| 1 |
/*------------------------------------------------------------------------- |
| 2 |
* |
| 3 |
* shmem.c |
| 4 |
* create shared memory and initialize shared memory data structures. |
| 5 |
* |
| 6 |
* Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group |
| 7 |
* Portions Copyright (c) 1994, Regents of the University of California |
| 8 |
* |
| 9 |
* |
| 10 |
* IDENTIFICATION |
| 11 |
* src/backend/storage/ipc/shmem.c |
| 12 |
* |
| 13 |
*------------------------------------------------------------------------- |
| 14 |
*/ |
| 15 |
/* |
| 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. |
| 21 |
* |
| 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). |
| 33 |
* |
| 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. |
| 49 |
* |
| 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. |
| 56 |
* |
| 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. |
| 64 |
*/ |
| 65 |
|
| 66 |
#include "postgres.h" |
| 67 |
|
| 68 |
#include "fmgr.h" |
| 69 |
#include "funcapi.h" |
| 70 |
#include "miscadmin.h" |
| 71 |
#include "port/pg_numa.h" |
| 72 |
#include "storage/lwlock.h" |
| 73 |
#include "storage/pg_shmem.h" |
| 74 |
#include "storage/shmem.h" |
| 75 |
#include "storage/spin.h" |
| 76 |
#include "utils/builtins.h" |
| 77 |
|
| 78 |
static void *ShmemAllocRaw(Size size, Size *allocated_size); |
| 79 |
|
| 80 |
/* shared memory global variables */ |
| 81 |
|
| 82 |
static PGShmemHeader *ShmemSegHdr; /* shared mem segment header */ |
| 83 |
|
| 84 |
static void *ShmemBase; /* start address of shared memory */ |
| 85 |
|
| 86 |
static void *ShmemEnd; /* end+1 address of shared memory */ |
| 87 |
|
| 88 |
slock_t *ShmemLock; /* spinlock for shared memory and LWLock |
| 89 |
* allocation */ |
| 90 |
|
| 91 |
static HTAB *ShmemIndex = NULL; /* primary index hashtable for shmem */ |
| 92 |
|
| 93 |
/* To get reliable results for NUMA inquiry we need to "touch pages" once */ |
| 94 |
static bool firstNumaTouch = true; |
| 95 |
|
| 96 |
Datum pg_numa_available(PG_FUNCTION_ARGS); |
| 97 |
|
| 98 |
/* |
| 99 |
* InitShmemAccess() --- set up basic pointers to shared memory. |
| 100 |
*/ |
| 101 |
void |
| 102 |
InitShmemAccess(PGShmemHeader *seghdr) |
| 103 |
{ |
| 104 |
ShmemSegHdr = seghdr; |
| 105 |
ShmemBase = seghdr; |
| 106 |
ShmemEnd = (char *) ShmemBase + seghdr->totalsize; |
| 107 |
} |
| 108 |
|
| 109 |
/* |
| 110 |
* InitShmemAllocation() --- set up shared-memory space allocation. |
| 111 |
* |
| 112 |
* This should be called only in the postmaster or a standalone backend. |
| 113 |
*/ |
| 114 |
void |
| 115 |
InitShmemAllocation(void) |
| 116 |
{ |
| 117 |
PGShmemHeader *shmhdr = ShmemSegHdr; |
| 118 |
char *aligned; |
| 119 |
|
| 120 |
Assert(shmhdr != NULL); |
| 121 |
|
| 122 |
/* |
| 123 |
* Initialize the spinlock used by ShmemAlloc. We must use |
| 124 |
* ShmemAllocUnlocked, since obviously ShmemAlloc can't be called yet. |
| 125 |
*/ |
| 126 |
ShmemLock = (slock_t *) ShmemAllocUnlocked(sizeof(slock_t)); |
| 127 |
|
| 128 |
SpinLockInit(ShmemLock); |
| 129 |
|
| 130 |
/* |
| 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. |
| 134 |
*/ |
| 135 |
aligned = (char *) |
| 136 |
(CACHELINEALIGN((((char *) shmhdr) + shmhdr->freeoffset))); |
| 137 |
shmhdr->freeoffset = aligned - (char *) shmhdr; |
| 138 |
|
| 139 |
/* ShmemIndex can't be set up yet (need LWLocks first) */ |
| 140 |
shmhdr->index = NULL; |
| 141 |
ShmemIndex = (HTAB *) NULL; |
| 142 |
} |
| 143 |
|
| 144 |
/* |
| 145 |
* ShmemAlloc -- allocate max-aligned chunk from shared memory |
| 146 |
* |
| 147 |
* Throws error if request cannot be satisfied. |
| 148 |
* |
| 149 |
* Assumes ShmemLock and ShmemSegHdr are initialized. |
| 150 |
*/ |
| 151 |
void * |
| 152 |
ShmemAlloc(Size size) |
| 153 |
{ |
| 154 |
void *newSpace; |
| 155 |
Size allocated_size; |
| 156 |
|
| 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; |
| 164 |
} |
| 165 |
|
| 166 |
/* |
| 167 |
* ShmemAllocNoError -- allocate max-aligned chunk from shared memory |
| 168 |
* |
| 169 |
* As ShmemAlloc, but returns NULL if out of space, rather than erroring. |
| 170 |
*/ |
| 171 |
void * |
| 172 |
ShmemAllocNoError(Size size) |
| 173 |
{ |
| 174 |
Size allocated_size; |
| 175 |
|
| 176 |
return ShmemAllocRaw(size, &allocated_size); |
| 177 |
} |
| 178 |
|
| 179 |
/* |
| 180 |
* ShmemAllocRaw -- allocate align chunk and return allocated size |
| 181 |
* |
| 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. |
| 184 |
*/ |
| 185 |
static void * |
| 186 |
ShmemAllocRaw(Size size, Size *allocated_size) |
| 187 |
{ |
| 188 |
Size newStart; |
| 189 |
Size newFree; |
| 190 |
void *newSpace; |
| 191 |
|
| 192 |
/* |
| 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. |
| 202 |
*/ |
| 203 |
size = CACHELINEALIGN(size); |
| 204 |
*allocated_size = size; |
| 205 |
|
| 206 |
Assert(ShmemSegHdr != NULL); |
| 207 |
|
| 208 |
SpinLockAcquire(ShmemLock); |
| 209 |
|
| 210 |
newStart = ShmemSegHdr->freeoffset; |
| 211 |
|
| 212 |
newFree = newStart + size; |
| 213 |
if (newFree <= ShmemSegHdr->totalsize) |
| 214 |
{ |
| 215 |
newSpace = (char *) ShmemBase + newStart; |
| 216 |
ShmemSegHdr->freeoffset = newFree; |
| 217 |
} |
| 218 |
else |
| 219 |
newSpace = NULL; |
| 220 |
|
| 221 |
SpinLockRelease(ShmemLock); |
| 222 |
|
| 223 |
/* note this assert is okay with newSpace == NULL */ |
| 224 |
Assert(newSpace == (void *) CACHELINEALIGN(newSpace)); |
| 225 |
|
| 226 |
return newSpace; |
| 227 |
} |
| 228 |
|
| 229 |
/* |
| 230 |
* ShmemAllocUnlocked -- allocate max-aligned chunk from shared memory |
| 231 |
* |
| 232 |
* Allocate space without locking ShmemLock. This should be used for, |
| 233 |
* and only for, allocations that must happen before ShmemLock is ready. |
| 234 |
* |
| 235 |
* We consider maxalign, rather than cachealign, sufficient here. |
| 236 |
*/ |
| 237 |
void * |
| 238 |
ShmemAllocUnlocked(Size size) |
| 239 |
{ |
| 240 |
Size newStart; |
| 241 |
Size newFree; |
| 242 |
void *newSpace; |
| 243 |
|
| 244 |
/* |
| 245 |
* Ensure allocated space is adequately aligned. |
| 246 |
*/ |
| 247 |
size = MAXALIGN(size); |
| 248 |
|
| 249 |
Assert(ShmemSegHdr != NULL); |
| 250 |
|
| 251 |
newStart = ShmemSegHdr->freeoffset; |
| 252 |
|
| 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; |
| 260 |
|
| 261 |
newSpace = (char *) ShmemBase + newStart; |
| 262 |
|
| 263 |
Assert(newSpace == (void *) MAXALIGN(newSpace)); |
| 264 |
|
| 265 |
return newSpace; |
| 266 |
} |
| 267 |
|
| 268 |
/* |
| 269 |
* ShmemAddrIsValid -- test if an address refers to shared memory |
| 270 |
* |
| 271 |
* Returns true if the pointer points within the shared memory segment. |
| 272 |
*/ |
| 273 |
bool |
| 274 |
ShmemAddrIsValid(const void *addr) |
| 275 |
{ |
| 276 |
return (addr >= ShmemBase) && (addr < ShmemEnd); |
| 277 |
} |
| 278 |
|
| 279 |
/* |
| 280 |
* InitShmemIndex() --- set up or attach to shmem index table. |
| 281 |
*/ |
| 282 |
void |
| 283 |
InitShmemIndex(void) |
| 284 |
{ |
| 285 |
HASHCTL info; |
| 286 |
|
| 287 |
/* |
| 288 |
* Create the shared memory shmem index. |
| 289 |
* |
| 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. |
| 294 |
*/ |
| 295 |
info.keysize = SHMEM_INDEX_KEYSIZE; |
| 296 |
info.entrysize = sizeof(ShmemIndexEnt); |
| 297 |
|
| 298 |
ShmemIndex = ShmemInitHash("ShmemIndex", |
| 299 |
SHMEM_INDEX_SIZE, SHMEM_INDEX_SIZE, |
| 300 |
&info, |
| 301 |
HASH_ELEM | HASH_STRINGS); |
| 302 |
} |
| 303 |
|
| 304 |
/* |
| 305 |
* ShmemInitHash -- Create and initialize, or attach to, a |
| 306 |
* shared memory hash table. |
| 307 |
* |
| 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.) |
| 312 |
* |
| 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). |
| 317 |
* |
| 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. |
| 321 |
* |
| 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. |
| 326 |
* |
| 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. |
| 330 |
*/ |
| 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 */ |
| 337 |
{ |
| 338 |
bool found; |
| 339 |
void *location; |
| 340 |
|
| 341 |
/* |
| 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. |
| 345 |
* |
| 346 |
* The shared memory allocator must be specified too. |
| 347 |
*/ |
| 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; |
| 351 |
|
| 352 |
/* look it up in the shmem index */ |
| 353 |
location = ShmemInitStruct(name, |
| 354 |
hash_get_shared_size(infoP, hash_flags), |
| 355 |
&found); |
| 356 |
|
| 357 |
/* |
| 358 |
* if it already exists, attach to it rather than allocate and initialize |
| 359 |
* new space |
| 360 |
*/ |
| 361 |
if (found) |
| 362 |
hash_flags |= HASH_ATTACH; |
| 363 |
|
| 364 |
/* Pass location of hashtable header to hash_create */ |
| 365 |
infoP->hctl = (HASHHDR *) location; |
| 366 |
|
| 367 |
return hash_create(name, init_size, infoP, hash_flags); |
| 368 |
} |
| 369 |
|
| 370 |
/* |
| 371 |
* ShmemInitStruct -- Create/attach to a structure in shared memory. |
| 372 |
* |
| 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. |
| 378 |
* |
| 379 |
* Returns: pointer to the object. *foundPtr is set true if the object was |
| 380 |
* already in the shmem index (hence, already initialized). |
| 381 |
* |
| 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. |
| 385 |
*/ |
| 386 |
void * |
| 387 |
ShmemInitStruct(const char *name, Size size, bool *foundPtr) |
| 388 |
{ |
| 389 |
ShmemIndexEnt *result; |
| 390 |
void *structPtr; |
| 391 |
|
| 392 |
LWLockAcquire(ShmemIndexLock, LW_EXCLUSIVE); |
| 393 |
|
| 394 |
if (!ShmemIndex) |
| 395 |
{ |
| 396 |
PGShmemHeader *shmemseghdr = ShmemSegHdr; |
| 397 |
|
| 398 |
/* Must be trying to create/attach to ShmemIndex itself */ |
| 399 |
Assert(strcmp(name, "ShmemIndex") == 0); |
| 400 |
|
| 401 |
if (IsUnderPostmaster) |
| 402 |
{ |
| 403 |
/* Must be initializing a (non-standalone) backend */ |
| 404 |
Assert(shmemseghdr->index != NULL); |
| 405 |
structPtr = shmemseghdr->index; |
| 406 |
*foundPtr = true; |
| 407 |
} |
| 408 |
else |
| 409 |
{ |
| 410 |
/* |
| 411 |
* If the shmem index doesn't exist, we are bootstrapping: we must |
| 412 |
* be trying to init the shmem index itself. |
| 413 |
* |
| 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. |
| 417 |
*/ |
| 418 |
Assert(shmemseghdr->index == NULL); |
| 419 |
structPtr = ShmemAlloc(size); |
| 420 |
shmemseghdr->index = structPtr; |
| 421 |
*foundPtr = false; |
| 422 |
} |
| 423 |
LWLockRelease(ShmemIndexLock); |
| 424 |
return structPtr; |
| 425 |
} |
| 426 |
|
| 427 |
/* look it up in the shmem index */ |
| 428 |
result = (ShmemIndexEnt *) |
| 429 |
hash_search(ShmemIndex, name, HASH_ENTER_NULL, foundPtr); |
| 430 |
|
| 431 |
if (!result) |
| 432 |
{ |
| 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))); |
| 438 |
} |
| 439 |
|
| 440 |
if (*foundPtr) |
| 441 |
{ |
| 442 |
/* |
| 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). |
| 446 |
*/ |
| 447 |
if (result->size != size) |
| 448 |
{ |
| 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))); |
| 454 |
} |
| 455 |
structPtr = result->location; |
| 456 |
} |
| 457 |
else |
| 458 |
{ |
| 459 |
Size allocated_size; |
| 460 |
|
| 461 |
/* It isn't in the table yet. allocate and initialize it */ |
| 462 |
structPtr = ShmemAllocRaw(size, &allocated_size); |
| 463 |
if (structPtr == NULL) |
| 464 |
{ |
| 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))); |
| 473 |
} |
| 474 |
result->size = size; |
| 475 |
result->allocated_size = allocated_size; |
| 476 |
result->location = structPtr; |
| 477 |
} |
| 478 |
|
| 479 |
LWLockRelease(ShmemIndexLock); |
| 480 |
|
| 481 |
Assert(ShmemAddrIsValid(structPtr)); |
| 482 |
|
| 483 |
Assert(structPtr == (void *) CACHELINEALIGN(structPtr)); |
| 484 |
|
| 485 |
return structPtr; |
| 486 |
} |
| 487 |
|
| 488 |
|
| 489 |
/* SQL SRF showing allocated shared memory */ |
| 490 |
Datum |
| 491 |
pg_get_shmem_allocations(PG_FUNCTION_ARGS) |
| 492 |
{ |
| 493 |
#define PG_GET_SHMEM_SIZES_COLS 4 |
| 494 |
ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo; |
| 495 |
HASH_SEQ_STATUS hstat; |
| 496 |
ShmemIndexEnt *ent; |
| 497 |
Size named_allocated = 0; |
| 498 |
Datum values[PG_GET_SHMEM_SIZES_COLS]; |
| 499 |
bool nulls[PG_GET_SHMEM_SIZES_COLS]; |
| 500 |
|
| 501 |
InitMaterializedSRF(fcinfo, 0); |
| 502 |
|
| 503 |
LWLockAcquire(ShmemIndexLock, LW_SHARED); |
| 504 |
|
| 505 |
hash_seq_init(&hstat, ShmemIndex); |
| 506 |
|
| 507 |
/* output all allocated entries */ |
| 508 |
memset(nulls, 0, sizeof(nulls)); |
| 509 |
while ((ent = (ShmemIndexEnt *) hash_seq_search(&hstat)) != NULL) |
| 510 |
{ |
| 511 |
values[0] = CStringGetTextDatum(ent->key); |
| 512 |
values[1] = Int64GetDatum((char *) ent->location - (char *) ShmemSegHdr); |
| 513 |
values[2] = Int64GetDatum(ent->size); |
| 514 |
values[3] = Int64GetDatum(ent->allocated_size); |
| 515 |
named_allocated += ent->allocated_size; |
| 516 |
|
| 517 |
tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc, |
| 518 |
values, nulls); |
| 519 |
} |
| 520 |
|
| 521 |
/* output shared memory allocated but not counted via the shmem index */ |
| 522 |
values[0] = CStringGetTextDatum("<anonymous>"); |
| 523 |
nulls[1] = true; |
| 524 |
values[2] = Int64GetDatum(ShmemSegHdr->freeoffset - named_allocated); |
| 525 |
values[3] = values[2]; |
| 526 |
tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc, values, nulls); |
| 527 |
|
| 528 |
/* output as-of-yet unused shared memory */ |
| 529 |
nulls[0] = true; |
| 530 |
values[1] = Int64GetDatum(ShmemSegHdr->freeoffset); |
| 531 |
nulls[1] = false; |
| 532 |
values[2] = Int64GetDatum(ShmemSegHdr->totalsize - ShmemSegHdr->freeoffset); |
| 533 |
values[3] = values[2]; |
| 534 |
tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc, values, nulls); |
| 535 |
|
| 536 |
LWLockRelease(ShmemIndexLock); |
| 537 |
|
| 538 |
return (Datum) 0; |
| 539 |
} |
| 540 |
|
| 541 |
/* |
| 542 |
* SQL SRF showing NUMA memory nodes for allocated shared memory |
| 543 |
* |
| 544 |
* Compared to pg_get_shmem_allocations(), this function does not return |
| 545 |
* information about shared anonymous allocations and unused shared memory. |
| 546 |
*/ |
| 547 |
Datum |
| 548 |
pg_get_shmem_allocations_numa(PG_FUNCTION_ARGS) |
| 549 |
{ |
| 550 |
#define PG_GET_SHMEM_NUMA_SIZES_COLS 3 |
| 551 |
ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo; |
| 552 |
HASH_SEQ_STATUS hstat; |
| 553 |
ShmemIndexEnt *ent; |
| 554 |
Datum values[PG_GET_SHMEM_NUMA_SIZES_COLS]; |
| 555 |
bool nulls[PG_GET_SHMEM_NUMA_SIZES_COLS]; |
| 556 |
Size os_page_size; |
| 557 |
void **page_ptrs; |
| 558 |
int *pages_status; |
| 559 |
uint64 shm_total_page_count, |
| 560 |
shm_ent_page_count, |
| 561 |
max_nodes; |
| 562 |
Size *nodes; |
| 563 |
|
| 564 |
if (pg_numa_init() == -1) |
| 565 |
elog(ERROR, "libnuma initialization failed or NUMA is not supported on this platform"); |
| 566 |
|
| 567 |
InitMaterializedSRF(fcinfo, 0); |
| 568 |
|
| 569 |
max_nodes = pg_numa_get_max_node(); |
| 570 |
nodes = palloc(sizeof(Size) * (max_nodes + 2)); |
| 571 |
|
| 572 |
/* |
| 573 |
* Shared memory allocations can vary in size and may not align with OS |
| 574 |
* memory page boundaries, while NUMA queries work on pages. |
| 575 |
* |
| 576 |
* To correctly map each allocation to NUMA nodes, we need to: 1. |
| 577 |
* Determine the OS memory page size. 2. Align each allocation's start/end |
| 578 |
* addresses to page boundaries. 3. Query NUMA node information for all |
| 579 |
* pages spanning the allocation. |
| 580 |
*/ |
| 581 |
os_page_size = pg_get_shmem_pagesize(); |
| 582 |
|
| 583 |
/* |
| 584 |
* Allocate memory for page pointers and status based on total shared |
| 585 |
* memory size. This simplified approach allocates enough space for all |
| 586 |
* pages in shared memory rather than calculating the exact requirements |
| 587 |
* for each segment. |
| 588 |
* |
| 589 |
* Add 1, because we don't know how exactly the segments align to OS |
| 590 |
* pages, so the allocation might use one more memory page. In practice |
| 591 |
* this is not very likely, and moreover we have more entries, each of |
| 592 |
* them using only fraction of the total pages. |
| 593 |
*/ |
| 594 |
shm_total_page_count = (ShmemSegHdr->totalsize / os_page_size) + 1; |
| 595 |
page_ptrs = palloc0(sizeof(void *) * shm_total_page_count); |
| 596 |
pages_status = palloc(sizeof(int) * shm_total_page_count); |
| 597 |
|
| 598 |
if (firstNumaTouch) |
| 599 |
elog(DEBUG1, "NUMA: page-faulting shared memory segments for proper NUMA readouts"); |
| 600 |
|
| 601 |
LWLockAcquire(ShmemIndexLock, LW_SHARED); |
| 602 |
|
| 603 |
hash_seq_init(&hstat, ShmemIndex); |
| 604 |
|
| 605 |
/* output all allocated entries */ |
| 606 |
while ((ent = (ShmemIndexEnt *) hash_seq_search(&hstat)) != NULL) |
| 607 |
{ |
| 608 |
int i; |
| 609 |
char *startptr, |
| 610 |
*endptr; |
| 611 |
Size total_len; |
| 612 |
|
| 613 |
/* |
| 614 |
* Calculate the range of OS pages used by this segment. The segment |
| 615 |
* may start / end half-way through a page, we want to count these |
| 616 |
* pages too. So we align the start/end pointers down/up, and then |
| 617 |
* calculate the number of pages from that. |
| 618 |
*/ |
| 619 |
startptr = (char *) TYPEALIGN_DOWN(os_page_size, ent->location); |
| 620 |
endptr = (char *) TYPEALIGN(os_page_size, |
| 621 |
(char *) ent->location + ent->allocated_size); |
| 622 |
total_len = (endptr - startptr); |
| 623 |
|
| 624 |
shm_ent_page_count = total_len / os_page_size; |
| 625 |
|
| 626 |
/* |
| 627 |
* If we ever get 0xff (-1) back from kernel inquiry, then we probably |
| 628 |
* have a bug in mapping buffers to OS pages. |
| 629 |
*/ |
| 630 |
memset(pages_status, 0xff, sizeof(int) * shm_ent_page_count); |
| 631 |
|
| 632 |
/* |
| 633 |
* Setup page_ptrs[] with pointers to all OS pages for this segment, |
| 634 |
* and get the NUMA status using pg_numa_query_pages. |
| 635 |
* |
| 636 |
* In order to get reliable results we also need to touch memory |
| 637 |
* pages, so that inquiry about NUMA memory node doesn't return -2 |
| 638 |
* (ENOENT, which indicates unmapped/unallocated pages). |
| 639 |
*/ |
| 640 |
for (i = 0; i < shm_ent_page_count; i++) |
| 641 |
{ |
| 642 |
page_ptrs[i] = startptr + (i * os_page_size); |
| 643 |
|
| 644 |
if (firstNumaTouch) |
| 645 |
pg_numa_touch_mem_if_required(page_ptrs[i]); |
| 646 |
|
| 647 |
CHECK_FOR_INTERRUPTS(); |
| 648 |
} |
| 649 |
|
| 650 |
if (pg_numa_query_pages(0, shm_ent_page_count, page_ptrs, pages_status) == -1) |
| 651 |
elog(ERROR, "failed NUMA pages inquiry status: %m"); |
| 652 |
|
| 653 |
/* Count number of NUMA nodes used for this shared memory entry */ |
| 654 |
memset(nodes, 0, sizeof(Size) * (max_nodes + 2)); |
| 655 |
|
| 656 |
for (i = 0; i < shm_ent_page_count; i++) |
| 657 |
{ |
| 658 |
int s = pages_status[i]; |
| 659 |
|
| 660 |
/* Ensure we are adding only valid index to the array */ |
| 661 |
if (s >= 0 && s <= max_nodes) |
| 662 |
{ |
| 663 |
/* valid NUMA node */ |
| 664 |
nodes[s]++; |
| 665 |
continue; |
| 666 |
} |
| 667 |
else if (s == -2) |
| 668 |
{ |
| 669 |
/* -2 means ENOENT (e.g. page was moved to swap) */ |
| 670 |
nodes[max_nodes + 1]++; |
| 671 |
continue; |
| 672 |
} |
| 673 |
|
| 674 |
elog(ERROR, "invalid NUMA node id outside of allowed range " |
| 675 |
"[0, " UINT64_FORMAT "]: %d", max_nodes, s); |
| 676 |
} |
| 677 |
|
| 678 |
/* no NULLs for regular nodes */ |
| 679 |
memset(nulls, 0, sizeof(nulls)); |
| 680 |
|
| 681 |
/* |
| 682 |
* Add one entry for each NUMA node, including those without allocated |
| 683 |
* memory for this segment. |
| 684 |
*/ |
| 685 |
for (i = 0; i <= max_nodes; i++) |
| 686 |
{ |
| 687 |
values[0] = CStringGetTextDatum(ent->key); |
| 688 |
values[1] = i; |
| 689 |
values[2] = Int64GetDatum(nodes[i] * os_page_size); |
| 690 |
|
| 691 |
tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc, |
| 692 |
values, nulls); |
| 693 |
} |
| 694 |
|
| 695 |
/* The last entry is used for pages without a NUMA node. */ |
| 696 |
nulls[1] = true; |
| 697 |
values[0] = CStringGetTextDatum(ent->key); |
| 698 |
values[2] = Int64GetDatum(nodes[max_nodes + 1] * os_page_size); |
| 699 |
|
| 700 |
tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc, |
| 701 |
values, nulls); |
| 702 |
} |
| 703 |
|
| 704 |
LWLockRelease(ShmemIndexLock); |
| 705 |
firstNumaTouch = false; |
| 706 |
|
| 707 |
return (Datum) 0; |
| 708 |
} |
| 709 |
|
| 710 |
/* |
| 711 |
* Determine the memory page size used for the shared memory segment. |
| 712 |
* |
| 713 |
* If the shared segment was allocated using huge pages, returns the size of |
| 714 |
* a huge page. Otherwise returns the size of regular memory page. |
| 715 |
* |
| 716 |
* This should be used only after the server is started. |
| 717 |
*/ |
| 718 |
Size |
| 719 |
pg_get_shmem_pagesize(void) |
| 720 |
{ |
| 721 |
Size os_page_size; |
| 722 |
#ifdef WIN32 |
| 723 |
SYSTEM_INFO sysinfo; |
| 724 |
|
| 725 |
GetSystemInfo(&sysinfo); |
| 726 |
os_page_size = sysinfo.dwPageSize; |
| 727 |
#else |
| 728 |
os_page_size = sysconf(_SC_PAGESIZE); |
| 729 |
#endif |
| 730 |
|
| 731 |
Assert(IsUnderPostmaster); |
| 732 |
Assert(huge_pages_status != HUGE_PAGES_UNKNOWN); |
| 733 |
|
| 734 |
if (huge_pages_status == HUGE_PAGES_ON) |
| 735 |
GetHugePageSize(&os_page_size, NULL); |
| 736 |
|
| 737 |
return os_page_size; |
| 738 |
} |
| 739 |
|
| 740 |
Datum |
| 741 |
pg_numa_available(PG_FUNCTION_ARGS) |
| 742 |
{ |
| 743 |
PG_RETURN_BOOL(pg_numa_init() != -1); |
| 744 |
} |