| 1 |
/*------------------------------------------------------------------------- |
| 2 |
* |
| 3 |
* dsm_impl.c |
| 4 |
* manage dynamic shared memory segments |
| 5 |
* |
| 6 |
* This file provides low-level APIs for creating and destroying shared |
| 7 |
* memory segments using several different possible techniques. We refer |
| 8 |
* to these segments as dynamic because they can be created, altered, and |
| 9 |
* destroyed at any point during the server life cycle. This is unlike |
| 10 |
* the main shared memory segment, of which there is always exactly one |
| 11 |
* and which is always mapped at a fixed address in every PostgreSQL |
| 12 |
* background process. |
| 13 |
* |
| 14 |
* Because not all systems provide the same primitives in this area, nor |
| 15 |
* do all primitives behave the same way on all systems, we provide |
| 16 |
* several implementations of this facility. Many systems implement |
| 17 |
* POSIX shared memory (shm_open etc.), which is well-suited to our needs |
| 18 |
* in this area, with the exception that shared memory identifiers live |
| 19 |
* in a flat system-wide namespace, raising the uncomfortable prospect of |
| 20 |
* name collisions with other processes (including other copies of |
| 21 |
* PostgreSQL) running on the same system. Some systems only support |
| 22 |
* the older System V shared memory interface (shmget etc.) which is |
| 23 |
* also usable; however, the default allocation limits are often quite |
| 24 |
* small, and the namespace is even more restricted. |
| 25 |
* |
| 26 |
* We also provide an mmap-based shared memory implementation. This may |
| 27 |
* be useful on systems that provide shared memory via a special-purpose |
| 28 |
* filesystem; by opting for this implementation, the user can even |
| 29 |
* control precisely where their shared memory segments are placed. It |
| 30 |
* can also be used as a fallback for systems where shm_open and shmget |
| 31 |
* are not available or can't be used for some reason. Of course, |
| 32 |
* mapping a file residing on an actual spinning disk is a fairly poor |
| 33 |
* approximation for shared memory because writeback may hurt performance |
| 34 |
* substantially, but there should be few systems where we must make do |
| 35 |
* with such poor tools. |
| 36 |
* |
| 37 |
* As ever, Windows requires its own implementation. |
| 38 |
* |
| 39 |
* Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group |
| 40 |
* Portions Copyright (c) 1994, Regents of the University of California |
| 41 |
* |
| 42 |
* |
| 43 |
* IDENTIFICATION |
| 44 |
* src/backend/storage/ipc/dsm_impl.c |
| 45 |
* |
| 46 |
*------------------------------------------------------------------------- |
| 47 |
*/ |
| 48 |
|
| 49 |
#include "postgres.h" |
| 50 |
|
| 51 |
#include <fcntl.h> |
| 52 |
#include <signal.h> |
| 53 |
#include <unistd.h> |
| 54 |
#ifndef WIN32 |
| 55 |
#include <sys/mman.h> |
| 56 |
#include <sys/ipc.h> |
| 57 |
#include <sys/shm.h> |
| 58 |
#include <sys/stat.h> |
| 59 |
#endif |
| 60 |
|
| 61 |
#include "common/file_perm.h" |
| 62 |
#include "libpq/pqsignal.h" |
| 63 |
#include "miscadmin.h" |
| 64 |
#include "pgstat.h" |
| 65 |
#include "portability/mem.h" |
| 66 |
#include "postmaster/postmaster.h" |
| 67 |
#include "storage/dsm_impl.h" |
| 68 |
#include "storage/fd.h" |
| 69 |
#include "utils/guc.h" |
| 70 |
#include "utils/memutils.h" |
| 71 |
|
| 72 |
#ifdef USE_DSM_POSIX |
| 73 |
static bool dsm_impl_posix(dsm_op op, dsm_handle handle, Size request_size, |
| 74 |
void **impl_private, void **mapped_address, |
| 75 |
Size *mapped_size, int elevel); |
| 76 |
static int dsm_impl_posix_resize(int fd, off_t size); |
| 77 |
#endif |
| 78 |
#ifdef USE_DSM_SYSV |
| 79 |
static bool dsm_impl_sysv(dsm_op op, dsm_handle handle, Size request_size, |
| 80 |
void **impl_private, void **mapped_address, |
| 81 |
Size *mapped_size, int elevel); |
| 82 |
#endif |
| 83 |
#ifdef USE_DSM_WINDOWS |
| 84 |
static bool dsm_impl_windows(dsm_op op, dsm_handle handle, Size request_size, |
| 85 |
void **impl_private, void **mapped_address, |
| 86 |
Size *mapped_size, int elevel); |
| 87 |
#endif |
| 88 |
#ifdef USE_DSM_MMAP |
| 89 |
static bool dsm_impl_mmap(dsm_op op, dsm_handle handle, Size request_size, |
| 90 |
void **impl_private, void **mapped_address, |
| 91 |
Size *mapped_size, int elevel); |
| 92 |
#endif |
| 93 |
static int errcode_for_dynamic_shared_memory(void); |
| 94 |
|
| 95 |
const struct config_enum_entry dynamic_shared_memory_options[] = { |
| 96 |
#ifdef USE_DSM_POSIX |
| 97 |
{"posix", DSM_IMPL_POSIX, false}, |
| 98 |
#endif |
| 99 |
#ifdef USE_DSM_SYSV |
| 100 |
{"sysv", DSM_IMPL_SYSV, false}, |
| 101 |
#endif |
| 102 |
#ifdef USE_DSM_WINDOWS |
| 103 |
{"windows", DSM_IMPL_WINDOWS, false}, |
| 104 |
#endif |
| 105 |
#ifdef USE_DSM_MMAP |
| 106 |
{"mmap", DSM_IMPL_MMAP, false}, |
| 107 |
#endif |
| 108 |
{NULL, 0, false} |
| 109 |
}; |
| 110 |
|
| 111 |
/* Implementation selector. */ |
| 112 |
int dynamic_shared_memory_type = DEFAULT_DYNAMIC_SHARED_MEMORY_TYPE; |
| 113 |
|
| 114 |
/* Amount of space reserved for DSM segments in the main area. */ |
| 115 |
int min_dynamic_shared_memory; |
| 116 |
|
| 117 |
/* Size of buffer to be used for zero-filling. */ |
| 118 |
#define ZBUFFER_SIZE 8192 |
| 119 |
|
| 120 |
#define SEGMENT_NAME_PREFIX "Global/PostgreSQL" |
| 121 |
|
| 122 |
/*------ |
| 123 |
* Perform a low-level shared memory operation in a platform-specific way, |
| 124 |
* as dictated by the selected implementation. Each implementation is |
| 125 |
* required to implement the following primitives. |
| 126 |
* |
| 127 |
* DSM_OP_CREATE. Create a segment whose size is the request_size and |
| 128 |
* map it. |
| 129 |
* |
| 130 |
* DSM_OP_ATTACH. Map the segment, whose size must be the request_size. |
| 131 |
* |
| 132 |
* DSM_OP_DETACH. Unmap the segment. |
| 133 |
* |
| 134 |
* DSM_OP_DESTROY. Unmap the segment, if it is mapped. Destroy the |
| 135 |
* segment. |
| 136 |
* |
| 137 |
* Arguments: |
| 138 |
* op: The operation to be performed. |
| 139 |
* handle: The handle of an existing object, or for DSM_OP_CREATE, the |
| 140 |
* identifier for the new handle the caller wants created. |
| 141 |
* request_size: For DSM_OP_CREATE, the requested size. Otherwise, 0. |
| 142 |
* impl_private: Private, implementation-specific data. Will be a pointer |
| 143 |
* to NULL for the first operation on a shared memory segment within this |
| 144 |
* backend; thereafter, it will point to the value to which it was set |
| 145 |
* on the previous call. |
| 146 |
* mapped_address: Pointer to start of current mapping; pointer to NULL |
| 147 |
* if none. Updated with new mapping address. |
| 148 |
* mapped_size: Pointer to size of current mapping; pointer to 0 if none. |
| 149 |
* Updated with new mapped size. |
| 150 |
* elevel: Level at which to log errors. |
| 151 |
* |
| 152 |
* Return value: true on success, false on failure. When false is returned, |
| 153 |
* a message should first be logged at the specified elevel, except in the |
| 154 |
* case where DSM_OP_CREATE experiences a name collision, which should |
| 155 |
* silently return false. |
| 156 |
*----- |
| 157 |
*/ |
| 158 |
bool |
| 159 |
dsm_impl_op(dsm_op op, dsm_handle handle, Size request_size, |
| 160 |
void **impl_private, void **mapped_address, Size *mapped_size, |
| 161 |
int elevel) |
| 162 |
{ |
| 163 |
Assert(op == DSM_OP_CREATE || request_size == 0); |
| 164 |
Assert((op != DSM_OP_CREATE && op != DSM_OP_ATTACH) || |
| 165 |
(*mapped_address == NULL && *mapped_size == 0)); |
| 166 |
|
| 167 |
switch (dynamic_shared_memory_type) |
| 168 |
{ |
| 169 |
#ifdef USE_DSM_POSIX |
| 170 |
case DSM_IMPL_POSIX: |
| 171 |
return dsm_impl_posix(op, handle, request_size, impl_private, |
| 172 |
mapped_address, mapped_size, elevel); |
| 173 |
#endif |
| 174 |
#ifdef USE_DSM_SYSV |
| 175 |
case DSM_IMPL_SYSV: |
| 176 |
return dsm_impl_sysv(op, handle, request_size, impl_private, |
| 177 |
mapped_address, mapped_size, elevel); |
| 178 |
#endif |
| 179 |
#ifdef USE_DSM_WINDOWS |
| 180 |
case DSM_IMPL_WINDOWS: |
| 181 |
return dsm_impl_windows(op, handle, request_size, impl_private, |
| 182 |
mapped_address, mapped_size, elevel); |
| 183 |
#endif |
| 184 |
#ifdef USE_DSM_MMAP |
| 185 |
case DSM_IMPL_MMAP: |
| 186 |
return dsm_impl_mmap(op, handle, request_size, impl_private, |
| 187 |
mapped_address, mapped_size, elevel); |
| 188 |
#endif |
| 189 |
default: |
| 190 |
elog(ERROR, "unexpected dynamic shared memory type: %d", |
| 191 |
dynamic_shared_memory_type); |
| 192 |
return false; |
| 193 |
} |
| 194 |
} |
| 195 |
|
| 196 |
#ifdef USE_DSM_POSIX |
| 197 |
/* |
| 198 |
* Operating system primitives to support POSIX shared memory. |
| 199 |
* |
| 200 |
* POSIX shared memory segments are created and attached using shm_open() |
| 201 |
* and shm_unlink(); other operations, such as sizing or mapping the |
| 202 |
* segment, are performed as if the shared memory segments were files. |
| 203 |
* |
| 204 |
* Indeed, on some platforms, they may be implemented that way. While |
| 205 |
* POSIX shared memory segments seem intended to exist in a flat namespace, |
| 206 |
* some operating systems may implement them as files, even going so far |
| 207 |
* to treat a request for /xyz as a request to create a file by that name |
| 208 |
* in the root directory. Users of such broken platforms should select |
| 209 |
* a different shared memory implementation. |
| 210 |
*/ |
| 211 |
static bool |
| 212 |
dsm_impl_posix(dsm_op op, dsm_handle handle, Size request_size, |
| 213 |
void **impl_private, void **mapped_address, Size *mapped_size, |
| 214 |
int elevel) |
| 215 |
{ |
| 216 |
char name[64]; |
| 217 |
int flags; |
| 218 |
int fd; |
| 219 |
char *address; |
| 220 |
|
| 221 |
snprintf(name, 64, "/PostgreSQL.%u", handle); |
| 222 |
|
| 223 |
/* Handle teardown cases. */ |
| 224 |
if (op == DSM_OP_DETACH || op == DSM_OP_DESTROY) |
| 225 |
{ |
| 226 |
if (*mapped_address != NULL |
| 227 |
&& munmap(*mapped_address, *mapped_size) != 0) |
| 228 |
{ |
| 229 |
ereport(elevel, |
| 230 |
(errcode_for_dynamic_shared_memory(), |
| 231 |
errmsg("could not unmap shared memory segment \"%s\": %m", |
| 232 |
name))); |
| 233 |
return false; |
| 234 |
} |
| 235 |
*mapped_address = NULL; |
| 236 |
*mapped_size = 0; |
| 237 |
if (op == DSM_OP_DESTROY && shm_unlink(name) != 0) |
| 238 |
{ |
| 239 |
ereport(elevel, |
| 240 |
(errcode_for_dynamic_shared_memory(), |
| 241 |
errmsg("could not remove shared memory segment \"%s\": %m", |
| 242 |
name))); |
| 243 |
return false; |
| 244 |
} |
| 245 |
return true; |
| 246 |
} |
| 247 |
|
| 248 |
/* |
| 249 |
* Create new segment or open an existing one for attach. |
| 250 |
* |
| 251 |
* Even though we will close the FD before returning, it seems desirable |
| 252 |
* to use Reserve/ReleaseExternalFD, to reduce the probability of EMFILE |
| 253 |
* failure. The fact that we won't hold the FD open long justifies using |
| 254 |
* ReserveExternalFD rather than AcquireExternalFD, though. |
| 255 |
*/ |
| 256 |
ReserveExternalFD(); |
| 257 |
|
| 258 |
flags = O_RDWR | (op == DSM_OP_CREATE ? O_CREAT | O_EXCL : 0); |
| 259 |
if ((fd = shm_open(name, flags, PG_FILE_MODE_OWNER)) == -1) |
| 260 |
{ |
| 261 |
ReleaseExternalFD(); |
| 262 |
if (op == DSM_OP_ATTACH || errno != EEXIST) |
| 263 |
ereport(elevel, |
| 264 |
(errcode_for_dynamic_shared_memory(), |
| 265 |
errmsg("could not open shared memory segment \"%s\": %m", |
| 266 |
name))); |
| 267 |
return false; |
| 268 |
} |
| 269 |
|
| 270 |
/* |
| 271 |
* If we're attaching the segment, determine the current size; if we are |
| 272 |
* creating the segment, set the size to the requested value. |
| 273 |
*/ |
| 274 |
if (op == DSM_OP_ATTACH) |
| 275 |
{ |
| 276 |
struct stat st; |
| 277 |
|
| 278 |
if (fstat(fd, &st) != 0) |
| 279 |
{ |
| 280 |
int save_errno; |
| 281 |
|
| 282 |
/* Back out what's already been done. */ |
| 283 |
save_errno = errno; |
| 284 |
close(fd); |
| 285 |
ReleaseExternalFD(); |
| 286 |
errno = save_errno; |
| 287 |
|
| 288 |
ereport(elevel, |
| 289 |
(errcode_for_dynamic_shared_memory(), |
| 290 |
errmsg("could not stat shared memory segment \"%s\": %m", |
| 291 |
name))); |
| 292 |
return false; |
| 293 |
} |
| 294 |
request_size = st.st_size; |
| 295 |
} |
| 296 |
else if (dsm_impl_posix_resize(fd, request_size) != 0) |
| 297 |
{ |
| 298 |
int save_errno; |
| 299 |
|
| 300 |
/* Back out what's already been done. */ |
| 301 |
save_errno = errno; |
| 302 |
close(fd); |
| 303 |
ReleaseExternalFD(); |
| 304 |
shm_unlink(name); |
| 305 |
errno = save_errno; |
| 306 |
|
| 307 |
ereport(elevel, |
| 308 |
(errcode_for_dynamic_shared_memory(), |
| 309 |
errmsg("could not resize shared memory segment \"%s\" to %zu bytes: %m", |
| 310 |
name, request_size))); |
| 311 |
return false; |
| 312 |
} |
| 313 |
|
| 314 |
/* Map it. */ |
| 315 |
address = mmap(NULL, request_size, PROT_READ | PROT_WRITE, |
| 316 |
MAP_SHARED | MAP_HASSEMAPHORE | MAP_NOSYNC, fd, 0); |
| 317 |
if (address == MAP_FAILED) |
| 318 |
{ |
| 319 |
int save_errno; |
| 320 |
|
| 321 |
/* Back out what's already been done. */ |
| 322 |
save_errno = errno; |
| 323 |
close(fd); |
| 324 |
ReleaseExternalFD(); |
| 325 |
if (op == DSM_OP_CREATE) |
| 326 |
shm_unlink(name); |
| 327 |
errno = save_errno; |
| 328 |
|
| 329 |
ereport(elevel, |
| 330 |
(errcode_for_dynamic_shared_memory(), |
| 331 |
errmsg("could not map shared memory segment \"%s\": %m", |
| 332 |
name))); |
| 333 |
return false; |
| 334 |
} |
| 335 |
*mapped_address = address; |
| 336 |
*mapped_size = request_size; |
| 337 |
close(fd); |
| 338 |
ReleaseExternalFD(); |
| 339 |
|
| 340 |
return true; |
| 341 |
} |
| 342 |
|
| 343 |
/* |
| 344 |
* Set the size of a virtual memory region associated with a file descriptor. |
| 345 |
* If necessary, also ensure that virtual memory is actually allocated by the |
| 346 |
* operating system, to avoid nasty surprises later. |
| 347 |
* |
| 348 |
* Returns non-zero if either truncation or allocation fails, and sets errno. |
| 349 |
*/ |
| 350 |
static int |
| 351 |
dsm_impl_posix_resize(int fd, off_t size) |
| 352 |
{ |
| 353 |
int rc; |
| 354 |
int save_errno; |
| 355 |
sigset_t save_sigmask; |
| 356 |
|
| 357 |
/* |
| 358 |
* Block all blockable signals, except SIGQUIT. posix_fallocate() can run |
| 359 |
* for quite a long time, and is an all-or-nothing operation. If we |
| 360 |
* allowed SIGUSR1 to interrupt us repeatedly (for example, due to |
| 361 |
* recovery conflicts), the retry loop might never succeed. |
| 362 |
*/ |
| 363 |
if (IsUnderPostmaster) |
| 364 |
sigprocmask(SIG_SETMASK, &BlockSig, &save_sigmask); |
| 365 |
|
| 366 |
pgstat_report_wait_start(WAIT_EVENT_DSM_ALLOCATE); |
| 367 |
#if defined(HAVE_POSIX_FALLOCATE) && defined(__linux__) |
| 368 |
|
| 369 |
/* |
| 370 |
* On Linux, a shm_open fd is backed by a tmpfs file. If we were to use |
| 371 |
* ftruncate, the file would contain a hole. Accessing memory backed by a |
| 372 |
* hole causes tmpfs to allocate pages, which fails with SIGBUS if there |
| 373 |
* is no more tmpfs space available. So we ask tmpfs to allocate pages |
| 374 |
* here, so we can fail gracefully with ENOSPC now rather than risking |
| 375 |
* SIGBUS later. |
| 376 |
* |
| 377 |
* We still use a traditional EINTR retry loop to handle SIGCONT. |
| 378 |
* posix_fallocate() doesn't restart automatically, and we don't want this |
| 379 |
* to fail if you attach a debugger. |
| 380 |
*/ |
| 381 |
do |
| 382 |
{ |
| 383 |
rc = posix_fallocate(fd, 0, size); |
| 384 |
} while (rc == EINTR); |
| 385 |
|
| 386 |
/* |
| 387 |
* The caller expects errno to be set, but posix_fallocate() doesn't set |
| 388 |
* it. Instead it returns error numbers directly. So set errno, even |
| 389 |
* though we'll also return rc to indicate success or failure. |
| 390 |
*/ |
| 391 |
errno = rc; |
| 392 |
#else |
| 393 |
/* Extend the file to the requested size. */ |
| 394 |
do |
| 395 |
{ |
| 396 |
rc = ftruncate(fd, size); |
| 397 |
} while (rc < 0 && errno == EINTR); |
| 398 |
#endif |
| 399 |
pgstat_report_wait_end(); |
| 400 |
|
| 401 |
if (IsUnderPostmaster) |
| 402 |
{ |
| 403 |
save_errno = errno; |
| 404 |
sigprocmask(SIG_SETMASK, &save_sigmask, NULL); |
| 405 |
errno = save_errno; |
| 406 |
} |
| 407 |
|
| 408 |
return rc; |
| 409 |
} |
| 410 |
|
| 411 |
#endif /* USE_DSM_POSIX */ |
| 412 |
|
| 413 |
#ifdef USE_DSM_SYSV |
| 414 |
/* |
| 415 |
* Operating system primitives to support System V shared memory. |
| 416 |
* |
| 417 |
* System V shared memory segments are manipulated using shmget(), shmat(), |
| 418 |
* shmdt(), and shmctl(). As the default allocation limits for System V |
| 419 |
* shared memory are usually quite low, the POSIX facilities may be |
| 420 |
* preferable; but those are not supported everywhere. |
| 421 |
*/ |
| 422 |
static bool |
| 423 |
dsm_impl_sysv(dsm_op op, dsm_handle handle, Size request_size, |
| 424 |
void **impl_private, void **mapped_address, Size *mapped_size, |
| 425 |
int elevel) |
| 426 |
{ |
| 427 |
key_t key; |
| 428 |
int ident; |
| 429 |
char *address; |
| 430 |
char name[64]; |
| 431 |
int *ident_cache; |
| 432 |
|
| 433 |
/* |
| 434 |
* POSIX shared memory and mmap-based shared memory identify segments with |
| 435 |
* names. To avoid needless error message variation, we use the handle as |
| 436 |
* the name. |
| 437 |
*/ |
| 438 |
snprintf(name, 64, "%u", handle); |
| 439 |
|
| 440 |
/* |
| 441 |
* The System V shared memory namespace is very restricted; names are of |
| 442 |
* type key_t, which is expected to be some sort of integer data type, but |
| 443 |
* not necessarily the same one as dsm_handle. Since we use dsm_handle to |
| 444 |
* identify shared memory segments across processes, this might seem like |
| 445 |
* a problem, but it's really not. If dsm_handle is bigger than key_t, |
| 446 |
* the cast below might truncate away some bits from the handle the |
| 447 |
* user-provided, but it'll truncate exactly the same bits away in exactly |
| 448 |
* the same fashion every time we use that handle, which is all that |
| 449 |
* really matters. Conversely, if dsm_handle is smaller than key_t, we |
| 450 |
* won't use the full range of available key space, but that's no big deal |
| 451 |
* either. |
| 452 |
* |
| 453 |
* We do make sure that the key isn't negative, because that might not be |
| 454 |
* portable. |
| 455 |
*/ |
| 456 |
key = (key_t) handle; |
| 457 |
if (key < 1) /* avoid compiler warning if type is unsigned */ |
| 458 |
key = -key; |
| 459 |
|
| 460 |
/* |
| 461 |
* There's one special key, IPC_PRIVATE, which can't be used. If we end |
| 462 |
* up with that value by chance during a create operation, just pretend it |
| 463 |
* already exists, so that caller will retry. If we run into it anywhere |
| 464 |
* else, the caller has passed a handle that doesn't correspond to |
| 465 |
* anything we ever created, which should not happen. |
| 466 |
*/ |
| 467 |
if (key == IPC_PRIVATE) |
| 468 |
{ |
| 469 |
if (op != DSM_OP_CREATE) |
| 470 |
elog(DEBUG4, "System V shared memory key may not be IPC_PRIVATE"); |
| 471 |
errno = EEXIST; |
| 472 |
return false; |
| 473 |
} |
| 474 |
|
| 475 |
/* |
| 476 |
* Before we can do anything with a shared memory segment, we have to map |
| 477 |
* the shared memory key to a shared memory identifier using shmget(). To |
| 478 |
* avoid repeated lookups, we store the key using impl_private. |
| 479 |
*/ |
| 480 |
if (*impl_private != NULL) |
| 481 |
{ |
| 482 |
ident_cache = *impl_private; |
| 483 |
ident = *ident_cache; |
| 484 |
} |
| 485 |
else |
| 486 |
{ |
| 487 |
int flags = IPCProtection; |
| 488 |
size_t segsize; |
| 489 |
|
| 490 |
/* |
| 491 |
* Allocate the memory BEFORE acquiring the resource, so that we don't |
| 492 |
* leak the resource if memory allocation fails. |
| 493 |
*/ |
| 494 |
ident_cache = MemoryContextAlloc(TopMemoryContext, sizeof(int)); |
| 495 |
|
| 496 |
/* |
| 497 |
* When using shmget to find an existing segment, we must pass the |
| 498 |
* size as 0. Passing a non-zero size which is greater than the |
| 499 |
* actual size will result in EINVAL. |
| 500 |
*/ |
| 501 |
segsize = 0; |
| 502 |
|
| 503 |
if (op == DSM_OP_CREATE) |
| 504 |
{ |
| 505 |
flags |= IPC_CREAT | IPC_EXCL; |
| 506 |
segsize = request_size; |
| 507 |
} |
| 508 |
|
| 509 |
if ((ident = shmget(key, segsize, flags)) == -1) |
| 510 |
{ |
| 511 |
if (op == DSM_OP_ATTACH || errno != EEXIST) |
| 512 |
{ |
| 513 |
int save_errno = errno; |
| 514 |
|
| 515 |
pfree(ident_cache); |
| 516 |
errno = save_errno; |
| 517 |
ereport(elevel, |
| 518 |
(errcode_for_dynamic_shared_memory(), |
| 519 |
errmsg("could not get shared memory segment: %m"))); |
| 520 |
} |
| 521 |
return false; |
| 522 |
} |
| 523 |
|
| 524 |
*ident_cache = ident; |
| 525 |
*impl_private = ident_cache; |
| 526 |
} |
| 527 |
|
| 528 |
/* Handle teardown cases. */ |
| 529 |
if (op == DSM_OP_DETACH || op == DSM_OP_DESTROY) |
| 530 |
{ |
| 531 |
pfree(ident_cache); |
| 532 |
*impl_private = NULL; |
| 533 |
if (*mapped_address != NULL && shmdt(*mapped_address) != 0) |
| 534 |
{ |
| 535 |
ereport(elevel, |
| 536 |
(errcode_for_dynamic_shared_memory(), |
| 537 |
errmsg("could not unmap shared memory segment \"%s\": %m", |
| 538 |
name))); |
| 539 |
return false; |
| 540 |
} |
| 541 |
*mapped_address = NULL; |
| 542 |
*mapped_size = 0; |
| 543 |
if (op == DSM_OP_DESTROY && shmctl(ident, IPC_RMID, NULL) < 0) |
| 544 |
{ |
| 545 |
ereport(elevel, |
| 546 |
(errcode_for_dynamic_shared_memory(), |
| 547 |
errmsg("could not remove shared memory segment \"%s\": %m", |
| 548 |
name))); |
| 549 |
return false; |
| 550 |
} |
| 551 |
return true; |
| 552 |
} |
| 553 |
|
| 554 |
/* If we're attaching it, we must use IPC_STAT to determine the size. */ |
| 555 |
if (op == DSM_OP_ATTACH) |
| 556 |
{ |
| 557 |
struct shmid_ds shm; |
| 558 |
|
| 559 |
if (shmctl(ident, IPC_STAT, &shm) != 0) |
| 560 |
{ |
| 561 |
ereport(elevel, |
| 562 |
(errcode_for_dynamic_shared_memory(), |
| 563 |
errmsg("could not stat shared memory segment \"%s\": %m", |
| 564 |
name))); |
| 565 |
return false; |
| 566 |
} |
| 567 |
request_size = shm.shm_segsz; |
| 568 |
} |
| 569 |
|
| 570 |
/* Map it. */ |
| 571 |
address = shmat(ident, NULL, PG_SHMAT_FLAGS); |
| 572 |
if (address == (void *) -1) |
| 573 |
{ |
| 574 |
int save_errno; |
| 575 |
|
| 576 |
/* Back out what's already been done. */ |
| 577 |
save_errno = errno; |
| 578 |
if (op == DSM_OP_CREATE) |
| 579 |
shmctl(ident, IPC_RMID, NULL); |
| 580 |
errno = save_errno; |
| 581 |
|
| 582 |
ereport(elevel, |
| 583 |
(errcode_for_dynamic_shared_memory(), |
| 584 |
errmsg("could not map shared memory segment \"%s\": %m", |
| 585 |
name))); |
| 586 |
return false; |
| 587 |
} |
| 588 |
*mapped_address = address; |
| 589 |
*mapped_size = request_size; |
| 590 |
|
| 591 |
return true; |
| 592 |
} |
| 593 |
#endif |
| 594 |
|
| 595 |
#ifdef USE_DSM_WINDOWS |
| 596 |
/* |
| 597 |
* Operating system primitives to support Windows shared memory. |
| 598 |
* |
| 599 |
* Windows shared memory implementation is done using file mapping |
| 600 |
* which can be backed by either physical file or system paging file. |
| 601 |
* Current implementation uses system paging file as other effects |
| 602 |
* like performance are not clear for physical file and it is used in similar |
| 603 |
* way for main shared memory in windows. |
| 604 |
* |
| 605 |
* A memory mapping object is a kernel object - they always get deleted when |
| 606 |
* the last reference to them goes away, either explicitly via a CloseHandle or |
| 607 |
* when the process containing the reference exits. |
| 608 |
*/ |
| 609 |
static bool |
| 610 |
dsm_impl_windows(dsm_op op, dsm_handle handle, Size request_size, |
| 611 |
void **impl_private, void **mapped_address, |
| 612 |
Size *mapped_size, int elevel) |
| 613 |
{ |
| 614 |
char *address; |
| 615 |
HANDLE hmap; |
| 616 |
char name[64]; |
| 617 |
MEMORY_BASIC_INFORMATION info; |
| 618 |
|
| 619 |
/* |
| 620 |
* Storing the shared memory segment in the Global\ namespace, can allow |
| 621 |
* any process running in any session to access that file mapping object |
| 622 |
* provided that the caller has the required access rights. But to avoid |
| 623 |
* issues faced in main shared memory, we are using the naming convention |
| 624 |
* similar to main shared memory. We can change here once issue mentioned |
| 625 |
* in GetSharedMemName is resolved. |
| 626 |
*/ |
| 627 |
snprintf(name, 64, "%s.%u", SEGMENT_NAME_PREFIX, handle); |
| 628 |
|
| 629 |
/* |
| 630 |
* Handle teardown cases. Since Windows automatically destroys the object |
| 631 |
* when no references remain, we can treat it the same as detach. |
| 632 |
*/ |
| 633 |
if (op == DSM_OP_DETACH || op == DSM_OP_DESTROY) |
| 634 |
{ |
| 635 |
if (*mapped_address != NULL |
| 636 |
&& UnmapViewOfFile(*mapped_address) == 0) |
| 637 |
{ |
| 638 |
_dosmaperr(GetLastError()); |
| 639 |
ereport(elevel, |
| 640 |
(errcode_for_dynamic_shared_memory(), |
| 641 |
errmsg("could not unmap shared memory segment \"%s\": %m", |
| 642 |
name))); |
| 643 |
return false; |
| 644 |
} |
| 645 |
if (*impl_private != NULL |
| 646 |
&& CloseHandle(*impl_private) == 0) |
| 647 |
{ |
| 648 |
_dosmaperr(GetLastError()); |
| 649 |
ereport(elevel, |
| 650 |
(errcode_for_dynamic_shared_memory(), |
| 651 |
errmsg("could not remove shared memory segment \"%s\": %m", |
| 652 |
name))); |
| 653 |
return false; |
| 654 |
} |
| 655 |
|
| 656 |
*impl_private = NULL; |
| 657 |
*mapped_address = NULL; |
| 658 |
*mapped_size = 0; |
| 659 |
return true; |
| 660 |
} |
| 661 |
|
| 662 |
/* Create new segment or open an existing one for attach. */ |
| 663 |
if (op == DSM_OP_CREATE) |
| 664 |
{ |
| 665 |
DWORD size_high; |
| 666 |
DWORD size_low; |
| 667 |
DWORD errcode; |
| 668 |
|
| 669 |
/* Shifts >= the width of the type are undefined. */ |
| 670 |
#ifdef _WIN64 |
| 671 |
size_high = request_size >> 32; |
| 672 |
#else |
| 673 |
size_high = 0; |
| 674 |
#endif |
| 675 |
size_low = (DWORD) request_size; |
| 676 |
|
| 677 |
/* CreateFileMapping might not clear the error code on success */ |
| 678 |
SetLastError(0); |
| 679 |
|
| 680 |
hmap = CreateFileMapping(INVALID_HANDLE_VALUE, /* Use the pagefile */ |
| 681 |
NULL, /* Default security attrs */ |
| 682 |
PAGE_READWRITE, /* Memory is read/write */ |
| 683 |
size_high, /* Upper 32 bits of size */ |
| 684 |
size_low, /* Lower 32 bits of size */ |
| 685 |
name); |
| 686 |
|
| 687 |
errcode = GetLastError(); |
| 688 |
if (errcode == ERROR_ALREADY_EXISTS || errcode == ERROR_ACCESS_DENIED) |
| 689 |
{ |
| 690 |
/* |
| 691 |
* On Windows, when the segment already exists, a handle for the |
| 692 |
* existing segment is returned. We must close it before |
| 693 |
* returning. However, if the existing segment is created by a |
| 694 |
* service, then it returns ERROR_ACCESS_DENIED. We don't do |
| 695 |
* _dosmaperr here, so errno won't be modified. |
| 696 |
*/ |
| 697 |
if (hmap) |
| 698 |
CloseHandle(hmap); |
| 699 |
return false; |
| 700 |
} |
| 701 |
|
| 702 |
if (!hmap) |
| 703 |
{ |
| 704 |
_dosmaperr(errcode); |
| 705 |
ereport(elevel, |
| 706 |
(errcode_for_dynamic_shared_memory(), |
| 707 |
errmsg("could not create shared memory segment \"%s\": %m", |
| 708 |
name))); |
| 709 |
return false; |
| 710 |
} |
| 711 |
} |
| 712 |
else |
| 713 |
{ |
| 714 |
hmap = OpenFileMapping(FILE_MAP_WRITE | FILE_MAP_READ, |
| 715 |
FALSE, /* do not inherit the name */ |
| 716 |
name); /* name of mapping object */ |
| 717 |
if (!hmap) |
| 718 |
{ |
| 719 |
_dosmaperr(GetLastError()); |
| 720 |
ereport(elevel, |
| 721 |
(errcode_for_dynamic_shared_memory(), |
| 722 |
errmsg("could not open shared memory segment \"%s\": %m", |
| 723 |
name))); |
| 724 |
return false; |
| 725 |
} |
| 726 |
} |
| 727 |
|
| 728 |
/* Map it. */ |
| 729 |
address = MapViewOfFile(hmap, FILE_MAP_WRITE | FILE_MAP_READ, |
| 730 |
0, 0, 0); |
| 731 |
if (!address) |
| 732 |
{ |
| 733 |
int save_errno; |
| 734 |
|
| 735 |
_dosmaperr(GetLastError()); |
| 736 |
/* Back out what's already been done. */ |
| 737 |
save_errno = errno; |
| 738 |
CloseHandle(hmap); |
| 739 |
errno = save_errno; |
| 740 |
|
| 741 |
ereport(elevel, |
| 742 |
(errcode_for_dynamic_shared_memory(), |
| 743 |
errmsg("could not map shared memory segment \"%s\": %m", |
| 744 |
name))); |
| 745 |
return false; |
| 746 |
} |
| 747 |
|
| 748 |
/* |
| 749 |
* VirtualQuery gives size in page_size units, which is 4K for Windows. We |
| 750 |
* need size only when we are attaching, but it's better to get the size |
| 751 |
* when creating new segment to keep size consistent both for |
| 752 |
* DSM_OP_CREATE and DSM_OP_ATTACH. |
| 753 |
*/ |
| 754 |
if (VirtualQuery(address, &info, sizeof(info)) == 0) |
| 755 |
{ |
| 756 |
int save_errno; |
| 757 |
|
| 758 |
_dosmaperr(GetLastError()); |
| 759 |
/* Back out what's already been done. */ |
| 760 |
save_errno = errno; |
| 761 |
UnmapViewOfFile(address); |
| 762 |
CloseHandle(hmap); |
| 763 |
errno = save_errno; |
| 764 |
|
| 765 |
ereport(elevel, |
| 766 |
(errcode_for_dynamic_shared_memory(), |
| 767 |
errmsg("could not stat shared memory segment \"%s\": %m", |
| 768 |
name))); |
| 769 |
return false; |
| 770 |
} |
| 771 |
|
| 772 |
*mapped_address = address; |
| 773 |
*mapped_size = info.RegionSize; |
| 774 |
*impl_private = hmap; |
| 775 |
|
| 776 |
return true; |
| 777 |
} |
| 778 |
#endif |
| 779 |
|
| 780 |
#ifdef USE_DSM_MMAP |
| 781 |
/* |
| 782 |
* Operating system primitives to support mmap-based shared memory. |
| 783 |
* |
| 784 |
* Calling this "shared memory" is somewhat of a misnomer, because what |
| 785 |
* we're really doing is creating a bunch of files and mapping them into |
| 786 |
* our address space. The operating system may feel obliged to |
| 787 |
* synchronize the contents to disk even if nothing is being paged out, |
| 788 |
* which will not serve us well. The user can relocate the pg_dynshmem |
| 789 |
* directory to a ramdisk to avoid this problem, if available. |
| 790 |
*/ |
| 791 |
static bool |
| 792 |
dsm_impl_mmap(dsm_op op, dsm_handle handle, Size request_size, |
| 793 |
void **impl_private, void **mapped_address, Size *mapped_size, |
| 794 |
int elevel) |
| 795 |
{ |
| 796 |
char name[64]; |
| 797 |
int flags; |
| 798 |
int fd; |
| 799 |
char *address; |
| 800 |
|
| 801 |
snprintf(name, 64, PG_DYNSHMEM_DIR "/" PG_DYNSHMEM_MMAP_FILE_PREFIX "%u", |
| 802 |
handle); |
| 803 |
|
| 804 |
/* Handle teardown cases. */ |
| 805 |
if (op == DSM_OP_DETACH || op == DSM_OP_DESTROY) |
| 806 |
{ |
| 807 |
if (*mapped_address != NULL |
| 808 |
&& munmap(*mapped_address, *mapped_size) != 0) |
| 809 |
{ |
| 810 |
ereport(elevel, |
| 811 |
(errcode_for_dynamic_shared_memory(), |
| 812 |
errmsg("could not unmap shared memory segment \"%s\": %m", |
| 813 |
name))); |
| 814 |
return false; |
| 815 |
} |
| 816 |
*mapped_address = NULL; |
| 817 |
*mapped_size = 0; |
| 818 |
if (op == DSM_OP_DESTROY && unlink(name) != 0) |
| 819 |
{ |
| 820 |
ereport(elevel, |
| 821 |
(errcode_for_dynamic_shared_memory(), |
| 822 |
errmsg("could not remove shared memory segment \"%s\": %m", |
| 823 |
name))); |
| 824 |
return false; |
| 825 |
} |
| 826 |
return true; |
| 827 |
} |
| 828 |
|
| 829 |
/* Create new segment or open an existing one for attach. */ |
| 830 |
flags = O_RDWR | (op == DSM_OP_CREATE ? O_CREAT | O_EXCL : 0); |
| 831 |
if ((fd = OpenTransientFile(name, flags)) == -1) |
| 832 |
{ |
| 833 |
if (op == DSM_OP_ATTACH || errno != EEXIST) |
| 834 |
ereport(elevel, |
| 835 |
(errcode_for_dynamic_shared_memory(), |
| 836 |
errmsg("could not open shared memory segment \"%s\": %m", |
| 837 |
name))); |
| 838 |
return false; |
| 839 |
} |
| 840 |
|
| 841 |
/* |
| 842 |
* If we're attaching the segment, determine the current size; if we are |
| 843 |
* creating the segment, set the size to the requested value. |
| 844 |
*/ |
| 845 |
if (op == DSM_OP_ATTACH) |
| 846 |
{ |
| 847 |
struct stat st; |
| 848 |
|
| 849 |
if (fstat(fd, &st) != 0) |
| 850 |
{ |
| 851 |
int save_errno; |
| 852 |
|
| 853 |
/* Back out what's already been done. */ |
| 854 |
save_errno = errno; |
| 855 |
CloseTransientFile(fd); |
| 856 |
errno = save_errno; |
| 857 |
|
| 858 |
ereport(elevel, |
| 859 |
(errcode_for_dynamic_shared_memory(), |
| 860 |
errmsg("could not stat shared memory segment \"%s\": %m", |
| 861 |
name))); |
| 862 |
return false; |
| 863 |
} |
| 864 |
request_size = st.st_size; |
| 865 |
} |
| 866 |
else |
| 867 |
{ |
| 868 |
/* |
| 869 |
* Allocate a buffer full of zeros. |
| 870 |
* |
| 871 |
* Note: palloc zbuffer, instead of just using a local char array, to |
| 872 |
* ensure it is reasonably well-aligned; this may save a few cycles |
| 873 |
* transferring data to the kernel. |
| 874 |
*/ |
| 875 |
char *zbuffer = (char *) palloc0(ZBUFFER_SIZE); |
| 876 |
Size remaining = request_size; |
| 877 |
bool success = true; |
| 878 |
|
| 879 |
/* |
| 880 |
* Zero-fill the file. We have to do this the hard way to ensure that |
| 881 |
* all the file space has really been allocated, so that we don't |
| 882 |
* later seg fault when accessing the memory mapping. This is pretty |
| 883 |
* pessimal. |
| 884 |
*/ |
| 885 |
while (success && remaining > 0) |
| 886 |
{ |
| 887 |
Size goal = remaining; |
| 888 |
|
| 889 |
if (goal > ZBUFFER_SIZE) |
| 890 |
goal = ZBUFFER_SIZE; |
| 891 |
pgstat_report_wait_start(WAIT_EVENT_DSM_FILL_ZERO_WRITE); |
| 892 |
if (write(fd, zbuffer, goal) == goal) |
| 893 |
remaining -= goal; |
| 894 |
else |
| 895 |
success = false; |
| 896 |
pgstat_report_wait_end(); |
| 897 |
} |
| 898 |
|
| 899 |
if (!success) |
| 900 |
{ |
| 901 |
int save_errno; |
| 902 |
|
| 903 |
/* Back out what's already been done. */ |
| 904 |
save_errno = errno; |
| 905 |
CloseTransientFile(fd); |
| 906 |
unlink(name); |
| 907 |
errno = save_errno ? save_errno : ENOSPC; |
| 908 |
|
| 909 |
ereport(elevel, |
| 910 |
(errcode_for_dynamic_shared_memory(), |
| 911 |
errmsg("could not resize shared memory segment \"%s\" to %zu bytes: %m", |
| 912 |
name, request_size))); |
| 913 |
return false; |
| 914 |
} |
| 915 |
} |
| 916 |
|
| 917 |
/* Map it. */ |
| 918 |
address = mmap(NULL, request_size, PROT_READ | PROT_WRITE, |
| 919 |
MAP_SHARED | MAP_HASSEMAPHORE | MAP_NOSYNC, fd, 0); |
| 920 |
if (address == MAP_FAILED) |
| 921 |
{ |
| 922 |
int save_errno; |
| 923 |
|
| 924 |
/* Back out what's already been done. */ |
| 925 |
save_errno = errno; |
| 926 |
CloseTransientFile(fd); |
| 927 |
if (op == DSM_OP_CREATE) |
| 928 |
unlink(name); |
| 929 |
errno = save_errno; |
| 930 |
|
| 931 |
ereport(elevel, |
| 932 |
(errcode_for_dynamic_shared_memory(), |
| 933 |
errmsg("could not map shared memory segment \"%s\": %m", |
| 934 |
name))); |
| 935 |
return false; |
| 936 |
} |
| 937 |
*mapped_address = address; |
| 938 |
*mapped_size = request_size; |
| 939 |
|
| 940 |
if (CloseTransientFile(fd) != 0) |
| 941 |
{ |
| 942 |
ereport(elevel, |
| 943 |
(errcode_for_file_access(), |
| 944 |
errmsg("could not close shared memory segment \"%s\": %m", |
| 945 |
name))); |
| 946 |
return false; |
| 947 |
} |
| 948 |
|
| 949 |
return true; |
| 950 |
} |
| 951 |
#endif |
| 952 |
|
| 953 |
/* |
| 954 |
* Implementation-specific actions that must be performed when a segment is to |
| 955 |
* be preserved even when no backend has it attached. |
| 956 |
* |
| 957 |
* Except on Windows, we don't need to do anything at all. But since Windows |
| 958 |
* cleans up segments automatically when no references remain, we duplicate |
| 959 |
* the segment handle into the postmaster process. The postmaster needn't |
| 960 |
* do anything to receive the handle; Windows transfers it automatically. |
| 961 |
*/ |
| 962 |
void |
| 963 |
dsm_impl_pin_segment(dsm_handle handle, void *impl_private, |
| 964 |
void **impl_private_pm_handle) |
| 965 |
{ |
| 966 |
switch (dynamic_shared_memory_type) |
| 967 |
{ |
| 968 |
#ifdef USE_DSM_WINDOWS |
| 969 |
case DSM_IMPL_WINDOWS: |
| 970 |
if (IsUnderPostmaster) |
| 971 |
{ |
| 972 |
HANDLE hmap; |
| 973 |
|
| 974 |
if (!DuplicateHandle(GetCurrentProcess(), impl_private, |
| 975 |
PostmasterHandle, &hmap, 0, FALSE, |
| 976 |
DUPLICATE_SAME_ACCESS)) |
| 977 |
{ |
| 978 |
char name[64]; |
| 979 |
|
| 980 |
snprintf(name, 64, "%s.%u", SEGMENT_NAME_PREFIX, handle); |
| 981 |
_dosmaperr(GetLastError()); |
| 982 |
ereport(ERROR, |
| 983 |
(errcode_for_dynamic_shared_memory(), |
| 984 |
errmsg("could not duplicate handle for \"%s\": %m", |
| 985 |
name))); |
| 986 |
} |
| 987 |
|
| 988 |
/* |
| 989 |
* Here, we remember the handle that we created in the |
| 990 |
* postmaster process. This handle isn't actually usable in |
| 991 |
* any process other than the postmaster, but that doesn't |
| 992 |
* matter. We're just holding onto it so that, if the segment |
| 993 |
* is unpinned, dsm_impl_unpin_segment can close it. |
| 994 |
*/ |
| 995 |
*impl_private_pm_handle = hmap; |
| 996 |
} |
| 997 |
break; |
| 998 |
#endif |
| 999 |
default: |
| 1000 |
break; |
| 1001 |
} |
| 1002 |
} |
| 1003 |
|
| 1004 |
/* |
| 1005 |
* Implementation-specific actions that must be performed when a segment is no |
| 1006 |
* longer to be preserved, so that it will be cleaned up when all backends |
| 1007 |
* have detached from it. |
| 1008 |
* |
| 1009 |
* Except on Windows, we don't need to do anything at all. For Windows, we |
| 1010 |
* close the extra handle that dsm_impl_pin_segment created in the |
| 1011 |
* postmaster's process space. |
| 1012 |
*/ |
| 1013 |
void |
| 1014 |
dsm_impl_unpin_segment(dsm_handle handle, void **impl_private) |
| 1015 |
{ |
| 1016 |
switch (dynamic_shared_memory_type) |
| 1017 |
{ |
| 1018 |
#ifdef USE_DSM_WINDOWS |
| 1019 |
case DSM_IMPL_WINDOWS: |
| 1020 |
if (IsUnderPostmaster) |
| 1021 |
{ |
| 1022 |
if (*impl_private && |
| 1023 |
!DuplicateHandle(PostmasterHandle, *impl_private, |
| 1024 |
NULL, NULL, 0, FALSE, |
| 1025 |
DUPLICATE_CLOSE_SOURCE)) |
| 1026 |
{ |
| 1027 |
char name[64]; |
| 1028 |
|
| 1029 |
snprintf(name, 64, "%s.%u", SEGMENT_NAME_PREFIX, handle); |
| 1030 |
_dosmaperr(GetLastError()); |
| 1031 |
ereport(ERROR, |
| 1032 |
(errcode_for_dynamic_shared_memory(), |
| 1033 |
errmsg("could not duplicate handle for \"%s\": %m", |
| 1034 |
name))); |
| 1035 |
} |
| 1036 |
|
| 1037 |
*impl_private = NULL; |
| 1038 |
} |
| 1039 |
break; |
| 1040 |
#endif |
| 1041 |
default: |
| 1042 |
break; |
| 1043 |
} |
| 1044 |
} |
| 1045 |
|
| 1046 |
static int |
| 1047 |
errcode_for_dynamic_shared_memory(void) |
| 1048 |
{ |
| 1049 |
if (errno == EFBIG || errno == ENOMEM) |
| 1050 |
return errcode(ERRCODE_OUT_OF_MEMORY); |
| 1051 |
else |
| 1052 |
return errcode_for_file_access(); |
| 1053 |
} |