| 1 |
/*------------------------------------------------------------------------- |
| 2 |
* |
| 3 |
* fd.c |
| 4 |
* Virtual file descriptor code. |
| 5 |
* |
| 6 |
* Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group |
| 7 |
* Portions Copyright (c) 1994, Regents of the University of California |
| 8 |
* |
| 9 |
* IDENTIFICATION |
| 10 |
* src/backend/storage/file/fd.c |
| 11 |
* |
| 12 |
* NOTES: |
| 13 |
* |
| 14 |
* This code manages a cache of 'virtual' file descriptors (VFDs). |
| 15 |
* The server opens many file descriptors for a variety of reasons, |
| 16 |
* including base tables, scratch files (e.g., sort and hash spool |
| 17 |
* files), and random calls to C library routines like system(3); it |
| 18 |
* is quite easy to exceed system limits on the number of open files a |
| 19 |
* single process can have. (This is around 1024 on many modern |
| 20 |
* operating systems, but may be lower on others.) |
| 21 |
* |
| 22 |
* VFDs are managed as an LRU pool, with actual OS file descriptors |
| 23 |
* being opened and closed as needed. Obviously, if a routine is |
| 24 |
* opened using these interfaces, all subsequent operations must also |
| 25 |
* be through these interfaces (the File type is not a real file |
| 26 |
* descriptor). |
| 27 |
* |
| 28 |
* For this scheme to work, most (if not all) routines throughout the |
| 29 |
* server should use these interfaces instead of calling the C library |
| 30 |
* routines (e.g., open(2) and fopen(3)) themselves. Otherwise, we |
| 31 |
* may find ourselves short of real file descriptors anyway. |
| 32 |
* |
| 33 |
* INTERFACE ROUTINES |
| 34 |
* |
| 35 |
* PathNameOpenFile and OpenTemporaryFile are used to open virtual files. |
| 36 |
* A File opened with OpenTemporaryFile is automatically deleted when the |
| 37 |
* File is closed, either explicitly or implicitly at end of transaction or |
| 38 |
* process exit. PathNameOpenFile is intended for files that are held open |
| 39 |
* for a long time, like relation files. It is the caller's responsibility |
| 40 |
* to close them, there is no automatic mechanism in fd.c for that. |
| 41 |
* |
| 42 |
* PathName(Create|Open|Delete)Temporary(File|Dir) are used to manage |
| 43 |
* temporary files that have names so that they can be shared between |
| 44 |
* backends. Such files are automatically closed and count against the |
| 45 |
* temporary file limit of the backend that creates them, but unlike anonymous |
| 46 |
* files they are not automatically deleted. See sharedfileset.c for a shared |
| 47 |
* ownership mechanism that provides automatic cleanup for shared files when |
| 48 |
* the last of a group of backends detaches. |
| 49 |
* |
| 50 |
* AllocateFile, AllocateDir, OpenPipeStream and OpenTransientFile are |
| 51 |
* wrappers around fopen(3), opendir(3), popen(3) and open(2), respectively. |
| 52 |
* They behave like the corresponding native functions, except that the handle |
| 53 |
* is registered with the current subtransaction, and will be automatically |
| 54 |
* closed at abort. These are intended mainly for short operations like |
| 55 |
* reading a configuration file; there is a limit on the number of files that |
| 56 |
* can be opened using these functions at any one time. |
| 57 |
* |
| 58 |
* Finally, BasicOpenFile is just a thin wrapper around open() that can |
| 59 |
* release file descriptors in use by the virtual file descriptors if |
| 60 |
* necessary. There is no automatic cleanup of file descriptors returned by |
| 61 |
* BasicOpenFile, it is solely the caller's responsibility to close the file |
| 62 |
* descriptor by calling close(2). |
| 63 |
* |
| 64 |
* If a non-virtual file descriptor needs to be held open for any length of |
| 65 |
* time, report it to fd.c by calling AcquireExternalFD or ReserveExternalFD |
| 66 |
* (and eventually ReleaseExternalFD), so that we can take it into account |
| 67 |
* while deciding how many VFDs can be open. This applies to FDs obtained |
| 68 |
* with BasicOpenFile as well as those obtained without use of any fd.c API. |
| 69 |
* |
| 70 |
*------------------------------------------------------------------------- |
| 71 |
*/ |
| 72 |
|
| 73 |
#include "postgres.h" |
| 74 |
|
| 75 |
#include <dirent.h> |
| 76 |
#include <sys/file.h> |
| 77 |
#include <sys/param.h> |
| 78 |
#include <sys/resource.h> /* for getrlimit */ |
| 79 |
#include <sys/stat.h> |
| 80 |
#include <sys/types.h> |
| 81 |
#ifndef WIN32 |
| 82 |
#include <sys/mman.h> |
| 83 |
#endif |
| 84 |
#include <limits.h> |
| 85 |
#include <unistd.h> |
| 86 |
#include <fcntl.h> |
| 87 |
|
| 88 |
#include "access/xact.h" |
| 89 |
#include "access/xlog.h" |
| 90 |
#include "catalog/pg_tablespace.h" |
| 91 |
#include "common/file_perm.h" |
| 92 |
#include "common/file_utils.h" |
| 93 |
#include "common/pg_prng.h" |
| 94 |
#include "miscadmin.h" |
| 95 |
#include "pgstat.h" |
| 96 |
#include "postmaster/startup.h" |
| 97 |
#include "storage/aio.h" |
| 98 |
#include "storage/fd.h" |
| 99 |
#include "storage/ipc.h" |
| 100 |
#include "utils/guc.h" |
| 101 |
#include "utils/guc_hooks.h" |
| 102 |
#include "utils/resowner.h" |
| 103 |
#include "utils/varlena.h" |
| 104 |
|
| 105 |
/* Define PG_FLUSH_DATA_WORKS if we have an implementation for pg_flush_data */ |
| 106 |
#if defined(HAVE_SYNC_FILE_RANGE) |
| 107 |
#define PG_FLUSH_DATA_WORKS 1 |
| 108 |
#elif !defined(WIN32) && defined(MS_ASYNC) |
| 109 |
#define PG_FLUSH_DATA_WORKS 1 |
| 110 |
#elif defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_DONTNEED) |
| 111 |
#define PG_FLUSH_DATA_WORKS 1 |
| 112 |
#endif |
| 113 |
|
| 114 |
/* |
| 115 |
* We must leave some file descriptors free for system(), the dynamic loader, |
| 116 |
* and other code that tries to open files without consulting fd.c. This |
| 117 |
* is the number left free. (While we try fairly hard to prevent EMFILE |
| 118 |
* errors, there's never any guarantee that we won't get ENFILE due to |
| 119 |
* other processes chewing up FDs. So it's a bad idea to try to open files |
| 120 |
* without consulting fd.c. Nonetheless we cannot control all code.) |
| 121 |
* |
| 122 |
* Because this is just a fixed setting, we are effectively assuming that |
| 123 |
* no such code will leave FDs open over the long term; otherwise the slop |
| 124 |
* is likely to be insufficient. Note in particular that we expect that |
| 125 |
* loading a shared library does not result in any permanent increase in |
| 126 |
* the number of open files. (This appears to be true on most if not |
| 127 |
* all platforms as of Feb 2004.) |
| 128 |
*/ |
| 129 |
#define NUM_RESERVED_FDS 10 |
| 130 |
|
| 131 |
/* |
| 132 |
* If we have fewer than this many usable FDs after allowing for the reserved |
| 133 |
* ones, choke. (This value is chosen to work with "ulimit -n 64", but not |
| 134 |
* much less than that. Note that this value ensures numExternalFDs can be |
| 135 |
* at least 16; as of this writing, the contrib/postgres_fdw regression tests |
| 136 |
* will not pass unless that can grow to at least 14.) |
| 137 |
*/ |
| 138 |
#define FD_MINFREE 48 |
| 139 |
|
| 140 |
/* |
| 141 |
* A number of platforms allow individual processes to open many more files |
| 142 |
* than they can really support when *many* processes do the same thing. |
| 143 |
* This GUC parameter lets the DBA limit max_safe_fds to something less than |
| 144 |
* what the postmaster's initial probe suggests will work. |
| 145 |
*/ |
| 146 |
int max_files_per_process = 1000; |
| 147 |
|
| 148 |
/* |
| 149 |
* Maximum number of file descriptors to open for operations that fd.c knows |
| 150 |
* about (VFDs, AllocateFile etc, or "external" FDs). This is initialized |
| 151 |
* to a conservative value, and remains that way indefinitely in bootstrap or |
| 152 |
* standalone-backend cases. In normal postmaster operation, the postmaster |
| 153 |
* calls set_max_safe_fds() late in initialization to update the value, and |
| 154 |
* that value is then inherited by forked subprocesses. |
| 155 |
* |
| 156 |
* Note: the value of max_files_per_process is taken into account while |
| 157 |
* setting this variable, and so need not be tested separately. |
| 158 |
*/ |
| 159 |
int max_safe_fds = FD_MINFREE; /* default if not changed */ |
| 160 |
|
| 161 |
/* Whether it is safe to continue running after fsync() fails. */ |
| 162 |
bool data_sync_retry = false; |
| 163 |
|
| 164 |
/* How SyncDataDirectory() should do its job. */ |
| 165 |
int recovery_init_sync_method = DATA_DIR_SYNC_METHOD_FSYNC; |
| 166 |
|
| 167 |
/* How data files should be bulk-extended with zeros. */ |
| 168 |
int file_extend_method = DEFAULT_FILE_EXTEND_METHOD; |
| 169 |
|
| 170 |
/* Which kinds of files should be opened with PG_O_DIRECT. */ |
| 171 |
int io_direct_flags; |
| 172 |
|
| 173 |
/* Debugging.... */ |
| 174 |
|
| 175 |
#ifdef FDDEBUG |
| 176 |
#define DO_DB(A) \ |
| 177 |
do { \ |
| 178 |
int _do_db_save_errno = errno; \ |
| 179 |
A; \ |
| 180 |
errno = _do_db_save_errno; \ |
| 181 |
} while (0) |
| 182 |
#else |
| 183 |
#define DO_DB(A) \ |
| 184 |
((void) 0) |
| 185 |
#endif |
| 186 |
|
| 187 |
#define VFD_CLOSED (-1) |
| 188 |
|
| 189 |
#define FileIsValid(file) \ |
| 190 |
((file) > 0 && (file) < (int) SizeVfdCache && VfdCache[file].fileName != NULL) |
| 191 |
|
| 192 |
#define FileIsNotOpen(file) (VfdCache[file].fd == VFD_CLOSED) |
| 193 |
|
| 194 |
/* these are the assigned bits in fdstate below: */ |
| 195 |
#define FD_DELETE_AT_CLOSE (1 << 0) /* T = delete when closed */ |
| 196 |
#define FD_CLOSE_AT_EOXACT (1 << 1) /* T = close at eoXact */ |
| 197 |
#define FD_TEMP_FILE_LIMIT (1 << 2) /* T = respect temp_file_limit */ |
| 198 |
|
| 199 |
typedef struct vfd |
| 200 |
{ |
| 201 |
int fd; /* current FD, or VFD_CLOSED if none */ |
| 202 |
unsigned short fdstate; /* bitflags for VFD's state */ |
| 203 |
ResourceOwner resowner; /* owner, for automatic cleanup */ |
| 204 |
File nextFree; /* link to next free VFD, if in freelist */ |
| 205 |
File lruMoreRecently; /* doubly linked recency-of-use list */ |
| 206 |
File lruLessRecently; |
| 207 |
off_t fileSize; /* current size of file (0 if not temporary) */ |
| 208 |
char *fileName; /* name of file, or NULL for unused VFD */ |
| 209 |
/* NB: fileName is malloc'd, and must be free'd when closing the VFD */ |
| 210 |
int fileFlags; /* open(2) flags for (re)opening the file */ |
| 211 |
mode_t fileMode; /* mode to pass to open(2) */ |
| 212 |
} Vfd; |
| 213 |
|
| 214 |
/* |
| 215 |
* Virtual File Descriptor array pointer and size. This grows as |
| 216 |
* needed. 'File' values are indexes into this array. |
| 217 |
* Note that VfdCache[0] is not a usable VFD, just a list header. |
| 218 |
*/ |
| 219 |
static Vfd *VfdCache; |
| 220 |
static Size SizeVfdCache = 0; |
| 221 |
|
| 222 |
/* |
| 223 |
* Number of file descriptors known to be in use by VFD entries. |
| 224 |
*/ |
| 225 |
static int nfile = 0; |
| 226 |
|
| 227 |
/* |
| 228 |
* Flag to tell whether it's worth scanning VfdCache looking for temp files |
| 229 |
* to close |
| 230 |
*/ |
| 231 |
static bool have_xact_temporary_files = false; |
| 232 |
|
| 233 |
/* |
| 234 |
* Tracks the total size of all temporary files. Note: when temp_file_limit |
| 235 |
* is being enforced, this cannot overflow since the limit cannot be more |
| 236 |
* than INT_MAX kilobytes. When not enforcing, it could theoretically |
| 237 |
* overflow, but we don't care. |
| 238 |
*/ |
| 239 |
static uint64 temporary_files_size = 0; |
| 240 |
|
| 241 |
/* Temporary file access initialized and not yet shut down? */ |
| 242 |
#ifdef USE_ASSERT_CHECKING |
| 243 |
static bool temporary_files_allowed = false; |
| 244 |
#endif |
| 245 |
|
| 246 |
/* |
| 247 |
* List of OS handles opened with AllocateFile, AllocateDir and |
| 248 |
* OpenTransientFile. |
| 249 |
*/ |
| 250 |
typedef enum |
| 251 |
{ |
| 252 |
AllocateDescFile, |
| 253 |
AllocateDescPipe, |
| 254 |
AllocateDescDir, |
| 255 |
AllocateDescRawFD, |
| 256 |
} AllocateDescKind; |
| 257 |
|
| 258 |
typedef struct |
| 259 |
{ |
| 260 |
AllocateDescKind kind; |
| 261 |
SubTransactionId create_subid; |
| 262 |
union |
| 263 |
{ |
| 264 |
FILE *file; |
| 265 |
DIR *dir; |
| 266 |
int fd; |
| 267 |
} desc; |
| 268 |
} AllocateDesc; |
| 269 |
|
| 270 |
static int numAllocatedDescs = 0; |
| 271 |
static int maxAllocatedDescs = 0; |
| 272 |
static AllocateDesc *allocatedDescs = NULL; |
| 273 |
|
| 274 |
/* |
| 275 |
* Number of open "external" FDs reported to Reserve/ReleaseExternalFD. |
| 276 |
*/ |
| 277 |
static int numExternalFDs = 0; |
| 278 |
|
| 279 |
/* |
| 280 |
* Number of temporary files opened during the current session; |
| 281 |
* this is used in generation of tempfile names. |
| 282 |
*/ |
| 283 |
static long tempFileCounter = 0; |
| 284 |
|
| 285 |
/* |
| 286 |
* Array of OIDs of temp tablespaces. (Some entries may be InvalidOid, |
| 287 |
* indicating that the current database's default tablespace should be used.) |
| 288 |
* When numTempTableSpaces is -1, this has not been set in the current |
| 289 |
* transaction. |
| 290 |
*/ |
| 291 |
static Oid *tempTableSpaces = NULL; |
| 292 |
static int numTempTableSpaces = -1; |
| 293 |
static int nextTempTableSpace = 0; |
| 294 |
|
| 295 |
|
| 296 |
/*-------------------- |
| 297 |
* |
| 298 |
* Private Routines |
| 299 |
* |
| 300 |
* Delete - delete a file from the Lru ring |
| 301 |
* LruDelete - remove a file from the Lru ring and close its FD |
| 302 |
* Insert - put a file at the front of the Lru ring |
| 303 |
* LruInsert - put a file at the front of the Lru ring and open it |
| 304 |
* ReleaseLruFile - Release an fd by closing the last entry in the Lru ring |
| 305 |
* ReleaseLruFiles - Release fd(s) until we're under the max_safe_fds limit |
| 306 |
* AllocateVfd - grab a free (or new) file record (from VfdCache) |
| 307 |
* FreeVfd - free a file record |
| 308 |
* |
| 309 |
* The Least Recently Used ring is a doubly linked list that begins and |
| 310 |
* ends on element zero. Element zero is special -- it doesn't represent |
| 311 |
* a file and its "fd" field always == VFD_CLOSED. Element zero is just an |
| 312 |
* anchor that shows us the beginning/end of the ring. |
| 313 |
* Only VFD elements that are currently really open (have an FD assigned) are |
| 314 |
* in the Lru ring. Elements that are "virtually" open can be recognized |
| 315 |
* by having a non-null fileName field. |
| 316 |
* |
| 317 |
* example: |
| 318 |
* |
| 319 |
* /--less----\ /---------\ |
| 320 |
* v \ v \ |
| 321 |
* #0 --more---> LeastRecentlyUsed --more-\ \ |
| 322 |
* ^\ | | |
| 323 |
* \\less--> MostRecentlyUsedFile <---/ | |
| 324 |
* \more---/ \--less--/ |
| 325 |
* |
| 326 |
*-------------------- |
| 327 |
*/ |
| 328 |
static void Delete(File file); |
| 329 |
static void LruDelete(File file); |
| 330 |
static void Insert(File file); |
| 331 |
static int LruInsert(File file); |
| 332 |
static bool ReleaseLruFile(void); |
| 333 |
static void ReleaseLruFiles(void); |
| 334 |
static File AllocateVfd(void); |
| 335 |
static void FreeVfd(File file); |
| 336 |
|
| 337 |
static int FileAccess(File file); |
| 338 |
static File OpenTemporaryFileInTablespace(Oid tblspcOid, bool rejectError); |
| 339 |
static bool reserveAllocatedDesc(void); |
| 340 |
static int FreeDesc(AllocateDesc *desc); |
| 341 |
|
| 342 |
static void BeforeShmemExit_Files(int code, Datum arg); |
| 343 |
static void CleanupTempFiles(bool isCommit, bool isProcExit); |
| 344 |
static void RemovePgTempRelationFiles(const char *tsdirname); |
| 345 |
static void RemovePgTempRelationFilesInDbspace(const char *dbspacedirname); |
| 346 |
|
| 347 |
static void walkdir(const char *path, |
| 348 |
void (*action) (const char *fname, bool isdir, int elevel), |
| 349 |
bool process_symlinks, |
| 350 |
int elevel); |
| 351 |
#ifdef PG_FLUSH_DATA_WORKS |
| 352 |
static void pre_sync_fname(const char *fname, bool isdir, int elevel); |
| 353 |
#endif |
| 354 |
static void datadir_fsync_fname(const char *fname, bool isdir, int elevel); |
| 355 |
static void unlink_if_exists_fname(const char *fname, bool isdir, int elevel); |
| 356 |
|
| 357 |
static int fsync_parent_path(const char *fname, int elevel); |
| 358 |
|
| 359 |
|
| 360 |
/* ResourceOwner callbacks to hold virtual file descriptors */ |
| 361 |
static void ResOwnerReleaseFile(Datum res); |
| 362 |
static char *ResOwnerPrintFile(Datum res); |
| 363 |
|
| 364 |
static const ResourceOwnerDesc file_resowner_desc = |
| 365 |
{ |
| 366 |
.name = "File", |
| 367 |
.release_phase = RESOURCE_RELEASE_AFTER_LOCKS, |
| 368 |
.release_priority = RELEASE_PRIO_FILES, |
| 369 |
.ReleaseResource = ResOwnerReleaseFile, |
| 370 |
.DebugPrint = ResOwnerPrintFile |
| 371 |
}; |
| 372 |
|
| 373 |
/* Convenience wrappers over ResourceOwnerRemember/Forget */ |
| 374 |
static inline void |
| 375 |
ResourceOwnerRememberFile(ResourceOwner owner, File file) |
| 376 |
{ |
| 377 |
ResourceOwnerRemember(owner, Int32GetDatum(file), &file_resowner_desc); |
| 378 |
} |
| 379 |
static inline void |
| 380 |
ResourceOwnerForgetFile(ResourceOwner owner, File file) |
| 381 |
{ |
| 382 |
ResourceOwnerForget(owner, Int32GetDatum(file), &file_resowner_desc); |
| 383 |
} |
| 384 |
|
| 385 |
/* |
| 386 |
* pg_fsync --- do fsync with or without writethrough |
| 387 |
*/ |
| 388 |
int |
| 389 |
pg_fsync(int fd) |
| 390 |
{ |
| 391 |
#if !defined(WIN32) && defined(USE_ASSERT_CHECKING) |
| 392 |
struct stat st; |
| 393 |
|
| 394 |
/* |
| 395 |
* Some operating system implementations of fsync() have requirements |
| 396 |
* about the file access modes that were used when their file descriptor |
| 397 |
* argument was opened, and these requirements differ depending on whether |
| 398 |
* the file descriptor is for a directory. |
| 399 |
* |
| 400 |
* For any file descriptor that may eventually be handed to fsync(), we |
| 401 |
* should have opened it with access modes that are compatible with |
| 402 |
* fsync() on all supported systems, otherwise the code may not be |
| 403 |
* portable, even if it runs ok on the current system. |
| 404 |
* |
| 405 |
* We assert here that a descriptor for a file was opened with write |
| 406 |
* permissions (i.e., not O_RDONLY) and for a directory without write |
| 407 |
* permissions (O_RDONLY). Notice that the assertion check is made even |
| 408 |
* if fsync() is disabled. |
| 409 |
* |
| 410 |
* If fstat() fails, ignore it and let the follow-up fsync() complain. |
| 411 |
*/ |
| 412 |
if (fstat(fd, &st) == 0) |
| 413 |
{ |
| 414 |
int desc_flags = fcntl(fd, F_GETFL); |
| 415 |
|
| 416 |
desc_flags &= O_ACCMODE; |
| 417 |
|
| 418 |
if (S_ISDIR(st.st_mode)) |
| 419 |
Assert(desc_flags == O_RDONLY); |
| 420 |
else |
| 421 |
Assert(desc_flags != O_RDONLY); |
| 422 |
} |
| 423 |
errno = 0; |
| 424 |
#endif |
| 425 |
|
| 426 |
/* #if is to skip the wal_sync_method test if there's no need for it */ |
| 427 |
#if defined(HAVE_FSYNC_WRITETHROUGH) |
| 428 |
if (wal_sync_method == WAL_SYNC_METHOD_FSYNC_WRITETHROUGH) |
| 429 |
return pg_fsync_writethrough(fd); |
| 430 |
else |
| 431 |
#endif |
| 432 |
return pg_fsync_no_writethrough(fd); |
| 433 |
} |
| 434 |
|
| 435 |
|
| 436 |
/* |
| 437 |
* pg_fsync_no_writethrough --- same as fsync except does nothing if |
| 438 |
* enableFsync is off |
| 439 |
*/ |
| 440 |
int |
| 441 |
pg_fsync_no_writethrough(int fd) |
| 442 |
{ |
| 443 |
int rc; |
| 444 |
|
| 445 |
if (!enableFsync) |
| 446 |
return 0; |
| 447 |
|
| 448 |
retry: |
| 449 |
rc = fsync(fd); |
| 450 |
|
| 451 |
if (rc == -1 && errno == EINTR) |
| 452 |
goto retry; |
| 453 |
|
| 454 |
return rc; |
| 455 |
} |
| 456 |
|
| 457 |
/* |
| 458 |
* pg_fsync_writethrough |
| 459 |
*/ |
| 460 |
int |
| 461 |
pg_fsync_writethrough(int fd) |
| 462 |
{ |
| 463 |
if (enableFsync) |
| 464 |
{ |
| 465 |
#if defined(F_FULLFSYNC) |
| 466 |
return (fcntl(fd, F_FULLFSYNC, 0) == -1) ? -1 : 0; |
| 467 |
#else |
| 468 |
errno = ENOSYS; |
| 469 |
return -1; |
| 470 |
#endif |
| 471 |
} |
| 472 |
else |
| 473 |
return 0; |
| 474 |
} |
| 475 |
|
| 476 |
/* |
| 477 |
* pg_fdatasync --- same as fdatasync except does nothing if enableFsync is off |
| 478 |
*/ |
| 479 |
int |
| 480 |
pg_fdatasync(int fd) |
| 481 |
{ |
| 482 |
int rc; |
| 483 |
|
| 484 |
if (!enableFsync) |
| 485 |
return 0; |
| 486 |
|
| 487 |
retry: |
| 488 |
rc = fdatasync(fd); |
| 489 |
|
| 490 |
if (rc == -1 && errno == EINTR) |
| 491 |
goto retry; |
| 492 |
|
| 493 |
return rc; |
| 494 |
} |
| 495 |
|
| 496 |
/* |
| 497 |
* pg_file_exists -- check that a file exists. |
| 498 |
* |
| 499 |
* This requires an absolute path to the file. Returns true if the file is |
| 500 |
* not a directory, false otherwise. |
| 501 |
*/ |
| 502 |
bool |
| 503 |
pg_file_exists(const char *name) |
| 504 |
{ |
| 505 |
struct stat st; |
| 506 |
|
| 507 |
Assert(name != NULL); |
| 508 |
|
| 509 |
if (stat(name, &st) == 0) |
| 510 |
return !S_ISDIR(st.st_mode); |
| 511 |
else if (!(errno == ENOENT || errno == ENOTDIR || errno == EACCES)) |
| 512 |
ereport(ERROR, |
| 513 |
(errcode_for_file_access(), |
| 514 |
errmsg("could not access file \"%s\": %m", name))); |
| 515 |
|
| 516 |
return false; |
| 517 |
} |
| 518 |
|
| 519 |
/* |
| 520 |
* pg_flush_data --- advise OS that the described dirty data should be flushed |
| 521 |
* |
| 522 |
* offset of 0 with nbytes 0 means that the entire file should be flushed |
| 523 |
*/ |
| 524 |
void |
| 525 |
pg_flush_data(int fd, off_t offset, off_t nbytes) |
| 526 |
{ |
| 527 |
/* |
| 528 |
* Right now file flushing is primarily used to avoid making later |
| 529 |
* fsync()/fdatasync() calls have less impact. Thus don't trigger flushes |
| 530 |
* if fsyncs are disabled - that's a decision we might want to make |
| 531 |
* configurable at some point. |
| 532 |
*/ |
| 533 |
if (!enableFsync) |
| 534 |
return; |
| 535 |
|
| 536 |
/* |
| 537 |
* We compile all alternatives that are supported on the current platform, |
| 538 |
* to find portability problems more easily. |
| 539 |
*/ |
| 540 |
#if defined(HAVE_SYNC_FILE_RANGE) |
| 541 |
{ |
| 542 |
int rc; |
| 543 |
static bool not_implemented_by_kernel = false; |
| 544 |
|
| 545 |
if (not_implemented_by_kernel) |
| 546 |
return; |
| 547 |
|
| 548 |
retry: |
| 549 |
|
| 550 |
/* |
| 551 |
* sync_file_range(SYNC_FILE_RANGE_WRITE), currently linux specific, |
| 552 |
* tells the OS that writeback for the specified blocks should be |
| 553 |
* started, but that we don't want to wait for completion. Note that |
| 554 |
* this call might block if too much dirty data exists in the range. |
| 555 |
* This is the preferable method on OSs supporting it, as it works |
| 556 |
* reliably when available (contrast to msync()) and doesn't flush out |
| 557 |
* clean data (like FADV_DONTNEED). |
| 558 |
*/ |
| 559 |
rc = sync_file_range(fd, offset, nbytes, |
| 560 |
SYNC_FILE_RANGE_WRITE); |
| 561 |
if (rc != 0) |
| 562 |
{ |
| 563 |
int elevel; |
| 564 |
|
| 565 |
if (errno == EINTR) |
| 566 |
goto retry; |
| 567 |
|
| 568 |
/* |
| 569 |
* For systems that don't have an implementation of |
| 570 |
* sync_file_range() such as Windows WSL, generate only one |
| 571 |
* warning and then suppress all further attempts by this process. |
| 572 |
*/ |
| 573 |
if (errno == ENOSYS) |
| 574 |
{ |
| 575 |
elevel = WARNING; |
| 576 |
not_implemented_by_kernel = true; |
| 577 |
} |
| 578 |
else |
| 579 |
elevel = data_sync_elevel(WARNING); |
| 580 |
|
| 581 |
ereport(elevel, |
| 582 |
(errcode_for_file_access(), |
| 583 |
errmsg("could not flush dirty data: %m"))); |
| 584 |
} |
| 585 |
|
| 586 |
return; |
| 587 |
} |
| 588 |
#endif |
| 589 |
#if !defined(WIN32) && defined(MS_ASYNC) |
| 590 |
{ |
| 591 |
void *p; |
| 592 |
static int pagesize = 0; |
| 593 |
|
| 594 |
/* |
| 595 |
* On several OSs msync(MS_ASYNC) on a mmap'ed file triggers |
| 596 |
* writeback. On linux it only does so if MS_SYNC is specified, but |
| 597 |
* then it does the writeback synchronously. Luckily all common linux |
| 598 |
* systems have sync_file_range(). This is preferable over |
| 599 |
* FADV_DONTNEED because it doesn't flush out clean data. |
| 600 |
* |
| 601 |
* We map the file (mmap()), tell the kernel to sync back the contents |
| 602 |
* (msync()), and then remove the mapping again (munmap()). |
| 603 |
*/ |
| 604 |
|
| 605 |
/* mmap() needs actual length if we want to map whole file */ |
| 606 |
if (offset == 0 && nbytes == 0) |
| 607 |
{ |
| 608 |
nbytes = lseek(fd, 0, SEEK_END); |
| 609 |
if (nbytes < 0) |
| 610 |
{ |
| 611 |
ereport(WARNING, |
| 612 |
(errcode_for_file_access(), |
| 613 |
errmsg("could not determine dirty data size: %m"))); |
| 614 |
return; |
| 615 |
} |
| 616 |
} |
| 617 |
|
| 618 |
/* |
| 619 |
* Some platforms reject partial-page mmap() attempts. To deal with |
| 620 |
* that, just truncate the request to a page boundary. If any extra |
| 621 |
* bytes don't get flushed, well, it's only a hint anyway. |
| 622 |
*/ |
| 623 |
|
| 624 |
/* fetch pagesize only once */ |
| 625 |
if (pagesize == 0) |
| 626 |
pagesize = sysconf(_SC_PAGESIZE); |
| 627 |
|
| 628 |
/* align length to pagesize, dropping any fractional page */ |
| 629 |
if (pagesize > 0) |
| 630 |
nbytes = (nbytes / pagesize) * pagesize; |
| 631 |
|
| 632 |
/* fractional-page request is a no-op */ |
| 633 |
if (nbytes <= 0) |
| 634 |
return; |
| 635 |
|
| 636 |
/* |
| 637 |
* mmap could well fail, particularly on 32-bit platforms where there |
| 638 |
* may simply not be enough address space. If so, silently fall |
| 639 |
* through to the next implementation. |
| 640 |
*/ |
| 641 |
if (nbytes <= (off_t) SSIZE_MAX) |
| 642 |
p = mmap(NULL, nbytes, PROT_READ, MAP_SHARED, fd, offset); |
| 643 |
else |
| 644 |
p = MAP_FAILED; |
| 645 |
|
| 646 |
if (p != MAP_FAILED) |
| 647 |
{ |
| 648 |
int rc; |
| 649 |
|
| 650 |
rc = msync(p, (size_t) nbytes, MS_ASYNC); |
| 651 |
if (rc != 0) |
| 652 |
{ |
| 653 |
ereport(data_sync_elevel(WARNING), |
| 654 |
(errcode_for_file_access(), |
| 655 |
errmsg("could not flush dirty data: %m"))); |
| 656 |
/* NB: need to fall through to munmap()! */ |
| 657 |
} |
| 658 |
|
| 659 |
rc = munmap(p, (size_t) nbytes); |
| 660 |
if (rc != 0) |
| 661 |
{ |
| 662 |
/* FATAL error because mapping would remain */ |
| 663 |
ereport(FATAL, |
| 664 |
(errcode_for_file_access(), |
| 665 |
errmsg("could not munmap() while flushing data: %m"))); |
| 666 |
} |
| 667 |
|
| 668 |
return; |
| 669 |
} |
| 670 |
} |
| 671 |
#endif |
| 672 |
#if defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_DONTNEED) |
| 673 |
{ |
| 674 |
int rc; |
| 675 |
|
| 676 |
/* |
| 677 |
* Signal the kernel that the passed in range should not be cached |
| 678 |
* anymore. This has the, desired, side effect of writing out dirty |
| 679 |
* data, and the, undesired, side effect of likely discarding useful |
| 680 |
* clean cached blocks. For the latter reason this is the least |
| 681 |
* preferable method. |
| 682 |
*/ |
| 683 |
|
| 684 |
rc = posix_fadvise(fd, offset, nbytes, POSIX_FADV_DONTNEED); |
| 685 |
|
| 686 |
if (rc != 0) |
| 687 |
{ |
| 688 |
/* don't error out, this is just a performance optimization */ |
| 689 |
ereport(WARNING, |
| 690 |
(errcode_for_file_access(), |
| 691 |
errmsg("could not flush dirty data: %m"))); |
| 692 |
} |
| 693 |
|
| 694 |
return; |
| 695 |
} |
| 696 |
#endif |
| 697 |
} |
| 698 |
|
| 699 |
/* |
| 700 |
* Truncate an open file to a given length. |
| 701 |
*/ |
| 702 |
static int |
| 703 |
pg_ftruncate(int fd, off_t length) |
| 704 |
{ |
| 705 |
int ret; |
| 706 |
|
| 707 |
retry: |
| 708 |
ret = ftruncate(fd, length); |
| 709 |
|
| 710 |
if (ret == -1 && errno == EINTR) |
| 711 |
goto retry; |
| 712 |
|
| 713 |
return ret; |
| 714 |
} |
| 715 |
|
| 716 |
/* |
| 717 |
* Truncate a file to a given length by name. |
| 718 |
*/ |
| 719 |
int |
| 720 |
pg_truncate(const char *path, off_t length) |
| 721 |
{ |
| 722 |
int ret; |
| 723 |
#ifdef WIN32 |
| 724 |
int save_errno; |
| 725 |
int fd; |
| 726 |
|
| 727 |
fd = OpenTransientFile(path, O_RDWR | PG_BINARY); |
| 728 |
if (fd >= 0) |
| 729 |
{ |
| 730 |
ret = pg_ftruncate(fd, length); |
| 731 |
save_errno = errno; |
| 732 |
CloseTransientFile(fd); |
| 733 |
errno = save_errno; |
| 734 |
} |
| 735 |
else |
| 736 |
ret = -1; |
| 737 |
#else |
| 738 |
|
| 739 |
retry: |
| 740 |
ret = truncate(path, length); |
| 741 |
|
| 742 |
if (ret == -1 && errno == EINTR) |
| 743 |
goto retry; |
| 744 |
#endif |
| 745 |
|
| 746 |
return ret; |
| 747 |
} |
| 748 |
|
| 749 |
/* |
| 750 |
* fsync_fname -- fsync a file or directory, handling errors properly |
| 751 |
* |
| 752 |
* Try to fsync a file or directory. When doing the latter, ignore errors that |
| 753 |
* indicate the OS just doesn't allow/require fsyncing directories. |
| 754 |
*/ |
| 755 |
void |
| 756 |
fsync_fname(const char *fname, bool isdir) |
| 757 |
{ |
| 758 |
fsync_fname_ext(fname, isdir, false, data_sync_elevel(ERROR)); |
| 759 |
} |
| 760 |
|
| 761 |
/* |
| 762 |
* durable_rename -- rename(2) wrapper, issuing fsyncs required for durability |
| 763 |
* |
| 764 |
* This routine ensures that, after returning, the effect of renaming file |
| 765 |
* persists in case of a crash. A crash while this routine is running will |
| 766 |
* leave you with either the pre-existing or the moved file in place of the |
| 767 |
* new file; no mixed state or truncated files are possible. |
| 768 |
* |
| 769 |
* It does so by using fsync on the old filename and the possibly existing |
| 770 |
* target filename before the rename, and the target file and directory after. |
| 771 |
* |
| 772 |
* Note that rename() cannot be used across arbitrary directories, as they |
| 773 |
* might not be on the same filesystem. Therefore this routine does not |
| 774 |
* support renaming across directories. |
| 775 |
* |
| 776 |
* Log errors with the caller specified severity. |
| 777 |
* |
| 778 |
* Returns 0 if the operation succeeded, -1 otherwise. Note that errno is not |
| 779 |
* valid upon return. |
| 780 |
*/ |
| 781 |
int |
| 782 |
durable_rename(const char *oldfile, const char *newfile, int elevel) |
| 783 |
{ |
| 784 |
int fd; |
| 785 |
|
| 786 |
/* |
| 787 |
* First fsync the old and target path (if it exists), to ensure that they |
| 788 |
* are properly persistent on disk. Syncing the target file is not |
| 789 |
* strictly necessary, but it makes it easier to reason about crashes; |
| 790 |
* because it's then guaranteed that either source or target file exists |
| 791 |
* after a crash. |
| 792 |
*/ |
| 793 |
if (fsync_fname_ext(oldfile, false, false, elevel) != 0) |
| 794 |
return -1; |
| 795 |
|
| 796 |
fd = OpenTransientFile(newfile, PG_BINARY | O_RDWR); |
| 797 |
if (fd < 0) |
| 798 |
{ |
| 799 |
if (errno != ENOENT) |
| 800 |
{ |
| 801 |
ereport(elevel, |
| 802 |
(errcode_for_file_access(), |
| 803 |
errmsg("could not open file \"%s\": %m", newfile))); |
| 804 |
return -1; |
| 805 |
} |
| 806 |
} |
| 807 |
else |
| 808 |
{ |
| 809 |
if (pg_fsync(fd) != 0) |
| 810 |
{ |
| 811 |
int save_errno; |
| 812 |
|
| 813 |
/* close file upon error, might not be in transaction context */ |
| 814 |
save_errno = errno; |
| 815 |
CloseTransientFile(fd); |
| 816 |
errno = save_errno; |
| 817 |
|
| 818 |
ereport(elevel, |
| 819 |
(errcode_for_file_access(), |
| 820 |
errmsg("could not fsync file \"%s\": %m", newfile))); |
| 821 |
return -1; |
| 822 |
} |
| 823 |
|
| 824 |
if (CloseTransientFile(fd) != 0) |
| 825 |
{ |
| 826 |
ereport(elevel, |
| 827 |
(errcode_for_file_access(), |
| 828 |
errmsg("could not close file \"%s\": %m", newfile))); |
| 829 |
return -1; |
| 830 |
} |
| 831 |
} |
| 832 |
|
| 833 |
/* Time to do the real deal... */ |
| 834 |
if (rename(oldfile, newfile) < 0) |
| 835 |
{ |
| 836 |
ereport(elevel, |
| 837 |
(errcode_for_file_access(), |
| 838 |
errmsg("could not rename file \"%s\" to \"%s\": %m", |
| 839 |
oldfile, newfile))); |
| 840 |
return -1; |
| 841 |
} |
| 842 |
|
| 843 |
/* |
| 844 |
* To guarantee renaming the file is persistent, fsync the file with its |
| 845 |
* new name, and its containing directory. |
| 846 |
*/ |
| 847 |
if (fsync_fname_ext(newfile, false, false, elevel) != 0) |
| 848 |
return -1; |
| 849 |
|
| 850 |
if (fsync_parent_path(newfile, elevel) != 0) |
| 851 |
return -1; |
| 852 |
|
| 853 |
return 0; |
| 854 |
} |
| 855 |
|
| 856 |
/* |
| 857 |
* durable_unlink -- remove a file in a durable manner |
| 858 |
* |
| 859 |
* This routine ensures that, after returning, the effect of removing file |
| 860 |
* persists in case of a crash. A crash while this routine is running will |
| 861 |
* leave the system in no mixed state. |
| 862 |
* |
| 863 |
* It does so by using fsync on the parent directory of the file after the |
| 864 |
* actual removal is done. |
| 865 |
* |
| 866 |
* Log errors with the severity specified by caller. |
| 867 |
* |
| 868 |
* Returns 0 if the operation succeeded, -1 otherwise. Note that errno is not |
| 869 |
* valid upon return. |
| 870 |
*/ |
| 871 |
int |
| 872 |
durable_unlink(const char *fname, int elevel) |
| 873 |
{ |
| 874 |
if (unlink(fname) < 0) |
| 875 |
{ |
| 876 |
ereport(elevel, |
| 877 |
(errcode_for_file_access(), |
| 878 |
errmsg("could not remove file \"%s\": %m", |
| 879 |
fname))); |
| 880 |
return -1; |
| 881 |
} |
| 882 |
|
| 883 |
/* |
| 884 |
* To guarantee that the removal of the file is persistent, fsync its |
| 885 |
* parent directory. |
| 886 |
*/ |
| 887 |
if (fsync_parent_path(fname, elevel) != 0) |
| 888 |
return -1; |
| 889 |
|
| 890 |
return 0; |
| 891 |
} |
| 892 |
|
| 893 |
/* |
| 894 |
* InitFileAccess --- initialize this module during backend startup |
| 895 |
* |
| 896 |
* This is called during either normal or standalone backend start. |
| 897 |
* It is *not* called in the postmaster. |
| 898 |
* |
| 899 |
* Note that this does not initialize temporary file access, that is |
| 900 |
* separately initialized via InitTemporaryFileAccess(). |
| 901 |
*/ |
| 902 |
void |
| 903 |
InitFileAccess(void) |
| 904 |
{ |
| 905 |
Assert(SizeVfdCache == 0); /* call me only once */ |
| 906 |
|
| 907 |
/* initialize cache header entry */ |
| 908 |
VfdCache = (Vfd *) malloc(sizeof(Vfd)); |
| 909 |
if (VfdCache == NULL) |
| 910 |
ereport(FATAL, |
| 911 |
(errcode(ERRCODE_OUT_OF_MEMORY), |
| 912 |
errmsg("out of memory"))); |
| 913 |
|
| 914 |
MemSet(&(VfdCache[0]), 0, sizeof(Vfd)); |
| 915 |
VfdCache->fd = VFD_CLOSED; |
| 916 |
|
| 917 |
SizeVfdCache = 1; |
| 918 |
} |
| 919 |
|
| 920 |
/* |
| 921 |
* InitTemporaryFileAccess --- initialize temporary file access during startup |
| 922 |
* |
| 923 |
* This is called during either normal or standalone backend start. |
| 924 |
* It is *not* called in the postmaster. |
| 925 |
* |
| 926 |
* This is separate from InitFileAccess() because temporary file cleanup can |
| 927 |
* cause pgstat reporting. As pgstat is shut down during before_shmem_exit(), |
| 928 |
* our reporting has to happen before that. Low level file access should be |
| 929 |
* available for longer, hence the separate initialization / shutdown of |
| 930 |
* temporary file handling. |
| 931 |
*/ |
| 932 |
void |
| 933 |
InitTemporaryFileAccess(void) |
| 934 |
{ |
| 935 |
Assert(SizeVfdCache != 0); /* InitFileAccess() needs to have run */ |
| 936 |
Assert(!temporary_files_allowed); /* call me only once */ |
| 937 |
|
| 938 |
/* |
| 939 |
* Register before-shmem-exit hook to ensure temp files are dropped while |
| 940 |
* we can still report stats. |
| 941 |
*/ |
| 942 |
before_shmem_exit(BeforeShmemExit_Files, 0); |
| 943 |
|
| 944 |
#ifdef USE_ASSERT_CHECKING |
| 945 |
temporary_files_allowed = true; |
| 946 |
#endif |
| 947 |
} |
| 948 |
|
| 949 |
/* |
| 950 |
* count_usable_fds --- count how many FDs the system will let us open, |
| 951 |
* and estimate how many are already open. |
| 952 |
* |
| 953 |
* We stop counting if usable_fds reaches max_to_probe. Note: a small |
| 954 |
* value of max_to_probe might result in an underestimate of already_open; |
| 955 |
* we must fill in any "gaps" in the set of used FDs before the calculation |
| 956 |
* of already_open will give the right answer. In practice, max_to_probe |
| 957 |
* of a couple of dozen should be enough to ensure good results. |
| 958 |
* |
| 959 |
* We assume stderr (FD 2) is available for dup'ing. While the calling |
| 960 |
* script could theoretically close that, it would be a really bad idea, |
| 961 |
* since then one risks loss of error messages from, e.g., libc. |
| 962 |
*/ |
| 963 |
static void |
| 964 |
count_usable_fds(int max_to_probe, int *usable_fds, int *already_open) |
| 965 |
{ |
| 966 |
int *fd; |
| 967 |
int size; |
| 968 |
int used = 0; |
| 969 |
int highestfd = 0; |
| 970 |
int j; |
| 971 |
|
| 972 |
#ifdef HAVE_GETRLIMIT |
| 973 |
struct rlimit rlim; |
| 974 |
int getrlimit_status; |
| 975 |
#endif |
| 976 |
|
| 977 |
size = 1024; |
| 978 |
fd = (int *) palloc(size * sizeof(int)); |
| 979 |
|
| 980 |
#ifdef HAVE_GETRLIMIT |
| 981 |
getrlimit_status = getrlimit(RLIMIT_NOFILE, &rlim); |
| 982 |
if (getrlimit_status != 0) |
| 983 |
ereport(WARNING, (errmsg("getrlimit failed: %m"))); |
| 984 |
#endif /* HAVE_GETRLIMIT */ |
| 985 |
|
| 986 |
/* dup until failure or probe limit reached */ |
| 987 |
for (;;) |
| 988 |
{ |
| 989 |
int thisfd; |
| 990 |
|
| 991 |
#ifdef HAVE_GETRLIMIT |
| 992 |
|
| 993 |
/* |
| 994 |
* don't go beyond RLIMIT_NOFILE; causes irritating kernel logs on |
| 995 |
* some platforms |
| 996 |
*/ |
| 997 |
if (getrlimit_status == 0 && highestfd >= rlim.rlim_cur - 1) |
| 998 |
break; |
| 999 |
#endif |
| 1000 |
|
| 1001 |
thisfd = dup(2); |
| 1002 |
if (thisfd < 0) |
| 1003 |
{ |
| 1004 |
/* Expect EMFILE or ENFILE, else it's fishy */ |
| 1005 |
if (errno != EMFILE && errno != ENFILE) |
| 1006 |
elog(WARNING, "duplicating stderr file descriptor failed after %d successes: %m", used); |
| 1007 |
break; |
| 1008 |
} |
| 1009 |
|
| 1010 |
if (used >= size) |
| 1011 |
{ |
| 1012 |
size *= 2; |
| 1013 |
fd = (int *) repalloc(fd, size * sizeof(int)); |
| 1014 |
} |
| 1015 |
fd[used++] = thisfd; |
| 1016 |
|
| 1017 |
if (highestfd < thisfd) |
| 1018 |
highestfd = thisfd; |
| 1019 |
|
| 1020 |
if (used >= max_to_probe) |
| 1021 |
break; |
| 1022 |
} |
| 1023 |
|
| 1024 |
/* release the files we opened */ |
| 1025 |
for (j = 0; j < used; j++) |
| 1026 |
close(fd[j]); |
| 1027 |
|
| 1028 |
pfree(fd); |
| 1029 |
|
| 1030 |
/* |
| 1031 |
* Return results. usable_fds is just the number of successful dups. We |
| 1032 |
* assume that the system limit is highestfd+1 (remember 0 is a legal FD |
| 1033 |
* number) and so already_open is highestfd+1 - usable_fds. |
| 1034 |
*/ |
| 1035 |
*usable_fds = used; |
| 1036 |
*already_open = highestfd + 1 - used; |
| 1037 |
} |
| 1038 |
|
| 1039 |
/* |
| 1040 |
* set_max_safe_fds |
| 1041 |
* Determine number of file descriptors that fd.c is allowed to use |
| 1042 |
*/ |
| 1043 |
void |
| 1044 |
set_max_safe_fds(void) |
| 1045 |
{ |
| 1046 |
int usable_fds; |
| 1047 |
int already_open; |
| 1048 |
|
| 1049 |
/*---------- |
| 1050 |
* We want to set max_safe_fds to |
| 1051 |
* MIN(usable_fds, max_files_per_process) |
| 1052 |
* less the slop factor for files that are opened without consulting |
| 1053 |
* fd.c. This ensures that we won't allow to open more than |
| 1054 |
* max_files_per_process, or the experimentally-determined EMFILE limit, |
| 1055 |
* additional files. |
| 1056 |
*---------- |
| 1057 |
*/ |
| 1058 |
count_usable_fds(max_files_per_process, |
| 1059 |
&usable_fds, &already_open); |
| 1060 |
|
| 1061 |
max_safe_fds = Min(usable_fds, max_files_per_process); |
| 1062 |
|
| 1063 |
/* |
| 1064 |
* Take off the FDs reserved for system() etc. |
| 1065 |
*/ |
| 1066 |
max_safe_fds -= NUM_RESERVED_FDS; |
| 1067 |
|
| 1068 |
/* |
| 1069 |
* Make sure we still have enough to get by. |
| 1070 |
*/ |
| 1071 |
if (max_safe_fds < FD_MINFREE) |
| 1072 |
ereport(FATAL, |
| 1073 |
(errcode(ERRCODE_INSUFFICIENT_RESOURCES), |
| 1074 |
errmsg("insufficient file descriptors available to start server process"), |
| 1075 |
errdetail("System allows %d, server needs at least %d, %d files are already open.", |
| 1076 |
max_safe_fds + NUM_RESERVED_FDS, |
| 1077 |
FD_MINFREE + NUM_RESERVED_FDS, |
| 1078 |
already_open))); |
| 1079 |
|
| 1080 |
elog(DEBUG2, "max_safe_fds = %d, usable_fds = %d, already_open = %d", |
| 1081 |
max_safe_fds, usable_fds, already_open); |
| 1082 |
} |
| 1083 |
|
| 1084 |
/* |
| 1085 |
* Open a file with BasicOpenFilePerm() and pass default file mode for the |
| 1086 |
* fileMode parameter. |
| 1087 |
*/ |
| 1088 |
int |
| 1089 |
BasicOpenFile(const char *fileName, int fileFlags) |
| 1090 |
{ |
| 1091 |
return BasicOpenFilePerm(fileName, fileFlags, pg_file_create_mode); |
| 1092 |
} |
| 1093 |
|
| 1094 |
/* |
| 1095 |
* BasicOpenFilePerm --- same as open(2) except can free other FDs if needed |
| 1096 |
* |
| 1097 |
* This is exported for use by places that really want a plain kernel FD, |
| 1098 |
* but need to be proof against running out of FDs. Once an FD has been |
| 1099 |
* successfully returned, it is the caller's responsibility to ensure that |
| 1100 |
* it will not be leaked on ereport()! Most users should *not* call this |
| 1101 |
* routine directly, but instead use the VFD abstraction level, which |
| 1102 |
* provides protection against descriptor leaks as well as management of |
| 1103 |
* files that need to be open for more than a short period of time. |
| 1104 |
* |
| 1105 |
* Ideally this should be the *only* direct call of open() in the backend. |
| 1106 |
* In practice, the postmaster calls open() directly, and there are some |
| 1107 |
* direct open() calls done early in backend startup. Those are OK since |
| 1108 |
* this module wouldn't have any open files to close at that point anyway. |
| 1109 |
*/ |
| 1110 |
int |
| 1111 |
BasicOpenFilePerm(const char *fileName, int fileFlags, mode_t fileMode) |
| 1112 |
{ |
| 1113 |
int fd; |
| 1114 |
|
| 1115 |
tryAgain: |
| 1116 |
#ifdef PG_O_DIRECT_USE_F_NOCACHE |
| 1117 |
|
| 1118 |
/* |
| 1119 |
* The value we defined to stand in for O_DIRECT when simulating it with |
| 1120 |
* F_NOCACHE had better not collide with any of the standard flags. |
| 1121 |
*/ |
| 1122 |
StaticAssertStmt((PG_O_DIRECT & |
| 1123 |
(O_APPEND | |
| 1124 |
O_CLOEXEC | |
| 1125 |
O_CREAT | |
| 1126 |
O_DSYNC | |
| 1127 |
O_EXCL | |
| 1128 |
O_RDWR | |
| 1129 |
O_RDONLY | |
| 1130 |
O_SYNC | |
| 1131 |
O_TRUNC | |
| 1132 |
O_WRONLY)) == 0, |
| 1133 |
"PG_O_DIRECT value collides with standard flag"); |
| 1134 |
fd = open(fileName, fileFlags & ~PG_O_DIRECT, fileMode); |
| 1135 |
#else |
| 1136 |
fd = open(fileName, fileFlags, fileMode); |
| 1137 |
#endif |
| 1138 |
|
| 1139 |
if (fd >= 0) |
| 1140 |
{ |
| 1141 |
#ifdef PG_O_DIRECT_USE_F_NOCACHE |
| 1142 |
if (fileFlags & PG_O_DIRECT) |
| 1143 |
{ |
| 1144 |
if (fcntl(fd, F_NOCACHE, 1) < 0) |
| 1145 |
{ |
| 1146 |
int save_errno = errno; |
| 1147 |
|
| 1148 |
close(fd); |
| 1149 |
errno = save_errno; |
| 1150 |
return -1; |
| 1151 |
} |
| 1152 |
} |
| 1153 |
#endif |
| 1154 |
|
| 1155 |
return fd; /* success! */ |
| 1156 |
} |
| 1157 |
|
| 1158 |
if (errno == EMFILE || errno == ENFILE) |
| 1159 |
{ |
| 1160 |
int save_errno = errno; |
| 1161 |
|
| 1162 |
ereport(LOG, |
| 1163 |
(errcode(ERRCODE_INSUFFICIENT_RESOURCES), |
| 1164 |
errmsg("out of file descriptors: %m; release and retry"))); |
| 1165 |
errno = 0; |
| 1166 |
if (ReleaseLruFile()) |
| 1167 |
goto tryAgain; |
| 1168 |
errno = save_errno; |
| 1169 |
} |
| 1170 |
|
| 1171 |
return -1; /* failure */ |
| 1172 |
} |
| 1173 |
|
| 1174 |
/* |
| 1175 |
* AcquireExternalFD - attempt to reserve an external file descriptor |
| 1176 |
* |
| 1177 |
* This should be used by callers that need to hold a file descriptor open |
| 1178 |
* over more than a short interval, but cannot use any of the other facilities |
| 1179 |
* provided by this module. |
| 1180 |
* |
| 1181 |
* The difference between this and the underlying ReserveExternalFD function |
| 1182 |
* is that this will report failure (by setting errno and returning false) |
| 1183 |
* if "too many" external FDs are already reserved. This should be used in |
| 1184 |
* any code where the total number of FDs to be reserved is not predictable |
| 1185 |
* and small. |
| 1186 |
*/ |
| 1187 |
bool |
| 1188 |
AcquireExternalFD(void) |
| 1189 |
{ |
| 1190 |
/* |
| 1191 |
* We don't want more than max_safe_fds / 3 FDs to be consumed for |
| 1192 |
* "external" FDs. |
| 1193 |
*/ |
| 1194 |
if (numExternalFDs < max_safe_fds / 3) |
| 1195 |
{ |
| 1196 |
ReserveExternalFD(); |
| 1197 |
return true; |
| 1198 |
} |
| 1199 |
errno = EMFILE; |
| 1200 |
return false; |
| 1201 |
} |
| 1202 |
|
| 1203 |
/* |
| 1204 |
* ReserveExternalFD - report external consumption of a file descriptor |
| 1205 |
* |
| 1206 |
* This should be used by callers that need to hold a file descriptor open |
| 1207 |
* over more than a short interval, but cannot use any of the other facilities |
| 1208 |
* provided by this module. This just tracks the use of the FD and closes |
| 1209 |
* VFDs if needed to ensure we keep NUM_RESERVED_FDS FDs available. |
| 1210 |
* |
| 1211 |
* Call this directly only in code where failure to reserve the FD would be |
| 1212 |
* fatal; for example, the WAL-writing code does so, since the alternative is |
| 1213 |
* session failure. Also, it's very unwise to do so in code that could |
| 1214 |
* consume more than one FD per process. |
| 1215 |
* |
| 1216 |
* Note: as long as everybody plays nice so that NUM_RESERVED_FDS FDs remain |
| 1217 |
* available, it doesn't matter too much whether this is called before or |
| 1218 |
* after actually opening the FD; but doing so beforehand reduces the risk of |
| 1219 |
* an EMFILE failure if not everybody played nice. In any case, it's solely |
| 1220 |
* caller's responsibility to keep the external-FD count in sync with reality. |
| 1221 |
*/ |
| 1222 |
void |
| 1223 |
ReserveExternalFD(void) |
| 1224 |
{ |
| 1225 |
/* |
| 1226 |
* Release VFDs if needed to stay safe. Because we do this before |
| 1227 |
* incrementing numExternalFDs, the final state will be as desired, i.e., |
| 1228 |
* nfile + numAllocatedDescs + numExternalFDs <= max_safe_fds. |
| 1229 |
*/ |
| 1230 |
ReleaseLruFiles(); |
| 1231 |
|
| 1232 |
numExternalFDs++; |
| 1233 |
} |
| 1234 |
|
| 1235 |
/* |
| 1236 |
* ReleaseExternalFD - report release of an external file descriptor |
| 1237 |
* |
| 1238 |
* This is guaranteed not to change errno, so it can be used in failure paths. |
| 1239 |
*/ |
| 1240 |
void |
| 1241 |
ReleaseExternalFD(void) |
| 1242 |
{ |
| 1243 |
Assert(numExternalFDs > 0); |
| 1244 |
numExternalFDs--; |
| 1245 |
} |
| 1246 |
|
| 1247 |
|
| 1248 |
#if defined(FDDEBUG) |
| 1249 |
|
| 1250 |
static void |
| 1251 |
_dump_lru(void) |
| 1252 |
{ |
| 1253 |
int mru = VfdCache[0].lruLessRecently; |
| 1254 |
Vfd *vfdP = &VfdCache[mru]; |
| 1255 |
char buf[2048]; |
| 1256 |
|
| 1257 |
snprintf(buf, sizeof(buf), "LRU: MOST %d ", mru); |
| 1258 |
while (mru != 0) |
| 1259 |
{ |
| 1260 |
mru = vfdP->lruLessRecently; |
| 1261 |
vfdP = &VfdCache[mru]; |
| 1262 |
snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), "%d ", mru); |
| 1263 |
} |
| 1264 |
snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), "LEAST"); |
| 1265 |
elog(LOG, "%s", buf); |
| 1266 |
} |
| 1267 |
#endif /* FDDEBUG */ |
| 1268 |
|
| 1269 |
static void |
| 1270 |
Delete(File file) |
| 1271 |
{ |
| 1272 |
Vfd *vfdP; |
| 1273 |
|
| 1274 |
Assert(file != 0); |
| 1275 |
|
| 1276 |
DO_DB(elog(LOG, "Delete %d (%s)", |
| 1277 |
file, VfdCache[file].fileName)); |
| 1278 |
DO_DB(_dump_lru()); |
| 1279 |
|
| 1280 |
vfdP = &VfdCache[file]; |
| 1281 |
|
| 1282 |
VfdCache[vfdP->lruLessRecently].lruMoreRecently = vfdP->lruMoreRecently; |
| 1283 |
VfdCache[vfdP->lruMoreRecently].lruLessRecently = vfdP->lruLessRecently; |
| 1284 |
|
| 1285 |
DO_DB(_dump_lru()); |
| 1286 |
} |
| 1287 |
|
| 1288 |
static void |
| 1289 |
LruDelete(File file) |
| 1290 |
{ |
| 1291 |
Vfd *vfdP; |
| 1292 |
|
| 1293 |
Assert(file != 0); |
| 1294 |
|
| 1295 |
DO_DB(elog(LOG, "LruDelete %d (%s)", |
| 1296 |
file, VfdCache[file].fileName)); |
| 1297 |
|
| 1298 |
vfdP = &VfdCache[file]; |
| 1299 |
|
| 1300 |
pgaio_closing_fd(vfdP->fd); |
| 1301 |
|
| 1302 |
/* |
| 1303 |
* Close the file. We aren't expecting this to fail; if it does, better |
| 1304 |
* to leak the FD than to mess up our internal state. |
| 1305 |
*/ |
| 1306 |
if (close(vfdP->fd) != 0) |
| 1307 |
elog(vfdP->fdstate & FD_TEMP_FILE_LIMIT ? LOG : data_sync_elevel(LOG), |
| 1308 |
"could not close file \"%s\": %m", vfdP->fileName); |
| 1309 |
vfdP->fd = VFD_CLOSED; |
| 1310 |
--nfile; |
| 1311 |
|
| 1312 |
/* delete the vfd record from the LRU ring */ |
| 1313 |
Delete(file); |
| 1314 |
} |
| 1315 |
|
| 1316 |
static void |
| 1317 |
Insert(File file) |
| 1318 |
{ |
| 1319 |
Vfd *vfdP; |
| 1320 |
|
| 1321 |
Assert(file != 0); |
| 1322 |
|
| 1323 |
DO_DB(elog(LOG, "Insert %d (%s)", |
| 1324 |
file, VfdCache[file].fileName)); |
| 1325 |
DO_DB(_dump_lru()); |
| 1326 |
|
| 1327 |
vfdP = &VfdCache[file]; |
| 1328 |
|
| 1329 |
vfdP->lruMoreRecently = 0; |
| 1330 |
vfdP->lruLessRecently = VfdCache[0].lruLessRecently; |
| 1331 |
VfdCache[0].lruLessRecently = file; |
| 1332 |
VfdCache[vfdP->lruLessRecently].lruMoreRecently = file; |
| 1333 |
|
| 1334 |
DO_DB(_dump_lru()); |
| 1335 |
} |
| 1336 |
|
| 1337 |
/* returns 0 on success, -1 on re-open failure (with errno set) */ |
| 1338 |
static int |
| 1339 |
LruInsert(File file) |
| 1340 |
{ |
| 1341 |
Vfd *vfdP; |
| 1342 |
|
| 1343 |
Assert(file != 0); |
| 1344 |
|
| 1345 |
DO_DB(elog(LOG, "LruInsert %d (%s)", |
| 1346 |
file, VfdCache[file].fileName)); |
| 1347 |
|
| 1348 |
vfdP = &VfdCache[file]; |
| 1349 |
|
| 1350 |
if (FileIsNotOpen(file)) |
| 1351 |
{ |
| 1352 |
/* Close excess kernel FDs. */ |
| 1353 |
ReleaseLruFiles(); |
| 1354 |
|
| 1355 |
/* |
| 1356 |
* The open could still fail for lack of file descriptors, eg due to |
| 1357 |
* overall system file table being full. So, be prepared to release |
| 1358 |
* another FD if necessary... |
| 1359 |
*/ |
| 1360 |
vfdP->fd = BasicOpenFilePerm(vfdP->fileName, vfdP->fileFlags, |
| 1361 |
vfdP->fileMode); |
| 1362 |
if (vfdP->fd < 0) |
| 1363 |
{ |
| 1364 |
DO_DB(elog(LOG, "re-open failed: %m")); |
| 1365 |
return -1; |
| 1366 |
} |
| 1367 |
else |
| 1368 |
{ |
| 1369 |
++nfile; |
| 1370 |
} |
| 1371 |
} |
| 1372 |
|
| 1373 |
/* |
| 1374 |
* put it at the head of the Lru ring |
| 1375 |
*/ |
| 1376 |
|
| 1377 |
Insert(file); |
| 1378 |
|
| 1379 |
return 0; |
| 1380 |
} |
| 1381 |
|
| 1382 |
/* |
| 1383 |
* Release one kernel FD by closing the least-recently-used VFD. |
| 1384 |
*/ |
| 1385 |
static bool |
| 1386 |
ReleaseLruFile(void) |
| 1387 |
{ |
| 1388 |
DO_DB(elog(LOG, "ReleaseLruFile. Opened %d", nfile)); |
| 1389 |
|
| 1390 |
if (nfile > 0) |
| 1391 |
{ |
| 1392 |
/* |
| 1393 |
* There are opened files and so there should be at least one used vfd |
| 1394 |
* in the ring. |
| 1395 |
*/ |
| 1396 |
Assert(VfdCache[0].lruMoreRecently != 0); |
| 1397 |
LruDelete(VfdCache[0].lruMoreRecently); |
| 1398 |
return true; /* freed a file */ |
| 1399 |
} |
| 1400 |
return false; /* no files available to free */ |
| 1401 |
} |
| 1402 |
|
| 1403 |
/* |
| 1404 |
* Release kernel FDs as needed to get under the max_safe_fds limit. |
| 1405 |
* After calling this, it's OK to try to open another file. |
| 1406 |
*/ |
| 1407 |
static void |
| 1408 |
ReleaseLruFiles(void) |
| 1409 |
{ |
| 1410 |
while (nfile + numAllocatedDescs + numExternalFDs >= max_safe_fds) |
| 1411 |
{ |
| 1412 |
if (!ReleaseLruFile()) |
| 1413 |
break; |
| 1414 |
} |
| 1415 |
} |
| 1416 |
|
| 1417 |
static File |
| 1418 |
AllocateVfd(void) |
| 1419 |
{ |
| 1420 |
Index i; |
| 1421 |
File file; |
| 1422 |
|
| 1423 |
DO_DB(elog(LOG, "AllocateVfd. Size %zu", SizeVfdCache)); |
| 1424 |
|
| 1425 |
Assert(SizeVfdCache > 0); /* InitFileAccess not called? */ |
| 1426 |
|
| 1427 |
if (VfdCache[0].nextFree == 0) |
| 1428 |
{ |
| 1429 |
/* |
| 1430 |
* The free list is empty so it is time to increase the size of the |
| 1431 |
* array. We choose to double it each time this happens. However, |
| 1432 |
* there's not much point in starting *real* small. |
| 1433 |
*/ |
| 1434 |
Size newCacheSize = SizeVfdCache * 2; |
| 1435 |
Vfd *newVfdCache; |
| 1436 |
|
| 1437 |
if (newCacheSize < 32) |
| 1438 |
newCacheSize = 32; |
| 1439 |
|
| 1440 |
/* |
| 1441 |
* Be careful not to clobber VfdCache ptr if realloc fails. |
| 1442 |
*/ |
| 1443 |
newVfdCache = (Vfd *) realloc(VfdCache, sizeof(Vfd) * newCacheSize); |
| 1444 |
if (newVfdCache == NULL) |
| 1445 |
ereport(ERROR, |
| 1446 |
(errcode(ERRCODE_OUT_OF_MEMORY), |
| 1447 |
errmsg("out of memory"))); |
| 1448 |
VfdCache = newVfdCache; |
| 1449 |
|
| 1450 |
/* |
| 1451 |
* Initialize the new entries and link them into the free list. |
| 1452 |
*/ |
| 1453 |
for (i = SizeVfdCache; i < newCacheSize; i++) |
| 1454 |
{ |
| 1455 |
MemSet(&(VfdCache[i]), 0, sizeof(Vfd)); |
| 1456 |
VfdCache[i].nextFree = i + 1; |
| 1457 |
VfdCache[i].fd = VFD_CLOSED; |
| 1458 |
} |
| 1459 |
VfdCache[newCacheSize - 1].nextFree = 0; |
| 1460 |
VfdCache[0].nextFree = SizeVfdCache; |
| 1461 |
|
| 1462 |
/* |
| 1463 |
* Record the new size |
| 1464 |
*/ |
| 1465 |
SizeVfdCache = newCacheSize; |
| 1466 |
} |
| 1467 |
|
| 1468 |
file = VfdCache[0].nextFree; |
| 1469 |
|
| 1470 |
VfdCache[0].nextFree = VfdCache[file].nextFree; |
| 1471 |
|
| 1472 |
return file; |
| 1473 |
} |
| 1474 |
|
| 1475 |
static void |
| 1476 |
FreeVfd(File file) |
| 1477 |
{ |
| 1478 |
Vfd *vfdP = &VfdCache[file]; |
| 1479 |
|
| 1480 |
DO_DB(elog(LOG, "FreeVfd: %d (%s)", |
| 1481 |
file, vfdP->fileName ? vfdP->fileName : "")); |
| 1482 |
|
| 1483 |
if (vfdP->fileName != NULL) |
| 1484 |
{ |
| 1485 |
free(vfdP->fileName); |
| 1486 |
vfdP->fileName = NULL; |
| 1487 |
} |
| 1488 |
vfdP->fdstate = 0x0; |
| 1489 |
|
| 1490 |
vfdP->nextFree = VfdCache[0].nextFree; |
| 1491 |
VfdCache[0].nextFree = file; |
| 1492 |
} |
| 1493 |
|
| 1494 |
/* returns 0 on success, -1 on re-open failure (with errno set) */ |
| 1495 |
static int |
| 1496 |
FileAccess(File file) |
| 1497 |
{ |
| 1498 |
int returnValue; |
| 1499 |
|
| 1500 |
DO_DB(elog(LOG, "FileAccess %d (%s)", |
| 1501 |
file, VfdCache[file].fileName)); |
| 1502 |
|
| 1503 |
/* |
| 1504 |
* Is the file open? If not, open it and put it at the head of the LRU |
| 1505 |
* ring (possibly closing the least recently used file to get an FD). |
| 1506 |
*/ |
| 1507 |
|
| 1508 |
if (FileIsNotOpen(file)) |
| 1509 |
{ |
| 1510 |
returnValue = LruInsert(file); |
| 1511 |
if (returnValue != 0) |
| 1512 |
return returnValue; |
| 1513 |
} |
| 1514 |
else if (VfdCache[0].lruLessRecently != file) |
| 1515 |
{ |
| 1516 |
/* |
| 1517 |
* We now know that the file is open and that it is not the last one |
| 1518 |
* accessed, so we need to move it to the head of the Lru ring. |
| 1519 |
*/ |
| 1520 |
|
| 1521 |
Delete(file); |
| 1522 |
Insert(file); |
| 1523 |
} |
| 1524 |
|
| 1525 |
return 0; |
| 1526 |
} |
| 1527 |
|
| 1528 |
/* |
| 1529 |
* Called whenever a temporary file is deleted to report its size. |
| 1530 |
*/ |
| 1531 |
static void |
| 1532 |
ReportTemporaryFileUsage(const char *path, off_t size) |
| 1533 |
{ |
| 1534 |
pgstat_report_tempfile(size); |
| 1535 |
|
| 1536 |
if (log_temp_files >= 0) |
| 1537 |
{ |
| 1538 |
if ((size / 1024) >= log_temp_files) |
| 1539 |
ereport(LOG, |
| 1540 |
(errmsg("temporary file: path \"%s\", size %lu", |
| 1541 |
path, (unsigned long) size))); |
| 1542 |
} |
| 1543 |
} |
| 1544 |
|
| 1545 |
/* |
| 1546 |
* Called to register a temporary file for automatic close. |
| 1547 |
* ResourceOwnerEnlarge(CurrentResourceOwner) must have been called |
| 1548 |
* before the file was opened. |
| 1549 |
*/ |
| 1550 |
static void |
| 1551 |
RegisterTemporaryFile(File file) |
| 1552 |
{ |
| 1553 |
ResourceOwnerRememberFile(CurrentResourceOwner, file); |
| 1554 |
VfdCache[file].resowner = CurrentResourceOwner; |
| 1555 |
|
| 1556 |
/* Backup mechanism for closing at end of xact. */ |
| 1557 |
VfdCache[file].fdstate |= FD_CLOSE_AT_EOXACT; |
| 1558 |
have_xact_temporary_files = true; |
| 1559 |
} |
| 1560 |
|
| 1561 |
/* |
| 1562 |
* Called when we get a shared invalidation message on some relation. |
| 1563 |
*/ |
| 1564 |
#ifdef NOT_USED |
| 1565 |
void |
| 1566 |
FileInvalidate(File file) |
| 1567 |
{ |
| 1568 |
Assert(FileIsValid(file)); |
| 1569 |
if (!FileIsNotOpen(file)) |
| 1570 |
LruDelete(file); |
| 1571 |
} |
| 1572 |
#endif |
| 1573 |
|
| 1574 |
/* |
| 1575 |
* Open a file with PathNameOpenFilePerm() and pass default file mode for the |
| 1576 |
* fileMode parameter. |
| 1577 |
*/ |
| 1578 |
File |
| 1579 |
PathNameOpenFile(const char *fileName, int fileFlags) |
| 1580 |
{ |
| 1581 |
return PathNameOpenFilePerm(fileName, fileFlags, pg_file_create_mode); |
| 1582 |
} |
| 1583 |
|
| 1584 |
/* |
| 1585 |
* open a file in an arbitrary directory |
| 1586 |
* |
| 1587 |
* NB: if the passed pathname is relative (which it usually is), |
| 1588 |
* it will be interpreted relative to the process' working directory |
| 1589 |
* (which should always be $PGDATA when this code is running). |
| 1590 |
*/ |
| 1591 |
File |
| 1592 |
PathNameOpenFilePerm(const char *fileName, int fileFlags, mode_t fileMode) |
| 1593 |
{ |
| 1594 |
char *fnamecopy; |
| 1595 |
File file; |
| 1596 |
Vfd *vfdP; |
| 1597 |
|
| 1598 |
DO_DB(elog(LOG, "PathNameOpenFilePerm: %s %x %o", |
| 1599 |
fileName, fileFlags, fileMode)); |
| 1600 |
|
| 1601 |
/* |
| 1602 |
* We need a malloc'd copy of the file name; fail cleanly if no room. |
| 1603 |
*/ |
| 1604 |
fnamecopy = strdup(fileName); |
| 1605 |
if (fnamecopy == NULL) |
| 1606 |
ereport(ERROR, |
| 1607 |
(errcode(ERRCODE_OUT_OF_MEMORY), |
| 1608 |
errmsg("out of memory"))); |
| 1609 |
|
| 1610 |
file = AllocateVfd(); |
| 1611 |
vfdP = &VfdCache[file]; |
| 1612 |
|
| 1613 |
/* Close excess kernel FDs. */ |
| 1614 |
ReleaseLruFiles(); |
| 1615 |
|
| 1616 |
/* |
| 1617 |
* Descriptors managed by VFDs are implicitly marked O_CLOEXEC. The |
| 1618 |
* client shouldn't be expected to know which kernel descriptors are |
| 1619 |
* currently open, so it wouldn't make sense for them to be inherited by |
| 1620 |
* executed subprograms. |
| 1621 |
*/ |
| 1622 |
fileFlags |= O_CLOEXEC; |
| 1623 |
|
| 1624 |
vfdP->fd = BasicOpenFilePerm(fileName, fileFlags, fileMode); |
| 1625 |
|
| 1626 |
if (vfdP->fd < 0) |
| 1627 |
{ |
| 1628 |
int save_errno = errno; |
| 1629 |
|
| 1630 |
FreeVfd(file); |
| 1631 |
free(fnamecopy); |
| 1632 |
errno = save_errno; |
| 1633 |
return -1; |
| 1634 |
} |
| 1635 |
++nfile; |
| 1636 |
DO_DB(elog(LOG, "PathNameOpenFile: success %d", |
| 1637 |
vfdP->fd)); |
| 1638 |
|
| 1639 |
vfdP->fileName = fnamecopy; |
| 1640 |
/* Saved flags are adjusted to be OK for re-opening file */ |
| 1641 |
vfdP->fileFlags = fileFlags & ~(O_CREAT | O_TRUNC | O_EXCL); |
| 1642 |
vfdP->fileMode = fileMode; |
| 1643 |
vfdP->fileSize = 0; |
| 1644 |
vfdP->fdstate = 0x0; |
| 1645 |
vfdP->resowner = NULL; |
| 1646 |
|
| 1647 |
Insert(file); |
| 1648 |
|
| 1649 |
return file; |
| 1650 |
} |
| 1651 |
|
| 1652 |
/* |
| 1653 |
* Create directory 'directory'. If necessary, create 'basedir', which must |
| 1654 |
* be the directory above it. This is designed for creating the top-level |
| 1655 |
* temporary directory on demand before creating a directory underneath it. |
| 1656 |
* Do nothing if the directory already exists. |
| 1657 |
* |
| 1658 |
* Directories created within the top-level temporary directory should begin |
| 1659 |
* with PG_TEMP_FILE_PREFIX, so that they can be identified as temporary and |
| 1660 |
* deleted at startup by RemovePgTempFiles(). Further subdirectories below |
| 1661 |
* that do not need any particular prefix. |
| 1662 |
*/ |
| 1663 |
void |
| 1664 |
PathNameCreateTemporaryDir(const char *basedir, const char *directory) |
| 1665 |
{ |
| 1666 |
if (MakePGDirectory(directory) < 0) |
| 1667 |
{ |
| 1668 |
if (errno == EEXIST) |
| 1669 |
return; |
| 1670 |
|
| 1671 |
/* |
| 1672 |
* Failed. Try to create basedir first in case it's missing. Tolerate |
| 1673 |
* EEXIST to close a race against another process following the same |
| 1674 |
* algorithm. |
| 1675 |
*/ |
| 1676 |
if (MakePGDirectory(basedir) < 0 && errno != EEXIST) |
| 1677 |
ereport(ERROR, |
| 1678 |
(errcode_for_file_access(), |
| 1679 |
errmsg("cannot create temporary directory \"%s\": %m", |
| 1680 |
basedir))); |
| 1681 |
|
| 1682 |
/* Try again. */ |
| 1683 |
if (MakePGDirectory(directory) < 0 && errno != EEXIST) |
| 1684 |
ereport(ERROR, |
| 1685 |
(errcode_for_file_access(), |
| 1686 |
errmsg("cannot create temporary subdirectory \"%s\": %m", |
| 1687 |
directory))); |
| 1688 |
} |
| 1689 |
} |
| 1690 |
|
| 1691 |
/* |
| 1692 |
* Delete a directory and everything in it, if it exists. |
| 1693 |
*/ |
| 1694 |
void |
| 1695 |
PathNameDeleteTemporaryDir(const char *dirname) |
| 1696 |
{ |
| 1697 |
struct stat statbuf; |
| 1698 |
|
| 1699 |
/* Silently ignore missing directory. */ |
| 1700 |
if (stat(dirname, &statbuf) != 0 && errno == ENOENT) |
| 1701 |
return; |
| 1702 |
|
| 1703 |
/* |
| 1704 |
* Currently, walkdir doesn't offer a way for our passed in function to |
| 1705 |
* maintain state. Perhaps it should, so that we could tell the caller |
| 1706 |
* whether this operation succeeded or failed. Since this operation is |
| 1707 |
* used in a cleanup path, we wouldn't actually behave differently: we'll |
| 1708 |
* just log failures. |
| 1709 |
*/ |
| 1710 |
walkdir(dirname, unlink_if_exists_fname, false, LOG); |
| 1711 |
} |
| 1712 |
|
| 1713 |
/* |
| 1714 |
* Open a temporary file that will disappear when we close it. |
| 1715 |
* |
| 1716 |
* This routine takes care of generating an appropriate tempfile name. |
| 1717 |
* There's no need to pass in fileFlags or fileMode either, since only |
| 1718 |
* one setting makes any sense for a temp file. |
| 1719 |
* |
| 1720 |
* Unless interXact is true, the file is remembered by CurrentResourceOwner |
| 1721 |
* to ensure it's closed and deleted when it's no longer needed, typically at |
| 1722 |
* the end-of-transaction. In most cases, you don't want temporary files to |
| 1723 |
* outlive the transaction that created them, so this should be false -- but |
| 1724 |
* if you need "somewhat" temporary storage, this might be useful. In either |
| 1725 |
* case, the file is removed when the File is explicitly closed. |
| 1726 |
*/ |
| 1727 |
File |
| 1728 |
OpenTemporaryFile(bool interXact) |
| 1729 |
{ |
| 1730 |
File file = 0; |
| 1731 |
|
| 1732 |
Assert(temporary_files_allowed); /* check temp file access is up */ |
| 1733 |
|
| 1734 |
/* |
| 1735 |
* Make sure the current resource owner has space for this File before we |
| 1736 |
* open it, if we'll be registering it below. |
| 1737 |
*/ |
| 1738 |
if (!interXact) |
| 1739 |
ResourceOwnerEnlarge(CurrentResourceOwner); |
| 1740 |
|
| 1741 |
/* |
| 1742 |
* If some temp tablespace(s) have been given to us, try to use the next |
| 1743 |
* one. If a given tablespace can't be found, we silently fall back to |
| 1744 |
* the database's default tablespace. |
| 1745 |
* |
| 1746 |
* BUT: if the temp file is slated to outlive the current transaction, |
| 1747 |
* force it into the database's default tablespace, so that it will not |
| 1748 |
* pose a threat to possible tablespace drop attempts. |
| 1749 |
*/ |
| 1750 |
if (numTempTableSpaces > 0 && !interXact) |
| 1751 |
{ |
| 1752 |
Oid tblspcOid = GetNextTempTableSpace(); |
| 1753 |
|
| 1754 |
if (OidIsValid(tblspcOid)) |
| 1755 |
file = OpenTemporaryFileInTablespace(tblspcOid, false); |
| 1756 |
} |
| 1757 |
|
| 1758 |
/* |
| 1759 |
* If not, or if tablespace is bad, create in database's default |
| 1760 |
* tablespace. MyDatabaseTableSpace should normally be set before we get |
| 1761 |
* here, but just in case it isn't, fall back to pg_default tablespace. |
| 1762 |
*/ |
| 1763 |
if (file <= 0) |
| 1764 |
file = OpenTemporaryFileInTablespace(MyDatabaseTableSpace ? |
| 1765 |
MyDatabaseTableSpace : |
| 1766 |
DEFAULTTABLESPACE_OID, |
| 1767 |
true); |
| 1768 |
|
| 1769 |
/* Mark it for deletion at close and temporary file size limit */ |
| 1770 |
VfdCache[file].fdstate |= FD_DELETE_AT_CLOSE | FD_TEMP_FILE_LIMIT; |
| 1771 |
|
| 1772 |
/* Register it with the current resource owner */ |
| 1773 |
if (!interXact) |
| 1774 |
RegisterTemporaryFile(file); |
| 1775 |
|
| 1776 |
return file; |
| 1777 |
} |
| 1778 |
|
| 1779 |
/* |
| 1780 |
* Return the path of the temp directory in a given tablespace. |
| 1781 |
*/ |
| 1782 |
void |
| 1783 |
TempTablespacePath(char *path, Oid tablespace) |
| 1784 |
{ |
| 1785 |
/* |
| 1786 |
* Identify the tempfile directory for this tablespace. |
| 1787 |
* |
| 1788 |
* If someone tries to specify pg_global, use pg_default instead. |
| 1789 |
*/ |
| 1790 |
if (tablespace == InvalidOid || |
| 1791 |
tablespace == DEFAULTTABLESPACE_OID || |
| 1792 |
tablespace == GLOBALTABLESPACE_OID) |
| 1793 |
snprintf(path, MAXPGPATH, "base/%s", PG_TEMP_FILES_DIR); |
| 1794 |
else |
| 1795 |
{ |
| 1796 |
/* All other tablespaces are accessed via symlinks */ |
| 1797 |
snprintf(path, MAXPGPATH, "%s/%u/%s/%s", |
| 1798 |
PG_TBLSPC_DIR, tablespace, TABLESPACE_VERSION_DIRECTORY, |
| 1799 |
PG_TEMP_FILES_DIR); |
| 1800 |
} |
| 1801 |
} |
| 1802 |
|
| 1803 |
/* |
| 1804 |
* Open a temporary file in a specific tablespace. |
| 1805 |
* Subroutine for OpenTemporaryFile, which see for details. |
| 1806 |
*/ |
| 1807 |
static File |
| 1808 |
OpenTemporaryFileInTablespace(Oid tblspcOid, bool rejectError) |
| 1809 |
{ |
| 1810 |
char tempdirpath[MAXPGPATH]; |
| 1811 |
char tempfilepath[MAXPGPATH]; |
| 1812 |
File file; |
| 1813 |
|
| 1814 |
TempTablespacePath(tempdirpath, tblspcOid); |
| 1815 |
|
| 1816 |
/* |
| 1817 |
* Generate a tempfile name that should be unique within the current |
| 1818 |
* database instance. |
| 1819 |
*/ |
| 1820 |
snprintf(tempfilepath, sizeof(tempfilepath), "%s/%s%d.%ld", |
| 1821 |
tempdirpath, PG_TEMP_FILE_PREFIX, MyProcPid, tempFileCounter++); |
| 1822 |
|
| 1823 |
/* |
| 1824 |
* Open the file. Note: we don't use O_EXCL, in case there is an orphaned |
| 1825 |
* temp file that can be reused. |
| 1826 |
*/ |
| 1827 |
file = PathNameOpenFile(tempfilepath, |
| 1828 |
O_RDWR | O_CREAT | O_TRUNC | PG_BINARY); |
| 1829 |
if (file <= 0) |
| 1830 |
{ |
| 1831 |
/* |
| 1832 |
* We might need to create the tablespace's tempfile directory, if no |
| 1833 |
* one has yet done so. |
| 1834 |
* |
| 1835 |
* Don't check for an error from MakePGDirectory; it could fail if |
| 1836 |
* someone else just did the same thing. If it doesn't work then |
| 1837 |
* we'll bomb out on the second create attempt, instead. |
| 1838 |
*/ |
| 1839 |
(void) MakePGDirectory(tempdirpath); |
| 1840 |
|
| 1841 |
file = PathNameOpenFile(tempfilepath, |
| 1842 |
O_RDWR | O_CREAT | O_TRUNC | PG_BINARY); |
| 1843 |
if (file <= 0 && rejectError) |
| 1844 |
elog(ERROR, "could not create temporary file \"%s\": %m", |
| 1845 |
tempfilepath); |
| 1846 |
} |
| 1847 |
|
| 1848 |
return file; |
| 1849 |
} |
| 1850 |
|
| 1851 |
|
| 1852 |
/* |
| 1853 |
* Create a new file. The directory containing it must already exist. Files |
| 1854 |
* created this way are subject to temp_file_limit and are automatically |
| 1855 |
* closed at end of transaction, but are not automatically deleted on close |
| 1856 |
* because they are intended to be shared between cooperating backends. |
| 1857 |
* |
| 1858 |
* If the file is inside the top-level temporary directory, its name should |
| 1859 |
* begin with PG_TEMP_FILE_PREFIX so that it can be identified as temporary |
| 1860 |
* and deleted at startup by RemovePgTempFiles(). Alternatively, it can be |
| 1861 |
* inside a directory created with PathNameCreateTemporaryDir(), in which case |
| 1862 |
* the prefix isn't needed. |
| 1863 |
*/ |
| 1864 |
File |
| 1865 |
PathNameCreateTemporaryFile(const char *path, bool error_on_failure) |
| 1866 |
{ |
| 1867 |
File file; |
| 1868 |
|
| 1869 |
Assert(temporary_files_allowed); /* check temp file access is up */ |
| 1870 |
|
| 1871 |
ResourceOwnerEnlarge(CurrentResourceOwner); |
| 1872 |
|
| 1873 |
/* |
| 1874 |
* Open the file. Note: we don't use O_EXCL, in case there is an orphaned |
| 1875 |
* temp file that can be reused. |
| 1876 |
*/ |
| 1877 |
file = PathNameOpenFile(path, O_RDWR | O_CREAT | O_TRUNC | PG_BINARY); |
| 1878 |
if (file <= 0) |
| 1879 |
{ |
| 1880 |
if (error_on_failure) |
| 1881 |
ereport(ERROR, |
| 1882 |
(errcode_for_file_access(), |
| 1883 |
errmsg("could not create temporary file \"%s\": %m", |
| 1884 |
path))); |
| 1885 |
else |
| 1886 |
return file; |
| 1887 |
} |
| 1888 |
|
| 1889 |
/* Mark it for temp_file_limit accounting. */ |
| 1890 |
VfdCache[file].fdstate |= FD_TEMP_FILE_LIMIT; |
| 1891 |
|
| 1892 |
/* Register it for automatic close. */ |
| 1893 |
RegisterTemporaryFile(file); |
| 1894 |
|
| 1895 |
return file; |
| 1896 |
} |
| 1897 |
|
| 1898 |
/* |
| 1899 |
* Open a file that was created with PathNameCreateTemporaryFile, possibly in |
| 1900 |
* another backend. Files opened this way don't count against the |
| 1901 |
* temp_file_limit of the caller, are automatically closed at the end of the |
| 1902 |
* transaction but are not deleted on close. |
| 1903 |
*/ |
| 1904 |
File |
| 1905 |
PathNameOpenTemporaryFile(const char *path, int mode) |
| 1906 |
{ |
| 1907 |
File file; |
| 1908 |
|
| 1909 |
Assert(temporary_files_allowed); /* check temp file access is up */ |
| 1910 |
|
| 1911 |
ResourceOwnerEnlarge(CurrentResourceOwner); |
| 1912 |
|
| 1913 |
file = PathNameOpenFile(path, mode | PG_BINARY); |
| 1914 |
|
| 1915 |
/* If no such file, then we don't raise an error. */ |
| 1916 |
if (file <= 0 && errno != ENOENT) |
| 1917 |
ereport(ERROR, |
| 1918 |
(errcode_for_file_access(), |
| 1919 |
errmsg("could not open temporary file \"%s\": %m", |
| 1920 |
path))); |
| 1921 |
|
| 1922 |
if (file > 0) |
| 1923 |
{ |
| 1924 |
/* Register it for automatic close. */ |
| 1925 |
RegisterTemporaryFile(file); |
| 1926 |
} |
| 1927 |
|
| 1928 |
return file; |
| 1929 |
} |
| 1930 |
|
| 1931 |
/* |
| 1932 |
* Delete a file by pathname. Return true if the file existed, false if |
| 1933 |
* didn't. |
| 1934 |
*/ |
| 1935 |
bool |
| 1936 |
PathNameDeleteTemporaryFile(const char *path, bool error_on_failure) |
| 1937 |
{ |
| 1938 |
struct stat filestats; |
| 1939 |
int stat_errno; |
| 1940 |
|
| 1941 |
/* Get the final size for pgstat reporting. */ |
| 1942 |
if (stat(path, &filestats) != 0) |
| 1943 |
stat_errno = errno; |
| 1944 |
else |
| 1945 |
stat_errno = 0; |
| 1946 |
|
| 1947 |
/* |
| 1948 |
* Unlike FileClose's automatic file deletion code, we tolerate |
| 1949 |
* non-existence to support BufFileDeleteFileSet which doesn't know how |
| 1950 |
* many segments it has to delete until it runs out. |
| 1951 |
*/ |
| 1952 |
if (stat_errno == ENOENT) |
| 1953 |
return false; |
| 1954 |
|
| 1955 |
if (unlink(path) < 0) |
| 1956 |
{ |
| 1957 |
if (errno != ENOENT) |
| 1958 |
ereport(error_on_failure ? ERROR : LOG, |
| 1959 |
(errcode_for_file_access(), |
| 1960 |
errmsg("could not unlink temporary file \"%s\": %m", |
| 1961 |
path))); |
| 1962 |
return false; |
| 1963 |
} |
| 1964 |
|
| 1965 |
if (stat_errno == 0) |
| 1966 |
ReportTemporaryFileUsage(path, filestats.st_size); |
| 1967 |
else |
| 1968 |
{ |
| 1969 |
errno = stat_errno; |
| 1970 |
ereport(LOG, |
| 1971 |
(errcode_for_file_access(), |
| 1972 |
errmsg("could not stat file \"%s\": %m", path))); |
| 1973 |
} |
| 1974 |
|
| 1975 |
return true; |
| 1976 |
} |
| 1977 |
|
| 1978 |
/* |
| 1979 |
* close a file when done with it |
| 1980 |
*/ |
| 1981 |
void |
| 1982 |
FileClose(File file) |
| 1983 |
{ |
| 1984 |
Vfd *vfdP; |
| 1985 |
|
| 1986 |
Assert(FileIsValid(file)); |
| 1987 |
|
| 1988 |
DO_DB(elog(LOG, "FileClose: %d (%s)", |
| 1989 |
file, VfdCache[file].fileName)); |
| 1990 |
|
| 1991 |
vfdP = &VfdCache[file]; |
| 1992 |
|
| 1993 |
if (!FileIsNotOpen(file)) |
| 1994 |
{ |
| 1995 |
pgaio_closing_fd(vfdP->fd); |
| 1996 |
|
| 1997 |
/* close the file */ |
| 1998 |
if (close(vfdP->fd) != 0) |
| 1999 |
{ |
| 2000 |
/* |
| 2001 |
* We may need to panic on failure to close non-temporary files; |
| 2002 |
* see LruDelete. |
| 2003 |
*/ |
| 2004 |
elog(vfdP->fdstate & FD_TEMP_FILE_LIMIT ? LOG : data_sync_elevel(LOG), |
| 2005 |
"could not close file \"%s\": %m", vfdP->fileName); |
| 2006 |
} |
| 2007 |
|
| 2008 |
--nfile; |
| 2009 |
vfdP->fd = VFD_CLOSED; |
| 2010 |
|
| 2011 |
/* remove the file from the lru ring */ |
| 2012 |
Delete(file); |
| 2013 |
} |
| 2014 |
|
| 2015 |
if (vfdP->fdstate & FD_TEMP_FILE_LIMIT) |
| 2016 |
{ |
| 2017 |
/* Subtract its size from current usage (do first in case of error) */ |
| 2018 |
temporary_files_size -= vfdP->fileSize; |
| 2019 |
vfdP->fileSize = 0; |
| 2020 |
} |
| 2021 |
|
| 2022 |
/* |
| 2023 |
* Delete the file if it was temporary, and make a log entry if wanted |
| 2024 |
*/ |
| 2025 |
if (vfdP->fdstate & FD_DELETE_AT_CLOSE) |
| 2026 |
{ |
| 2027 |
struct stat filestats; |
| 2028 |
int stat_errno; |
| 2029 |
|
| 2030 |
/* |
| 2031 |
* If we get an error, as could happen within the ereport/elog calls, |
| 2032 |
* we'll come right back here during transaction abort. Reset the |
| 2033 |
* flag to ensure that we can't get into an infinite loop. This code |
| 2034 |
* is arranged to ensure that the worst-case consequence is failing to |
| 2035 |
* emit log message(s), not failing to attempt the unlink. |
| 2036 |
*/ |
| 2037 |
vfdP->fdstate &= ~FD_DELETE_AT_CLOSE; |
| 2038 |
|
| 2039 |
|
| 2040 |
/* first try the stat() */ |
| 2041 |
if (stat(vfdP->fileName, &filestats)) |
| 2042 |
stat_errno = errno; |
| 2043 |
else |
| 2044 |
stat_errno = 0; |
| 2045 |
|
| 2046 |
/* in any case do the unlink */ |
| 2047 |
if (unlink(vfdP->fileName)) |
| 2048 |
ereport(LOG, |
| 2049 |
(errcode_for_file_access(), |
| 2050 |
errmsg("could not delete file \"%s\": %m", vfdP->fileName))); |
| 2051 |
|
| 2052 |
/* and last report the stat results */ |
| 2053 |
if (stat_errno == 0) |
| 2054 |
ReportTemporaryFileUsage(vfdP->fileName, filestats.st_size); |
| 2055 |
else |
| 2056 |
{ |
| 2057 |
errno = stat_errno; |
| 2058 |
ereport(LOG, |
| 2059 |
(errcode_for_file_access(), |
| 2060 |
errmsg("could not stat file \"%s\": %m", vfdP->fileName))); |
| 2061 |
} |
| 2062 |
} |
| 2063 |
|
| 2064 |
/* Unregister it from the resource owner */ |
| 2065 |
if (vfdP->resowner) |
| 2066 |
ResourceOwnerForgetFile(vfdP->resowner, file); |
| 2067 |
|
| 2068 |
/* |
| 2069 |
* Return the Vfd slot to the free list |
| 2070 |
*/ |
| 2071 |
FreeVfd(file); |
| 2072 |
} |
| 2073 |
|
| 2074 |
/* |
| 2075 |
* FilePrefetch - initiate asynchronous read of a given range of the file. |
| 2076 |
* |
| 2077 |
* Returns 0 on success, otherwise an errno error code (like posix_fadvise()). |
| 2078 |
* |
| 2079 |
* posix_fadvise() is the simplest standardized interface that accomplishes |
| 2080 |
* this. |
| 2081 |
*/ |
| 2082 |
int |
| 2083 |
FilePrefetch(File file, off_t offset, off_t amount, uint32 wait_event_info) |
| 2084 |
{ |
| 2085 |
Assert(FileIsValid(file)); |
| 2086 |
|
| 2087 |
DO_DB(elog(LOG, "FilePrefetch: %d (%s) " INT64_FORMAT " " INT64_FORMAT, |
| 2088 |
file, VfdCache[file].fileName, |
| 2089 |
(int64) offset, (int64) amount)); |
| 2090 |
|
| 2091 |
#if defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_WILLNEED) |
| 2092 |
{ |
| 2093 |
int returnCode; |
| 2094 |
|
| 2095 |
returnCode = FileAccess(file); |
| 2096 |
if (returnCode < 0) |
| 2097 |
return returnCode; |
| 2098 |
|
| 2099 |
retry: |
| 2100 |
pgstat_report_wait_start(wait_event_info); |
| 2101 |
returnCode = posix_fadvise(VfdCache[file].fd, offset, amount, |
| 2102 |
POSIX_FADV_WILLNEED); |
| 2103 |
pgstat_report_wait_end(); |
| 2104 |
|
| 2105 |
if (returnCode == EINTR) |
| 2106 |
goto retry; |
| 2107 |
|
| 2108 |
return returnCode; |
| 2109 |
} |
| 2110 |
#elif defined(__darwin__) |
| 2111 |
{ |
| 2112 |
struct radvisory |
| 2113 |
{ |
| 2114 |
off_t ra_offset; /* offset into the file */ |
| 2115 |
int ra_count; /* size of the read */ |
| 2116 |
} ra; |
| 2117 |
int returnCode; |
| 2118 |
|
| 2119 |
returnCode = FileAccess(file); |
| 2120 |
if (returnCode < 0) |
| 2121 |
return returnCode; |
| 2122 |
|
| 2123 |
ra.ra_offset = offset; |
| 2124 |
ra.ra_count = amount; |
| 2125 |
pgstat_report_wait_start(wait_event_info); |
| 2126 |
returnCode = fcntl(VfdCache[file].fd, F_RDADVISE, &ra); |
| 2127 |
pgstat_report_wait_end(); |
| 2128 |
if (returnCode != -1) |
| 2129 |
return 0; |
| 2130 |
else |
| 2131 |
return errno; |
| 2132 |
} |
| 2133 |
#else |
| 2134 |
return 0; |
| 2135 |
#endif |
| 2136 |
} |
| 2137 |
|
| 2138 |
void |
| 2139 |
FileWriteback(File file, off_t offset, off_t nbytes, uint32 wait_event_info) |
| 2140 |
{ |
| 2141 |
int returnCode; |
| 2142 |
|
| 2143 |
Assert(FileIsValid(file)); |
| 2144 |
|
| 2145 |
DO_DB(elog(LOG, "FileWriteback: %d (%s) " INT64_FORMAT " " INT64_FORMAT, |
| 2146 |
file, VfdCache[file].fileName, |
| 2147 |
(int64) offset, (int64) nbytes)); |
| 2148 |
|
| 2149 |
if (nbytes <= 0) |
| 2150 |
return; |
| 2151 |
|
| 2152 |
if (VfdCache[file].fileFlags & PG_O_DIRECT) |
| 2153 |
return; |
| 2154 |
|
| 2155 |
returnCode = FileAccess(file); |
| 2156 |
if (returnCode < 0) |
| 2157 |
return; |
| 2158 |
|
| 2159 |
pgstat_report_wait_start(wait_event_info); |
| 2160 |
pg_flush_data(VfdCache[file].fd, offset, nbytes); |
| 2161 |
pgstat_report_wait_end(); |
| 2162 |
} |
| 2163 |
|
| 2164 |
ssize_t |
| 2165 |
FileReadV(File file, const struct iovec *iov, int iovcnt, off_t offset, |
| 2166 |
uint32 wait_event_info) |
| 2167 |
{ |
| 2168 |
ssize_t returnCode; |
| 2169 |
Vfd *vfdP; |
| 2170 |
|
| 2171 |
Assert(FileIsValid(file)); |
| 2172 |
|
| 2173 |
DO_DB(elog(LOG, "FileReadV: %d (%s) " INT64_FORMAT " %d", |
| 2174 |
file, VfdCache[file].fileName, |
| 2175 |
(int64) offset, |
| 2176 |
iovcnt)); |
| 2177 |
|
| 2178 |
returnCode = FileAccess(file); |
| 2179 |
if (returnCode < 0) |
| 2180 |
return returnCode; |
| 2181 |
|
| 2182 |
vfdP = &VfdCache[file]; |
| 2183 |
|
| 2184 |
retry: |
| 2185 |
pgstat_report_wait_start(wait_event_info); |
| 2186 |
returnCode = pg_preadv(vfdP->fd, iov, iovcnt, offset); |
| 2187 |
pgstat_report_wait_end(); |
| 2188 |
|
| 2189 |
if (returnCode < 0) |
| 2190 |
{ |
| 2191 |
/* |
| 2192 |
* Windows may run out of kernel buffers and return "Insufficient |
| 2193 |
* system resources" error. Wait a bit and retry to solve it. |
| 2194 |
* |
| 2195 |
* It is rumored that EINTR is also possible on some Unix filesystems, |
| 2196 |
* in which case immediate retry is indicated. |
| 2197 |
*/ |
| 2198 |
#ifdef WIN32 |
| 2199 |
DWORD error = GetLastError(); |
| 2200 |
|
| 2201 |
switch (error) |
| 2202 |
{ |
| 2203 |
case ERROR_NO_SYSTEM_RESOURCES: |
| 2204 |
pg_usleep(1000L); |
| 2205 |
errno = EINTR; |
| 2206 |
break; |
| 2207 |
default: |
| 2208 |
_dosmaperr(error); |
| 2209 |
break; |
| 2210 |
} |
| 2211 |
#endif |
| 2212 |
/* OK to retry if interrupted */ |
| 2213 |
if (errno == EINTR) |
| 2214 |
goto retry; |
| 2215 |
} |
| 2216 |
|
| 2217 |
return returnCode; |
| 2218 |
} |
| 2219 |
|
| 2220 |
int |
| 2221 |
FileStartReadV(PgAioHandle *ioh, File file, |
| 2222 |
int iovcnt, off_t offset, |
| 2223 |
uint32 wait_event_info) |
| 2224 |
{ |
| 2225 |
int returnCode; |
| 2226 |
Vfd *vfdP; |
| 2227 |
|
| 2228 |
Assert(FileIsValid(file)); |
| 2229 |
|
| 2230 |
DO_DB(elog(LOG, "FileStartReadV: %d (%s) " INT64_FORMAT " %d", |
| 2231 |
file, VfdCache[file].fileName, |
| 2232 |
(int64) offset, |
| 2233 |
iovcnt)); |
| 2234 |
|
| 2235 |
returnCode = FileAccess(file); |
| 2236 |
if (returnCode < 0) |
| 2237 |
return returnCode; |
| 2238 |
|
| 2239 |
vfdP = &VfdCache[file]; |
| 2240 |
|
| 2241 |
pgaio_io_start_readv(ioh, vfdP->fd, iovcnt, offset); |
| 2242 |
|
| 2243 |
return 0; |
| 2244 |
} |
| 2245 |
|
| 2246 |
ssize_t |
| 2247 |
FileWriteV(File file, const struct iovec *iov, int iovcnt, off_t offset, |
| 2248 |
uint32 wait_event_info) |
| 2249 |
{ |
| 2250 |
ssize_t returnCode; |
| 2251 |
Vfd *vfdP; |
| 2252 |
|
| 2253 |
Assert(FileIsValid(file)); |
| 2254 |
|
| 2255 |
DO_DB(elog(LOG, "FileWriteV: %d (%s) " INT64_FORMAT " %d", |
| 2256 |
file, VfdCache[file].fileName, |
| 2257 |
(int64) offset, |
| 2258 |
iovcnt)); |
| 2259 |
|
| 2260 |
returnCode = FileAccess(file); |
| 2261 |
if (returnCode < 0) |
| 2262 |
return returnCode; |
| 2263 |
|
| 2264 |
vfdP = &VfdCache[file]; |
| 2265 |
|
| 2266 |
/* |
| 2267 |
* If enforcing temp_file_limit and it's a temp file, check to see if the |
| 2268 |
* write would overrun temp_file_limit, and throw error if so. Note: it's |
| 2269 |
* really a modularity violation to throw error here; we should set errno |
| 2270 |
* and return -1. However, there's no way to report a suitable error |
| 2271 |
* message if we do that. All current callers would just throw error |
| 2272 |
* immediately anyway, so this is safe at present. |
| 2273 |
*/ |
| 2274 |
if (temp_file_limit >= 0 && (vfdP->fdstate & FD_TEMP_FILE_LIMIT)) |
| 2275 |
{ |
| 2276 |
off_t past_write = offset; |
| 2277 |
|
| 2278 |
for (int i = 0; i < iovcnt; ++i) |
| 2279 |
past_write += iov[i].iov_len; |
| 2280 |
|
| 2281 |
if (past_write > vfdP->fileSize) |
| 2282 |
{ |
| 2283 |
uint64 newTotal = temporary_files_size; |
| 2284 |
|
| 2285 |
newTotal += past_write - vfdP->fileSize; |
| 2286 |
if (newTotal > (uint64) temp_file_limit * (uint64) 1024) |
| 2287 |
ereport(ERROR, |
| 2288 |
(errcode(ERRCODE_CONFIGURATION_LIMIT_EXCEEDED), |
| 2289 |
errmsg("temporary file size exceeds \"temp_file_limit\" (%dkB)", |
| 2290 |
temp_file_limit))); |
| 2291 |
} |
| 2292 |
} |
| 2293 |
|
| 2294 |
retry: |
| 2295 |
pgstat_report_wait_start(wait_event_info); |
| 2296 |
returnCode = pg_pwritev(vfdP->fd, iov, iovcnt, offset); |
| 2297 |
pgstat_report_wait_end(); |
| 2298 |
|
| 2299 |
if (returnCode >= 0) |
| 2300 |
{ |
| 2301 |
/* |
| 2302 |
* Some callers expect short writes to set errno, and traditionally we |
| 2303 |
* have assumed that they imply disk space shortage. We don't want to |
| 2304 |
* waste CPU cycles adding up the total size here, so we'll just set |
| 2305 |
* it for all successful writes in case such a caller determines that |
| 2306 |
* the write was short and ereports "%m". |
| 2307 |
*/ |
| 2308 |
errno = ENOSPC; |
| 2309 |
|
| 2310 |
/* |
| 2311 |
* Maintain fileSize and temporary_files_size if it's a temp file. |
| 2312 |
*/ |
| 2313 |
if (vfdP->fdstate & FD_TEMP_FILE_LIMIT) |
| 2314 |
{ |
| 2315 |
off_t past_write = offset + returnCode; |
| 2316 |
|
| 2317 |
if (past_write > vfdP->fileSize) |
| 2318 |
{ |
| 2319 |
temporary_files_size += past_write - vfdP->fileSize; |
| 2320 |
vfdP->fileSize = past_write; |
| 2321 |
} |
| 2322 |
} |
| 2323 |
} |
| 2324 |
else |
| 2325 |
{ |
| 2326 |
/* |
| 2327 |
* See comments in FileReadV() |
| 2328 |
*/ |
| 2329 |
#ifdef WIN32 |
| 2330 |
DWORD error = GetLastError(); |
| 2331 |
|
| 2332 |
switch (error) |
| 2333 |
{ |
| 2334 |
case ERROR_NO_SYSTEM_RESOURCES: |
| 2335 |
pg_usleep(1000L); |
| 2336 |
errno = EINTR; |
| 2337 |
break; |
| 2338 |
default: |
| 2339 |
_dosmaperr(error); |
| 2340 |
break; |
| 2341 |
} |
| 2342 |
#endif |
| 2343 |
/* OK to retry if interrupted */ |
| 2344 |
if (errno == EINTR) |
| 2345 |
goto retry; |
| 2346 |
} |
| 2347 |
|
| 2348 |
return returnCode; |
| 2349 |
} |
| 2350 |
|
| 2351 |
int |
| 2352 |
FileSync(File file, uint32 wait_event_info) |
| 2353 |
{ |
| 2354 |
int returnCode; |
| 2355 |
|
| 2356 |
Assert(FileIsValid(file)); |
| 2357 |
|
| 2358 |
DO_DB(elog(LOG, "FileSync: %d (%s)", |
| 2359 |
file, VfdCache[file].fileName)); |
| 2360 |
|
| 2361 |
returnCode = FileAccess(file); |
| 2362 |
if (returnCode < 0) |
| 2363 |
return returnCode; |
| 2364 |
|
| 2365 |
pgstat_report_wait_start(wait_event_info); |
| 2366 |
returnCode = pg_fsync(VfdCache[file].fd); |
| 2367 |
pgstat_report_wait_end(); |
| 2368 |
|
| 2369 |
return returnCode; |
| 2370 |
} |
| 2371 |
|
| 2372 |
/* |
| 2373 |
* Zero a region of the file. |
| 2374 |
* |
| 2375 |
* Returns 0 on success, -1 otherwise. In the latter case errno is set to the |
| 2376 |
* appropriate error. |
| 2377 |
*/ |
| 2378 |
int |
| 2379 |
FileZero(File file, off_t offset, off_t amount, uint32 wait_event_info) |
| 2380 |
{ |
| 2381 |
int returnCode; |
| 2382 |
ssize_t written; |
| 2383 |
|
| 2384 |
Assert(FileIsValid(file)); |
| 2385 |
|
| 2386 |
DO_DB(elog(LOG, "FileZero: %d (%s) " INT64_FORMAT " " INT64_FORMAT, |
| 2387 |
file, VfdCache[file].fileName, |
| 2388 |
(int64) offset, (int64) amount)); |
| 2389 |
|
| 2390 |
returnCode = FileAccess(file); |
| 2391 |
if (returnCode < 0) |
| 2392 |
return returnCode; |
| 2393 |
|
| 2394 |
pgstat_report_wait_start(wait_event_info); |
| 2395 |
written = pg_pwrite_zeros(VfdCache[file].fd, amount, offset); |
| 2396 |
pgstat_report_wait_end(); |
| 2397 |
|
| 2398 |
if (written < 0) |
| 2399 |
return -1; |
| 2400 |
else if (written != amount) |
| 2401 |
{ |
| 2402 |
/* if errno is unset, assume problem is no disk space */ |
| 2403 |
if (errno == 0) |
| 2404 |
errno = ENOSPC; |
| 2405 |
return -1; |
| 2406 |
} |
| 2407 |
|
| 2408 |
return 0; |
| 2409 |
} |
| 2410 |
|
| 2411 |
/* |
| 2412 |
* Try to reserve file space with posix_fallocate(). If posix_fallocate() is |
| 2413 |
* not implemented on the operating system or fails with EINVAL / EOPNOTSUPP, |
| 2414 |
* use FileZero() instead. |
| 2415 |
* |
| 2416 |
* Note that at least glibc() implements posix_fallocate() in userspace if not |
| 2417 |
* implemented by the filesystem. That's not the case for all environments |
| 2418 |
* though. |
| 2419 |
* |
| 2420 |
* Returns 0 on success, -1 otherwise. In the latter case errno is set to the |
| 2421 |
* appropriate error. |
| 2422 |
*/ |
| 2423 |
int |
| 2424 |
FileFallocate(File file, off_t offset, off_t amount, uint32 wait_event_info) |
| 2425 |
{ |
| 2426 |
#ifdef HAVE_POSIX_FALLOCATE |
| 2427 |
int returnCode; |
| 2428 |
|
| 2429 |
Assert(FileIsValid(file)); |
| 2430 |
|
| 2431 |
DO_DB(elog(LOG, "FileFallocate: %d (%s) " INT64_FORMAT " " INT64_FORMAT, |
| 2432 |
file, VfdCache[file].fileName, |
| 2433 |
(int64) offset, (int64) amount)); |
| 2434 |
|
| 2435 |
returnCode = FileAccess(file); |
| 2436 |
if (returnCode < 0) |
| 2437 |
return -1; |
| 2438 |
|
| 2439 |
retry: |
| 2440 |
pgstat_report_wait_start(wait_event_info); |
| 2441 |
returnCode = posix_fallocate(VfdCache[file].fd, offset, amount); |
| 2442 |
pgstat_report_wait_end(); |
| 2443 |
|
| 2444 |
if (returnCode == 0) |
| 2445 |
return 0; |
| 2446 |
else if (returnCode == EINTR) |
| 2447 |
goto retry; |
| 2448 |
|
| 2449 |
/* for compatibility with %m printing etc */ |
| 2450 |
errno = returnCode; |
| 2451 |
|
| 2452 |
/* |
| 2453 |
* Return in cases of a "real" failure, if fallocate is not supported, |
| 2454 |
* fall through to the FileZero() backed implementation. |
| 2455 |
*/ |
| 2456 |
if (returnCode != EINVAL && returnCode != EOPNOTSUPP) |
| 2457 |
return -1; |
| 2458 |
#endif |
| 2459 |
|
| 2460 |
return FileZero(file, offset, amount, wait_event_info); |
| 2461 |
} |
| 2462 |
|
| 2463 |
off_t |
| 2464 |
FileSize(File file) |
| 2465 |
{ |
| 2466 |
Assert(FileIsValid(file)); |
| 2467 |
|
| 2468 |
DO_DB(elog(LOG, "FileSize %d (%s)", |
| 2469 |
file, VfdCache[file].fileName)); |
| 2470 |
|
| 2471 |
if (FileIsNotOpen(file)) |
| 2472 |
{ |
| 2473 |
if (FileAccess(file) < 0) |
| 2474 |
return (off_t) -1; |
| 2475 |
} |
| 2476 |
|
| 2477 |
return lseek(VfdCache[file].fd, 0, SEEK_END); |
| 2478 |
} |
| 2479 |
|
| 2480 |
int |
| 2481 |
FileTruncate(File file, off_t offset, uint32 wait_event_info) |
| 2482 |
{ |
| 2483 |
int returnCode; |
| 2484 |
|
| 2485 |
Assert(FileIsValid(file)); |
| 2486 |
|
| 2487 |
DO_DB(elog(LOG, "FileTruncate %d (%s)", |
| 2488 |
file, VfdCache[file].fileName)); |
| 2489 |
|
| 2490 |
returnCode = FileAccess(file); |
| 2491 |
if (returnCode < 0) |
| 2492 |
return returnCode; |
| 2493 |
|
| 2494 |
pgstat_report_wait_start(wait_event_info); |
| 2495 |
returnCode = pg_ftruncate(VfdCache[file].fd, offset); |
| 2496 |
pgstat_report_wait_end(); |
| 2497 |
|
| 2498 |
if (returnCode == 0 && VfdCache[file].fileSize > offset) |
| 2499 |
{ |
| 2500 |
/* adjust our state for truncation of a temp file */ |
| 2501 |
Assert(VfdCache[file].fdstate & FD_TEMP_FILE_LIMIT); |
| 2502 |
temporary_files_size -= VfdCache[file].fileSize - offset; |
| 2503 |
VfdCache[file].fileSize = offset; |
| 2504 |
} |
| 2505 |
|
| 2506 |
return returnCode; |
| 2507 |
} |
| 2508 |
|
| 2509 |
/* |
| 2510 |
* Return the pathname associated with an open file. |
| 2511 |
* |
| 2512 |
* The returned string points to an internal buffer, which is valid until |
| 2513 |
* the file is closed. |
| 2514 |
*/ |
| 2515 |
char * |
| 2516 |
FilePathName(File file) |
| 2517 |
{ |
| 2518 |
Assert(FileIsValid(file)); |
| 2519 |
|
| 2520 |
return VfdCache[file].fileName; |
| 2521 |
} |
| 2522 |
|
| 2523 |
/* |
| 2524 |
* Return the raw file descriptor of an opened file. |
| 2525 |
* |
| 2526 |
* The returned file descriptor will be valid until the file is closed, but |
| 2527 |
* there are a lot of things that can make that happen. So the caller should |
| 2528 |
* be careful not to do much of anything else before it finishes using the |
| 2529 |
* returned file descriptor. |
| 2530 |
*/ |
| 2531 |
int |
| 2532 |
FileGetRawDesc(File file) |
| 2533 |
{ |
| 2534 |
int returnCode; |
| 2535 |
|
| 2536 |
returnCode = FileAccess(file); |
| 2537 |
if (returnCode < 0) |
| 2538 |
return returnCode; |
| 2539 |
|
| 2540 |
Assert(FileIsValid(file)); |
| 2541 |
return VfdCache[file].fd; |
| 2542 |
} |
| 2543 |
|
| 2544 |
/* |
| 2545 |
* FileGetRawFlags - returns the file flags on open(2) |
| 2546 |
*/ |
| 2547 |
int |
| 2548 |
FileGetRawFlags(File file) |
| 2549 |
{ |
| 2550 |
Assert(FileIsValid(file)); |
| 2551 |
return VfdCache[file].fileFlags; |
| 2552 |
} |
| 2553 |
|
| 2554 |
/* |
| 2555 |
* FileGetRawMode - returns the mode bitmask passed to open(2) |
| 2556 |
*/ |
| 2557 |
mode_t |
| 2558 |
FileGetRawMode(File file) |
| 2559 |
{ |
| 2560 |
Assert(FileIsValid(file)); |
| 2561 |
return VfdCache[file].fileMode; |
| 2562 |
} |
| 2563 |
|
| 2564 |
/* |
| 2565 |
* Make room for another allocatedDescs[] array entry if needed and possible. |
| 2566 |
* Returns true if an array element is available. |
| 2567 |
*/ |
| 2568 |
static bool |
| 2569 |
reserveAllocatedDesc(void) |
| 2570 |
{ |
| 2571 |
AllocateDesc *newDescs; |
| 2572 |
int newMax; |
| 2573 |
|
| 2574 |
/* Quick out if array already has a free slot. */ |
| 2575 |
if (numAllocatedDescs < maxAllocatedDescs) |
| 2576 |
return true; |
| 2577 |
|
| 2578 |
/* |
| 2579 |
* If the array hasn't yet been created in the current process, initialize |
| 2580 |
* it with FD_MINFREE / 3 elements. In many scenarios this is as many as |
| 2581 |
* we will ever need, anyway. We don't want to look at max_safe_fds |
| 2582 |
* immediately because set_max_safe_fds() may not have run yet. |
| 2583 |
*/ |
| 2584 |
if (allocatedDescs == NULL) |
| 2585 |
{ |
| 2586 |
newMax = FD_MINFREE / 3; |
| 2587 |
newDescs = (AllocateDesc *) malloc(newMax * sizeof(AllocateDesc)); |
| 2588 |
/* Out of memory already? Treat as fatal error. */ |
| 2589 |
if (newDescs == NULL) |
| 2590 |
ereport(ERROR, |
| 2591 |
(errcode(ERRCODE_OUT_OF_MEMORY), |
| 2592 |
errmsg("out of memory"))); |
| 2593 |
allocatedDescs = newDescs; |
| 2594 |
maxAllocatedDescs = newMax; |
| 2595 |
return true; |
| 2596 |
} |
| 2597 |
|
| 2598 |
/* |
| 2599 |
* Consider enlarging the array beyond the initial allocation used above. |
| 2600 |
* By the time this happens, max_safe_fds should be known accurately. |
| 2601 |
* |
| 2602 |
* We mustn't let allocated descriptors hog all the available FDs, and in |
| 2603 |
* practice we'd better leave a reasonable number of FDs for VFD use. So |
| 2604 |
* set the maximum to max_safe_fds / 3. (This should certainly be at |
| 2605 |
* least as large as the initial size, FD_MINFREE / 3, so we aren't |
| 2606 |
* tightening the restriction here.) Recall that "external" FDs are |
| 2607 |
* allowed to consume another third of max_safe_fds. |
| 2608 |
*/ |
| 2609 |
newMax = max_safe_fds / 3; |
| 2610 |
if (newMax > maxAllocatedDescs) |
| 2611 |
{ |
| 2612 |
newDescs = (AllocateDesc *) realloc(allocatedDescs, |
| 2613 |
newMax * sizeof(AllocateDesc)); |
| 2614 |
/* Treat out-of-memory as a non-fatal error. */ |
| 2615 |
if (newDescs == NULL) |
| 2616 |
return false; |
| 2617 |
allocatedDescs = newDescs; |
| 2618 |
maxAllocatedDescs = newMax; |
| 2619 |
return true; |
| 2620 |
} |
| 2621 |
|
| 2622 |
/* Can't enlarge allocatedDescs[] any more. */ |
| 2623 |
return false; |
| 2624 |
} |
| 2625 |
|
| 2626 |
/* |
| 2627 |
* Routines that want to use stdio (ie, FILE*) should use AllocateFile |
| 2628 |
* rather than plain fopen(). This lets fd.c deal with freeing FDs if |
| 2629 |
* necessary to open the file. When done, call FreeFile rather than fclose. |
| 2630 |
* |
| 2631 |
* Note that files that will be open for any significant length of time |
| 2632 |
* should NOT be handled this way, since they cannot share kernel file |
| 2633 |
* descriptors with other files; there is grave risk of running out of FDs |
| 2634 |
* if anyone locks down too many FDs. Most callers of this routine are |
| 2635 |
* simply reading a config file that they will read and close immediately. |
| 2636 |
* |
| 2637 |
* fd.c will automatically close all files opened with AllocateFile at |
| 2638 |
* transaction commit or abort; this prevents FD leakage if a routine |
| 2639 |
* that calls AllocateFile is terminated prematurely by ereport(ERROR). |
| 2640 |
* |
| 2641 |
* Ideally this should be the *only* direct call of fopen() in the backend. |
| 2642 |
*/ |
| 2643 |
FILE * |
| 2644 |
AllocateFile(const char *name, const char *mode) |
| 2645 |
{ |
| 2646 |
FILE *file; |
| 2647 |
|
| 2648 |
DO_DB(elog(LOG, "AllocateFile: Allocated %d (%s)", |
| 2649 |
numAllocatedDescs, name)); |
| 2650 |
|
| 2651 |
/* Can we allocate another non-virtual FD? */ |
| 2652 |
if (!reserveAllocatedDesc()) |
| 2653 |
ereport(ERROR, |
| 2654 |
(errcode(ERRCODE_INSUFFICIENT_RESOURCES), |
| 2655 |
errmsg("exceeded maxAllocatedDescs (%d) while trying to open file \"%s\"", |
| 2656 |
maxAllocatedDescs, name))); |
| 2657 |
|
| 2658 |
/* Close excess kernel FDs. */ |
| 2659 |
ReleaseLruFiles(); |
| 2660 |
|
| 2661 |
TryAgain: |
| 2662 |
if ((file = fopen(name, mode)) != NULL) |
| 2663 |
{ |
| 2664 |
AllocateDesc *desc = &allocatedDescs[numAllocatedDescs]; |
| 2665 |
|
| 2666 |
desc->kind = AllocateDescFile; |
| 2667 |
desc->desc.file = file; |
| 2668 |
desc->create_subid = GetCurrentSubTransactionId(); |
| 2669 |
numAllocatedDescs++; |
| 2670 |
return desc->desc.file; |
| 2671 |
} |
| 2672 |
|
| 2673 |
if (errno == EMFILE || errno == ENFILE) |
| 2674 |
{ |
| 2675 |
int save_errno = errno; |
| 2676 |
|
| 2677 |
ereport(LOG, |
| 2678 |
(errcode(ERRCODE_INSUFFICIENT_RESOURCES), |
| 2679 |
errmsg("out of file descriptors: %m; release and retry"))); |
| 2680 |
errno = 0; |
| 2681 |
if (ReleaseLruFile()) |
| 2682 |
goto TryAgain; |
| 2683 |
errno = save_errno; |
| 2684 |
} |
| 2685 |
|
| 2686 |
return NULL; |
| 2687 |
} |
| 2688 |
|
| 2689 |
/* |
| 2690 |
* Open a file with OpenTransientFilePerm() and pass default file mode for |
| 2691 |
* the fileMode parameter. |
| 2692 |
*/ |
| 2693 |
int |
| 2694 |
OpenTransientFile(const char *fileName, int fileFlags) |
| 2695 |
{ |
| 2696 |
return OpenTransientFilePerm(fileName, fileFlags, pg_file_create_mode); |
| 2697 |
} |
| 2698 |
|
| 2699 |
/* |
| 2700 |
* Like AllocateFile, but returns an unbuffered fd like open(2) |
| 2701 |
*/ |
| 2702 |
int |
| 2703 |
OpenTransientFilePerm(const char *fileName, int fileFlags, mode_t fileMode) |
| 2704 |
{ |
| 2705 |
int fd; |
| 2706 |
|
| 2707 |
DO_DB(elog(LOG, "OpenTransientFile: Allocated %d (%s)", |
| 2708 |
numAllocatedDescs, fileName)); |
| 2709 |
|
| 2710 |
/* Can we allocate another non-virtual FD? */ |
| 2711 |
if (!reserveAllocatedDesc()) |
| 2712 |
ereport(ERROR, |
| 2713 |
(errcode(ERRCODE_INSUFFICIENT_RESOURCES), |
| 2714 |
errmsg("exceeded maxAllocatedDescs (%d) while trying to open file \"%s\"", |
| 2715 |
maxAllocatedDescs, fileName))); |
| 2716 |
|
| 2717 |
/* Close excess kernel FDs. */ |
| 2718 |
ReleaseLruFiles(); |
| 2719 |
|
| 2720 |
fd = BasicOpenFilePerm(fileName, fileFlags, fileMode); |
| 2721 |
|
| 2722 |
if (fd >= 0) |
| 2723 |
{ |
| 2724 |
AllocateDesc *desc = &allocatedDescs[numAllocatedDescs]; |
| 2725 |
|
| 2726 |
desc->kind = AllocateDescRawFD; |
| 2727 |
desc->desc.fd = fd; |
| 2728 |
desc->create_subid = GetCurrentSubTransactionId(); |
| 2729 |
numAllocatedDescs++; |
| 2730 |
|
| 2731 |
return fd; |
| 2732 |
} |
| 2733 |
|
| 2734 |
return -1; /* failure */ |
| 2735 |
} |
| 2736 |
|
| 2737 |
/* |
| 2738 |
* Routines that want to initiate a pipe stream should use OpenPipeStream |
| 2739 |
* rather than plain popen(). This lets fd.c deal with freeing FDs if |
| 2740 |
* necessary. When done, call ClosePipeStream rather than pclose. |
| 2741 |
* |
| 2742 |
* This function also ensures that the popen'd program is run with default |
| 2743 |
* SIGPIPE processing, rather than the SIG_IGN setting the backend normally |
| 2744 |
* uses. This ensures desirable response to, eg, closing a read pipe early. |
| 2745 |
*/ |
| 2746 |
FILE * |
| 2747 |
OpenPipeStream(const char *command, const char *mode) |
| 2748 |
{ |
| 2749 |
FILE *file; |
| 2750 |
int save_errno; |
| 2751 |
|
| 2752 |
DO_DB(elog(LOG, "OpenPipeStream: Allocated %d (%s)", |
| 2753 |
numAllocatedDescs, command)); |
| 2754 |
|
| 2755 |
/* Can we allocate another non-virtual FD? */ |
| 2756 |
if (!reserveAllocatedDesc()) |
| 2757 |
ereport(ERROR, |
| 2758 |
(errcode(ERRCODE_INSUFFICIENT_RESOURCES), |
| 2759 |
errmsg("exceeded maxAllocatedDescs (%d) while trying to execute command \"%s\"", |
| 2760 |
maxAllocatedDescs, command))); |
| 2761 |
|
| 2762 |
/* Close excess kernel FDs. */ |
| 2763 |
ReleaseLruFiles(); |
| 2764 |
|
| 2765 |
TryAgain: |
| 2766 |
fflush(NULL); |
| 2767 |
pqsignal(SIGPIPE, SIG_DFL); |
| 2768 |
errno = 0; |
| 2769 |
file = popen(command, mode); |
| 2770 |
save_errno = errno; |
| 2771 |
pqsignal(SIGPIPE, SIG_IGN); |
| 2772 |
errno = save_errno; |
| 2773 |
if (file != NULL) |
| 2774 |
{ |
| 2775 |
AllocateDesc *desc = &allocatedDescs[numAllocatedDescs]; |
| 2776 |
|
| 2777 |
desc->kind = AllocateDescPipe; |
| 2778 |
desc->desc.file = file; |
| 2779 |
desc->create_subid = GetCurrentSubTransactionId(); |
| 2780 |
numAllocatedDescs++; |
| 2781 |
return desc->desc.file; |
| 2782 |
} |
| 2783 |
|
| 2784 |
if (errno == EMFILE || errno == ENFILE) |
| 2785 |
{ |
| 2786 |
ereport(LOG, |
| 2787 |
(errcode(ERRCODE_INSUFFICIENT_RESOURCES), |
| 2788 |
errmsg("out of file descriptors: %m; release and retry"))); |
| 2789 |
if (ReleaseLruFile()) |
| 2790 |
goto TryAgain; |
| 2791 |
errno = save_errno; |
| 2792 |
} |
| 2793 |
|
| 2794 |
return NULL; |
| 2795 |
} |
| 2796 |
|
| 2797 |
/* |
| 2798 |
* Free an AllocateDesc of any type. |
| 2799 |
* |
| 2800 |
* The argument *must* point into the allocatedDescs[] array. |
| 2801 |
*/ |
| 2802 |
static int |
| 2803 |
FreeDesc(AllocateDesc *desc) |
| 2804 |
{ |
| 2805 |
int result; |
| 2806 |
|
| 2807 |
/* Close the underlying object */ |
| 2808 |
switch (desc->kind) |
| 2809 |
{ |
| 2810 |
case AllocateDescFile: |
| 2811 |
result = fclose(desc->desc.file); |
| 2812 |
break; |
| 2813 |
case AllocateDescPipe: |
| 2814 |
result = pclose(desc->desc.file); |
| 2815 |
break; |
| 2816 |
case AllocateDescDir: |
| 2817 |
result = closedir(desc->desc.dir); |
| 2818 |
break; |
| 2819 |
case AllocateDescRawFD: |
| 2820 |
pgaio_closing_fd(desc->desc.fd); |
| 2821 |
result = close(desc->desc.fd); |
| 2822 |
break; |
| 2823 |
default: |
| 2824 |
elog(ERROR, "AllocateDesc kind not recognized"); |
| 2825 |
result = 0; /* keep compiler quiet */ |
| 2826 |
break; |
| 2827 |
} |
| 2828 |
|
| 2829 |
/* Compact storage in the allocatedDescs array */ |
| 2830 |
numAllocatedDescs--; |
| 2831 |
*desc = allocatedDescs[numAllocatedDescs]; |
| 2832 |
|
| 2833 |
return result; |
| 2834 |
} |
| 2835 |
|
| 2836 |
/* |
| 2837 |
* Close a file returned by AllocateFile. |
| 2838 |
* |
| 2839 |
* Note we do not check fclose's return value --- it is up to the caller |
| 2840 |
* to handle close errors. |
| 2841 |
*/ |
| 2842 |
int |
| 2843 |
FreeFile(FILE *file) |
| 2844 |
{ |
| 2845 |
int i; |
| 2846 |
|
| 2847 |
DO_DB(elog(LOG, "FreeFile: Allocated %d", numAllocatedDescs)); |
| 2848 |
|
| 2849 |
/* Remove file from list of allocated files, if it's present */ |
| 2850 |
for (i = numAllocatedDescs; --i >= 0;) |
| 2851 |
{ |
| 2852 |
AllocateDesc *desc = &allocatedDescs[i]; |
| 2853 |
|
| 2854 |
if (desc->kind == AllocateDescFile && desc->desc.file == file) |
| 2855 |
return FreeDesc(desc); |
| 2856 |
} |
| 2857 |
|
| 2858 |
/* Only get here if someone passes us a file not in allocatedDescs */ |
| 2859 |
elog(WARNING, "file passed to FreeFile was not obtained from AllocateFile"); |
| 2860 |
|
| 2861 |
return fclose(file); |
| 2862 |
} |
| 2863 |
|
| 2864 |
/* |
| 2865 |
* Close a file returned by OpenTransientFile. |
| 2866 |
* |
| 2867 |
* Note we do not check close's return value --- it is up to the caller |
| 2868 |
* to handle close errors. |
| 2869 |
*/ |
| 2870 |
int |
| 2871 |
CloseTransientFile(int fd) |
| 2872 |
{ |
| 2873 |
int i; |
| 2874 |
|
| 2875 |
DO_DB(elog(LOG, "CloseTransientFile: Allocated %d", numAllocatedDescs)); |
| 2876 |
|
| 2877 |
/* Remove fd from list of allocated files, if it's present */ |
| 2878 |
for (i = numAllocatedDescs; --i >= 0;) |
| 2879 |
{ |
| 2880 |
AllocateDesc *desc = &allocatedDescs[i]; |
| 2881 |
|
| 2882 |
if (desc->kind == AllocateDescRawFD && desc->desc.fd == fd) |
| 2883 |
return FreeDesc(desc); |
| 2884 |
} |
| 2885 |
|
| 2886 |
/* Only get here if someone passes us a file not in allocatedDescs */ |
| 2887 |
elog(WARNING, "fd passed to CloseTransientFile was not obtained from OpenTransientFile"); |
| 2888 |
|
| 2889 |
pgaio_closing_fd(fd); |
| 2890 |
|
| 2891 |
return close(fd); |
| 2892 |
} |
| 2893 |
|
| 2894 |
/* |
| 2895 |
* Routines that want to use <dirent.h> (ie, DIR*) should use AllocateDir |
| 2896 |
* rather than plain opendir(). This lets fd.c deal with freeing FDs if |
| 2897 |
* necessary to open the directory, and with closing it after an elog. |
| 2898 |
* When done, call FreeDir rather than closedir. |
| 2899 |
* |
| 2900 |
* Returns NULL, with errno set, on failure. Note that failure detection |
| 2901 |
* is commonly left to the following call of ReadDir or ReadDirExtended; |
| 2902 |
* see the comments for ReadDir. |
| 2903 |
* |
| 2904 |
* Ideally this should be the *only* direct call of opendir() in the backend. |
| 2905 |
*/ |
| 2906 |
DIR * |
| 2907 |
AllocateDir(const char *dirname) |
| 2908 |
{ |
| 2909 |
DIR *dir; |
| 2910 |
|
| 2911 |
DO_DB(elog(LOG, "AllocateDir: Allocated %d (%s)", |
| 2912 |
numAllocatedDescs, dirname)); |
| 2913 |
|
| 2914 |
/* Can we allocate another non-virtual FD? */ |
| 2915 |
if (!reserveAllocatedDesc()) |
| 2916 |
ereport(ERROR, |
| 2917 |
(errcode(ERRCODE_INSUFFICIENT_RESOURCES), |
| 2918 |
errmsg("exceeded maxAllocatedDescs (%d) while trying to open directory \"%s\"", |
| 2919 |
maxAllocatedDescs, dirname))); |
| 2920 |
|
| 2921 |
/* Close excess kernel FDs. */ |
| 2922 |
ReleaseLruFiles(); |
| 2923 |
|
| 2924 |
TryAgain: |
| 2925 |
if ((dir = opendir(dirname)) != NULL) |
| 2926 |
{ |
| 2927 |
AllocateDesc *desc = &allocatedDescs[numAllocatedDescs]; |
| 2928 |
|
| 2929 |
desc->kind = AllocateDescDir; |
| 2930 |
desc->desc.dir = dir; |
| 2931 |
desc->create_subid = GetCurrentSubTransactionId(); |
| 2932 |
numAllocatedDescs++; |
| 2933 |
return desc->desc.dir; |
| 2934 |
} |
| 2935 |
|
| 2936 |
if (errno == EMFILE || errno == ENFILE) |
| 2937 |
{ |
| 2938 |
int save_errno = errno; |
| 2939 |
|
| 2940 |
ereport(LOG, |
| 2941 |
(errcode(ERRCODE_INSUFFICIENT_RESOURCES), |
| 2942 |
errmsg("out of file descriptors: %m; release and retry"))); |
| 2943 |
errno = 0; |
| 2944 |
if (ReleaseLruFile()) |
| 2945 |
goto TryAgain; |
| 2946 |
errno = save_errno; |
| 2947 |
} |
| 2948 |
|
| 2949 |
return NULL; |
| 2950 |
} |
| 2951 |
|
| 2952 |
/* |
| 2953 |
* Read a directory opened with AllocateDir, ereport'ing any error. |
| 2954 |
* |
| 2955 |
* This is easier to use than raw readdir() since it takes care of some |
| 2956 |
* otherwise rather tedious and error-prone manipulation of errno. Also, |
| 2957 |
* if you are happy with a generic error message for AllocateDir failure, |
| 2958 |
* you can just do |
| 2959 |
* |
| 2960 |
* dir = AllocateDir(path); |
| 2961 |
* while ((dirent = ReadDir(dir, path)) != NULL) |
| 2962 |
* process dirent; |
| 2963 |
* FreeDir(dir); |
| 2964 |
* |
| 2965 |
* since a NULL dir parameter is taken as indicating AllocateDir failed. |
| 2966 |
* (Make sure errno isn't changed between AllocateDir and ReadDir if you |
| 2967 |
* use this shortcut.) |
| 2968 |
* |
| 2969 |
* The pathname passed to AllocateDir must be passed to this routine too, |
| 2970 |
* but it is only used for error reporting. |
| 2971 |
*/ |
| 2972 |
struct dirent * |
| 2973 |
ReadDir(DIR *dir, const char *dirname) |
| 2974 |
{ |
| 2975 |
return ReadDirExtended(dir, dirname, ERROR); |
| 2976 |
} |
| 2977 |
|
| 2978 |
/* |
| 2979 |
* Alternate version of ReadDir that allows caller to specify the elevel |
| 2980 |
* for any error report (whether it's reporting an initial failure of |
| 2981 |
* AllocateDir or a subsequent directory read failure). |
| 2982 |
* |
| 2983 |
* If elevel < ERROR, returns NULL after any error. With the normal coding |
| 2984 |
* pattern, this will result in falling out of the loop immediately as |
| 2985 |
* though the directory contained no (more) entries. |
| 2986 |
*/ |
| 2987 |
struct dirent * |
| 2988 |
ReadDirExtended(DIR *dir, const char *dirname, int elevel) |
| 2989 |
{ |
| 2990 |
struct dirent *dent; |
| 2991 |
|
| 2992 |
/* Give a generic message for AllocateDir failure, if caller didn't */ |
| 2993 |
if (dir == NULL) |
| 2994 |
{ |
| 2995 |
ereport(elevel, |
| 2996 |
(errcode_for_file_access(), |
| 2997 |
errmsg("could not open directory \"%s\": %m", |
| 2998 |
dirname))); |
| 2999 |
return NULL; |
| 3000 |
} |
| 3001 |
|
| 3002 |
errno = 0; |
| 3003 |
if ((dent = readdir(dir)) != NULL) |
| 3004 |
return dent; |
| 3005 |
|
| 3006 |
if (errno) |
| 3007 |
ereport(elevel, |
| 3008 |
(errcode_for_file_access(), |
| 3009 |
errmsg("could not read directory \"%s\": %m", |
| 3010 |
dirname))); |
| 3011 |
return NULL; |
| 3012 |
} |
| 3013 |
|
| 3014 |
/* |
| 3015 |
* Close a directory opened with AllocateDir. |
| 3016 |
* |
| 3017 |
* Returns closedir's return value (with errno set if it's not 0). |
| 3018 |
* Note we do not check the return value --- it is up to the caller |
| 3019 |
* to handle close errors if wanted. |
| 3020 |
* |
| 3021 |
* Does nothing if dir == NULL; we assume that directory open failure was |
| 3022 |
* already reported if desired. |
| 3023 |
*/ |
| 3024 |
int |
| 3025 |
FreeDir(DIR *dir) |
| 3026 |
{ |
| 3027 |
int i; |
| 3028 |
|
| 3029 |
/* Nothing to do if AllocateDir failed */ |
| 3030 |
if (dir == NULL) |
| 3031 |
return 0; |
| 3032 |
|
| 3033 |
DO_DB(elog(LOG, "FreeDir: Allocated %d", numAllocatedDescs)); |
| 3034 |
|
| 3035 |
/* Remove dir from list of allocated dirs, if it's present */ |
| 3036 |
for (i = numAllocatedDescs; --i >= 0;) |
| 3037 |
{ |
| 3038 |
AllocateDesc *desc = &allocatedDescs[i]; |
| 3039 |
|
| 3040 |
if (desc->kind == AllocateDescDir && desc->desc.dir == dir) |
| 3041 |
return FreeDesc(desc); |
| 3042 |
} |
| 3043 |
|
| 3044 |
/* Only get here if someone passes us a dir not in allocatedDescs */ |
| 3045 |
elog(WARNING, "dir passed to FreeDir was not obtained from AllocateDir"); |
| 3046 |
|
| 3047 |
return closedir(dir); |
| 3048 |
} |
| 3049 |
|
| 3050 |
|
| 3051 |
/* |
| 3052 |
* Close a pipe stream returned by OpenPipeStream. |
| 3053 |
*/ |
| 3054 |
int |
| 3055 |
ClosePipeStream(FILE *file) |
| 3056 |
{ |
| 3057 |
int i; |
| 3058 |
|
| 3059 |
DO_DB(elog(LOG, "ClosePipeStream: Allocated %d", numAllocatedDescs)); |
| 3060 |
|
| 3061 |
/* Remove file from list of allocated files, if it's present */ |
| 3062 |
for (i = numAllocatedDescs; --i >= 0;) |
| 3063 |
{ |
| 3064 |
AllocateDesc *desc = &allocatedDescs[i]; |
| 3065 |
|
| 3066 |
if (desc->kind == AllocateDescPipe && desc->desc.file == file) |
| 3067 |
return FreeDesc(desc); |
| 3068 |
} |
| 3069 |
|
| 3070 |
/* Only get here if someone passes us a file not in allocatedDescs */ |
| 3071 |
elog(WARNING, "file passed to ClosePipeStream was not obtained from OpenPipeStream"); |
| 3072 |
|
| 3073 |
return pclose(file); |
| 3074 |
} |
| 3075 |
|
| 3076 |
/* |
| 3077 |
* closeAllVfds |
| 3078 |
* |
| 3079 |
* Force all VFDs into the physically-closed state, so that the fewest |
| 3080 |
* possible number of kernel file descriptors are in use. There is no |
| 3081 |
* change in the logical state of the VFDs. |
| 3082 |
*/ |
| 3083 |
void |
| 3084 |
closeAllVfds(void) |
| 3085 |
{ |
| 3086 |
Index i; |
| 3087 |
|
| 3088 |
if (SizeVfdCache > 0) |
| 3089 |
{ |
| 3090 |
Assert(FileIsNotOpen(0)); /* Make sure ring not corrupted */ |
| 3091 |
for (i = 1; i < SizeVfdCache; i++) |
| 3092 |
{ |
| 3093 |
if (!FileIsNotOpen(i)) |
| 3094 |
LruDelete(i); |
| 3095 |
} |
| 3096 |
} |
| 3097 |
} |
| 3098 |
|
| 3099 |
|
| 3100 |
/* |
| 3101 |
* SetTempTablespaces |
| 3102 |
* |
| 3103 |
* Define a list (actually an array) of OIDs of tablespaces to use for |
| 3104 |
* temporary files. This list will be used until end of transaction, |
| 3105 |
* unless this function is called again before then. It is caller's |
| 3106 |
* responsibility that the passed-in array has adequate lifespan (typically |
| 3107 |
* it'd be allocated in TopTransactionContext). |
| 3108 |
* |
| 3109 |
* Some entries of the array may be InvalidOid, indicating that the current |
| 3110 |
* database's default tablespace should be used. |
| 3111 |
*/ |
| 3112 |
void |
| 3113 |
SetTempTablespaces(Oid *tableSpaces, int numSpaces) |
| 3114 |
{ |
| 3115 |
Assert(numSpaces >= 0); |
| 3116 |
tempTableSpaces = tableSpaces; |
| 3117 |
numTempTableSpaces = numSpaces; |
| 3118 |
|
| 3119 |
/* |
| 3120 |
* Select a random starting point in the list. This is to minimize |
| 3121 |
* conflicts between backends that are most likely sharing the same list |
| 3122 |
* of temp tablespaces. Note that if we create multiple temp files in the |
| 3123 |
* same transaction, we'll advance circularly through the list --- this |
| 3124 |
* ensures that large temporary sort files are nicely spread across all |
| 3125 |
* available tablespaces. |
| 3126 |
*/ |
| 3127 |
if (numSpaces > 1) |
| 3128 |
nextTempTableSpace = pg_prng_uint64_range(&pg_global_prng_state, |
| 3129 |
0, numSpaces - 1); |
| 3130 |
else |
| 3131 |
nextTempTableSpace = 0; |
| 3132 |
} |
| 3133 |
|
| 3134 |
/* |
| 3135 |
* TempTablespacesAreSet |
| 3136 |
* |
| 3137 |
* Returns true if SetTempTablespaces has been called in current transaction. |
| 3138 |
* (This is just so that tablespaces.c doesn't need its own per-transaction |
| 3139 |
* state.) |
| 3140 |
*/ |
| 3141 |
bool |
| 3142 |
TempTablespacesAreSet(void) |
| 3143 |
{ |
| 3144 |
return (numTempTableSpaces >= 0); |
| 3145 |
} |
| 3146 |
|
| 3147 |
/* |
| 3148 |
* GetTempTablespaces |
| 3149 |
* |
| 3150 |
* Populate an array with the OIDs of the tablespaces that should be used for |
| 3151 |
* temporary files. (Some entries may be InvalidOid, indicating that the |
| 3152 |
* current database's default tablespace should be used.) At most numSpaces |
| 3153 |
* entries will be filled. |
| 3154 |
* Returns the number of OIDs that were copied into the output array. |
| 3155 |
*/ |
| 3156 |
int |
| 3157 |
GetTempTablespaces(Oid *tableSpaces, int numSpaces) |
| 3158 |
{ |
| 3159 |
int i; |
| 3160 |
|
| 3161 |
Assert(TempTablespacesAreSet()); |
| 3162 |
for (i = 0; i < numTempTableSpaces && i < numSpaces; ++i) |
| 3163 |
tableSpaces[i] = tempTableSpaces[i]; |
| 3164 |
|
| 3165 |
return i; |
| 3166 |
} |
| 3167 |
|
| 3168 |
/* |
| 3169 |
* GetNextTempTableSpace |
| 3170 |
* |
| 3171 |
* Select the next temp tablespace to use. A result of InvalidOid means |
| 3172 |
* to use the current database's default tablespace. |
| 3173 |
*/ |
| 3174 |
Oid |
| 3175 |
GetNextTempTableSpace(void) |
| 3176 |
{ |
| 3177 |
if (numTempTableSpaces > 0) |
| 3178 |
{ |
| 3179 |
/* Advance nextTempTableSpace counter with wraparound */ |
| 3180 |
if (++nextTempTableSpace >= numTempTableSpaces) |
| 3181 |
nextTempTableSpace = 0; |
| 3182 |
return tempTableSpaces[nextTempTableSpace]; |
| 3183 |
} |
| 3184 |
return InvalidOid; |
| 3185 |
} |
| 3186 |
|
| 3187 |
|
| 3188 |
/* |
| 3189 |
* AtEOSubXact_Files |
| 3190 |
* |
| 3191 |
* Take care of subtransaction commit/abort. At abort, we close temp files |
| 3192 |
* that the subtransaction may have opened. At commit, we reassign the |
| 3193 |
* files that were opened to the parent subtransaction. |
| 3194 |
*/ |
| 3195 |
void |
| 3196 |
AtEOSubXact_Files(bool isCommit, SubTransactionId mySubid, |
| 3197 |
SubTransactionId parentSubid) |
| 3198 |
{ |
| 3199 |
Index i; |
| 3200 |
|
| 3201 |
for (i = 0; i < numAllocatedDescs; i++) |
| 3202 |
{ |
| 3203 |
if (allocatedDescs[i].create_subid == mySubid) |
| 3204 |
{ |
| 3205 |
if (isCommit) |
| 3206 |
allocatedDescs[i].create_subid = parentSubid; |
| 3207 |
else |
| 3208 |
{ |
| 3209 |
/* have to recheck the item after FreeDesc (ugly) */ |
| 3210 |
FreeDesc(&allocatedDescs[i--]); |
| 3211 |
} |
| 3212 |
} |
| 3213 |
} |
| 3214 |
} |
| 3215 |
|
| 3216 |
/* |
| 3217 |
* AtEOXact_Files |
| 3218 |
* |
| 3219 |
* This routine is called during transaction commit or abort. All still-open |
| 3220 |
* per-transaction temporary file VFDs are closed, which also causes the |
| 3221 |
* underlying files to be deleted (although they should've been closed already |
| 3222 |
* by the ResourceOwner cleanup). Furthermore, all "allocated" stdio files are |
| 3223 |
* closed. We also forget any transaction-local temp tablespace list. |
| 3224 |
* |
| 3225 |
* The isCommit flag is used only to decide whether to emit warnings about |
| 3226 |
* unclosed files. |
| 3227 |
*/ |
| 3228 |
void |
| 3229 |
AtEOXact_Files(bool isCommit) |
| 3230 |
{ |
| 3231 |
CleanupTempFiles(isCommit, false); |
| 3232 |
tempTableSpaces = NULL; |
| 3233 |
numTempTableSpaces = -1; |
| 3234 |
} |
| 3235 |
|
| 3236 |
/* |
| 3237 |
* BeforeShmemExit_Files |
| 3238 |
* |
| 3239 |
* before_shmem_exit hook to clean up temp files during backend shutdown. |
| 3240 |
* Here, we want to clean up *all* temp files including interXact ones. |
| 3241 |
*/ |
| 3242 |
static void |
| 3243 |
BeforeShmemExit_Files(int code, Datum arg) |
| 3244 |
{ |
| 3245 |
CleanupTempFiles(false, true); |
| 3246 |
|
| 3247 |
/* prevent further temp files from being created */ |
| 3248 |
#ifdef USE_ASSERT_CHECKING |
| 3249 |
temporary_files_allowed = false; |
| 3250 |
#endif |
| 3251 |
} |
| 3252 |
|
| 3253 |
/* |
| 3254 |
* Close temporary files and delete their underlying files. |
| 3255 |
* |
| 3256 |
* isCommit: if true, this is normal transaction commit, and we don't |
| 3257 |
* expect any remaining files; warn if there are some. |
| 3258 |
* |
| 3259 |
* isProcExit: if true, this is being called as the backend process is |
| 3260 |
* exiting. If that's the case, we should remove all temporary files; if |
| 3261 |
* that's not the case, we are being called for transaction commit/abort |
| 3262 |
* and should only remove transaction-local temp files. In either case, |
| 3263 |
* also clean up "allocated" stdio files, dirs and fds. |
| 3264 |
*/ |
| 3265 |
static void |
| 3266 |
CleanupTempFiles(bool isCommit, bool isProcExit) |
| 3267 |
{ |
| 3268 |
Index i; |
| 3269 |
|
| 3270 |
/* |
| 3271 |
* Careful here: at proc_exit we need extra cleanup, not just |
| 3272 |
* xact_temporary files. |
| 3273 |
*/ |
| 3274 |
if (isProcExit || have_xact_temporary_files) |
| 3275 |
{ |
| 3276 |
Assert(FileIsNotOpen(0)); /* Make sure ring not corrupted */ |
| 3277 |
for (i = 1; i < SizeVfdCache; i++) |
| 3278 |
{ |
| 3279 |
unsigned short fdstate = VfdCache[i].fdstate; |
| 3280 |
|
| 3281 |
if (((fdstate & FD_DELETE_AT_CLOSE) || (fdstate & FD_CLOSE_AT_EOXACT)) && |
| 3282 |
VfdCache[i].fileName != NULL) |
| 3283 |
{ |
| 3284 |
/* |
| 3285 |
* If we're in the process of exiting a backend process, close |
| 3286 |
* all temporary files. Otherwise, only close temporary files |
| 3287 |
* local to the current transaction. They should be closed by |
| 3288 |
* the ResourceOwner mechanism already, so this is just a |
| 3289 |
* debugging cross-check. |
| 3290 |
*/ |
| 3291 |
if (isProcExit) |
| 3292 |
FileClose(i); |
| 3293 |
else if (fdstate & FD_CLOSE_AT_EOXACT) |
| 3294 |
{ |
| 3295 |
elog(WARNING, |
| 3296 |
"temporary file %s not closed at end-of-transaction", |
| 3297 |
VfdCache[i].fileName); |
| 3298 |
FileClose(i); |
| 3299 |
} |
| 3300 |
} |
| 3301 |
} |
| 3302 |
|
| 3303 |
have_xact_temporary_files = false; |
| 3304 |
} |
| 3305 |
|
| 3306 |
/* Complain if any allocated files remain open at commit. */ |
| 3307 |
if (isCommit && numAllocatedDescs > 0) |
| 3308 |
elog(WARNING, "%d temporary files and directories not closed at end-of-transaction", |
| 3309 |
numAllocatedDescs); |
| 3310 |
|
| 3311 |
/* Clean up "allocated" stdio files, dirs and fds. */ |
| 3312 |
while (numAllocatedDescs > 0) |
| 3313 |
FreeDesc(&allocatedDescs[0]); |
| 3314 |
} |
| 3315 |
|
| 3316 |
|
| 3317 |
/* |
| 3318 |
* Remove temporary and temporary relation files left over from a prior |
| 3319 |
* postmaster session |
| 3320 |
* |
| 3321 |
* This should be called during postmaster startup. It will forcibly |
| 3322 |
* remove any leftover files created by OpenTemporaryFile and any leftover |
| 3323 |
* temporary relation files created by mdcreate. |
| 3324 |
* |
| 3325 |
* During post-backend-crash restart cycle, this routine is called when |
| 3326 |
* remove_temp_files_after_crash GUC is enabled. Multiple crashes while |
| 3327 |
* queries are using temp files could result in useless storage usage that can |
| 3328 |
* only be reclaimed by a service restart. The argument against enabling it is |
| 3329 |
* that someone might want to examine the temporary files for debugging |
| 3330 |
* purposes. This does however mean that OpenTemporaryFile had better allow for |
| 3331 |
* collision with an existing temp file name. |
| 3332 |
* |
| 3333 |
* NOTE: this function and its subroutines generally report syscall failures |
| 3334 |
* with ereport(LOG) and keep going. Removing temp files is not so critical |
| 3335 |
* that we should fail to start the database when we can't do it. |
| 3336 |
*/ |
| 3337 |
void |
| 3338 |
RemovePgTempFiles(void) |
| 3339 |
{ |
| 3340 |
char temp_path[MAXPGPATH + sizeof(PG_TBLSPC_DIR) + sizeof(TABLESPACE_VERSION_DIRECTORY) + sizeof(PG_TEMP_FILES_DIR)]; |
| 3341 |
DIR *spc_dir; |
| 3342 |
struct dirent *spc_de; |
| 3343 |
|
| 3344 |
/* |
| 3345 |
* First process temp files in pg_default ($PGDATA/base) |
| 3346 |
*/ |
| 3347 |
snprintf(temp_path, sizeof(temp_path), "base/%s", PG_TEMP_FILES_DIR); |
| 3348 |
RemovePgTempFilesInDir(temp_path, true, false); |
| 3349 |
RemovePgTempRelationFiles("base"); |
| 3350 |
|
| 3351 |
/* |
| 3352 |
* Cycle through temp directories for all non-default tablespaces. |
| 3353 |
*/ |
| 3354 |
spc_dir = AllocateDir(PG_TBLSPC_DIR); |
| 3355 |
|
| 3356 |
while ((spc_de = ReadDirExtended(spc_dir, PG_TBLSPC_DIR, LOG)) != NULL) |
| 3357 |
{ |
| 3358 |
if (strcmp(spc_de->d_name, ".") == 0 || |
| 3359 |
strcmp(spc_de->d_name, "..") == 0) |
| 3360 |
continue; |
| 3361 |
|
| 3362 |
snprintf(temp_path, sizeof(temp_path), "%s/%s/%s/%s", |
| 3363 |
PG_TBLSPC_DIR, spc_de->d_name, TABLESPACE_VERSION_DIRECTORY, |
| 3364 |
PG_TEMP_FILES_DIR); |
| 3365 |
RemovePgTempFilesInDir(temp_path, true, false); |
| 3366 |
|
| 3367 |
snprintf(temp_path, sizeof(temp_path), "%s/%s/%s", |
| 3368 |
PG_TBLSPC_DIR, spc_de->d_name, TABLESPACE_VERSION_DIRECTORY); |
| 3369 |
RemovePgTempRelationFiles(temp_path); |
| 3370 |
} |
| 3371 |
|
| 3372 |
FreeDir(spc_dir); |
| 3373 |
|
| 3374 |
/* |
| 3375 |
* In EXEC_BACKEND case there is a pgsql_tmp directory at the top level of |
| 3376 |
* DataDir as well. However, that is *not* cleaned here because doing so |
| 3377 |
* would create a race condition. It's done separately, earlier in |
| 3378 |
* postmaster startup. |
| 3379 |
*/ |
| 3380 |
} |
| 3381 |
|
| 3382 |
/* |
| 3383 |
* Process one pgsql_tmp directory for RemovePgTempFiles. |
| 3384 |
* |
| 3385 |
* If missing_ok is true, it's all right for the named directory to not exist. |
| 3386 |
* Any other problem results in a LOG message. (missing_ok should be true at |
| 3387 |
* the top level, since pgsql_tmp directories are not created until needed.) |
| 3388 |
* |
| 3389 |
* At the top level, this should be called with unlink_all = false, so that |
| 3390 |
* only files matching the temporary name prefix will be unlinked. When |
| 3391 |
* recursing it will be called with unlink_all = true to unlink everything |
| 3392 |
* under a top-level temporary directory. |
| 3393 |
* |
| 3394 |
* (These two flags could be replaced by one, but it seems clearer to keep |
| 3395 |
* them separate.) |
| 3396 |
*/ |
| 3397 |
void |
| 3398 |
RemovePgTempFilesInDir(const char *tmpdirname, bool missing_ok, bool unlink_all) |
| 3399 |
{ |
| 3400 |
DIR *temp_dir; |
| 3401 |
struct dirent *temp_de; |
| 3402 |
char rm_path[MAXPGPATH * 2]; |
| 3403 |
|
| 3404 |
temp_dir = AllocateDir(tmpdirname); |
| 3405 |
|
| 3406 |
if (temp_dir == NULL && errno == ENOENT && missing_ok) |
| 3407 |
return; |
| 3408 |
|
| 3409 |
while ((temp_de = ReadDirExtended(temp_dir, tmpdirname, LOG)) != NULL) |
| 3410 |
{ |
| 3411 |
if (strcmp(temp_de->d_name, ".") == 0 || |
| 3412 |
strcmp(temp_de->d_name, "..") == 0) |
| 3413 |
continue; |
| 3414 |
|
| 3415 |
snprintf(rm_path, sizeof(rm_path), "%s/%s", |
| 3416 |
tmpdirname, temp_de->d_name); |
| 3417 |
|
| 3418 |
if (unlink_all || |
| 3419 |
strncmp(temp_de->d_name, |
| 3420 |
PG_TEMP_FILE_PREFIX, |
| 3421 |
strlen(PG_TEMP_FILE_PREFIX)) == 0) |
| 3422 |
{ |
| 3423 |
PGFileType type = get_dirent_type(rm_path, temp_de, false, LOG); |
| 3424 |
|
| 3425 |
if (type == PGFILETYPE_ERROR) |
| 3426 |
continue; |
| 3427 |
else if (type == PGFILETYPE_DIR) |
| 3428 |
{ |
| 3429 |
/* recursively remove contents, then directory itself */ |
| 3430 |
RemovePgTempFilesInDir(rm_path, false, true); |
| 3431 |
|
| 3432 |
if (rmdir(rm_path) < 0) |
| 3433 |
ereport(LOG, |
| 3434 |
(errcode_for_file_access(), |
| 3435 |
errmsg("could not remove directory \"%s\": %m", |
| 3436 |
rm_path))); |
| 3437 |
} |
| 3438 |
else |
| 3439 |
{ |
| 3440 |
if (unlink(rm_path) < 0) |
| 3441 |
ereport(LOG, |
| 3442 |
(errcode_for_file_access(), |
| 3443 |
errmsg("could not remove file \"%s\": %m", |
| 3444 |
rm_path))); |
| 3445 |
} |
| 3446 |
} |
| 3447 |
else |
| 3448 |
ereport(LOG, |
| 3449 |
(errmsg("unexpected file found in temporary-files directory: \"%s\"", |
| 3450 |
rm_path))); |
| 3451 |
} |
| 3452 |
|
| 3453 |
FreeDir(temp_dir); |
| 3454 |
} |
| 3455 |
|
| 3456 |
/* Process one tablespace directory, look for per-DB subdirectories */ |
| 3457 |
static void |
| 3458 |
RemovePgTempRelationFiles(const char *tsdirname) |
| 3459 |
{ |
| 3460 |
DIR *ts_dir; |
| 3461 |
struct dirent *de; |
| 3462 |
char dbspace_path[MAXPGPATH * 2]; |
| 3463 |
|
| 3464 |
ts_dir = AllocateDir(tsdirname); |
| 3465 |
|
| 3466 |
while ((de = ReadDirExtended(ts_dir, tsdirname, LOG)) != NULL) |
| 3467 |
{ |
| 3468 |
/* |
| 3469 |
* We're only interested in the per-database directories, which have |
| 3470 |
* numeric names. Note that this code will also (properly) ignore "." |
| 3471 |
* and "..". |
| 3472 |
*/ |
| 3473 |
if (strspn(de->d_name, "0123456789") != strlen(de->d_name)) |
| 3474 |
continue; |
| 3475 |
|
| 3476 |
snprintf(dbspace_path, sizeof(dbspace_path), "%s/%s", |
| 3477 |
tsdirname, de->d_name); |
| 3478 |
RemovePgTempRelationFilesInDbspace(dbspace_path); |
| 3479 |
} |
| 3480 |
|
| 3481 |
FreeDir(ts_dir); |
| 3482 |
} |
| 3483 |
|
| 3484 |
/* Process one per-dbspace directory for RemovePgTempRelationFiles */ |
| 3485 |
static void |
| 3486 |
RemovePgTempRelationFilesInDbspace(const char *dbspacedirname) |
| 3487 |
{ |
| 3488 |
DIR *dbspace_dir; |
| 3489 |
struct dirent *de; |
| 3490 |
char rm_path[MAXPGPATH * 2]; |
| 3491 |
|
| 3492 |
dbspace_dir = AllocateDir(dbspacedirname); |
| 3493 |
|
| 3494 |
while ((de = ReadDirExtended(dbspace_dir, dbspacedirname, LOG)) != NULL) |
| 3495 |
{ |
| 3496 |
if (!looks_like_temp_rel_name(de->d_name)) |
| 3497 |
continue; |
| 3498 |
|
| 3499 |
snprintf(rm_path, sizeof(rm_path), "%s/%s", |
| 3500 |
dbspacedirname, de->d_name); |
| 3501 |
|
| 3502 |
if (unlink(rm_path) < 0) |
| 3503 |
ereport(LOG, |
| 3504 |
(errcode_for_file_access(), |
| 3505 |
errmsg("could not remove file \"%s\": %m", |
| 3506 |
rm_path))); |
| 3507 |
} |
| 3508 |
|
| 3509 |
FreeDir(dbspace_dir); |
| 3510 |
} |
| 3511 |
|
| 3512 |
/* t<digits>_<digits>, or t<digits>_<digits>_<forkname> */ |
| 3513 |
bool |
| 3514 |
looks_like_temp_rel_name(const char *name) |
| 3515 |
{ |
| 3516 |
int pos; |
| 3517 |
int savepos; |
| 3518 |
|
| 3519 |
/* Must start with "t". */ |
| 3520 |
if (name[0] != 't') |
| 3521 |
return false; |
| 3522 |
|
| 3523 |
/* Followed by a non-empty string of digits and then an underscore. */ |
| 3524 |
for (pos = 1; isdigit((unsigned char) name[pos]); ++pos) |
| 3525 |
; |
| 3526 |
if (pos == 1 || name[pos] != '_') |
| 3527 |
return false; |
| 3528 |
|
| 3529 |
/* Followed by another nonempty string of digits. */ |
| 3530 |
for (savepos = ++pos; isdigit((unsigned char) name[pos]); ++pos) |
| 3531 |
; |
| 3532 |
if (savepos == pos) |
| 3533 |
return false; |
| 3534 |
|
| 3535 |
/* We might have _forkname or .segment or both. */ |
| 3536 |
if (name[pos] == '_') |
| 3537 |
{ |
| 3538 |
int forkchar = forkname_chars(&name[pos + 1], NULL); |
| 3539 |
|
| 3540 |
if (forkchar <= 0) |
| 3541 |
return false; |
| 3542 |
pos += forkchar + 1; |
| 3543 |
} |
| 3544 |
if (name[pos] == '.') |
| 3545 |
{ |
| 3546 |
int segchar; |
| 3547 |
|
| 3548 |
for (segchar = 1; isdigit((unsigned char) name[pos + segchar]); ++segchar) |
| 3549 |
; |
| 3550 |
if (segchar <= 1) |
| 3551 |
return false; |
| 3552 |
pos += segchar; |
| 3553 |
} |
| 3554 |
|
| 3555 |
/* Now we should be at the end. */ |
| 3556 |
if (name[pos] != '\0') |
| 3557 |
return false; |
| 3558 |
return true; |
| 3559 |
} |
| 3560 |
|
| 3561 |
#ifdef HAVE_SYNCFS |
| 3562 |
static void |
| 3563 |
do_syncfs(const char *path) |
| 3564 |
{ |
| 3565 |
int fd; |
| 3566 |
|
| 3567 |
ereport_startup_progress("syncing data directory (syncfs), elapsed time: %ld.%02d s, current path: %s", |
| 3568 |
path); |
| 3569 |
|
| 3570 |
fd = OpenTransientFile(path, O_RDONLY); |
| 3571 |
if (fd < 0) |
| 3572 |
{ |
| 3573 |
ereport(LOG, |
| 3574 |
(errcode_for_file_access(), |
| 3575 |
errmsg("could not open file \"%s\": %m", path))); |
| 3576 |
return; |
| 3577 |
} |
| 3578 |
if (syncfs(fd) < 0) |
| 3579 |
ereport(LOG, |
| 3580 |
(errcode_for_file_access(), |
| 3581 |
errmsg("could not synchronize file system for file \"%s\": %m", path))); |
| 3582 |
CloseTransientFile(fd); |
| 3583 |
} |
| 3584 |
#endif |
| 3585 |
|
| 3586 |
/* |
| 3587 |
* Issue fsync recursively on PGDATA and all its contents, or issue syncfs for |
| 3588 |
* all potential filesystem, depending on recovery_init_sync_method setting. |
| 3589 |
* |
| 3590 |
* We fsync regular files and directories wherever they are, but we |
| 3591 |
* follow symlinks only for pg_wal and immediately under pg_tblspc. |
| 3592 |
* Other symlinks are presumed to point at files we're not responsible |
| 3593 |
* for fsyncing, and might not have privileges to write at all. |
| 3594 |
* |
| 3595 |
* Errors are logged but not considered fatal; that's because this is used |
| 3596 |
* only during database startup, to deal with the possibility that there are |
| 3597 |
* issued-but-unsynced writes pending against the data directory. We want to |
| 3598 |
* ensure that such writes reach disk before anything that's done in the new |
| 3599 |
* run. However, aborting on error would result in failure to start for |
| 3600 |
* harmless cases such as read-only files in the data directory, and that's |
| 3601 |
* not good either. |
| 3602 |
* |
| 3603 |
* Note that if we previously crashed due to a PANIC on fsync(), we'll be |
| 3604 |
* rewriting all changes again during recovery. |
| 3605 |
* |
| 3606 |
* Note we assume we're chdir'd into PGDATA to begin with. |
| 3607 |
*/ |
| 3608 |
void |
| 3609 |
SyncDataDirectory(void) |
| 3610 |
{ |
| 3611 |
bool xlog_is_symlink; |
| 3612 |
|
| 3613 |
/* We can skip this whole thing if fsync is disabled. */ |
| 3614 |
if (!enableFsync) |
| 3615 |
return; |
| 3616 |
|
| 3617 |
/* |
| 3618 |
* If pg_wal is a symlink, we'll need to recurse into it separately, |
| 3619 |
* because the first walkdir below will ignore it. |
| 3620 |
*/ |
| 3621 |
xlog_is_symlink = false; |
| 3622 |
|
| 3623 |
{ |
| 3624 |
struct stat st; |
| 3625 |
|
| 3626 |
if (lstat("pg_wal", &st) < 0) |
| 3627 |
ereport(LOG, |
| 3628 |
(errcode_for_file_access(), |
| 3629 |
errmsg("could not stat file \"%s\": %m", |
| 3630 |
"pg_wal"))); |
| 3631 |
else if (S_ISLNK(st.st_mode)) |
| 3632 |
xlog_is_symlink = true; |
| 3633 |
} |
| 3634 |
|
| 3635 |
#ifdef HAVE_SYNCFS |
| 3636 |
if (recovery_init_sync_method == DATA_DIR_SYNC_METHOD_SYNCFS) |
| 3637 |
{ |
| 3638 |
DIR *dir; |
| 3639 |
struct dirent *de; |
| 3640 |
|
| 3641 |
/* |
| 3642 |
* On Linux, we don't have to open every single file one by one. We |
| 3643 |
* can use syncfs() to sync whole filesystems. We only expect |
| 3644 |
* filesystem boundaries to exist where we tolerate symlinks, namely |
| 3645 |
* pg_wal and the tablespaces, so we call syncfs() for each of those |
| 3646 |
* directories. |
| 3647 |
*/ |
| 3648 |
|
| 3649 |
/* Prepare to report progress syncing the data directory via syncfs. */ |
| 3650 |
begin_startup_progress_phase(); |
| 3651 |
|
| 3652 |
/* Sync the top level pgdata directory. */ |
| 3653 |
do_syncfs("."); |
| 3654 |
/* If any tablespaces are configured, sync each of those. */ |
| 3655 |
dir = AllocateDir(PG_TBLSPC_DIR); |
| 3656 |
while ((de = ReadDirExtended(dir, PG_TBLSPC_DIR, LOG))) |
| 3657 |
{ |
| 3658 |
char path[MAXPGPATH]; |
| 3659 |
|
| 3660 |
if (strcmp(de->d_name, ".") == 0 || strcmp(de->d_name, "..") == 0) |
| 3661 |
continue; |
| 3662 |
|
| 3663 |
snprintf(path, MAXPGPATH, "%s/%s", PG_TBLSPC_DIR, de->d_name); |
| 3664 |
do_syncfs(path); |
| 3665 |
} |
| 3666 |
FreeDir(dir); |
| 3667 |
/* If pg_wal is a symlink, process that too. */ |
| 3668 |
if (xlog_is_symlink) |
| 3669 |
do_syncfs("pg_wal"); |
| 3670 |
return; |
| 3671 |
} |
| 3672 |
#endif /* !HAVE_SYNCFS */ |
| 3673 |
|
| 3674 |
#ifdef PG_FLUSH_DATA_WORKS |
| 3675 |
/* Prepare to report progress of the pre-fsync phase. */ |
| 3676 |
begin_startup_progress_phase(); |
| 3677 |
|
| 3678 |
/* |
| 3679 |
* If possible, hint to the kernel that we're soon going to fsync the data |
| 3680 |
* directory and its contents. Errors in this step are even less |
| 3681 |
* interesting than normal, so log them only at DEBUG1. |
| 3682 |
*/ |
| 3683 |
walkdir(".", pre_sync_fname, false, DEBUG1); |
| 3684 |
if (xlog_is_symlink) |
| 3685 |
walkdir("pg_wal", pre_sync_fname, false, DEBUG1); |
| 3686 |
walkdir(PG_TBLSPC_DIR, pre_sync_fname, true, DEBUG1); |
| 3687 |
#endif |
| 3688 |
|
| 3689 |
/* Prepare to report progress syncing the data directory via fsync. */ |
| 3690 |
begin_startup_progress_phase(); |
| 3691 |
|
| 3692 |
/* |
| 3693 |
* Now we do the fsync()s in the same order. |
| 3694 |
* |
| 3695 |
* The main call ignores symlinks, so in addition to specially processing |
| 3696 |
* pg_wal if it's a symlink, pg_tblspc has to be visited separately with |
| 3697 |
* process_symlinks = true. Note that if there are any plain directories |
| 3698 |
* in pg_tblspc, they'll get fsync'd twice. That's not an expected case |
| 3699 |
* so we don't worry about optimizing it. |
| 3700 |
*/ |
| 3701 |
walkdir(".", datadir_fsync_fname, false, LOG); |
| 3702 |
if (xlog_is_symlink) |
| 3703 |
walkdir("pg_wal", datadir_fsync_fname, false, LOG); |
| 3704 |
walkdir(PG_TBLSPC_DIR, datadir_fsync_fname, true, LOG); |
| 3705 |
} |
| 3706 |
|
| 3707 |
/* |
| 3708 |
* walkdir: recursively walk a directory, applying the action to each |
| 3709 |
* regular file and directory (including the named directory itself). |
| 3710 |
* |
| 3711 |
* If process_symlinks is true, the action and recursion are also applied |
| 3712 |
* to regular files and directories that are pointed to by symlinks in the |
| 3713 |
* given directory; otherwise symlinks are ignored. Symlinks are always |
| 3714 |
* ignored in subdirectories, ie we intentionally don't pass down the |
| 3715 |
* process_symlinks flag to recursive calls. |
| 3716 |
* |
| 3717 |
* Errors are reported at level elevel, which might be ERROR or less. |
| 3718 |
* |
| 3719 |
* See also walkdir in file_utils.c, which is a frontend version of this |
| 3720 |
* logic. |
| 3721 |
*/ |
| 3722 |
static void |
| 3723 |
walkdir(const char *path, |
| 3724 |
void (*action) (const char *fname, bool isdir, int elevel), |
| 3725 |
bool process_symlinks, |
| 3726 |
int elevel) |
| 3727 |
{ |
| 3728 |
DIR *dir; |
| 3729 |
struct dirent *de; |
| 3730 |
|
| 3731 |
dir = AllocateDir(path); |
| 3732 |
|
| 3733 |
while ((de = ReadDirExtended(dir, path, elevel)) != NULL) |
| 3734 |
{ |
| 3735 |
char subpath[MAXPGPATH * 2]; |
| 3736 |
|
| 3737 |
CHECK_FOR_INTERRUPTS(); |
| 3738 |
|
| 3739 |
if (strcmp(de->d_name, ".") == 0 || |
| 3740 |
strcmp(de->d_name, "..") == 0) |
| 3741 |
continue; |
| 3742 |
|
| 3743 |
snprintf(subpath, sizeof(subpath), "%s/%s", path, de->d_name); |
| 3744 |
|
| 3745 |
switch (get_dirent_type(subpath, de, process_symlinks, elevel)) |
| 3746 |
{ |
| 3747 |
case PGFILETYPE_REG: |
| 3748 |
(*action) (subpath, false, elevel); |
| 3749 |
break; |
| 3750 |
case PGFILETYPE_DIR: |
| 3751 |
walkdir(subpath, action, false, elevel); |
| 3752 |
break; |
| 3753 |
default: |
| 3754 |
|
| 3755 |
/* |
| 3756 |
* Errors are already reported directly by get_dirent_type(), |
| 3757 |
* and any remaining symlinks and unknown file types are |
| 3758 |
* ignored. |
| 3759 |
*/ |
| 3760 |
break; |
| 3761 |
} |
| 3762 |
} |
| 3763 |
|
| 3764 |
FreeDir(dir); /* we ignore any error here */ |
| 3765 |
|
| 3766 |
/* |
| 3767 |
* It's important to fsync the destination directory itself as individual |
| 3768 |
* file fsyncs don't guarantee that the directory entry for the file is |
| 3769 |
* synced. However, skip this if AllocateDir failed; the action function |
| 3770 |
* might not be robust against that. |
| 3771 |
*/ |
| 3772 |
if (dir) |
| 3773 |
(*action) (path, true, elevel); |
| 3774 |
} |
| 3775 |
|
| 3776 |
|
| 3777 |
/* |
| 3778 |
* Hint to the OS that it should get ready to fsync() this file. |
| 3779 |
* |
| 3780 |
* Ignores errors trying to open unreadable files, and logs other errors at a |
| 3781 |
* caller-specified level. |
| 3782 |
*/ |
| 3783 |
#ifdef PG_FLUSH_DATA_WORKS |
| 3784 |
|
| 3785 |
static void |
| 3786 |
pre_sync_fname(const char *fname, bool isdir, int elevel) |
| 3787 |
{ |
| 3788 |
int fd; |
| 3789 |
|
| 3790 |
/* Don't try to flush directories, it'll likely just fail */ |
| 3791 |
if (isdir) |
| 3792 |
return; |
| 3793 |
|
| 3794 |
ereport_startup_progress("syncing data directory (pre-fsync), elapsed time: %ld.%02d s, current path: %s", |
| 3795 |
fname); |
| 3796 |
|
| 3797 |
fd = OpenTransientFile(fname, O_RDONLY | PG_BINARY); |
| 3798 |
|
| 3799 |
if (fd < 0) |
| 3800 |
{ |
| 3801 |
if (errno == EACCES) |
| 3802 |
return; |
| 3803 |
ereport(elevel, |
| 3804 |
(errcode_for_file_access(), |
| 3805 |
errmsg("could not open file \"%s\": %m", fname))); |
| 3806 |
return; |
| 3807 |
} |
| 3808 |
|
| 3809 |
/* |
| 3810 |
* pg_flush_data() ignores errors, which is ok because this is only a |
| 3811 |
* hint. |
| 3812 |
*/ |
| 3813 |
pg_flush_data(fd, 0, 0); |
| 3814 |
|
| 3815 |
if (CloseTransientFile(fd) != 0) |
| 3816 |
ereport(elevel, |
| 3817 |
(errcode_for_file_access(), |
| 3818 |
errmsg("could not close file \"%s\": %m", fname))); |
| 3819 |
} |
| 3820 |
|
| 3821 |
#endif /* PG_FLUSH_DATA_WORKS */ |
| 3822 |
|
| 3823 |
static void |
| 3824 |
datadir_fsync_fname(const char *fname, bool isdir, int elevel) |
| 3825 |
{ |
| 3826 |
ereport_startup_progress("syncing data directory (fsync), elapsed time: %ld.%02d s, current path: %s", |
| 3827 |
fname); |
| 3828 |
|
| 3829 |
/* |
| 3830 |
* We want to silently ignoring errors about unreadable files. Pass that |
| 3831 |
* desire on to fsync_fname_ext(). |
| 3832 |
*/ |
| 3833 |
fsync_fname_ext(fname, isdir, true, elevel); |
| 3834 |
} |
| 3835 |
|
| 3836 |
static void |
| 3837 |
unlink_if_exists_fname(const char *fname, bool isdir, int elevel) |
| 3838 |
{ |
| 3839 |
if (isdir) |
| 3840 |
{ |
| 3841 |
if (rmdir(fname) != 0 && errno != ENOENT) |
| 3842 |
ereport(elevel, |
| 3843 |
(errcode_for_file_access(), |
| 3844 |
errmsg("could not remove directory \"%s\": %m", fname))); |
| 3845 |
} |
| 3846 |
else |
| 3847 |
{ |
| 3848 |
/* Use PathNameDeleteTemporaryFile to report filesize */ |
| 3849 |
PathNameDeleteTemporaryFile(fname, false); |
| 3850 |
} |
| 3851 |
} |
| 3852 |
|
| 3853 |
/* |
| 3854 |
* fsync_fname_ext -- Try to fsync a file or directory |
| 3855 |
* |
| 3856 |
* If ignore_perm is true, ignore errors upon trying to open unreadable |
| 3857 |
* files. Logs other errors at a caller-specified level. |
| 3858 |
* |
| 3859 |
* Returns 0 if the operation succeeded, -1 otherwise. |
| 3860 |
*/ |
| 3861 |
int |
| 3862 |
fsync_fname_ext(const char *fname, bool isdir, bool ignore_perm, int elevel) |
| 3863 |
{ |
| 3864 |
int fd; |
| 3865 |
int flags; |
| 3866 |
int returncode; |
| 3867 |
|
| 3868 |
/* |
| 3869 |
* Some OSs require directories to be opened read-only whereas other |
| 3870 |
* systems don't allow us to fsync files opened read-only; so we need both |
| 3871 |
* cases here. Using O_RDWR will cause us to fail to fsync files that are |
| 3872 |
* not writable by our userid, but we assume that's OK. |
| 3873 |
*/ |
| 3874 |
flags = PG_BINARY; |
| 3875 |
if (!isdir) |
| 3876 |
flags |= O_RDWR; |
| 3877 |
else |
| 3878 |
flags |= O_RDONLY; |
| 3879 |
|
| 3880 |
fd = OpenTransientFile(fname, flags); |
| 3881 |
|
| 3882 |
/* |
| 3883 |
* Some OSs don't allow us to open directories at all (Windows returns |
| 3884 |
* EACCES), just ignore the error in that case. If desired also silently |
| 3885 |
* ignoring errors about unreadable files. Log others. |
| 3886 |
*/ |
| 3887 |
if (fd < 0 && isdir && (errno == EISDIR || errno == EACCES)) |
| 3888 |
return 0; |
| 3889 |
else if (fd < 0 && ignore_perm && errno == EACCES) |
| 3890 |
return 0; |
| 3891 |
else if (fd < 0) |
| 3892 |
{ |
| 3893 |
ereport(elevel, |
| 3894 |
(errcode_for_file_access(), |
| 3895 |
errmsg("could not open file \"%s\": %m", fname))); |
| 3896 |
return -1; |
| 3897 |
} |
| 3898 |
|
| 3899 |
returncode = pg_fsync(fd); |
| 3900 |
|
| 3901 |
/* |
| 3902 |
* Some OSes don't allow us to fsync directories at all, so we can ignore |
| 3903 |
* those errors. Anything else needs to be logged. |
| 3904 |
*/ |
| 3905 |
if (returncode != 0 && !(isdir && (errno == EBADF || errno == EINVAL))) |
| 3906 |
{ |
| 3907 |
int save_errno; |
| 3908 |
|
| 3909 |
/* close file upon error, might not be in transaction context */ |
| 3910 |
save_errno = errno; |
| 3911 |
(void) CloseTransientFile(fd); |
| 3912 |
errno = save_errno; |
| 3913 |
|
| 3914 |
ereport(elevel, |
| 3915 |
(errcode_for_file_access(), |
| 3916 |
errmsg("could not fsync file \"%s\": %m", fname))); |
| 3917 |
return -1; |
| 3918 |
} |
| 3919 |
|
| 3920 |
if (CloseTransientFile(fd) != 0) |
| 3921 |
{ |
| 3922 |
ereport(elevel, |
| 3923 |
(errcode_for_file_access(), |
| 3924 |
errmsg("could not close file \"%s\": %m", fname))); |
| 3925 |
return -1; |
| 3926 |
} |
| 3927 |
|
| 3928 |
return 0; |
| 3929 |
} |
| 3930 |
|
| 3931 |
/* |
| 3932 |
* fsync_parent_path -- fsync the parent path of a file or directory |
| 3933 |
* |
| 3934 |
* This is aimed at making file operations persistent on disk in case of |
| 3935 |
* an OS crash or power failure. |
| 3936 |
*/ |
| 3937 |
static int |
| 3938 |
fsync_parent_path(const char *fname, int elevel) |
| 3939 |
{ |
| 3940 |
char parentpath[MAXPGPATH]; |
| 3941 |
|
| 3942 |
strlcpy(parentpath, fname, MAXPGPATH); |
| 3943 |
get_parent_directory(parentpath); |
| 3944 |
|
| 3945 |
/* |
| 3946 |
* get_parent_directory() returns an empty string if the input argument is |
| 3947 |
* just a file name (see comments in path.c), so handle that as being the |
| 3948 |
* current directory. |
| 3949 |
*/ |
| 3950 |
if (strlen(parentpath) == 0) |
| 3951 |
strlcpy(parentpath, ".", MAXPGPATH); |
| 3952 |
|
| 3953 |
if (fsync_fname_ext(parentpath, true, false, elevel) != 0) |
| 3954 |
return -1; |
| 3955 |
|
| 3956 |
return 0; |
| 3957 |
} |
| 3958 |
|
| 3959 |
/* |
| 3960 |
* Create a PostgreSQL data sub-directory |
| 3961 |
* |
| 3962 |
* The data directory itself, and most of its sub-directories, are created at |
| 3963 |
* initdb time, but we do have some occasions when we create directories in |
| 3964 |
* the backend (CREATE TABLESPACE, for example). In those cases, we want to |
| 3965 |
* make sure that those directories are created consistently. Today, that means |
| 3966 |
* making sure that the created directory has the correct permissions, which is |
| 3967 |
* what pg_dir_create_mode tracks for us. |
| 3968 |
* |
| 3969 |
* Note that we also set the umask() based on what we understand the correct |
| 3970 |
* permissions to be (see file_perm.c). |
| 3971 |
* |
| 3972 |
* For permissions other than the default, mkdir() can be used directly, but |
| 3973 |
* be sure to consider carefully such cases -- a sub-directory with incorrect |
| 3974 |
* permissions in a PostgreSQL data directory could cause backups and other |
| 3975 |
* processes to fail. |
| 3976 |
*/ |
| 3977 |
int |
| 3978 |
MakePGDirectory(const char *directoryName) |
| 3979 |
{ |
| 3980 |
return mkdir(directoryName, pg_dir_create_mode); |
| 3981 |
} |
| 3982 |
|
| 3983 |
/* |
| 3984 |
* Return the passed-in error level, or PANIC if data_sync_retry is off. |
| 3985 |
* |
| 3986 |
* Failure to fsync any data file is cause for immediate panic, unless |
| 3987 |
* data_sync_retry is enabled. Data may have been written to the operating |
| 3988 |
* system and removed from our buffer pool already, and if we are running on |
| 3989 |
* an operating system that forgets dirty data on write-back failure, there |
| 3990 |
* may be only one copy of the data remaining: in the WAL. A later attempt to |
| 3991 |
* fsync again might falsely report success. Therefore we must not allow any |
| 3992 |
* further checkpoints to be attempted. data_sync_retry can in theory be |
| 3993 |
* enabled on systems known not to drop dirty buffered data on write-back |
| 3994 |
* failure (with the likely outcome that checkpoints will continue to fail |
| 3995 |
* until the underlying problem is fixed). |
| 3996 |
* |
| 3997 |
* Any code that reports a failure from fsync() or related functions should |
| 3998 |
* filter the error level with this function. |
| 3999 |
*/ |
| 4000 |
int |
| 4001 |
data_sync_elevel(int elevel) |
| 4002 |
{ |
| 4003 |
return data_sync_retry ? elevel : PANIC; |
| 4004 |
} |
| 4005 |
|
| 4006 |
bool |
| 4007 |
check_debug_io_direct(char **newval, void **extra, GucSource source) |
| 4008 |
{ |
| 4009 |
bool result = true; |
| 4010 |
int flags; |
| 4011 |
|
| 4012 |
#if PG_O_DIRECT == 0 |
| 4013 |
if (strcmp(*newval, "") != 0) |
| 4014 |
{ |
| 4015 |
GUC_check_errdetail("\"%s\" is not supported on this platform.", |
| 4016 |
"debug_io_direct"); |
| 4017 |
result = false; |
| 4018 |
} |
| 4019 |
flags = 0; |
| 4020 |
#else |
| 4021 |
List *elemlist; |
| 4022 |
ListCell *l; |
| 4023 |
char *rawstring; |
| 4024 |
|
| 4025 |
/* Need a modifiable copy of string */ |
| 4026 |
rawstring = pstrdup(*newval); |
| 4027 |
|
| 4028 |
if (!SplitGUCList(rawstring, ',', &elemlist)) |
| 4029 |
{ |
| 4030 |
GUC_check_errdetail("Invalid list syntax in parameter \"%s\".", |
| 4031 |
"debug_io_direct"); |
| 4032 |
pfree(rawstring); |
| 4033 |
list_free(elemlist); |
| 4034 |
return false; |
| 4035 |
} |
| 4036 |
|
| 4037 |
flags = 0; |
| 4038 |
foreach(l, elemlist) |
| 4039 |
{ |
| 4040 |
char *item = (char *) lfirst(l); |
| 4041 |
|
| 4042 |
if (pg_strcasecmp(item, "data") == 0) |
| 4043 |
flags |= IO_DIRECT_DATA; |
| 4044 |
else if (pg_strcasecmp(item, "wal") == 0) |
| 4045 |
flags |= IO_DIRECT_WAL; |
| 4046 |
else if (pg_strcasecmp(item, "wal_init") == 0) |
| 4047 |
flags |= IO_DIRECT_WAL_INIT; |
| 4048 |
else |
| 4049 |
{ |
| 4050 |
GUC_check_errdetail("Invalid option \"%s\".", item); |
| 4051 |
result = false; |
| 4052 |
break; |
| 4053 |
} |
| 4054 |
} |
| 4055 |
|
| 4056 |
/* |
| 4057 |
* It's possible to configure block sizes smaller than our assumed I/O |
| 4058 |
* alignment size, which could result in invalid I/O requests. |
| 4059 |
*/ |
| 4060 |
#if XLOG_BLCKSZ < PG_IO_ALIGN_SIZE |
| 4061 |
if (result && (flags & (IO_DIRECT_WAL | IO_DIRECT_WAL_INIT))) |
| 4062 |
{ |
| 4063 |
GUC_check_errdetail("\"%s\" is not supported for WAL because %s is too small.", |
| 4064 |
"debug_io_direct", "XLOG_BLCKSZ"); |
| 4065 |
result = false; |
| 4066 |
} |
| 4067 |
#endif |
| 4068 |
#if BLCKSZ < PG_IO_ALIGN_SIZE |
| 4069 |
if (result && (flags & IO_DIRECT_DATA)) |
| 4070 |
{ |
| 4071 |
GUC_check_errdetail("\"%s\" is not supported for data because %s is too small.", |
| 4072 |
"debug_io_direct", "BLCKSZ"); |
| 4073 |
result = false; |
| 4074 |
} |
| 4075 |
#endif |
| 4076 |
|
| 4077 |
pfree(rawstring); |
| 4078 |
list_free(elemlist); |
| 4079 |
#endif |
| 4080 |
|
| 4081 |
if (!result) |
| 4082 |
return result; |
| 4083 |
|
| 4084 |
/* Save the flags in *extra, for use by assign_debug_io_direct */ |
| 4085 |
*extra = guc_malloc(LOG, sizeof(int)); |
| 4086 |
if (!*extra) |
| 4087 |
return false; |
| 4088 |
*((int *) *extra) = flags; |
| 4089 |
|
| 4090 |
return result; |
| 4091 |
} |
| 4092 |
|
| 4093 |
void |
| 4094 |
assign_debug_io_direct(const char *newval, void *extra) |
| 4095 |
{ |
| 4096 |
int *flags = (int *) extra; |
| 4097 |
|
| 4098 |
io_direct_flags = *flags; |
| 4099 |
} |
| 4100 |
|
| 4101 |
/* ResourceOwner callbacks */ |
| 4102 |
|
| 4103 |
static void |
| 4104 |
ResOwnerReleaseFile(Datum res) |
| 4105 |
{ |
| 4106 |
File file = (File) DatumGetInt32(res); |
| 4107 |
Vfd *vfdP; |
| 4108 |
|
| 4109 |
Assert(FileIsValid(file)); |
| 4110 |
|
| 4111 |
vfdP = &VfdCache[file]; |
| 4112 |
vfdP->resowner = NULL; |
| 4113 |
|
| 4114 |
FileClose(file); |
| 4115 |
} |
| 4116 |
|
| 4117 |
static char * |
| 4118 |
ResOwnerPrintFile(Datum res) |
| 4119 |
{ |
| 4120 |
return psprintf("File %d", DatumGetInt32(res)); |
| 4121 |
} |