Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * fd.c
4 : * Virtual file descriptor code.
5 : *
6 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : * IDENTIFICATION
10 : * src/backend/storage/file/fd.c
11 : *
12 : * NOTES:
13 : *
14 : * This code manages a cache of 'virtual' file descriptors (VFDs).
15 : * The server opens many file descriptors for a variety of reasons,
16 : * including base tables, scratch files (e.g., sort and hash spool
17 : * files), and random calls to C library routines like system(3); it
18 : * is quite easy to exceed system limits on the number of open files a
19 : * single process can have. (This is around 1024 on many modern
20 : * operating systems, but may be lower on others.)
21 : *
22 : * VFDs are managed as an LRU pool, with actual OS file descriptors
23 : * being opened and closed as needed. Obviously, if a routine is
24 : * opened using these interfaces, all subsequent operations must also
25 : * be through these interfaces (the File type is not a real file
26 : * descriptor).
27 : *
28 : * For this scheme to work, most (if not all) routines throughout the
29 : * server should use these interfaces instead of calling the C library
30 : * routines (e.g., open(2) and fopen(3)) themselves. Otherwise, we
31 : * may find ourselves short of real file descriptors anyway.
32 : *
33 : * INTERFACE ROUTINES
34 : *
35 : * PathNameOpenFile and OpenTemporaryFile are used to open virtual files.
36 : * A File opened with OpenTemporaryFile is automatically deleted when the
37 : * File is closed, either explicitly or implicitly at end of transaction or
38 : * process exit. PathNameOpenFile is intended for files that are held open
39 : * for a long time, like relation files. It is the caller's responsibility
40 : * to close them, there is no automatic mechanism in fd.c for that.
41 : *
42 : * PathName(Create|Open|Delete)Temporary(File|Dir) are used to manage
43 : * temporary files that have names so that they can be shared between
44 : * backends. Such files are automatically closed and count against the
45 : * temporary file limit of the backend that creates them, but unlike anonymous
46 : * files they are not automatically deleted. See sharedfileset.c for a shared
47 : * ownership mechanism that provides automatic cleanup for shared files when
48 : * the last of a group of backends detaches.
49 : *
50 : * AllocateFile, AllocateDir, OpenPipeStream and OpenTransientFile are
51 : * wrappers around fopen(3), opendir(3), popen(3) and open(2), respectively.
52 : * They behave like the corresponding native functions, except that the handle
53 : * is registered with the current subtransaction, and will be automatically
54 : * closed at abort. These are intended mainly for short operations like
55 : * reading a configuration file; there is a limit on the number of files that
56 : * can be opened using these functions at any one time.
57 : *
58 : * Finally, BasicOpenFile is just a thin wrapper around open() that can
59 : * release file descriptors in use by the virtual file descriptors if
60 : * necessary. There is no automatic cleanup of file descriptors returned by
61 : * BasicOpenFile, it is solely the caller's responsibility to close the file
62 : * descriptor by calling close(2).
63 : *
64 : * If a non-virtual file descriptor needs to be held open for any length of
65 : * time, report it to fd.c by calling AcquireExternalFD or ReserveExternalFD
66 : * (and eventually ReleaseExternalFD), so that we can take it into account
67 : * while deciding how many VFDs can be open. This applies to FDs obtained
68 : * with BasicOpenFile as well as those obtained without use of any fd.c API.
69 : *
70 : *-------------------------------------------------------------------------
71 : */
72 :
73 : #include "postgres.h"
74 :
75 : #include <dirent.h>
76 : #include <sys/file.h>
77 : #include <sys/param.h>
78 : #include <sys/resource.h> /* for getrlimit */
79 : #include <sys/stat.h>
80 : #include <sys/types.h>
81 : #ifndef WIN32
82 : #include <sys/mman.h>
83 : #endif
84 : #include <limits.h>
85 : #include <unistd.h>
86 : #include <fcntl.h>
87 :
88 : #include "access/xact.h"
89 : #include "access/xlog.h"
90 : #include "catalog/pg_tablespace.h"
91 : #include "common/file_perm.h"
92 : #include "common/file_utils.h"
93 : #include "common/pg_prng.h"
94 : #include "miscadmin.h"
95 : #include "pgstat.h"
96 : #include "postmaster/startup.h"
97 : #include "storage/aio.h"
98 : #include "storage/fd.h"
99 : #include "storage/ipc.h"
100 : #include "utils/guc.h"
101 : #include "utils/guc_hooks.h"
102 : #include "utils/resowner.h"
103 : #include "utils/varlena.h"
104 : #include "utils/wait_event.h"
105 :
106 : /* Define PG_FLUSH_DATA_WORKS if we have an implementation for pg_flush_data */
107 : #if defined(HAVE_SYNC_FILE_RANGE)
108 : #define PG_FLUSH_DATA_WORKS 1
109 : #elif !defined(WIN32) && defined(MS_ASYNC)
110 : #define PG_FLUSH_DATA_WORKS 1
111 : #elif defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_DONTNEED)
112 : #define PG_FLUSH_DATA_WORKS 1
113 : #endif
114 :
115 : /*
116 : * We must leave some file descriptors free for system(), the dynamic loader,
117 : * and other code that tries to open files without consulting fd.c. This
118 : * is the number left free. (While we try fairly hard to prevent EMFILE
119 : * errors, there's never any guarantee that we won't get ENFILE due to
120 : * other processes chewing up FDs. So it's a bad idea to try to open files
121 : * without consulting fd.c. Nonetheless we cannot control all code.)
122 : *
123 : * Because this is just a fixed setting, we are effectively assuming that
124 : * no such code will leave FDs open over the long term; otherwise the slop
125 : * is likely to be insufficient. Note in particular that we expect that
126 : * loading a shared library does not result in any permanent increase in
127 : * the number of open files. (This appears to be true on most if not
128 : * all platforms as of Feb 2004.)
129 : */
130 : #define NUM_RESERVED_FDS 10
131 :
132 : /*
133 : * If we have fewer than this many usable FDs after allowing for the reserved
134 : * ones, choke. (This value is chosen to work with "ulimit -n 64", but not
135 : * much less than that. Note that this value ensures numExternalFDs can be
136 : * at least 16; as of this writing, the contrib/postgres_fdw regression tests
137 : * will not pass unless that can grow to at least 14.)
138 : */
139 : #define FD_MINFREE 48
140 :
141 : /*
142 : * A number of platforms allow individual processes to open many more files
143 : * than they can really support when *many* processes do the same thing.
144 : * This GUC parameter lets the DBA limit max_safe_fds to something less than
145 : * what the postmaster's initial probe suggests will work.
146 : */
147 : int max_files_per_process = 1000;
148 :
149 : /*
150 : * Maximum number of file descriptors to open for operations that fd.c knows
151 : * about (VFDs, AllocateFile etc, or "external" FDs). This is initialized
152 : * to a conservative value, and remains that way indefinitely in bootstrap or
153 : * standalone-backend cases. In normal postmaster operation, the postmaster
154 : * calls set_max_safe_fds() late in initialization to update the value, and
155 : * that value is then inherited by forked subprocesses.
156 : *
157 : * Note: the value of max_files_per_process is taken into account while
158 : * setting this variable, and so need not be tested separately.
159 : */
160 : int max_safe_fds = FD_MINFREE; /* default if not changed */
161 :
162 : /* Whether it is safe to continue running after fsync() fails. */
163 : bool data_sync_retry = false;
164 :
165 : /* How SyncDataDirectory() should do its job. */
166 : int recovery_init_sync_method = DATA_DIR_SYNC_METHOD_FSYNC;
167 :
168 : /* How data files should be bulk-extended with zeros. */
169 : int file_extend_method = DEFAULT_FILE_EXTEND_METHOD;
170 :
171 : /* Which kinds of files should be opened with PG_O_DIRECT. */
172 : int io_direct_flags;
173 :
174 : /* Debugging.... */
175 :
176 : #ifdef FDDEBUG
177 : #define DO_DB(A) \
178 : do { \
179 : int _do_db_save_errno = errno; \
180 : A; \
181 : errno = _do_db_save_errno; \
182 : } while (0)
183 : #else
184 : #define DO_DB(A) \
185 : ((void) 0)
186 : #endif
187 :
188 : #define VFD_CLOSED (-1)
189 :
190 : #define FileIsValid(file) \
191 : ((file) > 0 && (file) < (int) SizeVfdCache && VfdCache[file].fileName != NULL)
192 :
193 : #define FileIsNotOpen(file) (VfdCache[file].fd == VFD_CLOSED)
194 :
195 : /* these are the assigned bits in fdstate below: */
196 : #define FD_DELETE_AT_CLOSE (1 << 0) /* T = delete when closed */
197 : #define FD_CLOSE_AT_EOXACT (1 << 1) /* T = close at eoXact */
198 : #define FD_TEMP_FILE_LIMIT (1 << 2) /* T = respect temp_file_limit */
199 :
200 : typedef struct vfd
201 : {
202 : int fd; /* current FD, or VFD_CLOSED if none */
203 : unsigned short fdstate; /* bitflags for VFD's state */
204 : ResourceOwner resowner; /* owner, for automatic cleanup */
205 : File nextFree; /* link to next free VFD, if in freelist */
206 : File lruMoreRecently; /* doubly linked recency-of-use list */
207 : File lruLessRecently;
208 : pgoff_t fileSize; /* current size of file (0 if not temporary) */
209 : char *fileName; /* name of file, or NULL for unused VFD */
210 : /* NB: fileName is malloc'd, and must be free'd when closing the VFD */
211 : int fileFlags; /* open(2) flags for (re)opening the file */
212 : mode_t fileMode; /* mode to pass to open(2) */
213 : } Vfd;
214 :
215 : /*
216 : * Virtual File Descriptor array pointer and size. This grows as
217 : * needed. 'File' values are indexes into this array.
218 : * Note that VfdCache[0] is not a usable VFD, just a list header.
219 : */
220 : static Vfd *VfdCache;
221 : static Size SizeVfdCache = 0;
222 :
223 : /*
224 : * Number of file descriptors known to be in use by VFD entries.
225 : */
226 : static int nfile = 0;
227 :
228 : /*
229 : * Flag to tell whether it's worth scanning VfdCache looking for temp files
230 : * to close
231 : */
232 : static bool have_xact_temporary_files = false;
233 :
234 : /*
235 : * Tracks the total size of all temporary files. Note: when temp_file_limit
236 : * is being enforced, this cannot overflow since the limit cannot be more
237 : * than INT_MAX kilobytes. When not enforcing, it could theoretically
238 : * overflow, but we don't care.
239 : */
240 : static uint64 temporary_files_size = 0;
241 :
242 : /* Temporary file access initialized and not yet shut down? */
243 : #ifdef USE_ASSERT_CHECKING
244 : static bool temporary_files_allowed = false;
245 : #endif
246 :
247 : /*
248 : * List of OS handles opened with AllocateFile, AllocateDir and
249 : * OpenTransientFile.
250 : */
251 : typedef enum
252 : {
253 : AllocateDescFile,
254 : AllocateDescPipe,
255 : AllocateDescDir,
256 : AllocateDescRawFD,
257 : } AllocateDescKind;
258 :
259 : typedef struct
260 : {
261 : AllocateDescKind kind;
262 : SubTransactionId create_subid;
263 : union
264 : {
265 : FILE *file;
266 : DIR *dir;
267 : int fd;
268 : } desc;
269 : } AllocateDesc;
270 :
271 : static int numAllocatedDescs = 0;
272 : static int maxAllocatedDescs = 0;
273 : static AllocateDesc *allocatedDescs = NULL;
274 :
275 : /*
276 : * Number of open "external" FDs reported to Reserve/ReleaseExternalFD.
277 : */
278 : static int numExternalFDs = 0;
279 :
280 : /*
281 : * Number of temporary files opened during the current session;
282 : * this is used in generation of tempfile names.
283 : */
284 : static long tempFileCounter = 0;
285 :
286 : /*
287 : * Array of OIDs of temp tablespaces. (Some entries may be InvalidOid,
288 : * indicating that the current database's default tablespace should be used.)
289 : * When numTempTableSpaces is -1, this has not been set in the current
290 : * transaction.
291 : */
292 : static Oid *tempTableSpaces = NULL;
293 : static int numTempTableSpaces = -1;
294 : static int nextTempTableSpace = 0;
295 :
296 :
297 : /*--------------------
298 : *
299 : * Private Routines
300 : *
301 : * Delete - delete a file from the Lru ring
302 : * LruDelete - remove a file from the Lru ring and close its FD
303 : * Insert - put a file at the front of the Lru ring
304 : * LruInsert - put a file at the front of the Lru ring and open it
305 : * ReleaseLruFile - Release an fd by closing the last entry in the Lru ring
306 : * ReleaseLruFiles - Release fd(s) until we're under the max_safe_fds limit
307 : * AllocateVfd - grab a free (or new) file record (from VfdCache)
308 : * FreeVfd - free a file record
309 : *
310 : * The Least Recently Used ring is a doubly linked list that begins and
311 : * ends on element zero. Element zero is special -- it doesn't represent
312 : * a file and its "fd" field always == VFD_CLOSED. Element zero is just an
313 : * anchor that shows us the beginning/end of the ring.
314 : * Only VFD elements that are currently really open (have an FD assigned) are
315 : * in the Lru ring. Elements that are "virtually" open can be recognized
316 : * by having a non-null fileName field.
317 : *
318 : * example:
319 : *
320 : * /--less----\ /---------\
321 : * v \ v \
322 : * #0 --more---> LeastRecentlyUsed --more-\ \
323 : * ^\ | |
324 : * \\less--> MostRecentlyUsedFile <---/ |
325 : * \more---/ \--less--/
326 : *
327 : *--------------------
328 : */
329 : static void Delete(File file);
330 : static void LruDelete(File file);
331 : static void Insert(File file);
332 : static int LruInsert(File file);
333 : static bool ReleaseLruFile(void);
334 : static void ReleaseLruFiles(void);
335 : static File AllocateVfd(void);
336 : static void FreeVfd(File file);
337 :
338 : static int FileAccess(File file);
339 : static File OpenTemporaryFileInTablespace(Oid tblspcOid, bool rejectError);
340 : static bool reserveAllocatedDesc(void);
341 : static int FreeDesc(AllocateDesc *desc);
342 :
343 : static void BeforeShmemExit_Files(int code, Datum arg);
344 : static void CleanupTempFiles(bool isCommit, bool isProcExit);
345 : static void RemovePgTempRelationFiles(const char *tsdirname);
346 : static void RemovePgTempRelationFilesInDbspace(const char *dbspacedirname);
347 :
348 : static void walkdir(const char *path,
349 : void (*action) (const char *fname, bool isdir, int elevel),
350 : bool process_symlinks,
351 : int elevel);
352 : #ifdef PG_FLUSH_DATA_WORKS
353 : static void pre_sync_fname(const char *fname, bool isdir, int elevel);
354 : #endif
355 : static void datadir_fsync_fname(const char *fname, bool isdir, int elevel);
356 : static void unlink_if_exists_fname(const char *fname, bool isdir, int elevel);
357 :
358 : static int fsync_parent_path(const char *fname, int elevel);
359 :
360 :
361 : /* ResourceOwner callbacks to hold virtual file descriptors */
362 : static void ResOwnerReleaseFile(Datum res);
363 : static char *ResOwnerPrintFile(Datum res);
364 :
365 : static const ResourceOwnerDesc file_resowner_desc =
366 : {
367 : .name = "File",
368 : .release_phase = RESOURCE_RELEASE_AFTER_LOCKS,
369 : .release_priority = RELEASE_PRIO_FILES,
370 : .ReleaseResource = ResOwnerReleaseFile,
371 : .DebugPrint = ResOwnerPrintFile
372 : };
373 :
374 : /* Convenience wrappers over ResourceOwnerRemember/Forget */
375 : static inline void
376 6135 : ResourceOwnerRememberFile(ResourceOwner owner, File file)
377 : {
378 6135 : ResourceOwnerRemember(owner, Int32GetDatum(file), &file_resowner_desc);
379 6135 : }
380 : static inline void
381 6130 : ResourceOwnerForgetFile(ResourceOwner owner, File file)
382 : {
383 6130 : ResourceOwnerForget(owner, Int32GetDatum(file), &file_resowner_desc);
384 6130 : }
385 :
386 : /*
387 : * pg_fsync --- do fsync with or without writethrough
388 : */
389 : int
390 70037 : pg_fsync(int fd)
391 : {
392 : #if !defined(WIN32) && defined(USE_ASSERT_CHECKING)
393 : struct stat st;
394 :
395 : /*
396 : * Some operating system implementations of fsync() have requirements
397 : * about the file access modes that were used when their file descriptor
398 : * argument was opened, and these requirements differ depending on whether
399 : * the file descriptor is for a directory.
400 : *
401 : * For any file descriptor that may eventually be handed to fsync(), we
402 : * should have opened it with access modes that are compatible with
403 : * fsync() on all supported systems, otherwise the code may not be
404 : * portable, even if it runs ok on the current system.
405 : *
406 : * We assert here that a descriptor for a file was opened with write
407 : * permissions (i.e., not O_RDONLY) and for a directory without write
408 : * permissions (O_RDONLY). Notice that the assertion check is made even
409 : * if fsync() is disabled.
410 : *
411 : * If fstat() fails, ignore it and let the follow-up fsync() complain.
412 : */
413 : if (fstat(fd, &st) == 0)
414 : {
415 : int desc_flags = fcntl(fd, F_GETFL);
416 :
417 : desc_flags &= O_ACCMODE;
418 :
419 : if (S_ISDIR(st.st_mode))
420 : Assert(desc_flags == O_RDONLY);
421 : else
422 : Assert(desc_flags != O_RDONLY);
423 : }
424 : errno = 0;
425 : #endif
426 :
427 : /* #if is to skip the wal_sync_method test if there's no need for it */
428 : #if defined(HAVE_FSYNC_WRITETHROUGH)
429 : if (wal_sync_method == WAL_SYNC_METHOD_FSYNC_WRITETHROUGH)
430 : return pg_fsync_writethrough(fd);
431 : else
432 : #endif
433 70037 : return pg_fsync_no_writethrough(fd);
434 : }
435 :
436 :
437 : /*
438 : * pg_fsync_no_writethrough --- same as fsync except does nothing if
439 : * enableFsync is off
440 : */
441 : int
442 70037 : pg_fsync_no_writethrough(int fd)
443 : {
444 : int rc;
445 :
446 70037 : if (!enableFsync)
447 70037 : return 0;
448 :
449 0 : retry:
450 0 : rc = fsync(fd);
451 :
452 0 : if (rc == -1 && errno == EINTR)
453 0 : goto retry;
454 :
455 0 : return rc;
456 : }
457 :
458 : /*
459 : * pg_fsync_writethrough
460 : */
461 : int
462 0 : pg_fsync_writethrough(int fd)
463 : {
464 0 : if (enableFsync)
465 : {
466 : #if defined(F_FULLFSYNC)
467 : return (fcntl(fd, F_FULLFSYNC, 0) == -1) ? -1 : 0;
468 : #else
469 0 : errno = ENOSYS;
470 0 : return -1;
471 : #endif
472 : }
473 : else
474 0 : return 0;
475 : }
476 :
477 : /*
478 : * pg_fdatasync --- same as fdatasync except does nothing if enableFsync is off
479 : */
480 : int
481 0 : pg_fdatasync(int fd)
482 : {
483 : int rc;
484 :
485 0 : if (!enableFsync)
486 0 : return 0;
487 :
488 0 : retry:
489 0 : rc = fdatasync(fd);
490 :
491 0 : if (rc == -1 && errno == EINTR)
492 0 : goto retry;
493 :
494 0 : return rc;
495 : }
496 :
497 : /*
498 : * pg_file_exists -- check that a file exists.
499 : *
500 : * This requires an absolute path to the file. Returns true if the file is
501 : * not a directory, false otherwise.
502 : */
503 : bool
504 22401 : pg_file_exists(const char *name)
505 : {
506 : struct stat st;
507 :
508 : Assert(name != NULL);
509 :
510 22401 : if (stat(name, &st) == 0)
511 11785 : return !S_ISDIR(st.st_mode);
512 10616 : else if (!(errno == ENOENT || errno == ENOTDIR || errno == EACCES))
513 0 : ereport(ERROR,
514 : (errcode_for_file_access(),
515 : errmsg("could not access file \"%s\": %m", name)));
516 :
517 10616 : return false;
518 : }
519 :
520 : /*
521 : * pg_flush_data --- advise OS that the described dirty data should be flushed
522 : *
523 : * offset of 0 with nbytes 0 means that the entire file should be flushed
524 : */
525 : void
526 41598 : pg_flush_data(int fd, pgoff_t offset, pgoff_t nbytes)
527 : {
528 : /*
529 : * Right now file flushing is primarily used to avoid making later
530 : * fsync()/fdatasync() calls have less impact. Thus don't trigger flushes
531 : * if fsyncs are disabled - that's a decision we might want to make
532 : * configurable at some point.
533 : */
534 41598 : if (!enableFsync)
535 41598 : return;
536 :
537 : /*
538 : * We compile all alternatives that are supported on the current platform,
539 : * to find portability problems more easily.
540 : */
541 : #if defined(HAVE_SYNC_FILE_RANGE)
542 : {
543 : int rc;
544 : static bool not_implemented_by_kernel = false;
545 :
546 0 : if (not_implemented_by_kernel)
547 0 : return;
548 :
549 0 : retry:
550 :
551 : /*
552 : * sync_file_range(SYNC_FILE_RANGE_WRITE), currently linux specific,
553 : * tells the OS that writeback for the specified blocks should be
554 : * started, but that we don't want to wait for completion. Note that
555 : * this call might block if too much dirty data exists in the range.
556 : * This is the preferable method on OSs supporting it, as it works
557 : * reliably when available (contrast to msync()) and doesn't flush out
558 : * clean data (like FADV_DONTNEED).
559 : */
560 0 : rc = sync_file_range(fd, offset, nbytes,
561 : SYNC_FILE_RANGE_WRITE);
562 0 : if (rc != 0)
563 : {
564 : int elevel;
565 :
566 0 : if (rc == EINTR)
567 0 : goto retry;
568 :
569 : /*
570 : * For systems that don't have an implementation of
571 : * sync_file_range() such as Windows WSL, generate only one
572 : * warning and then suppress all further attempts by this process.
573 : */
574 0 : if (errno == ENOSYS)
575 : {
576 0 : elevel = WARNING;
577 0 : not_implemented_by_kernel = true;
578 : }
579 : else
580 0 : elevel = data_sync_elevel(WARNING);
581 :
582 0 : ereport(elevel,
583 : (errcode_for_file_access(),
584 : errmsg("could not flush dirty data: %m")));
585 : }
586 :
587 0 : return;
588 : }
589 : #endif
590 : #if !defined(WIN32) && defined(MS_ASYNC)
591 : {
592 : void *p;
593 : static int pagesize = 0;
594 :
595 : /*
596 : * On several OSs msync(MS_ASYNC) on a mmap'ed file triggers
597 : * writeback. On linux it only does so if MS_SYNC is specified, but
598 : * then it does the writeback synchronously. Luckily all common linux
599 : * systems have sync_file_range(). This is preferable over
600 : * FADV_DONTNEED because it doesn't flush out clean data.
601 : *
602 : * We map the file (mmap()), tell the kernel to sync back the contents
603 : * (msync()), and then remove the mapping again (munmap()).
604 : */
605 :
606 : /* mmap() needs actual length if we want to map whole file */
607 : if (offset == 0 && nbytes == 0)
608 : {
609 : nbytes = lseek(fd, 0, SEEK_END);
610 : if (nbytes < 0)
611 : {
612 : ereport(WARNING,
613 : (errcode_for_file_access(),
614 : errmsg("could not determine dirty data size: %m")));
615 : return;
616 : }
617 : }
618 :
619 : /*
620 : * Some platforms reject partial-page mmap() attempts. To deal with
621 : * that, just truncate the request to a page boundary. If any extra
622 : * bytes don't get flushed, well, it's only a hint anyway.
623 : */
624 :
625 : /* fetch pagesize only once */
626 : if (pagesize == 0)
627 : pagesize = sysconf(_SC_PAGESIZE);
628 :
629 : /* align length to pagesize, dropping any fractional page */
630 : if (pagesize > 0)
631 : nbytes = (nbytes / pagesize) * pagesize;
632 :
633 : /* fractional-page request is a no-op */
634 : if (nbytes <= 0)
635 : return;
636 :
637 : /*
638 : * mmap could well fail, particularly on 32-bit platforms where there
639 : * may simply not be enough address space. If so, silently fall
640 : * through to the next implementation.
641 : */
642 : if (nbytes <= (pgoff_t) SSIZE_MAX)
643 : p = mmap(NULL, nbytes, PROT_READ, MAP_SHARED, fd, offset);
644 : else
645 : p = MAP_FAILED;
646 :
647 : if (p != MAP_FAILED)
648 : {
649 : int rc;
650 :
651 : rc = msync(p, (size_t) nbytes, MS_ASYNC);
652 : if (rc != 0)
653 : {
654 : ereport(data_sync_elevel(WARNING),
655 : (errcode_for_file_access(),
656 : errmsg("could not flush dirty data: %m")));
657 : /* NB: need to fall through to munmap()! */
658 : }
659 :
660 : rc = munmap(p, (size_t) nbytes);
661 : if (rc != 0)
662 : {
663 : /* FATAL error because mapping would remain */
664 : ereport(FATAL,
665 : (errcode_for_file_access(),
666 : errmsg("could not munmap() while flushing data: %m")));
667 : }
668 :
669 : return;
670 : }
671 : }
672 : #endif
673 : #if defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_DONTNEED)
674 : {
675 : int rc;
676 :
677 : /*
678 : * Signal the kernel that the passed in range should not be cached
679 : * anymore. This has the, desired, side effect of writing out dirty
680 : * data, and the, undesired, side effect of likely discarding useful
681 : * clean cached blocks. For the latter reason this is the least
682 : * preferable method.
683 : */
684 :
685 : rc = posix_fadvise(fd, offset, nbytes, POSIX_FADV_DONTNEED);
686 :
687 : if (rc != 0)
688 : {
689 : /* don't error out, this is just a performance optimization */
690 : ereport(WARNING,
691 : (errcode_for_file_access(),
692 : errmsg("could not flush dirty data: %m")));
693 : }
694 :
695 : return;
696 : }
697 : #endif
698 : }
699 :
700 : /*
701 : * Truncate an open file to a given length.
702 : */
703 : static int
704 672 : pg_ftruncate(int fd, pgoff_t length)
705 : {
706 : int ret;
707 :
708 672 : retry:
709 672 : ret = ftruncate(fd, length);
710 :
711 672 : if (ret == -1 && errno == EINTR)
712 0 : goto retry;
713 :
714 672 : return ret;
715 : }
716 :
717 : /*
718 : * Truncate a file to a given length by name.
719 : */
720 : int
721 271654 : pg_truncate(const char *path, pgoff_t length)
722 : {
723 : int ret;
724 : #ifdef WIN32
725 : int save_errno;
726 : int fd;
727 :
728 : fd = OpenTransientFile(path, O_RDWR | PG_BINARY);
729 : if (fd >= 0)
730 : {
731 : ret = pg_ftruncate(fd, length);
732 : save_errno = errno;
733 : CloseTransientFile(fd);
734 : errno = save_errno;
735 : }
736 : else
737 : ret = -1;
738 : #else
739 :
740 271654 : retry:
741 271654 : ret = truncate(path, length);
742 :
743 271654 : if (ret == -1 && errno == EINTR)
744 0 : goto retry;
745 : #endif
746 :
747 271654 : return ret;
748 : }
749 :
750 : /*
751 : * fsync_fname -- fsync a file or directory, handling errors properly
752 : *
753 : * Try to fsync a file or directory. When doing the latter, ignore errors that
754 : * indicate the OS just doesn't allow/require fsyncing directories.
755 : */
756 : void
757 22892 : fsync_fname(const char *fname, bool isdir)
758 : {
759 22892 : fsync_fname_ext(fname, isdir, false, data_sync_elevel(ERROR));
760 22892 : }
761 :
762 : /*
763 : * durable_rename -- rename(2) wrapper, issuing fsyncs required for durability
764 : *
765 : * This routine ensures that, after returning, the effect of renaming file
766 : * persists in case of a crash. A crash while this routine is running will
767 : * leave you with either the pre-existing or the moved file in place of the
768 : * new file; no mixed state or truncated files are possible.
769 : *
770 : * It does so by using fsync on the old filename and the possibly existing
771 : * target filename before the rename, and the target file and directory after.
772 : *
773 : * Note that rename() cannot be used across arbitrary directories, as they
774 : * might not be on the same filesystem. Therefore this routine does not
775 : * support renaming across directories.
776 : *
777 : * Log errors with the caller specified severity.
778 : *
779 : * Returns 0 if the operation succeeded, -1 otherwise. Note that errno is not
780 : * valid upon return.
781 : */
782 : int
783 6734 : durable_rename(const char *oldfile, const char *newfile, int elevel)
784 : {
785 : int fd;
786 :
787 : /*
788 : * First fsync the old and target path (if it exists), to ensure that they
789 : * are properly persistent on disk. Syncing the target file is not
790 : * strictly necessary, but it makes it easier to reason about crashes;
791 : * because it's then guaranteed that either source or target file exists
792 : * after a crash.
793 : */
794 6734 : if (fsync_fname_ext(oldfile, false, false, elevel) != 0)
795 0 : return -1;
796 :
797 6734 : fd = OpenTransientFile(newfile, PG_BINARY | O_RDWR);
798 6734 : if (fd < 0)
799 : {
800 4661 : if (errno != ENOENT)
801 : {
802 0 : ereport(elevel,
803 : (errcode_for_file_access(),
804 : errmsg("could not open file \"%s\": %m", newfile)));
805 0 : return -1;
806 : }
807 : }
808 : else
809 : {
810 2073 : if (pg_fsync(fd) != 0)
811 : {
812 : int save_errno;
813 :
814 : /* close file upon error, might not be in transaction context */
815 0 : save_errno = errno;
816 0 : CloseTransientFile(fd);
817 0 : errno = save_errno;
818 :
819 0 : ereport(elevel,
820 : (errcode_for_file_access(),
821 : errmsg("could not fsync file \"%s\": %m", newfile)));
822 0 : return -1;
823 : }
824 :
825 2073 : if (CloseTransientFile(fd) != 0)
826 : {
827 0 : ereport(elevel,
828 : (errcode_for_file_access(),
829 : errmsg("could not close file \"%s\": %m", newfile)));
830 0 : return -1;
831 : }
832 : }
833 :
834 : /* Time to do the real deal... */
835 6734 : if (rename(oldfile, newfile) < 0)
836 : {
837 0 : ereport(elevel,
838 : (errcode_for_file_access(),
839 : errmsg("could not rename file \"%s\" to \"%s\": %m",
840 : oldfile, newfile)));
841 0 : return -1;
842 : }
843 :
844 : /*
845 : * To guarantee renaming the file is persistent, fsync the file with its
846 : * new name, and its containing directory.
847 : */
848 6734 : if (fsync_fname_ext(newfile, false, false, elevel) != 0)
849 0 : return -1;
850 :
851 6734 : if (fsync_parent_path(newfile, elevel) != 0)
852 0 : return -1;
853 :
854 6734 : return 0;
855 : }
856 :
857 : /*
858 : * durable_unlink -- remove a file in a durable manner
859 : *
860 : * This routine ensures that, after returning, the effect of removing file
861 : * persists in case of a crash. A crash while this routine is running will
862 : * leave the system in no mixed state.
863 : *
864 : * It does so by using fsync on the parent directory of the file after the
865 : * actual removal is done.
866 : *
867 : * Log errors with the severity specified by caller.
868 : *
869 : * Returns 0 if the operation succeeded, -1 otherwise. Note that errno is not
870 : * valid upon return.
871 : */
872 : int
873 1345 : durable_unlink(const char *fname, int elevel)
874 : {
875 1345 : if (unlink(fname) < 0)
876 : {
877 43 : ereport(elevel,
878 : (errcode_for_file_access(),
879 : errmsg("could not remove file \"%s\": %m",
880 : fname)));
881 43 : return -1;
882 : }
883 :
884 : /*
885 : * To guarantee that the removal of the file is persistent, fsync its
886 : * parent directory.
887 : */
888 1302 : if (fsync_parent_path(fname, elevel) != 0)
889 0 : return -1;
890 :
891 1302 : return 0;
892 : }
893 :
894 : /*
895 : * InitFileAccess --- initialize this module during backend startup
896 : *
897 : * This is called during either normal or standalone backend start.
898 : * It is *not* called in the postmaster.
899 : *
900 : * Note that this does not initialize temporary file access, that is
901 : * separately initialized via InitTemporaryFileAccess().
902 : */
903 : void
904 24251 : InitFileAccess(void)
905 : {
906 : Assert(SizeVfdCache == 0); /* call me only once */
907 :
908 : /* initialize cache header entry */
909 24251 : VfdCache = (Vfd *) malloc(sizeof(Vfd));
910 24251 : if (VfdCache == NULL)
911 0 : ereport(FATAL,
912 : (errcode(ERRCODE_OUT_OF_MEMORY),
913 : errmsg("out of memory")));
914 :
915 194008 : MemSet(&(VfdCache[0]), 0, sizeof(Vfd));
916 24251 : VfdCache->fd = VFD_CLOSED;
917 :
918 24251 : SizeVfdCache = 1;
919 24251 : }
920 :
921 : /*
922 : * InitTemporaryFileAccess --- initialize temporary file access during startup
923 : *
924 : * This is called during either normal or standalone backend start.
925 : * It is *not* called in the postmaster.
926 : *
927 : * This is separate from InitFileAccess() because temporary file cleanup can
928 : * cause pgstat reporting. As pgstat is shut down during before_shmem_exit(),
929 : * our reporting has to happen before that. Low level file access should be
930 : * available for longer, hence the separate initialization / shutdown of
931 : * temporary file handling.
932 : */
933 : void
934 24251 : InitTemporaryFileAccess(void)
935 : {
936 : Assert(SizeVfdCache != 0); /* InitFileAccess() needs to have run */
937 : Assert(!temporary_files_allowed); /* call me only once */
938 :
939 : /*
940 : * Register before-shmem-exit hook to ensure temp files are dropped while
941 : * we can still report stats.
942 : */
943 24251 : before_shmem_exit(BeforeShmemExit_Files, 0);
944 :
945 : #ifdef USE_ASSERT_CHECKING
946 : temporary_files_allowed = true;
947 : #endif
948 24251 : }
949 :
950 : /*
951 : * count_usable_fds --- count how many FDs the system will let us open,
952 : * and estimate how many are already open.
953 : *
954 : * We stop counting if usable_fds reaches max_to_probe. Note: a small
955 : * value of max_to_probe might result in an underestimate of already_open;
956 : * we must fill in any "gaps" in the set of used FDs before the calculation
957 : * of already_open will give the right answer. In practice, max_to_probe
958 : * of a couple of dozen should be enough to ensure good results.
959 : *
960 : * We assume stderr (FD 2) is available for dup'ing. While the calling
961 : * script could theoretically close that, it would be a really bad idea,
962 : * since then one risks loss of error messages from, e.g., libc.
963 : */
964 : static void
965 1175 : count_usable_fds(int max_to_probe, int *usable_fds, int *already_open)
966 : {
967 : int *fd;
968 : int size;
969 1175 : int used = 0;
970 1175 : int highestfd = 0;
971 : int j;
972 :
973 : #ifdef HAVE_GETRLIMIT
974 : struct rlimit rlim;
975 : int getrlimit_status;
976 : #endif
977 :
978 1175 : size = 1024;
979 1175 : fd = (int *) palloc(size * sizeof(int));
980 :
981 : #ifdef HAVE_GETRLIMIT
982 1175 : getrlimit_status = getrlimit(RLIMIT_NOFILE, &rlim);
983 1175 : if (getrlimit_status != 0)
984 0 : ereport(WARNING, (errmsg("getrlimit failed: %m")));
985 : #endif /* HAVE_GETRLIMIT */
986 :
987 : /* dup until failure or probe limit reached */
988 : for (;;)
989 1173825 : {
990 : int thisfd;
991 :
992 : #ifdef HAVE_GETRLIMIT
993 :
994 : /*
995 : * don't go beyond RLIMIT_NOFILE; causes irritating kernel logs on
996 : * some platforms
997 : */
998 1175000 : if (getrlimit_status == 0 && highestfd >= rlim.rlim_cur - 1)
999 0 : break;
1000 : #endif
1001 :
1002 1175000 : thisfd = dup(2);
1003 1175000 : if (thisfd < 0)
1004 : {
1005 : /* Expect EMFILE or ENFILE, else it's fishy */
1006 0 : if (errno != EMFILE && errno != ENFILE)
1007 0 : elog(WARNING, "duplicating stderr file descriptor failed after %d successes: %m", used);
1008 0 : break;
1009 : }
1010 :
1011 1175000 : if (used >= size)
1012 : {
1013 0 : size *= 2;
1014 0 : fd = (int *) repalloc(fd, size * sizeof(int));
1015 : }
1016 1175000 : fd[used++] = thisfd;
1017 :
1018 1175000 : if (highestfd < thisfd)
1019 1175000 : highestfd = thisfd;
1020 :
1021 1175000 : if (used >= max_to_probe)
1022 1175 : break;
1023 : }
1024 :
1025 : /* release the files we opened */
1026 1176175 : for (j = 0; j < used; j++)
1027 1175000 : close(fd[j]);
1028 :
1029 1175 : pfree(fd);
1030 :
1031 : /*
1032 : * Return results. usable_fds is just the number of successful dups. We
1033 : * assume that the system limit is highestfd+1 (remember 0 is a legal FD
1034 : * number) and so already_open is highestfd+1 - usable_fds.
1035 : */
1036 1175 : *usable_fds = used;
1037 1175 : *already_open = highestfd + 1 - used;
1038 1175 : }
1039 :
1040 : /*
1041 : * set_max_safe_fds
1042 : * Determine number of file descriptors that fd.c is allowed to use
1043 : */
1044 : void
1045 1175 : set_max_safe_fds(void)
1046 : {
1047 : int usable_fds;
1048 : int already_open;
1049 :
1050 : /*----------
1051 : * We want to set max_safe_fds to
1052 : * MIN(usable_fds, max_files_per_process)
1053 : * less the slop factor for files that are opened without consulting
1054 : * fd.c. This ensures that we won't allow to open more than
1055 : * max_files_per_process, or the experimentally-determined EMFILE limit,
1056 : * additional files.
1057 : *----------
1058 : */
1059 1175 : count_usable_fds(max_files_per_process,
1060 : &usable_fds, &already_open);
1061 :
1062 1175 : max_safe_fds = Min(usable_fds, max_files_per_process);
1063 :
1064 : /*
1065 : * Take off the FDs reserved for system() etc.
1066 : */
1067 1175 : max_safe_fds -= NUM_RESERVED_FDS;
1068 :
1069 : /*
1070 : * Make sure we still have enough to get by.
1071 : */
1072 1175 : if (max_safe_fds < FD_MINFREE)
1073 0 : ereport(FATAL,
1074 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
1075 : errmsg("insufficient file descriptors available to start server process"),
1076 : errdetail("System allows %d, server needs at least %d, %d files are already open.",
1077 : max_safe_fds + NUM_RESERVED_FDS,
1078 : FD_MINFREE + NUM_RESERVED_FDS,
1079 : already_open)));
1080 :
1081 1175 : elog(DEBUG2, "max_safe_fds = %d, usable_fds = %d, already_open = %d",
1082 : max_safe_fds, usable_fds, already_open);
1083 1175 : }
1084 :
1085 : /*
1086 : * Open a file with BasicOpenFilePerm() and pass default file mode for the
1087 : * fileMode parameter.
1088 : */
1089 : int
1090 44326 : BasicOpenFile(const char *fileName, int fileFlags)
1091 : {
1092 44326 : return BasicOpenFilePerm(fileName, fileFlags, pg_file_create_mode);
1093 : }
1094 :
1095 : /*
1096 : * BasicOpenFilePerm --- same as open(2) except can free other FDs if needed
1097 : *
1098 : * This is exported for use by places that really want a plain kernel FD,
1099 : * but need to be proof against running out of FDs. Once an FD has been
1100 : * successfully returned, it is the caller's responsibility to ensure that
1101 : * it will not be leaked on ereport()! Most users should *not* call this
1102 : * routine directly, but instead use the VFD abstraction level, which
1103 : * provides protection against descriptor leaks as well as management of
1104 : * files that need to be open for more than a short period of time.
1105 : *
1106 : * Ideally this should be the *only* direct call of open() in the backend.
1107 : * In practice, the postmaster calls open() directly, and there are some
1108 : * direct open() calls done early in backend startup. Those are OK since
1109 : * this module wouldn't have any open files to close at that point anyway.
1110 : */
1111 : int
1112 9612037 : BasicOpenFilePerm(const char *fileName, int fileFlags, mode_t fileMode)
1113 : {
1114 : int fd;
1115 :
1116 9612037 : tryAgain:
1117 : #ifdef PG_O_DIRECT_USE_F_NOCACHE
1118 : fd = open(fileName, fileFlags & ~PG_O_DIRECT, fileMode);
1119 : #else
1120 9612037 : fd = open(fileName, fileFlags, fileMode);
1121 : #endif
1122 :
1123 9612037 : if (fd >= 0)
1124 : {
1125 : #ifdef PG_O_DIRECT_USE_F_NOCACHE
1126 : if (fileFlags & PG_O_DIRECT)
1127 : {
1128 : if (fcntl(fd, F_NOCACHE, 1) < 0)
1129 : {
1130 : int save_errno = errno;
1131 :
1132 : close(fd);
1133 : errno = save_errno;
1134 : return -1;
1135 : }
1136 : }
1137 : #endif
1138 :
1139 9073681 : return fd; /* success! */
1140 : }
1141 :
1142 538356 : if (errno == EMFILE || errno == ENFILE)
1143 : {
1144 0 : int save_errno = errno;
1145 :
1146 0 : ereport(LOG,
1147 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
1148 : errmsg("out of file descriptors: %m; release and retry")));
1149 0 : errno = 0;
1150 0 : if (ReleaseLruFile())
1151 0 : goto tryAgain;
1152 0 : errno = save_errno;
1153 : }
1154 :
1155 538356 : return -1; /* failure */
1156 : }
1157 :
1158 : /*
1159 : * AcquireExternalFD - attempt to reserve an external file descriptor
1160 : *
1161 : * This should be used by callers that need to hold a file descriptor open
1162 : * over more than a short interval, but cannot use any of the other facilities
1163 : * provided by this module.
1164 : *
1165 : * The difference between this and the underlying ReserveExternalFD function
1166 : * is that this will report failure (by setting errno and returning false)
1167 : * if "too many" external FDs are already reserved. This should be used in
1168 : * any code where the total number of FDs to be reserved is not predictable
1169 : * and small.
1170 : */
1171 : bool
1172 144638 : AcquireExternalFD(void)
1173 : {
1174 : /*
1175 : * We don't want more than max_safe_fds / 3 FDs to be consumed for
1176 : * "external" FDs.
1177 : */
1178 144638 : if (numExternalFDs < max_safe_fds / 3)
1179 : {
1180 144638 : ReserveExternalFD();
1181 144638 : return true;
1182 : }
1183 0 : errno = EMFILE;
1184 0 : return false;
1185 : }
1186 :
1187 : /*
1188 : * ReserveExternalFD - report external consumption of a file descriptor
1189 : *
1190 : * This should be used by callers that need to hold a file descriptor open
1191 : * over more than a short interval, but cannot use any of the other facilities
1192 : * provided by this module. This just tracks the use of the FD and closes
1193 : * VFDs if needed to ensure we keep NUM_RESERVED_FDS FDs available.
1194 : *
1195 : * Call this directly only in code where failure to reserve the FD would be
1196 : * fatal; for example, the WAL-writing code does so, since the alternative is
1197 : * session failure. Also, it's very unwise to do so in code that could
1198 : * consume more than one FD per process.
1199 : *
1200 : * Note: as long as everybody plays nice so that NUM_RESERVED_FDS FDs remain
1201 : * available, it doesn't matter too much whether this is called before or
1202 : * after actually opening the FD; but doing so beforehand reduces the risk of
1203 : * an EMFILE failure if not everybody played nice. In any case, it's solely
1204 : * caller's responsibility to keep the external-FD count in sync with reality.
1205 : */
1206 : void
1207 230685 : ReserveExternalFD(void)
1208 : {
1209 : /*
1210 : * Release VFDs if needed to stay safe. Because we do this before
1211 : * incrementing numExternalFDs, the final state will be as desired, i.e.,
1212 : * nfile + numAllocatedDescs + numExternalFDs <= max_safe_fds.
1213 : */
1214 230685 : ReleaseLruFiles();
1215 :
1216 230685 : numExternalFDs++;
1217 230685 : }
1218 :
1219 : /*
1220 : * ReleaseExternalFD - report release of an external file descriptor
1221 : *
1222 : * This is guaranteed not to change errno, so it can be used in failure paths.
1223 : */
1224 : void
1225 209785 : ReleaseExternalFD(void)
1226 : {
1227 : Assert(numExternalFDs > 0);
1228 209785 : numExternalFDs--;
1229 209785 : }
1230 :
1231 :
1232 : #if defined(FDDEBUG)
1233 :
1234 : static void
1235 : _dump_lru(void)
1236 : {
1237 : int mru = VfdCache[0].lruLessRecently;
1238 : Vfd *vfdP = &VfdCache[mru];
1239 : char buf[2048];
1240 :
1241 : snprintf(buf, sizeof(buf), "LRU: MOST %d ", mru);
1242 : while (mru != 0)
1243 : {
1244 : mru = vfdP->lruLessRecently;
1245 : vfdP = &VfdCache[mru];
1246 : snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), "%d ", mru);
1247 : }
1248 : snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), "LEAST");
1249 : elog(LOG, "%s", buf);
1250 : }
1251 : #endif /* FDDEBUG */
1252 :
1253 : static void
1254 1556614 : Delete(File file)
1255 : {
1256 : Vfd *vfdP;
1257 :
1258 : Assert(file != 0);
1259 :
1260 : DO_DB(elog(LOG, "Delete %d (%s)",
1261 : file, VfdCache[file].fileName));
1262 : DO_DB(_dump_lru());
1263 :
1264 1556614 : vfdP = &VfdCache[file];
1265 :
1266 1556614 : VfdCache[vfdP->lruLessRecently].lruMoreRecently = vfdP->lruMoreRecently;
1267 1556614 : VfdCache[vfdP->lruMoreRecently].lruLessRecently = vfdP->lruLessRecently;
1268 :
1269 : DO_DB(_dump_lru());
1270 1556614 : }
1271 :
1272 : static void
1273 4650 : LruDelete(File file)
1274 : {
1275 : Vfd *vfdP;
1276 :
1277 : Assert(file != 0);
1278 :
1279 : DO_DB(elog(LOG, "LruDelete %d (%s)",
1280 : file, VfdCache[file].fileName));
1281 :
1282 4650 : vfdP = &VfdCache[file];
1283 :
1284 4650 : pgaio_closing_fd(vfdP->fd);
1285 :
1286 : /*
1287 : * Close the file. We aren't expecting this to fail; if it does, better
1288 : * to leak the FD than to mess up our internal state.
1289 : */
1290 4650 : if (close(vfdP->fd) != 0)
1291 0 : elog(vfdP->fdstate & FD_TEMP_FILE_LIMIT ? LOG : data_sync_elevel(LOG),
1292 : "could not close file \"%s\": %m", vfdP->fileName);
1293 4650 : vfdP->fd = VFD_CLOSED;
1294 4650 : --nfile;
1295 :
1296 : /* delete the vfd record from the LRU ring */
1297 4650 : Delete(file);
1298 4650 : }
1299 :
1300 : static void
1301 2143346 : Insert(File file)
1302 : {
1303 : Vfd *vfdP;
1304 :
1305 : Assert(file != 0);
1306 :
1307 : DO_DB(elog(LOG, "Insert %d (%s)",
1308 : file, VfdCache[file].fileName));
1309 : DO_DB(_dump_lru());
1310 :
1311 2143346 : vfdP = &VfdCache[file];
1312 :
1313 2143346 : vfdP->lruMoreRecently = 0;
1314 2143346 : vfdP->lruLessRecently = VfdCache[0].lruLessRecently;
1315 2143346 : VfdCache[0].lruLessRecently = file;
1316 2143346 : VfdCache[vfdP->lruLessRecently].lruMoreRecently = file;
1317 :
1318 : DO_DB(_dump_lru());
1319 2143346 : }
1320 :
1321 : /* returns 0 on success, -1 on re-open failure (with errno set) */
1322 : static int
1323 41 : LruInsert(File file)
1324 : {
1325 : Vfd *vfdP;
1326 :
1327 : Assert(file != 0);
1328 :
1329 : DO_DB(elog(LOG, "LruInsert %d (%s)",
1330 : file, VfdCache[file].fileName));
1331 :
1332 41 : vfdP = &VfdCache[file];
1333 :
1334 41 : if (FileIsNotOpen(file))
1335 : {
1336 : /* Close excess kernel FDs. */
1337 41 : ReleaseLruFiles();
1338 :
1339 : /*
1340 : * The open could still fail for lack of file descriptors, eg due to
1341 : * overall system file table being full. So, be prepared to release
1342 : * another FD if necessary...
1343 : */
1344 41 : vfdP->fd = BasicOpenFilePerm(vfdP->fileName, vfdP->fileFlags,
1345 : vfdP->fileMode);
1346 41 : if (vfdP->fd < 0)
1347 : {
1348 : DO_DB(elog(LOG, "re-open failed: %m"));
1349 0 : return -1;
1350 : }
1351 : else
1352 : {
1353 41 : ++nfile;
1354 : }
1355 : }
1356 :
1357 : /*
1358 : * put it at the head of the Lru ring
1359 : */
1360 :
1361 41 : Insert(file);
1362 :
1363 41 : return 0;
1364 : }
1365 :
1366 : /*
1367 : * Release one kernel FD by closing the least-recently-used VFD.
1368 : */
1369 : static bool
1370 4502 : ReleaseLruFile(void)
1371 : {
1372 : DO_DB(elog(LOG, "ReleaseLruFile. Opened %d", nfile));
1373 :
1374 4502 : if (nfile > 0)
1375 : {
1376 : /*
1377 : * There are opened files and so there should be at least one used vfd
1378 : * in the ring.
1379 : */
1380 : Assert(VfdCache[0].lruMoreRecently != 0);
1381 4502 : LruDelete(VfdCache[0].lruMoreRecently);
1382 4502 : return true; /* freed a file */
1383 : }
1384 0 : return false; /* no files available to free */
1385 : }
1386 :
1387 : /*
1388 : * Release kernel FDs as needed to get under the max_safe_fds limit.
1389 : * After calling this, it's OK to try to open another file.
1390 : */
1391 : static void
1392 9948736 : ReleaseLruFiles(void)
1393 : {
1394 9953238 : while (nfile + numAllocatedDescs + numExternalFDs >= max_safe_fds)
1395 : {
1396 4502 : if (!ReleaseLruFile())
1397 0 : break;
1398 : }
1399 9948736 : }
1400 :
1401 : static File
1402 1781520 : AllocateVfd(void)
1403 : {
1404 : Index i;
1405 : File file;
1406 :
1407 : DO_DB(elog(LOG, "AllocateVfd. Size %zu", SizeVfdCache));
1408 :
1409 : Assert(SizeVfdCache > 0); /* InitFileAccess not called? */
1410 :
1411 1781520 : if (VfdCache[0].nextFree == 0)
1412 : {
1413 : /*
1414 : * The free list is empty so it is time to increase the size of the
1415 : * array. We choose to double it each time this happens. However,
1416 : * there's not much point in starting *real* small.
1417 : */
1418 30559 : Size newCacheSize = SizeVfdCache * 2;
1419 : Vfd *newVfdCache;
1420 :
1421 30559 : if (newCacheSize < 32)
1422 21023 : newCacheSize = 32;
1423 :
1424 : /*
1425 : * Be careful not to clobber VfdCache ptr if realloc fails.
1426 : */
1427 30559 : newVfdCache = (Vfd *) realloc(VfdCache, sizeof(Vfd) * newCacheSize);
1428 30559 : if (newVfdCache == NULL)
1429 0 : ereport(ERROR,
1430 : (errcode(ERRCODE_OUT_OF_MEMORY),
1431 : errmsg("out of memory")));
1432 30559 : VfdCache = newVfdCache;
1433 :
1434 : /*
1435 : * Initialize the new entries and link them into the free list.
1436 : */
1437 1534048 : for (i = SizeVfdCache; i < newCacheSize; i++)
1438 : {
1439 12027912 : MemSet(&(VfdCache[i]), 0, sizeof(Vfd));
1440 1503489 : VfdCache[i].nextFree = i + 1;
1441 1503489 : VfdCache[i].fd = VFD_CLOSED;
1442 : }
1443 30559 : VfdCache[newCacheSize - 1].nextFree = 0;
1444 30559 : VfdCache[0].nextFree = SizeVfdCache;
1445 :
1446 : /*
1447 : * Record the new size
1448 : */
1449 30559 : SizeVfdCache = newCacheSize;
1450 : }
1451 :
1452 1781520 : file = VfdCache[0].nextFree;
1453 :
1454 1781520 : VfdCache[0].nextFree = VfdCache[file].nextFree;
1455 :
1456 1781520 : return file;
1457 : }
1458 :
1459 : static void
1460 1190605 : FreeVfd(File file)
1461 : {
1462 1190605 : Vfd *vfdP = &VfdCache[file];
1463 :
1464 : DO_DB(elog(LOG, "FreeVfd: %d (%s)",
1465 : file, vfdP->fileName ? vfdP->fileName : ""));
1466 :
1467 1190605 : if (vfdP->fileName != NULL)
1468 : {
1469 659120 : free(vfdP->fileName);
1470 659120 : vfdP->fileName = NULL;
1471 : }
1472 1190605 : vfdP->fdstate = 0x0;
1473 :
1474 1190605 : vfdP->nextFree = VfdCache[0].nextFree;
1475 1190605 : VfdCache[0].nextFree = file;
1476 1190605 : }
1477 :
1478 : /* returns 0 on success, -1 on re-open failure (with errno set) */
1479 : static int
1480 3392523 : FileAccess(File file)
1481 : {
1482 : int returnValue;
1483 :
1484 : DO_DB(elog(LOG, "FileAccess %d (%s)",
1485 : file, VfdCache[file].fileName));
1486 :
1487 : /*
1488 : * Is the file open? If not, open it and put it at the head of the LRU
1489 : * ring (possibly closing the least recently used file to get an FD).
1490 : */
1491 :
1492 3392523 : if (FileIsNotOpen(file))
1493 : {
1494 41 : returnValue = LruInsert(file);
1495 41 : if (returnValue != 0)
1496 0 : return returnValue;
1497 : }
1498 3392482 : else if (VfdCache[0].lruLessRecently != file)
1499 : {
1500 : /*
1501 : * We now know that the file is open and that it is not the last one
1502 : * accessed, so we need to move it to the head of the Lru ring.
1503 : */
1504 :
1505 893270 : Delete(file);
1506 893270 : Insert(file);
1507 : }
1508 :
1509 3392523 : return 0;
1510 : }
1511 :
1512 : /*
1513 : * Called whenever a temporary file is deleted to report its size.
1514 : */
1515 : static void
1516 3856 : ReportTemporaryFileUsage(const char *path, pgoff_t size)
1517 : {
1518 3856 : pgstat_report_tempfile(size);
1519 :
1520 3856 : if (log_temp_files >= 0)
1521 : {
1522 842 : if ((size / 1024) >= log_temp_files)
1523 132 : ereport(LOG,
1524 : (errmsg("temporary file: path \"%s\", size %lu",
1525 : path, (unsigned long) size)));
1526 : }
1527 3856 : }
1528 :
1529 : /*
1530 : * Called to register a temporary file for automatic close.
1531 : * ResourceOwnerEnlarge(CurrentResourceOwner) must have been called
1532 : * before the file was opened.
1533 : */
1534 : static void
1535 6135 : RegisterTemporaryFile(File file)
1536 : {
1537 6135 : ResourceOwnerRememberFile(CurrentResourceOwner, file);
1538 6135 : VfdCache[file].resowner = CurrentResourceOwner;
1539 :
1540 : /* Backup mechanism for closing at end of xact. */
1541 6135 : VfdCache[file].fdstate |= FD_CLOSE_AT_EOXACT;
1542 6135 : have_xact_temporary_files = true;
1543 6135 : }
1544 :
1545 : /*
1546 : * Called when we get a shared invalidation message on some relation.
1547 : */
1548 : #ifdef NOT_USED
1549 : void
1550 : FileInvalidate(File file)
1551 : {
1552 : Assert(FileIsValid(file));
1553 : if (!FileIsNotOpen(file))
1554 : LruDelete(file);
1555 : }
1556 : #endif
1557 :
1558 : /*
1559 : * Open a file with PathNameOpenFilePerm() and pass default file mode for the
1560 : * fileMode parameter.
1561 : */
1562 : File
1563 1781520 : PathNameOpenFile(const char *fileName, int fileFlags)
1564 : {
1565 1781520 : return PathNameOpenFilePerm(fileName, fileFlags, pg_file_create_mode);
1566 : }
1567 :
1568 : /*
1569 : * open a file in an arbitrary directory
1570 : *
1571 : * NB: if the passed pathname is relative (which it usually is),
1572 : * it will be interpreted relative to the process' working directory
1573 : * (which should always be $PGDATA when this code is running).
1574 : */
1575 : File
1576 1781520 : PathNameOpenFilePerm(const char *fileName, int fileFlags, mode_t fileMode)
1577 : {
1578 : char *fnamecopy;
1579 : File file;
1580 : Vfd *vfdP;
1581 :
1582 : DO_DB(elog(LOG, "PathNameOpenFilePerm: %s %x %o",
1583 : fileName, fileFlags, fileMode));
1584 :
1585 : /*
1586 : * We need a malloc'd copy of the file name; fail cleanly if no room.
1587 : */
1588 1781520 : fnamecopy = strdup(fileName);
1589 1781520 : if (fnamecopy == NULL)
1590 0 : ereport(ERROR,
1591 : (errcode(ERRCODE_OUT_OF_MEMORY),
1592 : errmsg("out of memory")));
1593 :
1594 1781520 : file = AllocateVfd();
1595 1781520 : vfdP = &VfdCache[file];
1596 :
1597 : /* Close excess kernel FDs. */
1598 1781520 : ReleaseLruFiles();
1599 :
1600 : /*
1601 : * Descriptors managed by VFDs are implicitly marked O_CLOEXEC. The
1602 : * client shouldn't be expected to know which kernel descriptors are
1603 : * currently open, so it wouldn't make sense for them to be inherited by
1604 : * executed subprograms.
1605 : */
1606 1781520 : fileFlags |= O_CLOEXEC;
1607 :
1608 1781520 : vfdP->fd = BasicOpenFilePerm(fileName, fileFlags, fileMode);
1609 :
1610 1781520 : if (vfdP->fd < 0)
1611 : {
1612 531485 : int save_errno = errno;
1613 :
1614 531485 : FreeVfd(file);
1615 531485 : free(fnamecopy);
1616 531485 : errno = save_errno;
1617 531485 : return -1;
1618 : }
1619 1250035 : ++nfile;
1620 : DO_DB(elog(LOG, "PathNameOpenFile: success %d",
1621 : vfdP->fd));
1622 :
1623 1250035 : vfdP->fileName = fnamecopy;
1624 : /* Saved flags are adjusted to be OK for re-opening file */
1625 1250035 : vfdP->fileFlags = fileFlags & ~(O_CREAT | O_TRUNC | O_EXCL);
1626 1250035 : vfdP->fileMode = fileMode;
1627 1250035 : vfdP->fileSize = 0;
1628 1250035 : vfdP->fdstate = 0x0;
1629 1250035 : vfdP->resowner = NULL;
1630 :
1631 1250035 : Insert(file);
1632 :
1633 1250035 : return file;
1634 : }
1635 :
1636 : /*
1637 : * Create directory 'directory'. If necessary, create 'basedir', which must
1638 : * be the directory above it. This is designed for creating the top-level
1639 : * temporary directory on demand before creating a directory underneath it.
1640 : * Do nothing if the directory already exists.
1641 : *
1642 : * Directories created within the top-level temporary directory should begin
1643 : * with PG_TEMP_FILE_PREFIX, so that they can be identified as temporary and
1644 : * deleted at startup by RemovePgTempFiles(). Further subdirectories below
1645 : * that do not need any particular prefix.
1646 : */
1647 : void
1648 262 : PathNameCreateTemporaryDir(const char *basedir, const char *directory)
1649 : {
1650 262 : if (MakePGDirectory(directory) < 0)
1651 : {
1652 22 : if (errno == EEXIST)
1653 9 : return;
1654 :
1655 : /*
1656 : * Failed. Try to create basedir first in case it's missing. Tolerate
1657 : * EEXIST to close a race against another process following the same
1658 : * algorithm.
1659 : */
1660 13 : if (MakePGDirectory(basedir) < 0 && errno != EEXIST)
1661 0 : ereport(ERROR,
1662 : (errcode_for_file_access(),
1663 : errmsg("cannot create temporary directory \"%s\": %m",
1664 : basedir)));
1665 :
1666 : /* Try again. */
1667 13 : if (MakePGDirectory(directory) < 0 && errno != EEXIST)
1668 0 : ereport(ERROR,
1669 : (errcode_for_file_access(),
1670 : errmsg("cannot create temporary subdirectory \"%s\": %m",
1671 : directory)));
1672 : }
1673 : }
1674 :
1675 : /*
1676 : * Delete a directory and everything in it, if it exists.
1677 : */
1678 : void
1679 309 : PathNameDeleteTemporaryDir(const char *dirname)
1680 : {
1681 : struct stat statbuf;
1682 :
1683 : /* Silently ignore missing directory. */
1684 309 : if (stat(dirname, &statbuf) != 0 && errno == ENOENT)
1685 57 : return;
1686 :
1687 : /*
1688 : * Currently, walkdir doesn't offer a way for our passed in function to
1689 : * maintain state. Perhaps it should, so that we could tell the caller
1690 : * whether this operation succeeded or failed. Since this operation is
1691 : * used in a cleanup path, we wouldn't actually behave differently: we'll
1692 : * just log failures.
1693 : */
1694 252 : walkdir(dirname, unlink_if_exists_fname, false, LOG);
1695 : }
1696 :
1697 : /*
1698 : * Open a temporary file that will disappear when we close it.
1699 : *
1700 : * This routine takes care of generating an appropriate tempfile name.
1701 : * There's no need to pass in fileFlags or fileMode either, since only
1702 : * one setting makes any sense for a temp file.
1703 : *
1704 : * Unless interXact is true, the file is remembered by CurrentResourceOwner
1705 : * to ensure it's closed and deleted when it's no longer needed, typically at
1706 : * the end-of-transaction. In most cases, you don't want temporary files to
1707 : * outlive the transaction that created them, so this should be false -- but
1708 : * if you need "somewhat" temporary storage, this might be useful. In either
1709 : * case, the file is removed when the File is explicitly closed.
1710 : */
1711 : File
1712 2001 : OpenTemporaryFile(bool interXact)
1713 : {
1714 2001 : File file = 0;
1715 :
1716 : Assert(temporary_files_allowed); /* check temp file access is up */
1717 :
1718 : /*
1719 : * Make sure the current resource owner has space for this File before we
1720 : * open it, if we'll be registering it below.
1721 : */
1722 2001 : if (!interXact)
1723 2001 : ResourceOwnerEnlarge(CurrentResourceOwner);
1724 :
1725 : /*
1726 : * If some temp tablespace(s) have been given to us, try to use the next
1727 : * one. If a given tablespace can't be found, we silently fall back to
1728 : * the database's default tablespace.
1729 : *
1730 : * BUT: if the temp file is slated to outlive the current transaction,
1731 : * force it into the database's default tablespace, so that it will not
1732 : * pose a threat to possible tablespace drop attempts.
1733 : */
1734 2001 : if (numTempTableSpaces > 0 && !interXact)
1735 : {
1736 1 : Oid tblspcOid = GetNextTempTableSpace();
1737 :
1738 1 : if (OidIsValid(tblspcOid))
1739 1 : file = OpenTemporaryFileInTablespace(tblspcOid, false);
1740 : }
1741 :
1742 : /*
1743 : * If not, or if tablespace is bad, create in database's default
1744 : * tablespace. MyDatabaseTableSpace should normally be set before we get
1745 : * here, but just in case it isn't, fall back to pg_default tablespace.
1746 : */
1747 2001 : if (file <= 0)
1748 2000 : file = OpenTemporaryFileInTablespace(MyDatabaseTableSpace ?
1749 : MyDatabaseTableSpace :
1750 : DEFAULTTABLESPACE_OID,
1751 : true);
1752 :
1753 : /* Mark it for deletion at close and temporary file size limit */
1754 2001 : VfdCache[file].fdstate |= FD_DELETE_AT_CLOSE | FD_TEMP_FILE_LIMIT;
1755 :
1756 : /* Register it with the current resource owner */
1757 2001 : if (!interXact)
1758 2001 : RegisterTemporaryFile(file);
1759 :
1760 2001 : return file;
1761 : }
1762 :
1763 : /*
1764 : * Return the path of the temp directory in a given tablespace.
1765 : */
1766 : void
1767 11766 : TempTablespacePath(char *path, Oid tablespace)
1768 : {
1769 : /*
1770 : * Identify the tempfile directory for this tablespace.
1771 : *
1772 : * If someone tries to specify pg_global, use pg_default instead.
1773 : */
1774 11766 : if (tablespace == InvalidOid ||
1775 1 : tablespace == DEFAULTTABLESPACE_OID ||
1776 : tablespace == GLOBALTABLESPACE_OID)
1777 11765 : snprintf(path, MAXPGPATH, "base/%s", PG_TEMP_FILES_DIR);
1778 : else
1779 : {
1780 : /* All other tablespaces are accessed via symlinks */
1781 1 : snprintf(path, MAXPGPATH, "%s/%u/%s/%s",
1782 : PG_TBLSPC_DIR, tablespace, TABLESPACE_VERSION_DIRECTORY,
1783 : PG_TEMP_FILES_DIR);
1784 : }
1785 11766 : }
1786 :
1787 : /*
1788 : * Open a temporary file in a specific tablespace.
1789 : * Subroutine for OpenTemporaryFile, which see for details.
1790 : */
1791 : static File
1792 2001 : OpenTemporaryFileInTablespace(Oid tblspcOid, bool rejectError)
1793 : {
1794 : char tempdirpath[MAXPGPATH];
1795 : char tempfilepath[MAXPGPATH];
1796 : File file;
1797 :
1798 2001 : TempTablespacePath(tempdirpath, tblspcOid);
1799 :
1800 : /*
1801 : * Generate a tempfile name that should be unique within the current
1802 : * database instance.
1803 : */
1804 2001 : snprintf(tempfilepath, sizeof(tempfilepath), "%s/%s%d.%ld",
1805 : tempdirpath, PG_TEMP_FILE_PREFIX, MyProcPid, tempFileCounter++);
1806 :
1807 : /*
1808 : * Open the file. Note: we don't use O_EXCL, in case there is an orphaned
1809 : * temp file that can be reused.
1810 : */
1811 2001 : file = PathNameOpenFile(tempfilepath,
1812 : O_RDWR | O_CREAT | O_TRUNC | PG_BINARY);
1813 2001 : if (file <= 0)
1814 : {
1815 : /*
1816 : * We might need to create the tablespace's tempfile directory, if no
1817 : * one has yet done so.
1818 : *
1819 : * Don't check for an error from MakePGDirectory; it could fail if
1820 : * someone else just did the same thing. If it doesn't work then
1821 : * we'll bomb out on the second create attempt, instead.
1822 : */
1823 103 : (void) MakePGDirectory(tempdirpath);
1824 :
1825 103 : file = PathNameOpenFile(tempfilepath,
1826 : O_RDWR | O_CREAT | O_TRUNC | PG_BINARY);
1827 103 : if (file <= 0 && rejectError)
1828 0 : elog(ERROR, "could not create temporary file \"%s\": %m",
1829 : tempfilepath);
1830 : }
1831 :
1832 2001 : return file;
1833 : }
1834 :
1835 :
1836 : /*
1837 : * Create a new file. The directory containing it must already exist. Files
1838 : * created this way are subject to temp_file_limit and are automatically
1839 : * closed at end of transaction, but are not automatically deleted on close
1840 : * because they are intended to be shared between cooperating backends.
1841 : *
1842 : * If the file is inside the top-level temporary directory, its name should
1843 : * begin with PG_TEMP_FILE_PREFIX so that it can be identified as temporary
1844 : * and deleted at startup by RemovePgTempFiles(). Alternatively, it can be
1845 : * inside a directory created with PathNameCreateTemporaryDir(), in which case
1846 : * the prefix isn't needed.
1847 : */
1848 : File
1849 2117 : PathNameCreateTemporaryFile(const char *path, bool error_on_failure)
1850 : {
1851 : File file;
1852 :
1853 : Assert(temporary_files_allowed); /* check temp file access is up */
1854 :
1855 2117 : ResourceOwnerEnlarge(CurrentResourceOwner);
1856 :
1857 : /*
1858 : * Open the file. Note: we don't use O_EXCL, in case there is an orphaned
1859 : * temp file that can be reused.
1860 : */
1861 2117 : file = PathNameOpenFile(path, O_RDWR | O_CREAT | O_TRUNC | PG_BINARY);
1862 2117 : if (file <= 0)
1863 : {
1864 262 : if (error_on_failure)
1865 0 : ereport(ERROR,
1866 : (errcode_for_file_access(),
1867 : errmsg("could not create temporary file \"%s\": %m",
1868 : path)));
1869 : else
1870 262 : return file;
1871 : }
1872 :
1873 : /* Mark it for temp_file_limit accounting. */
1874 1855 : VfdCache[file].fdstate |= FD_TEMP_FILE_LIMIT;
1875 :
1876 : /* Register it for automatic close. */
1877 1855 : RegisterTemporaryFile(file);
1878 :
1879 1855 : return file;
1880 : }
1881 :
1882 : /*
1883 : * Open a file that was created with PathNameCreateTemporaryFile, possibly in
1884 : * another backend. Files opened this way don't count against the
1885 : * temp_file_limit of the caller, are automatically closed at the end of the
1886 : * transaction but are not deleted on close.
1887 : */
1888 : File
1889 4831 : PathNameOpenTemporaryFile(const char *path, int mode)
1890 : {
1891 : File file;
1892 :
1893 : Assert(temporary_files_allowed); /* check temp file access is up */
1894 :
1895 4831 : ResourceOwnerEnlarge(CurrentResourceOwner);
1896 :
1897 4831 : file = PathNameOpenFile(path, mode | PG_BINARY);
1898 :
1899 : /* If no such file, then we don't raise an error. */
1900 4831 : if (file <= 0 && errno != ENOENT)
1901 0 : ereport(ERROR,
1902 : (errcode_for_file_access(),
1903 : errmsg("could not open temporary file \"%s\": %m",
1904 : path)));
1905 :
1906 4831 : if (file > 0)
1907 : {
1908 : /* Register it for automatic close. */
1909 2279 : RegisterTemporaryFile(file);
1910 : }
1911 :
1912 4831 : return file;
1913 : }
1914 :
1915 : /*
1916 : * Delete a file by pathname. Return true if the file existed, false if
1917 : * didn't.
1918 : */
1919 : bool
1920 4062 : PathNameDeleteTemporaryFile(const char *path, bool error_on_failure)
1921 : {
1922 : struct stat filestats;
1923 : int stat_errno;
1924 :
1925 : /* Get the final size for pgstat reporting. */
1926 4062 : if (stat(path, &filestats) != 0)
1927 2207 : stat_errno = errno;
1928 : else
1929 1855 : stat_errno = 0;
1930 :
1931 : /*
1932 : * Unlike FileClose's automatic file deletion code, we tolerate
1933 : * non-existence to support BufFileDeleteFileSet which doesn't know how
1934 : * many segments it has to delete until it runs out.
1935 : */
1936 4062 : if (stat_errno == ENOENT)
1937 2207 : return false;
1938 :
1939 1855 : if (unlink(path) < 0)
1940 : {
1941 0 : if (errno != ENOENT)
1942 0 : ereport(error_on_failure ? ERROR : LOG,
1943 : (errcode_for_file_access(),
1944 : errmsg("could not unlink temporary file \"%s\": %m",
1945 : path)));
1946 0 : return false;
1947 : }
1948 :
1949 1855 : if (stat_errno == 0)
1950 1855 : ReportTemporaryFileUsage(path, filestats.st_size);
1951 : else
1952 : {
1953 0 : errno = stat_errno;
1954 0 : ereport(LOG,
1955 : (errcode_for_file_access(),
1956 : errmsg("could not stat file \"%s\": %m", path)));
1957 : }
1958 :
1959 1855 : return true;
1960 : }
1961 :
1962 : /*
1963 : * close a file when done with it
1964 : */
1965 : void
1966 659120 : FileClose(File file)
1967 : {
1968 : Vfd *vfdP;
1969 :
1970 : Assert(FileIsValid(file));
1971 :
1972 : DO_DB(elog(LOG, "FileClose: %d (%s)",
1973 : file, VfdCache[file].fileName));
1974 :
1975 659120 : vfdP = &VfdCache[file];
1976 :
1977 659120 : if (!FileIsNotOpen(file))
1978 : {
1979 658694 : pgaio_closing_fd(vfdP->fd);
1980 :
1981 : /* close the file */
1982 658694 : if (close(vfdP->fd) != 0)
1983 : {
1984 : /*
1985 : * We may need to panic on failure to close non-temporary files;
1986 : * see LruDelete.
1987 : */
1988 0 : elog(vfdP->fdstate & FD_TEMP_FILE_LIMIT ? LOG : data_sync_elevel(LOG),
1989 : "could not close file \"%s\": %m", vfdP->fileName);
1990 : }
1991 :
1992 658694 : --nfile;
1993 658694 : vfdP->fd = VFD_CLOSED;
1994 :
1995 : /* remove the file from the lru ring */
1996 658694 : Delete(file);
1997 : }
1998 :
1999 659120 : if (vfdP->fdstate & FD_TEMP_FILE_LIMIT)
2000 : {
2001 : /* Subtract its size from current usage (do first in case of error) */
2002 3856 : temporary_files_size -= vfdP->fileSize;
2003 3856 : vfdP->fileSize = 0;
2004 : }
2005 :
2006 : /*
2007 : * Delete the file if it was temporary, and make a log entry if wanted
2008 : */
2009 659120 : if (vfdP->fdstate & FD_DELETE_AT_CLOSE)
2010 : {
2011 : struct stat filestats;
2012 : int stat_errno;
2013 :
2014 : /*
2015 : * If we get an error, as could happen within the ereport/elog calls,
2016 : * we'll come right back here during transaction abort. Reset the
2017 : * flag to ensure that we can't get into an infinite loop. This code
2018 : * is arranged to ensure that the worst-case consequence is failing to
2019 : * emit log message(s), not failing to attempt the unlink.
2020 : */
2021 2001 : vfdP->fdstate &= ~FD_DELETE_AT_CLOSE;
2022 :
2023 :
2024 : /* first try the stat() */
2025 2001 : if (stat(vfdP->fileName, &filestats))
2026 0 : stat_errno = errno;
2027 : else
2028 2001 : stat_errno = 0;
2029 :
2030 : /* in any case do the unlink */
2031 2001 : if (unlink(vfdP->fileName))
2032 0 : ereport(LOG,
2033 : (errcode_for_file_access(),
2034 : errmsg("could not delete file \"%s\": %m", vfdP->fileName)));
2035 :
2036 : /* and last report the stat results */
2037 2001 : if (stat_errno == 0)
2038 2001 : ReportTemporaryFileUsage(vfdP->fileName, filestats.st_size);
2039 : else
2040 : {
2041 0 : errno = stat_errno;
2042 0 : ereport(LOG,
2043 : (errcode_for_file_access(),
2044 : errmsg("could not stat file \"%s\": %m", vfdP->fileName)));
2045 : }
2046 : }
2047 :
2048 : /* Unregister it from the resource owner */
2049 659120 : if (vfdP->resowner)
2050 6130 : ResourceOwnerForgetFile(vfdP->resowner, file);
2051 :
2052 : /*
2053 : * Return the Vfd slot to the free list
2054 : */
2055 659120 : FreeVfd(file);
2056 659120 : }
2057 :
2058 : /*
2059 : * FilePrefetch - initiate asynchronous read of a given range of the file.
2060 : *
2061 : * Returns 0 on success, otherwise an errno error code (like posix_fadvise()).
2062 : *
2063 : * posix_fadvise() is the simplest standardized interface that accomplishes
2064 : * this.
2065 : */
2066 : int
2067 8778 : FilePrefetch(File file, pgoff_t offset, pgoff_t amount, uint32 wait_event_info)
2068 : {
2069 : Assert(FileIsValid(file));
2070 :
2071 : DO_DB(elog(LOG, "FilePrefetch: %d (%s) " INT64_FORMAT " " INT64_FORMAT,
2072 : file, VfdCache[file].fileName,
2073 : (int64) offset, (int64) amount));
2074 :
2075 : #if defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_WILLNEED)
2076 : {
2077 : int returnCode;
2078 :
2079 8778 : returnCode = FileAccess(file);
2080 8778 : if (returnCode < 0)
2081 0 : return returnCode;
2082 :
2083 8778 : retry:
2084 8778 : pgstat_report_wait_start(wait_event_info);
2085 8778 : returnCode = posix_fadvise(VfdCache[file].fd, offset, amount,
2086 : POSIX_FADV_WILLNEED);
2087 8778 : pgstat_report_wait_end();
2088 :
2089 8778 : if (returnCode == EINTR)
2090 0 : goto retry;
2091 :
2092 8778 : return returnCode;
2093 : }
2094 : #elif defined(__darwin__)
2095 : {
2096 : struct radvisory
2097 : {
2098 : off_t ra_offset; /* offset into the file */
2099 : int ra_count; /* size of the read */
2100 : } ra;
2101 : int returnCode;
2102 :
2103 : returnCode = FileAccess(file);
2104 : if (returnCode < 0)
2105 : return returnCode;
2106 :
2107 : ra.ra_offset = offset;
2108 : ra.ra_count = amount;
2109 : pgstat_report_wait_start(wait_event_info);
2110 : returnCode = fcntl(VfdCache[file].fd, F_RDADVISE, &ra);
2111 : pgstat_report_wait_end();
2112 : if (returnCode != -1)
2113 : return 0;
2114 : else
2115 : return errno;
2116 : }
2117 : #else
2118 : return 0;
2119 : #endif
2120 : }
2121 :
2122 : void
2123 0 : FileWriteback(File file, pgoff_t offset, pgoff_t nbytes, uint32 wait_event_info)
2124 : {
2125 : int returnCode;
2126 :
2127 : Assert(FileIsValid(file));
2128 :
2129 : DO_DB(elog(LOG, "FileWriteback: %d (%s) " INT64_FORMAT " " INT64_FORMAT,
2130 : file, VfdCache[file].fileName,
2131 : (int64) offset, (int64) nbytes));
2132 :
2133 0 : if (nbytes <= 0)
2134 0 : return;
2135 :
2136 0 : if (VfdCache[file].fileFlags & PG_O_DIRECT)
2137 0 : return;
2138 :
2139 0 : returnCode = FileAccess(file);
2140 0 : if (returnCode < 0)
2141 0 : return;
2142 :
2143 0 : pgstat_report_wait_start(wait_event_info);
2144 0 : pg_flush_data(VfdCache[file].fd, offset, nbytes);
2145 0 : pgstat_report_wait_end();
2146 : }
2147 :
2148 : ssize_t
2149 420091 : FileReadV(File file, const struct iovec *iov, int iovcnt, pgoff_t offset,
2150 : uint32 wait_event_info)
2151 : {
2152 : ssize_t returnCode;
2153 : Vfd *vfdP;
2154 :
2155 : Assert(FileIsValid(file));
2156 :
2157 : DO_DB(elog(LOG, "FileReadV: %d (%s) " INT64_FORMAT " %d",
2158 : file, VfdCache[file].fileName,
2159 : (int64) offset,
2160 : iovcnt));
2161 :
2162 420091 : returnCode = FileAccess(file);
2163 420091 : if (returnCode < 0)
2164 0 : return returnCode;
2165 :
2166 420091 : vfdP = &VfdCache[file];
2167 :
2168 420091 : retry:
2169 420091 : pgstat_report_wait_start(wait_event_info);
2170 420091 : returnCode = pg_preadv(vfdP->fd, iov, iovcnt, offset);
2171 420091 : pgstat_report_wait_end();
2172 :
2173 420091 : if (returnCode < 0)
2174 : {
2175 : /*
2176 : * Windows may run out of kernel buffers and return "Insufficient
2177 : * system resources" error. Wait a bit and retry to solve it.
2178 : *
2179 : * It is rumored that EINTR is also possible on some Unix filesystems,
2180 : * in which case immediate retry is indicated.
2181 : */
2182 : #ifdef WIN32
2183 : DWORD error = GetLastError();
2184 :
2185 : switch (error)
2186 : {
2187 : case ERROR_NO_SYSTEM_RESOURCES:
2188 : pg_usleep(1000L);
2189 : errno = EINTR;
2190 : break;
2191 : default:
2192 : _dosmaperr(error);
2193 : break;
2194 : }
2195 : #endif
2196 : /* OK to retry if interrupted */
2197 0 : if (errno == EINTR)
2198 0 : goto retry;
2199 : }
2200 :
2201 420091 : return returnCode;
2202 : }
2203 :
2204 : int
2205 1378704 : FileStartReadV(PgAioHandle *ioh, File file,
2206 : int iovcnt, pgoff_t offset,
2207 : uint32 wait_event_info)
2208 : {
2209 : int returnCode;
2210 : Vfd *vfdP;
2211 :
2212 : Assert(FileIsValid(file));
2213 :
2214 : DO_DB(elog(LOG, "FileStartReadV: %d (%s) " INT64_FORMAT " %d",
2215 : file, VfdCache[file].fileName,
2216 : (int64) offset,
2217 : iovcnt));
2218 :
2219 1378704 : returnCode = FileAccess(file);
2220 1378704 : if (returnCode < 0)
2221 0 : return returnCode;
2222 :
2223 1378704 : vfdP = &VfdCache[file];
2224 :
2225 1378704 : pgaio_io_start_readv(ioh, vfdP->fd, iovcnt, offset);
2226 :
2227 1378704 : return 0;
2228 : }
2229 :
2230 : ssize_t
2231 831411 : FileWriteV(File file, const struct iovec *iov, int iovcnt, pgoff_t offset,
2232 : uint32 wait_event_info)
2233 : {
2234 : ssize_t returnCode;
2235 : Vfd *vfdP;
2236 :
2237 : Assert(FileIsValid(file));
2238 :
2239 : DO_DB(elog(LOG, "FileWriteV: %d (%s) " INT64_FORMAT " %d",
2240 : file, VfdCache[file].fileName,
2241 : (int64) offset,
2242 : iovcnt));
2243 :
2244 831411 : returnCode = FileAccess(file);
2245 831411 : if (returnCode < 0)
2246 0 : return returnCode;
2247 :
2248 831411 : vfdP = &VfdCache[file];
2249 :
2250 : /*
2251 : * If enforcing temp_file_limit and it's a temp file, check to see if the
2252 : * write would overrun temp_file_limit, and throw error if so. Note: it's
2253 : * really a modularity violation to throw error here; we should set errno
2254 : * and return -1. However, there's no way to report a suitable error
2255 : * message if we do that. All current callers would just throw error
2256 : * immediately anyway, so this is safe at present.
2257 : */
2258 831411 : if (temp_file_limit >= 0 && (vfdP->fdstate & FD_TEMP_FILE_LIMIT))
2259 : {
2260 0 : pgoff_t past_write = offset;
2261 :
2262 0 : for (int i = 0; i < iovcnt; ++i)
2263 0 : past_write += iov[i].iov_len;
2264 :
2265 0 : if (past_write > vfdP->fileSize)
2266 : {
2267 0 : uint64 newTotal = temporary_files_size;
2268 :
2269 0 : newTotal += past_write - vfdP->fileSize;
2270 0 : if (newTotal > (uint64) temp_file_limit * (uint64) 1024)
2271 0 : ereport(ERROR,
2272 : (errcode(ERRCODE_CONFIGURATION_LIMIT_EXCEEDED),
2273 : errmsg("temporary file size exceeds \"temp_file_limit\" (%dkB)",
2274 : temp_file_limit)));
2275 : }
2276 : }
2277 :
2278 831411 : retry:
2279 831411 : pgstat_report_wait_start(wait_event_info);
2280 831411 : returnCode = pg_pwritev(vfdP->fd, iov, iovcnt, offset);
2281 831411 : pgstat_report_wait_end();
2282 :
2283 831411 : if (returnCode >= 0)
2284 : {
2285 : /*
2286 : * Some callers expect short writes to set errno, and traditionally we
2287 : * have assumed that they imply disk space shortage. We don't want to
2288 : * waste CPU cycles adding up the total size here, so we'll just set
2289 : * it for all successful writes in case such a caller determines that
2290 : * the write was short and ereports "%m".
2291 : */
2292 831411 : errno = ENOSPC;
2293 :
2294 : /*
2295 : * Maintain fileSize and temporary_files_size if it's a temp file.
2296 : */
2297 831411 : if (vfdP->fdstate & FD_TEMP_FILE_LIMIT)
2298 : {
2299 71442 : pgoff_t past_write = offset + returnCode;
2300 :
2301 71442 : if (past_write > vfdP->fileSize)
2302 : {
2303 49382 : temporary_files_size += past_write - vfdP->fileSize;
2304 49382 : vfdP->fileSize = past_write;
2305 : }
2306 : }
2307 : }
2308 : else
2309 : {
2310 : /*
2311 : * See comments in FileReadV()
2312 : */
2313 : #ifdef WIN32
2314 : DWORD error = GetLastError();
2315 :
2316 : switch (error)
2317 : {
2318 : case ERROR_NO_SYSTEM_RESOURCES:
2319 : pg_usleep(1000L);
2320 : errno = EINTR;
2321 : break;
2322 : default:
2323 : _dosmaperr(error);
2324 : break;
2325 : }
2326 : #endif
2327 : /* OK to retry if interrupted */
2328 0 : if (errno == EINTR)
2329 0 : goto retry;
2330 : }
2331 :
2332 831411 : return returnCode;
2333 : }
2334 :
2335 : int
2336 413 : FileSync(File file, uint32 wait_event_info)
2337 : {
2338 : int returnCode;
2339 :
2340 : Assert(FileIsValid(file));
2341 :
2342 : DO_DB(elog(LOG, "FileSync: %d (%s)",
2343 : file, VfdCache[file].fileName));
2344 :
2345 413 : returnCode = FileAccess(file);
2346 413 : if (returnCode < 0)
2347 0 : return returnCode;
2348 :
2349 413 : pgstat_report_wait_start(wait_event_info);
2350 413 : returnCode = pg_fsync(VfdCache[file].fd);
2351 413 : pgstat_report_wait_end();
2352 :
2353 413 : return returnCode;
2354 : }
2355 :
2356 : /*
2357 : * Zero a region of the file.
2358 : *
2359 : * Returns 0 on success, -1 otherwise. In the latter case errno is set to the
2360 : * appropriate error.
2361 : */
2362 : int
2363 245333 : FileZero(File file, pgoff_t offset, pgoff_t amount, uint32 wait_event_info)
2364 : {
2365 : int returnCode;
2366 : ssize_t written;
2367 :
2368 : Assert(FileIsValid(file));
2369 :
2370 : DO_DB(elog(LOG, "FileZero: %d (%s) " INT64_FORMAT " " INT64_FORMAT,
2371 : file, VfdCache[file].fileName,
2372 : (int64) offset, (int64) amount));
2373 :
2374 245333 : returnCode = FileAccess(file);
2375 245333 : if (returnCode < 0)
2376 0 : return returnCode;
2377 :
2378 245333 : pgstat_report_wait_start(wait_event_info);
2379 245333 : written = pg_pwrite_zeros(VfdCache[file].fd, amount, offset);
2380 245333 : pgstat_report_wait_end();
2381 :
2382 245333 : if (written < 0)
2383 0 : return -1;
2384 245333 : else if (written != amount)
2385 : {
2386 : /* if errno is unset, assume problem is no disk space */
2387 0 : if (errno == 0)
2388 0 : errno = ENOSPC;
2389 0 : return -1;
2390 : }
2391 :
2392 245333 : return 0;
2393 : }
2394 :
2395 : /*
2396 : * Try to reserve file space with posix_fallocate(). If posix_fallocate() is
2397 : * not implemented on the operating system or fails with EINVAL / EOPNOTSUPP,
2398 : * use FileZero() instead.
2399 : *
2400 : * Note that at least glibc() implements posix_fallocate() in userspace if not
2401 : * implemented by the filesystem. That's not the case for all environments
2402 : * though.
2403 : *
2404 : * Returns 0 on success, -1 otherwise. In the latter case errno is set to the
2405 : * appropriate error.
2406 : */
2407 : int
2408 694 : FileFallocate(File file, pgoff_t offset, pgoff_t amount, uint32 wait_event_info)
2409 : {
2410 : #ifdef HAVE_POSIX_FALLOCATE
2411 : int returnCode;
2412 :
2413 : Assert(FileIsValid(file));
2414 :
2415 : DO_DB(elog(LOG, "FileFallocate: %d (%s) " INT64_FORMAT " " INT64_FORMAT,
2416 : file, VfdCache[file].fileName,
2417 : (int64) offset, (int64) amount));
2418 :
2419 694 : returnCode = FileAccess(file);
2420 694 : if (returnCode < 0)
2421 0 : return -1;
2422 :
2423 694 : retry:
2424 694 : pgstat_report_wait_start(wait_event_info);
2425 694 : returnCode = posix_fallocate(VfdCache[file].fd, offset, amount);
2426 694 : pgstat_report_wait_end();
2427 :
2428 694 : if (returnCode == 0)
2429 694 : return 0;
2430 0 : else if (returnCode == EINTR)
2431 0 : goto retry;
2432 :
2433 : /* for compatibility with %m printing etc */
2434 0 : errno = returnCode;
2435 :
2436 : /*
2437 : * Return in cases of a "real" failure, if fallocate is not supported,
2438 : * fall through to the FileZero() backed implementation.
2439 : */
2440 0 : if (returnCode != EINVAL && returnCode != EOPNOTSUPP)
2441 0 : return -1;
2442 : #endif
2443 :
2444 0 : return FileZero(file, offset, amount, wait_event_info);
2445 : }
2446 :
2447 : pgoff_t
2448 3457890 : FileSize(File file)
2449 : {
2450 : Assert(FileIsValid(file));
2451 :
2452 : DO_DB(elog(LOG, "FileSize %d (%s)",
2453 : file, VfdCache[file].fileName));
2454 :
2455 3457890 : if (FileIsNotOpen(file))
2456 : {
2457 29 : if (FileAccess(file) < 0)
2458 0 : return (pgoff_t) -1;
2459 : }
2460 :
2461 3457890 : return lseek(VfdCache[file].fd, 0, SEEK_END);
2462 : }
2463 :
2464 : int
2465 672 : FileTruncate(File file, pgoff_t offset, uint32 wait_event_info)
2466 : {
2467 : int returnCode;
2468 :
2469 : Assert(FileIsValid(file));
2470 :
2471 : DO_DB(elog(LOG, "FileTruncate %d (%s)",
2472 : file, VfdCache[file].fileName));
2473 :
2474 672 : returnCode = FileAccess(file);
2475 672 : if (returnCode < 0)
2476 0 : return returnCode;
2477 :
2478 672 : pgstat_report_wait_start(wait_event_info);
2479 672 : returnCode = pg_ftruncate(VfdCache[file].fd, offset);
2480 672 : pgstat_report_wait_end();
2481 :
2482 672 : if (returnCode == 0 && VfdCache[file].fileSize > offset)
2483 : {
2484 : /* adjust our state for truncation of a temp file */
2485 : Assert(VfdCache[file].fdstate & FD_TEMP_FILE_LIMIT);
2486 0 : temporary_files_size -= VfdCache[file].fileSize - offset;
2487 0 : VfdCache[file].fileSize = offset;
2488 : }
2489 :
2490 672 : return returnCode;
2491 : }
2492 :
2493 : /*
2494 : * Return the pathname associated with an open file.
2495 : *
2496 : * The returned string points to an internal buffer, which is valid until
2497 : * the file is closed.
2498 : */
2499 : char *
2500 32 : FilePathName(File file)
2501 : {
2502 : Assert(FileIsValid(file));
2503 :
2504 32 : return VfdCache[file].fileName;
2505 : }
2506 :
2507 : /*
2508 : * Return the raw file descriptor of an opened file.
2509 : *
2510 : * The returned file descriptor will be valid until the file is closed, but
2511 : * there are a lot of things that can make that happen. So the caller should
2512 : * be careful not to do much of anything else before it finishes using the
2513 : * returned file descriptor.
2514 : */
2515 : int
2516 506398 : FileGetRawDesc(File file)
2517 : {
2518 : int returnCode;
2519 :
2520 506398 : returnCode = FileAccess(file);
2521 506398 : if (returnCode < 0)
2522 0 : return returnCode;
2523 :
2524 : Assert(FileIsValid(file));
2525 506398 : return VfdCache[file].fd;
2526 : }
2527 :
2528 : /*
2529 : * FileGetRawFlags - returns the file flags on open(2)
2530 : */
2531 : int
2532 0 : FileGetRawFlags(File file)
2533 : {
2534 : Assert(FileIsValid(file));
2535 0 : return VfdCache[file].fileFlags;
2536 : }
2537 :
2538 : /*
2539 : * FileGetRawMode - returns the mode bitmask passed to open(2)
2540 : */
2541 : mode_t
2542 0 : FileGetRawMode(File file)
2543 : {
2544 : Assert(FileIsValid(file));
2545 0 : return VfdCache[file].fileMode;
2546 : }
2547 :
2548 : /*
2549 : * Make room for another allocatedDescs[] array entry if needed and possible.
2550 : * Returns true if an array element is available.
2551 : */
2552 : static bool
2553 7936490 : reserveAllocatedDesc(void)
2554 : {
2555 : AllocateDesc *newDescs;
2556 : int newMax;
2557 :
2558 : /* Quick out if array already has a free slot. */
2559 7936490 : if (numAllocatedDescs < maxAllocatedDescs)
2560 7935292 : return true;
2561 :
2562 : /*
2563 : * If the array hasn't yet been created in the current process, initialize
2564 : * it with FD_MINFREE / 3 elements. In many scenarios this is as many as
2565 : * we will ever need, anyway. We don't want to look at max_safe_fds
2566 : * immediately because set_max_safe_fds() may not have run yet.
2567 : */
2568 1198 : if (allocatedDescs == NULL)
2569 : {
2570 1198 : newMax = FD_MINFREE / 3;
2571 1198 : newDescs = (AllocateDesc *) malloc(newMax * sizeof(AllocateDesc));
2572 : /* Out of memory already? Treat as fatal error. */
2573 1198 : if (newDescs == NULL)
2574 0 : ereport(ERROR,
2575 : (errcode(ERRCODE_OUT_OF_MEMORY),
2576 : errmsg("out of memory")));
2577 1198 : allocatedDescs = newDescs;
2578 1198 : maxAllocatedDescs = newMax;
2579 1198 : return true;
2580 : }
2581 :
2582 : /*
2583 : * Consider enlarging the array beyond the initial allocation used above.
2584 : * By the time this happens, max_safe_fds should be known accurately.
2585 : *
2586 : * We mustn't let allocated descriptors hog all the available FDs, and in
2587 : * practice we'd better leave a reasonable number of FDs for VFD use. So
2588 : * set the maximum to max_safe_fds / 3. (This should certainly be at
2589 : * least as large as the initial size, FD_MINFREE / 3, so we aren't
2590 : * tightening the restriction here.) Recall that "external" FDs are
2591 : * allowed to consume another third of max_safe_fds.
2592 : */
2593 0 : newMax = max_safe_fds / 3;
2594 0 : if (newMax > maxAllocatedDescs)
2595 : {
2596 0 : newDescs = (AllocateDesc *) realloc(allocatedDescs,
2597 : newMax * sizeof(AllocateDesc));
2598 : /* Treat out-of-memory as a non-fatal error. */
2599 0 : if (newDescs == NULL)
2600 0 : return false;
2601 0 : allocatedDescs = newDescs;
2602 0 : maxAllocatedDescs = newMax;
2603 0 : return true;
2604 : }
2605 :
2606 : /* Can't enlarge allocatedDescs[] any more. */
2607 0 : return false;
2608 : }
2609 :
2610 : /*
2611 : * Routines that want to use stdio (ie, FILE*) should use AllocateFile
2612 : * rather than plain fopen(). This lets fd.c deal with freeing FDs if
2613 : * necessary to open the file. When done, call FreeFile rather than fclose.
2614 : *
2615 : * Note that files that will be open for any significant length of time
2616 : * should NOT be handled this way, since they cannot share kernel file
2617 : * descriptors with other files; there is grave risk of running out of FDs
2618 : * if anyone locks down too many FDs. Most callers of this routine are
2619 : * simply reading a config file that they will read and close immediately.
2620 : *
2621 : * fd.c will automatically close all files opened with AllocateFile at
2622 : * transaction commit or abort; this prevents FD leakage if a routine
2623 : * that calls AllocateFile is terminated prematurely by ereport(ERROR).
2624 : *
2625 : * Ideally this should be the *only* direct call of fopen() in the backend.
2626 : */
2627 : FILE *
2628 99141 : AllocateFile(const char *name, const char *mode)
2629 : {
2630 : FILE *file;
2631 :
2632 : DO_DB(elog(LOG, "AllocateFile: Allocated %d (%s)",
2633 : numAllocatedDescs, name));
2634 :
2635 : /* Can we allocate another non-virtual FD? */
2636 99141 : if (!reserveAllocatedDesc())
2637 0 : ereport(ERROR,
2638 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
2639 : errmsg("exceeded maxAllocatedDescs (%d) while trying to open file \"%s\"",
2640 : maxAllocatedDescs, name)));
2641 :
2642 : /* Close excess kernel FDs. */
2643 99141 : ReleaseLruFiles();
2644 :
2645 99141 : TryAgain:
2646 99141 : if ((file = fopen(name, mode)) != NULL)
2647 : {
2648 91437 : AllocateDesc *desc = &allocatedDescs[numAllocatedDescs];
2649 :
2650 91437 : desc->kind = AllocateDescFile;
2651 91437 : desc->desc.file = file;
2652 91437 : desc->create_subid = GetCurrentSubTransactionId();
2653 91437 : numAllocatedDescs++;
2654 91437 : return desc->desc.file;
2655 : }
2656 :
2657 7704 : if (errno == EMFILE || errno == ENFILE)
2658 : {
2659 0 : int save_errno = errno;
2660 :
2661 0 : ereport(LOG,
2662 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
2663 : errmsg("out of file descriptors: %m; release and retry")));
2664 0 : errno = 0;
2665 0 : if (ReleaseLruFile())
2666 0 : goto TryAgain;
2667 0 : errno = save_errno;
2668 : }
2669 :
2670 7704 : return NULL;
2671 : }
2672 :
2673 : /*
2674 : * Open a file with OpenTransientFilePerm() and pass default file mode for
2675 : * the fileMode parameter.
2676 : */
2677 : int
2678 7786024 : OpenTransientFile(const char *fileName, int fileFlags)
2679 : {
2680 7786024 : return OpenTransientFilePerm(fileName, fileFlags, pg_file_create_mode);
2681 : }
2682 :
2683 : /*
2684 : * Like AllocateFile, but returns an unbuffered fd like open(2)
2685 : */
2686 : int
2687 7786032 : OpenTransientFilePerm(const char *fileName, int fileFlags, mode_t fileMode)
2688 : {
2689 : int fd;
2690 :
2691 : DO_DB(elog(LOG, "OpenTransientFile: Allocated %d (%s)",
2692 : numAllocatedDescs, fileName));
2693 :
2694 : /* Can we allocate another non-virtual FD? */
2695 7786032 : if (!reserveAllocatedDesc())
2696 0 : ereport(ERROR,
2697 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
2698 : errmsg("exceeded maxAllocatedDescs (%d) while trying to open file \"%s\"",
2699 : maxAllocatedDescs, fileName)));
2700 :
2701 : /* Close excess kernel FDs. */
2702 7786032 : ReleaseLruFiles();
2703 :
2704 7786032 : fd = BasicOpenFilePerm(fileName, fileFlags, fileMode);
2705 :
2706 7786032 : if (fd >= 0)
2707 : {
2708 7780914 : AllocateDesc *desc = &allocatedDescs[numAllocatedDescs];
2709 :
2710 7780914 : desc->kind = AllocateDescRawFD;
2711 7780914 : desc->desc.fd = fd;
2712 7780914 : desc->create_subid = GetCurrentSubTransactionId();
2713 7780914 : numAllocatedDescs++;
2714 :
2715 7780914 : return fd;
2716 : }
2717 :
2718 5118 : return -1; /* failure */
2719 : }
2720 :
2721 : /*
2722 : * Routines that want to initiate a pipe stream should use OpenPipeStream
2723 : * rather than plain popen(). This lets fd.c deal with freeing FDs if
2724 : * necessary. When done, call ClosePipeStream rather than pclose.
2725 : *
2726 : * This function also ensures that the popen'd program is run with default
2727 : * SIGPIPE processing, rather than the SIG_IGN setting the backend normally
2728 : * uses. This ensures desirable response to, eg, closing a read pipe early.
2729 : */
2730 : FILE *
2731 69 : OpenPipeStream(const char *command, const char *mode)
2732 : {
2733 : FILE *file;
2734 : int save_errno;
2735 :
2736 : DO_DB(elog(LOG, "OpenPipeStream: Allocated %d (%s)",
2737 : numAllocatedDescs, command));
2738 :
2739 : /* Can we allocate another non-virtual FD? */
2740 69 : if (!reserveAllocatedDesc())
2741 0 : ereport(ERROR,
2742 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
2743 : errmsg("exceeded maxAllocatedDescs (%d) while trying to execute command \"%s\"",
2744 : maxAllocatedDescs, command)));
2745 :
2746 : /* Close excess kernel FDs. */
2747 69 : ReleaseLruFiles();
2748 :
2749 69 : TryAgain:
2750 69 : fflush(NULL);
2751 69 : pqsignal(SIGPIPE, SIG_DFL);
2752 69 : errno = 0;
2753 69 : file = popen(command, mode);
2754 69 : save_errno = errno;
2755 69 : pqsignal(SIGPIPE, SIG_IGN);
2756 69 : errno = save_errno;
2757 69 : if (file != NULL)
2758 : {
2759 69 : AllocateDesc *desc = &allocatedDescs[numAllocatedDescs];
2760 :
2761 69 : desc->kind = AllocateDescPipe;
2762 69 : desc->desc.file = file;
2763 69 : desc->create_subid = GetCurrentSubTransactionId();
2764 69 : numAllocatedDescs++;
2765 69 : return desc->desc.file;
2766 : }
2767 :
2768 0 : if (errno == EMFILE || errno == ENFILE)
2769 : {
2770 0 : ereport(LOG,
2771 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
2772 : errmsg("out of file descriptors: %m; release and retry")));
2773 0 : if (ReleaseLruFile())
2774 0 : goto TryAgain;
2775 0 : errno = save_errno;
2776 : }
2777 :
2778 0 : return NULL;
2779 : }
2780 :
2781 : /*
2782 : * Free an AllocateDesc of any type.
2783 : *
2784 : * The argument *must* point into the allocatedDescs[] array.
2785 : */
2786 : static int
2787 7922691 : FreeDesc(AllocateDesc *desc)
2788 : {
2789 : int result;
2790 :
2791 : /* Close the underlying object */
2792 7922691 : switch (desc->kind)
2793 : {
2794 91440 : case AllocateDescFile:
2795 91440 : result = fclose(desc->desc.file);
2796 91440 : break;
2797 69 : case AllocateDescPipe:
2798 69 : result = pclose(desc->desc.file);
2799 69 : break;
2800 50266 : case AllocateDescDir:
2801 50266 : result = closedir(desc->desc.dir);
2802 50266 : break;
2803 7780916 : case AllocateDescRawFD:
2804 7780916 : pgaio_closing_fd(desc->desc.fd);
2805 7780916 : result = close(desc->desc.fd);
2806 7780916 : break;
2807 0 : default:
2808 0 : elog(ERROR, "AllocateDesc kind not recognized");
2809 : result = 0; /* keep compiler quiet */
2810 : break;
2811 : }
2812 :
2813 : /* Compact storage in the allocatedDescs array */
2814 7922691 : numAllocatedDescs--;
2815 7922691 : *desc = allocatedDescs[numAllocatedDescs];
2816 :
2817 7922691 : return result;
2818 : }
2819 :
2820 : /*
2821 : * Close a file returned by AllocateFile.
2822 : *
2823 : * Note we do not check fclose's return value --- it is up to the caller
2824 : * to handle close errors.
2825 : */
2826 : int
2827 91420 : FreeFile(FILE *file)
2828 : {
2829 : int i;
2830 :
2831 : DO_DB(elog(LOG, "FreeFile: Allocated %d", numAllocatedDescs));
2832 :
2833 : /* Remove file from list of allocated files, if it's present */
2834 91423 : for (i = numAllocatedDescs; --i >= 0;)
2835 : {
2836 91423 : AllocateDesc *desc = &allocatedDescs[i];
2837 :
2838 91423 : if (desc->kind == AllocateDescFile && desc->desc.file == file)
2839 91420 : return FreeDesc(desc);
2840 : }
2841 :
2842 : /* Only get here if someone passes us a file not in allocatedDescs */
2843 0 : elog(WARNING, "file passed to FreeFile was not obtained from AllocateFile");
2844 :
2845 0 : return fclose(file);
2846 : }
2847 :
2848 : /*
2849 : * Close a file returned by OpenTransientFile.
2850 : *
2851 : * Note we do not check close's return value --- it is up to the caller
2852 : * to handle close errors.
2853 : */
2854 : int
2855 7780915 : CloseTransientFile(int fd)
2856 : {
2857 : int i;
2858 :
2859 : DO_DB(elog(LOG, "CloseTransientFile: Allocated %d", numAllocatedDescs));
2860 :
2861 : /* Remove fd from list of allocated files, if it's present */
2862 7780925 : for (i = numAllocatedDescs; --i >= 0;)
2863 : {
2864 7780925 : AllocateDesc *desc = &allocatedDescs[i];
2865 :
2866 7780925 : if (desc->kind == AllocateDescRawFD && desc->desc.fd == fd)
2867 7780915 : return FreeDesc(desc);
2868 : }
2869 :
2870 : /* Only get here if someone passes us a file not in allocatedDescs */
2871 0 : elog(WARNING, "fd passed to CloseTransientFile was not obtained from OpenTransientFile");
2872 :
2873 0 : pgaio_closing_fd(fd);
2874 :
2875 0 : return close(fd);
2876 : }
2877 :
2878 : /*
2879 : * Routines that want to use <dirent.h> (ie, DIR*) should use AllocateDir
2880 : * rather than plain opendir(). This lets fd.c deal with freeing FDs if
2881 : * necessary to open the directory, and with closing it after an elog.
2882 : * When done, call FreeDir rather than closedir.
2883 : *
2884 : * Returns NULL, with errno set, on failure. Note that failure detection
2885 : * is commonly left to the following call of ReadDir or ReadDirExtended;
2886 : * see the comments for ReadDir.
2887 : *
2888 : * Ideally this should be the *only* direct call of opendir() in the backend.
2889 : */
2890 : DIR *
2891 51248 : AllocateDir(const char *dirname)
2892 : {
2893 : DIR *dir;
2894 :
2895 : DO_DB(elog(LOG, "AllocateDir: Allocated %d (%s)",
2896 : numAllocatedDescs, dirname));
2897 :
2898 : /* Can we allocate another non-virtual FD? */
2899 51248 : if (!reserveAllocatedDesc())
2900 0 : ereport(ERROR,
2901 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
2902 : errmsg("exceeded maxAllocatedDescs (%d) while trying to open directory \"%s\"",
2903 : maxAllocatedDescs, dirname)));
2904 :
2905 : /* Close excess kernel FDs. */
2906 51248 : ReleaseLruFiles();
2907 :
2908 51248 : TryAgain:
2909 51248 : if ((dir = opendir(dirname)) != NULL)
2910 : {
2911 50266 : AllocateDesc *desc = &allocatedDescs[numAllocatedDescs];
2912 :
2913 50266 : desc->kind = AllocateDescDir;
2914 50266 : desc->desc.dir = dir;
2915 50266 : desc->create_subid = GetCurrentSubTransactionId();
2916 50266 : numAllocatedDescs++;
2917 50266 : return desc->desc.dir;
2918 : }
2919 :
2920 982 : if (errno == EMFILE || errno == ENFILE)
2921 : {
2922 0 : int save_errno = errno;
2923 :
2924 0 : ereport(LOG,
2925 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
2926 : errmsg("out of file descriptors: %m; release and retry")));
2927 0 : errno = 0;
2928 0 : if (ReleaseLruFile())
2929 0 : goto TryAgain;
2930 0 : errno = save_errno;
2931 : }
2932 :
2933 982 : return NULL;
2934 : }
2935 :
2936 : /*
2937 : * Read a directory opened with AllocateDir, ereport'ing any error.
2938 : *
2939 : * This is easier to use than raw readdir() since it takes care of some
2940 : * otherwise rather tedious and error-prone manipulation of errno. Also,
2941 : * if you are happy with a generic error message for AllocateDir failure,
2942 : * you can just do
2943 : *
2944 : * dir = AllocateDir(path);
2945 : * while ((dirent = ReadDir(dir, path)) != NULL)
2946 : * process dirent;
2947 : * FreeDir(dir);
2948 : *
2949 : * since a NULL dir parameter is taken as indicating AllocateDir failed.
2950 : * (Make sure errno isn't changed between AllocateDir and ReadDir if you
2951 : * use this shortcut.)
2952 : *
2953 : * The pathname passed to AllocateDir must be passed to this routine too,
2954 : * but it is only used for error reporting.
2955 : */
2956 : struct dirent *
2957 2545582 : ReadDir(DIR *dir, const char *dirname)
2958 : {
2959 2545582 : return ReadDirExtended(dir, dirname, ERROR);
2960 : }
2961 :
2962 : /*
2963 : * Alternate version of ReadDir that allows caller to specify the elevel
2964 : * for any error report (whether it's reporting an initial failure of
2965 : * AllocateDir or a subsequent directory read failure).
2966 : *
2967 : * If elevel < ERROR, returns NULL after any error. With the normal coding
2968 : * pattern, this will result in falling out of the loop immediately as
2969 : * though the directory contained no (more) entries.
2970 : */
2971 : struct dirent *
2972 4500446 : ReadDirExtended(DIR *dir, const char *dirname, int elevel)
2973 : {
2974 : struct dirent *dent;
2975 :
2976 : /* Give a generic message for AllocateDir failure, if caller didn't */
2977 4500446 : if (dir == NULL)
2978 : {
2979 4 : ereport(elevel,
2980 : (errcode_for_file_access(),
2981 : errmsg("could not open directory \"%s\": %m",
2982 : dirname)));
2983 0 : return NULL;
2984 : }
2985 :
2986 4500442 : errno = 0;
2987 4500442 : if ((dent = readdir(dir)) != NULL)
2988 4464025 : return dent;
2989 :
2990 36417 : if (errno)
2991 0 : ereport(elevel,
2992 : (errcode_for_file_access(),
2993 : errmsg("could not read directory \"%s\": %m",
2994 : dirname)));
2995 36417 : return NULL;
2996 : }
2997 :
2998 : /*
2999 : * Close a directory opened with AllocateDir.
3000 : *
3001 : * Returns closedir's return value (with errno set if it's not 0).
3002 : * Note we do not check the return value --- it is up to the caller
3003 : * to handle close errors if wanted.
3004 : *
3005 : * Does nothing if dir == NULL; we assume that directory open failure was
3006 : * already reported if desired.
3007 : */
3008 : int
3009 50133 : FreeDir(DIR *dir)
3010 : {
3011 : int i;
3012 :
3013 : /* Nothing to do if AllocateDir failed */
3014 50133 : if (dir == NULL)
3015 0 : return 0;
3016 :
3017 : DO_DB(elog(LOG, "FreeDir: Allocated %d", numAllocatedDescs));
3018 :
3019 : /* Remove dir from list of allocated dirs, if it's present */
3020 50133 : for (i = numAllocatedDescs; --i >= 0;)
3021 : {
3022 50133 : AllocateDesc *desc = &allocatedDescs[i];
3023 :
3024 50133 : if (desc->kind == AllocateDescDir && desc->desc.dir == dir)
3025 50133 : return FreeDesc(desc);
3026 : }
3027 :
3028 : /* Only get here if someone passes us a dir not in allocatedDescs */
3029 0 : elog(WARNING, "dir passed to FreeDir was not obtained from AllocateDir");
3030 :
3031 0 : return closedir(dir);
3032 : }
3033 :
3034 :
3035 : /*
3036 : * Close a pipe stream returned by OpenPipeStream.
3037 : */
3038 : int
3039 69 : ClosePipeStream(FILE *file)
3040 : {
3041 : int i;
3042 :
3043 : DO_DB(elog(LOG, "ClosePipeStream: Allocated %d", numAllocatedDescs));
3044 :
3045 : /* Remove file from list of allocated files, if it's present */
3046 69 : for (i = numAllocatedDescs; --i >= 0;)
3047 : {
3048 69 : AllocateDesc *desc = &allocatedDescs[i];
3049 :
3050 69 : if (desc->kind == AllocateDescPipe && desc->desc.file == file)
3051 69 : return FreeDesc(desc);
3052 : }
3053 :
3054 : /* Only get here if someone passes us a file not in allocatedDescs */
3055 0 : elog(WARNING, "file passed to ClosePipeStream was not obtained from OpenPipeStream");
3056 :
3057 0 : return pclose(file);
3058 : }
3059 :
3060 : /*
3061 : * closeAllVfds
3062 : *
3063 : * Force all VFDs into the physically-closed state, so that the fewest
3064 : * possible number of kernel file descriptors are in use. There is no
3065 : * change in the logical state of the VFDs.
3066 : */
3067 : void
3068 43 : closeAllVfds(void)
3069 : {
3070 : Index i;
3071 :
3072 43 : if (SizeVfdCache > 0)
3073 : {
3074 : Assert(FileIsNotOpen(0)); /* Make sure ring not corrupted */
3075 1376 : for (i = 1; i < SizeVfdCache; i++)
3076 : {
3077 1333 : if (!FileIsNotOpen(i))
3078 148 : LruDelete(i);
3079 : }
3080 : }
3081 43 : }
3082 :
3083 :
3084 : /*
3085 : * SetTempTablespaces
3086 : *
3087 : * Define a list (actually an array) of OIDs of tablespaces to use for
3088 : * temporary files. This list will be used until end of transaction,
3089 : * unless this function is called again before then. It is caller's
3090 : * responsibility that the passed-in array has adequate lifespan (typically
3091 : * it'd be allocated in TopTransactionContext).
3092 : *
3093 : * Some entries of the array may be InvalidOid, indicating that the current
3094 : * database's default tablespace should be used.
3095 : */
3096 : void
3097 4078 : SetTempTablespaces(Oid *tableSpaces, int numSpaces)
3098 : {
3099 : Assert(numSpaces >= 0);
3100 4078 : tempTableSpaces = tableSpaces;
3101 4078 : numTempTableSpaces = numSpaces;
3102 :
3103 : /*
3104 : * Select a random starting point in the list. This is to minimize
3105 : * conflicts between backends that are most likely sharing the same list
3106 : * of temp tablespaces. Note that if we create multiple temp files in the
3107 : * same transaction, we'll advance circularly through the list --- this
3108 : * ensures that large temporary sort files are nicely spread across all
3109 : * available tablespaces.
3110 : */
3111 4078 : if (numSpaces > 1)
3112 0 : nextTempTableSpace = pg_prng_uint64_range(&pg_global_prng_state,
3113 0 : 0, numSpaces - 1);
3114 : else
3115 4078 : nextTempTableSpace = 0;
3116 4078 : }
3117 :
3118 : /*
3119 : * TempTablespacesAreSet
3120 : *
3121 : * Returns true if SetTempTablespaces has been called in current transaction.
3122 : * (This is just so that tablespaces.c doesn't need its own per-transaction
3123 : * state.)
3124 : */
3125 : bool
3126 6114 : TempTablespacesAreSet(void)
3127 : {
3128 6114 : return (numTempTableSpaces >= 0);
3129 : }
3130 :
3131 : /*
3132 : * GetTempTablespaces
3133 : *
3134 : * Populate an array with the OIDs of the tablespaces that should be used for
3135 : * temporary files. (Some entries may be InvalidOid, indicating that the
3136 : * current database's default tablespace should be used.) At most numSpaces
3137 : * entries will be filled.
3138 : * Returns the number of OIDs that were copied into the output array.
3139 : */
3140 : int
3141 277 : GetTempTablespaces(Oid *tableSpaces, int numSpaces)
3142 : {
3143 : int i;
3144 :
3145 : Assert(TempTablespacesAreSet());
3146 277 : for (i = 0; i < numTempTableSpaces && i < numSpaces; ++i)
3147 0 : tableSpaces[i] = tempTableSpaces[i];
3148 :
3149 277 : return i;
3150 : }
3151 :
3152 : /*
3153 : * GetNextTempTableSpace
3154 : *
3155 : * Select the next temp tablespace to use. A result of InvalidOid means
3156 : * to use the current database's default tablespace.
3157 : */
3158 : Oid
3159 3023 : GetNextTempTableSpace(void)
3160 : {
3161 3023 : if (numTempTableSpaces > 0)
3162 : {
3163 : /* Advance nextTempTableSpace counter with wraparound */
3164 1 : if (++nextTempTableSpace >= numTempTableSpaces)
3165 1 : nextTempTableSpace = 0;
3166 1 : return tempTableSpaces[nextTempTableSpace];
3167 : }
3168 3022 : return InvalidOid;
3169 : }
3170 :
3171 :
3172 : /*
3173 : * AtEOSubXact_Files
3174 : *
3175 : * Take care of subtransaction commit/abort. At abort, we close AllocateDescs
3176 : * that the subtransaction may have opened. At commit, we reassign them to
3177 : * the parent subtransaction. (Temporary files are tracked by ResourceOwners
3178 : * instead.)
3179 : */
3180 : void
3181 11803 : AtEOSubXact_Files(bool isCommit, SubTransactionId mySubid,
3182 : SubTransactionId parentSubid)
3183 : {
3184 : Index i;
3185 :
3186 11803 : for (i = 0; i < numAllocatedDescs; i++)
3187 : {
3188 0 : if (allocatedDescs[i].create_subid == mySubid)
3189 : {
3190 0 : if (isCommit)
3191 0 : allocatedDescs[i].create_subid = parentSubid;
3192 : else
3193 : {
3194 : /* have to recheck the item after FreeDesc (ugly) */
3195 0 : FreeDesc(&allocatedDescs[i--]);
3196 : }
3197 : }
3198 : }
3199 11803 : }
3200 :
3201 : /*
3202 : * AtEOXact_Files
3203 : *
3204 : * This routine is called during transaction commit or abort. All still-open
3205 : * per-transaction temporary file VFDs are closed, which also causes the
3206 : * underlying files to be deleted (although they should've been closed already
3207 : * by the ResourceOwner cleanup). Furthermore, all "allocated" stdio files are
3208 : * closed. We also forget any transaction-local temp tablespace list.
3209 : *
3210 : * The isCommit flag is used only to decide whether to emit warnings about
3211 : * unclosed files.
3212 : */
3213 : void
3214 622864 : AtEOXact_Files(bool isCommit)
3215 : {
3216 622864 : CleanupTempFiles(isCommit, false);
3217 622864 : tempTableSpaces = NULL;
3218 622864 : numTempTableSpaces = -1;
3219 622864 : }
3220 :
3221 : /*
3222 : * BeforeShmemExit_Files
3223 : *
3224 : * before_shmem_exit hook to clean up temp files during backend shutdown.
3225 : * Here, we want to clean up *all* temp files including interXact ones.
3226 : */
3227 : static void
3228 24252 : BeforeShmemExit_Files(int code, Datum arg)
3229 : {
3230 24252 : CleanupTempFiles(false, true);
3231 :
3232 : /* prevent further temp files from being created */
3233 : #ifdef USE_ASSERT_CHECKING
3234 : temporary_files_allowed = false;
3235 : #endif
3236 24252 : }
3237 :
3238 : /*
3239 : * Close temporary files and delete their underlying files.
3240 : *
3241 : * isCommit: if true, this is normal transaction commit, and we don't
3242 : * expect any remaining files; warn if there are some.
3243 : *
3244 : * isProcExit: if true, this is being called as the backend process is
3245 : * exiting. If that's the case, we should remove all temporary files; if
3246 : * that's not the case, we are being called for transaction commit/abort
3247 : * and should only remove transaction-local temp files. In either case,
3248 : * also clean up "allocated" stdio files, dirs and fds.
3249 : */
3250 : static void
3251 647116 : CleanupTempFiles(bool isCommit, bool isProcExit)
3252 : {
3253 : Index i;
3254 :
3255 : /*
3256 : * Careful here: at proc_exit we need extra cleanup, not just
3257 : * xact_temporary files.
3258 : */
3259 647116 : if (isProcExit || have_xact_temporary_files)
3260 : {
3261 : Assert(FileIsNotOpen(0)); /* Make sure ring not corrupted */
3262 1575228 : for (i = 1; i < SizeVfdCache; i++)
3263 : {
3264 1550016 : unsigned short fdstate = VfdCache[i].fdstate;
3265 :
3266 1550016 : if (((fdstate & FD_DELETE_AT_CLOSE) || (fdstate & FD_CLOSE_AT_EOXACT)) &&
3267 4 : VfdCache[i].fileName != NULL)
3268 : {
3269 : /*
3270 : * If we're in the process of exiting a backend process, close
3271 : * all temporary files. Otherwise, only close temporary files
3272 : * local to the current transaction. They should be closed by
3273 : * the ResourceOwner mechanism already, so this is just a
3274 : * debugging cross-check.
3275 : */
3276 4 : if (isProcExit)
3277 4 : FileClose(i);
3278 0 : else if (fdstate & FD_CLOSE_AT_EOXACT)
3279 : {
3280 0 : elog(WARNING,
3281 : "temporary file %s not closed at end-of-transaction",
3282 : VfdCache[i].fileName);
3283 0 : FileClose(i);
3284 : }
3285 : }
3286 : }
3287 :
3288 25212 : have_xact_temporary_files = false;
3289 : }
3290 :
3291 : /* Complain if any allocated files remain open at commit. */
3292 647116 : if (isCommit && numAllocatedDescs > 0)
3293 0 : elog(WARNING, "%d temporary files and directories not closed at end-of-transaction",
3294 : numAllocatedDescs);
3295 :
3296 : /* Clean up "allocated" stdio files, dirs and fds. */
3297 647270 : while (numAllocatedDescs > 0)
3298 154 : FreeDesc(&allocatedDescs[0]);
3299 647116 : }
3300 :
3301 :
3302 : /*
3303 : * Remove temporary and temporary relation files left over from a prior
3304 : * postmaster session
3305 : *
3306 : * This should be called during postmaster startup. It will forcibly
3307 : * remove any leftover files created by OpenTemporaryFile and any leftover
3308 : * temporary relation files created by mdcreate.
3309 : *
3310 : * During post-backend-crash restart cycle, this routine is called when
3311 : * remove_temp_files_after_crash GUC is enabled. Multiple crashes while
3312 : * queries are using temp files could result in useless storage usage that can
3313 : * only be reclaimed by a service restart. The argument against enabling it is
3314 : * that someone might want to examine the temporary files for debugging
3315 : * purposes. This does however mean that OpenTemporaryFile had better allow for
3316 : * collision with an existing temp file name.
3317 : *
3318 : * NOTE: this function and its subroutines generally report syscall failures
3319 : * with ereport(LOG) and keep going. Removing temp files is not so critical
3320 : * that we should fail to start the database when we can't do it.
3321 : */
3322 : void
3323 956 : RemovePgTempFiles(void)
3324 : {
3325 : char temp_path[MAXPGPATH + sizeof(PG_TBLSPC_DIR) + sizeof(TABLESPACE_VERSION_DIRECTORY) + sizeof(PG_TEMP_FILES_DIR)];
3326 : DIR *spc_dir;
3327 : struct dirent *spc_de;
3328 :
3329 : /*
3330 : * First process temp files in pg_default ($PGDATA/base)
3331 : */
3332 956 : snprintf(temp_path, sizeof(temp_path), "base/%s", PG_TEMP_FILES_DIR);
3333 956 : RemovePgTempFilesInDir(temp_path, true, false);
3334 956 : RemovePgTempRelationFiles("base");
3335 :
3336 : /*
3337 : * Cycle through temp directories for all non-default tablespaces.
3338 : */
3339 956 : spc_dir = AllocateDir(PG_TBLSPC_DIR);
3340 :
3341 2941 : while ((spc_de = ReadDirExtended(spc_dir, PG_TBLSPC_DIR, LOG)) != NULL)
3342 : {
3343 1985 : if (strcmp(spc_de->d_name, ".") == 0 ||
3344 1029 : strcmp(spc_de->d_name, "..") == 0)
3345 1912 : continue;
3346 :
3347 73 : snprintf(temp_path, sizeof(temp_path), "%s/%s/%s/%s",
3348 73 : PG_TBLSPC_DIR, spc_de->d_name, TABLESPACE_VERSION_DIRECTORY,
3349 : PG_TEMP_FILES_DIR);
3350 73 : RemovePgTempFilesInDir(temp_path, true, false);
3351 :
3352 73 : snprintf(temp_path, sizeof(temp_path), "%s/%s/%s",
3353 73 : PG_TBLSPC_DIR, spc_de->d_name, TABLESPACE_VERSION_DIRECTORY);
3354 73 : RemovePgTempRelationFiles(temp_path);
3355 : }
3356 :
3357 956 : FreeDir(spc_dir);
3358 :
3359 : /*
3360 : * In EXEC_BACKEND case there is a pgsql_tmp directory at the top level of
3361 : * DataDir as well. However, that is *not* cleaned here because doing so
3362 : * would create a race condition. It's done separately, earlier in
3363 : * postmaster startup.
3364 : */
3365 956 : }
3366 :
3367 : /*
3368 : * Process one pgsql_tmp directory for RemovePgTempFiles.
3369 : *
3370 : * If missing_ok is true, it's all right for the named directory to not exist.
3371 : * Any other problem results in a LOG message. (missing_ok should be true at
3372 : * the top level, since pgsql_tmp directories are not created until needed.)
3373 : *
3374 : * At the top level, this should be called with unlink_all = false, so that
3375 : * only files matching the temporary name prefix will be unlinked. When
3376 : * recursing it will be called with unlink_all = true to unlink everything
3377 : * under a top-level temporary directory.
3378 : *
3379 : * (These two flags could be replaced by one, but it seems clearer to keep
3380 : * them separate.)
3381 : */
3382 : void
3383 1030 : RemovePgTempFilesInDir(const char *tmpdirname, bool missing_ok, bool unlink_all)
3384 : {
3385 : DIR *temp_dir;
3386 : struct dirent *temp_de;
3387 : char rm_path[MAXPGPATH * 2];
3388 :
3389 1030 : temp_dir = AllocateDir(tmpdirname);
3390 :
3391 1030 : if (temp_dir == NULL && errno == ENOENT && missing_ok)
3392 958 : return;
3393 :
3394 219 : while ((temp_de = ReadDirExtended(temp_dir, tmpdirname, LOG)) != NULL)
3395 : {
3396 147 : if (strcmp(temp_de->d_name, ".") == 0 ||
3397 75 : strcmp(temp_de->d_name, "..") == 0)
3398 144 : continue;
3399 :
3400 3 : snprintf(rm_path, sizeof(rm_path), "%s/%s",
3401 3 : tmpdirname, temp_de->d_name);
3402 :
3403 3 : if (unlink_all ||
3404 3 : strncmp(temp_de->d_name,
3405 : PG_TEMP_FILE_PREFIX,
3406 : strlen(PG_TEMP_FILE_PREFIX)) == 0)
3407 3 : {
3408 3 : PGFileType type = get_dirent_type(rm_path, temp_de, false, LOG);
3409 :
3410 3 : if (type == PGFILETYPE_ERROR)
3411 0 : continue;
3412 3 : else if (type == PGFILETYPE_DIR)
3413 : {
3414 : /* recursively remove contents, then directory itself */
3415 1 : RemovePgTempFilesInDir(rm_path, false, true);
3416 :
3417 1 : if (rmdir(rm_path) < 0)
3418 0 : ereport(LOG,
3419 : (errcode_for_file_access(),
3420 : errmsg("could not remove directory \"%s\": %m",
3421 : rm_path)));
3422 : }
3423 : else
3424 : {
3425 2 : if (unlink(rm_path) < 0)
3426 0 : ereport(LOG,
3427 : (errcode_for_file_access(),
3428 : errmsg("could not remove file \"%s\": %m",
3429 : rm_path)));
3430 : }
3431 : }
3432 : else
3433 0 : ereport(LOG,
3434 : (errmsg("unexpected file found in temporary-files directory: \"%s\"",
3435 : rm_path)));
3436 : }
3437 :
3438 72 : FreeDir(temp_dir);
3439 : }
3440 :
3441 : /* Process one tablespace directory, look for per-DB subdirectories */
3442 : static void
3443 1029 : RemovePgTempRelationFiles(const char *tsdirname)
3444 : {
3445 : DIR *ts_dir;
3446 : struct dirent *de;
3447 : char dbspace_path[MAXPGPATH * 2];
3448 :
3449 1029 : ts_dir = AllocateDir(tsdirname);
3450 :
3451 6459 : while ((de = ReadDirExtended(ts_dir, tsdirname, LOG)) != NULL)
3452 : {
3453 : /*
3454 : * We're only interested in the per-database directories, which have
3455 : * numeric names. Note that this code will also (properly) ignore "."
3456 : * and "..".
3457 : */
3458 5430 : if (strspn(de->d_name, "0123456789") != strlen(de->d_name))
3459 2129 : continue;
3460 :
3461 3301 : snprintf(dbspace_path, sizeof(dbspace_path), "%s/%s",
3462 3301 : tsdirname, de->d_name);
3463 3301 : RemovePgTempRelationFilesInDbspace(dbspace_path);
3464 : }
3465 :
3466 1029 : FreeDir(ts_dir);
3467 1029 : }
3468 :
3469 : /* Process one per-dbspace directory for RemovePgTempRelationFiles */
3470 : static void
3471 3301 : RemovePgTempRelationFilesInDbspace(const char *dbspacedirname)
3472 : {
3473 : DIR *dbspace_dir;
3474 : struct dirent *de;
3475 : char rm_path[MAXPGPATH * 2];
3476 :
3477 3301 : dbspace_dir = AllocateDir(dbspacedirname);
3478 :
3479 1066360 : while ((de = ReadDirExtended(dbspace_dir, dbspacedirname, LOG)) != NULL)
3480 : {
3481 1063059 : if (!looks_like_temp_rel_name(de->d_name))
3482 1063055 : continue;
3483 :
3484 4 : snprintf(rm_path, sizeof(rm_path), "%s/%s",
3485 4 : dbspacedirname, de->d_name);
3486 :
3487 4 : if (unlink(rm_path) < 0)
3488 0 : ereport(LOG,
3489 : (errcode_for_file_access(),
3490 : errmsg("could not remove file \"%s\": %m",
3491 : rm_path)));
3492 : }
3493 :
3494 3301 : FreeDir(dbspace_dir);
3495 3301 : }
3496 :
3497 : /* t<digits>_<digits>, or t<digits>_<digits>_<forkname> */
3498 : bool
3499 1400601 : looks_like_temp_rel_name(const char *name)
3500 : {
3501 : int pos;
3502 : int savepos;
3503 :
3504 : /* Must start with "t". */
3505 1400601 : if (name[0] != 't')
3506 1400561 : return false;
3507 :
3508 : /* Followed by a non-empty string of digits and then an underscore. */
3509 196 : for (pos = 1; isdigit((unsigned char) name[pos]); ++pos)
3510 : ;
3511 40 : if (pos == 1 || name[pos] != '_')
3512 0 : return false;
3513 :
3514 : /* Followed by another nonempty string of digits. */
3515 196 : for (savepos = ++pos; isdigit((unsigned char) name[pos]); ++pos)
3516 : ;
3517 40 : if (savepos == pos)
3518 0 : return false;
3519 :
3520 : /* We might have _forkname or .segment or both. */
3521 40 : if (name[pos] == '_')
3522 : {
3523 20 : int forkchar = forkname_chars(&name[pos + 1], NULL);
3524 :
3525 20 : if (forkchar <= 0)
3526 0 : return false;
3527 20 : pos += forkchar + 1;
3528 : }
3529 40 : if (name[pos] == '.')
3530 : {
3531 : int segchar;
3532 :
3533 40 : for (segchar = 1; isdigit((unsigned char) name[pos + segchar]); ++segchar)
3534 : ;
3535 20 : if (segchar <= 1)
3536 0 : return false;
3537 20 : pos += segchar;
3538 : }
3539 :
3540 : /* Now we should be at the end. */
3541 40 : if (name[pos] != '\0')
3542 0 : return false;
3543 40 : return true;
3544 : }
3545 :
3546 : #ifdef HAVE_SYNCFS
3547 : static void
3548 0 : do_syncfs(const char *path)
3549 : {
3550 : int fd;
3551 :
3552 0 : ereport_startup_progress("syncing data directory (syncfs), elapsed time: %ld.%02d s, current path: %s",
3553 : path);
3554 :
3555 0 : fd = OpenTransientFile(path, O_RDONLY);
3556 0 : if (fd < 0)
3557 : {
3558 0 : ereport(LOG,
3559 : (errcode_for_file_access(),
3560 : errmsg("could not open file \"%s\": %m", path)));
3561 0 : return;
3562 : }
3563 0 : if (syncfs(fd) < 0)
3564 0 : ereport(LOG,
3565 : (errcode_for_file_access(),
3566 : errmsg("could not synchronize file system for file \"%s\": %m", path)));
3567 0 : CloseTransientFile(fd);
3568 : }
3569 : #endif
3570 :
3571 : /*
3572 : * Issue fsync recursively on PGDATA and all its contents, or issue syncfs for
3573 : * all potential filesystem, depending on recovery_init_sync_method setting.
3574 : *
3575 : * We fsync regular files and directories wherever they are, but we
3576 : * follow symlinks only for pg_wal and immediately under pg_tblspc.
3577 : * Other symlinks are presumed to point at files we're not responsible
3578 : * for fsyncing, and might not have privileges to write at all.
3579 : *
3580 : * Errors are logged but not considered fatal; that's because this is used
3581 : * only during database startup, to deal with the possibility that there are
3582 : * issued-but-unsynced writes pending against the data directory. We want to
3583 : * ensure that such writes reach disk before anything that's done in the new
3584 : * run. However, aborting on error would result in failure to start for
3585 : * harmless cases such as read-only files in the data directory, and that's
3586 : * not good either.
3587 : *
3588 : * Note that if we previously crashed due to a PANIC on fsync(), we'll be
3589 : * rewriting all changes again during recovery.
3590 : *
3591 : * Note we assume we're chdir'd into PGDATA to begin with.
3592 : */
3593 : void
3594 186 : SyncDataDirectory(void)
3595 : {
3596 : bool xlog_is_symlink;
3597 :
3598 : /* We can skip this whole thing if fsync is disabled. */
3599 186 : if (!enableFsync)
3600 186 : return;
3601 :
3602 : /*
3603 : * If pg_wal is a symlink, we'll need to recurse into it separately,
3604 : * because the first walkdir below will ignore it.
3605 : */
3606 0 : xlog_is_symlink = false;
3607 :
3608 : {
3609 : struct stat st;
3610 :
3611 0 : if (lstat("pg_wal", &st) < 0)
3612 0 : ereport(LOG,
3613 : (errcode_for_file_access(),
3614 : errmsg("could not stat file \"%s\": %m",
3615 : "pg_wal")));
3616 0 : else if (S_ISLNK(st.st_mode))
3617 0 : xlog_is_symlink = true;
3618 : }
3619 :
3620 : #ifdef HAVE_SYNCFS
3621 0 : if (recovery_init_sync_method == DATA_DIR_SYNC_METHOD_SYNCFS)
3622 : {
3623 : DIR *dir;
3624 : struct dirent *de;
3625 :
3626 : /*
3627 : * On Linux, we don't have to open every single file one by one. We
3628 : * can use syncfs() to sync whole filesystems. We only expect
3629 : * filesystem boundaries to exist where we tolerate symlinks, namely
3630 : * pg_wal and the tablespaces, so we call syncfs() for each of those
3631 : * directories.
3632 : */
3633 :
3634 : /* Prepare to report progress syncing the data directory via syncfs. */
3635 0 : begin_startup_progress_phase();
3636 :
3637 : /* Sync the top level pgdata directory. */
3638 0 : do_syncfs(".");
3639 : /* If any tablespaces are configured, sync each of those. */
3640 0 : dir = AllocateDir(PG_TBLSPC_DIR);
3641 0 : while ((de = ReadDirExtended(dir, PG_TBLSPC_DIR, LOG)))
3642 : {
3643 : char path[MAXPGPATH];
3644 :
3645 0 : if (strcmp(de->d_name, ".") == 0 || strcmp(de->d_name, "..") == 0)
3646 0 : continue;
3647 :
3648 0 : snprintf(path, MAXPGPATH, "%s/%s", PG_TBLSPC_DIR, de->d_name);
3649 0 : do_syncfs(path);
3650 : }
3651 0 : FreeDir(dir);
3652 : /* If pg_wal is a symlink, process that too. */
3653 0 : if (xlog_is_symlink)
3654 0 : do_syncfs("pg_wal");
3655 0 : return;
3656 : }
3657 : #endif /* !HAVE_SYNCFS */
3658 :
3659 : #ifdef PG_FLUSH_DATA_WORKS
3660 : /* Prepare to report progress of the pre-fsync phase. */
3661 0 : begin_startup_progress_phase();
3662 :
3663 : /*
3664 : * If possible, hint to the kernel that we're soon going to fsync the data
3665 : * directory and its contents. Errors in this step are even less
3666 : * interesting than normal, so log them only at DEBUG1.
3667 : */
3668 0 : walkdir(".", pre_sync_fname, false, DEBUG1);
3669 0 : if (xlog_is_symlink)
3670 0 : walkdir("pg_wal", pre_sync_fname, false, DEBUG1);
3671 0 : walkdir(PG_TBLSPC_DIR, pre_sync_fname, true, DEBUG1);
3672 : #endif
3673 :
3674 : /* Prepare to report progress syncing the data directory via fsync. */
3675 0 : begin_startup_progress_phase();
3676 :
3677 : /*
3678 : * Now we do the fsync()s in the same order.
3679 : *
3680 : * The main call ignores symlinks, so in addition to specially processing
3681 : * pg_wal if it's a symlink, pg_tblspc has to be visited separately with
3682 : * process_symlinks = true. Note that if there are any plain directories
3683 : * in pg_tblspc, they'll get fsync'd twice. That's not an expected case
3684 : * so we don't worry about optimizing it.
3685 : */
3686 0 : walkdir(".", datadir_fsync_fname, false, LOG);
3687 0 : if (xlog_is_symlink)
3688 0 : walkdir("pg_wal", datadir_fsync_fname, false, LOG);
3689 0 : walkdir(PG_TBLSPC_DIR, datadir_fsync_fname, true, LOG);
3690 : }
3691 :
3692 : /*
3693 : * walkdir: recursively walk a directory, applying the action to each
3694 : * regular file and directory (including the named directory itself).
3695 : *
3696 : * If process_symlinks is true, the action and recursion are also applied
3697 : * to regular files and directories that are pointed to by symlinks in the
3698 : * given directory; otherwise symlinks are ignored. Symlinks are always
3699 : * ignored in subdirectories, ie we intentionally don't pass down the
3700 : * process_symlinks flag to recursive calls.
3701 : *
3702 : * Errors are reported at level elevel, which might be ERROR or less.
3703 : *
3704 : * See also walkdir in file_utils.c, which is a frontend version of this
3705 : * logic.
3706 : */
3707 : static void
3708 252 : walkdir(const char *path,
3709 : void (*action) (const char *fname, bool isdir, int elevel),
3710 : bool process_symlinks,
3711 : int elevel)
3712 : {
3713 : DIR *dir;
3714 : struct dirent *de;
3715 :
3716 252 : dir = AllocateDir(path);
3717 :
3718 2572 : while ((de = ReadDirExtended(dir, path, elevel)) != NULL)
3719 : {
3720 : char subpath[MAXPGPATH * 2];
3721 :
3722 2320 : CHECK_FOR_INTERRUPTS();
3723 :
3724 2320 : if (strcmp(de->d_name, ".") == 0 ||
3725 2068 : strcmp(de->d_name, "..") == 0)
3726 504 : continue;
3727 :
3728 1816 : snprintf(subpath, sizeof(subpath), "%s/%s", path, de->d_name);
3729 :
3730 1816 : switch (get_dirent_type(subpath, de, process_symlinks, elevel))
3731 : {
3732 1816 : case PGFILETYPE_REG:
3733 1816 : (*action) (subpath, false, elevel);
3734 1816 : break;
3735 0 : case PGFILETYPE_DIR:
3736 0 : walkdir(subpath, action, false, elevel);
3737 0 : break;
3738 0 : default:
3739 :
3740 : /*
3741 : * Errors are already reported directly by get_dirent_type(),
3742 : * and any remaining symlinks and unknown file types are
3743 : * ignored.
3744 : */
3745 0 : break;
3746 : }
3747 : }
3748 :
3749 252 : FreeDir(dir); /* we ignore any error here */
3750 :
3751 : /*
3752 : * It's important to fsync the destination directory itself as individual
3753 : * file fsyncs don't guarantee that the directory entry for the file is
3754 : * synced. However, skip this if AllocateDir failed; the action function
3755 : * might not be robust against that.
3756 : */
3757 252 : if (dir)
3758 252 : (*action) (path, true, elevel);
3759 252 : }
3760 :
3761 :
3762 : /*
3763 : * Hint to the OS that it should get ready to fsync() this file.
3764 : *
3765 : * Ignores errors trying to open unreadable files, and logs other errors at a
3766 : * caller-specified level.
3767 : */
3768 : #ifdef PG_FLUSH_DATA_WORKS
3769 :
3770 : static void
3771 0 : pre_sync_fname(const char *fname, bool isdir, int elevel)
3772 : {
3773 : int fd;
3774 :
3775 : /* Don't try to flush directories, it'll likely just fail */
3776 0 : if (isdir)
3777 0 : return;
3778 :
3779 0 : ereport_startup_progress("syncing data directory (pre-fsync), elapsed time: %ld.%02d s, current path: %s",
3780 : fname);
3781 :
3782 0 : fd = OpenTransientFile(fname, O_RDONLY | PG_BINARY);
3783 :
3784 0 : if (fd < 0)
3785 : {
3786 0 : if (errno == EACCES)
3787 0 : return;
3788 0 : ereport(elevel,
3789 : (errcode_for_file_access(),
3790 : errmsg("could not open file \"%s\": %m", fname)));
3791 0 : return;
3792 : }
3793 :
3794 : /*
3795 : * pg_flush_data() ignores errors, which is ok because this is only a
3796 : * hint.
3797 : */
3798 0 : pg_flush_data(fd, 0, 0);
3799 :
3800 0 : if (CloseTransientFile(fd) != 0)
3801 0 : ereport(elevel,
3802 : (errcode_for_file_access(),
3803 : errmsg("could not close file \"%s\": %m", fname)));
3804 : }
3805 :
3806 : #endif /* PG_FLUSH_DATA_WORKS */
3807 :
3808 : static void
3809 0 : datadir_fsync_fname(const char *fname, bool isdir, int elevel)
3810 : {
3811 0 : ereport_startup_progress("syncing data directory (fsync), elapsed time: %ld.%02d s, current path: %s",
3812 : fname);
3813 :
3814 : /*
3815 : * We want to silently ignoring errors about unreadable files. Pass that
3816 : * desire on to fsync_fname_ext().
3817 : */
3818 0 : fsync_fname_ext(fname, isdir, true, elevel);
3819 0 : }
3820 :
3821 : static void
3822 2068 : unlink_if_exists_fname(const char *fname, bool isdir, int elevel)
3823 : {
3824 2068 : if (isdir)
3825 : {
3826 252 : if (rmdir(fname) != 0 && errno != ENOENT)
3827 0 : ereport(elevel,
3828 : (errcode_for_file_access(),
3829 : errmsg("could not remove directory \"%s\": %m", fname)));
3830 : }
3831 : else
3832 : {
3833 : /* Use PathNameDeleteTemporaryFile to report filesize */
3834 1816 : PathNameDeleteTemporaryFile(fname, false);
3835 : }
3836 2068 : }
3837 :
3838 : /*
3839 : * fsync_fname_ext -- Try to fsync a file or directory
3840 : *
3841 : * If ignore_perm is true, ignore errors upon trying to open unreadable
3842 : * files. Logs other errors at a caller-specified level.
3843 : *
3844 : * Returns 0 if the operation succeeded, -1 otherwise.
3845 : */
3846 : int
3847 44396 : fsync_fname_ext(const char *fname, bool isdir, bool ignore_perm, int elevel)
3848 : {
3849 : int fd;
3850 : int flags;
3851 : int returncode;
3852 :
3853 : /*
3854 : * Some OSs require directories to be opened read-only whereas other
3855 : * systems don't allow us to fsync files opened read-only; so we need both
3856 : * cases here. Using O_RDWR will cause us to fail to fsync files that are
3857 : * not writable by our userid, but we assume that's OK.
3858 : */
3859 44396 : flags = PG_BINARY;
3860 44396 : if (!isdir)
3861 16514 : flags |= O_RDWR;
3862 : else
3863 27882 : flags |= O_RDONLY;
3864 :
3865 44396 : fd = OpenTransientFile(fname, flags);
3866 :
3867 : /*
3868 : * Some OSs don't allow us to open directories at all (Windows returns
3869 : * EACCES), just ignore the error in that case. If desired also silently
3870 : * ignoring errors about unreadable files. Log others.
3871 : */
3872 44396 : if (fd < 0 && isdir && (errno == EISDIR || errno == EACCES))
3873 0 : return 0;
3874 44396 : else if (fd < 0 && ignore_perm && errno == EACCES)
3875 0 : return 0;
3876 44396 : else if (fd < 0)
3877 : {
3878 0 : ereport(elevel,
3879 : (errcode_for_file_access(),
3880 : errmsg("could not open file \"%s\": %m", fname)));
3881 0 : return -1;
3882 : }
3883 :
3884 44396 : returncode = pg_fsync(fd);
3885 :
3886 : /*
3887 : * Some OSes don't allow us to fsync directories at all, so we can ignore
3888 : * those errors. Anything else needs to be logged.
3889 : */
3890 44396 : if (returncode != 0 && !(isdir && (errno == EBADF || errno == EINVAL)))
3891 : {
3892 : int save_errno;
3893 :
3894 : /* close file upon error, might not be in transaction context */
3895 0 : save_errno = errno;
3896 0 : (void) CloseTransientFile(fd);
3897 0 : errno = save_errno;
3898 :
3899 0 : ereport(elevel,
3900 : (errcode_for_file_access(),
3901 : errmsg("could not fsync file \"%s\": %m", fname)));
3902 0 : return -1;
3903 : }
3904 :
3905 44396 : if (CloseTransientFile(fd) != 0)
3906 : {
3907 0 : ereport(elevel,
3908 : (errcode_for_file_access(),
3909 : errmsg("could not close file \"%s\": %m", fname)));
3910 0 : return -1;
3911 : }
3912 :
3913 44396 : return 0;
3914 : }
3915 :
3916 : /*
3917 : * fsync_parent_path -- fsync the parent path of a file or directory
3918 : *
3919 : * This is aimed at making file operations persistent on disk in case of
3920 : * an OS crash or power failure.
3921 : */
3922 : static int
3923 8036 : fsync_parent_path(const char *fname, int elevel)
3924 : {
3925 : char parentpath[MAXPGPATH];
3926 :
3927 8036 : strlcpy(parentpath, fname, MAXPGPATH);
3928 8036 : get_parent_directory(parentpath);
3929 :
3930 : /*
3931 : * get_parent_directory() returns an empty string if the input argument is
3932 : * just a file name (see comments in path.c), so handle that as being the
3933 : * current directory.
3934 : */
3935 8036 : if (strlen(parentpath) == 0)
3936 223 : strlcpy(parentpath, ".", MAXPGPATH);
3937 :
3938 8036 : if (fsync_fname_ext(parentpath, true, false, elevel) != 0)
3939 0 : return -1;
3940 :
3941 8036 : return 0;
3942 : }
3943 :
3944 : /*
3945 : * Create a PostgreSQL data sub-directory
3946 : *
3947 : * The data directory itself, and most of its sub-directories, are created at
3948 : * initdb time, but we do have some occasions when we create directories in
3949 : * the backend (CREATE TABLESPACE, for example). In those cases, we want to
3950 : * make sure that those directories are created consistently. Today, that means
3951 : * making sure that the created directory has the correct permissions, which is
3952 : * what pg_dir_create_mode tracks for us.
3953 : *
3954 : * Note that we also set the umask() based on what we understand the correct
3955 : * permissions to be (see file_perm.c).
3956 : *
3957 : * For permissions other than the default, mkdir() can be used directly, but
3958 : * be sure to consider carefully such cases -- a sub-directory with incorrect
3959 : * permissions in a PostgreSQL data directory could cause backups and other
3960 : * processes to fail.
3961 : */
3962 : int
3963 1690 : MakePGDirectory(const char *directoryName)
3964 : {
3965 1690 : return mkdir(directoryName, pg_dir_create_mode);
3966 : }
3967 :
3968 : /*
3969 : * Return the passed-in error level, or PANIC if data_sync_retry is off.
3970 : *
3971 : * Failure to fsync any data file is cause for immediate panic, unless
3972 : * data_sync_retry is enabled. Data may have been written to the operating
3973 : * system and removed from our buffer pool already, and if we are running on
3974 : * an operating system that forgets dirty data on write-back failure, there
3975 : * may be only one copy of the data remaining: in the WAL. A later attempt to
3976 : * fsync again might falsely report success. Therefore we must not allow any
3977 : * further checkpoints to be attempted. data_sync_retry can in theory be
3978 : * enabled on systems known not to drop dirty buffered data on write-back
3979 : * failure (with the likely outcome that checkpoints will continue to fail
3980 : * until the underlying problem is fixed).
3981 : *
3982 : * Any code that reports a failure from fsync() or related functions should
3983 : * filter the error level with this function.
3984 : */
3985 : int
3986 22892 : data_sync_elevel(int elevel)
3987 : {
3988 22892 : return data_sync_retry ? elevel : PANIC;
3989 : }
3990 :
3991 : bool
3992 1224 : check_debug_io_direct(char **newval, void **extra, GucSource source)
3993 : {
3994 1224 : bool result = true;
3995 : int flags;
3996 :
3997 : #if PG_O_DIRECT == 0
3998 : if (strcmp(*newval, "") != 0)
3999 : {
4000 : GUC_check_errdetail("\"%s\" is not supported on this platform.",
4001 : "debug_io_direct");
4002 : result = false;
4003 : }
4004 : flags = 0;
4005 : #else
4006 : List *elemlist;
4007 : ListCell *l;
4008 : char *rawstring;
4009 :
4010 : /* Need a modifiable copy of string */
4011 1224 : rawstring = pstrdup(*newval);
4012 :
4013 1224 : if (!SplitGUCList(rawstring, ',', &elemlist))
4014 : {
4015 0 : GUC_check_errdetail("Invalid list syntax in parameter \"%s\".",
4016 : "debug_io_direct");
4017 0 : pfree(rawstring);
4018 0 : list_free(elemlist);
4019 0 : return false;
4020 : }
4021 :
4022 1224 : flags = 0;
4023 1230 : foreach(l, elemlist)
4024 : {
4025 6 : char *item = (char *) lfirst(l);
4026 :
4027 6 : if (pg_strcasecmp(item, "data") == 0)
4028 2 : flags |= IO_DIRECT_DATA;
4029 4 : else if (pg_strcasecmp(item, "wal") == 0)
4030 2 : flags |= IO_DIRECT_WAL;
4031 2 : else if (pg_strcasecmp(item, "wal_init") == 0)
4032 2 : flags |= IO_DIRECT_WAL_INIT;
4033 : else
4034 : {
4035 0 : GUC_check_errdetail("Invalid option \"%s\".", item);
4036 0 : result = false;
4037 0 : break;
4038 : }
4039 : }
4040 :
4041 : /*
4042 : * It's possible to configure block sizes smaller than our assumed I/O
4043 : * alignment size, which could result in invalid I/O requests.
4044 : */
4045 : #if XLOG_BLCKSZ < PG_IO_ALIGN_SIZE
4046 : if (result && (flags & (IO_DIRECT_WAL | IO_DIRECT_WAL_INIT)))
4047 : {
4048 : GUC_check_errdetail("\"%s\" is not supported for WAL because %s is too small.",
4049 : "debug_io_direct", "XLOG_BLCKSZ");
4050 : result = false;
4051 : }
4052 : #endif
4053 : #if BLCKSZ < PG_IO_ALIGN_SIZE
4054 : if (result && (flags & IO_DIRECT_DATA))
4055 : {
4056 : GUC_check_errdetail("\"%s\" is not supported for data because %s is too small.",
4057 : "debug_io_direct", "BLCKSZ");
4058 : result = false;
4059 : }
4060 : #endif
4061 :
4062 1224 : pfree(rawstring);
4063 1224 : list_free(elemlist);
4064 : #endif
4065 :
4066 1224 : if (!result)
4067 0 : return result;
4068 :
4069 : /* Save the flags in *extra, for use by assign_debug_io_direct */
4070 1224 : *extra = guc_malloc(LOG, sizeof(int));
4071 1224 : if (!*extra)
4072 0 : return false;
4073 1224 : *((int *) *extra) = flags;
4074 :
4075 1224 : return result;
4076 : }
4077 :
4078 : void
4079 1224 : assign_debug_io_direct(const char *newval, void *extra)
4080 : {
4081 1224 : int *flags = (int *) extra;
4082 :
4083 1224 : io_direct_flags = *flags;
4084 1224 : }
4085 :
4086 : /* ResourceOwner callbacks */
4087 :
4088 : static void
4089 5 : ResOwnerReleaseFile(Datum res)
4090 : {
4091 5 : File file = (File) DatumGetInt32(res);
4092 : Vfd *vfdP;
4093 :
4094 : Assert(FileIsValid(file));
4095 :
4096 5 : vfdP = &VfdCache[file];
4097 5 : vfdP->resowner = NULL;
4098 :
4099 5 : FileClose(file);
4100 5 : }
4101 :
4102 : static char *
4103 0 : ResOwnerPrintFile(Datum res)
4104 : {
4105 0 : return psprintf("File %d", DatumGetInt32(res));
4106 : }
|
