Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * fd.c
4 : * Virtual file descriptor code.
5 : *
6 : * Portions Copyright (c) 1996-2024, 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 "portability/mem.h"
97 : #include "postmaster/startup.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 : /* Which kinds of files should be opened with PG_O_DIRECT. */
168 : int io_direct_flags;
169 :
170 : /* Debugging.... */
171 :
172 : #ifdef FDDEBUG
173 : #define DO_DB(A) \
174 : do { \
175 : int _do_db_save_errno = errno; \
176 : A; \
177 : errno = _do_db_save_errno; \
178 : } while (0)
179 : #else
180 : #define DO_DB(A) \
181 : ((void) 0)
182 : #endif
183 :
184 : #define VFD_CLOSED (-1)
185 :
186 : #define FileIsValid(file) \
187 : ((file) > 0 && (file) < (int) SizeVfdCache && VfdCache[file].fileName != NULL)
188 :
189 : #define FileIsNotOpen(file) (VfdCache[file].fd == VFD_CLOSED)
190 :
191 : /* these are the assigned bits in fdstate below: */
192 : #define FD_DELETE_AT_CLOSE (1 << 0) /* T = delete when closed */
193 : #define FD_CLOSE_AT_EOXACT (1 << 1) /* T = close at eoXact */
194 : #define FD_TEMP_FILE_LIMIT (1 << 2) /* T = respect temp_file_limit */
195 :
196 : typedef struct vfd
197 : {
198 : int fd; /* current FD, or VFD_CLOSED if none */
199 : unsigned short fdstate; /* bitflags for VFD's state */
200 : ResourceOwner resowner; /* owner, for automatic cleanup */
201 : File nextFree; /* link to next free VFD, if in freelist */
202 : File lruMoreRecently; /* doubly linked recency-of-use list */
203 : File lruLessRecently;
204 : off_t fileSize; /* current size of file (0 if not temporary) */
205 : char *fileName; /* name of file, or NULL for unused VFD */
206 : /* NB: fileName is malloc'd, and must be free'd when closing the VFD */
207 : int fileFlags; /* open(2) flags for (re)opening the file */
208 : mode_t fileMode; /* mode to pass to open(2) */
209 : } Vfd;
210 :
211 : /*
212 : * Virtual File Descriptor array pointer and size. This grows as
213 : * needed. 'File' values are indexes into this array.
214 : * Note that VfdCache[0] is not a usable VFD, just a list header.
215 : */
216 : static Vfd *VfdCache;
217 : static Size SizeVfdCache = 0;
218 :
219 : /*
220 : * Number of file descriptors known to be in use by VFD entries.
221 : */
222 : static int nfile = 0;
223 :
224 : /*
225 : * Flag to tell whether it's worth scanning VfdCache looking for temp files
226 : * to close
227 : */
228 : static bool have_xact_temporary_files = false;
229 :
230 : /*
231 : * Tracks the total size of all temporary files. Note: when temp_file_limit
232 : * is being enforced, this cannot overflow since the limit cannot be more
233 : * than INT_MAX kilobytes. When not enforcing, it could theoretically
234 : * overflow, but we don't care.
235 : */
236 : static uint64 temporary_files_size = 0;
237 :
238 : /* Temporary file access initialized and not yet shut down? */
239 : #ifdef USE_ASSERT_CHECKING
240 : static bool temporary_files_allowed = false;
241 : #endif
242 :
243 : /*
244 : * List of OS handles opened with AllocateFile, AllocateDir and
245 : * OpenTransientFile.
246 : */
247 : typedef enum
248 : {
249 : AllocateDescFile,
250 : AllocateDescPipe,
251 : AllocateDescDir,
252 : AllocateDescRawFD,
253 : } AllocateDescKind;
254 :
255 : typedef struct
256 : {
257 : AllocateDescKind kind;
258 : SubTransactionId create_subid;
259 : union
260 : {
261 : FILE *file;
262 : DIR *dir;
263 : int fd;
264 : } desc;
265 : } AllocateDesc;
266 :
267 : static int numAllocatedDescs = 0;
268 : static int maxAllocatedDescs = 0;
269 : static AllocateDesc *allocatedDescs = NULL;
270 :
271 : /*
272 : * Number of open "external" FDs reported to Reserve/ReleaseExternalFD.
273 : */
274 : static int numExternalFDs = 0;
275 :
276 : /*
277 : * Number of temporary files opened during the current session;
278 : * this is used in generation of tempfile names.
279 : */
280 : static long tempFileCounter = 0;
281 :
282 : /*
283 : * Array of OIDs of temp tablespaces. (Some entries may be InvalidOid,
284 : * indicating that the current database's default tablespace should be used.)
285 : * When numTempTableSpaces is -1, this has not been set in the current
286 : * transaction.
287 : */
288 : static Oid *tempTableSpaces = NULL;
289 : static int numTempTableSpaces = -1;
290 : static int nextTempTableSpace = 0;
291 :
292 :
293 : /*--------------------
294 : *
295 : * Private Routines
296 : *
297 : * Delete - delete a file from the Lru ring
298 : * LruDelete - remove a file from the Lru ring and close its FD
299 : * Insert - put a file at the front of the Lru ring
300 : * LruInsert - put a file at the front of the Lru ring and open it
301 : * ReleaseLruFile - Release an fd by closing the last entry in the Lru ring
302 : * ReleaseLruFiles - Release fd(s) until we're under the max_safe_fds limit
303 : * AllocateVfd - grab a free (or new) file record (from VfdCache)
304 : * FreeVfd - free a file record
305 : *
306 : * The Least Recently Used ring is a doubly linked list that begins and
307 : * ends on element zero. Element zero is special -- it doesn't represent
308 : * a file and its "fd" field always == VFD_CLOSED. Element zero is just an
309 : * anchor that shows us the beginning/end of the ring.
310 : * Only VFD elements that are currently really open (have an FD assigned) are
311 : * in the Lru ring. Elements that are "virtually" open can be recognized
312 : * by having a non-null fileName field.
313 : *
314 : * example:
315 : *
316 : * /--less----\ /---------\
317 : * v \ v \
318 : * #0 --more---> LeastRecentlyUsed --more-\ \
319 : * ^\ | |
320 : * \\less--> MostRecentlyUsedFile <---/ |
321 : * \more---/ \--less--/
322 : *
323 : *--------------------
324 : */
325 : static void Delete(File file);
326 : static void LruDelete(File file);
327 : static void Insert(File file);
328 : static int LruInsert(File file);
329 : static bool ReleaseLruFile(void);
330 : static void ReleaseLruFiles(void);
331 : static File AllocateVfd(void);
332 : static void FreeVfd(File file);
333 :
334 : static int FileAccess(File file);
335 : static File OpenTemporaryFileInTablespace(Oid tblspcOid, bool rejectError);
336 : static bool reserveAllocatedDesc(void);
337 : static int FreeDesc(AllocateDesc *desc);
338 :
339 : static void BeforeShmemExit_Files(int code, Datum arg);
340 : static void CleanupTempFiles(bool isCommit, bool isProcExit);
341 : static void RemovePgTempRelationFiles(const char *tsdirname);
342 : static void RemovePgTempRelationFilesInDbspace(const char *dbspacedirname);
343 :
344 : static void walkdir(const char *path,
345 : void (*action) (const char *fname, bool isdir, int elevel),
346 : bool process_symlinks,
347 : int elevel);
348 : #ifdef PG_FLUSH_DATA_WORKS
349 : static void pre_sync_fname(const char *fname, bool isdir, int elevel);
350 : #endif
351 : static void datadir_fsync_fname(const char *fname, bool isdir, int elevel);
352 : static void unlink_if_exists_fname(const char *fname, bool isdir, int elevel);
353 :
354 : static int fsync_parent_path(const char *fname, int elevel);
355 :
356 :
357 : /* ResourceOwner callbacks to hold virtual file descriptors */
358 : static void ResOwnerReleaseFile(Datum res);
359 : static char *ResOwnerPrintFile(Datum res);
360 :
361 : static const ResourceOwnerDesc file_resowner_desc =
362 : {
363 : .name = "File",
364 : .release_phase = RESOURCE_RELEASE_AFTER_LOCKS,
365 : .release_priority = RELEASE_PRIO_FILES,
366 : .ReleaseResource = ResOwnerReleaseFile,
367 : .DebugPrint = ResOwnerPrintFile
368 : };
369 :
370 : /* Convenience wrappers over ResourceOwnerRemember/Forget */
371 : static inline void
372 8396 : ResourceOwnerRememberFile(ResourceOwner owner, File file)
373 : {
374 8396 : ResourceOwnerRemember(owner, Int32GetDatum(file), &file_resowner_desc);
375 8396 : }
376 : static inline void
377 8388 : ResourceOwnerForgetFile(ResourceOwner owner, File file)
378 : {
379 8388 : ResourceOwnerForget(owner, Int32GetDatum(file), &file_resowner_desc);
380 8388 : }
381 :
382 : /*
383 : * pg_fsync --- do fsync with or without writethrough
384 : */
385 : int
386 82354 : pg_fsync(int fd)
387 : {
388 : #if !defined(WIN32) && defined(USE_ASSERT_CHECKING)
389 : struct stat st;
390 :
391 : /*
392 : * Some operating system implementations of fsync() have requirements
393 : * about the file access modes that were used when their file descriptor
394 : * argument was opened, and these requirements differ depending on whether
395 : * the file descriptor is for a directory.
396 : *
397 : * For any file descriptor that may eventually be handed to fsync(), we
398 : * should have opened it with access modes that are compatible with
399 : * fsync() on all supported systems, otherwise the code may not be
400 : * portable, even if it runs ok on the current system.
401 : *
402 : * We assert here that a descriptor for a file was opened with write
403 : * permissions (either O_RDWR or O_WRONLY) and for a directory without
404 : * write permissions (O_RDONLY).
405 : *
406 : * Ignore any fstat errors and let the follow-up fsync() do its work.
407 : * Doing this sanity check here counts for the case where fsync() is
408 : * disabled.
409 : */
410 : if (fstat(fd, &st) == 0)
411 : {
412 : int desc_flags = fcntl(fd, F_GETFL);
413 :
414 : /*
415 : * O_RDONLY is historically 0, so just make sure that for directories
416 : * no write flags are used.
417 : */
418 : if (S_ISDIR(st.st_mode))
419 : Assert((desc_flags & (O_RDWR | O_WRONLY)) == 0);
420 : else
421 : Assert((desc_flags & (O_RDWR | O_WRONLY)) != 0);
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 82354 : 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 82354 : pg_fsync_no_writethrough(int fd)
442 : {
443 : int rc;
444 :
445 82354 : if (!enableFsync)
446 82354 : return 0;
447 :
448 0 : retry:
449 0 : rc = fsync(fd);
450 :
451 0 : if (rc == -1 && errno == EINTR)
452 0 : goto retry;
453 :
454 0 : return rc;
455 : }
456 :
457 : /*
458 : * pg_fsync_writethrough
459 : */
460 : int
461 0 : pg_fsync_writethrough(int fd)
462 : {
463 0 : if (enableFsync)
464 : {
465 : #if defined(F_FULLFSYNC)
466 : return (fcntl(fd, F_FULLFSYNC, 0) == -1) ? -1 : 0;
467 : #else
468 0 : errno = ENOSYS;
469 0 : return -1;
470 : #endif
471 : }
472 : else
473 0 : return 0;
474 : }
475 :
476 : /*
477 : * pg_fdatasync --- same as fdatasync except does nothing if enableFsync is off
478 : */
479 : int
480 0 : pg_fdatasync(int fd)
481 : {
482 : int rc;
483 :
484 0 : if (!enableFsync)
485 0 : return 0;
486 :
487 0 : retry:
488 0 : rc = fdatasync(fd);
489 :
490 0 : if (rc == -1 && errno == EINTR)
491 0 : goto retry;
492 :
493 0 : 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 30816 : pg_file_exists(const char *name)
504 : {
505 : struct stat st;
506 :
507 : Assert(name != NULL);
508 :
509 30816 : if (stat(name, &st) == 0)
510 16370 : return !S_ISDIR(st.st_mode);
511 14446 : else if (!(errno == ENOENT || errno == ENOTDIR || errno == EACCES))
512 0 : ereport(ERROR,
513 : (errcode_for_file_access(),
514 : errmsg("could not access file \"%s\": %m", name)));
515 :
516 14446 : 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 180036 : 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 180036 : if (!enableFsync)
534 180036 : 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 0 : if (not_implemented_by_kernel)
546 0 : return;
547 :
548 0 : 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 0 : rc = sync_file_range(fd, offset, nbytes,
560 : SYNC_FILE_RANGE_WRITE);
561 0 : if (rc != 0)
562 : {
563 : int elevel;
564 :
565 0 : if (rc == EINTR)
566 0 : 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 0 : if (errno == ENOSYS)
574 : {
575 0 : elevel = WARNING;
576 0 : not_implemented_by_kernel = true;
577 : }
578 : else
579 0 : elevel = data_sync_elevel(WARNING);
580 :
581 0 : ereport(elevel,
582 : (errcode_for_file_access(),
583 : errmsg("could not flush dirty data: %m")));
584 : }
585 :
586 0 : 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
703 918 : pg_ftruncate(int fd, off_t length)
704 : {
705 : int ret;
706 :
707 918 : retry:
708 918 : ret = ftruncate(fd, length);
709 :
710 918 : if (ret == -1 && errno == EINTR)
711 0 : goto retry;
712 :
713 918 : return ret;
714 : }
715 :
716 : /*
717 : * Truncate a file to a given length by name.
718 : */
719 : int
720 401376 : 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 401376 : retry:
740 401376 : ret = truncate(path, length);
741 :
742 401376 : if (ret == -1 && errno == EINTR)
743 0 : goto retry;
744 : #endif
745 :
746 401376 : 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 28132 : fsync_fname(const char *fname, bool isdir)
757 : {
758 28132 : fsync_fname_ext(fname, isdir, false, data_sync_elevel(ERROR));
759 28132 : }
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 5192 : 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 5192 : if (fsync_fname_ext(oldfile, false, false, elevel) != 0)
794 0 : return -1;
795 :
796 5192 : fd = OpenTransientFile(newfile, PG_BINARY | O_RDWR);
797 5192 : if (fd < 0)
798 : {
799 3112 : if (errno != ENOENT)
800 : {
801 0 : ereport(elevel,
802 : (errcode_for_file_access(),
803 : errmsg("could not open file \"%s\": %m", newfile)));
804 0 : return -1;
805 : }
806 : }
807 : else
808 : {
809 2080 : if (pg_fsync(fd) != 0)
810 : {
811 : int save_errno;
812 :
813 : /* close file upon error, might not be in transaction context */
814 0 : save_errno = errno;
815 0 : CloseTransientFile(fd);
816 0 : errno = save_errno;
817 :
818 0 : ereport(elevel,
819 : (errcode_for_file_access(),
820 : errmsg("could not fsync file \"%s\": %m", newfile)));
821 0 : return -1;
822 : }
823 :
824 2080 : if (CloseTransientFile(fd) != 0)
825 : {
826 0 : ereport(elevel,
827 : (errcode_for_file_access(),
828 : errmsg("could not close file \"%s\": %m", newfile)));
829 0 : return -1;
830 : }
831 : }
832 :
833 : /* Time to do the real deal... */
834 5192 : if (rename(oldfile, newfile) < 0)
835 : {
836 0 : ereport(elevel,
837 : (errcode_for_file_access(),
838 : errmsg("could not rename file \"%s\" to \"%s\": %m",
839 : oldfile, newfile)));
840 0 : 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 5192 : if (fsync_fname_ext(newfile, false, false, elevel) != 0)
848 0 : return -1;
849 :
850 5192 : if (fsync_parent_path(newfile, elevel) != 0)
851 0 : return -1;
852 :
853 5192 : 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
872 254 : durable_unlink(const char *fname, int elevel)
873 : {
874 254 : if (unlink(fname) < 0)
875 : {
876 74 : ereport(elevel,
877 : (errcode_for_file_access(),
878 : errmsg("could not remove file \"%s\": %m",
879 : fname)));
880 74 : return -1;
881 : }
882 :
883 : /*
884 : * To guarantee that the removal of the file is persistent, fsync its
885 : * parent directory.
886 : */
887 180 : if (fsync_parent_path(fname, elevel) != 0)
888 0 : return -1;
889 :
890 180 : 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 32562 : InitFileAccess(void)
904 : {
905 : Assert(SizeVfdCache == 0); /* call me only once */
906 :
907 : /* initialize cache header entry */
908 32562 : VfdCache = (Vfd *) malloc(sizeof(Vfd));
909 32562 : if (VfdCache == NULL)
910 0 : ereport(FATAL,
911 : (errcode(ERRCODE_OUT_OF_MEMORY),
912 : errmsg("out of memory")));
913 :
914 260496 : MemSet((char *) &(VfdCache[0]), 0, sizeof(Vfd));
915 32562 : VfdCache->fd = VFD_CLOSED;
916 :
917 32562 : SizeVfdCache = 1;
918 32562 : }
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 32562 : 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 32562 : before_shmem_exit(BeforeShmemExit_Files, 0);
943 :
944 : #ifdef USE_ASSERT_CHECKING
945 : temporary_files_allowed = true;
946 : #endif
947 32562 : }
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 1422 : count_usable_fds(int max_to_probe, int *usable_fds, int *already_open)
965 : {
966 : int *fd;
967 : int size;
968 1422 : int used = 0;
969 1422 : int highestfd = 0;
970 : int j;
971 :
972 : #ifdef HAVE_GETRLIMIT
973 : struct rlimit rlim;
974 : int getrlimit_status;
975 : #endif
976 :
977 1422 : size = 1024;
978 1422 : fd = (int *) palloc(size * sizeof(int));
979 :
980 : #ifdef HAVE_GETRLIMIT
981 1422 : getrlimit_status = getrlimit(RLIMIT_NOFILE, &rlim);
982 1422 : if (getrlimit_status != 0)
983 0 : ereport(WARNING, (errmsg("getrlimit failed: %m")));
984 : #endif /* HAVE_GETRLIMIT */
985 :
986 : /* dup until failure or probe limit reached */
987 : for (;;)
988 1420578 : {
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 1422000 : if (getrlimit_status == 0 && highestfd >= rlim.rlim_cur - 1)
998 0 : break;
999 : #endif
1000 :
1001 1422000 : thisfd = dup(2);
1002 1422000 : if (thisfd < 0)
1003 : {
1004 : /* Expect EMFILE or ENFILE, else it's fishy */
1005 0 : if (errno != EMFILE && errno != ENFILE)
1006 0 : elog(WARNING, "duplicating stderr file descriptor failed after %d successes: %m", used);
1007 0 : break;
1008 : }
1009 :
1010 1422000 : if (used >= size)
1011 : {
1012 0 : size *= 2;
1013 0 : fd = (int *) repalloc(fd, size * sizeof(int));
1014 : }
1015 1422000 : fd[used++] = thisfd;
1016 :
1017 1422000 : if (highestfd < thisfd)
1018 1422000 : highestfd = thisfd;
1019 :
1020 1422000 : if (used >= max_to_probe)
1021 1422 : break;
1022 : }
1023 :
1024 : /* release the files we opened */
1025 1423422 : for (j = 0; j < used; j++)
1026 1422000 : close(fd[j]);
1027 :
1028 1422 : 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 1422 : *usable_fds = used;
1036 1422 : *already_open = highestfd + 1 - used;
1037 1422 : }
1038 :
1039 : /*
1040 : * set_max_safe_fds
1041 : * Determine number of file descriptors that fd.c is allowed to use
1042 : */
1043 : void
1044 1422 : 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 - already_open)
1052 : * less the slop factor for files that are opened without consulting
1053 : * fd.c. This ensures that we won't exceed either max_files_per_process
1054 : * or the experimentally-determined EMFILE limit.
1055 : *----------
1056 : */
1057 1422 : count_usable_fds(max_files_per_process,
1058 : &usable_fds, &already_open);
1059 :
1060 1422 : max_safe_fds = Min(usable_fds, max_files_per_process - already_open);
1061 :
1062 : /*
1063 : * Take off the FDs reserved for system() etc.
1064 : */
1065 1422 : max_safe_fds -= NUM_RESERVED_FDS;
1066 :
1067 : /*
1068 : * Make sure we still have enough to get by.
1069 : */
1070 1422 : if (max_safe_fds < FD_MINFREE)
1071 0 : ereport(FATAL,
1072 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
1073 : errmsg("insufficient file descriptors available to start server process"),
1074 : errdetail("System allows %d, server needs at least %d.",
1075 : max_safe_fds + NUM_RESERVED_FDS,
1076 : FD_MINFREE + NUM_RESERVED_FDS)));
1077 :
1078 1422 : elog(DEBUG2, "max_safe_fds = %d, usable_fds = %d, already_open = %d",
1079 : max_safe_fds, usable_fds, already_open);
1080 1422 : }
1081 :
1082 : /*
1083 : * Open a file with BasicOpenFilePerm() and pass default file mode for the
1084 : * fileMode parameter.
1085 : */
1086 : int
1087 48876 : BasicOpenFile(const char *fileName, int fileFlags)
1088 : {
1089 48876 : return BasicOpenFilePerm(fileName, fileFlags, pg_file_create_mode);
1090 : }
1091 :
1092 : /*
1093 : * BasicOpenFilePerm --- same as open(2) except can free other FDs if needed
1094 : *
1095 : * This is exported for use by places that really want a plain kernel FD,
1096 : * but need to be proof against running out of FDs. Once an FD has been
1097 : * successfully returned, it is the caller's responsibility to ensure that
1098 : * it will not be leaked on ereport()! Most users should *not* call this
1099 : * routine directly, but instead use the VFD abstraction level, which
1100 : * provides protection against descriptor leaks as well as management of
1101 : * files that need to be open for more than a short period of time.
1102 : *
1103 : * Ideally this should be the *only* direct call of open() in the backend.
1104 : * In practice, the postmaster calls open() directly, and there are some
1105 : * direct open() calls done early in backend startup. Those are OK since
1106 : * this module wouldn't have any open files to close at that point anyway.
1107 : */
1108 : int
1109 2758528 : BasicOpenFilePerm(const char *fileName, int fileFlags, mode_t fileMode)
1110 : {
1111 : int fd;
1112 :
1113 2758528 : tryAgain:
1114 : #ifdef PG_O_DIRECT_USE_F_NOCACHE
1115 :
1116 : /*
1117 : * The value we defined to stand in for O_DIRECT when simulating it with
1118 : * F_NOCACHE had better not collide with any of the standard flags.
1119 : */
1120 : StaticAssertStmt((PG_O_DIRECT &
1121 : (O_APPEND |
1122 : O_CLOEXEC |
1123 : O_CREAT |
1124 : O_DSYNC |
1125 : O_EXCL |
1126 : O_RDWR |
1127 : O_RDONLY |
1128 : O_SYNC |
1129 : O_TRUNC |
1130 : O_WRONLY)) == 0,
1131 : "PG_O_DIRECT value collides with standard flag");
1132 : fd = open(fileName, fileFlags & ~PG_O_DIRECT, fileMode);
1133 : #else
1134 2758528 : fd = open(fileName, fileFlags, fileMode);
1135 : #endif
1136 :
1137 2758528 : if (fd >= 0)
1138 : {
1139 : #ifdef PG_O_DIRECT_USE_F_NOCACHE
1140 : if (fileFlags & PG_O_DIRECT)
1141 : {
1142 : if (fcntl(fd, F_NOCACHE, 1) < 0)
1143 : {
1144 : int save_errno = errno;
1145 :
1146 : close(fd);
1147 : errno = save_errno;
1148 : return -1;
1149 : }
1150 : }
1151 : #endif
1152 :
1153 2068538 : return fd; /* success! */
1154 : }
1155 :
1156 689990 : if (errno == EMFILE || errno == ENFILE)
1157 : {
1158 0 : int save_errno = errno;
1159 :
1160 0 : ereport(LOG,
1161 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
1162 : errmsg("out of file descriptors: %m; release and retry")));
1163 0 : errno = 0;
1164 0 : if (ReleaseLruFile())
1165 0 : goto tryAgain;
1166 0 : errno = save_errno;
1167 : }
1168 :
1169 689990 : return -1; /* failure */
1170 : }
1171 :
1172 : /*
1173 : * AcquireExternalFD - attempt to reserve an external file descriptor
1174 : *
1175 : * This should be used by callers that need to hold a file descriptor open
1176 : * over more than a short interval, but cannot use any of the other facilities
1177 : * provided by this module.
1178 : *
1179 : * The difference between this and the underlying ReserveExternalFD function
1180 : * is that this will report failure (by setting errno and returning false)
1181 : * if "too many" external FDs are already reserved. This should be used in
1182 : * any code where the total number of FDs to be reserved is not predictable
1183 : * and small.
1184 : */
1185 : bool
1186 223536 : AcquireExternalFD(void)
1187 : {
1188 : /*
1189 : * We don't want more than max_safe_fds / 3 FDs to be consumed for
1190 : * "external" FDs.
1191 : */
1192 223536 : if (numExternalFDs < max_safe_fds / 3)
1193 : {
1194 223536 : ReserveExternalFD();
1195 223536 : return true;
1196 : }
1197 0 : errno = EMFILE;
1198 0 : return false;
1199 : }
1200 :
1201 : /*
1202 : * ReserveExternalFD - report external consumption of a file descriptor
1203 : *
1204 : * This should be used by callers that need to hold a file descriptor open
1205 : * over more than a short interval, but cannot use any of the other facilities
1206 : * provided by this module. This just tracks the use of the FD and closes
1207 : * VFDs if needed to ensure we keep NUM_RESERVED_FDS FDs available.
1208 : *
1209 : * Call this directly only in code where failure to reserve the FD would be
1210 : * fatal; for example, the WAL-writing code does so, since the alternative is
1211 : * session failure. Also, it's very unwise to do so in code that could
1212 : * consume more than one FD per process.
1213 : *
1214 : * Note: as long as everybody plays nice so that NUM_RESERVED_FDS FDs remain
1215 : * available, it doesn't matter too much whether this is called before or
1216 : * after actually opening the FD; but doing so beforehand reduces the risk of
1217 : * an EMFILE failure if not everybody played nice. In any case, it's solely
1218 : * caller's responsibility to keep the external-FD count in sync with reality.
1219 : */
1220 : void
1221 338432 : ReserveExternalFD(void)
1222 : {
1223 : /*
1224 : * Release VFDs if needed to stay safe. Because we do this before
1225 : * incrementing numExternalFDs, the final state will be as desired, i.e.,
1226 : * nfile + numAllocatedDescs + numExternalFDs <= max_safe_fds.
1227 : */
1228 338432 : ReleaseLruFiles();
1229 :
1230 338432 : numExternalFDs++;
1231 338432 : }
1232 :
1233 : /*
1234 : * ReleaseExternalFD - report release of an external file descriptor
1235 : *
1236 : * This is guaranteed not to change errno, so it can be used in failure paths.
1237 : */
1238 : void
1239 303026 : ReleaseExternalFD(void)
1240 : {
1241 : Assert(numExternalFDs > 0);
1242 303026 : numExternalFDs--;
1243 303026 : }
1244 :
1245 :
1246 : #if defined(FDDEBUG)
1247 :
1248 : static void
1249 : _dump_lru(void)
1250 : {
1251 : int mru = VfdCache[0].lruLessRecently;
1252 : Vfd *vfdP = &VfdCache[mru];
1253 : char buf[2048];
1254 :
1255 : snprintf(buf, sizeof(buf), "LRU: MOST %d ", mru);
1256 : while (mru != 0)
1257 : {
1258 : mru = vfdP->lruLessRecently;
1259 : vfdP = &VfdCache[mru];
1260 : snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), "%d ", mru);
1261 : }
1262 : snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), "LEAST");
1263 : elog(LOG, "%s", buf);
1264 : }
1265 : #endif /* FDDEBUG */
1266 :
1267 : static void
1268 2354136 : Delete(File file)
1269 : {
1270 : Vfd *vfdP;
1271 :
1272 : Assert(file != 0);
1273 :
1274 : DO_DB(elog(LOG, "Delete %d (%s)",
1275 : file, VfdCache[file].fileName));
1276 : DO_DB(_dump_lru());
1277 :
1278 2354136 : vfdP = &VfdCache[file];
1279 :
1280 2354136 : VfdCache[vfdP->lruLessRecently].lruMoreRecently = vfdP->lruMoreRecently;
1281 2354136 : VfdCache[vfdP->lruMoreRecently].lruLessRecently = vfdP->lruLessRecently;
1282 :
1283 : DO_DB(_dump_lru());
1284 2354136 : }
1285 :
1286 : static void
1287 78286 : LruDelete(File file)
1288 : {
1289 : Vfd *vfdP;
1290 :
1291 : Assert(file != 0);
1292 :
1293 : DO_DB(elog(LOG, "LruDelete %d (%s)",
1294 : file, VfdCache[file].fileName));
1295 :
1296 78286 : vfdP = &VfdCache[file];
1297 :
1298 : /*
1299 : * Close the file. We aren't expecting this to fail; if it does, better
1300 : * to leak the FD than to mess up our internal state.
1301 : */
1302 78286 : if (close(vfdP->fd) != 0)
1303 0 : elog(vfdP->fdstate & FD_TEMP_FILE_LIMIT ? LOG : data_sync_elevel(LOG),
1304 : "could not close file \"%s\": %m", vfdP->fileName);
1305 78286 : vfdP->fd = VFD_CLOSED;
1306 78286 : --nfile;
1307 :
1308 : /* delete the vfd record from the LRU ring */
1309 78286 : Delete(file);
1310 78286 : }
1311 :
1312 : static void
1313 2862022 : Insert(File file)
1314 : {
1315 : Vfd *vfdP;
1316 :
1317 : Assert(file != 0);
1318 :
1319 : DO_DB(elog(LOG, "Insert %d (%s)",
1320 : file, VfdCache[file].fileName));
1321 : DO_DB(_dump_lru());
1322 :
1323 2862022 : vfdP = &VfdCache[file];
1324 :
1325 2862022 : vfdP->lruMoreRecently = 0;
1326 2862022 : vfdP->lruLessRecently = VfdCache[0].lruLessRecently;
1327 2862022 : VfdCache[0].lruLessRecently = file;
1328 2862022 : VfdCache[vfdP->lruLessRecently].lruMoreRecently = file;
1329 :
1330 : DO_DB(_dump_lru());
1331 2862022 : }
1332 :
1333 : /* returns 0 on success, -1 on re-open failure (with errno set) */
1334 : static int
1335 29452 : LruInsert(File file)
1336 : {
1337 : Vfd *vfdP;
1338 :
1339 : Assert(file != 0);
1340 :
1341 : DO_DB(elog(LOG, "LruInsert %d (%s)",
1342 : file, VfdCache[file].fileName));
1343 :
1344 29452 : vfdP = &VfdCache[file];
1345 :
1346 29452 : if (FileIsNotOpen(file))
1347 : {
1348 : /* Close excess kernel FDs. */
1349 29452 : ReleaseLruFiles();
1350 :
1351 : /*
1352 : * The open could still fail for lack of file descriptors, eg due to
1353 : * overall system file table being full. So, be prepared to release
1354 : * another FD if necessary...
1355 : */
1356 29452 : vfdP->fd = BasicOpenFilePerm(vfdP->fileName, vfdP->fileFlags,
1357 : vfdP->fileMode);
1358 29452 : if (vfdP->fd < 0)
1359 : {
1360 : DO_DB(elog(LOG, "re-open failed: %m"));
1361 0 : return -1;
1362 : }
1363 : else
1364 : {
1365 29452 : ++nfile;
1366 : }
1367 : }
1368 :
1369 : /*
1370 : * put it at the head of the Lru ring
1371 : */
1372 :
1373 29452 : Insert(file);
1374 :
1375 29452 : return 0;
1376 : }
1377 :
1378 : /*
1379 : * Release one kernel FD by closing the least-recently-used VFD.
1380 : */
1381 : static bool
1382 78030 : ReleaseLruFile(void)
1383 : {
1384 : DO_DB(elog(LOG, "ReleaseLruFile. Opened %d", nfile));
1385 :
1386 78030 : if (nfile > 0)
1387 : {
1388 : /*
1389 : * There are opened files and so there should be at least one used vfd
1390 : * in the ring.
1391 : */
1392 : Assert(VfdCache[0].lruMoreRecently != 0);
1393 78030 : LruDelete(VfdCache[0].lruMoreRecently);
1394 78030 : return true; /* freed a file */
1395 : }
1396 0 : return false; /* no files available to free */
1397 : }
1398 :
1399 : /*
1400 : * Release kernel FDs as needed to get under the max_safe_fds limit.
1401 : * After calling this, it's OK to try to open another file.
1402 : */
1403 : static void
1404 3242756 : ReleaseLruFiles(void)
1405 : {
1406 3320786 : while (nfile + numAllocatedDescs + numExternalFDs >= max_safe_fds)
1407 : {
1408 78030 : if (!ReleaseLruFile())
1409 0 : break;
1410 : }
1411 3242756 : }
1412 :
1413 : static File
1414 2098090 : AllocateVfd(void)
1415 : {
1416 : Index i;
1417 : File file;
1418 :
1419 : DO_DB(elog(LOG, "AllocateVfd. Size %zu", SizeVfdCache));
1420 :
1421 : Assert(SizeVfdCache > 0); /* InitFileAccess not called? */
1422 :
1423 2098090 : if (VfdCache[0].nextFree == 0)
1424 : {
1425 : /*
1426 : * The free list is empty so it is time to increase the size of the
1427 : * array. We choose to double it each time this happens. However,
1428 : * there's not much point in starting *real* small.
1429 : */
1430 38056 : Size newCacheSize = SizeVfdCache * 2;
1431 : Vfd *newVfdCache;
1432 :
1433 38056 : if (newCacheSize < 32)
1434 28954 : newCacheSize = 32;
1435 :
1436 : /*
1437 : * Be careful not to clobber VfdCache ptr if realloc fails.
1438 : */
1439 38056 : newVfdCache = (Vfd *) realloc(VfdCache, sizeof(Vfd) * newCacheSize);
1440 38056 : if (newVfdCache == NULL)
1441 0 : ereport(ERROR,
1442 : (errcode(ERRCODE_OUT_OF_MEMORY),
1443 : errmsg("out of memory")));
1444 38056 : VfdCache = newVfdCache;
1445 :
1446 : /*
1447 : * Initialize the new entries and link them into the free list.
1448 : */
1449 1817166 : for (i = SizeVfdCache; i < newCacheSize; i++)
1450 : {
1451 14232880 : MemSet((char *) &(VfdCache[i]), 0, sizeof(Vfd));
1452 1779110 : VfdCache[i].nextFree = i + 1;
1453 1779110 : VfdCache[i].fd = VFD_CLOSED;
1454 : }
1455 38056 : VfdCache[newCacheSize - 1].nextFree = 0;
1456 38056 : VfdCache[0].nextFree = SizeVfdCache;
1457 :
1458 : /*
1459 : * Record the new size
1460 : */
1461 38056 : SizeVfdCache = newCacheSize;
1462 : }
1463 :
1464 2098090 : file = VfdCache[0].nextFree;
1465 :
1466 2098090 : VfdCache[0].nextFree = VfdCache[file].nextFree;
1467 :
1468 2098090 : return file;
1469 : }
1470 :
1471 : static void
1472 1557222 : FreeVfd(File file)
1473 : {
1474 1557222 : Vfd *vfdP = &VfdCache[file];
1475 :
1476 : DO_DB(elog(LOG, "FreeVfd: %d (%s)",
1477 : file, vfdP->fileName ? vfdP->fileName : ""));
1478 :
1479 1557222 : if (vfdP->fileName != NULL)
1480 : {
1481 871816 : free(vfdP->fileName);
1482 871816 : vfdP->fileName = NULL;
1483 : }
1484 1557222 : vfdP->fdstate = 0x0;
1485 :
1486 1557222 : vfdP->nextFree = VfdCache[0].nextFree;
1487 1557222 : VfdCache[0].nextFree = file;
1488 1557222 : }
1489 :
1490 : /* returns 0 on success, -1 on re-open failure (with errno set) */
1491 : static int
1492 4923366 : FileAccess(File file)
1493 : {
1494 : int returnValue;
1495 :
1496 : DO_DB(elog(LOG, "FileAccess %d (%s)",
1497 : file, VfdCache[file].fileName));
1498 :
1499 : /*
1500 : * Is the file open? If not, open it and put it at the head of the LRU
1501 : * ring (possibly closing the least recently used file to get an FD).
1502 : */
1503 :
1504 4923366 : if (FileIsNotOpen(file))
1505 : {
1506 29452 : returnValue = LruInsert(file);
1507 29452 : if (returnValue != 0)
1508 0 : return returnValue;
1509 : }
1510 4893914 : else if (VfdCache[0].lruLessRecently != file)
1511 : {
1512 : /*
1513 : * We now know that the file is open and that it is not the last one
1514 : * accessed, so we need to move it to the head of the Lru ring.
1515 : */
1516 :
1517 1419886 : Delete(file);
1518 1419886 : Insert(file);
1519 : }
1520 :
1521 4923366 : return 0;
1522 : }
1523 :
1524 : /*
1525 : * Called whenever a temporary file is deleted to report its size.
1526 : */
1527 : static void
1528 5264 : ReportTemporaryFileUsage(const char *path, off_t size)
1529 : {
1530 5264 : pgstat_report_tempfile(size);
1531 :
1532 5264 : if (log_temp_files >= 0)
1533 : {
1534 1612 : if ((size / 1024) >= log_temp_files)
1535 246 : ereport(LOG,
1536 : (errmsg("temporary file: path \"%s\", size %lu",
1537 : path, (unsigned long) size)));
1538 : }
1539 5264 : }
1540 :
1541 : /*
1542 : * Called to register a temporary file for automatic close.
1543 : * ResourceOwnerEnlarge(CurrentResourceOwner) must have been called
1544 : * before the file was opened.
1545 : */
1546 : static void
1547 8396 : RegisterTemporaryFile(File file)
1548 : {
1549 8396 : ResourceOwnerRememberFile(CurrentResourceOwner, file);
1550 8396 : VfdCache[file].resowner = CurrentResourceOwner;
1551 :
1552 : /* Backup mechanism for closing at end of xact. */
1553 8396 : VfdCache[file].fdstate |= FD_CLOSE_AT_EOXACT;
1554 8396 : have_xact_temporary_files = true;
1555 8396 : }
1556 :
1557 : /*
1558 : * Called when we get a shared invalidation message on some relation.
1559 : */
1560 : #ifdef NOT_USED
1561 : void
1562 : FileInvalidate(File file)
1563 : {
1564 : Assert(FileIsValid(file));
1565 : if (!FileIsNotOpen(file))
1566 : LruDelete(file);
1567 : }
1568 : #endif
1569 :
1570 : /*
1571 : * Open a file with PathNameOpenFilePerm() and pass default file mode for the
1572 : * fileMode parameter.
1573 : */
1574 : File
1575 2098090 : PathNameOpenFile(const char *fileName, int fileFlags)
1576 : {
1577 2098090 : return PathNameOpenFilePerm(fileName, fileFlags, pg_file_create_mode);
1578 : }
1579 :
1580 : /*
1581 : * open a file in an arbitrary directory
1582 : *
1583 : * NB: if the passed pathname is relative (which it usually is),
1584 : * it will be interpreted relative to the process' working directory
1585 : * (which should always be $PGDATA when this code is running).
1586 : */
1587 : File
1588 2098090 : PathNameOpenFilePerm(const char *fileName, int fileFlags, mode_t fileMode)
1589 : {
1590 : char *fnamecopy;
1591 : File file;
1592 : Vfd *vfdP;
1593 :
1594 : DO_DB(elog(LOG, "PathNameOpenFilePerm: %s %x %o",
1595 : fileName, fileFlags, fileMode));
1596 :
1597 : /*
1598 : * We need a malloc'd copy of the file name; fail cleanly if no room.
1599 : */
1600 2098090 : fnamecopy = strdup(fileName);
1601 2098090 : if (fnamecopy == NULL)
1602 0 : ereport(ERROR,
1603 : (errcode(ERRCODE_OUT_OF_MEMORY),
1604 : errmsg("out of memory")));
1605 :
1606 2098090 : file = AllocateVfd();
1607 2098090 : vfdP = &VfdCache[file];
1608 :
1609 : /* Close excess kernel FDs. */
1610 2098090 : ReleaseLruFiles();
1611 :
1612 : /*
1613 : * Descriptors managed by VFDs are implicitly marked O_CLOEXEC. The
1614 : * client shouldn't be expected to know which kernel descriptors are
1615 : * currently open, so it wouldn't make sense for them to be inherited by
1616 : * executed subprograms.
1617 : */
1618 2098090 : fileFlags |= O_CLOEXEC;
1619 :
1620 2098090 : vfdP->fd = BasicOpenFilePerm(fileName, fileFlags, fileMode);
1621 :
1622 2098090 : if (vfdP->fd < 0)
1623 : {
1624 685406 : int save_errno = errno;
1625 :
1626 685406 : FreeVfd(file);
1627 685406 : free(fnamecopy);
1628 685406 : errno = save_errno;
1629 685406 : return -1;
1630 : }
1631 1412684 : ++nfile;
1632 : DO_DB(elog(LOG, "PathNameOpenFile: success %d",
1633 : vfdP->fd));
1634 :
1635 1412684 : vfdP->fileName = fnamecopy;
1636 : /* Saved flags are adjusted to be OK for re-opening file */
1637 1412684 : vfdP->fileFlags = fileFlags & ~(O_CREAT | O_TRUNC | O_EXCL);
1638 1412684 : vfdP->fileMode = fileMode;
1639 1412684 : vfdP->fileSize = 0;
1640 1412684 : vfdP->fdstate = 0x0;
1641 1412684 : vfdP->resowner = NULL;
1642 :
1643 1412684 : Insert(file);
1644 :
1645 1412684 : return file;
1646 : }
1647 :
1648 : /*
1649 : * Create directory 'directory'. If necessary, create 'basedir', which must
1650 : * be the directory above it. This is designed for creating the top-level
1651 : * temporary directory on demand before creating a directory underneath it.
1652 : * Do nothing if the directory already exists.
1653 : *
1654 : * Directories created within the top-level temporary directory should begin
1655 : * with PG_TEMP_FILE_PREFIX, so that they can be identified as temporary and
1656 : * deleted at startup by RemovePgTempFiles(). Further subdirectories below
1657 : * that do not need any particular prefix.
1658 : */
1659 : void
1660 330 : PathNameCreateTemporaryDir(const char *basedir, const char *directory)
1661 : {
1662 330 : if (MakePGDirectory(directory) < 0)
1663 : {
1664 24 : if (errno == EEXIST)
1665 4 : return;
1666 :
1667 : /*
1668 : * Failed. Try to create basedir first in case it's missing. Tolerate
1669 : * EEXIST to close a race against another process following the same
1670 : * algorithm.
1671 : */
1672 20 : if (MakePGDirectory(basedir) < 0 && errno != EEXIST)
1673 0 : ereport(ERROR,
1674 : (errcode_for_file_access(),
1675 : errmsg("cannot create temporary directory \"%s\": %m",
1676 : basedir)));
1677 :
1678 : /* Try again. */
1679 20 : if (MakePGDirectory(directory) < 0 && errno != EEXIST)
1680 0 : ereport(ERROR,
1681 : (errcode_for_file_access(),
1682 : errmsg("cannot create temporary subdirectory \"%s\": %m",
1683 : directory)));
1684 : }
1685 : }
1686 :
1687 : /*
1688 : * Delete a directory and everything in it, if it exists.
1689 : */
1690 : void
1691 406 : PathNameDeleteTemporaryDir(const char *dirname)
1692 : {
1693 : struct stat statbuf;
1694 :
1695 : /* Silently ignore missing directory. */
1696 406 : if (stat(dirname, &statbuf) != 0 && errno == ENOENT)
1697 80 : return;
1698 :
1699 : /*
1700 : * Currently, walkdir doesn't offer a way for our passed in function to
1701 : * maintain state. Perhaps it should, so that we could tell the caller
1702 : * whether this operation succeeded or failed. Since this operation is
1703 : * used in a cleanup path, we wouldn't actually behave differently: we'll
1704 : * just log failures.
1705 : */
1706 326 : walkdir(dirname, unlink_if_exists_fname, false, LOG);
1707 : }
1708 :
1709 : /*
1710 : * Open a temporary file that will disappear when we close it.
1711 : *
1712 : * This routine takes care of generating an appropriate tempfile name.
1713 : * There's no need to pass in fileFlags or fileMode either, since only
1714 : * one setting makes any sense for a temp file.
1715 : *
1716 : * Unless interXact is true, the file is remembered by CurrentResourceOwner
1717 : * to ensure it's closed and deleted when it's no longer needed, typically at
1718 : * the end-of-transaction. In most cases, you don't want temporary files to
1719 : * outlive the transaction that created them, so this should be false -- but
1720 : * if you need "somewhat" temporary storage, this might be useful. In either
1721 : * case, the file is removed when the File is explicitly closed.
1722 : */
1723 : File
1724 3002 : OpenTemporaryFile(bool interXact)
1725 : {
1726 3002 : File file = 0;
1727 :
1728 : Assert(temporary_files_allowed); /* check temp file access is up */
1729 :
1730 : /*
1731 : * Make sure the current resource owner has space for this File before we
1732 : * open it, if we'll be registering it below.
1733 : */
1734 3002 : if (!interXact)
1735 2990 : ResourceOwnerEnlarge(CurrentResourceOwner);
1736 :
1737 : /*
1738 : * If some temp tablespace(s) have been given to us, try to use the next
1739 : * one. If a given tablespace can't be found, we silently fall back to
1740 : * the database's default tablespace.
1741 : *
1742 : * BUT: if the temp file is slated to outlive the current transaction,
1743 : * force it into the database's default tablespace, so that it will not
1744 : * pose a threat to possible tablespace drop attempts.
1745 : */
1746 3002 : if (numTempTableSpaces > 0 && !interXact)
1747 : {
1748 2 : Oid tblspcOid = GetNextTempTableSpace();
1749 :
1750 2 : if (OidIsValid(tblspcOid))
1751 2 : file = OpenTemporaryFileInTablespace(tblspcOid, false);
1752 : }
1753 :
1754 : /*
1755 : * If not, or if tablespace is bad, create in database's default
1756 : * tablespace. MyDatabaseTableSpace should normally be set before we get
1757 : * here, but just in case it isn't, fall back to pg_default tablespace.
1758 : */
1759 3002 : if (file <= 0)
1760 3000 : file = OpenTemporaryFileInTablespace(MyDatabaseTableSpace ?
1761 : MyDatabaseTableSpace :
1762 : DEFAULTTABLESPACE_OID,
1763 : true);
1764 :
1765 : /* Mark it for deletion at close and temporary file size limit */
1766 3002 : VfdCache[file].fdstate |= FD_DELETE_AT_CLOSE | FD_TEMP_FILE_LIMIT;
1767 :
1768 : /* Register it with the current resource owner */
1769 3002 : if (!interXact)
1770 2990 : RegisterTemporaryFile(file);
1771 :
1772 3002 : return file;
1773 : }
1774 :
1775 : /*
1776 : * Return the path of the temp directory in a given tablespace.
1777 : */
1778 : void
1779 16208 : TempTablespacePath(char *path, Oid tablespace)
1780 : {
1781 : /*
1782 : * Identify the tempfile directory for this tablespace.
1783 : *
1784 : * If someone tries to specify pg_global, use pg_default instead.
1785 : */
1786 16208 : if (tablespace == InvalidOid ||
1787 2 : tablespace == DEFAULTTABLESPACE_OID ||
1788 : tablespace == GLOBALTABLESPACE_OID)
1789 16206 : snprintf(path, MAXPGPATH, "base/%s", PG_TEMP_FILES_DIR);
1790 : else
1791 : {
1792 : /* All other tablespaces are accessed via symlinks */
1793 2 : snprintf(path, MAXPGPATH, "pg_tblspc/%u/%s/%s",
1794 : tablespace, TABLESPACE_VERSION_DIRECTORY,
1795 : PG_TEMP_FILES_DIR);
1796 : }
1797 16208 : }
1798 :
1799 : /*
1800 : * Open a temporary file in a specific tablespace.
1801 : * Subroutine for OpenTemporaryFile, which see for details.
1802 : */
1803 : static File
1804 3002 : OpenTemporaryFileInTablespace(Oid tblspcOid, bool rejectError)
1805 : {
1806 : char tempdirpath[MAXPGPATH];
1807 : char tempfilepath[MAXPGPATH];
1808 : File file;
1809 :
1810 3002 : TempTablespacePath(tempdirpath, tblspcOid);
1811 :
1812 : /*
1813 : * Generate a tempfile name that should be unique within the current
1814 : * database instance.
1815 : */
1816 3002 : snprintf(tempfilepath, sizeof(tempfilepath), "%s/%s%d.%ld",
1817 : tempdirpath, PG_TEMP_FILE_PREFIX, MyProcPid, tempFileCounter++);
1818 :
1819 : /*
1820 : * Open the file. Note: we don't use O_EXCL, in case there is an orphaned
1821 : * temp file that can be reused.
1822 : */
1823 3002 : file = PathNameOpenFile(tempfilepath,
1824 : O_RDWR | O_CREAT | O_TRUNC | PG_BINARY);
1825 3002 : if (file <= 0)
1826 : {
1827 : /*
1828 : * We might need to create the tablespace's tempfile directory, if no
1829 : * one has yet done so.
1830 : *
1831 : * Don't check for an error from MakePGDirectory; it could fail if
1832 : * someone else just did the same thing. If it doesn't work then
1833 : * we'll bomb out on the second create attempt, instead.
1834 : */
1835 168 : (void) MakePGDirectory(tempdirpath);
1836 :
1837 168 : file = PathNameOpenFile(tempfilepath,
1838 : O_RDWR | O_CREAT | O_TRUNC | PG_BINARY);
1839 168 : if (file <= 0 && rejectError)
1840 0 : elog(ERROR, "could not create temporary file \"%s\": %m",
1841 : tempfilepath);
1842 : }
1843 :
1844 3002 : return file;
1845 : }
1846 :
1847 :
1848 : /*
1849 : * Create a new file. The directory containing it must already exist. Files
1850 : * created this way are subject to temp_file_limit and are automatically
1851 : * closed at end of transaction, but are not automatically deleted on close
1852 : * because they are intended to be shared between cooperating backends.
1853 : *
1854 : * If the file is inside the top-level temporary directory, its name should
1855 : * begin with PG_TEMP_FILE_PREFIX so that it can be identified as temporary
1856 : * and deleted at startup by RemovePgTempFiles(). Alternatively, it can be
1857 : * inside a directory created with PathNameCreateTemporaryDir(), in which case
1858 : * the prefix isn't needed.
1859 : */
1860 : File
1861 2592 : PathNameCreateTemporaryFile(const char *path, bool error_on_failure)
1862 : {
1863 : File file;
1864 :
1865 : Assert(temporary_files_allowed); /* check temp file access is up */
1866 :
1867 2592 : ResourceOwnerEnlarge(CurrentResourceOwner);
1868 :
1869 : /*
1870 : * Open the file. Note: we don't use O_EXCL, in case there is an orphaned
1871 : * temp file that can be reused.
1872 : */
1873 2592 : file = PathNameOpenFile(path, O_RDWR | O_CREAT | O_TRUNC | PG_BINARY);
1874 2592 : if (file <= 0)
1875 : {
1876 330 : if (error_on_failure)
1877 0 : ereport(ERROR,
1878 : (errcode_for_file_access(),
1879 : errmsg("could not create temporary file \"%s\": %m",
1880 : path)));
1881 : else
1882 330 : return file;
1883 : }
1884 :
1885 : /* Mark it for temp_file_limit accounting. */
1886 2262 : VfdCache[file].fdstate |= FD_TEMP_FILE_LIMIT;
1887 :
1888 : /* Register it for automatic close. */
1889 2262 : RegisterTemporaryFile(file);
1890 :
1891 2262 : return file;
1892 : }
1893 :
1894 : /*
1895 : * Open a file that was created with PathNameCreateTemporaryFile, possibly in
1896 : * another backend. Files opened this way don't count against the
1897 : * temp_file_limit of the caller, are automatically closed at the end of the
1898 : * transaction but are not deleted on close.
1899 : */
1900 : File
1901 6834 : PathNameOpenTemporaryFile(const char *path, int mode)
1902 : {
1903 : File file;
1904 :
1905 : Assert(temporary_files_allowed); /* check temp file access is up */
1906 :
1907 6834 : ResourceOwnerEnlarge(CurrentResourceOwner);
1908 :
1909 6834 : file = PathNameOpenFile(path, mode | PG_BINARY);
1910 :
1911 : /* If no such file, then we don't raise an error. */
1912 6834 : if (file <= 0 && errno != ENOENT)
1913 0 : ereport(ERROR,
1914 : (errcode_for_file_access(),
1915 : errmsg("could not open temporary file \"%s\": %m",
1916 : path)));
1917 :
1918 6834 : if (file > 0)
1919 : {
1920 : /* Register it for automatic close. */
1921 3144 : RegisterTemporaryFile(file);
1922 : }
1923 :
1924 6834 : return file;
1925 : }
1926 :
1927 : /*
1928 : * Delete a file by pathname. Return true if the file existed, false if
1929 : * didn't.
1930 : */
1931 : bool
1932 5228 : PathNameDeleteTemporaryFile(const char *path, bool error_on_failure)
1933 : {
1934 : struct stat filestats;
1935 : int stat_errno;
1936 :
1937 : /* Get the final size for pgstat reporting. */
1938 5228 : if (stat(path, &filestats) != 0)
1939 2966 : stat_errno = errno;
1940 : else
1941 2262 : stat_errno = 0;
1942 :
1943 : /*
1944 : * Unlike FileClose's automatic file deletion code, we tolerate
1945 : * non-existence to support BufFileDeleteFileSet which doesn't know how
1946 : * many segments it has to delete until it runs out.
1947 : */
1948 5228 : if (stat_errno == ENOENT)
1949 2966 : return false;
1950 :
1951 2262 : if (unlink(path) < 0)
1952 : {
1953 0 : if (errno != ENOENT)
1954 0 : ereport(error_on_failure ? ERROR : LOG,
1955 : (errcode_for_file_access(),
1956 : errmsg("could not unlink temporary file \"%s\": %m",
1957 : path)));
1958 0 : return false;
1959 : }
1960 :
1961 2262 : if (stat_errno == 0)
1962 2262 : ReportTemporaryFileUsage(path, filestats.st_size);
1963 : else
1964 : {
1965 0 : errno = stat_errno;
1966 0 : ereport(LOG,
1967 : (errcode_for_file_access(),
1968 : errmsg("could not stat file \"%s\": %m", path)));
1969 : }
1970 :
1971 2262 : return true;
1972 : }
1973 :
1974 : /*
1975 : * close a file when done with it
1976 : */
1977 : void
1978 871816 : FileClose(File file)
1979 : {
1980 : Vfd *vfdP;
1981 :
1982 : Assert(FileIsValid(file));
1983 :
1984 : DO_DB(elog(LOG, "FileClose: %d (%s)",
1985 : file, VfdCache[file].fileName));
1986 :
1987 871816 : vfdP = &VfdCache[file];
1988 :
1989 871816 : if (!FileIsNotOpen(file))
1990 : {
1991 : /* close the file */
1992 855964 : if (close(vfdP->fd) != 0)
1993 : {
1994 : /*
1995 : * We may need to panic on failure to close non-temporary files;
1996 : * see LruDelete.
1997 : */
1998 0 : elog(vfdP->fdstate & FD_TEMP_FILE_LIMIT ? LOG : data_sync_elevel(LOG),
1999 : "could not close file \"%s\": %m", vfdP->fileName);
2000 : }
2001 :
2002 855964 : --nfile;
2003 855964 : vfdP->fd = VFD_CLOSED;
2004 :
2005 : /* remove the file from the lru ring */
2006 855964 : Delete(file);
2007 : }
2008 :
2009 871816 : if (vfdP->fdstate & FD_TEMP_FILE_LIMIT)
2010 : {
2011 : /* Subtract its size from current usage (do first in case of error) */
2012 5264 : temporary_files_size -= vfdP->fileSize;
2013 5264 : vfdP->fileSize = 0;
2014 : }
2015 :
2016 : /*
2017 : * Delete the file if it was temporary, and make a log entry if wanted
2018 : */
2019 871816 : if (vfdP->fdstate & FD_DELETE_AT_CLOSE)
2020 : {
2021 : struct stat filestats;
2022 : int stat_errno;
2023 :
2024 : /*
2025 : * If we get an error, as could happen within the ereport/elog calls,
2026 : * we'll come right back here during transaction abort. Reset the
2027 : * flag to ensure that we can't get into an infinite loop. This code
2028 : * is arranged to ensure that the worst-case consequence is failing to
2029 : * emit log message(s), not failing to attempt the unlink.
2030 : */
2031 3002 : vfdP->fdstate &= ~FD_DELETE_AT_CLOSE;
2032 :
2033 :
2034 : /* first try the stat() */
2035 3002 : if (stat(vfdP->fileName, &filestats))
2036 0 : stat_errno = errno;
2037 : else
2038 3002 : stat_errno = 0;
2039 :
2040 : /* in any case do the unlink */
2041 3002 : if (unlink(vfdP->fileName))
2042 0 : ereport(LOG,
2043 : (errcode_for_file_access(),
2044 : errmsg("could not delete file \"%s\": %m", vfdP->fileName)));
2045 :
2046 : /* and last report the stat results */
2047 3002 : if (stat_errno == 0)
2048 3002 : ReportTemporaryFileUsage(vfdP->fileName, filestats.st_size);
2049 : else
2050 : {
2051 0 : errno = stat_errno;
2052 0 : ereport(LOG,
2053 : (errcode_for_file_access(),
2054 : errmsg("could not stat file \"%s\": %m", vfdP->fileName)));
2055 : }
2056 : }
2057 :
2058 : /* Unregister it from the resource owner */
2059 871816 : if (vfdP->resowner)
2060 8388 : ResourceOwnerForgetFile(vfdP->resowner, file);
2061 :
2062 : /*
2063 : * Return the Vfd slot to the free list
2064 : */
2065 871816 : FreeVfd(file);
2066 871816 : }
2067 :
2068 : /*
2069 : * FilePrefetch - initiate asynchronous read of a given range of the file.
2070 : *
2071 : * Currently the only implementation of this function is using posix_fadvise
2072 : * which is the simplest standardized interface that accomplishes this.
2073 : * We could add an implementation using libaio in the future; but note that
2074 : * this API is inappropriate for libaio, which wants to have a buffer provided
2075 : * to read into.
2076 : */
2077 : int
2078 243296 : FilePrefetch(File file, off_t offset, off_t amount, uint32 wait_event_info)
2079 : {
2080 : #if defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_WILLNEED)
2081 : int returnCode;
2082 :
2083 : Assert(FileIsValid(file));
2084 :
2085 : DO_DB(elog(LOG, "FilePrefetch: %d (%s) " INT64_FORMAT " " INT64_FORMAT,
2086 : file, VfdCache[file].fileName,
2087 : (int64) offset, (int64) amount));
2088 :
2089 243296 : returnCode = FileAccess(file);
2090 243296 : if (returnCode < 0)
2091 0 : return returnCode;
2092 :
2093 243296 : retry:
2094 243296 : pgstat_report_wait_start(wait_event_info);
2095 243296 : returnCode = posix_fadvise(VfdCache[file].fd, offset, amount,
2096 : POSIX_FADV_WILLNEED);
2097 243296 : pgstat_report_wait_end();
2098 :
2099 243296 : if (returnCode == EINTR)
2100 0 : goto retry;
2101 :
2102 243296 : return returnCode;
2103 : #else
2104 : Assert(FileIsValid(file));
2105 : return 0;
2106 : #endif
2107 : }
2108 :
2109 : void
2110 133426 : FileWriteback(File file, off_t offset, off_t nbytes, uint32 wait_event_info)
2111 : {
2112 : int returnCode;
2113 :
2114 : Assert(FileIsValid(file));
2115 :
2116 : DO_DB(elog(LOG, "FileWriteback: %d (%s) " INT64_FORMAT " " INT64_FORMAT,
2117 : file, VfdCache[file].fileName,
2118 : (int64) offset, (int64) nbytes));
2119 :
2120 133426 : if (nbytes <= 0)
2121 0 : return;
2122 :
2123 133426 : if (VfdCache[file].fileFlags & PG_O_DIRECT)
2124 0 : return;
2125 :
2126 133426 : returnCode = FileAccess(file);
2127 133426 : if (returnCode < 0)
2128 0 : return;
2129 :
2130 133426 : pgstat_report_wait_start(wait_event_info);
2131 133426 : pg_flush_data(VfdCache[file].fd, offset, nbytes);
2132 133426 : pgstat_report_wait_end();
2133 : }
2134 :
2135 : ssize_t
2136 2912166 : FileReadV(File file, const struct iovec *iov, int iovcnt, off_t offset,
2137 : uint32 wait_event_info)
2138 : {
2139 : ssize_t returnCode;
2140 : Vfd *vfdP;
2141 :
2142 : Assert(FileIsValid(file));
2143 :
2144 : DO_DB(elog(LOG, "FileReadV: %d (%s) " INT64_FORMAT " %d",
2145 : file, VfdCache[file].fileName,
2146 : (int64) offset,
2147 : iovcnt));
2148 :
2149 2912166 : returnCode = FileAccess(file);
2150 2912166 : if (returnCode < 0)
2151 0 : return returnCode;
2152 :
2153 2912166 : vfdP = &VfdCache[file];
2154 :
2155 2912166 : retry:
2156 2912166 : pgstat_report_wait_start(wait_event_info);
2157 2912166 : returnCode = pg_preadv(vfdP->fd, iov, iovcnt, offset);
2158 2912166 : pgstat_report_wait_end();
2159 :
2160 2912166 : if (returnCode < 0)
2161 : {
2162 : /*
2163 : * Windows may run out of kernel buffers and return "Insufficient
2164 : * system resources" error. Wait a bit and retry to solve it.
2165 : *
2166 : * It is rumored that EINTR is also possible on some Unix filesystems,
2167 : * in which case immediate retry is indicated.
2168 : */
2169 : #ifdef WIN32
2170 : DWORD error = GetLastError();
2171 :
2172 : switch (error)
2173 : {
2174 : case ERROR_NO_SYSTEM_RESOURCES:
2175 : pg_usleep(1000L);
2176 : errno = EINTR;
2177 : break;
2178 : default:
2179 : _dosmaperr(error);
2180 : break;
2181 : }
2182 : #endif
2183 : /* OK to retry if interrupted */
2184 0 : if (errno == EINTR)
2185 0 : goto retry;
2186 : }
2187 :
2188 2912166 : return returnCode;
2189 : }
2190 :
2191 : ssize_t
2192 1240132 : FileWriteV(File file, const struct iovec *iov, int iovcnt, off_t offset,
2193 : uint32 wait_event_info)
2194 : {
2195 : ssize_t returnCode;
2196 : Vfd *vfdP;
2197 :
2198 : Assert(FileIsValid(file));
2199 :
2200 : DO_DB(elog(LOG, "FileWriteV: %d (%s) " INT64_FORMAT " %d",
2201 : file, VfdCache[file].fileName,
2202 : (int64) offset,
2203 : iovcnt));
2204 :
2205 1240132 : returnCode = FileAccess(file);
2206 1240132 : if (returnCode < 0)
2207 0 : return returnCode;
2208 :
2209 1240132 : vfdP = &VfdCache[file];
2210 :
2211 : /*
2212 : * If enforcing temp_file_limit and it's a temp file, check to see if the
2213 : * write would overrun temp_file_limit, and throw error if so. Note: it's
2214 : * really a modularity violation to throw error here; we should set errno
2215 : * and return -1. However, there's no way to report a suitable error
2216 : * message if we do that. All current callers would just throw error
2217 : * immediately anyway, so this is safe at present.
2218 : */
2219 1240132 : if (temp_file_limit >= 0 && (vfdP->fdstate & FD_TEMP_FILE_LIMIT))
2220 : {
2221 0 : off_t past_write = offset;
2222 :
2223 0 : for (int i = 0; i < iovcnt; ++i)
2224 0 : past_write += iov[i].iov_len;
2225 :
2226 0 : if (past_write > vfdP->fileSize)
2227 : {
2228 0 : uint64 newTotal = temporary_files_size;
2229 :
2230 0 : newTotal += past_write - vfdP->fileSize;
2231 0 : if (newTotal > (uint64) temp_file_limit * (uint64) 1024)
2232 0 : ereport(ERROR,
2233 : (errcode(ERRCODE_CONFIGURATION_LIMIT_EXCEEDED),
2234 : errmsg("temporary file size exceeds temp_file_limit (%dkB)",
2235 : temp_file_limit)));
2236 : }
2237 : }
2238 :
2239 1240132 : retry:
2240 1240132 : pgstat_report_wait_start(wait_event_info);
2241 1240132 : returnCode = pg_pwritev(vfdP->fd, iov, iovcnt, offset);
2242 1240132 : pgstat_report_wait_end();
2243 :
2244 1240132 : if (returnCode >= 0)
2245 : {
2246 : /*
2247 : * Some callers expect short writes to set errno, and traditionally we
2248 : * have assumed that they imply disk space shortage. We don't want to
2249 : * waste CPU cycles adding up the total size here, so we'll just set
2250 : * it for all successful writes in case such a caller determines that
2251 : * the write was short and ereports "%m".
2252 : */
2253 1240132 : errno = ENOSPC;
2254 :
2255 : /*
2256 : * Maintain fileSize and temporary_files_size if it's a temp file.
2257 : */
2258 1240132 : if (vfdP->fdstate & FD_TEMP_FILE_LIMIT)
2259 : {
2260 110170 : off_t past_write = offset + returnCode;
2261 :
2262 110170 : if (past_write > vfdP->fileSize)
2263 : {
2264 78016 : temporary_files_size += past_write - vfdP->fileSize;
2265 78016 : vfdP->fileSize = past_write;
2266 : }
2267 : }
2268 : }
2269 : else
2270 : {
2271 : /*
2272 : * See comments in FileReadV()
2273 : */
2274 : #ifdef WIN32
2275 : DWORD error = GetLastError();
2276 :
2277 : switch (error)
2278 : {
2279 : case ERROR_NO_SYSTEM_RESOURCES:
2280 : pg_usleep(1000L);
2281 : errno = EINTR;
2282 : break;
2283 : default:
2284 : _dosmaperr(error);
2285 : break;
2286 : }
2287 : #endif
2288 : /* OK to retry if interrupted */
2289 0 : if (errno == EINTR)
2290 0 : goto retry;
2291 : }
2292 :
2293 1240132 : return returnCode;
2294 : }
2295 :
2296 : int
2297 278 : FileSync(File file, uint32 wait_event_info)
2298 : {
2299 : int returnCode;
2300 :
2301 : Assert(FileIsValid(file));
2302 :
2303 : DO_DB(elog(LOG, "FileSync: %d (%s)",
2304 : file, VfdCache[file].fileName));
2305 :
2306 278 : returnCode = FileAccess(file);
2307 278 : if (returnCode < 0)
2308 0 : return returnCode;
2309 :
2310 278 : pgstat_report_wait_start(wait_event_info);
2311 278 : returnCode = pg_fsync(VfdCache[file].fd);
2312 278 : pgstat_report_wait_end();
2313 :
2314 278 : return returnCode;
2315 : }
2316 :
2317 : /*
2318 : * Zero a region of the file.
2319 : *
2320 : * Returns 0 on success, -1 otherwise. In the latter case errno is set to the
2321 : * appropriate error.
2322 : */
2323 : int
2324 371276 : FileZero(File file, off_t offset, off_t amount, uint32 wait_event_info)
2325 : {
2326 : int returnCode;
2327 : ssize_t written;
2328 :
2329 : Assert(FileIsValid(file));
2330 :
2331 : DO_DB(elog(LOG, "FileZero: %d (%s) " INT64_FORMAT " " INT64_FORMAT,
2332 : file, VfdCache[file].fileName,
2333 : (int64) offset, (int64) amount));
2334 :
2335 371276 : returnCode = FileAccess(file);
2336 371276 : if (returnCode < 0)
2337 0 : return returnCode;
2338 :
2339 371276 : pgstat_report_wait_start(wait_event_info);
2340 371276 : written = pg_pwrite_zeros(VfdCache[file].fd, amount, offset);
2341 371276 : pgstat_report_wait_end();
2342 :
2343 371276 : if (written < 0)
2344 0 : return -1;
2345 371276 : else if (written != amount)
2346 : {
2347 : /* if errno is unset, assume problem is no disk space */
2348 0 : if (errno == 0)
2349 0 : errno = ENOSPC;
2350 0 : return -1;
2351 : }
2352 :
2353 371276 : return 0;
2354 : }
2355 :
2356 : /*
2357 : * Try to reserve file space with posix_fallocate(). If posix_fallocate() is
2358 : * not implemented on the operating system or fails with EINVAL / EOPNOTSUPP,
2359 : * use FileZero() instead.
2360 : *
2361 : * Note that at least glibc() implements posix_fallocate() in userspace if not
2362 : * implemented by the filesystem. That's not the case for all environments
2363 : * though.
2364 : *
2365 : * Returns 0 on success, -1 otherwise. In the latter case errno is set to the
2366 : * appropriate error.
2367 : */
2368 : int
2369 982 : FileFallocate(File file, off_t offset, off_t amount, uint32 wait_event_info)
2370 : {
2371 : #ifdef HAVE_POSIX_FALLOCATE
2372 : int returnCode;
2373 :
2374 : Assert(FileIsValid(file));
2375 :
2376 : DO_DB(elog(LOG, "FileFallocate: %d (%s) " INT64_FORMAT " " INT64_FORMAT,
2377 : file, VfdCache[file].fileName,
2378 : (int64) offset, (int64) amount));
2379 :
2380 982 : returnCode = FileAccess(file);
2381 982 : if (returnCode < 0)
2382 0 : return -1;
2383 :
2384 982 : retry:
2385 982 : pgstat_report_wait_start(wait_event_info);
2386 982 : returnCode = posix_fallocate(VfdCache[file].fd, offset, amount);
2387 982 : pgstat_report_wait_end();
2388 :
2389 982 : if (returnCode == 0)
2390 982 : return 0;
2391 0 : else if (returnCode == EINTR)
2392 0 : goto retry;
2393 :
2394 : /* for compatibility with %m printing etc */
2395 0 : errno = returnCode;
2396 :
2397 : /*
2398 : * Return in cases of a "real" failure, if fallocate is not supported,
2399 : * fall through to the FileZero() backed implementation.
2400 : */
2401 0 : if (returnCode != EINVAL && returnCode != EOPNOTSUPP)
2402 0 : return -1;
2403 : #endif
2404 :
2405 0 : return FileZero(file, offset, amount, wait_event_info);
2406 : }
2407 :
2408 : off_t
2409 3567232 : FileSize(File file)
2410 : {
2411 : Assert(FileIsValid(file));
2412 :
2413 : DO_DB(elog(LOG, "FileSize %d (%s)",
2414 : file, VfdCache[file].fileName));
2415 :
2416 3567232 : if (FileIsNotOpen(file))
2417 : {
2418 20892 : if (FileAccess(file) < 0)
2419 0 : return (off_t) -1;
2420 : }
2421 :
2422 3567232 : return lseek(VfdCache[file].fd, 0, SEEK_END);
2423 : }
2424 :
2425 : int
2426 918 : FileTruncate(File file, off_t offset, uint32 wait_event_info)
2427 : {
2428 : int returnCode;
2429 :
2430 : Assert(FileIsValid(file));
2431 :
2432 : DO_DB(elog(LOG, "FileTruncate %d (%s)",
2433 : file, VfdCache[file].fileName));
2434 :
2435 918 : returnCode = FileAccess(file);
2436 918 : if (returnCode < 0)
2437 0 : return returnCode;
2438 :
2439 918 : pgstat_report_wait_start(wait_event_info);
2440 918 : returnCode = pg_ftruncate(VfdCache[file].fd, offset);
2441 918 : pgstat_report_wait_end();
2442 :
2443 918 : if (returnCode == 0 && VfdCache[file].fileSize > offset)
2444 : {
2445 : /* adjust our state for truncation of a temp file */
2446 : Assert(VfdCache[file].fdstate & FD_TEMP_FILE_LIMIT);
2447 0 : temporary_files_size -= VfdCache[file].fileSize - offset;
2448 0 : VfdCache[file].fileSize = offset;
2449 : }
2450 :
2451 918 : return returnCode;
2452 : }
2453 :
2454 : /*
2455 : * Return the pathname associated with an open file.
2456 : *
2457 : * The returned string points to an internal buffer, which is valid until
2458 : * the file is closed.
2459 : */
2460 : char *
2461 24 : FilePathName(File file)
2462 : {
2463 : Assert(FileIsValid(file));
2464 :
2465 24 : return VfdCache[file].fileName;
2466 : }
2467 :
2468 : /*
2469 : * Return the raw file descriptor of an opened file.
2470 : *
2471 : * The returned file descriptor will be valid until the file is closed, but
2472 : * there are a lot of things that can make that happen. So the caller should
2473 : * be careful not to do much of anything else before it finishes using the
2474 : * returned file descriptor.
2475 : */
2476 : int
2477 0 : FileGetRawDesc(File file)
2478 : {
2479 : Assert(FileIsValid(file));
2480 0 : return VfdCache[file].fd;
2481 : }
2482 :
2483 : /*
2484 : * FileGetRawFlags - returns the file flags on open(2)
2485 : */
2486 : int
2487 0 : FileGetRawFlags(File file)
2488 : {
2489 : Assert(FileIsValid(file));
2490 0 : return VfdCache[file].fileFlags;
2491 : }
2492 :
2493 : /*
2494 : * FileGetRawMode - returns the mode bitmask passed to open(2)
2495 : */
2496 : mode_t
2497 0 : FileGetRawMode(File file)
2498 : {
2499 : Assert(FileIsValid(file));
2500 0 : return VfdCache[file].fileMode;
2501 : }
2502 :
2503 : /*
2504 : * Make room for another allocatedDescs[] array entry if needed and possible.
2505 : * Returns true if an array element is available.
2506 : */
2507 : static bool
2508 776782 : reserveAllocatedDesc(void)
2509 : {
2510 : AllocateDesc *newDescs;
2511 : int newMax;
2512 :
2513 : /* Quick out if array already has a free slot. */
2514 776782 : if (numAllocatedDescs < maxAllocatedDescs)
2515 775000 : return true;
2516 :
2517 : /*
2518 : * If the array hasn't yet been created in the current process, initialize
2519 : * it with FD_MINFREE / 3 elements. In many scenarios this is as many as
2520 : * we will ever need, anyway. We don't want to look at max_safe_fds
2521 : * immediately because set_max_safe_fds() may not have run yet.
2522 : */
2523 1782 : if (allocatedDescs == NULL)
2524 : {
2525 1782 : newMax = FD_MINFREE / 3;
2526 1782 : newDescs = (AllocateDesc *) malloc(newMax * sizeof(AllocateDesc));
2527 : /* Out of memory already? Treat as fatal error. */
2528 1782 : if (newDescs == NULL)
2529 0 : ereport(ERROR,
2530 : (errcode(ERRCODE_OUT_OF_MEMORY),
2531 : errmsg("out of memory")));
2532 1782 : allocatedDescs = newDescs;
2533 1782 : maxAllocatedDescs = newMax;
2534 1782 : return true;
2535 : }
2536 :
2537 : /*
2538 : * Consider enlarging the array beyond the initial allocation used above.
2539 : * By the time this happens, max_safe_fds should be known accurately.
2540 : *
2541 : * We mustn't let allocated descriptors hog all the available FDs, and in
2542 : * practice we'd better leave a reasonable number of FDs for VFD use. So
2543 : * set the maximum to max_safe_fds / 3. (This should certainly be at
2544 : * least as large as the initial size, FD_MINFREE / 3, so we aren't
2545 : * tightening the restriction here.) Recall that "external" FDs are
2546 : * allowed to consume another third of max_safe_fds.
2547 : */
2548 0 : newMax = max_safe_fds / 3;
2549 0 : if (newMax > maxAllocatedDescs)
2550 : {
2551 0 : newDescs = (AllocateDesc *) realloc(allocatedDescs,
2552 : newMax * sizeof(AllocateDesc));
2553 : /* Treat out-of-memory as a non-fatal error. */
2554 0 : if (newDescs == NULL)
2555 0 : return false;
2556 0 : allocatedDescs = newDescs;
2557 0 : maxAllocatedDescs = newMax;
2558 0 : return true;
2559 : }
2560 :
2561 : /* Can't enlarge allocatedDescs[] any more. */
2562 0 : return false;
2563 : }
2564 :
2565 : /*
2566 : * Routines that want to use stdio (ie, FILE*) should use AllocateFile
2567 : * rather than plain fopen(). This lets fd.c deal with freeing FDs if
2568 : * necessary to open the file. When done, call FreeFile rather than fclose.
2569 : *
2570 : * Note that files that will be open for any significant length of time
2571 : * should NOT be handled this way, since they cannot share kernel file
2572 : * descriptors with other files; there is grave risk of running out of FDs
2573 : * if anyone locks down too many FDs. Most callers of this routine are
2574 : * simply reading a config file that they will read and close immediately.
2575 : *
2576 : * fd.c will automatically close all files opened with AllocateFile at
2577 : * transaction commit or abort; this prevents FD leakage if a routine
2578 : * that calls AllocateFile is terminated prematurely by ereport(ERROR).
2579 : *
2580 : * Ideally this should be the *only* direct call of fopen() in the backend.
2581 : */
2582 : FILE *
2583 131436 : AllocateFile(const char *name, const char *mode)
2584 : {
2585 : FILE *file;
2586 :
2587 : DO_DB(elog(LOG, "AllocateFile: Allocated %d (%s)",
2588 : numAllocatedDescs, name));
2589 :
2590 : /* Can we allocate another non-virtual FD? */
2591 131436 : if (!reserveAllocatedDesc())
2592 0 : ereport(ERROR,
2593 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
2594 : errmsg("exceeded maxAllocatedDescs (%d) while trying to open file \"%s\"",
2595 : maxAllocatedDescs, name)));
2596 :
2597 : /* Close excess kernel FDs. */
2598 131436 : ReleaseLruFiles();
2599 :
2600 131436 : TryAgain:
2601 131436 : if ((file = fopen(name, mode)) != NULL)
2602 : {
2603 121984 : AllocateDesc *desc = &allocatedDescs[numAllocatedDescs];
2604 :
2605 121984 : desc->kind = AllocateDescFile;
2606 121984 : desc->desc.file = file;
2607 121984 : desc->create_subid = GetCurrentSubTransactionId();
2608 121984 : numAllocatedDescs++;
2609 121984 : return desc->desc.file;
2610 : }
2611 :
2612 9452 : if (errno == EMFILE || errno == ENFILE)
2613 : {
2614 0 : int save_errno = errno;
2615 :
2616 0 : ereport(LOG,
2617 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
2618 : errmsg("out of file descriptors: %m; release and retry")));
2619 0 : errno = 0;
2620 0 : if (ReleaseLruFile())
2621 0 : goto TryAgain;
2622 0 : errno = save_errno;
2623 : }
2624 :
2625 9452 : return NULL;
2626 : }
2627 :
2628 : /*
2629 : * Open a file with OpenTransientFilePerm() and pass default file mode for
2630 : * the fileMode parameter.
2631 : */
2632 : int
2633 581908 : OpenTransientFile(const char *fileName, int fileFlags)
2634 : {
2635 581908 : return OpenTransientFilePerm(fileName, fileFlags, pg_file_create_mode);
2636 : }
2637 :
2638 : /*
2639 : * Like AllocateFile, but returns an unbuffered fd like open(2)
2640 : */
2641 : int
2642 581920 : OpenTransientFilePerm(const char *fileName, int fileFlags, mode_t fileMode)
2643 : {
2644 : int fd;
2645 :
2646 : DO_DB(elog(LOG, "OpenTransientFile: Allocated %d (%s)",
2647 : numAllocatedDescs, fileName));
2648 :
2649 : /* Can we allocate another non-virtual FD? */
2650 581920 : if (!reserveAllocatedDesc())
2651 0 : ereport(ERROR,
2652 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
2653 : errmsg("exceeded maxAllocatedDescs (%d) while trying to open file \"%s\"",
2654 : maxAllocatedDescs, fileName)));
2655 :
2656 : /* Close excess kernel FDs. */
2657 581920 : ReleaseLruFiles();
2658 :
2659 581920 : fd = BasicOpenFilePerm(fileName, fileFlags, fileMode);
2660 :
2661 581920 : if (fd >= 0)
2662 : {
2663 578490 : AllocateDesc *desc = &allocatedDescs[numAllocatedDescs];
2664 :
2665 578490 : desc->kind = AllocateDescRawFD;
2666 578490 : desc->desc.fd = fd;
2667 578490 : desc->create_subid = GetCurrentSubTransactionId();
2668 578490 : numAllocatedDescs++;
2669 :
2670 578490 : return fd;
2671 : }
2672 :
2673 3430 : return -1; /* failure */
2674 : }
2675 :
2676 : /*
2677 : * Routines that want to initiate a pipe stream should use OpenPipeStream
2678 : * rather than plain popen(). This lets fd.c deal with freeing FDs if
2679 : * necessary. When done, call ClosePipeStream rather than pclose.
2680 : *
2681 : * This function also ensures that the popen'd program is run with default
2682 : * SIGPIPE processing, rather than the SIG_IGN setting the backend normally
2683 : * uses. This ensures desirable response to, eg, closing a read pipe early.
2684 : */
2685 : FILE *
2686 92 : OpenPipeStream(const char *command, const char *mode)
2687 : {
2688 : FILE *file;
2689 : int save_errno;
2690 :
2691 : DO_DB(elog(LOG, "OpenPipeStream: Allocated %d (%s)",
2692 : numAllocatedDescs, command));
2693 :
2694 : /* Can we allocate another non-virtual FD? */
2695 92 : if (!reserveAllocatedDesc())
2696 0 : ereport(ERROR,
2697 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
2698 : errmsg("exceeded maxAllocatedDescs (%d) while trying to execute command \"%s\"",
2699 : maxAllocatedDescs, command)));
2700 :
2701 : /* Close excess kernel FDs. */
2702 92 : ReleaseLruFiles();
2703 :
2704 92 : TryAgain:
2705 92 : fflush(NULL);
2706 92 : pqsignal(SIGPIPE, SIG_DFL);
2707 92 : errno = 0;
2708 92 : file = popen(command, mode);
2709 92 : save_errno = errno;
2710 92 : pqsignal(SIGPIPE, SIG_IGN);
2711 92 : errno = save_errno;
2712 92 : if (file != NULL)
2713 : {
2714 92 : AllocateDesc *desc = &allocatedDescs[numAllocatedDescs];
2715 :
2716 92 : desc->kind = AllocateDescPipe;
2717 92 : desc->desc.file = file;
2718 92 : desc->create_subid = GetCurrentSubTransactionId();
2719 92 : numAllocatedDescs++;
2720 92 : return desc->desc.file;
2721 : }
2722 :
2723 0 : if (errno == EMFILE || errno == ENFILE)
2724 : {
2725 0 : ereport(LOG,
2726 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
2727 : errmsg("out of file descriptors: %m; release and retry")));
2728 0 : if (ReleaseLruFile())
2729 0 : goto TryAgain;
2730 0 : errno = save_errno;
2731 : }
2732 :
2733 0 : return NULL;
2734 : }
2735 :
2736 : /*
2737 : * Free an AllocateDesc of any type.
2738 : *
2739 : * The argument *must* point into the allocatedDescs[] array.
2740 : */
2741 : static int
2742 762440 : FreeDesc(AllocateDesc *desc)
2743 : {
2744 : int result;
2745 :
2746 : /* Close the underlying object */
2747 762440 : switch (desc->kind)
2748 : {
2749 121984 : case AllocateDescFile:
2750 121984 : result = fclose(desc->desc.file);
2751 121984 : break;
2752 92 : case AllocateDescPipe:
2753 92 : result = pclose(desc->desc.file);
2754 92 : break;
2755 61874 : case AllocateDescDir:
2756 61874 : result = closedir(desc->desc.dir);
2757 61874 : break;
2758 578490 : case AllocateDescRawFD:
2759 578490 : result = close(desc->desc.fd);
2760 578490 : break;
2761 0 : default:
2762 0 : elog(ERROR, "AllocateDesc kind not recognized");
2763 : result = 0; /* keep compiler quiet */
2764 : break;
2765 : }
2766 :
2767 : /* Compact storage in the allocatedDescs array */
2768 762440 : numAllocatedDescs--;
2769 762440 : *desc = allocatedDescs[numAllocatedDescs];
2770 :
2771 762440 : return result;
2772 : }
2773 :
2774 : /*
2775 : * Close a file returned by AllocateFile.
2776 : *
2777 : * Note we do not check fclose's return value --- it is up to the caller
2778 : * to handle close errors.
2779 : */
2780 : int
2781 121966 : FreeFile(FILE *file)
2782 : {
2783 : int i;
2784 :
2785 : DO_DB(elog(LOG, "FreeFile: Allocated %d", numAllocatedDescs));
2786 :
2787 : /* Remove file from list of allocated files, if it's present */
2788 121968 : for (i = numAllocatedDescs; --i >= 0;)
2789 : {
2790 121968 : AllocateDesc *desc = &allocatedDescs[i];
2791 :
2792 121968 : if (desc->kind == AllocateDescFile && desc->desc.file == file)
2793 121966 : return FreeDesc(desc);
2794 : }
2795 :
2796 : /* Only get here if someone passes us a file not in allocatedDescs */
2797 0 : elog(WARNING, "file passed to FreeFile was not obtained from AllocateFile");
2798 :
2799 0 : return fclose(file);
2800 : }
2801 :
2802 : /*
2803 : * Close a file returned by OpenTransientFile.
2804 : *
2805 : * Note we do not check close's return value --- it is up to the caller
2806 : * to handle close errors.
2807 : */
2808 : int
2809 578488 : CloseTransientFile(int fd)
2810 : {
2811 : int i;
2812 :
2813 : DO_DB(elog(LOG, "CloseTransientFile: Allocated %d", numAllocatedDescs));
2814 :
2815 : /* Remove fd from list of allocated files, if it's present */
2816 578488 : for (i = numAllocatedDescs; --i >= 0;)
2817 : {
2818 578488 : AllocateDesc *desc = &allocatedDescs[i];
2819 :
2820 578488 : if (desc->kind == AllocateDescRawFD && desc->desc.fd == fd)
2821 578488 : return FreeDesc(desc);
2822 : }
2823 :
2824 : /* Only get here if someone passes us a file not in allocatedDescs */
2825 0 : elog(WARNING, "fd passed to CloseTransientFile was not obtained from OpenTransientFile");
2826 :
2827 0 : return close(fd);
2828 : }
2829 :
2830 : /*
2831 : * Routines that want to use <dirent.h> (ie, DIR*) should use AllocateDir
2832 : * rather than plain opendir(). This lets fd.c deal with freeing FDs if
2833 : * necessary to open the directory, and with closing it after an elog.
2834 : * When done, call FreeDir rather than closedir.
2835 : *
2836 : * Returns NULL, with errno set, on failure. Note that failure detection
2837 : * is commonly left to the following call of ReadDir or ReadDirExtended;
2838 : * see the comments for ReadDir.
2839 : *
2840 : * Ideally this should be the *only* direct call of opendir() in the backend.
2841 : */
2842 : DIR *
2843 63334 : AllocateDir(const char *dirname)
2844 : {
2845 : DIR *dir;
2846 :
2847 : DO_DB(elog(LOG, "AllocateDir: Allocated %d (%s)",
2848 : numAllocatedDescs, dirname));
2849 :
2850 : /* Can we allocate another non-virtual FD? */
2851 63334 : if (!reserveAllocatedDesc())
2852 0 : ereport(ERROR,
2853 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
2854 : errmsg("exceeded maxAllocatedDescs (%d) while trying to open directory \"%s\"",
2855 : maxAllocatedDescs, dirname)));
2856 :
2857 : /* Close excess kernel FDs. */
2858 63334 : ReleaseLruFiles();
2859 :
2860 63334 : TryAgain:
2861 63334 : if ((dir = opendir(dirname)) != NULL)
2862 : {
2863 61874 : AllocateDesc *desc = &allocatedDescs[numAllocatedDescs];
2864 :
2865 61874 : desc->kind = AllocateDescDir;
2866 61874 : desc->desc.dir = dir;
2867 61874 : desc->create_subid = GetCurrentSubTransactionId();
2868 61874 : numAllocatedDescs++;
2869 61874 : return desc->desc.dir;
2870 : }
2871 :
2872 1460 : if (errno == EMFILE || errno == ENFILE)
2873 : {
2874 0 : int save_errno = errno;
2875 :
2876 0 : ereport(LOG,
2877 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
2878 : errmsg("out of file descriptors: %m; release and retry")));
2879 0 : errno = 0;
2880 0 : if (ReleaseLruFile())
2881 0 : goto TryAgain;
2882 0 : errno = save_errno;
2883 : }
2884 :
2885 1460 : return NULL;
2886 : }
2887 :
2888 : /*
2889 : * Read a directory opened with AllocateDir, ereport'ing any error.
2890 : *
2891 : * This is easier to use than raw readdir() since it takes care of some
2892 : * otherwise rather tedious and error-prone manipulation of errno. Also,
2893 : * if you are happy with a generic error message for AllocateDir failure,
2894 : * you can just do
2895 : *
2896 : * dir = AllocateDir(path);
2897 : * while ((dirent = ReadDir(dir, path)) != NULL)
2898 : * process dirent;
2899 : * FreeDir(dir);
2900 : *
2901 : * since a NULL dir parameter is taken as indicating AllocateDir failed.
2902 : * (Make sure errno isn't changed between AllocateDir and ReadDir if you
2903 : * use this shortcut.)
2904 : *
2905 : * The pathname passed to AllocateDir must be passed to this routine too,
2906 : * but it is only used for error reporting.
2907 : */
2908 : struct dirent *
2909 1952168 : ReadDir(DIR *dir, const char *dirname)
2910 : {
2911 1952168 : return ReadDirExtended(dir, dirname, ERROR);
2912 : }
2913 :
2914 : /*
2915 : * Alternate version of ReadDir that allows caller to specify the elevel
2916 : * for any error report (whether it's reporting an initial failure of
2917 : * AllocateDir or a subsequent directory read failure).
2918 : *
2919 : * If elevel < ERROR, returns NULL after any error. With the normal coding
2920 : * pattern, this will result in falling out of the loop immediately as
2921 : * though the directory contained no (more) entries.
2922 : */
2923 : struct dirent *
2924 3926010 : ReadDirExtended(DIR *dir, const char *dirname, int elevel)
2925 : {
2926 : struct dirent *dent;
2927 :
2928 : /* Give a generic message for AllocateDir failure, if caller didn't */
2929 3926010 : if (dir == NULL)
2930 : {
2931 6 : ereport(elevel,
2932 : (errcode_for_file_access(),
2933 : errmsg("could not open directory \"%s\": %m",
2934 : dirname)));
2935 0 : return NULL;
2936 : }
2937 :
2938 3926004 : errno = 0;
2939 3926004 : if ((dent = readdir(dir)) != NULL)
2940 3877596 : return dent;
2941 :
2942 48408 : if (errno)
2943 0 : ereport(elevel,
2944 : (errcode_for_file_access(),
2945 : errmsg("could not read directory \"%s\": %m",
2946 : dirname)));
2947 48408 : return NULL;
2948 : }
2949 :
2950 : /*
2951 : * Close a directory opened with AllocateDir.
2952 : *
2953 : * Returns closedir's return value (with errno set if it's not 0).
2954 : * Note we do not check the return value --- it is up to the caller
2955 : * to handle close errors if wanted.
2956 : *
2957 : * Does nothing if dir == NULL; we assume that directory open failure was
2958 : * already reported if desired.
2959 : */
2960 : int
2961 61650 : FreeDir(DIR *dir)
2962 : {
2963 : int i;
2964 :
2965 : /* Nothing to do if AllocateDir failed */
2966 61650 : if (dir == NULL)
2967 0 : return 0;
2968 :
2969 : DO_DB(elog(LOG, "FreeDir: Allocated %d", numAllocatedDescs));
2970 :
2971 : /* Remove dir from list of allocated dirs, if it's present */
2972 61650 : for (i = numAllocatedDescs; --i >= 0;)
2973 : {
2974 61650 : AllocateDesc *desc = &allocatedDescs[i];
2975 :
2976 61650 : if (desc->kind == AllocateDescDir && desc->desc.dir == dir)
2977 61650 : return FreeDesc(desc);
2978 : }
2979 :
2980 : /* Only get here if someone passes us a dir not in allocatedDescs */
2981 0 : elog(WARNING, "dir passed to FreeDir was not obtained from AllocateDir");
2982 :
2983 0 : return closedir(dir);
2984 : }
2985 :
2986 :
2987 : /*
2988 : * Close a pipe stream returned by OpenPipeStream.
2989 : */
2990 : int
2991 92 : ClosePipeStream(FILE *file)
2992 : {
2993 : int i;
2994 :
2995 : DO_DB(elog(LOG, "ClosePipeStream: Allocated %d", numAllocatedDescs));
2996 :
2997 : /* Remove file from list of allocated files, if it's present */
2998 92 : for (i = numAllocatedDescs; --i >= 0;)
2999 : {
3000 92 : AllocateDesc *desc = &allocatedDescs[i];
3001 :
3002 92 : if (desc->kind == AllocateDescPipe && desc->desc.file == file)
3003 92 : return FreeDesc(desc);
3004 : }
3005 :
3006 : /* Only get here if someone passes us a file not in allocatedDescs */
3007 0 : elog(WARNING, "file passed to ClosePipeStream was not obtained from OpenPipeStream");
3008 :
3009 0 : return pclose(file);
3010 : }
3011 :
3012 : /*
3013 : * closeAllVfds
3014 : *
3015 : * Force all VFDs into the physically-closed state, so that the fewest
3016 : * possible number of kernel file descriptors are in use. There is no
3017 : * change in the logical state of the VFDs.
3018 : */
3019 : void
3020 58 : closeAllVfds(void)
3021 : {
3022 : Index i;
3023 :
3024 58 : if (SizeVfdCache > 0)
3025 : {
3026 : Assert(FileIsNotOpen(0)); /* Make sure ring not corrupted */
3027 1856 : for (i = 1; i < SizeVfdCache; i++)
3028 : {
3029 1798 : if (!FileIsNotOpen(i))
3030 256 : LruDelete(i);
3031 : }
3032 : }
3033 58 : }
3034 :
3035 :
3036 : /*
3037 : * SetTempTablespaces
3038 : *
3039 : * Define a list (actually an array) of OIDs of tablespaces to use for
3040 : * temporary files. This list will be used until end of transaction,
3041 : * unless this function is called again before then. It is caller's
3042 : * responsibility that the passed-in array has adequate lifespan (typically
3043 : * it'd be allocated in TopTransactionContext).
3044 : *
3045 : * Some entries of the array may be InvalidOid, indicating that the current
3046 : * database's default tablespace should be used.
3047 : */
3048 : void
3049 5440 : SetTempTablespaces(Oid *tableSpaces, int numSpaces)
3050 : {
3051 : Assert(numSpaces >= 0);
3052 5440 : tempTableSpaces = tableSpaces;
3053 5440 : numTempTableSpaces = numSpaces;
3054 :
3055 : /*
3056 : * Select a random starting point in the list. This is to minimize
3057 : * conflicts between backends that are most likely sharing the same list
3058 : * of temp tablespaces. Note that if we create multiple temp files in the
3059 : * same transaction, we'll advance circularly through the list --- this
3060 : * ensures that large temporary sort files are nicely spread across all
3061 : * available tablespaces.
3062 : */
3063 5440 : if (numSpaces > 1)
3064 0 : nextTempTableSpace = pg_prng_uint64_range(&pg_global_prng_state,
3065 0 : 0, numSpaces - 1);
3066 : else
3067 5440 : nextTempTableSpace = 0;
3068 5440 : }
3069 :
3070 : /*
3071 : * TempTablespacesAreSet
3072 : *
3073 : * Returns true if SetTempTablespaces has been called in current transaction.
3074 : * (This is just so that tablespaces.c doesn't need its own per-transaction
3075 : * state.)
3076 : */
3077 : bool
3078 8204 : TempTablespacesAreSet(void)
3079 : {
3080 8204 : return (numTempTableSpaces >= 0);
3081 : }
3082 :
3083 : /*
3084 : * GetTempTablespaces
3085 : *
3086 : * Populate an array with the OIDs of the tablespaces that should be used for
3087 : * temporary files. (Some entries may be InvalidOid, indicating that the
3088 : * current database's default tablespace should be used.) At most numSpaces
3089 : * entries will be filled.
3090 : * Returns the number of OIDs that were copied into the output array.
3091 : */
3092 : int
3093 358 : GetTempTablespaces(Oid *tableSpaces, int numSpaces)
3094 : {
3095 : int i;
3096 :
3097 : Assert(TempTablespacesAreSet());
3098 358 : for (i = 0; i < numTempTableSpaces && i < numSpaces; ++i)
3099 0 : tableSpaces[i] = tempTableSpaces[i];
3100 :
3101 358 : return i;
3102 : }
3103 :
3104 : /*
3105 : * GetNextTempTableSpace
3106 : *
3107 : * Select the next temp tablespace to use. A result of InvalidOid means
3108 : * to use the current database's default tablespace.
3109 : */
3110 : Oid
3111 3896 : GetNextTempTableSpace(void)
3112 : {
3113 3896 : if (numTempTableSpaces > 0)
3114 : {
3115 : /* Advance nextTempTableSpace counter with wraparound */
3116 2 : if (++nextTempTableSpace >= numTempTableSpaces)
3117 2 : nextTempTableSpace = 0;
3118 2 : return tempTableSpaces[nextTempTableSpace];
3119 : }
3120 3894 : return InvalidOid;
3121 : }
3122 :
3123 :
3124 : /*
3125 : * AtEOSubXact_Files
3126 : *
3127 : * Take care of subtransaction commit/abort. At abort, we close temp files
3128 : * that the subtransaction may have opened. At commit, we reassign the
3129 : * files that were opened to the parent subtransaction.
3130 : */
3131 : void
3132 18022 : AtEOSubXact_Files(bool isCommit, SubTransactionId mySubid,
3133 : SubTransactionId parentSubid)
3134 : {
3135 : Index i;
3136 :
3137 18022 : for (i = 0; i < numAllocatedDescs; i++)
3138 : {
3139 0 : if (allocatedDescs[i].create_subid == mySubid)
3140 : {
3141 0 : if (isCommit)
3142 0 : allocatedDescs[i].create_subid = parentSubid;
3143 : else
3144 : {
3145 : /* have to recheck the item after FreeDesc (ugly) */
3146 0 : FreeDesc(&allocatedDescs[i--]);
3147 : }
3148 : }
3149 : }
3150 18022 : }
3151 :
3152 : /*
3153 : * AtEOXact_Files
3154 : *
3155 : * This routine is called during transaction commit or abort. All still-open
3156 : * per-transaction temporary file VFDs are closed, which also causes the
3157 : * underlying files to be deleted (although they should've been closed already
3158 : * by the ResourceOwner cleanup). Furthermore, all "allocated" stdio files are
3159 : * closed. We also forget any transaction-local temp tablespace list.
3160 : *
3161 : * The isCommit flag is used only to decide whether to emit warnings about
3162 : * unclosed files.
3163 : */
3164 : void
3165 571468 : AtEOXact_Files(bool isCommit)
3166 : {
3167 571468 : CleanupTempFiles(isCommit, false);
3168 571468 : tempTableSpaces = NULL;
3169 571468 : numTempTableSpaces = -1;
3170 571468 : }
3171 :
3172 : /*
3173 : * BeforeShmemExit_Files
3174 : *
3175 : * before_shmem_exit hook to clean up temp files during backend shutdown.
3176 : * Here, we want to clean up *all* temp files including interXact ones.
3177 : */
3178 : static void
3179 32562 : BeforeShmemExit_Files(int code, Datum arg)
3180 : {
3181 32562 : CleanupTempFiles(false, true);
3182 :
3183 : /* prevent further temp files from being created */
3184 : #ifdef USE_ASSERT_CHECKING
3185 : temporary_files_allowed = false;
3186 : #endif
3187 32562 : }
3188 :
3189 : /*
3190 : * Close temporary files and delete their underlying files.
3191 : *
3192 : * isCommit: if true, this is normal transaction commit, and we don't
3193 : * expect any remaining files; warn if there are some.
3194 : *
3195 : * isProcExit: if true, this is being called as the backend process is
3196 : * exiting. If that's the case, we should remove all temporary files; if
3197 : * that's not the case, we are being called for transaction commit/abort
3198 : * and should only remove transaction-local temp files. In either case,
3199 : * also clean up "allocated" stdio files, dirs and fds.
3200 : */
3201 : static void
3202 604030 : CleanupTempFiles(bool isCommit, bool isProcExit)
3203 : {
3204 : Index i;
3205 :
3206 : /*
3207 : * Careful here: at proc_exit we need extra cleanup, not just
3208 : * xact_temporary files.
3209 : */
3210 604030 : if (isProcExit || have_xact_temporary_files)
3211 : {
3212 : Assert(FileIsNotOpen(0)); /* Make sure ring not corrupted */
3213 1881432 : for (i = 1; i < SizeVfdCache; i++)
3214 : {
3215 1847434 : unsigned short fdstate = VfdCache[i].fdstate;
3216 :
3217 1847434 : if (((fdstate & FD_DELETE_AT_CLOSE) || (fdstate & FD_CLOSE_AT_EOXACT)) &&
3218 10 : VfdCache[i].fileName != NULL)
3219 : {
3220 : /*
3221 : * If we're in the process of exiting a backend process, close
3222 : * all temporary files. Otherwise, only close temporary files
3223 : * local to the current transaction. They should be closed by
3224 : * the ResourceOwner mechanism already, so this is just a
3225 : * debugging cross-check.
3226 : */
3227 10 : if (isProcExit)
3228 10 : FileClose(i);
3229 0 : else if (fdstate & FD_CLOSE_AT_EOXACT)
3230 : {
3231 0 : elog(WARNING,
3232 : "temporary file %s not closed at end-of-transaction",
3233 : VfdCache[i].fileName);
3234 0 : FileClose(i);
3235 : }
3236 : }
3237 : }
3238 :
3239 33998 : have_xact_temporary_files = false;
3240 : }
3241 :
3242 : /* Complain if any allocated files remain open at commit. */
3243 604030 : if (isCommit && numAllocatedDescs > 0)
3244 0 : elog(WARNING, "%d temporary files and directories not closed at end-of-transaction",
3245 : numAllocatedDescs);
3246 :
3247 : /* Clean up "allocated" stdio files, dirs and fds. */
3248 604274 : while (numAllocatedDescs > 0)
3249 244 : FreeDesc(&allocatedDescs[0]);
3250 604030 : }
3251 :
3252 :
3253 : /*
3254 : * Remove temporary and temporary relation files left over from a prior
3255 : * postmaster session
3256 : *
3257 : * This should be called during postmaster startup. It will forcibly
3258 : * remove any leftover files created by OpenTemporaryFile and any leftover
3259 : * temporary relation files created by mdcreate.
3260 : *
3261 : * During post-backend-crash restart cycle, this routine is called when
3262 : * remove_temp_files_after_crash GUC is enabled. Multiple crashes while
3263 : * queries are using temp files could result in useless storage usage that can
3264 : * only be reclaimed by a service restart. The argument against enabling it is
3265 : * that someone might want to examine the temporary files for debugging
3266 : * purposes. This does however mean that OpenTemporaryFile had better allow for
3267 : * collision with an existing temp file name.
3268 : *
3269 : * NOTE: this function and its subroutines generally report syscall failures
3270 : * with ereport(LOG) and keep going. Removing temp files is not so critical
3271 : * that we should fail to start the database when we can't do it.
3272 : */
3273 : void
3274 1428 : RemovePgTempFiles(void)
3275 : {
3276 : char temp_path[MAXPGPATH + 10 + sizeof(TABLESPACE_VERSION_DIRECTORY) + sizeof(PG_TEMP_FILES_DIR)];
3277 : DIR *spc_dir;
3278 : struct dirent *spc_de;
3279 :
3280 : /*
3281 : * First process temp files in pg_default ($PGDATA/base)
3282 : */
3283 1428 : snprintf(temp_path, sizeof(temp_path), "base/%s", PG_TEMP_FILES_DIR);
3284 1428 : RemovePgTempFilesInDir(temp_path, true, false);
3285 1428 : RemovePgTempRelationFiles("base");
3286 :
3287 : /*
3288 : * Cycle through temp directories for all non-default tablespaces.
3289 : */
3290 1428 : spc_dir = AllocateDir("pg_tblspc");
3291 :
3292 4404 : while ((spc_de = ReadDirExtended(spc_dir, "pg_tblspc", LOG)) != NULL)
3293 : {
3294 2976 : if (strcmp(spc_de->d_name, ".") == 0 ||
3295 1548 : strcmp(spc_de->d_name, "..") == 0)
3296 2856 : continue;
3297 :
3298 120 : snprintf(temp_path, sizeof(temp_path), "pg_tblspc/%s/%s/%s",
3299 120 : spc_de->d_name, TABLESPACE_VERSION_DIRECTORY, PG_TEMP_FILES_DIR);
3300 120 : RemovePgTempFilesInDir(temp_path, true, false);
3301 :
3302 120 : snprintf(temp_path, sizeof(temp_path), "pg_tblspc/%s/%s",
3303 120 : spc_de->d_name, TABLESPACE_VERSION_DIRECTORY);
3304 120 : RemovePgTempRelationFiles(temp_path);
3305 : }
3306 :
3307 1428 : FreeDir(spc_dir);
3308 :
3309 : /*
3310 : * In EXEC_BACKEND case there is a pgsql_tmp directory at the top level of
3311 : * DataDir as well. However, that is *not* cleaned here because doing so
3312 : * would create a race condition. It's done separately, earlier in
3313 : * postmaster startup.
3314 : */
3315 1428 : }
3316 :
3317 : /*
3318 : * Process one pgsql_tmp directory for RemovePgTempFiles.
3319 : *
3320 : * If missing_ok is true, it's all right for the named directory to not exist.
3321 : * Any other problem results in a LOG message. (missing_ok should be true at
3322 : * the top level, since pgsql_tmp directories are not created until needed.)
3323 : *
3324 : * At the top level, this should be called with unlink_all = false, so that
3325 : * only files matching the temporary name prefix will be unlinked. When
3326 : * recursing it will be called with unlink_all = true to unlink everything
3327 : * under a top-level temporary directory.
3328 : *
3329 : * (These two flags could be replaced by one, but it seems clearer to keep
3330 : * them separate.)
3331 : */
3332 : void
3333 1550 : RemovePgTempFilesInDir(const char *tmpdirname, bool missing_ok, bool unlink_all)
3334 : {
3335 : DIR *temp_dir;
3336 : struct dirent *temp_de;
3337 : char rm_path[MAXPGPATH * 2];
3338 :
3339 1550 : temp_dir = AllocateDir(tmpdirname);
3340 :
3341 1550 : if (temp_dir == NULL && errno == ENOENT && missing_ok)
3342 1434 : return;
3343 :
3344 354 : while ((temp_de = ReadDirExtended(temp_dir, tmpdirname, LOG)) != NULL)
3345 : {
3346 238 : if (strcmp(temp_de->d_name, ".") == 0 ||
3347 122 : strcmp(temp_de->d_name, "..") == 0)
3348 232 : continue;
3349 :
3350 6 : snprintf(rm_path, sizeof(rm_path), "%s/%s",
3351 6 : tmpdirname, temp_de->d_name);
3352 :
3353 6 : if (unlink_all ||
3354 6 : strncmp(temp_de->d_name,
3355 : PG_TEMP_FILE_PREFIX,
3356 : strlen(PG_TEMP_FILE_PREFIX)) == 0)
3357 6 : {
3358 6 : PGFileType type = get_dirent_type(rm_path, temp_de, false, LOG);
3359 :
3360 6 : if (type == PGFILETYPE_ERROR)
3361 0 : continue;
3362 6 : else if (type == PGFILETYPE_DIR)
3363 : {
3364 : /* recursively remove contents, then directory itself */
3365 2 : RemovePgTempFilesInDir(rm_path, false, true);
3366 :
3367 2 : if (rmdir(rm_path) < 0)
3368 0 : ereport(LOG,
3369 : (errcode_for_file_access(),
3370 : errmsg("could not remove directory \"%s\": %m",
3371 : rm_path)));
3372 : }
3373 : else
3374 : {
3375 4 : if (unlink(rm_path) < 0)
3376 0 : ereport(LOG,
3377 : (errcode_for_file_access(),
3378 : errmsg("could not remove file \"%s\": %m",
3379 : rm_path)));
3380 : }
3381 : }
3382 : else
3383 0 : ereport(LOG,
3384 : (errmsg("unexpected file found in temporary-files directory: \"%s\"",
3385 : rm_path)));
3386 : }
3387 :
3388 116 : FreeDir(temp_dir);
3389 : }
3390 :
3391 : /* Process one tablespace directory, look for per-DB subdirectories */
3392 : static void
3393 1548 : RemovePgTempRelationFiles(const char *tsdirname)
3394 : {
3395 : DIR *ts_dir;
3396 : struct dirent *de;
3397 : char dbspace_path[MAXPGPATH * 2];
3398 :
3399 1548 : ts_dir = AllocateDir(tsdirname);
3400 :
3401 9632 : while ((de = ReadDirExtended(ts_dir, tsdirname, LOG)) != NULL)
3402 : {
3403 : /*
3404 : * We're only interested in the per-database directories, which have
3405 : * numeric names. Note that this code will also (properly) ignore "."
3406 : * and "..".
3407 : */
3408 8084 : if (strspn(de->d_name, "0123456789") != strlen(de->d_name))
3409 3210 : continue;
3410 :
3411 4874 : snprintf(dbspace_path, sizeof(dbspace_path), "%s/%s",
3412 4874 : tsdirname, de->d_name);
3413 4874 : RemovePgTempRelationFilesInDbspace(dbspace_path);
3414 : }
3415 :
3416 1548 : FreeDir(ts_dir);
3417 1548 : }
3418 :
3419 : /* Process one per-dbspace directory for RemovePgTempRelationFiles */
3420 : static void
3421 4874 : RemovePgTempRelationFilesInDbspace(const char *dbspacedirname)
3422 : {
3423 : DIR *dbspace_dir;
3424 : struct dirent *de;
3425 : char rm_path[MAXPGPATH * 2];
3426 :
3427 4874 : dbspace_dir = AllocateDir(dbspacedirname);
3428 :
3429 1465784 : while ((de = ReadDirExtended(dbspace_dir, dbspacedirname, LOG)) != NULL)
3430 : {
3431 1460910 : if (!looks_like_temp_rel_name(de->d_name))
3432 1460902 : continue;
3433 :
3434 8 : snprintf(rm_path, sizeof(rm_path), "%s/%s",
3435 8 : dbspacedirname, de->d_name);
3436 :
3437 8 : if (unlink(rm_path) < 0)
3438 0 : ereport(LOG,
3439 : (errcode_for_file_access(),
3440 : errmsg("could not remove file \"%s\": %m",
3441 : rm_path)));
3442 : }
3443 :
3444 4874 : FreeDir(dbspace_dir);
3445 4874 : }
3446 :
3447 : /* t<digits>_<digits>, or t<digits>_<digits>_<forkname> */
3448 : bool
3449 1974740 : looks_like_temp_rel_name(const char *name)
3450 : {
3451 : int pos;
3452 : int savepos;
3453 :
3454 : /* Must start with "t". */
3455 1974740 : if (name[0] != 't')
3456 1974660 : return false;
3457 :
3458 : /* Followed by a non-empty string of digits and then an underscore. */
3459 392 : for (pos = 1; isdigit((unsigned char) name[pos]); ++pos)
3460 : ;
3461 80 : if (pos == 1 || name[pos] != '_')
3462 0 : return false;
3463 :
3464 : /* Followed by another nonempty string of digits. */
3465 392 : for (savepos = ++pos; isdigit((unsigned char) name[pos]); ++pos)
3466 : ;
3467 80 : if (savepos == pos)
3468 0 : return false;
3469 :
3470 : /* We might have _forkname or .segment or both. */
3471 80 : if (name[pos] == '_')
3472 : {
3473 40 : int forkchar = forkname_chars(&name[pos + 1], NULL);
3474 :
3475 40 : if (forkchar <= 0)
3476 0 : return false;
3477 40 : pos += forkchar + 1;
3478 : }
3479 80 : if (name[pos] == '.')
3480 : {
3481 : int segchar;
3482 :
3483 80 : for (segchar = 1; isdigit((unsigned char) name[pos + segchar]); ++segchar)
3484 : ;
3485 40 : if (segchar <= 1)
3486 0 : return false;
3487 40 : pos += segchar;
3488 : }
3489 :
3490 : /* Now we should be at the end. */
3491 80 : if (name[pos] != '\0')
3492 0 : return false;
3493 80 : return true;
3494 : }
3495 :
3496 : #ifdef HAVE_SYNCFS
3497 : static void
3498 0 : do_syncfs(const char *path)
3499 : {
3500 : int fd;
3501 :
3502 0 : ereport_startup_progress("syncing data directory (syncfs), elapsed time: %ld.%02d s, current path: %s",
3503 : path);
3504 :
3505 0 : fd = OpenTransientFile(path, O_RDONLY);
3506 0 : if (fd < 0)
3507 : {
3508 0 : ereport(LOG,
3509 : (errcode_for_file_access(),
3510 : errmsg("could not open file \"%s\": %m", path)));
3511 0 : return;
3512 : }
3513 0 : if (syncfs(fd) < 0)
3514 0 : ereport(LOG,
3515 : (errcode_for_file_access(),
3516 : errmsg("could not synchronize file system for file \"%s\": %m", path)));
3517 0 : CloseTransientFile(fd);
3518 : }
3519 : #endif
3520 :
3521 : /*
3522 : * Issue fsync recursively on PGDATA and all its contents, or issue syncfs for
3523 : * all potential filesystem, depending on recovery_init_sync_method setting.
3524 : *
3525 : * We fsync regular files and directories wherever they are, but we
3526 : * follow symlinks only for pg_wal and immediately under pg_tblspc.
3527 : * Other symlinks are presumed to point at files we're not responsible
3528 : * for fsyncing, and might not have privileges to write at all.
3529 : *
3530 : * Errors are logged but not considered fatal; that's because this is used
3531 : * only during database startup, to deal with the possibility that there are
3532 : * issued-but-unsynced writes pending against the data directory. We want to
3533 : * ensure that such writes reach disk before anything that's done in the new
3534 : * run. However, aborting on error would result in failure to start for
3535 : * harmless cases such as read-only files in the data directory, and that's
3536 : * not good either.
3537 : *
3538 : * Note that if we previously crashed due to a PANIC on fsync(), we'll be
3539 : * rewriting all changes again during recovery.
3540 : *
3541 : * Note we assume we're chdir'd into PGDATA to begin with.
3542 : */
3543 : void
3544 326 : SyncDataDirectory(void)
3545 : {
3546 : bool xlog_is_symlink;
3547 :
3548 : /* We can skip this whole thing if fsync is disabled. */
3549 326 : if (!enableFsync)
3550 326 : return;
3551 :
3552 : /*
3553 : * If pg_wal is a symlink, we'll need to recurse into it separately,
3554 : * because the first walkdir below will ignore it.
3555 : */
3556 0 : xlog_is_symlink = false;
3557 :
3558 : {
3559 : struct stat st;
3560 :
3561 0 : if (lstat("pg_wal", &st) < 0)
3562 0 : ereport(LOG,
3563 : (errcode_for_file_access(),
3564 : errmsg("could not stat file \"%s\": %m",
3565 : "pg_wal")));
3566 0 : else if (S_ISLNK(st.st_mode))
3567 0 : xlog_is_symlink = true;
3568 : }
3569 :
3570 : #ifdef HAVE_SYNCFS
3571 0 : if (recovery_init_sync_method == DATA_DIR_SYNC_METHOD_SYNCFS)
3572 : {
3573 : DIR *dir;
3574 : struct dirent *de;
3575 :
3576 : /*
3577 : * On Linux, we don't have to open every single file one by one. We
3578 : * can use syncfs() to sync whole filesystems. We only expect
3579 : * filesystem boundaries to exist where we tolerate symlinks, namely
3580 : * pg_wal and the tablespaces, so we call syncfs() for each of those
3581 : * directories.
3582 : */
3583 :
3584 : /* Prepare to report progress syncing the data directory via syncfs. */
3585 0 : begin_startup_progress_phase();
3586 :
3587 : /* Sync the top level pgdata directory. */
3588 0 : do_syncfs(".");
3589 : /* If any tablespaces are configured, sync each of those. */
3590 0 : dir = AllocateDir("pg_tblspc");
3591 0 : while ((de = ReadDirExtended(dir, "pg_tblspc", LOG)))
3592 : {
3593 : char path[MAXPGPATH];
3594 :
3595 0 : if (strcmp(de->d_name, ".") == 0 || strcmp(de->d_name, "..") == 0)
3596 0 : continue;
3597 :
3598 0 : snprintf(path, MAXPGPATH, "pg_tblspc/%s", de->d_name);
3599 0 : do_syncfs(path);
3600 : }
3601 0 : FreeDir(dir);
3602 : /* If pg_wal is a symlink, process that too. */
3603 0 : if (xlog_is_symlink)
3604 0 : do_syncfs("pg_wal");
3605 0 : return;
3606 : }
3607 : #endif /* !HAVE_SYNCFS */
3608 :
3609 : #ifdef PG_FLUSH_DATA_WORKS
3610 : /* Prepare to report progress of the pre-fsync phase. */
3611 0 : begin_startup_progress_phase();
3612 :
3613 : /*
3614 : * If possible, hint to the kernel that we're soon going to fsync the data
3615 : * directory and its contents. Errors in this step are even less
3616 : * interesting than normal, so log them only at DEBUG1.
3617 : */
3618 0 : walkdir(".", pre_sync_fname, false, DEBUG1);
3619 0 : if (xlog_is_symlink)
3620 0 : walkdir("pg_wal", pre_sync_fname, false, DEBUG1);
3621 0 : walkdir("pg_tblspc", pre_sync_fname, true, DEBUG1);
3622 : #endif
3623 :
3624 : /* Prepare to report progress syncing the data directory via fsync. */
3625 0 : begin_startup_progress_phase();
3626 :
3627 : /*
3628 : * Now we do the fsync()s in the same order.
3629 : *
3630 : * The main call ignores symlinks, so in addition to specially processing
3631 : * pg_wal if it's a symlink, pg_tblspc has to be visited separately with
3632 : * process_symlinks = true. Note that if there are any plain directories
3633 : * in pg_tblspc, they'll get fsync'd twice. That's not an expected case
3634 : * so we don't worry about optimizing it.
3635 : */
3636 0 : walkdir(".", datadir_fsync_fname, false, LOG);
3637 0 : if (xlog_is_symlink)
3638 0 : walkdir("pg_wal", datadir_fsync_fname, false, LOG);
3639 0 : walkdir("pg_tblspc", datadir_fsync_fname, true, LOG);
3640 : }
3641 :
3642 : /*
3643 : * walkdir: recursively walk a directory, applying the action to each
3644 : * regular file and directory (including the named directory itself).
3645 : *
3646 : * If process_symlinks is true, the action and recursion are also applied
3647 : * to regular files and directories that are pointed to by symlinks in the
3648 : * given directory; otherwise symlinks are ignored. Symlinks are always
3649 : * ignored in subdirectories, ie we intentionally don't pass down the
3650 : * process_symlinks flag to recursive calls.
3651 : *
3652 : * Errors are reported at level elevel, which might be ERROR or less.
3653 : *
3654 : * See also walkdir in file_utils.c, which is a frontend version of this
3655 : * logic.
3656 : */
3657 : static void
3658 326 : walkdir(const char *path,
3659 : void (*action) (const char *fname, bool isdir, int elevel),
3660 : bool process_symlinks,
3661 : int elevel)
3662 : {
3663 : DIR *dir;
3664 : struct dirent *de;
3665 :
3666 326 : dir = AllocateDir(path);
3667 :
3668 3162 : while ((de = ReadDirExtended(dir, path, elevel)) != NULL)
3669 : {
3670 : char subpath[MAXPGPATH * 2];
3671 :
3672 2836 : CHECK_FOR_INTERRUPTS();
3673 :
3674 2836 : if (strcmp(de->d_name, ".") == 0 ||
3675 2510 : strcmp(de->d_name, "..") == 0)
3676 652 : continue;
3677 :
3678 2184 : snprintf(subpath, sizeof(subpath), "%s/%s", path, de->d_name);
3679 :
3680 2184 : switch (get_dirent_type(subpath, de, process_symlinks, elevel))
3681 : {
3682 2184 : case PGFILETYPE_REG:
3683 2184 : (*action) (subpath, false, elevel);
3684 2184 : break;
3685 0 : case PGFILETYPE_DIR:
3686 0 : walkdir(subpath, action, false, elevel);
3687 0 : break;
3688 0 : default:
3689 :
3690 : /*
3691 : * Errors are already reported directly by get_dirent_type(),
3692 : * and any remaining symlinks and unknown file types are
3693 : * ignored.
3694 : */
3695 0 : break;
3696 : }
3697 : }
3698 :
3699 326 : FreeDir(dir); /* we ignore any error here */
3700 :
3701 : /*
3702 : * It's important to fsync the destination directory itself as individual
3703 : * file fsyncs don't guarantee that the directory entry for the file is
3704 : * synced. However, skip this if AllocateDir failed; the action function
3705 : * might not be robust against that.
3706 : */
3707 326 : if (dir)
3708 326 : (*action) (path, true, elevel);
3709 326 : }
3710 :
3711 :
3712 : /*
3713 : * Hint to the OS that it should get ready to fsync() this file.
3714 : *
3715 : * Ignores errors trying to open unreadable files, and logs other errors at a
3716 : * caller-specified level.
3717 : */
3718 : #ifdef PG_FLUSH_DATA_WORKS
3719 :
3720 : static void
3721 0 : pre_sync_fname(const char *fname, bool isdir, int elevel)
3722 : {
3723 : int fd;
3724 :
3725 : /* Don't try to flush directories, it'll likely just fail */
3726 0 : if (isdir)
3727 0 : return;
3728 :
3729 0 : ereport_startup_progress("syncing data directory (pre-fsync), elapsed time: %ld.%02d s, current path: %s",
3730 : fname);
3731 :
3732 0 : fd = OpenTransientFile(fname, O_RDONLY | PG_BINARY);
3733 :
3734 0 : if (fd < 0)
3735 : {
3736 0 : if (errno == EACCES)
3737 0 : return;
3738 0 : ereport(elevel,
3739 : (errcode_for_file_access(),
3740 : errmsg("could not open file \"%s\": %m", fname)));
3741 0 : return;
3742 : }
3743 :
3744 : /*
3745 : * pg_flush_data() ignores errors, which is ok because this is only a
3746 : * hint.
3747 : */
3748 0 : pg_flush_data(fd, 0, 0);
3749 :
3750 0 : if (CloseTransientFile(fd) != 0)
3751 0 : ereport(elevel,
3752 : (errcode_for_file_access(),
3753 : errmsg("could not close file \"%s\": %m", fname)));
3754 : }
3755 :
3756 : #endif /* PG_FLUSH_DATA_WORKS */
3757 :
3758 : static void
3759 0 : datadir_fsync_fname(const char *fname, bool isdir, int elevel)
3760 : {
3761 0 : ereport_startup_progress("syncing data directory (fsync), elapsed time: %ld.%02d s, current path: %s",
3762 : fname);
3763 :
3764 : /*
3765 : * We want to silently ignoring errors about unreadable files. Pass that
3766 : * desire on to fsync_fname_ext().
3767 : */
3768 0 : fsync_fname_ext(fname, isdir, true, elevel);
3769 0 : }
3770 :
3771 : static void
3772 2510 : unlink_if_exists_fname(const char *fname, bool isdir, int elevel)
3773 : {
3774 2510 : if (isdir)
3775 : {
3776 326 : if (rmdir(fname) != 0 && errno != ENOENT)
3777 0 : ereport(elevel,
3778 : (errcode_for_file_access(),
3779 : errmsg("could not remove directory \"%s\": %m", fname)));
3780 : }
3781 : else
3782 : {
3783 : /* Use PathNameDeleteTemporaryFile to report filesize */
3784 2184 : PathNameDeleteTemporaryFile(fname, false);
3785 : }
3786 2510 : }
3787 :
3788 : /*
3789 : * fsync_fname_ext -- Try to fsync a file or directory
3790 : *
3791 : * If ignore_perm is true, ignore errors upon trying to open unreadable
3792 : * files. Logs other errors at a caller-specified level.
3793 : *
3794 : * Returns 0 if the operation succeeded, -1 otherwise.
3795 : */
3796 : int
3797 43888 : fsync_fname_ext(const char *fname, bool isdir, bool ignore_perm, int elevel)
3798 : {
3799 : int fd;
3800 : int flags;
3801 : int returncode;
3802 :
3803 : /*
3804 : * Some OSs require directories to be opened read-only whereas other
3805 : * systems don't allow us to fsync files opened read-only; so we need both
3806 : * cases here. Using O_RDWR will cause us to fail to fsync files that are
3807 : * not writable by our userid, but we assume that's OK.
3808 : */
3809 43888 : flags = PG_BINARY;
3810 43888 : if (!isdir)
3811 14588 : flags |= O_RDWR;
3812 : else
3813 29300 : flags |= O_RDONLY;
3814 :
3815 43888 : fd = OpenTransientFile(fname, flags);
3816 :
3817 : /*
3818 : * Some OSs don't allow us to open directories at all (Windows returns
3819 : * EACCES), just ignore the error in that case. If desired also silently
3820 : * ignoring errors about unreadable files. Log others.
3821 : */
3822 43888 : if (fd < 0 && isdir && (errno == EISDIR || errno == EACCES))
3823 0 : return 0;
3824 43888 : else if (fd < 0 && ignore_perm && errno == EACCES)
3825 0 : return 0;
3826 43888 : else if (fd < 0)
3827 : {
3828 0 : ereport(elevel,
3829 : (errcode_for_file_access(),
3830 : errmsg("could not open file \"%s\": %m", fname)));
3831 0 : return -1;
3832 : }
3833 :
3834 43888 : returncode = pg_fsync(fd);
3835 :
3836 : /*
3837 : * Some OSes don't allow us to fsync directories at all, so we can ignore
3838 : * those errors. Anything else needs to be logged.
3839 : */
3840 43888 : if (returncode != 0 && !(isdir && (errno == EBADF || errno == EINVAL)))
3841 : {
3842 : int save_errno;
3843 :
3844 : /* close file upon error, might not be in transaction context */
3845 0 : save_errno = errno;
3846 0 : (void) CloseTransientFile(fd);
3847 0 : errno = save_errno;
3848 :
3849 0 : ereport(elevel,
3850 : (errcode_for_file_access(),
3851 : errmsg("could not fsync file \"%s\": %m", fname)));
3852 0 : return -1;
3853 : }
3854 :
3855 43888 : if (CloseTransientFile(fd) != 0)
3856 : {
3857 0 : ereport(elevel,
3858 : (errcode_for_file_access(),
3859 : errmsg("could not close file \"%s\": %m", fname)));
3860 0 : return -1;
3861 : }
3862 :
3863 43888 : return 0;
3864 : }
3865 :
3866 : /*
3867 : * fsync_parent_path -- fsync the parent path of a file or directory
3868 : *
3869 : * This is aimed at making file operations persistent on disk in case of
3870 : * an OS crash or power failure.
3871 : */
3872 : static int
3873 5372 : fsync_parent_path(const char *fname, int elevel)
3874 : {
3875 : char parentpath[MAXPGPATH];
3876 :
3877 5372 : strlcpy(parentpath, fname, MAXPGPATH);
3878 5372 : get_parent_directory(parentpath);
3879 :
3880 : /*
3881 : * get_parent_directory() returns an empty string if the input argument is
3882 : * just a file name (see comments in path.c), so handle that as being the
3883 : * current directory.
3884 : */
3885 5372 : if (strlen(parentpath) == 0)
3886 342 : strlcpy(parentpath, ".", MAXPGPATH);
3887 :
3888 5372 : if (fsync_fname_ext(parentpath, true, false, elevel) != 0)
3889 0 : return -1;
3890 :
3891 5372 : return 0;
3892 : }
3893 :
3894 : /*
3895 : * Create a PostgreSQL data sub-directory
3896 : *
3897 : * The data directory itself, and most of its sub-directories, are created at
3898 : * initdb time, but we do have some occasions when we create directories in
3899 : * the backend (CREATE TABLESPACE, for example). In those cases, we want to
3900 : * make sure that those directories are created consistently. Today, that means
3901 : * making sure that the created directory has the correct permissions, which is
3902 : * what pg_dir_create_mode tracks for us.
3903 : *
3904 : * Note that we also set the umask() based on what we understand the correct
3905 : * permissions to be (see file_perm.c).
3906 : *
3907 : * For permissions other than the default, mkdir() can be used directly, but
3908 : * be sure to consider carefully such cases -- a sub-directory with incorrect
3909 : * permissions in a PostgreSQL data directory could cause backups and other
3910 : * processes to fail.
3911 : */
3912 : int
3913 2522 : MakePGDirectory(const char *directoryName)
3914 : {
3915 2522 : return mkdir(directoryName, pg_dir_create_mode);
3916 : }
3917 :
3918 : /*
3919 : * Return the passed-in error level, or PANIC if data_sync_retry is off.
3920 : *
3921 : * Failure to fsync any data file is cause for immediate panic, unless
3922 : * data_sync_retry is enabled. Data may have been written to the operating
3923 : * system and removed from our buffer pool already, and if we are running on
3924 : * an operating system that forgets dirty data on write-back failure, there
3925 : * may be only one copy of the data remaining: in the WAL. A later attempt to
3926 : * fsync again might falsely report success. Therefore we must not allow any
3927 : * further checkpoints to be attempted. data_sync_retry can in theory be
3928 : * enabled on systems known not to drop dirty buffered data on write-back
3929 : * failure (with the likely outcome that checkpoints will continue to fail
3930 : * until the underlying problem is fixed).
3931 : *
3932 : * Any code that reports a failure from fsync() or related functions should
3933 : * filter the error level with this function.
3934 : */
3935 : int
3936 28132 : data_sync_elevel(int elevel)
3937 : {
3938 28132 : return data_sync_retry ? elevel : PANIC;
3939 : }
3940 :
3941 : bool
3942 1834 : check_debug_io_direct(char **newval, void **extra, GucSource source)
3943 : {
3944 1834 : bool result = true;
3945 : int flags;
3946 :
3947 : #if PG_O_DIRECT == 0
3948 : if (strcmp(*newval, "") != 0)
3949 : {
3950 : GUC_check_errdetail("debug_io_direct is not supported on this platform.");
3951 : result = false;
3952 : }
3953 : flags = 0;
3954 : #else
3955 : List *elemlist;
3956 : ListCell *l;
3957 : char *rawstring;
3958 :
3959 : /* Need a modifiable copy of string */
3960 1834 : rawstring = pstrdup(*newval);
3961 :
3962 1834 : if (!SplitGUCList(rawstring, ',', &elemlist))
3963 : {
3964 0 : GUC_check_errdetail("Invalid list syntax in parameter %s",
3965 : "debug_io_direct");
3966 0 : pfree(rawstring);
3967 0 : list_free(elemlist);
3968 0 : return false;
3969 : }
3970 :
3971 1834 : flags = 0;
3972 1846 : foreach(l, elemlist)
3973 : {
3974 12 : char *item = (char *) lfirst(l);
3975 :
3976 12 : if (pg_strcasecmp(item, "data") == 0)
3977 4 : flags |= IO_DIRECT_DATA;
3978 8 : else if (pg_strcasecmp(item, "wal") == 0)
3979 4 : flags |= IO_DIRECT_WAL;
3980 4 : else if (pg_strcasecmp(item, "wal_init") == 0)
3981 4 : flags |= IO_DIRECT_WAL_INIT;
3982 : else
3983 : {
3984 0 : GUC_check_errdetail("Invalid option \"%s\"", item);
3985 0 : result = false;
3986 0 : break;
3987 : }
3988 : }
3989 :
3990 : /*
3991 : * It's possible to configure block sizes smaller than our assumed I/O
3992 : * alignment size, which could result in invalid I/O requests.
3993 : */
3994 : #if XLOG_BLCKSZ < PG_IO_ALIGN_SIZE
3995 : if (result && (flags & (IO_DIRECT_WAL | IO_DIRECT_WAL_INIT)))
3996 : {
3997 : GUC_check_errdetail("debug_io_direct is not supported for WAL because XLOG_BLCKSZ is too small");
3998 : result = false;
3999 : }
4000 : #endif
4001 : #if BLCKSZ < PG_IO_ALIGN_SIZE
4002 : if (result && (flags & IO_DIRECT_DATA))
4003 : {
4004 : GUC_check_errdetail("debug_io_direct is not supported for data because BLCKSZ is too small");
4005 : result = false;
4006 : }
4007 : #endif
4008 :
4009 1834 : pfree(rawstring);
4010 1834 : list_free(elemlist);
4011 : #endif
4012 :
4013 1834 : if (!result)
4014 0 : return result;
4015 :
4016 : /* Save the flags in *extra, for use by assign_debug_io_direct */
4017 1834 : *extra = guc_malloc(ERROR, sizeof(int));
4018 1834 : *((int *) *extra) = flags;
4019 :
4020 1834 : return result;
4021 : }
4022 :
4023 : extern void
4024 1834 : assign_debug_io_direct(const char *newval, void *extra)
4025 : {
4026 1834 : int *flags = (int *) extra;
4027 :
4028 1834 : io_direct_flags = *flags;
4029 1834 : }
4030 :
4031 : /* ResourceOwner callbacks */
4032 :
4033 : static void
4034 8 : ResOwnerReleaseFile(Datum res)
4035 : {
4036 8 : File file = (File) DatumGetInt32(res);
4037 : Vfd *vfdP;
4038 :
4039 : Assert(FileIsValid(file));
4040 :
4041 8 : vfdP = &VfdCache[file];
4042 8 : vfdP->resowner = NULL;
4043 :
4044 8 : FileClose(file);
4045 8 : }
4046 :
4047 : static char *
4048 0 : ResOwnerPrintFile(Datum res)
4049 : {
4050 0 : return psprintf("File %d", DatumGetInt32(res));
4051 : }
|