1 /*-------------------------------------------------------------------------
2 *
3 * md.c
4 * This code manages relations that reside on magnetic disk.
5 *
6 * Or at least, that was what the Berkeley folk had in mind when they named
7 * this file. In reality, what this code provides is an interface from
8 * the smgr API to Unix-like filesystem APIs, so it will work with any type
9 * of device for which the operating system provides filesystem support.
10 * It doesn't matter whether the bits are on spinning rust or some other
11 * storage technology.
12 *
13 * Portions Copyright (c) 1996-2019, PostgreSQL Global Development Group
14 * Portions Copyright (c) 1994, Regents of the University of California
15 *
16 *
17 * IDENTIFICATION
18 * src/backend/storage/smgr/md.c
19 *
20 *-------------------------------------------------------------------------
21 */
22 #include "postgres.h"
23
24 #include <unistd.h>
25 #include <fcntl.h>
26 #include <sys/file.h>
27
28 #include "miscadmin.h"
29 #include "access/xlogutils.h"
30 #include "access/xlog.h"
31 #include "pgstat.h"
32 #include "postmaster/bgwriter.h"
33 #include "storage/fd.h"
34 #include "storage/bufmgr.h"
35 #include "storage/md.h"
36 #include "storage/relfilenode.h"
37 #include "storage/smgr.h"
38 #include "storage/sync.h"
39 #include "utils/hsearch.h"
40 #include "utils/memutils.h"
41 #include "pg_trace.h"
42
43 /*
44 * The magnetic disk storage manager keeps track of open file
45 * descriptors in its own descriptor pool. This is done to make it
46 * easier to support relations that are larger than the operating
47 * system's file size limit (often 2GBytes). In order to do that,
48 * we break relations up into "segment" files that are each shorter than
49 * the OS file size limit. The segment size is set by the RELSEG_SIZE
50 * configuration constant in pg_config.h.
51 *
52 * On disk, a relation must consist of consecutively numbered segment
53 * files in the pattern
54 * -- Zero or more full segments of exactly RELSEG_SIZE blocks each
55 * -- Exactly one partial segment of size 0 <= size < RELSEG_SIZE blocks
56 * -- Optionally, any number of inactive segments of size 0 blocks.
57 * The full and partial segments are collectively the "active" segments.
58 * Inactive segments are those that once contained data but are currently
59 * not needed because of an mdtruncate() operation. The reason for leaving
60 * them present at size zero, rather than unlinking them, is that other
61 * backends and/or the checkpointer might be holding open file references to
62 * such segments. If the relation expands again after mdtruncate(), such
63 * that a deactivated segment becomes active again, it is important that
64 * such file references still be valid --- else data might get written
65 * out to an unlinked old copy of a segment file that will eventually
66 * disappear.
67 *
68 * File descriptors are stored in the per-fork md_seg_fds arrays inside
69 * SMgrRelation. The length of these arrays is stored in md_num_open_segs.
70 * Note that a fork's md_num_open_segs having a specific value does not
71 * necessarily mean the relation doesn't have additional segments; we may
72 * just not have opened the next segment yet. (We could not have "all
73 * segments are in the array" as an invariant anyway, since another backend
74 * could extend the relation while we aren't looking.) We do not have
75 * entries for inactive segments, however; as soon as we find a partial
76 * segment, we assume that any subsequent segments are inactive.
77 *
78 * The entire MdfdVec array is palloc'd in the MdCxt memory context.
79 */
80
81 typedef struct _MdfdVec
82 {
83 File mdfd_vfd; /* fd number in fd.c's pool */
84 BlockNumber mdfd_segno; /* segment number, from 0 */
85 } MdfdVec;
86
87 static MemoryContext MdCxt; /* context for all MdfdVec objects */
88
89
90 /* Populate a file tag describing an md.c segment file. */
91 #define INIT_MD_FILETAG(a,xx_rnode,xx_forknum,xx_segno) \
92 ( \
93 memset(&(a), 0, sizeof(FileTag)), \
94 (a).handler = SYNC_HANDLER_MD, \
95 (a).rnode = (xx_rnode), \
96 (a).forknum = (xx_forknum), \
97 (a).segno = (xx_segno) \
98 )
99
100
101 /*** behavior for mdopen & _mdfd_getseg ***/
102 /* ereport if segment not present */
103 #define EXTENSION_FAIL (1 << 0)
104 /* return NULL if segment not present */
105 #define EXTENSION_RETURN_NULL (1 << 1)
106 /* create new segments as needed */
107 #define EXTENSION_CREATE (1 << 2)
108 /* create new segments if needed during recovery */
109 #define EXTENSION_CREATE_RECOVERY (1 << 3)
110 /*
111 * Allow opening segments which are preceded by segments smaller than
112 * RELSEG_SIZE, e.g. inactive segments (see above). Note that this breaks
113 * mdnblocks() and related functionality henceforth - which currently is ok,
114 * because this is only required in the checkpointer which never uses
115 * mdnblocks().
116 */
117 #define EXTENSION_DONT_CHECK_SIZE (1 << 4)
118
119
120 /* local routines */
121 static void mdunlinkfork(RelFileNodeBackend rnode, ForkNumber forkNum,
122 bool isRedo);
123 static MdfdVec *mdopen(SMgrRelation reln, ForkNumber forknum, int behavior);
124 static void register_dirty_segment(SMgrRelation reln, ForkNumber forknum,
125 MdfdVec *seg);
126 static void register_unlink_segment(RelFileNodeBackend rnode, ForkNumber forknum,
127 BlockNumber segno);
128 static void register_forget_request(RelFileNodeBackend rnode, ForkNumber forknum,
129 BlockNumber segno);
130 static void _fdvec_resize(SMgrRelation reln,
131 ForkNumber forknum,
132 int nseg);
133 static char *_mdfd_segpath(SMgrRelation reln, ForkNumber forknum,
134 BlockNumber segno);
135 static MdfdVec *_mdfd_openseg(SMgrRelation reln, ForkNumber forkno,
136 BlockNumber segno, int oflags);
137 static MdfdVec *_mdfd_getseg(SMgrRelation reln, ForkNumber forkno,
138 BlockNumber blkno, bool skipFsync, int behavior);
139 static BlockNumber _mdnblocks(SMgrRelation reln, ForkNumber forknum,
140 MdfdVec *seg);
141
142
143 /*
144 * mdinit() -- Initialize private state for magnetic disk storage manager.
145 */
146 void
mdinit(void)147 mdinit(void)
148 {
149 MdCxt = AllocSetContextCreate(TopMemoryContext,
150 "MdSmgr",
151 ALLOCSET_DEFAULT_SIZES);
152 }
153
154 /*
155 * mdexists() -- Does the physical file exist?
156 *
157 * Note: this will return true for lingering files, with pending deletions
158 */
159 bool
mdexists(SMgrRelation reln,ForkNumber forkNum)160 mdexists(SMgrRelation reln, ForkNumber forkNum)
161 {
162 /*
163 * Close it first, to ensure that we notice if the fork has been unlinked
164 * since we opened it.
165 */
166 mdclose(reln, forkNum);
167
168 return (mdopen(reln, forkNum, EXTENSION_RETURN_NULL) != NULL);
169 }
170
171 /*
172 * mdcreate() -- Create a new relation on magnetic disk.
173 *
174 * If isRedo is true, it's okay for the relation to exist already.
175 */
176 void
mdcreate(SMgrRelation reln,ForkNumber forkNum,bool isRedo)177 mdcreate(SMgrRelation reln, ForkNumber forkNum, bool isRedo)
178 {
179 MdfdVec *mdfd;
180 char *path;
181 File fd;
182
183 if (isRedo && reln->md_num_open_segs[forkNum] > 0)
184 return; /* created and opened already... */
185
186 Assert(reln->md_num_open_segs[forkNum] == 0);
187
188 path = relpath(reln->smgr_rnode, forkNum);
189
190 fd = PathNameOpenFile(path, O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
191
192 if (fd < 0)
193 {
194 int save_errno = errno;
195
196 if (isRedo)
197 fd = PathNameOpenFile(path, O_RDWR | PG_BINARY);
198 if (fd < 0)
199 {
200 /* be sure to report the error reported by create, not open */
201 errno = save_errno;
202 ereport(ERROR,
203 (errcode_for_file_access(),
204 errmsg("could not create file \"%s\": %m", path)));
205 }
206 }
207
208 pfree(path);
209
210 _fdvec_resize(reln, forkNum, 1);
211 mdfd = &reln->md_seg_fds[forkNum][0];
212 mdfd->mdfd_vfd = fd;
213 mdfd->mdfd_segno = 0;
214 }
215
216 /*
217 * mdunlink() -- Unlink a relation.
218 *
219 * Note that we're passed a RelFileNodeBackend --- by the time this is called,
220 * there won't be an SMgrRelation hashtable entry anymore.
221 *
222 * forkNum can be a fork number to delete a specific fork, or InvalidForkNumber
223 * to delete all forks.
224 *
225 * For regular relations, we don't unlink the first segment file of the rel,
226 * but just truncate it to zero length, and record a request to unlink it after
227 * the next checkpoint. Additional segments can be unlinked immediately,
228 * however. Leaving the empty file in place prevents that relfilenode
229 * number from being reused. The scenario this protects us from is:
230 * 1. We delete a relation (and commit, and actually remove its file).
231 * 2. We create a new relation, which by chance gets the same relfilenode as
232 * the just-deleted one (OIDs must've wrapped around for that to happen).
233 * 3. We crash before another checkpoint occurs.
234 * During replay, we would delete the file and then recreate it, which is fine
235 * if the contents of the file were repopulated by subsequent WAL entries.
236 * But if we didn't WAL-log insertions, but instead relied on fsyncing the
237 * file after populating it (as for instance CLUSTER and CREATE INDEX do),
238 * the contents of the file would be lost forever. By leaving the empty file
239 * until after the next checkpoint, we prevent reassignment of the relfilenode
240 * number until it's safe, because relfilenode assignment skips over any
241 * existing file.
242 *
243 * We do not need to go through this dance for temp relations, though, because
244 * we never make WAL entries for temp rels, and so a temp rel poses no threat
245 * to the health of a regular rel that has taken over its relfilenode number.
246 * The fact that temp rels and regular rels have different file naming
247 * patterns provides additional safety.
248 *
249 * All the above applies only to the relation's main fork; other forks can
250 * just be removed immediately, since they are not needed to prevent the
251 * relfilenode number from being recycled. Also, we do not carefully
252 * track whether other forks have been created or not, but just attempt to
253 * unlink them unconditionally; so we should never complain about ENOENT.
254 *
255 * If isRedo is true, it's unsurprising for the relation to be already gone.
256 * Also, we should remove the file immediately instead of queuing a request
257 * for later, since during redo there's no possibility of creating a
258 * conflicting relation.
259 *
260 * Note: any failure should be reported as WARNING not ERROR, because
261 * we are usually not in a transaction anymore when this is called.
262 */
263 void
mdunlink(RelFileNodeBackend rnode,ForkNumber forkNum,bool isRedo)264 mdunlink(RelFileNodeBackend rnode, ForkNumber forkNum, bool isRedo)
265 {
266 /* Now do the per-fork work */
267 if (forkNum == InvalidForkNumber)
268 {
269 for (forkNum = 0; forkNum <= MAX_FORKNUM; forkNum++)
270 mdunlinkfork(rnode, forkNum, isRedo);
271 }
272 else
273 mdunlinkfork(rnode, forkNum, isRedo);
274 }
275
276 /*
277 * Truncate a file to release disk space.
278 */
279 static int
do_truncate(const char * path)280 do_truncate(const char *path)
281 {
282 int save_errno;
283 int ret;
284 int fd;
285
286 /* truncate(2) would be easier here, but Windows hasn't got it */
287 fd = OpenTransientFile(path, O_RDWR | PG_BINARY);
288 if (fd >= 0)
289 {
290 ret = ftruncate(fd, 0);
291 save_errno = errno;
292 CloseTransientFile(fd);
293 errno = save_errno;
294 }
295 else
296 ret = -1;
297
298 /* Log a warning here to avoid repetition in callers. */
299 if (ret < 0 && errno != ENOENT)
300 {
301 save_errno = errno;
302 ereport(WARNING,
303 (errcode_for_file_access(),
304 errmsg("could not truncate file \"%s\": %m", path)));
305 errno = save_errno;
306 }
307
308 return ret;
309 }
310
311 static void
mdunlinkfork(RelFileNodeBackend rnode,ForkNumber forkNum,bool isRedo)312 mdunlinkfork(RelFileNodeBackend rnode, ForkNumber forkNum, bool isRedo)
313 {
314 char *path;
315 int ret;
316
317 path = relpath(rnode, forkNum);
318
319 /*
320 * Delete or truncate the first segment.
321 */
322 if (isRedo || forkNum != MAIN_FORKNUM || RelFileNodeBackendIsTemp(rnode))
323 {
324 if (!RelFileNodeBackendIsTemp(rnode))
325 {
326 /* Prevent other backends' fds from holding on to the disk space */
327 ret = do_truncate(path);
328
329 /* Forget any pending sync requests for the first segment */
330 register_forget_request(rnode, forkNum, 0 /* first seg */ );
331 }
332 else
333 ret = 0;
334
335 /* Next unlink the file, unless it was already found to be missing */
336 if (ret == 0 || errno != ENOENT)
337 {
338 ret = unlink(path);
339 if (ret < 0 && errno != ENOENT)
340 ereport(WARNING,
341 (errcode_for_file_access(),
342 errmsg("could not remove file \"%s\": %m", path)));
343 }
344 }
345 else
346 {
347 /* Prevent other backends' fds from holding on to the disk space */
348 ret = do_truncate(path);
349
350 /* Register request to unlink first segment later */
351 register_unlink_segment(rnode, forkNum, 0 /* first seg */ );
352 }
353
354 /*
355 * Delete any additional segments.
356 */
357 if (ret >= 0)
358 {
359 char *segpath = (char *) palloc(strlen(path) + 12);
360 BlockNumber segno;
361
362 /*
363 * Note that because we loop until getting ENOENT, we will correctly
364 * remove all inactive segments as well as active ones.
365 */
366 for (segno = 1;; segno++)
367 {
368 sprintf(segpath, "%s.%u", path, segno);
369
370 if (!RelFileNodeBackendIsTemp(rnode))
371 {
372 /*
373 * Prevent other backends' fds from holding on to the disk
374 * space.
375 */
376 if (do_truncate(segpath) < 0 && errno == ENOENT)
377 break;
378
379 /*
380 * Forget any pending sync requests for this segment before we
381 * try to unlink.
382 */
383 register_forget_request(rnode, forkNum, segno);
384 }
385
386 if (unlink(segpath) < 0)
387 {
388 /* ENOENT is expected after the last segment... */
389 if (errno != ENOENT)
390 ereport(WARNING,
391 (errcode_for_file_access(),
392 errmsg("could not remove file \"%s\": %m", segpath)));
393 break;
394 }
395 }
396 pfree(segpath);
397 }
398
399 pfree(path);
400 }
401
402 /*
403 * mdextend() -- Add a block to the specified relation.
404 *
405 * The semantics are nearly the same as mdwrite(): write at the
406 * specified position. However, this is to be used for the case of
407 * extending a relation (i.e., blocknum is at or beyond the current
408 * EOF). Note that we assume writing a block beyond current EOF
409 * causes intervening file space to become filled with zeroes.
410 */
411 void
mdextend(SMgrRelation reln,ForkNumber forknum,BlockNumber blocknum,char * buffer,bool skipFsync)412 mdextend(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
413 char *buffer, bool skipFsync)
414 {
415 off_t seekpos;
416 int nbytes;
417 MdfdVec *v;
418
419 /* This assert is too expensive to have on normally ... */
420 #ifdef CHECK_WRITE_VS_EXTEND
421 Assert(blocknum >= mdnblocks(reln, forknum));
422 #endif
423
424 /*
425 * If a relation manages to grow to 2^32-1 blocks, refuse to extend it any
426 * more --- we mustn't create a block whose number actually is
427 * InvalidBlockNumber. (Note that this failure should be unreachable
428 * because of upstream checks in bufmgr.c.)
429 */
430 if (blocknum == InvalidBlockNumber)
431 ereport(ERROR,
432 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
433 errmsg("cannot extend file \"%s\" beyond %u blocks",
434 relpath(reln->smgr_rnode, forknum),
435 InvalidBlockNumber)));
436
437 v = _mdfd_getseg(reln, forknum, blocknum, skipFsync, EXTENSION_CREATE);
438
439 seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
440
441 Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
442
443 if ((nbytes = FileWrite(v->mdfd_vfd, buffer, BLCKSZ, seekpos, WAIT_EVENT_DATA_FILE_EXTEND)) != BLCKSZ)
444 {
445 if (nbytes < 0)
446 ereport(ERROR,
447 (errcode_for_file_access(),
448 errmsg("could not extend file \"%s\": %m",
449 FilePathName(v->mdfd_vfd)),
450 errhint("Check free disk space.")));
451 /* short write: complain appropriately */
452 ereport(ERROR,
453 (errcode(ERRCODE_DISK_FULL),
454 errmsg("could not extend file \"%s\": wrote only %d of %d bytes at block %u",
455 FilePathName(v->mdfd_vfd),
456 nbytes, BLCKSZ, blocknum),
457 errhint("Check free disk space.")));
458 }
459
460 if (!skipFsync && !SmgrIsTemp(reln))
461 register_dirty_segment(reln, forknum, v);
462
463 Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
464 }
465
466 /*
467 * mdopen() -- Open the specified relation.
468 *
469 * Note we only open the first segment, when there are multiple segments.
470 *
471 * If first segment is not present, either ereport or return NULL according
472 * to "behavior". We treat EXTENSION_CREATE the same as EXTENSION_FAIL;
473 * EXTENSION_CREATE means it's OK to extend an existing relation, not to
474 * invent one out of whole cloth.
475 */
476 static MdfdVec *
mdopen(SMgrRelation reln,ForkNumber forknum,int behavior)477 mdopen(SMgrRelation reln, ForkNumber forknum, int behavior)
478 {
479 MdfdVec *mdfd;
480 char *path;
481 File fd;
482
483 /* No work if already open */
484 if (reln->md_num_open_segs[forknum] > 0)
485 return &reln->md_seg_fds[forknum][0];
486
487 path = relpath(reln->smgr_rnode, forknum);
488
489 fd = PathNameOpenFile(path, O_RDWR | PG_BINARY);
490
491 if (fd < 0)
492 {
493 if ((behavior & EXTENSION_RETURN_NULL) &&
494 FILE_POSSIBLY_DELETED(errno))
495 {
496 pfree(path);
497 return NULL;
498 }
499 ereport(ERROR,
500 (errcode_for_file_access(),
501 errmsg("could not open file \"%s\": %m", path)));
502 }
503
504 pfree(path);
505
506 _fdvec_resize(reln, forknum, 1);
507 mdfd = &reln->md_seg_fds[forknum][0];
508 mdfd->mdfd_vfd = fd;
509 mdfd->mdfd_segno = 0;
510
511 Assert(_mdnblocks(reln, forknum, mdfd) <= ((BlockNumber) RELSEG_SIZE));
512
513 return mdfd;
514 }
515
516 /*
517 * mdclose() -- Close the specified relation, if it isn't closed already.
518 */
519 void
mdclose(SMgrRelation reln,ForkNumber forknum)520 mdclose(SMgrRelation reln, ForkNumber forknum)
521 {
522 int nopensegs = reln->md_num_open_segs[forknum];
523
524 /* No work if already closed */
525 if (nopensegs == 0)
526 return;
527
528 /* close segments starting from the end */
529 while (nopensegs > 0)
530 {
531 MdfdVec *v = &reln->md_seg_fds[forknum][nopensegs - 1];
532
533 FileClose(v->mdfd_vfd);
534 _fdvec_resize(reln, forknum, nopensegs - 1);
535 nopensegs--;
536 }
537 }
538
539 /*
540 * mdprefetch() -- Initiate asynchronous read of the specified block of a relation
541 */
542 void
mdprefetch(SMgrRelation reln,ForkNumber forknum,BlockNumber blocknum)543 mdprefetch(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum)
544 {
545 #ifdef USE_PREFETCH
546 off_t seekpos;
547 MdfdVec *v;
548
549 v = _mdfd_getseg(reln, forknum, blocknum, false, EXTENSION_FAIL);
550
551 seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
552
553 Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
554
555 (void) FilePrefetch(v->mdfd_vfd, seekpos, BLCKSZ, WAIT_EVENT_DATA_FILE_PREFETCH);
556 #endif /* USE_PREFETCH */
557 }
558
559 /*
560 * mdwriteback() -- Tell the kernel to write pages back to storage.
561 *
562 * This accepts a range of blocks because flushing several pages at once is
563 * considerably more efficient than doing so individually.
564 */
565 void
mdwriteback(SMgrRelation reln,ForkNumber forknum,BlockNumber blocknum,BlockNumber nblocks)566 mdwriteback(SMgrRelation reln, ForkNumber forknum,
567 BlockNumber blocknum, BlockNumber nblocks)
568 {
569 /*
570 * Issue flush requests in as few requests as possible; have to split at
571 * segment boundaries though, since those are actually separate files.
572 */
573 while (nblocks > 0)
574 {
575 BlockNumber nflush = nblocks;
576 off_t seekpos;
577 MdfdVec *v;
578 int segnum_start,
579 segnum_end;
580
581 v = _mdfd_getseg(reln, forknum, blocknum, true /* not used */ ,
582 EXTENSION_RETURN_NULL);
583
584 /*
585 * We might be flushing buffers of already removed relations, that's
586 * ok, just ignore that case.
587 */
588 if (!v)
589 return;
590
591 /* compute offset inside the current segment */
592 segnum_start = blocknum / RELSEG_SIZE;
593
594 /* compute number of desired writes within the current segment */
595 segnum_end = (blocknum + nblocks - 1) / RELSEG_SIZE;
596 if (segnum_start != segnum_end)
597 nflush = RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE));
598
599 Assert(nflush >= 1);
600 Assert(nflush <= nblocks);
601
602 seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
603
604 FileWriteback(v->mdfd_vfd, seekpos, (off_t) BLCKSZ * nflush, WAIT_EVENT_DATA_FILE_FLUSH);
605
606 nblocks -= nflush;
607 blocknum += nflush;
608 }
609 }
610
611 /*
612 * mdread() -- Read the specified block from a relation.
613 */
614 void
mdread(SMgrRelation reln,ForkNumber forknum,BlockNumber blocknum,char * buffer)615 mdread(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
616 char *buffer)
617 {
618 off_t seekpos;
619 int nbytes;
620 MdfdVec *v;
621
622 TRACE_POSTGRESQL_SMGR_MD_READ_START(forknum, blocknum,
623 reln->smgr_rnode.node.spcNode,
624 reln->smgr_rnode.node.dbNode,
625 reln->smgr_rnode.node.relNode,
626 reln->smgr_rnode.backend);
627
628 v = _mdfd_getseg(reln, forknum, blocknum, false,
629 EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
630
631 seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
632
633 Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
634
635 nbytes = FileRead(v->mdfd_vfd, buffer, BLCKSZ, seekpos, WAIT_EVENT_DATA_FILE_READ);
636
637 TRACE_POSTGRESQL_SMGR_MD_READ_DONE(forknum, blocknum,
638 reln->smgr_rnode.node.spcNode,
639 reln->smgr_rnode.node.dbNode,
640 reln->smgr_rnode.node.relNode,
641 reln->smgr_rnode.backend,
642 nbytes,
643 BLCKSZ);
644
645 if (nbytes != BLCKSZ)
646 {
647 if (nbytes < 0)
648 ereport(ERROR,
649 (errcode_for_file_access(),
650 errmsg("could not read block %u in file \"%s\": %m",
651 blocknum, FilePathName(v->mdfd_vfd))));
652
653 /*
654 * Short read: we are at or past EOF, or we read a partial block at
655 * EOF. Normally this is an error; upper levels should never try to
656 * read a nonexistent block. However, if zero_damaged_pages is ON or
657 * we are InRecovery, we should instead return zeroes without
658 * complaining. This allows, for example, the case of trying to
659 * update a block that was later truncated away.
660 */
661 if (zero_damaged_pages || InRecovery)
662 MemSet(buffer, 0, BLCKSZ);
663 else
664 ereport(ERROR,
665 (errcode(ERRCODE_DATA_CORRUPTED),
666 errmsg("could not read block %u in file \"%s\": read only %d of %d bytes",
667 blocknum, FilePathName(v->mdfd_vfd),
668 nbytes, BLCKSZ)));
669 }
670 }
671
672 /*
673 * mdwrite() -- Write the supplied block at the appropriate location.
674 *
675 * This is to be used only for updating already-existing blocks of a
676 * relation (ie, those before the current EOF). To extend a relation,
677 * use mdextend().
678 */
679 void
mdwrite(SMgrRelation reln,ForkNumber forknum,BlockNumber blocknum,char * buffer,bool skipFsync)680 mdwrite(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
681 char *buffer, bool skipFsync)
682 {
683 off_t seekpos;
684 int nbytes;
685 MdfdVec *v;
686
687 /* This assert is too expensive to have on normally ... */
688 #ifdef CHECK_WRITE_VS_EXTEND
689 Assert(blocknum < mdnblocks(reln, forknum));
690 #endif
691
692 TRACE_POSTGRESQL_SMGR_MD_WRITE_START(forknum, blocknum,
693 reln->smgr_rnode.node.spcNode,
694 reln->smgr_rnode.node.dbNode,
695 reln->smgr_rnode.node.relNode,
696 reln->smgr_rnode.backend);
697
698 v = _mdfd_getseg(reln, forknum, blocknum, skipFsync,
699 EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
700
701 seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
702
703 Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
704
705 nbytes = FileWrite(v->mdfd_vfd, buffer, BLCKSZ, seekpos, WAIT_EVENT_DATA_FILE_WRITE);
706
707 TRACE_POSTGRESQL_SMGR_MD_WRITE_DONE(forknum, blocknum,
708 reln->smgr_rnode.node.spcNode,
709 reln->smgr_rnode.node.dbNode,
710 reln->smgr_rnode.node.relNode,
711 reln->smgr_rnode.backend,
712 nbytes,
713 BLCKSZ);
714
715 if (nbytes != BLCKSZ)
716 {
717 if (nbytes < 0)
718 ereport(ERROR,
719 (errcode_for_file_access(),
720 errmsg("could not write block %u in file \"%s\": %m",
721 blocknum, FilePathName(v->mdfd_vfd))));
722 /* short write: complain appropriately */
723 ereport(ERROR,
724 (errcode(ERRCODE_DISK_FULL),
725 errmsg("could not write block %u in file \"%s\": wrote only %d of %d bytes",
726 blocknum,
727 FilePathName(v->mdfd_vfd),
728 nbytes, BLCKSZ),
729 errhint("Check free disk space.")));
730 }
731
732 if (!skipFsync && !SmgrIsTemp(reln))
733 register_dirty_segment(reln, forknum, v);
734 }
735
736 /*
737 * mdnblocks() -- Get the number of blocks stored in a relation.
738 *
739 * Important side effect: all active segments of the relation are opened
740 * and added to the mdfd_seg_fds array. If this routine has not been
741 * called, then only segments up to the last one actually touched
742 * are present in the array.
743 */
744 BlockNumber
mdnblocks(SMgrRelation reln,ForkNumber forknum)745 mdnblocks(SMgrRelation reln, ForkNumber forknum)
746 {
747 MdfdVec *v = mdopen(reln, forknum, EXTENSION_FAIL);
748 BlockNumber nblocks;
749 BlockNumber segno = 0;
750
751 /* mdopen has opened the first segment */
752 Assert(reln->md_num_open_segs[forknum] > 0);
753
754 /*
755 * Start from the last open segments, to avoid redundant seeks. We have
756 * previously verified that these segments are exactly RELSEG_SIZE long,
757 * and it's useless to recheck that each time.
758 *
759 * NOTE: this assumption could only be wrong if another backend has
760 * truncated the relation. We rely on higher code levels to handle that
761 * scenario by closing and re-opening the md fd, which is handled via
762 * relcache flush. (Since the checkpointer doesn't participate in
763 * relcache flush, it could have segment entries for inactive segments;
764 * that's OK because the checkpointer never needs to compute relation
765 * size.)
766 */
767 segno = reln->md_num_open_segs[forknum] - 1;
768 v = &reln->md_seg_fds[forknum][segno];
769
770 for (;;)
771 {
772 nblocks = _mdnblocks(reln, forknum, v);
773 if (nblocks > ((BlockNumber) RELSEG_SIZE))
774 elog(FATAL, "segment too big");
775 if (nblocks < ((BlockNumber) RELSEG_SIZE))
776 return (segno * ((BlockNumber) RELSEG_SIZE)) + nblocks;
777
778 /*
779 * If segment is exactly RELSEG_SIZE, advance to next one.
780 */
781 segno++;
782
783 /*
784 * We used to pass O_CREAT here, but that has the disadvantage that it
785 * might create a segment which has vanished through some operating
786 * system misadventure. In such a case, creating the segment here
787 * undermines _mdfd_getseg's attempts to notice and report an error
788 * upon access to a missing segment.
789 */
790 v = _mdfd_openseg(reln, forknum, segno, 0);
791 if (v == NULL)
792 return segno * ((BlockNumber) RELSEG_SIZE);
793 }
794 }
795
796 /*
797 * mdtruncate() -- Truncate relation to specified number of blocks.
798 */
799 void
mdtruncate(SMgrRelation reln,ForkNumber forknum,BlockNumber nblocks)800 mdtruncate(SMgrRelation reln, ForkNumber forknum, BlockNumber nblocks)
801 {
802 BlockNumber curnblk;
803 BlockNumber priorblocks;
804 int curopensegs;
805
806 /*
807 * NOTE: mdnblocks makes sure we have opened all active segments, so that
808 * truncation loop will get them all!
809 */
810 curnblk = mdnblocks(reln, forknum);
811 if (nblocks > curnblk)
812 {
813 /* Bogus request ... but no complaint if InRecovery */
814 if (InRecovery)
815 return;
816 ereport(ERROR,
817 (errmsg("could not truncate file \"%s\" to %u blocks: it's only %u blocks now",
818 relpath(reln->smgr_rnode, forknum),
819 nblocks, curnblk)));
820 }
821 if (nblocks == curnblk)
822 return; /* no work */
823
824 /*
825 * Truncate segments, starting at the last one. Starting at the end makes
826 * managing the memory for the fd array easier, should there be errors.
827 */
828 curopensegs = reln->md_num_open_segs[forknum];
829 while (curopensegs > 0)
830 {
831 MdfdVec *v;
832
833 priorblocks = (curopensegs - 1) * RELSEG_SIZE;
834
835 v = &reln->md_seg_fds[forknum][curopensegs - 1];
836
837 if (priorblocks > nblocks)
838 {
839 /*
840 * This segment is no longer active. We truncate the file, but do
841 * not delete it, for reasons explained in the header comments.
842 */
843 if (FileTruncate(v->mdfd_vfd, 0, WAIT_EVENT_DATA_FILE_TRUNCATE) < 0)
844 ereport(ERROR,
845 (errcode_for_file_access(),
846 errmsg("could not truncate file \"%s\": %m",
847 FilePathName(v->mdfd_vfd))));
848
849 if (!SmgrIsTemp(reln))
850 register_dirty_segment(reln, forknum, v);
851
852 /* we never drop the 1st segment */
853 Assert(v != &reln->md_seg_fds[forknum][0]);
854
855 FileClose(v->mdfd_vfd);
856 _fdvec_resize(reln, forknum, curopensegs - 1);
857 }
858 else if (priorblocks + ((BlockNumber) RELSEG_SIZE) > nblocks)
859 {
860 /*
861 * This is the last segment we want to keep. Truncate the file to
862 * the right length. NOTE: if nblocks is exactly a multiple K of
863 * RELSEG_SIZE, we will truncate the K+1st segment to 0 length but
864 * keep it. This adheres to the invariant given in the header
865 * comments.
866 */
867 BlockNumber lastsegblocks = nblocks - priorblocks;
868
869 if (FileTruncate(v->mdfd_vfd, (off_t) lastsegblocks * BLCKSZ, WAIT_EVENT_DATA_FILE_TRUNCATE) < 0)
870 ereport(ERROR,
871 (errcode_for_file_access(),
872 errmsg("could not truncate file \"%s\" to %u blocks: %m",
873 FilePathName(v->mdfd_vfd),
874 nblocks)));
875 if (!SmgrIsTemp(reln))
876 register_dirty_segment(reln, forknum, v);
877 }
878 else
879 {
880 /*
881 * We still need this segment, so nothing to do for this and any
882 * earlier segment.
883 */
884 break;
885 }
886 curopensegs--;
887 }
888 }
889
890 /*
891 * mdimmedsync() -- Immediately sync a relation to stable storage.
892 *
893 * Note that only writes already issued are synced; this routine knows
894 * nothing of dirty buffers that may exist inside the buffer manager.
895 */
896 void
mdimmedsync(SMgrRelation reln,ForkNumber forknum)897 mdimmedsync(SMgrRelation reln, ForkNumber forknum)
898 {
899 int segno;
900
901 /*
902 * NOTE: mdnblocks makes sure we have opened all active segments, so that
903 * fsync loop will get them all!
904 */
905 mdnblocks(reln, forknum);
906
907 segno = reln->md_num_open_segs[forknum];
908
909 while (segno > 0)
910 {
911 MdfdVec *v = &reln->md_seg_fds[forknum][segno - 1];
912
913 if (FileSync(v->mdfd_vfd, WAIT_EVENT_DATA_FILE_IMMEDIATE_SYNC) < 0)
914 ereport(data_sync_elevel(ERROR),
915 (errcode_for_file_access(),
916 errmsg("could not fsync file \"%s\": %m",
917 FilePathName(v->mdfd_vfd))));
918 segno--;
919 }
920 }
921
922 /*
923 * register_dirty_segment() -- Mark a relation segment as needing fsync
924 *
925 * If there is a local pending-ops table, just make an entry in it for
926 * ProcessSyncRequests to process later. Otherwise, try to pass off the
927 * fsync request to the checkpointer process. If that fails, just do the
928 * fsync locally before returning (we hope this will not happen often
929 * enough to be a performance problem).
930 */
931 static void
register_dirty_segment(SMgrRelation reln,ForkNumber forknum,MdfdVec * seg)932 register_dirty_segment(SMgrRelation reln, ForkNumber forknum, MdfdVec *seg)
933 {
934 FileTag tag;
935
936 INIT_MD_FILETAG(tag, reln->smgr_rnode.node, forknum, seg->mdfd_segno);
937
938 /* Temp relations should never be fsync'd */
939 Assert(!SmgrIsTemp(reln));
940
941 if (!RegisterSyncRequest(&tag, SYNC_REQUEST, false /* retryOnError */ ))
942 {
943 ereport(DEBUG1,
944 (errmsg("could not forward fsync request because request queue is full")));
945
946 if (FileSync(seg->mdfd_vfd, WAIT_EVENT_DATA_FILE_SYNC) < 0)
947 ereport(data_sync_elevel(ERROR),
948 (errcode_for_file_access(),
949 errmsg("could not fsync file \"%s\": %m",
950 FilePathName(seg->mdfd_vfd))));
951 }
952 }
953
954 /*
955 * register_unlink_segment() -- Schedule a file to be deleted after next checkpoint
956 */
957 static void
register_unlink_segment(RelFileNodeBackend rnode,ForkNumber forknum,BlockNumber segno)958 register_unlink_segment(RelFileNodeBackend rnode, ForkNumber forknum,
959 BlockNumber segno)
960 {
961 FileTag tag;
962
963 INIT_MD_FILETAG(tag, rnode.node, forknum, segno);
964
965 /* Should never be used with temp relations */
966 Assert(!RelFileNodeBackendIsTemp(rnode));
967
968 RegisterSyncRequest(&tag, SYNC_UNLINK_REQUEST, true /* retryOnError */ );
969 }
970
971 /*
972 * register_forget_request() -- forget any fsyncs for a relation fork's segment
973 */
974 static void
register_forget_request(RelFileNodeBackend rnode,ForkNumber forknum,BlockNumber segno)975 register_forget_request(RelFileNodeBackend rnode, ForkNumber forknum,
976 BlockNumber segno)
977 {
978 FileTag tag;
979
980 INIT_MD_FILETAG(tag, rnode.node, forknum, segno);
981
982 RegisterSyncRequest(&tag, SYNC_FORGET_REQUEST, true /* retryOnError */ );
983 }
984
985 /*
986 * ForgetDatabaseSyncRequests -- forget any fsyncs and unlinks for a DB
987 */
988 void
ForgetDatabaseSyncRequests(Oid dbid)989 ForgetDatabaseSyncRequests(Oid dbid)
990 {
991 FileTag tag;
992 RelFileNode rnode;
993
994 rnode.dbNode = dbid;
995 rnode.spcNode = 0;
996 rnode.relNode = 0;
997
998 INIT_MD_FILETAG(tag, rnode, InvalidForkNumber, InvalidBlockNumber);
999
1000 RegisterSyncRequest(&tag, SYNC_FILTER_REQUEST, true /* retryOnError */ );
1001 }
1002
1003 /*
1004 * DropRelationFiles -- drop files of all given relations
1005 */
1006 void
DropRelationFiles(RelFileNode * delrels,int ndelrels,bool isRedo)1007 DropRelationFiles(RelFileNode *delrels, int ndelrels, bool isRedo)
1008 {
1009 SMgrRelation *srels;
1010 int i;
1011
1012 srels = palloc(sizeof(SMgrRelation) * ndelrels);
1013 for (i = 0; i < ndelrels; i++)
1014 {
1015 SMgrRelation srel = smgropen(delrels[i], InvalidBackendId);
1016
1017 if (isRedo)
1018 {
1019 ForkNumber fork;
1020
1021 for (fork = 0; fork <= MAX_FORKNUM; fork++)
1022 XLogDropRelation(delrels[i], fork);
1023 }
1024 srels[i] = srel;
1025 }
1026
1027 smgrdounlinkall(srels, ndelrels, isRedo);
1028
1029 for (i = 0; i < ndelrels; i++)
1030 smgrclose(srels[i]);
1031 pfree(srels);
1032 }
1033
1034
1035 /*
1036 * _fdvec_resize() -- Resize the fork's open segments array
1037 */
1038 static void
_fdvec_resize(SMgrRelation reln,ForkNumber forknum,int nseg)1039 _fdvec_resize(SMgrRelation reln,
1040 ForkNumber forknum,
1041 int nseg)
1042 {
1043 if (nseg == 0)
1044 {
1045 if (reln->md_num_open_segs[forknum] > 0)
1046 {
1047 pfree(reln->md_seg_fds[forknum]);
1048 reln->md_seg_fds[forknum] = NULL;
1049 }
1050 }
1051 else if (reln->md_num_open_segs[forknum] == 0)
1052 {
1053 reln->md_seg_fds[forknum] =
1054 MemoryContextAlloc(MdCxt, sizeof(MdfdVec) * nseg);
1055 }
1056 else
1057 {
1058 /*
1059 * It doesn't seem worthwhile complicating the code to amortize
1060 * repalloc() calls. Those are far faster than PathNameOpenFile() or
1061 * FileClose(), and the memory context internally will sometimes avoid
1062 * doing an actual reallocation.
1063 */
1064 reln->md_seg_fds[forknum] =
1065 repalloc(reln->md_seg_fds[forknum],
1066 sizeof(MdfdVec) * nseg);
1067 }
1068
1069 reln->md_num_open_segs[forknum] = nseg;
1070 }
1071
1072 /*
1073 * Return the filename for the specified segment of the relation. The
1074 * returned string is palloc'd.
1075 */
1076 static char *
_mdfd_segpath(SMgrRelation reln,ForkNumber forknum,BlockNumber segno)1077 _mdfd_segpath(SMgrRelation reln, ForkNumber forknum, BlockNumber segno)
1078 {
1079 char *path,
1080 *fullpath;
1081
1082 path = relpath(reln->smgr_rnode, forknum);
1083
1084 if (segno > 0)
1085 {
1086 fullpath = psprintf("%s.%u", path, segno);
1087 pfree(path);
1088 }
1089 else
1090 fullpath = path;
1091
1092 return fullpath;
1093 }
1094
1095 /*
1096 * Open the specified segment of the relation,
1097 * and make a MdfdVec object for it. Returns NULL on failure.
1098 */
1099 static MdfdVec *
_mdfd_openseg(SMgrRelation reln,ForkNumber forknum,BlockNumber segno,int oflags)1100 _mdfd_openseg(SMgrRelation reln, ForkNumber forknum, BlockNumber segno,
1101 int oflags)
1102 {
1103 MdfdVec *v;
1104 int fd;
1105 char *fullpath;
1106
1107 fullpath = _mdfd_segpath(reln, forknum, segno);
1108
1109 /* open the file */
1110 fd = PathNameOpenFile(fullpath, O_RDWR | PG_BINARY | oflags);
1111
1112 pfree(fullpath);
1113
1114 if (fd < 0)
1115 return NULL;
1116
1117 if (segno <= reln->md_num_open_segs[forknum])
1118 _fdvec_resize(reln, forknum, segno + 1);
1119
1120 /* fill the entry */
1121 v = &reln->md_seg_fds[forknum][segno];
1122 v->mdfd_vfd = fd;
1123 v->mdfd_segno = segno;
1124
1125 Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
1126
1127 /* all done */
1128 return v;
1129 }
1130
1131 /*
1132 * _mdfd_getseg() -- Find the segment of the relation holding the
1133 * specified block.
1134 *
1135 * If the segment doesn't exist, we ereport, return NULL, or create the
1136 * segment, according to "behavior". Note: skipFsync is only used in the
1137 * EXTENSION_CREATE case.
1138 */
1139 static MdfdVec *
_mdfd_getseg(SMgrRelation reln,ForkNumber forknum,BlockNumber blkno,bool skipFsync,int behavior)1140 _mdfd_getseg(SMgrRelation reln, ForkNumber forknum, BlockNumber blkno,
1141 bool skipFsync, int behavior)
1142 {
1143 MdfdVec *v;
1144 BlockNumber targetseg;
1145 BlockNumber nextsegno;
1146
1147 /* some way to handle non-existent segments needs to be specified */
1148 Assert(behavior &
1149 (EXTENSION_FAIL | EXTENSION_CREATE | EXTENSION_RETURN_NULL));
1150
1151 targetseg = blkno / ((BlockNumber) RELSEG_SIZE);
1152
1153 /* if an existing and opened segment, we're done */
1154 if (targetseg < reln->md_num_open_segs[forknum])
1155 {
1156 v = &reln->md_seg_fds[forknum][targetseg];
1157 return v;
1158 }
1159
1160 /*
1161 * The target segment is not yet open. Iterate over all the segments
1162 * between the last opened and the target segment. This way missing
1163 * segments either raise an error, or get created (according to
1164 * 'behavior'). Start with either the last opened, or the first segment if
1165 * none was opened before.
1166 */
1167 if (reln->md_num_open_segs[forknum] > 0)
1168 v = &reln->md_seg_fds[forknum][reln->md_num_open_segs[forknum] - 1];
1169 else
1170 {
1171 v = mdopen(reln, forknum, behavior);
1172 if (!v)
1173 return NULL; /* if behavior & EXTENSION_RETURN_NULL */
1174 }
1175
1176 for (nextsegno = reln->md_num_open_segs[forknum];
1177 nextsegno <= targetseg; nextsegno++)
1178 {
1179 BlockNumber nblocks = _mdnblocks(reln, forknum, v);
1180 int flags = 0;
1181
1182 Assert(nextsegno == v->mdfd_segno + 1);
1183
1184 if (nblocks > ((BlockNumber) RELSEG_SIZE))
1185 elog(FATAL, "segment too big");
1186
1187 if ((behavior & EXTENSION_CREATE) ||
1188 (InRecovery && (behavior & EXTENSION_CREATE_RECOVERY)))
1189 {
1190 /*
1191 * Normally we will create new segments only if authorized by the
1192 * caller (i.e., we are doing mdextend()). But when doing WAL
1193 * recovery, create segments anyway; this allows cases such as
1194 * replaying WAL data that has a write into a high-numbered
1195 * segment of a relation that was later deleted. We want to go
1196 * ahead and create the segments so we can finish out the replay.
1197 * However if the caller has specified
1198 * EXTENSION_REALLY_RETURN_NULL, then extension is not desired
1199 * even in recovery; we won't reach this point in that case.
1200 *
1201 * We have to maintain the invariant that segments before the last
1202 * active segment are of size RELSEG_SIZE; therefore, if
1203 * extending, pad them out with zeroes if needed. (This only
1204 * matters if in recovery, or if the caller is extending the
1205 * relation discontiguously, but that can happen in hash indexes.)
1206 */
1207 if (nblocks < ((BlockNumber) RELSEG_SIZE))
1208 {
1209 char *zerobuf = palloc0(BLCKSZ);
1210
1211 mdextend(reln, forknum,
1212 nextsegno * ((BlockNumber) RELSEG_SIZE) - 1,
1213 zerobuf, skipFsync);
1214 pfree(zerobuf);
1215 }
1216 flags = O_CREAT;
1217 }
1218 else if (!(behavior & EXTENSION_DONT_CHECK_SIZE) &&
1219 nblocks < ((BlockNumber) RELSEG_SIZE))
1220 {
1221 /*
1222 * When not extending (or explicitly including truncated
1223 * segments), only open the next segment if the current one is
1224 * exactly RELSEG_SIZE. If not (this branch), either return NULL
1225 * or fail.
1226 */
1227 if (behavior & EXTENSION_RETURN_NULL)
1228 {
1229 /*
1230 * Some callers discern between reasons for _mdfd_getseg()
1231 * returning NULL based on errno. As there's no failing
1232 * syscall involved in this case, explicitly set errno to
1233 * ENOENT, as that seems the closest interpretation.
1234 */
1235 errno = ENOENT;
1236 return NULL;
1237 }
1238
1239 ereport(ERROR,
1240 (errcode_for_file_access(),
1241 errmsg("could not open file \"%s\" (target block %u): previous segment is only %u blocks",
1242 _mdfd_segpath(reln, forknum, nextsegno),
1243 blkno, nblocks)));
1244 }
1245
1246 v = _mdfd_openseg(reln, forknum, nextsegno, flags);
1247
1248 if (v == NULL)
1249 {
1250 if ((behavior & EXTENSION_RETURN_NULL) &&
1251 FILE_POSSIBLY_DELETED(errno))
1252 return NULL;
1253 ereport(ERROR,
1254 (errcode_for_file_access(),
1255 errmsg("could not open file \"%s\" (target block %u): %m",
1256 _mdfd_segpath(reln, forknum, nextsegno),
1257 blkno)));
1258 }
1259 }
1260
1261 return v;
1262 }
1263
1264 /*
1265 * Get number of blocks present in a single disk file
1266 */
1267 static BlockNumber
_mdnblocks(SMgrRelation reln,ForkNumber forknum,MdfdVec * seg)1268 _mdnblocks(SMgrRelation reln, ForkNumber forknum, MdfdVec *seg)
1269 {
1270 off_t len;
1271
1272 len = FileSize(seg->mdfd_vfd);
1273 if (len < 0)
1274 ereport(ERROR,
1275 (errcode_for_file_access(),
1276 errmsg("could not seek to end of file \"%s\": %m",
1277 FilePathName(seg->mdfd_vfd))));
1278 /* note that this calculation will ignore any partial block at EOF */
1279 return (BlockNumber) (len / BLCKSZ);
1280 }
1281
1282 /*
1283 * Sync a file to disk, given a file tag. Write the path into an output
1284 * buffer so the caller can use it in error messages.
1285 *
1286 * Return 0 on success, -1 on failure, with errno set.
1287 */
1288 int
mdsyncfiletag(const FileTag * ftag,char * path)1289 mdsyncfiletag(const FileTag *ftag, char *path)
1290 {
1291 SMgrRelation reln = smgropen(ftag->rnode, InvalidBackendId);
1292 File file;
1293 bool need_to_close;
1294 int result,
1295 save_errno;
1296
1297 /* See if we already have the file open, or need to open it. */
1298 if (ftag->segno < reln->md_num_open_segs[ftag->forknum])
1299 {
1300 file = reln->md_seg_fds[ftag->forknum][ftag->segno].mdfd_vfd;
1301 strlcpy(path, FilePathName(file), MAXPGPATH);
1302 need_to_close = false;
1303 }
1304 else
1305 {
1306 char *p;
1307
1308 p = _mdfd_segpath(reln, ftag->forknum, ftag->segno);
1309 strlcpy(path, p, MAXPGPATH);
1310 pfree(p);
1311
1312 file = PathNameOpenFile(path, O_RDWR | PG_BINARY);
1313 if (file < 0)
1314 return -1;
1315 need_to_close = true;
1316 }
1317
1318 /* Sync the file. */
1319 result = FileSync(file, WAIT_EVENT_DATA_FILE_SYNC);
1320 save_errno = errno;
1321
1322 if (need_to_close)
1323 FileClose(file);
1324
1325 errno = save_errno;
1326 return result;
1327 }
1328
1329 /*
1330 * Unlink a file, given a file tag. Write the path into an output
1331 * buffer so the caller can use it in error messages.
1332 *
1333 * Return 0 on success, -1 on failure, with errno set.
1334 */
1335 int
mdunlinkfiletag(const FileTag * ftag,char * path)1336 mdunlinkfiletag(const FileTag *ftag, char *path)
1337 {
1338 char *p;
1339
1340 /* Compute the path. */
1341 p = relpathperm(ftag->rnode, MAIN_FORKNUM);
1342 strlcpy(path, p, MAXPGPATH);
1343 pfree(p);
1344
1345 /* Try to unlink the file. */
1346 return unlink(path);
1347 }
1348
1349 /*
1350 * Check if a given candidate request matches a given tag, when processing
1351 * a SYNC_FILTER_REQUEST request. This will be called for all pending
1352 * requests to find out whether to forget them.
1353 */
1354 bool
mdfiletagmatches(const FileTag * ftag,const FileTag * candidate)1355 mdfiletagmatches(const FileTag *ftag, const FileTag *candidate)
1356 {
1357 /*
1358 * For now we only use filter requests as a way to drop all scheduled
1359 * callbacks relating to a given database, when dropping the database.
1360 * We'll return true for all candidates that have the same database OID as
1361 * the ftag from the SYNC_FILTER_REQUEST request, so they're forgotten.
1362 */
1363 return ftag->rnode.dbNode == candidate->rnode.dbNode;
1364 }
1365