nbtree: fix read page recheck typo.
[pgsql.git] / src / backend / storage / ipc / shmem.c
blob6d5f08398641cf0d35256590cb13449c3d1c08dc
1 /*-------------------------------------------------------------------------
3 * shmem.c
4 * create shared memory and initialize shared memory data structures.
6 * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
10 * IDENTIFICATION
11 * src/backend/storage/ipc/shmem.c
13 *-------------------------------------------------------------------------
16 * POSTGRES processes share one or more regions of shared memory.
17 * The shared memory is created by a postmaster and is inherited
18 * by each backend via fork() (or, in some ports, via other OS-specific
19 * methods). The routines in this file are used for allocating and
20 * binding to shared memory data structures.
22 * NOTES:
23 * (a) There are three kinds of shared memory data structures
24 * available to POSTGRES: fixed-size structures, queues and hash
25 * tables. Fixed-size structures contain things like global variables
26 * for a module and should never be allocated after the shared memory
27 * initialization phase. Hash tables have a fixed maximum size, but
28 * their actual size can vary dynamically. When entries are added
29 * to the table, more space is allocated. Queues link data structures
30 * that have been allocated either within fixed-size structures or as hash
31 * buckets. Each shared data structure has a string name to identify
32 * it (assigned in the module that declares it).
34 * (b) During initialization, each module looks for its
35 * shared data structures in a hash table called the "Shmem Index".
36 * If the data structure is not present, the caller can allocate
37 * a new one and initialize it. If the data structure is present,
38 * the caller "attaches" to the structure by initializing a pointer
39 * in the local address space.
40 * The shmem index has two purposes: first, it gives us
41 * a simple model of how the world looks when a backend process
42 * initializes. If something is present in the shmem index,
43 * it is initialized. If it is not, it is uninitialized. Second,
44 * the shmem index allows us to allocate shared memory on demand
45 * instead of trying to preallocate structures and hard-wire the
46 * sizes and locations in header files. If you are using a lot
47 * of shared memory in a lot of different places (and changing
48 * things during development), this is important.
50 * (c) In standard Unix-ish environments, individual backends do not
51 * need to re-establish their local pointers into shared memory, because
52 * they inherit correct values of those variables via fork() from the
53 * postmaster. However, this does not work in the EXEC_BACKEND case.
54 * In ports using EXEC_BACKEND, new backends have to set up their local
55 * pointers using the method described in (b) above.
57 * (d) memory allocation model: shared memory can never be
58 * freed, once allocated. Each hash table has its own free list,
59 * so hash buckets can be reused when an item is deleted. However,
60 * if one hash table grows very large and then shrinks, its space
61 * cannot be redistributed to other tables. We could build a simple
62 * hash bucket garbage collector if need be. Right now, it seems
63 * unnecessary.
66 #include "postgres.h"
68 #include "fmgr.h"
69 #include "funcapi.h"
70 #include "miscadmin.h"
71 #include "storage/lwlock.h"
72 #include "storage/pg_shmem.h"
73 #include "storage/shmem.h"
74 #include "storage/spin.h"
75 #include "utils/builtins.h"
77 static void *ShmemAllocRaw(Size size, Size *allocated_size);
79 /* shared memory global variables */
81 static PGShmemHeader *ShmemSegHdr; /* shared mem segment header */
83 static void *ShmemBase; /* start address of shared memory */
85 static void *ShmemEnd; /* end+1 address of shared memory */
87 slock_t *ShmemLock; /* spinlock for shared memory and LWLock
88 * allocation */
90 static HTAB *ShmemIndex = NULL; /* primary index hashtable for shmem */
94 * InitShmemAccess() --- set up basic pointers to shared memory.
96 * Note: the argument should be declared "PGShmemHeader *seghdr",
97 * but we use void to avoid having to include ipc.h in shmem.h.
99 void
100 InitShmemAccess(void *seghdr)
102 PGShmemHeader *shmhdr = (PGShmemHeader *) seghdr;
104 ShmemSegHdr = shmhdr;
105 ShmemBase = (void *) shmhdr;
106 ShmemEnd = (char *) ShmemBase + shmhdr->totalsize;
110 * InitShmemAllocation() --- set up shared-memory space allocation.
112 * This should be called only in the postmaster or a standalone backend.
114 void
115 InitShmemAllocation(void)
117 PGShmemHeader *shmhdr = ShmemSegHdr;
118 char *aligned;
120 Assert(shmhdr != NULL);
123 * Initialize the spinlock used by ShmemAlloc. We must use
124 * ShmemAllocUnlocked, since obviously ShmemAlloc can't be called yet.
126 ShmemLock = (slock_t *) ShmemAllocUnlocked(sizeof(slock_t));
128 SpinLockInit(ShmemLock);
131 * Allocations after this point should go through ShmemAlloc, which
132 * expects to allocate everything on cache line boundaries. Make sure the
133 * first allocation begins on a cache line boundary.
135 aligned = (char *)
136 (CACHELINEALIGN((((char *) shmhdr) + shmhdr->freeoffset)));
137 shmhdr->freeoffset = aligned - (char *) shmhdr;
139 /* ShmemIndex can't be set up yet (need LWLocks first) */
140 shmhdr->index = NULL;
141 ShmemIndex = (HTAB *) NULL;
145 * ShmemAlloc -- allocate max-aligned chunk from shared memory
147 * Throws error if request cannot be satisfied.
149 * Assumes ShmemLock and ShmemSegHdr are initialized.
151 void *
152 ShmemAlloc(Size size)
154 void *newSpace;
155 Size allocated_size;
157 newSpace = ShmemAllocRaw(size, &allocated_size);
158 if (!newSpace)
159 ereport(ERROR,
160 (errcode(ERRCODE_OUT_OF_MEMORY),
161 errmsg("out of shared memory (%zu bytes requested)",
162 size)));
163 return newSpace;
167 * ShmemAllocNoError -- allocate max-aligned chunk from shared memory
169 * As ShmemAlloc, but returns NULL if out of space, rather than erroring.
171 void *
172 ShmemAllocNoError(Size size)
174 Size allocated_size;
176 return ShmemAllocRaw(size, &allocated_size);
180 * ShmemAllocRaw -- allocate align chunk and return allocated size
182 * Also sets *allocated_size to the number of bytes allocated, which will
183 * be equal to the number requested plus any padding we choose to add.
185 static void *
186 ShmemAllocRaw(Size size, Size *allocated_size)
188 Size newStart;
189 Size newFree;
190 void *newSpace;
193 * Ensure all space is adequately aligned. We used to only MAXALIGN this
194 * space but experience has proved that on modern systems that is not good
195 * enough. Many parts of the system are very sensitive to critical data
196 * structures getting split across cache line boundaries. To avoid that,
197 * attempt to align the beginning of the allocation to a cache line
198 * boundary. The calling code will still need to be careful about how it
199 * uses the allocated space - e.g. by padding each element in an array of
200 * structures out to a power-of-two size - but without this, even that
201 * won't be sufficient.
203 size = CACHELINEALIGN(size);
204 *allocated_size = size;
206 Assert(ShmemSegHdr != NULL);
208 SpinLockAcquire(ShmemLock);
210 newStart = ShmemSegHdr->freeoffset;
212 newFree = newStart + size;
213 if (newFree <= ShmemSegHdr->totalsize)
215 newSpace = (void *) ((char *) ShmemBase + newStart);
216 ShmemSegHdr->freeoffset = newFree;
218 else
219 newSpace = NULL;
221 SpinLockRelease(ShmemLock);
223 /* note this assert is okay with newSpace == NULL */
224 Assert(newSpace == (void *) CACHELINEALIGN(newSpace));
226 return newSpace;
230 * ShmemAllocUnlocked -- allocate max-aligned chunk from shared memory
232 * Allocate space without locking ShmemLock. This should be used for,
233 * and only for, allocations that must happen before ShmemLock is ready.
235 * We consider maxalign, rather than cachealign, sufficient here.
237 void *
238 ShmemAllocUnlocked(Size size)
240 Size newStart;
241 Size newFree;
242 void *newSpace;
245 * Ensure allocated space is adequately aligned.
247 size = MAXALIGN(size);
249 Assert(ShmemSegHdr != NULL);
251 newStart = ShmemSegHdr->freeoffset;
253 newFree = newStart + size;
254 if (newFree > ShmemSegHdr->totalsize)
255 ereport(ERROR,
256 (errcode(ERRCODE_OUT_OF_MEMORY),
257 errmsg("out of shared memory (%zu bytes requested)",
258 size)));
259 ShmemSegHdr->freeoffset = newFree;
261 newSpace = (void *) ((char *) ShmemBase + newStart);
263 Assert(newSpace == (void *) MAXALIGN(newSpace));
265 return newSpace;
269 * ShmemAddrIsValid -- test if an address refers to shared memory
271 * Returns true if the pointer points within the shared memory segment.
273 bool
274 ShmemAddrIsValid(const void *addr)
276 return (addr >= ShmemBase) && (addr < ShmemEnd);
280 * InitShmemIndex() --- set up or attach to shmem index table.
282 void
283 InitShmemIndex(void)
285 HASHCTL info;
288 * Create the shared memory shmem index.
290 * Since ShmemInitHash calls ShmemInitStruct, which expects the ShmemIndex
291 * hashtable to exist already, we have a bit of a circularity problem in
292 * initializing the ShmemIndex itself. The special "ShmemIndex" hash
293 * table name will tell ShmemInitStruct to fake it.
295 info.keysize = SHMEM_INDEX_KEYSIZE;
296 info.entrysize = sizeof(ShmemIndexEnt);
298 ShmemIndex = ShmemInitHash("ShmemIndex",
299 SHMEM_INDEX_SIZE, SHMEM_INDEX_SIZE,
300 &info,
301 HASH_ELEM | HASH_STRINGS);
305 * ShmemInitHash -- Create and initialize, or attach to, a
306 * shared memory hash table.
308 * We assume caller is doing some kind of synchronization
309 * so that two processes don't try to create/initialize the same
310 * table at once. (In practice, all creations are done in the postmaster
311 * process; child processes should always be attaching to existing tables.)
313 * max_size is the estimated maximum number of hashtable entries. This is
314 * not a hard limit, but the access efficiency will degrade if it is
315 * exceeded substantially (since it's used to compute directory size and
316 * the hash table buckets will get overfull).
318 * init_size is the number of hashtable entries to preallocate. For a table
319 * whose maximum size is certain, this should be equal to max_size; that
320 * ensures that no run-time out-of-shared-memory failures can occur.
322 * *infoP and hash_flags must specify at least the entry sizes and key
323 * comparison semantics (see hash_create()). Flag bits and values specific
324 * to shared-memory hash tables are added here, except that callers may
325 * choose to specify HASH_PARTITION and/or HASH_FIXED_SIZE.
327 * Note: before Postgres 9.0, this function returned NULL for some failure
328 * cases. Now, it always throws error instead, so callers need not check
329 * for NULL.
331 HTAB *
332 ShmemInitHash(const char *name, /* table string name for shmem index */
333 long init_size, /* initial table size */
334 long max_size, /* max size of the table */
335 HASHCTL *infoP, /* info about key and bucket size */
336 int hash_flags) /* info about infoP */
338 bool found;
339 void *location;
342 * Hash tables allocated in shared memory have a fixed directory; it can't
343 * grow or other backends wouldn't be able to find it. So, make sure we
344 * make it big enough to start with.
346 * The shared memory allocator must be specified too.
348 infoP->dsize = infoP->max_dsize = hash_select_dirsize(max_size);
349 infoP->alloc = ShmemAllocNoError;
350 hash_flags |= HASH_SHARED_MEM | HASH_ALLOC | HASH_DIRSIZE;
352 /* look it up in the shmem index */
353 location = ShmemInitStruct(name,
354 hash_get_shared_size(infoP, hash_flags),
355 &found);
358 * if it already exists, attach to it rather than allocate and initialize
359 * new space
361 if (found)
362 hash_flags |= HASH_ATTACH;
364 /* Pass location of hashtable header to hash_create */
365 infoP->hctl = (HASHHDR *) location;
367 return hash_create(name, init_size, infoP, hash_flags);
371 * ShmemInitStruct -- Create/attach to a structure in shared memory.
373 * This is called during initialization to find or allocate
374 * a data structure in shared memory. If no other process
375 * has created the structure, this routine allocates space
376 * for it. If it exists already, a pointer to the existing
377 * structure is returned.
379 * Returns: pointer to the object. *foundPtr is set true if the object was
380 * already in the shmem index (hence, already initialized).
382 * Note: before Postgres 9.0, this function returned NULL for some failure
383 * cases. Now, it always throws error instead, so callers need not check
384 * for NULL.
386 void *
387 ShmemInitStruct(const char *name, Size size, bool *foundPtr)
389 ShmemIndexEnt *result;
390 void *structPtr;
392 LWLockAcquire(ShmemIndexLock, LW_EXCLUSIVE);
394 if (!ShmemIndex)
396 PGShmemHeader *shmemseghdr = ShmemSegHdr;
398 /* Must be trying to create/attach to ShmemIndex itself */
399 Assert(strcmp(name, "ShmemIndex") == 0);
401 if (IsUnderPostmaster)
403 /* Must be initializing a (non-standalone) backend */
404 Assert(shmemseghdr->index != NULL);
405 structPtr = shmemseghdr->index;
406 *foundPtr = true;
408 else
411 * If the shmem index doesn't exist, we are bootstrapping: we must
412 * be trying to init the shmem index itself.
414 * Notice that the ShmemIndexLock is released before the shmem
415 * index has been initialized. This should be OK because no other
416 * process can be accessing shared memory yet.
418 Assert(shmemseghdr->index == NULL);
419 structPtr = ShmemAlloc(size);
420 shmemseghdr->index = structPtr;
421 *foundPtr = false;
423 LWLockRelease(ShmemIndexLock);
424 return structPtr;
427 /* look it up in the shmem index */
428 result = (ShmemIndexEnt *)
429 hash_search(ShmemIndex, name, HASH_ENTER_NULL, foundPtr);
431 if (!result)
433 LWLockRelease(ShmemIndexLock);
434 ereport(ERROR,
435 (errcode(ERRCODE_OUT_OF_MEMORY),
436 errmsg("could not create ShmemIndex entry for data structure \"%s\"",
437 name)));
440 if (*foundPtr)
443 * Structure is in the shmem index so someone else has allocated it
444 * already. The size better be the same as the size we are trying to
445 * initialize to, or there is a name conflict (or worse).
447 if (result->size != size)
449 LWLockRelease(ShmemIndexLock);
450 ereport(ERROR,
451 (errmsg("ShmemIndex entry size is wrong for data structure"
452 " \"%s\": expected %zu, actual %zu",
453 name, size, result->size)));
455 structPtr = result->location;
457 else
459 Size allocated_size;
461 /* It isn't in the table yet. allocate and initialize it */
462 structPtr = ShmemAllocRaw(size, &allocated_size);
463 if (structPtr == NULL)
465 /* out of memory; remove the failed ShmemIndex entry */
466 hash_search(ShmemIndex, name, HASH_REMOVE, NULL);
467 LWLockRelease(ShmemIndexLock);
468 ereport(ERROR,
469 (errcode(ERRCODE_OUT_OF_MEMORY),
470 errmsg("not enough shared memory for data structure"
471 " \"%s\" (%zu bytes requested)",
472 name, size)));
474 result->size = size;
475 result->allocated_size = allocated_size;
476 result->location = structPtr;
479 LWLockRelease(ShmemIndexLock);
481 Assert(ShmemAddrIsValid(structPtr));
483 Assert(structPtr == (void *) CACHELINEALIGN(structPtr));
485 return structPtr;
490 * Add two Size values, checking for overflow
492 Size
493 add_size(Size s1, Size s2)
495 Size result;
497 result = s1 + s2;
498 /* We are assuming Size is an unsigned type here... */
499 if (result < s1 || result < s2)
500 ereport(ERROR,
501 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
502 errmsg("requested shared memory size overflows size_t")));
503 return result;
507 * Multiply two Size values, checking for overflow
509 Size
510 mul_size(Size s1, Size s2)
512 Size result;
514 if (s1 == 0 || s2 == 0)
515 return 0;
516 result = s1 * s2;
517 /* We are assuming Size is an unsigned type here... */
518 if (result / s2 != s1)
519 ereport(ERROR,
520 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
521 errmsg("requested shared memory size overflows size_t")));
522 return result;
525 /* SQL SRF showing allocated shared memory */
526 Datum
527 pg_get_shmem_allocations(PG_FUNCTION_ARGS)
529 #define PG_GET_SHMEM_SIZES_COLS 4
530 ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
531 HASH_SEQ_STATUS hstat;
532 ShmemIndexEnt *ent;
533 Size named_allocated = 0;
534 Datum values[PG_GET_SHMEM_SIZES_COLS];
535 bool nulls[PG_GET_SHMEM_SIZES_COLS];
537 InitMaterializedSRF(fcinfo, 0);
539 LWLockAcquire(ShmemIndexLock, LW_SHARED);
541 hash_seq_init(&hstat, ShmemIndex);
543 /* output all allocated entries */
544 memset(nulls, 0, sizeof(nulls));
545 while ((ent = (ShmemIndexEnt *) hash_seq_search(&hstat)) != NULL)
547 values[0] = CStringGetTextDatum(ent->key);
548 values[1] = Int64GetDatum((char *) ent->location - (char *) ShmemSegHdr);
549 values[2] = Int64GetDatum(ent->size);
550 values[3] = Int64GetDatum(ent->allocated_size);
551 named_allocated += ent->allocated_size;
553 tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc,
554 values, nulls);
557 /* output shared memory allocated but not counted via the shmem index */
558 values[0] = CStringGetTextDatum("<anonymous>");
559 nulls[1] = true;
560 values[2] = Int64GetDatum(ShmemSegHdr->freeoffset - named_allocated);
561 values[3] = values[2];
562 tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc, values, nulls);
564 /* output as-of-yet unused shared memory */
565 nulls[0] = true;
566 values[1] = Int64GetDatum(ShmemSegHdr->freeoffset);
567 nulls[1] = false;
568 values[2] = Int64GetDatum(ShmemSegHdr->totalsize - ShmemSegHdr->freeoffset);
569 values[3] = values[2];
570 tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc, values, nulls);
572 LWLockRelease(ShmemIndexLock);
574 return (Datum) 0;