Line data Source code
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 ejp@ausmelb.oz, 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 ejp@ausmelb.oz. WARNING: relies on pre-processor
87 : * concatenation property, in probably unnecessary code 'optimization'.
88 : *
89 : * Modified margo@postgres.berkeley.edu February 1990
90 : * added multiple table interface
91 : * Modified by sullivan@postgres.berkeley.edu 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 2884254 : DynaHashAlloc(Size size)
292 : {
293 : Assert(MemoryContextIsValid(CurrentDynaHashCxt));
294 2884254 : 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 1035028 : string_compare(const char *key1, const char *key2, Size keysize)
308 : {
309 1035028 : 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 538824 : 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 538824 : if (flags & HASH_SHARED_MEM)
376 : {
377 : /* Set up to allocate the hash header */
378 18302 : CurrentDynaHashCxt = TopMemoryContext;
379 : }
380 : else
381 : {
382 : /* Create the hash table's private memory context */
383 520522 : if (flags & HASH_CONTEXT)
384 213702 : CurrentDynaHashCxt = info->hcxt;
385 : else
386 306820 : CurrentDynaHashCxt = TopMemoryContext;
387 520522 : CurrentDynaHashCxt = AllocSetContextCreate(CurrentDynaHashCxt,
388 : "dynahash",
389 : ALLOCSET_DEFAULT_SIZES);
390 : }
391 :
392 : /* Initialize the hash header, plus a copy of the table name */
393 538824 : hashp = (HTAB *) DynaHashAlloc(sizeof(HTAB) + strlen(tabname) + 1);
394 7004712 : MemSet(hashp, 0, sizeof(HTAB));
395 :
396 538824 : hashp->tabname = (char *) (hashp + 1);
397 538824 : strcpy(hashp->tabname, tabname);
398 :
399 : /* If we have a private context, label it with hashtable's name */
400 538824 : if (!(flags & HASH_SHARED_MEM))
401 520522 : MemoryContextSetIdentifier(CurrentDynaHashCxt, hashp->tabname);
402 :
403 : /*
404 : * Select the appropriate hash function (see comments at head of file).
405 : */
406 538824 : if (flags & HASH_FUNCTION)
407 : {
408 : Assert(!(flags & (HASH_BLOBS | HASH_STRINGS)));
409 26584 : hashp->hash = info->hash;
410 : }
411 512240 : else if (flags & HASH_BLOBS)
412 : {
413 : Assert(!(flags & HASH_STRINGS));
414 : /* We can optimize hashing for common key sizes */
415 410824 : if (info->keysize == sizeof(uint32))
416 237136 : hashp->hash = uint32_hash;
417 : else
418 173688 : hashp->hash = tag_hash;
419 : }
420 : else
421 : {
422 : /*
423 : * string_hash used to be considered the default hash method, and in a
424 : * non-assert build it effectively still is. But we now consider it
425 : * an assertion error to not say HASH_STRINGS explicitly. To help
426 : * catch mistaken usage of HASH_STRINGS, we also insist on a
427 : * reasonably long string length: if the keysize is only 4 or 8 bytes,
428 : * it's almost certainly an integer or pointer not a string.
429 : */
430 : Assert(flags & HASH_STRINGS);
431 : Assert(info->keysize > 8);
432 :
433 101416 : hashp->hash = string_hash;
434 : }
435 :
436 : /*
437 : * If you don't specify a match function, it defaults to string_compare if
438 : * you used string_hash, and to memcmp otherwise.
439 : *
440 : * Note: explicitly specifying string_hash is deprecated, because this
441 : * might not work for callers in loadable modules on some platforms due to
442 : * referencing a trampoline instead of the string_hash function proper.
443 : * Specify HASH_STRINGS instead.
444 : */
445 538824 : if (flags & HASH_COMPARE)
446 11804 : hashp->match = info->match;
447 527020 : else if (hashp->hash == string_hash)
448 101416 : hashp->match = (HashCompareFunc) string_compare;
449 : else
450 425604 : hashp->match = memcmp;
451 :
452 : /*
453 : * Similarly, the key-copying function defaults to strlcpy or memcpy.
454 : */
455 538824 : if (flags & HASH_KEYCOPY)
456 0 : hashp->keycopy = info->keycopy;
457 538824 : else if (hashp->hash == string_hash)
458 : {
459 : /*
460 : * The signature of keycopy is meant for memcpy(), which returns
461 : * void*, but strlcpy() returns size_t. Since we never use the return
462 : * value of keycopy, and size_t is pretty much always the same size as
463 : * void *, this should be safe. The extra cast in the middle is to
464 : * avoid warnings from -Wcast-function-type.
465 : */
466 101416 : hashp->keycopy = (HashCopyFunc) (pg_funcptr_t) strlcpy;
467 : }
468 : else
469 437408 : hashp->keycopy = memcpy;
470 :
471 : /* And select the entry allocation function, too. */
472 538824 : if (flags & HASH_ALLOC)
473 18302 : hashp->alloc = info->alloc;
474 : else
475 520522 : hashp->alloc = DynaHashAlloc;
476 :
477 538824 : if (flags & HASH_SHARED_MEM)
478 : {
479 : /*
480 : * ctl structure and directory are preallocated for shared memory
481 : * tables. Note that HASH_DIRSIZE and HASH_ALLOC had better be set as
482 : * well.
483 : */
484 18302 : hashp->hctl = info->hctl;
485 18302 : hashp->dir = (HASHSEGMENT *) (((char *) info->hctl) + sizeof(HASHHDR));
486 18302 : hashp->hcxt = NULL;
487 18302 : hashp->isshared = true;
488 :
489 : /* hash table already exists, we're just attaching to it */
490 18302 : if (flags & HASH_ATTACH)
491 : {
492 : /* make local copies of some heavily-used values */
493 0 : hctl = hashp->hctl;
494 0 : hashp->keysize = hctl->keysize;
495 0 : hashp->ssize = hctl->ssize;
496 0 : hashp->sshift = hctl->sshift;
497 :
498 0 : return hashp;
499 : }
500 : }
501 : else
502 : {
503 : /* setup hash table defaults */
504 520522 : hashp->hctl = NULL;
505 520522 : hashp->dir = NULL;
506 520522 : hashp->hcxt = CurrentDynaHashCxt;
507 520522 : hashp->isshared = false;
508 : }
509 :
510 538824 : if (!hashp->hctl)
511 : {
512 520522 : hashp->hctl = (HASHHDR *) hashp->alloc(sizeof(HASHHDR));
513 520522 : if (!hashp->hctl)
514 0 : ereport(ERROR,
515 : (errcode(ERRCODE_OUT_OF_MEMORY),
516 : errmsg("out of memory")));
517 : }
518 :
519 538824 : hashp->frozen = false;
520 :
521 538824 : hdefault(hashp);
522 :
523 538824 : hctl = hashp->hctl;
524 :
525 538824 : if (flags & HASH_PARTITION)
526 : {
527 : /* Doesn't make sense to partition a local hash table */
528 : Assert(flags & HASH_SHARED_MEM);
529 :
530 : /*
531 : * The number of partitions had better be a power of 2. Also, it must
532 : * be less than INT_MAX (see init_htab()), so call the int version of
533 : * next_pow2.
534 : */
535 : Assert(info->num_partitions == next_pow2_int(info->num_partitions));
536 :
537 10160 : hctl->num_partitions = info->num_partitions;
538 : }
539 :
540 538824 : if (flags & HASH_SEGMENT)
541 : {
542 0 : hctl->ssize = info->ssize;
543 0 : hctl->sshift = my_log2(info->ssize);
544 : /* ssize had better be a power of 2 */
545 : Assert(hctl->ssize == (1L << hctl->sshift));
546 : }
547 :
548 : /*
549 : * SHM hash tables have fixed directory size passed by the caller.
550 : */
551 538824 : if (flags & HASH_DIRSIZE)
552 : {
553 18302 : hctl->max_dsize = info->max_dsize;
554 18302 : hctl->dsize = info->dsize;
555 : }
556 :
557 : /* remember the entry sizes, too */
558 538824 : hctl->keysize = info->keysize;
559 538824 : hctl->entrysize = info->entrysize;
560 :
561 : /* make local copies of heavily-used constant fields */
562 538824 : hashp->keysize = hctl->keysize;
563 538824 : hashp->ssize = hctl->ssize;
564 538824 : hashp->sshift = hctl->sshift;
565 :
566 : /* Build the hash directory structure */
567 538824 : if (!init_htab(hashp, nelem))
568 0 : elog(ERROR, "failed to initialize hash table \"%s\"", hashp->tabname);
569 :
570 : /*
571 : * For a shared hash table, preallocate the requested number of elements.
572 : * This reduces problems with run-time out-of-shared-memory conditions.
573 : *
574 : * For a non-shared hash table, preallocate the requested number of
575 : * elements if it's less than our chosen nelem_alloc. This avoids wasting
576 : * space if the caller correctly estimates a small table size.
577 : */
578 538824 : if ((flags & HASH_SHARED_MEM) ||
579 520522 : nelem < hctl->nelem_alloc)
580 : {
581 : int i,
582 : freelist_partitions,
583 : nelem_alloc,
584 : nelem_alloc_first;
585 :
586 : /*
587 : * If hash table is partitioned, give each freelist an equal share of
588 : * the initial allocation. Otherwise only freeList[0] is used.
589 : */
590 247622 : if (IS_PARTITIONED(hashp->hctl))
591 10160 : freelist_partitions = NUM_FREELISTS;
592 : else
593 237462 : freelist_partitions = 1;
594 :
595 247622 : nelem_alloc = nelem / freelist_partitions;
596 247622 : if (nelem_alloc <= 0)
597 0 : nelem_alloc = 1;
598 :
599 : /*
600 : * Make sure we'll allocate all the requested elements; freeList[0]
601 : * gets the excess if the request isn't divisible by NUM_FREELISTS.
602 : */
603 247622 : if (nelem_alloc * freelist_partitions < nelem)
604 94 : nelem_alloc_first =
605 94 : nelem - nelem_alloc * (freelist_partitions - 1);
606 : else
607 247528 : nelem_alloc_first = nelem_alloc;
608 :
609 810204 : for (i = 0; i < freelist_partitions; i++)
610 : {
611 562582 : int temp = (i == 0) ? nelem_alloc_first : nelem_alloc;
612 :
613 562582 : if (!element_alloc(hashp, temp, i))
614 0 : ereport(ERROR,
615 : (errcode(ERRCODE_OUT_OF_MEMORY),
616 : errmsg("out of memory")));
617 : }
618 : }
619 :
620 538824 : if (flags & HASH_FIXED_SIZE)
621 6096 : hashp->isfixed = true;
622 538824 : return hashp;
623 : }
624 :
625 : /*
626 : * Set default HASHHDR parameters.
627 : */
628 : static void
629 538824 : hdefault(HTAB *hashp)
630 : {
631 538824 : HASHHDR *hctl = hashp->hctl;
632 :
633 57654168 : MemSet(hctl, 0, sizeof(HASHHDR));
634 :
635 538824 : hctl->dsize = DEF_DIRSIZE;
636 538824 : hctl->nsegs = 0;
637 :
638 538824 : hctl->num_partitions = 0; /* not partitioned */
639 :
640 : /* table has no fixed maximum size */
641 538824 : hctl->max_dsize = NO_MAX_DSIZE;
642 :
643 538824 : hctl->ssize = DEF_SEGSIZE;
644 538824 : hctl->sshift = DEF_SEGSIZE_SHIFT;
645 :
646 : #ifdef HASH_STATISTICS
647 : hctl->accesses = hctl->collisions = 0;
648 : #endif
649 538824 : }
650 :
651 : /*
652 : * Given the user-specified entry size, choose nelem_alloc, ie, how many
653 : * elements to add to the hash table when we need more.
654 : */
655 : static int
656 572984 : choose_nelem_alloc(Size entrysize)
657 : {
658 : int nelem_alloc;
659 : Size elementSize;
660 : Size allocSize;
661 :
662 : /* Each element has a HASHELEMENT header plus user data. */
663 : /* NB: this had better match element_alloc() */
664 572984 : elementSize = MAXALIGN(sizeof(HASHELEMENT)) + MAXALIGN(entrysize);
665 :
666 : /*
667 : * The idea here is to choose nelem_alloc at least 32, but round up so
668 : * that the allocation request will be a power of 2 or just less. This
669 : * makes little difference for hash tables in shared memory, but for hash
670 : * tables managed by palloc, the allocation request will be rounded up to
671 : * a power of 2 anyway. If we fail to take this into account, we'll waste
672 : * as much as half the allocated space.
673 : */
674 572984 : allocSize = 32 * 4; /* assume elementSize at least 8 */
675 : do
676 : {
677 2357552 : allocSize <<= 1;
678 2357552 : nelem_alloc = allocSize / elementSize;
679 2357552 : } while (nelem_alloc < 32);
680 :
681 572984 : return nelem_alloc;
682 : }
683 :
684 : /*
685 : * Compute derived fields of hctl and build the initial directory/segment
686 : * arrays
687 : */
688 : static bool
689 538824 : init_htab(HTAB *hashp, long nelem)
690 : {
691 538824 : HASHHDR *hctl = hashp->hctl;
692 : HASHSEGMENT *segp;
693 : int nbuckets;
694 : int nsegs;
695 : int i;
696 :
697 : /*
698 : * initialize mutexes if it's a partitioned table
699 : */
700 538824 : if (IS_PARTITIONED(hctl))
701 335280 : for (i = 0; i < NUM_FREELISTS; i++)
702 325120 : SpinLockInit(&(hctl->freeList[i].mutex));
703 :
704 : /*
705 : * Allocate space for the next greater power of two number of buckets,
706 : * assuming a desired maximum load factor of 1.
707 : */
708 538824 : nbuckets = next_pow2_int(nelem);
709 :
710 : /*
711 : * In a partitioned table, nbuckets must be at least equal to
712 : * num_partitions; were it less, keys with apparently different partition
713 : * numbers would map to the same bucket, breaking partition independence.
714 : * (Normally nbuckets will be much bigger; this is just a safety check.)
715 : */
716 538824 : while (nbuckets < hctl->num_partitions)
717 0 : nbuckets <<= 1;
718 :
719 538824 : hctl->max_bucket = hctl->low_mask = nbuckets - 1;
720 538824 : hctl->high_mask = (nbuckets << 1) - 1;
721 :
722 : /*
723 : * Figure number of directory segments needed, round up to a power of 2
724 : */
725 538824 : nsegs = (nbuckets - 1) / hctl->ssize + 1;
726 538824 : nsegs = next_pow2_int(nsegs);
727 :
728 : /*
729 : * Make sure directory is big enough. If pre-allocated directory is too
730 : * small, choke (caller screwed up).
731 : */
732 538824 : if (nsegs > hctl->dsize)
733 : {
734 0 : if (!(hashp->dir))
735 0 : hctl->dsize = nsegs;
736 : else
737 0 : return false;
738 : }
739 :
740 : /* Allocate a directory */
741 538824 : if (!(hashp->dir))
742 : {
743 520522 : CurrentDynaHashCxt = hashp->hcxt;
744 520522 : hashp->dir = (HASHSEGMENT *)
745 520522 : hashp->alloc(hctl->dsize * sizeof(HASHSEGMENT));
746 520522 : if (!hashp->dir)
747 0 : return false;
748 : }
749 :
750 : /* Allocate initial segments */
751 1646946 : for (segp = hashp->dir; hctl->nsegs < nsegs; hctl->nsegs++, segp++)
752 : {
753 1108122 : *segp = seg_alloc(hashp);
754 1108122 : if (*segp == NULL)
755 0 : return false;
756 : }
757 :
758 : /* Choose number of entries to allocate at a time */
759 538824 : hctl->nelem_alloc = choose_nelem_alloc(hctl->entrysize);
760 :
761 : #ifdef HASH_DEBUG
762 : fprintf(stderr, "init_htab:\n%s%p\n%s%ld\n%s%ld\n%s%d\n%s%ld\n%s%u\n%s%x\n%s%x\n%s%ld\n",
763 : "TABLE POINTER ", hashp,
764 : "DIRECTORY SIZE ", hctl->dsize,
765 : "SEGMENT SIZE ", hctl->ssize,
766 : "SEGMENT SHIFT ", hctl->sshift,
767 : "MAX BUCKET ", hctl->max_bucket,
768 : "HIGH MASK ", hctl->high_mask,
769 : "LOW MASK ", hctl->low_mask,
770 : "NSEGS ", hctl->nsegs);
771 : #endif
772 538824 : return true;
773 : }
774 :
775 : /*
776 : * Estimate the space needed for a hashtable containing the given number
777 : * of entries of given size.
778 : * NOTE: this is used to estimate the footprint of hashtables in shared
779 : * memory; therefore it does not count HTAB which is in local memory.
780 : * NB: assumes that all hash structure parameters have default values!
781 : */
782 : Size
783 34160 : hash_estimate_size(long num_entries, Size entrysize)
784 : {
785 : Size size;
786 : long nBuckets,
787 : nSegments,
788 : nDirEntries,
789 : nElementAllocs,
790 : elementSize,
791 : elementAllocCnt;
792 :
793 : /* estimate number of buckets wanted */
794 34160 : nBuckets = next_pow2_long(num_entries);
795 : /* # of segments needed for nBuckets */
796 34160 : nSegments = next_pow2_long((nBuckets - 1) / DEF_SEGSIZE + 1);
797 : /* directory entries */
798 34160 : nDirEntries = DEF_DIRSIZE;
799 34160 : while (nDirEntries < nSegments)
800 0 : nDirEntries <<= 1; /* dir_alloc doubles dsize at each call */
801 :
802 : /* fixed control info */
803 34160 : size = MAXALIGN(sizeof(HASHHDR)); /* but not HTAB, per above */
804 : /* directory */
805 34160 : size = add_size(size, mul_size(nDirEntries, sizeof(HASHSEGMENT)));
806 : /* segments */
807 34160 : size = add_size(size, mul_size(nSegments,
808 : MAXALIGN(DEF_SEGSIZE * sizeof(HASHBUCKET))));
809 : /* elements --- allocated in groups of choose_nelem_alloc() entries */
810 34160 : elementAllocCnt = choose_nelem_alloc(entrysize);
811 34160 : nElementAllocs = (num_entries - 1) / elementAllocCnt + 1;
812 34160 : elementSize = MAXALIGN(sizeof(HASHELEMENT)) + MAXALIGN(entrysize);
813 34160 : size = add_size(size,
814 : mul_size(nElementAllocs,
815 : mul_size(elementAllocCnt, elementSize)));
816 :
817 34160 : return size;
818 : }
819 :
820 : /*
821 : * Select an appropriate directory size for a hashtable with the given
822 : * maximum number of entries.
823 : * This is only needed for hashtables in shared memory, whose directories
824 : * cannot be expanded dynamically.
825 : * NB: assumes that all hash structure parameters have default values!
826 : *
827 : * XXX this had better agree with the behavior of init_htab()...
828 : */
829 : long
830 18302 : hash_select_dirsize(long num_entries)
831 : {
832 : long nBuckets,
833 : nSegments,
834 : nDirEntries;
835 :
836 : /* estimate number of buckets wanted */
837 18302 : nBuckets = next_pow2_long(num_entries);
838 : /* # of segments needed for nBuckets */
839 18302 : nSegments = next_pow2_long((nBuckets - 1) / DEF_SEGSIZE + 1);
840 : /* directory entries */
841 18302 : nDirEntries = DEF_DIRSIZE;
842 18302 : while (nDirEntries < nSegments)
843 0 : nDirEntries <<= 1; /* dir_alloc doubles dsize at each call */
844 :
845 18302 : return nDirEntries;
846 : }
847 :
848 : /*
849 : * Compute the required initial memory allocation for a shared-memory
850 : * hashtable with the given parameters. We need space for the HASHHDR
851 : * and for the (non expansible) directory.
852 : */
853 : Size
854 18302 : hash_get_shared_size(HASHCTL *info, int flags)
855 : {
856 : Assert(flags & HASH_DIRSIZE);
857 : Assert(info->dsize == info->max_dsize);
858 18302 : return sizeof(HASHHDR) + info->dsize * sizeof(HASHSEGMENT);
859 : }
860 :
861 :
862 : /********************** DESTROY ROUTINES ************************/
863 :
864 : void
865 120310 : hash_destroy(HTAB *hashp)
866 : {
867 120310 : if (hashp != NULL)
868 : {
869 : /* allocation method must be one we know how to free, too */
870 : Assert(hashp->alloc == DynaHashAlloc);
871 : /* so this hashtable must have its own context */
872 : Assert(hashp->hcxt != NULL);
873 :
874 120310 : hash_stats("destroy", hashp);
875 :
876 : /*
877 : * Free everything by destroying the hash table's memory context.
878 : */
879 120310 : MemoryContextDelete(hashp->hcxt);
880 : }
881 120310 : }
882 :
883 : void
884 120310 : hash_stats(const char *where, HTAB *hashp)
885 : {
886 : #ifdef HASH_STATISTICS
887 : fprintf(stderr, "%s: this HTAB -- accesses %ld collisions %ld\n",
888 : where, hashp->hctl->accesses, hashp->hctl->collisions);
889 :
890 : fprintf(stderr, "hash_stats: entries %ld keysize %ld maxp %u segmentcount %ld\n",
891 : hash_get_num_entries(hashp), (long) hashp->hctl->keysize,
892 : hashp->hctl->max_bucket, hashp->hctl->nsegs);
893 : fprintf(stderr, "%s: total accesses %ld total collisions %ld\n",
894 : where, hash_accesses, hash_collisions);
895 : fprintf(stderr, "hash_stats: total expansions %ld\n",
896 : hash_expansions);
897 : #endif
898 120310 : }
899 :
900 : /*******************************SEARCH ROUTINES *****************************/
901 :
902 :
903 : /*
904 : * get_hash_value -- exported routine to calculate a key's hash value
905 : *
906 : * We export this because for partitioned tables, callers need to compute
907 : * the partition number (from the low-order bits of the hash value) before
908 : * searching.
909 : */
910 : uint32
911 152259748 : get_hash_value(HTAB *hashp, const void *keyPtr)
912 : {
913 152259748 : return hashp->hash(keyPtr, hashp->keysize);
914 : }
915 :
916 : /* Convert a hash value to a bucket number */
917 : static inline uint32
918 351149428 : calc_bucket(HASHHDR *hctl, uint32 hash_val)
919 : {
920 : uint32 bucket;
921 :
922 351149428 : bucket = hash_val & hctl->high_mask;
923 351149428 : if (bucket > hctl->max_bucket)
924 164267530 : bucket = bucket & hctl->low_mask;
925 :
926 351149428 : return bucket;
927 : }
928 :
929 : /*
930 : * hash_search -- look up key in table and perform action
931 : * hash_search_with_hash_value -- same, with key's hash value already computed
932 : *
933 : * action is one of:
934 : * HASH_FIND: look up key in table
935 : * HASH_ENTER: look up key in table, creating entry if not present
936 : * HASH_ENTER_NULL: same, but return NULL if out of memory
937 : * HASH_REMOVE: look up key in table, remove entry if present
938 : *
939 : * Return value is a pointer to the element found/entered/removed if any,
940 : * or NULL if no match was found. (NB: in the case of the REMOVE action,
941 : * the result is a dangling pointer that shouldn't be dereferenced!)
942 : *
943 : * HASH_ENTER will normally ereport a generic "out of memory" error if
944 : * it is unable to create a new entry. The HASH_ENTER_NULL operation is
945 : * the same except it will return NULL if out of memory.
946 : *
947 : * If foundPtr isn't NULL, then *foundPtr is set true if we found an
948 : * existing entry in the table, false otherwise. This is needed in the
949 : * HASH_ENTER case, but is redundant with the return value otherwise.
950 : *
951 : * For hash_search_with_hash_value, the hashvalue parameter must have been
952 : * calculated with get_hash_value().
953 : */
954 : void *
955 211440594 : hash_search(HTAB *hashp,
956 : const void *keyPtr,
957 : HASHACTION action,
958 : bool *foundPtr)
959 : {
960 211440594 : return hash_search_with_hash_value(hashp,
961 : keyPtr,
962 211440594 : hashp->hash(keyPtr, hashp->keysize),
963 : action,
964 : foundPtr);
965 : }
966 :
967 : void *
968 348684720 : hash_search_with_hash_value(HTAB *hashp,
969 : const void *keyPtr,
970 : uint32 hashvalue,
971 : HASHACTION action,
972 : bool *foundPtr)
973 : {
974 348684720 : HASHHDR *hctl = hashp->hctl;
975 348684720 : int freelist_idx = FREELIST_IDX(hctl, hashvalue);
976 : Size keysize;
977 : HASHBUCKET currBucket;
978 : HASHBUCKET *prevBucketPtr;
979 : HashCompareFunc match;
980 :
981 : #ifdef HASH_STATISTICS
982 : hash_accesses++;
983 : hctl->accesses++;
984 : #endif
985 :
986 : /*
987 : * If inserting, check if it is time to split a bucket.
988 : *
989 : * NOTE: failure to expand table is not a fatal error, it just means we
990 : * have to run at higher fill factor than we wanted. However, if we're
991 : * using the palloc allocator then it will throw error anyway on
992 : * out-of-memory, so we must do this before modifying the table.
993 : */
994 348684720 : if (action == HASH_ENTER || action == HASH_ENTER_NULL)
995 : {
996 : /*
997 : * Can't split if running in partitioned mode, nor if frozen, nor if
998 : * table is the subject of any active hash_seq_search scans.
999 : */
1000 94389184 : if (hctl->freeList[0].nentries > (long) hctl->max_bucket &&
1001 723756 : !IS_PARTITIONED(hctl) && !hashp->frozen &&
1002 723756 : !has_seq_scans(hashp))
1003 723756 : (void) expand_table(hashp);
1004 : }
1005 :
1006 : /*
1007 : * Do the initial lookup
1008 : */
1009 348684720 : (void) hash_initial_lookup(hashp, hashvalue, &prevBucketPtr);
1010 348684720 : currBucket = *prevBucketPtr;
1011 :
1012 : /*
1013 : * Follow collision chain looking for matching key
1014 : */
1015 348684720 : match = hashp->match; /* save one fetch in inner loop */
1016 348684720 : keysize = hashp->keysize; /* ditto */
1017 :
1018 430883014 : while (currBucket != NULL)
1019 : {
1020 642327674 : if (currBucket->hashvalue == hashvalue &&
1021 280067368 : match(ELEMENTKEY(currBucket), keyPtr, keysize) == 0)
1022 280062012 : break;
1023 82198294 : prevBucketPtr = &(currBucket->link);
1024 82198294 : currBucket = *prevBucketPtr;
1025 : #ifdef HASH_STATISTICS
1026 : hash_collisions++;
1027 : hctl->collisions++;
1028 : #endif
1029 : }
1030 :
1031 348684720 : if (foundPtr)
1032 97061764 : *foundPtr = (bool) (currBucket != NULL);
1033 :
1034 : /*
1035 : * OK, now what?
1036 : */
1037 348684720 : switch (action)
1038 : {
1039 207031100 : case HASH_FIND:
1040 207031100 : if (currBucket != NULL)
1041 195511340 : return ELEMENTKEY(currBucket);
1042 11519760 : return NULL;
1043 :
1044 47264436 : case HASH_REMOVE:
1045 47264436 : if (currBucket != NULL)
1046 : {
1047 : /* if partitioned, must lock to touch nentries and freeList */
1048 47257672 : if (IS_PARTITIONED(hctl))
1049 10425176 : SpinLockAcquire(&(hctl->freeList[freelist_idx].mutex));
1050 :
1051 : /* delete the record from the appropriate nentries counter. */
1052 : Assert(hctl->freeList[freelist_idx].nentries > 0);
1053 47257672 : hctl->freeList[freelist_idx].nentries--;
1054 :
1055 : /* remove record from hash bucket's chain. */
1056 47257672 : *prevBucketPtr = currBucket->link;
1057 :
1058 : /* add the record to the appropriate freelist. */
1059 47257672 : currBucket->link = hctl->freeList[freelist_idx].freeList;
1060 47257672 : hctl->freeList[freelist_idx].freeList = currBucket;
1061 :
1062 47257672 : if (IS_PARTITIONED(hctl))
1063 10425176 : SpinLockRelease(&hctl->freeList[freelist_idx].mutex);
1064 :
1065 : /*
1066 : * better hope the caller is synchronizing access to this
1067 : * element, because someone else is going to reuse it the next
1068 : * time something is added to the table
1069 : */
1070 47257672 : return ELEMENTKEY(currBucket);
1071 : }
1072 6764 : return NULL;
1073 :
1074 94389184 : case HASH_ENTER:
1075 : case HASH_ENTER_NULL:
1076 : /* Return existing element if found, else create one */
1077 94389184 : if (currBucket != NULL)
1078 37293000 : return ELEMENTKEY(currBucket);
1079 :
1080 : /* disallow inserts if frozen */
1081 57096184 : if (hashp->frozen)
1082 0 : elog(ERROR, "cannot insert into frozen hashtable \"%s\"",
1083 : hashp->tabname);
1084 :
1085 57096184 : currBucket = get_hash_entry(hashp, freelist_idx);
1086 57096184 : if (currBucket == NULL)
1087 : {
1088 : /* out of memory */
1089 0 : if (action == HASH_ENTER_NULL)
1090 0 : return NULL;
1091 : /* report a generic message */
1092 0 : if (hashp->isshared)
1093 0 : ereport(ERROR,
1094 : (errcode(ERRCODE_OUT_OF_MEMORY),
1095 : errmsg("out of shared memory")));
1096 : else
1097 0 : ereport(ERROR,
1098 : (errcode(ERRCODE_OUT_OF_MEMORY),
1099 : errmsg("out of memory")));
1100 : }
1101 :
1102 : /* link into hashbucket chain */
1103 57096184 : *prevBucketPtr = currBucket;
1104 57096184 : currBucket->link = NULL;
1105 :
1106 : /* copy key into record */
1107 57096184 : currBucket->hashvalue = hashvalue;
1108 57096184 : hashp->keycopy(ELEMENTKEY(currBucket), keyPtr, keysize);
1109 :
1110 : /*
1111 : * Caller is expected to fill the data field on return. DO NOT
1112 : * insert any code that could possibly throw error here, as doing
1113 : * so would leave the table entry incomplete and hence corrupt the
1114 : * caller's data structure.
1115 : */
1116 :
1117 57096184 : return ELEMENTKEY(currBucket);
1118 : }
1119 :
1120 0 : elog(ERROR, "unrecognized hash action code: %d", (int) action);
1121 :
1122 : return NULL; /* keep compiler quiet */
1123 : }
1124 :
1125 : /*
1126 : * hash_update_hash_key -- change the hash key of an existing table entry
1127 : *
1128 : * This is equivalent to removing the entry, making a new entry, and copying
1129 : * over its data, except that the entry never goes to the table's freelist.
1130 : * Therefore this cannot suffer an out-of-memory failure, even if there are
1131 : * other processes operating in other partitions of the hashtable.
1132 : *
1133 : * Returns true if successful, false if the requested new hash key is already
1134 : * present. Throws error if the specified entry pointer isn't actually a
1135 : * table member.
1136 : *
1137 : * NB: currently, there is no special case for old and new hash keys being
1138 : * identical, which means we'll report false for that situation. This is
1139 : * preferable for existing uses.
1140 : *
1141 : * NB: for a partitioned hashtable, caller must hold lock on both relevant
1142 : * partitions, if the new hash key would belong to a different partition.
1143 : */
1144 : bool
1145 1330 : hash_update_hash_key(HTAB *hashp,
1146 : void *existingEntry,
1147 : const void *newKeyPtr)
1148 : {
1149 1330 : HASHELEMENT *existingElement = ELEMENT_FROM_KEY(existingEntry);
1150 : uint32 newhashvalue;
1151 : Size keysize;
1152 : uint32 bucket;
1153 : uint32 newbucket;
1154 : HASHBUCKET currBucket;
1155 : HASHBUCKET *prevBucketPtr;
1156 : HASHBUCKET *oldPrevPtr;
1157 : HashCompareFunc match;
1158 :
1159 : #ifdef HASH_STATISTICS
1160 : hash_accesses++;
1161 : hctl->accesses++;
1162 : #endif
1163 :
1164 : /* disallow updates if frozen */
1165 1330 : if (hashp->frozen)
1166 0 : elog(ERROR, "cannot update in frozen hashtable \"%s\"",
1167 : hashp->tabname);
1168 :
1169 : /*
1170 : * Lookup the existing element using its saved hash value. We need to do
1171 : * this to be able to unlink it from its hash chain, but as a side benefit
1172 : * we can verify the validity of the passed existingEntry pointer.
1173 : */
1174 1330 : bucket = hash_initial_lookup(hashp, existingElement->hashvalue,
1175 : &prevBucketPtr);
1176 1330 : currBucket = *prevBucketPtr;
1177 :
1178 1334 : while (currBucket != NULL)
1179 : {
1180 1334 : if (currBucket == existingElement)
1181 1330 : break;
1182 4 : prevBucketPtr = &(currBucket->link);
1183 4 : currBucket = *prevBucketPtr;
1184 : }
1185 :
1186 1330 : if (currBucket == NULL)
1187 0 : elog(ERROR, "hash_update_hash_key argument is not in hashtable \"%s\"",
1188 : hashp->tabname);
1189 :
1190 1330 : oldPrevPtr = prevBucketPtr;
1191 :
1192 : /*
1193 : * Now perform the equivalent of a HASH_ENTER operation to locate the hash
1194 : * chain we want to put the entry into.
1195 : */
1196 1330 : newhashvalue = hashp->hash(newKeyPtr, hashp->keysize);
1197 1330 : newbucket = hash_initial_lookup(hashp, newhashvalue, &prevBucketPtr);
1198 1330 : currBucket = *prevBucketPtr;
1199 :
1200 : /*
1201 : * Follow collision chain looking for matching key
1202 : */
1203 1330 : match = hashp->match; /* save one fetch in inner loop */
1204 1330 : keysize = hashp->keysize; /* ditto */
1205 :
1206 1472 : while (currBucket != NULL)
1207 : {
1208 142 : if (currBucket->hashvalue == newhashvalue &&
1209 0 : match(ELEMENTKEY(currBucket), newKeyPtr, keysize) == 0)
1210 0 : break;
1211 142 : prevBucketPtr = &(currBucket->link);
1212 142 : currBucket = *prevBucketPtr;
1213 : #ifdef HASH_STATISTICS
1214 : hash_collisions++;
1215 : hctl->collisions++;
1216 : #endif
1217 : }
1218 :
1219 1330 : if (currBucket != NULL)
1220 0 : return false; /* collision with an existing entry */
1221 :
1222 1330 : currBucket = existingElement;
1223 :
1224 : /*
1225 : * If old and new hash values belong to the same bucket, we need not
1226 : * change any chain links, and indeed should not since this simplistic
1227 : * update will corrupt the list if currBucket is the last element. (We
1228 : * cannot fall out earlier, however, since we need to scan the bucket to
1229 : * check for duplicate keys.)
1230 : */
1231 1330 : if (bucket != newbucket)
1232 : {
1233 : /* OK to remove record from old hash bucket's chain. */
1234 1190 : *oldPrevPtr = currBucket->link;
1235 :
1236 : /* link into new hashbucket chain */
1237 1190 : *prevBucketPtr = currBucket;
1238 1190 : currBucket->link = NULL;
1239 : }
1240 :
1241 : /* copy new key into record */
1242 1330 : currBucket->hashvalue = newhashvalue;
1243 1330 : hashp->keycopy(ELEMENTKEY(currBucket), newKeyPtr, keysize);
1244 :
1245 : /* rest of record is untouched */
1246 :
1247 1330 : return true;
1248 : }
1249 :
1250 : /*
1251 : * Allocate a new hashtable entry if possible; return NULL if out of memory.
1252 : * (Or, if the underlying space allocator throws error for out-of-memory,
1253 : * we won't return at all.)
1254 : */
1255 : static HASHBUCKET
1256 57096184 : get_hash_entry(HTAB *hashp, int freelist_idx)
1257 : {
1258 57096184 : HASHHDR *hctl = hashp->hctl;
1259 : HASHBUCKET newElement;
1260 :
1261 : for (;;)
1262 : {
1263 : /* if partitioned, must lock to touch nentries and freeList */
1264 57557040 : if (IS_PARTITIONED(hctl))
1265 11566496 : SpinLockAcquire(&hctl->freeList[freelist_idx].mutex);
1266 :
1267 : /* try to get an entry from the freelist */
1268 57557040 : newElement = hctl->freeList[freelist_idx].freeList;
1269 :
1270 57557040 : if (newElement != NULL)
1271 57096184 : break;
1272 :
1273 460856 : if (IS_PARTITIONED(hctl))
1274 2732 : SpinLockRelease(&hctl->freeList[freelist_idx].mutex);
1275 :
1276 : /*
1277 : * No free elements in this freelist. In a partitioned table, there
1278 : * might be entries in other freelists, but to reduce contention we
1279 : * prefer to first try to get another chunk of buckets from the main
1280 : * shmem allocator. If that fails, though, we *MUST* root through all
1281 : * the other freelists before giving up. There are multiple callers
1282 : * that assume that they can allocate every element in the initially
1283 : * requested table size, or that deleting an element guarantees they
1284 : * can insert a new element, even if shared memory is entirely full.
1285 : * Failing because the needed element is in a different freelist is
1286 : * not acceptable.
1287 : */
1288 460856 : if (!element_alloc(hashp, hctl->nelem_alloc, freelist_idx))
1289 : {
1290 : int borrow_from_idx;
1291 :
1292 0 : if (!IS_PARTITIONED(hctl))
1293 0 : return NULL; /* out of memory */
1294 :
1295 : /* try to borrow element from another freelist */
1296 0 : borrow_from_idx = freelist_idx;
1297 : for (;;)
1298 : {
1299 0 : borrow_from_idx = (borrow_from_idx + 1) % NUM_FREELISTS;
1300 0 : if (borrow_from_idx == freelist_idx)
1301 0 : break; /* examined all freelists, fail */
1302 :
1303 0 : SpinLockAcquire(&(hctl->freeList[borrow_from_idx].mutex));
1304 0 : newElement = hctl->freeList[borrow_from_idx].freeList;
1305 :
1306 0 : if (newElement != NULL)
1307 : {
1308 0 : hctl->freeList[borrow_from_idx].freeList = newElement->link;
1309 0 : SpinLockRelease(&(hctl->freeList[borrow_from_idx].mutex));
1310 :
1311 : /* careful: count the new element in its proper freelist */
1312 0 : SpinLockAcquire(&hctl->freeList[freelist_idx].mutex);
1313 0 : hctl->freeList[freelist_idx].nentries++;
1314 0 : SpinLockRelease(&hctl->freeList[freelist_idx].mutex);
1315 :
1316 0 : return newElement;
1317 : }
1318 :
1319 0 : SpinLockRelease(&(hctl->freeList[borrow_from_idx].mutex));
1320 : }
1321 :
1322 : /* no elements available to borrow either, so out of memory */
1323 0 : return NULL;
1324 : }
1325 : }
1326 :
1327 : /* remove entry from freelist, bump nentries */
1328 57096184 : hctl->freeList[freelist_idx].freeList = newElement->link;
1329 57096184 : hctl->freeList[freelist_idx].nentries++;
1330 :
1331 57096184 : if (IS_PARTITIONED(hctl))
1332 11563764 : SpinLockRelease(&hctl->freeList[freelist_idx].mutex);
1333 :
1334 57096184 : return newElement;
1335 : }
1336 :
1337 : /*
1338 : * hash_get_num_entries -- get the number of entries in a hashtable
1339 : */
1340 : long
1341 117678 : hash_get_num_entries(HTAB *hashp)
1342 : {
1343 : int i;
1344 117678 : long sum = hashp->hctl->freeList[0].nentries;
1345 :
1346 : /*
1347 : * We currently don't bother with acquiring the mutexes; it's only
1348 : * sensible to call this function if you've got lock on all partitions of
1349 : * the table.
1350 : */
1351 117678 : if (IS_PARTITIONED(hashp->hctl))
1352 : {
1353 95936 : for (i = 1; i < NUM_FREELISTS; i++)
1354 92938 : sum += hashp->hctl->freeList[i].nentries;
1355 : }
1356 :
1357 117678 : return sum;
1358 : }
1359 :
1360 : /*
1361 : * hash_seq_init/_search/_term
1362 : * Sequentially search through hash table and return
1363 : * all the elements one by one, return NULL when no more.
1364 : *
1365 : * hash_seq_term should be called if and only if the scan is abandoned before
1366 : * completion; if hash_seq_search returns NULL then it has already done the
1367 : * end-of-scan cleanup.
1368 : *
1369 : * NOTE: caller may delete the returned element before continuing the scan.
1370 : * However, deleting any other element while the scan is in progress is
1371 : * UNDEFINED (it might be the one that curIndex is pointing at!). Also,
1372 : * if elements are added to the table while the scan is in progress, it is
1373 : * unspecified whether they will be visited by the scan or not.
1374 : *
1375 : * NOTE: it is possible to use hash_seq_init/hash_seq_search without any
1376 : * worry about hash_seq_term cleanup, if the hashtable is first locked against
1377 : * further insertions by calling hash_freeze.
1378 : *
1379 : * NOTE: to use this with a partitioned hashtable, caller had better hold
1380 : * at least shared lock on all partitions of the table throughout the scan!
1381 : * We can cope with insertions or deletions by our own backend, but *not*
1382 : * with concurrent insertions or deletions by another.
1383 : */
1384 : void
1385 4309614 : hash_seq_init(HASH_SEQ_STATUS *status, HTAB *hashp)
1386 : {
1387 4309614 : status->hashp = hashp;
1388 4309614 : status->curBucket = 0;
1389 4309614 : status->curEntry = NULL;
1390 4309614 : status->hasHashvalue = false;
1391 4309614 : if (!hashp->frozen)
1392 4309614 : register_seq_scan(hashp);
1393 4309614 : }
1394 :
1395 : /*
1396 : * Same as above but scan by the given hash value.
1397 : * See also hash_seq_search().
1398 : *
1399 : * NOTE: the default hash function doesn't match syscache hash function.
1400 : * Thus, if you're going to use this function in syscache callback, make sure
1401 : * you're using custom hash function. See relatt_cache_syshash()
1402 : * for example.
1403 : */
1404 : void
1405 1442964 : hash_seq_init_with_hash_value(HASH_SEQ_STATUS *status, HTAB *hashp,
1406 : uint32 hashvalue)
1407 : {
1408 : HASHBUCKET *bucketPtr;
1409 :
1410 1442964 : hash_seq_init(status, hashp);
1411 :
1412 1442964 : status->hasHashvalue = true;
1413 1442964 : status->hashvalue = hashvalue;
1414 :
1415 1442964 : status->curBucket = hash_initial_lookup(hashp, hashvalue, &bucketPtr);
1416 1442964 : status->curEntry = *bucketPtr;
1417 1442964 : }
1418 :
1419 : void *
1420 53153960 : hash_seq_search(HASH_SEQ_STATUS *status)
1421 : {
1422 : HTAB *hashp;
1423 : HASHHDR *hctl;
1424 : uint32 max_bucket;
1425 : long ssize;
1426 : long segment_num;
1427 : long segment_ndx;
1428 : HASHSEGMENT segp;
1429 : uint32 curBucket;
1430 : HASHELEMENT *curElem;
1431 :
1432 53153960 : if (status->hasHashvalue)
1433 : {
1434 : /*
1435 : * Scan entries only in the current bucket because only this bucket
1436 : * can contain entries with the given hash value.
1437 : */
1438 1640374 : while ((curElem = status->curEntry) != NULL)
1439 : {
1440 197410 : status->curEntry = curElem->link;
1441 197410 : if (status->hashvalue != curElem->hashvalue)
1442 188462 : continue;
1443 8948 : return (void *) ELEMENTKEY(curElem);
1444 : }
1445 :
1446 1442964 : hash_seq_term(status);
1447 1442964 : return NULL;
1448 : }
1449 :
1450 51702048 : if ((curElem = status->curEntry) != NULL)
1451 : {
1452 : /* Continuing scan of curBucket... */
1453 15234692 : status->curEntry = curElem->link;
1454 15234692 : if (status->curEntry == NULL) /* end of this bucket */
1455 10658464 : ++status->curBucket;
1456 15234692 : return ELEMENTKEY(curElem);
1457 : }
1458 :
1459 : /*
1460 : * Search for next nonempty bucket starting at curBucket.
1461 : */
1462 36467356 : curBucket = status->curBucket;
1463 36467356 : hashp = status->hashp;
1464 36467356 : hctl = hashp->hctl;
1465 36467356 : ssize = hashp->ssize;
1466 36467356 : max_bucket = hctl->max_bucket;
1467 :
1468 36467356 : if (curBucket > max_bucket)
1469 : {
1470 101562 : hash_seq_term(status);
1471 101562 : return NULL; /* search is done */
1472 : }
1473 :
1474 : /*
1475 : * first find the right segment in the table directory.
1476 : */
1477 36365794 : segment_num = curBucket >> hashp->sshift;
1478 36365794 : segment_ndx = MOD(curBucket, ssize);
1479 :
1480 36365794 : segp = hashp->dir[segment_num];
1481 :
1482 : /*
1483 : * Pick up the first item in this bucket's chain. If chain is not empty
1484 : * we can begin searching it. Otherwise we have to advance to find the
1485 : * next nonempty bucket. We try to optimize that case since searching a
1486 : * near-empty hashtable has to iterate this loop a lot.
1487 : */
1488 177897502 : while ((curElem = segp[segment_ndx]) == NULL)
1489 : {
1490 : /* empty bucket, advance to next */
1491 144261740 : if (++curBucket > max_bucket)
1492 : {
1493 2730032 : status->curBucket = curBucket;
1494 2730032 : hash_seq_term(status);
1495 2730032 : return NULL; /* search is done */
1496 : }
1497 141531708 : if (++segment_ndx >= ssize)
1498 : {
1499 269324 : segment_num++;
1500 269324 : segment_ndx = 0;
1501 269324 : segp = hashp->dir[segment_num];
1502 : }
1503 : }
1504 :
1505 : /* Begin scan of curBucket... */
1506 33635762 : status->curEntry = curElem->link;
1507 33635762 : if (status->curEntry == NULL) /* end of this bucket */
1508 22976908 : ++curBucket;
1509 33635762 : status->curBucket = curBucket;
1510 33635762 : return ELEMENTKEY(curElem);
1511 : }
1512 :
1513 : void
1514 4309594 : hash_seq_term(HASH_SEQ_STATUS *status)
1515 : {
1516 4309594 : if (!status->hashp->frozen)
1517 4309594 : deregister_seq_scan(status->hashp);
1518 4309594 : }
1519 :
1520 : /*
1521 : * hash_freeze
1522 : * Freeze a hashtable against future insertions (deletions are
1523 : * still allowed)
1524 : *
1525 : * The reason for doing this is that by preventing any more bucket splits,
1526 : * we no longer need to worry about registering hash_seq_search scans,
1527 : * and thus caller need not be careful about ensuring hash_seq_term gets
1528 : * called at the right times.
1529 : *
1530 : * Multiple calls to hash_freeze() are allowed, but you can't freeze a table
1531 : * with active scans (since hash_seq_term would then do the wrong thing).
1532 : */
1533 : void
1534 0 : hash_freeze(HTAB *hashp)
1535 : {
1536 0 : if (hashp->isshared)
1537 0 : elog(ERROR, "cannot freeze shared hashtable \"%s\"", hashp->tabname);
1538 0 : if (!hashp->frozen && has_seq_scans(hashp))
1539 0 : elog(ERROR, "cannot freeze hashtable \"%s\" because it has active scans",
1540 : hashp->tabname);
1541 0 : hashp->frozen = true;
1542 0 : }
1543 :
1544 :
1545 : /********************************* UTILITIES ************************/
1546 :
1547 : /*
1548 : * Expand the table by adding one more hash bucket.
1549 : */
1550 : static bool
1551 723756 : expand_table(HTAB *hashp)
1552 : {
1553 723756 : HASHHDR *hctl = hashp->hctl;
1554 : HASHSEGMENT old_seg,
1555 : new_seg;
1556 : long old_bucket,
1557 : new_bucket;
1558 : long new_segnum,
1559 : new_segndx;
1560 : long old_segnum,
1561 : old_segndx;
1562 : HASHBUCKET *oldlink,
1563 : *newlink;
1564 : HASHBUCKET currElement,
1565 : nextElement;
1566 :
1567 : Assert(!IS_PARTITIONED(hctl));
1568 :
1569 : #ifdef HASH_STATISTICS
1570 : hash_expansions++;
1571 : #endif
1572 :
1573 723756 : new_bucket = hctl->max_bucket + 1;
1574 723756 : new_segnum = new_bucket >> hashp->sshift;
1575 723756 : new_segndx = MOD(new_bucket, hashp->ssize);
1576 :
1577 723756 : if (new_segnum >= hctl->nsegs)
1578 : {
1579 : /* Allocate new segment if necessary -- could fail if dir full */
1580 2380 : if (new_segnum >= hctl->dsize)
1581 0 : if (!dir_realloc(hashp))
1582 0 : return false;
1583 2380 : if (!(hashp->dir[new_segnum] = seg_alloc(hashp)))
1584 0 : return false;
1585 2380 : hctl->nsegs++;
1586 : }
1587 :
1588 : /* OK, we created a new bucket */
1589 723756 : hctl->max_bucket++;
1590 :
1591 : /*
1592 : * *Before* changing masks, find old bucket corresponding to same hash
1593 : * values; values in that bucket may need to be relocated to new bucket.
1594 : * Note that new_bucket is certainly larger than low_mask at this point,
1595 : * so we can skip the first step of the regular hash mask calc.
1596 : */
1597 723756 : old_bucket = (new_bucket & hctl->low_mask);
1598 :
1599 : /*
1600 : * If we crossed a power of 2, readjust masks.
1601 : */
1602 723756 : if ((uint32) new_bucket > hctl->high_mask)
1603 : {
1604 4362 : hctl->low_mask = hctl->high_mask;
1605 4362 : hctl->high_mask = (uint32) new_bucket | hctl->low_mask;
1606 : }
1607 :
1608 : /*
1609 : * Relocate records to the new bucket. NOTE: because of the way the hash
1610 : * masking is done in calc_bucket, only one old bucket can need to be
1611 : * split at this point. With a different way of reducing the hash value,
1612 : * that might not be true!
1613 : */
1614 723756 : old_segnum = old_bucket >> hashp->sshift;
1615 723756 : old_segndx = MOD(old_bucket, hashp->ssize);
1616 :
1617 723756 : old_seg = hashp->dir[old_segnum];
1618 723756 : new_seg = hashp->dir[new_segnum];
1619 :
1620 723756 : oldlink = &old_seg[old_segndx];
1621 723756 : newlink = &new_seg[new_segndx];
1622 :
1623 1742840 : for (currElement = *oldlink;
1624 : currElement != NULL;
1625 1019084 : currElement = nextElement)
1626 : {
1627 1019084 : nextElement = currElement->link;
1628 1019084 : if ((long) calc_bucket(hctl, currElement->hashvalue) == old_bucket)
1629 : {
1630 502278 : *oldlink = currElement;
1631 502278 : oldlink = &currElement->link;
1632 : }
1633 : else
1634 : {
1635 516806 : *newlink = currElement;
1636 516806 : newlink = &currElement->link;
1637 : }
1638 : }
1639 : /* don't forget to terminate the rebuilt hash chains... */
1640 723756 : *oldlink = NULL;
1641 723756 : *newlink = NULL;
1642 :
1643 723756 : return true;
1644 : }
1645 :
1646 :
1647 : static bool
1648 0 : dir_realloc(HTAB *hashp)
1649 : {
1650 : HASHSEGMENT *p;
1651 : HASHSEGMENT *old_p;
1652 : long new_dsize;
1653 : long old_dirsize;
1654 : long new_dirsize;
1655 :
1656 0 : if (hashp->hctl->max_dsize != NO_MAX_DSIZE)
1657 0 : return false;
1658 :
1659 : /* Reallocate directory */
1660 0 : new_dsize = hashp->hctl->dsize << 1;
1661 0 : old_dirsize = hashp->hctl->dsize * sizeof(HASHSEGMENT);
1662 0 : new_dirsize = new_dsize * sizeof(HASHSEGMENT);
1663 :
1664 0 : old_p = hashp->dir;
1665 0 : CurrentDynaHashCxt = hashp->hcxt;
1666 0 : p = (HASHSEGMENT *) hashp->alloc((Size) new_dirsize);
1667 :
1668 0 : if (p != NULL)
1669 : {
1670 0 : memcpy(p, old_p, old_dirsize);
1671 0 : MemSet(((char *) p) + old_dirsize, 0, new_dirsize - old_dirsize);
1672 0 : hashp->dir = p;
1673 0 : hashp->hctl->dsize = new_dsize;
1674 :
1675 : /* XXX assume the allocator is palloc, so we know how to free */
1676 : Assert(hashp->alloc == DynaHashAlloc);
1677 0 : pfree(old_p);
1678 :
1679 0 : return true;
1680 : }
1681 :
1682 0 : return false;
1683 : }
1684 :
1685 :
1686 : static HASHSEGMENT
1687 1110502 : seg_alloc(HTAB *hashp)
1688 : {
1689 : HASHSEGMENT segp;
1690 :
1691 1110502 : CurrentDynaHashCxt = hashp->hcxt;
1692 1110502 : segp = (HASHSEGMENT) hashp->alloc(sizeof(HASHBUCKET) * hashp->ssize);
1693 :
1694 1110502 : if (!segp)
1695 0 : return NULL;
1696 :
1697 1110502 : MemSet(segp, 0, sizeof(HASHBUCKET) * hashp->ssize);
1698 :
1699 1110502 : return segp;
1700 : }
1701 :
1702 : /*
1703 : * allocate some new elements and link them into the indicated free list
1704 : */
1705 : static bool
1706 1023438 : element_alloc(HTAB *hashp, int nelem, int freelist_idx)
1707 : {
1708 1023438 : HASHHDR *hctl = hashp->hctl;
1709 : Size elementSize;
1710 : HASHELEMENT *firstElement;
1711 : HASHELEMENT *tmpElement;
1712 : HASHELEMENT *prevElement;
1713 : int i;
1714 :
1715 1023438 : if (hashp->isfixed)
1716 0 : return false;
1717 :
1718 : /* Each element has a HASHELEMENT header plus user data. */
1719 1023438 : elementSize = MAXALIGN(sizeof(HASHELEMENT)) + MAXALIGN(hctl->entrysize);
1720 :
1721 1023438 : CurrentDynaHashCxt = hashp->hcxt;
1722 1023438 : firstElement = (HASHELEMENT *) hashp->alloc(nelem * elementSize);
1723 :
1724 1023438 : if (!firstElement)
1725 0 : return false;
1726 :
1727 : /* prepare to link all the new entries into the freelist */
1728 1023438 : prevElement = NULL;
1729 1023438 : tmpElement = firstElement;
1730 98503722 : for (i = 0; i < nelem; i++)
1731 : {
1732 97480284 : tmpElement->link = prevElement;
1733 97480284 : prevElement = tmpElement;
1734 97480284 : tmpElement = (HASHELEMENT *) (((char *) tmpElement) + elementSize);
1735 : }
1736 :
1737 : /* if partitioned, must lock to touch freeList */
1738 1023438 : if (IS_PARTITIONED(hctl))
1739 327852 : SpinLockAcquire(&hctl->freeList[freelist_idx].mutex);
1740 :
1741 : /* freelist could be nonempty if two backends did this concurrently */
1742 1023438 : firstElement->link = hctl->freeList[freelist_idx].freeList;
1743 1023438 : hctl->freeList[freelist_idx].freeList = prevElement;
1744 :
1745 1023438 : if (IS_PARTITIONED(hctl))
1746 327852 : SpinLockRelease(&hctl->freeList[freelist_idx].mutex);
1747 :
1748 1023438 : return true;
1749 : }
1750 :
1751 : /*
1752 : * Do initial lookup of a bucket for the given hash value, retrieving its
1753 : * bucket number and its hash bucket.
1754 : */
1755 : static inline uint32
1756 350130344 : hash_initial_lookup(HTAB *hashp, uint32 hashvalue, HASHBUCKET **bucketptr)
1757 : {
1758 350130344 : HASHHDR *hctl = hashp->hctl;
1759 : HASHSEGMENT segp;
1760 : long segment_num;
1761 : long segment_ndx;
1762 : uint32 bucket;
1763 :
1764 350130344 : bucket = calc_bucket(hctl, hashvalue);
1765 :
1766 350130344 : segment_num = bucket >> hashp->sshift;
1767 350130344 : segment_ndx = MOD(bucket, hashp->ssize);
1768 :
1769 350130344 : segp = hashp->dir[segment_num];
1770 :
1771 350130344 : if (segp == NULL)
1772 0 : hash_corrupted(hashp);
1773 :
1774 350130344 : *bucketptr = &segp[segment_ndx];
1775 350130344 : return bucket;
1776 : }
1777 :
1778 : /* complain when we have detected a corrupted hashtable */
1779 : static void
1780 0 : hash_corrupted(HTAB *hashp)
1781 : {
1782 : /*
1783 : * If the corruption is in a shared hashtable, we'd better force a
1784 : * systemwide restart. Otherwise, just shut down this one backend.
1785 : */
1786 0 : if (hashp->isshared)
1787 0 : elog(PANIC, "hash table \"%s\" corrupted", hashp->tabname);
1788 : else
1789 0 : elog(FATAL, "hash table \"%s\" corrupted", hashp->tabname);
1790 : }
1791 :
1792 : /* calculate ceil(log base 2) of num */
1793 : int
1794 1220688 : my_log2(long num)
1795 : {
1796 : /*
1797 : * guard against too-large input, which would be invalid for
1798 : * pg_ceil_log2_*()
1799 : */
1800 1220688 : if (num > LONG_MAX / 2)
1801 0 : num = LONG_MAX / 2;
1802 :
1803 : #if SIZEOF_LONG < 8
1804 : return pg_ceil_log2_32(num);
1805 : #else
1806 1220688 : return pg_ceil_log2_64(num);
1807 : #endif
1808 : }
1809 :
1810 : /* calculate first power of 2 >= num, bounded to what will fit in a long */
1811 : static long
1812 104924 : next_pow2_long(long num)
1813 : {
1814 : /* my_log2's internal range check is sufficient */
1815 104924 : return 1L << my_log2(num);
1816 : }
1817 :
1818 : /* calculate first power of 2 >= num, bounded to what will fit in an int */
1819 : static int
1820 1077648 : next_pow2_int(long num)
1821 : {
1822 1077648 : if (num > INT_MAX / 2)
1823 0 : num = INT_MAX / 2;
1824 1077648 : return 1 << my_log2(num);
1825 : }
1826 :
1827 :
1828 : /************************* SEQ SCAN TRACKING ************************/
1829 :
1830 : /*
1831 : * We track active hash_seq_search scans here. The need for this mechanism
1832 : * comes from the fact that a scan will get confused if a bucket split occurs
1833 : * while it's in progress: it might visit entries twice, or even miss some
1834 : * entirely (if it's partway through the same bucket that splits). Hence
1835 : * we want to inhibit bucket splits if there are any active scans on the
1836 : * table being inserted into. This is a fairly rare case in current usage,
1837 : * so just postponing the split until the next insertion seems sufficient.
1838 : *
1839 : * Given present usages of the function, only a few scans are likely to be
1840 : * open concurrently; so a finite-size stack of open scans seems sufficient,
1841 : * and we don't worry that linear search is too slow. Note that we do
1842 : * allow multiple scans of the same hashtable to be open concurrently.
1843 : *
1844 : * This mechanism can support concurrent scan and insertion in a shared
1845 : * hashtable if it's the same backend doing both. It would fail otherwise,
1846 : * but locking reasons seem to preclude any such scenario anyway, so we don't
1847 : * worry.
1848 : *
1849 : * This arrangement is reasonably robust if a transient hashtable is deleted
1850 : * without notifying us. The absolute worst case is we might inhibit splits
1851 : * in another table created later at exactly the same address. We will give
1852 : * a warning at transaction end for reference leaks, so any bugs leading to
1853 : * lack of notification should be easy to catch.
1854 : */
1855 :
1856 : #define MAX_SEQ_SCANS 100
1857 :
1858 : static HTAB *seq_scan_tables[MAX_SEQ_SCANS]; /* tables being scanned */
1859 : static int seq_scan_level[MAX_SEQ_SCANS]; /* subtransaction nest level */
1860 : static int num_seq_scans = 0;
1861 :
1862 :
1863 : /* Register a table as having an active hash_seq_search scan */
1864 : static void
1865 4309614 : register_seq_scan(HTAB *hashp)
1866 : {
1867 4309614 : if (num_seq_scans >= MAX_SEQ_SCANS)
1868 0 : elog(ERROR, "too many active hash_seq_search scans, cannot start one on \"%s\"",
1869 : hashp->tabname);
1870 4309614 : seq_scan_tables[num_seq_scans] = hashp;
1871 4309614 : seq_scan_level[num_seq_scans] = GetCurrentTransactionNestLevel();
1872 4309614 : num_seq_scans++;
1873 4309614 : }
1874 :
1875 : /* Deregister an active scan */
1876 : static void
1877 4309594 : deregister_seq_scan(HTAB *hashp)
1878 : {
1879 : int i;
1880 :
1881 : /* Search backward since it's most likely at the stack top */
1882 4309594 : for (i = num_seq_scans - 1; i >= 0; i--)
1883 : {
1884 4309594 : if (seq_scan_tables[i] == hashp)
1885 : {
1886 4309594 : seq_scan_tables[i] = seq_scan_tables[num_seq_scans - 1];
1887 4309594 : seq_scan_level[i] = seq_scan_level[num_seq_scans - 1];
1888 4309594 : num_seq_scans--;
1889 4309594 : return;
1890 : }
1891 : }
1892 0 : elog(ERROR, "no hash_seq_search scan for hash table \"%s\"",
1893 : hashp->tabname);
1894 : }
1895 :
1896 : /* Check if a table has any active scan */
1897 : static bool
1898 723756 : has_seq_scans(HTAB *hashp)
1899 : {
1900 : int i;
1901 :
1902 723758 : for (i = 0; i < num_seq_scans; i++)
1903 : {
1904 2 : if (seq_scan_tables[i] == hashp)
1905 0 : return true;
1906 : }
1907 723756 : return false;
1908 : }
1909 :
1910 : /* Clean up any open scans at end of transaction */
1911 : void
1912 818606 : AtEOXact_HashTables(bool isCommit)
1913 : {
1914 : /*
1915 : * During abort cleanup, open scans are expected; just silently clean 'em
1916 : * out. An open scan at commit means someone forgot a hash_seq_term()
1917 : * call, so complain.
1918 : *
1919 : * Note: it's tempting to try to print the tabname here, but refrain for
1920 : * fear of touching deallocated memory. This isn't a user-facing message
1921 : * anyway, so it needn't be pretty.
1922 : */
1923 818606 : if (isCommit)
1924 : {
1925 : int i;
1926 :
1927 770226 : for (i = 0; i < num_seq_scans; i++)
1928 : {
1929 0 : elog(WARNING, "leaked hash_seq_search scan for hash table %p",
1930 : seq_scan_tables[i]);
1931 : }
1932 : }
1933 818606 : num_seq_scans = 0;
1934 818606 : }
1935 :
1936 : /* Clean up any open scans at end of subtransaction */
1937 : void
1938 19966 : AtEOSubXact_HashTables(bool isCommit, int nestDepth)
1939 : {
1940 : int i;
1941 :
1942 : /*
1943 : * Search backward to make cleanup easy. Note we must check all entries,
1944 : * not only those at the end of the array, because deletion technique
1945 : * doesn't keep them in order.
1946 : */
1947 19966 : for (i = num_seq_scans - 1; i >= 0; i--)
1948 : {
1949 0 : if (seq_scan_level[i] >= nestDepth)
1950 : {
1951 0 : if (isCommit)
1952 0 : elog(WARNING, "leaked hash_seq_search scan for hash table %p",
1953 : seq_scan_tables[i]);
1954 0 : seq_scan_tables[i] = seq_scan_tables[num_seq_scans - 1];
1955 0 : seq_scan_level[i] = seq_scan_level[num_seq_scans - 1];
1956 0 : num_seq_scans--;
1957 : }
1958 : }
1959 19966 : }
|
