if (currBucket == existingElement)
1184 break;
1185 prevBucketPtr = &(currBucket->link);
1186 currBucket = *prevBucketPtr;
1187 }
1188
1189 if (currBucket == NULL)
1190 elog(ERROR, "hash_update_hash_key argument is not in hashtable \"%s\"",
1191 hashp->tabname);
1192
1193 oldPrevPtr = prevBucketPtr;
1194
1195 /*
1196 * Now perform the equivalent of a HASH_ENTER operation to locate the hash
1197 * chain we want to put the entry into.
1198 */
1199 newhashvalue = hashp->hash(newKeyPtr, hashp->keysize);
1200 newbucket = hash_initial_lookup(hashp, newhashvalue, &prevBucketPtr);
1201 currBucket = *prevBucketPtr;
1202
1203 /*
1204 * Follow collision chain looking for matching key
1205 */
1206 match = hashp->match; /* save one fetch in inner loop */
1207 keysize = hashp->keysize; /* ditto */
1208
1209 while (currBucket != NULL)
1210 {
1211 if (currBucket->hashvalue == newhashvalue &&
1212 match(ELEMENTKEY(currBucket), newKeyPtr, keysize) == 0)
1213 break;
1214 prevBucketPtr = &(currBucket->link);
1215 currBucket = *prevBucketPtr;
1216 #ifdef HASH_STATISTICS
1217 hash_collisions++;
1218 hctl->collisions++;
1219 #endif
1220 }
1221
1222 if (currBucket != NULL)
1223 return false; /* collision with an existing entry */
1224
1225 currBucket = existingElement;
1226
1227 /*
1228 * If old and new hash values belong to the same bucket, we need not
1229 * change any chain links, and indeed should not since this simplistic
1230 * update will corrupt the list if currBucket is the last element. (We
1231 * cannot fall out earlier, however, since we need to scan the bucket to
1232 * check for duplicate keys.)
1233 */
1234 if (bucket != newbucket)
1235 {
1236 /* OK to remove record from old hash bucket's chain. */
1237 *oldPrevPtr = currBucket->link;
1238
1239 /* link into new hashbucket chain */
1240 *prevBucketPtr = currBucket;
1241 currBucket->link = NULL;
1242 }
1243
1244 /* copy new key into record */
1245 currBucket->hashvalue = newhashvalue;
1246 hashp->keycopy(ELEMENTKEY(currBucket), newKeyPtr, keysize);
1247
1248 /* rest of record is untouched */
1249
1250 return true;
1251 }
1252
1253 /*
1254 * Allocate a new hashtable entry if possible; return NULL if out of memory.
1255 * (Or, if the underlying space allocator throws error for out-of-memory,
1256 * we won't return at all.)
1257 */
1258 static HASHBUCKET
1259 get_hash_entry(HTAB *hashp, int freelist_idx)
1260 {
1261 HASHHDR *hctl = hashp->hctl;
1262 HASHBUCKET newElement;
1263
1264 for (;;)
1265 {
1266 /* if partitioned, must lock to touch nentries and freeList */
1267 if (IS_PARTITIONED(hctl))
1268 SpinLockAcquire(&hctl->freeList[freelist_idx].mutex);
1269
1270 /* try to get an entry from the freelist */
1271 newElement = hctl->freeList[freelist_idx].freeList;
1272
1273 if (newElement != NULL)
1274 break;
1275
1276 if (IS_PARTITIONED(hctl))
1277 SpinLockRelease(&hctl->freeList[freelist_idx].mutex);
1278
1279 /*
1280 * No free elements in this freelist. In a partitioned table, there
1281 * might be entries in other freelists, but to reduce contention we
1282 * prefer to first try to get another chunk of buckets from the main
1283 * shmem allocator. If that fails, though, we *MUST* root through all
1284 * the other freelists before giving up. There are multiple callers
1285 * that assume that they can allocate every element in the initially
1286 * requested table size, or that deleting an element guarantees they
1287 * can insert a new element, even if shared memory is entirely full.
1288 * Failing because the needed element is in a different freelist is
1289 * not acceptable.
1290 */
1291 if (!element_alloc(hashp, hctl->nelem_alloc, freelist_idx))
1292 {
1293 int borrow_from_idx;
1294
1295 if (!IS_PARTITIONED(hctl))
1296 return NULL; /* out of memory */
1297
1298 /* try to borrow element from another freelist */
1299 borrow_from_idx = freelist_idx;
1300 for (;;)
1301 {
1302 borrow_from_idx = (borrow_from_idx + 1) % NUM_FREELISTS;
1303 if (borrow_from_idx == freelist_idx)
1304 break; /* examined all freelists, fail */
1305
1306 SpinLockAcquire(&(hctl->freeList[borrow_from_idx].mutex));
1307 newElement = hctl->freeList[borrow_from_idx].freeList;
1308
1309 if (newElement != NULL)
1310 {
1311 hctl->freeList[borrow_from_idx].freeList = newElement->link;
1312 SpinLockRelease(&(hctl->freeList[borrow_from_idx].mutex));
1313
1314 /* careful: count the new element in its proper freelist */
1315 SpinLockAcquire(&hctl->freeList[freelist_idx].mutex);
1316 hctl->freeList[freelist_idx].nentries++;
1317 SpinLockRelease(&hctl->freeList[freelist_idx].mutex);
1318
1319 return newElement;
1320 }
1321
1322 SpinLockRelease(&(hctl->freeList[borrow_from_idx].mutex));
1323 }
1324
1325 /* no elements available to borrow either, so out of memory */
1326 return NULL;
1327 }
1328 }
1329
1330 /* remove entry from freelist, bump nentries */
1331 hctl->freeList[freelist_idx].freeList = newElement->link;
1332 hctl->freeList[freelist_idx].nentries++;
1333
1334 if (IS_PARTITIONED(hctl))
1335 SpinLockRelease(&hctl->freeList[freelist_idx].mutex);
1336
1337 return newElement;
1338 }
1339
1340 /*
1341 * hash_get_num_entries -- get the number of entries in a hashtable
1342 */
1343 long
1344 hash_get_num_entries(HTAB *hashp)
1345 {
1346 int i;
1347 long sum = hashp->hctl->freeList[0].nentries;
1348
1349 /*
1350 * We currently don't bother with acquiring the mutexes; it's only
1351 * sensible to call this function if you've got lock on all partitions of
1352 * the table.
1353 */
1354 if (IS_PARTITIONED(hashp->hctl))
1355 {
1356 for (i = 1; i < NUM_FREELISTS; i++)
1357 sum += hashp->hctl->freeList[i].nentries;
1358 }
1359
1360 return sum;
1361 }
1362
1363 /*
1364 * hash_seq_init/_search/_term
1365 * Sequentially search through hash table and return
1366 * all the elements one by one, return NULL when no more.
1367 *
1368 * hash_seq_term should be called if and only if the scan is abandoned before
1369 * completion; if hash_seq_search returns NULL then it has already done the
1370 * end-of-scan cleanup.
1371 *
1372 * NOTE: caller may delete the returned element before continuing the scan.
1373 * However, deleting any other element while the scan is in progress is
1374 * UNDEFINED (it might be the one that curIndex is pointing at!). Also,
1375 * if elements are added to the table while the scan is in progress, it is
1376 * unspecified whether they will be visited by the scan or not.
1377 *
1378 * NOTE: it is possible to use hash_seq_init/hash_seq_search without any
1379 * worry about hash_seq_term cleanup, if the hashtable is first locked against
1380 * further insertions by calling hash_freeze.
1381 *
1382 * NOTE: to use this with a partitioned hashtable, caller had better hold
1383 * at least shared lock on all partitions of the table throughout the scan!
1384 * We can cope with insertions or deletions by our own backend, but *not*
1385 * with concurrent insertions or deletions by another.
1386 */
1387 void
1388 hash_seq_init(HASH_SEQ_STATUS *status, HTAB *hashp)
1389 {
1390 status->hashp = hashp;
1391 status->curBucket = 0;
1392 status->curEntry = NULL;
1393 status->hasHashvalue = false;
1394 if (!hashp->frozen)
1395 register_seq_scan(hashp);
1396 }
1397
1398 /*
1399 * Same as above but scan by the given hash value.
1400 * See also hash_seq_search().
1401 *
1402 * NOTE: the default hash function doesn't match syscache hash function.
1403 * Thus, if you're going to use this function in syscache callback, make sure
1404 * you're using custom hash function. See relatt_cache_syshash()
1405 * for example.
1406 */
1407 void
1408 hash_seq_init_with_hash_value(HASH_SEQ_STATUS *status, HTAB *hashp,
1409 uint32 hashvalue)
1410 {
1411 HASHBUCKET *bucketPtr;
1412
1413 hash_seq_init(status, hashp);
1414
1415 status->hasHashvalue = true;
1416 status->hashvalue = hashvalue;
1417
1418 status->curBucket = hash_initial_lookup(hashp, hashvalue, &bucketPtr);
1419 status->curEntry = *bucketPtr;
1420 }
1421
1422 void *
1423 hash_seq_search(HASH_SEQ_STATUS *status)
1424 {
1425 HTAB *hashp;
1426 HASHHDR *hctl;
1427 uint32 max_bucket;
1428 long ssize;
1429 long segment_num;
1430 long segment_ndx;
1431 HASHSEGMENT segp;
1432 uint32 curBucket;
1433 HASHELEMENT *curElem;
1434
1435 if (status->hasHashvalue)
1436 {
1437 /*
1438 * Scan entries only in the current bucket because only this bucket
1439 * can contain entries with the given hash value.
1440 */
1441 while ((curElem = status->curEntry) != NULL)
1442 {
1443 status->curEntry = curElem->link;
1444 if (status->hashvalue != curElem->hashvalue)
1445 continue;
1446 return (void *) ELEMENTKEY(curElem);
1447 }
1448
1449 hash_seq_term(status);
1450 return NULL;
1451 }
1452
1453 if ((curElem = status->curEntry) != NULL)
1454 {
1455 /* Continuing scan of curBucket... */
1456 status->curEntry = curElem->link;
1457 if (status->curEntry == NULL) /* end of this bucket */
1458 ++status->curBucket;
1459 return ELEMENTKEY(curElem);
1460 }
1461
1462 /*
1463 * Search for next nonempty bucket starting at curBucket.
1464 */
1465 curBucket = status->curBucket;
1466 hashp = status->hashp;
1467 hctl = hashp->hctl;
1468 ssize = hashp->ssize;
1469 max_bucket = hctl->max_bucket;
1470
1471 if (curBucket > max_bucket)
1472 {
1473 hash_seq_term(status);
1474 return NULL; /* search is done */
1475 }
1476
1477 /*
1478 * first find the right segment in the table directory.
1479 */
1480 segment_num = curBucket >> hashp->sshift;
1481 segment_ndx = MOD(curBucket, ssize);
1482
1483 segp = hashp->dir[segment_num];
1484
1485 /*
1486 * Pick up the first item in this bucket's chain. If chain is not empty
1487 * we can begin searching it. Otherwise we have to advance to find the
1488 * next nonempty bucket. We try to optimize that case since searching a
1489 * near-empty hashtable has to iterate this loop a lot.
1490 */
1491 while ((curElem = segp[segment_ndx]) == NULL)
1492 {
1493 /* empty bucket, advance to next */
1494 if (++curBucket > max_bucket)
1495 {
1496 status->curBucket = curBucket;
1497 hash_seq_term(status);
1498 return NULL; /* search is done */
1499 }
1500 if (++segment_ndx >= ssize)
1501 {
1502 segment_num++;
1503 segment_ndx = 0;
1504 segp = hashp->dir[segment_num];
1505 }
1506 }
1507
1508 /* Begin scan of curBucket... */
1509 status->curEntry = curElem->link;
1510 if (status->curEntry == NULL) /* end of this bucket */
1511 ++curBucket;
1512 status->curBucket = curBucket;
1513 return ELEMENTKEY(curElem);
1514 }
1515
1516 void
1517 hash_seq_term(HASH_SEQ_STATUS *status)
1518 {
1519 if (!status->hashp->frozen)
1520 deregister_seq_scan(status->hashp);
1521 }
1522
1523 /*
1524 * hash_freeze
1525 * Freeze a hashtable against future insertions (deletions are
1526 * still allowed)
1527 *
1528 * The reason for doing this is that by preventing any more bucket splits,
1529 * we no longer need to worry about registering hash_seq_search scans,
1530 * and thus caller need not be careful about ensuring hash_seq_term gets
1531 * called at the right times.
1532 *
1533 * Multiple calls to hash_freeze() are allowed, but you can't freeze a table
1534 * with active scans (since hash_seq_term would then do the wrong thing).
1535 */
1536 void
1537 hash_freeze(HTAB *hashp)
1538 {
1539 if (hashp->isshared)
1540 elog(ERROR, "cannot freeze shared hashtable \"%s\"", hashp->tabname);
1541 if (!hashp->frozen && has_seq_scans(hashp))
1542 elog(ERROR, "cannot freeze hashtable \"%s\" because it has active scans",
1543 hashp->tabname);
1544 hashp->frozen = true;
1545 }
1546
1547
1548 /********************************* UTILITIES ************************/
1549
1550 /*
1551 * Expand the table by adding one more hash bucket.
1552 */
1553 static bool
1554 expand_table(HTAB *hashp)
1555 {
1556 HASHHDR *hctl = hashp->hctl;
1557 HASHSEGMENT old_seg,
1558 new_seg;
1559 long old_bucket,
1560 new_bucket;
1561 long new_segnum,
1562 new_segndx;
1563 long old_segnum,
1564 old_segndx;
1565 HASHBUCKET *oldlink,
1566 *newlink;
1567 HASHBUCKET currElement,
1568 nextElement;
1569
1570 Assert(!IS_PARTITIONED(hctl));
1571
1572 #ifdef HASH_STATISTICS
1573 hash_expansions++;
1574 #endif
1575
1576 new_bucket = hctl->max_bucket + 1;
1577 new_segnum = new_bucket >> hashp->sshift;
1578 new_segndx = MOD(new_bucket, hashp->ssize);
1579
1580 if (new_segnum >= hctl->nsegs)
1581 {
1582 /* Allocate new segment if necessary -- could fail if dir full */
1583 if (new_segnum >= hctl->dsize)
1584 if (!dir_realloc(hashp))
1585 return false;
1586 if (!(hashp->dir[new_segnum] = seg_alloc(hashp)))
1587 return false;
1588 hctl->nsegs++;
1589 }
1590
1591 /* OK, we created a new bucket */
1592 hctl->max_bucket++;
1593
1594 /*
1595 * *Before* changing masks, find old bucket corresponding to same hash
1596 * values; values in that bucket may need to be relocated to new bucket.
1597 * Note that new_bucket is certainly larger than low_mask at this point,
1598 * so we can skip the first step of the regular hash mask calc.
1599 */
1600 old_bucket = (new_bucket & hctl->low_mask);
1601
1602 /*
1603 * If we crossed a power of 2, readjust masks.
1604 */
1605 if ((uint32) new_bucket > hctl->high_mask)
1606 {
1607 hctl->low_mask = hctl->high_mask;
1608 hctl->high_mask = (uint32) new_bucket | hctl->low_mask;
1609 }
1610
1611 /*
1612 * Relocate records to the new bucket. NOTE: because of the way the hash
1613 * masking is done in calc_bucket, only one old bucket can need to be
1614 * split at this point. With a different way of reducing the hash value,
1615 * that might not be true!
1616 */
1617 old_segnum = old_bucket >> hashp->sshift;
1618 old_segndx = MOD(old_bucket, hashp->ssize);
1619
1620 old_seg = hashp->dir[old_segnum];
1621 new_seg = hashp->dir[new_segnum];
1622
1623 oldlink = &old_seg[old_segndx];
1624 newlink = &new_seg[new_segndx];
1625
1626 for (currElement = *oldlink;
1627 currElement != NULL;
1628 currElement = nextElement)
1629 {
1630 nextElement = currElement->link;
1631 if ((long) calc_bucket(hctl, currElement->hashvalue) == old_bucket)
1632 {
1633 *oldlink = currElement;
1634 oldlink = &currElement->link;
1635 }
1636 else
1637 {
1638 *newlink = currElement;
1639 newlink = &currElement->link;
1640 }
1641 }
1642 /* don't forget to terminate the rebuilt hash chains... */
1643 *oldlink = NULL;
1644 *newlink = NULL;
1645
1646 return true;
1647 }
1648
1649
1650 static bool
1651 dir_realloc(HTAB *hashp)
1652 {
1653 HASHSEGMENT *p;
1654 HASHSEGMENT *old_p;
1655 long new_dsize;
1656 long old_dirsize;
1657 long new_dirsize;
1658
1659 if (hashp->hctl->max_dsize != NO_MAX_DSIZE)
1660 return false;
1661
1662 /* Reallocate directory */
1663 new_dsize = hashp->hctl->dsize << 1;
1664 old_dirsize = hashp->hctl->dsize * sizeof(HASHSEGMENT);
1665 new_dirsize = new_dsize * sizeof(HASHSEGMENT);
1666
1667 old_p = hashp->dir;
1668 CurrentDynaHashCxt = hashp->hcxt;
1669 p = (HASHSEGMENT *) hashp->alloc((Size) new_dirsize);
1670
1671 if (p != NULL)
1672 {
1673 memcpy(p, old_p, old_dirsize);
1674 MemSet(((char *) p) + old_dirsize, 0, new_dirsize - old_dirsize);
1675 hashp->dir = p;
1676 hashp->hctl->dsize = new_dsize;
1677
1678 /* XXX assume the allocator is palloc, so we know how to free */
1679 Assert(hashp->alloc == DynaHashAlloc);
1680 pfree(old_p);
1681
1682 return true;
1683 }
1684
1685 return false;
1686 }
1687
1688
1689 static HASHSEGMENT
1690 seg_alloc(HTAB *hashp)
1691 {
1692 HASHSEGMENT segp;
1693
1694 CurrentDynaHashCxt = hashp->hcxt;
1695 segp = (HASHSEGMENT) hashp->alloc(sizeof(HASHBUCKET) * hashp->ssize);
1696
1697 if (!segp)
1698 return NULL;
1699
1700 MemSet(segp, 0, sizeof(HASHBUCKET) * hashp->ssize);
1701
1702 return segp;
1703 }
1704
1705 /*
1706 * allocate some new elements and link them into the indicated free list
1707 */
1708 static bool
1709 element_alloc(HTAB *hashp, int nelem, int freelist_idx)
1710 {
1711 HASHHDR *hctl = hashp->hctl;
1712 Size elementSize;
1713 HASHELEMENT *firstElement;
1714 HASHELEMENT *tmpElement;
1715 HASHELEMENT *prevElement;
1716 int i;
1717
1718 if (hashp->isfixed)
1719 return false;
1720
1721 /* Each element has a HASHELEMENT header plus user data. */
1722 elementSize = MAXALIGN(sizeof(HASHELEMENT)) + MAXALIGN(hctl->entrysize);
1723
1724 CurrentDynaHashCxt = hashp->hcxt;
1725 firstElement = (HASHELEMENT *) hashp->alloc(nelem * elementSize);
1726
1727 if (!firstElement)
1728 return false;
1729
1730 /* prepare to link all the new entries into the freelist */
1731 prevElement = NULL;
1732 tmpElement = firstElement;
1733 for (i = 0; i < nelem; i++)
1734 {
1735 tmpElement->link = prevElement;
1736 prevElement = tmpElement;
1737 tmpElement = (HASHELEMENT *) (((char *) tmpElement) + elementSize);
1738 }
1739
1740 /* if partitioned, must lock to touch freeList */
1741 if (IS_PARTITIONED(hctl))
1742 SpinLockAcquire(&hctl->freeList[freelist_idx].mutex);
1743
1744 /* freelist could be nonempty if two backends did this concurrently */
1745 firstElement->link = hctl->freeList[freelist_idx].freeList;
1746 hctl->freeList[freelist_idx].freeList = prevElement;
1747
1748 if (IS_PARTITIONED(hctl))
1749 SpinLockRelease(&hctl->freeList[freelist_idx].mutex);
1750
1751 return true;
1752 }
1753
1754 /*
1755 * Do initial lookup of a bucket for the given hash value, retrieving its
1756 * bucket number and its hash bucket.
1757 */
1758 static inline uint32
1759 hash_initial_lookup(HTAB *hashp, uint32 hashvalue, HASHBUCKET **bucketptr)
1760 {
1761 HASHHDR *hctl = hashp->hctl;
1762 HASHSEGMENT segp;
1763 long segment_num;
1764 long segment_ndx;
1765 uint32 bucket;
1766
1767 bucket = calc_bucket(hctl, hashvalue);
1768
1769 segment_num = bucket >> hashp->sshift;
1770 segment_ndx = MOD(bucket, hashp->ssize);
1771
1772 segp = hashp->dir[segment_num];
1773
1774 if (segp == NULL)
1775 hash_corrupted(hashp);
1776
1777 *bucketptr = &segp[segment_ndx];
1778 return bucket;
1779 }
1780
1781 /* complain when we have detected a corrupted hashtable */
1782 static void
1783 hash_corrupted(HTAB *hashp)
1784 {
1785 /*
1786 * If the corruption is in a shared hashtable, we'd better force a
1787 * systemwide restart. Otherwise, just shut down this one backend.
1788 */
1789 if (hashp->isshared)
1790 elog(PANIC, "hash table \"%s\" corrupted", hashp->tabname);
1791 else
1792 elog(FATAL, "hash table \"%s\" corrupted", hashp->tabname);
1793 }
1794
1795 /* calculate ceil(log base 2) of num */
1796 int
1797 my_log2(long num)
1798 {
1799 /*
1800 * guard against too-large input, which would be invalid for
1801 * pg_ceil_log2_*()
1802 */
1803 if (num > LONG_MAX / 2)
1804 num = LONG_MAX / 2;
1805
1806 #if SIZEOF_LONG < 8
1807 return pg_ceil_log2_32(num);
1808 #else
1809 return pg_ceil_log2_64(num);
1810 #endif
1811 }
1812
1813 /* calculate first power of 2 >= num, bounded to what will fit in a long */
1814 static long
1815 next_pow2_long(long num)
1816 {
1817 /* my_log2's internal range check is sufficient */
1818 return 1L << my_log2(num);
1819 }
1820
1821 /* calculate first power of 2 >= num, bounded to what will fit in an int */
1822 static int
1823 next_pow2_int(long num)
1824 {
1825 if (num > INT_MAX / 2)
1826 num = INT_MAX / 2;
1827 return 1 << my_log2(num);
1828 }
1829
1830
1831 /************************* SEQ SCAN TRACKING ************************/
1832
1833 /*
1834 * We track active hash_seq_search scans here. The need for this mechanism
1835 * comes from the fact that a scan will get confused if a bucket split occurs
1836 * while it's in progress: it might visit entries twice, or even miss some
1837 * entirely (if it's partway through the same bucket that splits). Hence
1838 * we want to inhibit bucket splits if there are any active scans on the
1839 * table being inserted into. This is a fairly rare case in current usage,
1840 * so just postponing the split until the next insertion seems sufficient.
1841 *
1842 * Given present usages of the function, only a few scans are likely to be
1843 * open concurrently; so a finite-size stack of open scans seems sufficient,
1844 * and we don't worry that linear search is too slow. Note that we do
1845 * allow multiple scans of the same hashtable to be open concurrently.
1846 *
1847 * This mechanism can support concurrent scan and insertion in a shared
1848 * hashtable if it's the same backend doing both. It would fail otherwise,
1849 * but locking reasons seem to preclude any such scenario anyway, so we don't
1850 * worry.
1851 *
1852 * This arrangement is reasonably robust if a transient hashtable is deleted
1853 * without notifying us. The absolute worst case is we might inhibit splits
1854 * in another table created later at exactly the same address. We will give
1855 * a warning at transaction end for reference leaks, so any bugs leading to
1856 * lack of notification should be easy to catch.
1857 */
1858
1859 #define MAX_SEQ_SCANS 100
1860
1861 static HTAB *seq_scan_tables[MAX_SEQ_SCANS]; /* tables being scanned */
1862 static int seq_scan_level[MAX_SEQ_SCANS]; /* subtransaction nest level */
1863 static int num_seq_scans = 0;
1864
1865
1866 /* Register a table as having an active hash_seq_search scan */
1867 static void
1868 register_seq_scan(HTAB *hashp)
1869 {
1870 if (num_seq_scans >= MAX_SEQ_SCANS)
1871 elog(ERROR, "too many active hash_seq_search scans, cannot start one on \"%s\"",
1872 hashp->tabname);
1873 seq_scan_tables[num_seq_scans] = hashp;
1874 seq_scan_level[num_seq_scans] = GetCurrentTransactionNestLevel();
1875 num_seq_scans++;
1876 }
1877
1878 /* Deregister an active scan */
1879 static void
1880 deregister_seq_scan(HTAB *hashp)
1881 {
1882 int i;
1883
1884 /* Search backward since it's most likely at the stack top */
1885 for (i = num_seq_scans - 1; i >= 0; i--)
1886 {
1887 if (seq_scan_tables[i] == hashp)
1888 {
1889 seq_scan_tables[i] = seq_scan_tables[num_seq_scans - 1];
1890 seq_scan_level[i] = seq_scan_level[num_seq_scans - 1];
1891 num_seq_scans--;
1892 return;
1893 }
1894 }
1895 elog(ERROR, "no hash_seq_search scan for hash table \"%s\"",
1896 hashp->tabname);
1897 }
1898
1899 /* Check if a table has any active scan */
1900 static bool
1901 has_seq_scans(HTAB *hashp)
1902 {
1903 int i;
1904
1905 for (i = 0; i < num_seq_scans; i++)
1906 {
1907 if (seq_scan_tables[i] == hashp)
1908 return true;
1909 }
1910 return false;
1911 }
1912
1913 /* Clean up any open scans at end of transaction */
1914 void
1915 AtEOXact_HashTables(bool isCommit)
1916 {
1917 /*
1918 * During abort cleanup, open scans are expected; just silently clean 'em
1919 * out. An open scan at commit means someone forgot a hash_seq_term()
1920 * call, so complain.
1921 *
1922 * Note: it's tempting to try to print the tabname here, but refrain for
1923 * fear of touching deallocated memory. This isn't a user-facing message
1924 * anyway, so it needn't be pretty.
1925 */
1926 if (isCommit)
1927 {
1928 int i;
1929
1930 for (i = 0; i < num_seq_scans; i++)
1931 {
1932 elog(WARNING, "leaked hash_seq_search scan for hash table %p",
1933 seq_scan_tables[i]);
1934 }
1935 }
1936 num_seq_scans = 0;
1937 }
1938
1939 /* Clean up any open scans at end of subtransaction */
1940 void
1941 AtEOSubXact_HashTables(bool isCommit, int nestDepth)
1942 {
1943 int i;
1944
1945 /*
1946 * Search backward to make cleanup easy. Note we must check all entries,
1947 * not only those at the end of the array, because deletion technique
1948 * doesn't keep them in order.
1949 */
1950 for (i = num_seq_scans - 1; i >= 0; i--)
1951 {
1952 if (seq_scan_level[i] >= nestDepth)
1953 {
1954 if (isCommit)
1955 elog(WARNING, "leaked hash_seq_search scan for hash table %p",
1956 seq_scan_tables[i]);
1957 seq_scan_tables[i] = seq_scan_tables[num_seq_scans - 1];
1958 seq_scan_level[i] = seq_scan_level[num_seq_scans - 1];
1959 num_seq_scans--;
1960 }
1961 }
1962 }