Line data Source code
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) 1996-2026, 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 <limits.h>
25 : #include <unistd.h>
26 : #include <fcntl.h>
27 : #include <sys/file.h>
28 :
29 : #include "access/xlogutils.h"
30 : #include "commands/tablespace.h"
31 : #include "common/file_utils.h"
32 : #include "miscadmin.h"
33 : #include "pg_trace.h"
34 : #include "pgstat.h"
35 : #include "storage/aio.h"
36 : #include "storage/bufmgr.h"
37 : #include "storage/fd.h"
38 : #include "storage/md.h"
39 : #include "storage/relfilelocator.h"
40 : #include "storage/smgr.h"
41 : #include "storage/sync.h"
42 : #include "utils/memutils.h"
43 :
44 : /*
45 : * The magnetic disk storage manager keeps track of open file
46 : * descriptors in its own descriptor pool. This is done to make it
47 : * easier to support relations that are larger than the operating
48 : * system's file size limit (often 2GBytes). In order to do that,
49 : * we break relations up into "segment" files that are each shorter than
50 : * the OS file size limit. The segment size is set by the RELSEG_SIZE
51 : * configuration constant in pg_config.h.
52 : *
53 : * On disk, a relation must consist of consecutively numbered segment
54 : * files in the pattern
55 : * -- Zero or more full segments of exactly RELSEG_SIZE blocks each
56 : * -- Exactly one partial segment of size 0 <= size < RELSEG_SIZE blocks
57 : * -- Optionally, any number of inactive segments of size 0 blocks.
58 : * The full and partial segments are collectively the "active" segments.
59 : * Inactive segments are those that once contained data but are currently
60 : * not needed because of an mdtruncate() operation. The reason for leaving
61 : * them present at size zero, rather than unlinking them, is that other
62 : * backends and/or the checkpointer might be holding open file references to
63 : * such segments. If the relation expands again after mdtruncate(), such
64 : * that a deactivated segment becomes active again, it is important that
65 : * such file references still be valid --- else data might get written
66 : * out to an unlinked old copy of a segment file that will eventually
67 : * disappear.
68 : *
69 : * RELSEG_SIZE must fit into BlockNumber; but since we expose its value
70 : * as an integer GUC, it actually needs to fit in signed int. It's worth
71 : * having a cross-check for this since configure's --with-segsize options
72 : * could let people select insane values.
73 : */
74 : StaticAssertDecl(RELSEG_SIZE > 0 && RELSEG_SIZE <= INT_MAX,
75 : "RELSEG_SIZE must fit in an integer");
76 :
77 : /*
78 : * File descriptors are stored in the per-fork md_seg_fds arrays inside
79 : * SMgrRelation. The length of these arrays is stored in md_num_open_segs.
80 : * Note that a fork's md_num_open_segs having a specific value does not
81 : * necessarily mean the relation doesn't have additional segments; we may
82 : * just not have opened the next segment yet. (We could not have "all
83 : * segments are in the array" as an invariant anyway, since another backend
84 : * could extend the relation while we aren't looking.) We do not have
85 : * entries for inactive segments, however; as soon as we find a partial
86 : * segment, we assume that any subsequent segments are inactive.
87 : *
88 : * The entire MdfdVec array is palloc'd in the MdCxt memory context.
89 : */
90 :
91 : typedef struct _MdfdVec
92 : {
93 : File mdfd_vfd; /* fd number in fd.c's pool */
94 : BlockNumber mdfd_segno; /* segment number, from 0 */
95 : } MdfdVec;
96 :
97 : static MemoryContext MdCxt; /* context for all MdfdVec objects */
98 :
99 :
100 : /* Populate a file tag describing an md.c segment file. */
101 : #define INIT_MD_FILETAG(a,xx_rlocator,xx_forknum,xx_segno) \
102 : ( \
103 : memset(&(a), 0, sizeof(FileTag)), \
104 : (a).handler = SYNC_HANDLER_MD, \
105 : (a).rlocator = (xx_rlocator), \
106 : (a).forknum = (xx_forknum), \
107 : (a).segno = (xx_segno) \
108 : )
109 :
110 :
111 : /*** behavior for mdopen & _mdfd_getseg ***/
112 : /* ereport if segment not present */
113 : #define EXTENSION_FAIL (1 << 0)
114 : /* return NULL if segment not present */
115 : #define EXTENSION_RETURN_NULL (1 << 1)
116 : /* create new segments as needed */
117 : #define EXTENSION_CREATE (1 << 2)
118 : /* create new segments if needed during recovery */
119 : #define EXTENSION_CREATE_RECOVERY (1 << 3)
120 : /* don't try to open a segment, if not already open */
121 : #define EXTENSION_DONT_OPEN (1 << 5)
122 :
123 :
124 : /*
125 : * Fixed-length string to represent paths to files that need to be built by
126 : * md.c.
127 : *
128 : * The maximum number of segments is MaxBlockNumber / RELSEG_SIZE, where
129 : * RELSEG_SIZE can be set to 1 (for testing only).
130 : */
131 : #define SEGMENT_CHARS OIDCHARS
132 : #define MD_PATH_STR_MAXLEN \
133 : (\
134 : REL_PATH_STR_MAXLEN \
135 : + sizeof((char)'.') \
136 : + SEGMENT_CHARS \
137 : )
138 : typedef struct MdPathStr
139 : {
140 : char str[MD_PATH_STR_MAXLEN + 1];
141 : } MdPathStr;
142 :
143 :
144 : /* local routines */
145 : static void mdunlinkfork(RelFileLocatorBackend rlocator, ForkNumber forknum,
146 : bool isRedo);
147 : static MdfdVec *mdopenfork(SMgrRelation reln, ForkNumber forknum, int behavior);
148 : static void register_dirty_segment(SMgrRelation reln, ForkNumber forknum,
149 : MdfdVec *seg);
150 : static void register_unlink_segment(RelFileLocatorBackend rlocator, ForkNumber forknum,
151 : BlockNumber segno);
152 : static void register_forget_request(RelFileLocatorBackend rlocator, ForkNumber forknum,
153 : BlockNumber segno);
154 : static void _fdvec_resize(SMgrRelation reln,
155 : ForkNumber forknum,
156 : int nseg);
157 : static MdPathStr _mdfd_segpath(SMgrRelation reln, ForkNumber forknum,
158 : BlockNumber segno);
159 : static MdfdVec *_mdfd_openseg(SMgrRelation reln, ForkNumber forknum,
160 : BlockNumber segno, int oflags);
161 : static MdfdVec *_mdfd_getseg(SMgrRelation reln, ForkNumber forknum,
162 : BlockNumber blkno, bool skipFsync, int behavior);
163 : static BlockNumber _mdnblocks(SMgrRelation reln, ForkNumber forknum,
164 : MdfdVec *seg);
165 :
166 : static PgAioResult md_readv_complete(PgAioHandle *ioh, PgAioResult prior_result, uint8 cb_data);
167 : static void md_readv_report(PgAioResult result, const PgAioTargetData *td, int elevel);
168 :
169 : const PgAioHandleCallbacks aio_md_readv_cb = {
170 : .complete_shared = md_readv_complete,
171 : .report = md_readv_report,
172 : };
173 :
174 :
175 : static inline int
176 1577930 : _mdfd_open_flags(void)
177 : {
178 1577930 : int flags = O_RDWR | PG_BINARY;
179 :
180 1577930 : if (io_direct_flags & IO_DIRECT_DATA)
181 306 : flags |= PG_O_DIRECT;
182 :
183 1577930 : return flags;
184 : }
185 :
186 : /*
187 : * mdinit() -- Initialize private state for magnetic disk storage manager.
188 : */
189 : void
190 23480 : mdinit(void)
191 : {
192 23480 : MdCxt = AllocSetContextCreate(TopMemoryContext,
193 : "MdSmgr",
194 : ALLOCSET_DEFAULT_SIZES);
195 23480 : }
196 :
197 : /*
198 : * mdexists() -- Does the physical file exist?
199 : *
200 : * Note: this will return true for lingering files, with pending deletions
201 : */
202 : bool
203 634598 : mdexists(SMgrRelation reln, ForkNumber forknum)
204 : {
205 : /*
206 : * Close it first, to ensure that we notice if the fork has been unlinked
207 : * since we opened it. As an optimization, we can skip that in recovery,
208 : * which already closes relations when dropping them.
209 : */
210 634598 : if (!InRecovery)
211 612647 : mdclose(reln, forknum);
212 :
213 634598 : return (mdopenfork(reln, forknum, EXTENSION_RETURN_NULL) != NULL);
214 : }
215 :
216 : /*
217 : * mdcreate() -- Create a new relation on magnetic disk.
218 : *
219 : * If isRedo is true, it's okay for the relation to exist already.
220 : */
221 : void
222 5764293 : mdcreate(SMgrRelation reln, ForkNumber forknum, bool isRedo)
223 : {
224 : MdfdVec *mdfd;
225 : RelPathStr path;
226 : File fd;
227 :
228 5764293 : if (isRedo && reln->md_num_open_segs[forknum] > 0)
229 5589965 : return; /* created and opened already... */
230 :
231 : Assert(reln->md_num_open_segs[forknum] == 0);
232 :
233 : /*
234 : * We may be using the target table space for the first time in this
235 : * database, so create a per-database subdirectory if needed.
236 : *
237 : * XXX this is a fairly ugly violation of module layering, but this seems
238 : * to be the best place to put the check. Maybe TablespaceCreateDbspace
239 : * should be here and not in commands/tablespace.c? But that would imply
240 : * importing a lot of stuff that smgr.c oughtn't know, either.
241 : */
242 174328 : TablespaceCreateDbspace(reln->smgr_rlocator.locator.spcOid,
243 : reln->smgr_rlocator.locator.dbOid,
244 : isRedo);
245 :
246 174328 : path = relpath(reln->smgr_rlocator, forknum);
247 :
248 174328 : fd = PathNameOpenFile(path.str, _mdfd_open_flags() | O_CREAT | O_EXCL);
249 :
250 174328 : if (fd < 0)
251 : {
252 13001 : int save_errno = errno;
253 :
254 13001 : if (isRedo)
255 13001 : fd = PathNameOpenFile(path.str, _mdfd_open_flags());
256 13001 : if (fd < 0)
257 : {
258 : /* be sure to report the error reported by create, not open */
259 0 : errno = save_errno;
260 0 : ereport(ERROR,
261 : (errcode_for_file_access(),
262 : errmsg("could not create file \"%s\": %m", path.str)));
263 : }
264 : }
265 :
266 174328 : _fdvec_resize(reln, forknum, 1);
267 174328 : mdfd = &reln->md_seg_fds[forknum][0];
268 174328 : mdfd->mdfd_vfd = fd;
269 174328 : mdfd->mdfd_segno = 0;
270 :
271 174328 : if (!SmgrIsTemp(reln))
272 170885 : register_dirty_segment(reln, forknum, mdfd);
273 : }
274 :
275 : /*
276 : * mdunlink() -- Unlink a relation.
277 : *
278 : * Note that we're passed a RelFileLocatorBackend --- by the time this is called,
279 : * there won't be an SMgrRelation hashtable entry anymore.
280 : *
281 : * forknum can be a fork number to delete a specific fork, or InvalidForkNumber
282 : * to delete all forks.
283 : *
284 : * For regular relations, we don't unlink the first segment file of the rel,
285 : * but just truncate it to zero length, and record a request to unlink it after
286 : * the next checkpoint. Additional segments can be unlinked immediately,
287 : * however. Leaving the empty file in place prevents that relfilenumber
288 : * from being reused. The scenario this protects us from is:
289 : * 1. We delete a relation (and commit, and actually remove its file).
290 : * 2. We create a new relation, which by chance gets the same relfilenumber as
291 : * the just-deleted one (OIDs must've wrapped around for that to happen).
292 : * 3. We crash before another checkpoint occurs.
293 : * During replay, we would delete the file and then recreate it, which is fine
294 : * if the contents of the file were repopulated by subsequent WAL entries.
295 : * But if we didn't WAL-log insertions, but instead relied on fsyncing the
296 : * file after populating it (as we do at wal_level=minimal), the contents of
297 : * the file would be lost forever. By leaving the empty file until after the
298 : * next checkpoint, we prevent reassignment of the relfilenumber until it's
299 : * safe, because relfilenumber assignment skips over any existing file.
300 : *
301 : * Additional segments, if any, are truncated and then unlinked. The reason
302 : * for truncating is that other backends may still hold open FDs for these at
303 : * the smgr level, so that the kernel can't remove the file yet. We want to
304 : * reclaim the disk space right away despite that.
305 : *
306 : * We do not need to go through this dance for temp relations, though, because
307 : * we never make WAL entries for temp rels, and so a temp rel poses no threat
308 : * to the health of a regular rel that has taken over its relfilenumber.
309 : * The fact that temp rels and regular rels have different file naming
310 : * patterns provides additional safety. Other backends shouldn't have open
311 : * FDs for them, either.
312 : *
313 : * We also don't do it while performing a binary upgrade. There is no reuse
314 : * hazard in that case, since after a crash or even a simple ERROR, the
315 : * upgrade fails and the whole cluster must be recreated from scratch.
316 : * Furthermore, it is important to remove the files from disk immediately,
317 : * because we may be about to reuse the same relfilenumber.
318 : *
319 : * All the above applies only to the relation's main fork; other forks can
320 : * just be removed immediately, since they are not needed to prevent the
321 : * relfilenumber from being recycled. Also, we do not carefully
322 : * track whether other forks have been created or not, but just attempt to
323 : * unlink them unconditionally; so we should never complain about ENOENT.
324 : *
325 : * If isRedo is true, it's unsurprising for the relation to be already gone.
326 : * Also, we should remove the file immediately instead of queuing a request
327 : * for later, since during redo there's no possibility of creating a
328 : * conflicting relation.
329 : *
330 : * Note: we currently just never warn about ENOENT at all. We could warn in
331 : * the main-fork, non-isRedo case, but it doesn't seem worth the trouble.
332 : *
333 : * Note: any failure should be reported as WARNING not ERROR, because
334 : * we are usually not in a transaction anymore when this is called.
335 : */
336 : void
337 193436 : mdunlink(RelFileLocatorBackend rlocator, ForkNumber forknum, bool isRedo)
338 : {
339 : /* Now do the per-fork work */
340 193436 : if (forknum == InvalidForkNumber)
341 : {
342 0 : for (forknum = 0; forknum <= MAX_FORKNUM; forknum++)
343 0 : mdunlinkfork(rlocator, forknum, isRedo);
344 : }
345 : else
346 193436 : mdunlinkfork(rlocator, forknum, isRedo);
347 193436 : }
348 :
349 : /*
350 : * Truncate a file to release disk space.
351 : */
352 : static int
353 227377 : do_truncate(const char *path)
354 : {
355 : int save_errno;
356 : int ret;
357 :
358 227377 : ret = pg_truncate(path, 0);
359 :
360 : /* Log a warning here to avoid repetition in callers. */
361 227377 : if (ret < 0 && errno != ENOENT)
362 : {
363 0 : save_errno = errno;
364 0 : ereport(WARNING,
365 : (errcode_for_file_access(),
366 : errmsg("could not truncate file \"%s\": %m", path)));
367 0 : errno = save_errno;
368 : }
369 :
370 227377 : return ret;
371 : }
372 :
373 : static void
374 193436 : mdunlinkfork(RelFileLocatorBackend rlocator, ForkNumber forknum, bool isRedo)
375 : {
376 : RelPathStr path;
377 : int ret;
378 : int save_errno;
379 :
380 193436 : path = relpath(rlocator, forknum);
381 :
382 : /*
383 : * Truncate and then unlink the first segment, or just register a request
384 : * to unlink it later, as described in the comments for mdunlink().
385 : */
386 193436 : if (isRedo || IsBinaryUpgrade || forknum != MAIN_FORKNUM ||
387 40212 : RelFileLocatorBackendIsTemp(rlocator))
388 : {
389 156496 : if (!RelFileLocatorBackendIsTemp(rlocator))
390 : {
391 : /* Prevent other backends' fds from holding on to the disk space */
392 143408 : ret = do_truncate(path.str);
393 :
394 : /* Forget any pending sync requests for the first segment */
395 143408 : save_errno = errno;
396 143408 : register_forget_request(rlocator, forknum, 0 /* first seg */ );
397 143408 : errno = save_errno;
398 : }
399 : else
400 13088 : ret = 0;
401 :
402 : /* Next unlink the file, unless it was already found to be missing */
403 156496 : if (ret >= 0 || errno != ENOENT)
404 : {
405 23181 : ret = unlink(path.str);
406 23181 : if (ret < 0 && errno != ENOENT)
407 : {
408 0 : save_errno = errno;
409 0 : ereport(WARNING,
410 : (errcode_for_file_access(),
411 : errmsg("could not remove file \"%s\": %m", path.str)));
412 0 : errno = save_errno;
413 : }
414 : }
415 : }
416 : else
417 : {
418 : /* Prevent other backends' fds from holding on to the disk space */
419 36940 : ret = do_truncate(path.str);
420 :
421 : /* Register request to unlink first segment later */
422 36940 : save_errno = errno;
423 36940 : register_unlink_segment(rlocator, forknum, 0 /* first seg */ );
424 36940 : errno = save_errno;
425 : }
426 :
427 : /*
428 : * Delete any additional segments.
429 : *
430 : * Note that because we loop until getting ENOENT, we will correctly
431 : * remove all inactive segments as well as active ones. Ideally we'd
432 : * continue the loop until getting exactly that errno, but that risks an
433 : * infinite loop if the problem is directory-wide (for instance, if we
434 : * suddenly can't read the data directory itself). We compromise by
435 : * continuing after a non-ENOENT truncate error, but stopping after any
436 : * unlink error. If there is indeed a directory-wide problem, additional
437 : * unlink attempts wouldn't work anyway.
438 : */
439 193436 : if (ret >= 0 || errno != ENOENT)
440 : {
441 : MdPathStr segpath;
442 : BlockNumber segno;
443 :
444 50472 : for (segno = 1;; segno++)
445 : {
446 50472 : sprintf(segpath.str, "%s.%u", path.str, segno);
447 :
448 50472 : if (!RelFileLocatorBackendIsTemp(rlocator))
449 : {
450 : /*
451 : * Prevent other backends' fds from holding on to the disk
452 : * space. We're done if we see ENOENT, though.
453 : */
454 47029 : if (do_truncate(segpath.str) < 0 && errno == ENOENT)
455 47029 : break;
456 :
457 : /*
458 : * Forget any pending sync requests for this segment before we
459 : * try to unlink.
460 : */
461 0 : register_forget_request(rlocator, forknum, segno);
462 : }
463 :
464 3443 : if (unlink(segpath.str) < 0)
465 : {
466 : /* ENOENT is expected after the last segment... */
467 3443 : if (errno != ENOENT)
468 0 : ereport(WARNING,
469 : (errcode_for_file_access(),
470 : errmsg("could not remove file \"%s\": %m", segpath.str)));
471 3443 : break;
472 : }
473 : }
474 : }
475 193436 : }
476 :
477 : /*
478 : * mdextend() -- Add a block to the specified relation.
479 : *
480 : * The semantics are nearly the same as mdwrite(): write at the
481 : * specified position. However, this is to be used for the case of
482 : * extending a relation (i.e., blocknum is at or beyond the current
483 : * EOF). Note that we assume writing a block beyond current EOF
484 : * causes intervening file space to become filled with zeroes.
485 : */
486 : void
487 123306 : mdextend(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
488 : const void *buffer, bool skipFsync)
489 : {
490 : pgoff_t seekpos;
491 : int nbytes;
492 : MdfdVec *v;
493 :
494 : /* If this build supports direct I/O, the buffer must be I/O aligned. */
495 : if (PG_O_DIRECT != 0 && PG_IO_ALIGN_SIZE <= BLCKSZ)
496 : Assert((uintptr_t) buffer == TYPEALIGN(PG_IO_ALIGN_SIZE, buffer));
497 :
498 : /* This assert is too expensive to have on normally ... */
499 : #ifdef CHECK_WRITE_VS_EXTEND
500 : Assert(blocknum >= mdnblocks(reln, forknum));
501 : #endif
502 :
503 : /*
504 : * If a relation manages to grow to 2^32-1 blocks, refuse to extend it any
505 : * more --- we mustn't create a block whose number actually is
506 : * InvalidBlockNumber. (Note that this failure should be unreachable
507 : * because of upstream checks in bufmgr.c.)
508 : */
509 123306 : if (blocknum == InvalidBlockNumber)
510 0 : ereport(ERROR,
511 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
512 : errmsg("cannot extend file \"%s\" beyond %u blocks",
513 : relpath(reln->smgr_rlocator, forknum).str,
514 : InvalidBlockNumber)));
515 :
516 123306 : v = _mdfd_getseg(reln, forknum, blocknum, skipFsync, EXTENSION_CREATE);
517 :
518 123306 : seekpos = (pgoff_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
519 :
520 : Assert(seekpos < (pgoff_t) BLCKSZ * RELSEG_SIZE);
521 :
522 123306 : if ((nbytes = FileWrite(v->mdfd_vfd, buffer, BLCKSZ, seekpos, WAIT_EVENT_DATA_FILE_EXTEND)) != BLCKSZ)
523 : {
524 0 : if (nbytes < 0)
525 0 : ereport(ERROR,
526 : (errcode_for_file_access(),
527 : errmsg("could not extend file \"%s\": %m",
528 : FilePathName(v->mdfd_vfd)),
529 : errhint("Check free disk space.")));
530 : /* short write: complain appropriately */
531 0 : ereport(ERROR,
532 : (errcode(ERRCODE_DISK_FULL),
533 : errmsg("could not extend file \"%s\": wrote only %d of %d bytes at block %u",
534 : FilePathName(v->mdfd_vfd),
535 : nbytes, BLCKSZ, blocknum),
536 : errhint("Check free disk space.")));
537 : }
538 :
539 123306 : if (!skipFsync && !SmgrIsTemp(reln))
540 29 : register_dirty_segment(reln, forknum, v);
541 :
542 : Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
543 123306 : }
544 :
545 : /*
546 : * mdzeroextend() -- Add new zeroed out blocks to the specified relation.
547 : *
548 : * Similar to mdextend(), except the relation can be extended by multiple
549 : * blocks at once and the added blocks will be filled with zeroes.
550 : */
551 : void
552 215467 : mdzeroextend(SMgrRelation reln, ForkNumber forknum,
553 : BlockNumber blocknum, int nblocks, bool skipFsync)
554 : {
555 : MdfdVec *v;
556 215467 : BlockNumber curblocknum = blocknum;
557 215467 : int remblocks = nblocks;
558 :
559 : Assert(nblocks > 0);
560 :
561 : /* This assert is too expensive to have on normally ... */
562 : #ifdef CHECK_WRITE_VS_EXTEND
563 : Assert(blocknum >= mdnblocks(reln, forknum));
564 : #endif
565 :
566 : /*
567 : * If a relation manages to grow to 2^32-1 blocks, refuse to extend it any
568 : * more --- we mustn't create a block whose number actually is
569 : * InvalidBlockNumber or larger.
570 : */
571 215467 : if ((uint64) blocknum + nblocks >= (uint64) InvalidBlockNumber)
572 0 : ereport(ERROR,
573 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
574 : errmsg("cannot extend file \"%s\" beyond %u blocks",
575 : relpath(reln->smgr_rlocator, forknum).str,
576 : InvalidBlockNumber)));
577 :
578 430934 : while (remblocks > 0)
579 : {
580 215467 : BlockNumber segstartblock = curblocknum % ((BlockNumber) RELSEG_SIZE);
581 215467 : pgoff_t seekpos = (pgoff_t) BLCKSZ * segstartblock;
582 : int numblocks;
583 :
584 215467 : if (segstartblock + remblocks > RELSEG_SIZE)
585 0 : numblocks = RELSEG_SIZE - segstartblock;
586 : else
587 215467 : numblocks = remblocks;
588 :
589 215467 : v = _mdfd_getseg(reln, forknum, curblocknum, skipFsync, EXTENSION_CREATE);
590 :
591 : Assert(segstartblock < RELSEG_SIZE);
592 : Assert(segstartblock + numblocks <= RELSEG_SIZE);
593 :
594 : /*
595 : * If available and useful, use posix_fallocate() (via
596 : * FileFallocate()) to extend the relation. That's often more
597 : * efficient than using write(), as it commonly won't cause the kernel
598 : * to allocate page cache space for the extended pages.
599 : *
600 : * However, we don't use FileFallocate() for small extensions, as it
601 : * defeats delayed allocation on some filesystems. Not clear where
602 : * that decision should be made though? For now just use a cutoff of
603 : * 8, anything between 4 and 8 worked OK in some local testing.
604 : */
605 215467 : if (numblocks > 8 &&
606 506 : file_extend_method != FILE_EXTEND_METHOD_WRITE_ZEROS)
607 506 : {
608 506 : int ret = 0;
609 :
610 : #ifdef HAVE_POSIX_FALLOCATE
611 506 : if (file_extend_method == FILE_EXTEND_METHOD_POSIX_FALLOCATE)
612 : {
613 506 : ret = FileFallocate(v->mdfd_vfd,
614 : seekpos, (pgoff_t) BLCKSZ * numblocks,
615 : WAIT_EVENT_DATA_FILE_EXTEND);
616 : }
617 : else
618 : #endif
619 : {
620 0 : elog(ERROR, "unsupported file_extend_method: %d",
621 : file_extend_method);
622 : }
623 506 : if (ret != 0)
624 : {
625 0 : ereport(ERROR,
626 : errcode_for_file_access(),
627 : errmsg("could not extend file \"%s\" with FileFallocate(): %m",
628 : FilePathName(v->mdfd_vfd)),
629 : errhint("Check free disk space."));
630 : }
631 : }
632 : else
633 : {
634 : int ret;
635 :
636 : /*
637 : * Even if we don't want to use fallocate, we can still extend a
638 : * bit more efficiently than writing each 8kB block individually.
639 : * pg_pwrite_zeros() (via FileZero()) uses pg_pwritev_with_retry()
640 : * to avoid multiple writes or needing a zeroed buffer for the
641 : * whole length of the extension.
642 : */
643 214961 : ret = FileZero(v->mdfd_vfd,
644 : seekpos, (pgoff_t) BLCKSZ * numblocks,
645 : WAIT_EVENT_DATA_FILE_EXTEND);
646 214961 : if (ret < 0)
647 0 : ereport(ERROR,
648 : errcode_for_file_access(),
649 : errmsg("could not extend file \"%s\": %m",
650 : FilePathName(v->mdfd_vfd)),
651 : errhint("Check free disk space."));
652 : }
653 :
654 215467 : if (!skipFsync && !SmgrIsTemp(reln))
655 204018 : register_dirty_segment(reln, forknum, v);
656 :
657 : Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
658 :
659 215467 : remblocks -= numblocks;
660 215467 : curblocknum += numblocks;
661 : }
662 215467 : }
663 :
664 : /*
665 : * mdopenfork() -- Open one fork of the specified relation.
666 : *
667 : * Note we only open the first segment, when there are multiple segments.
668 : *
669 : * If first segment is not present, either ereport or return NULL according
670 : * to "behavior". We treat EXTENSION_CREATE the same as EXTENSION_FAIL;
671 : * EXTENSION_CREATE means it's OK to extend an existing relation, not to
672 : * invent one out of whole cloth.
673 : */
674 : static MdfdVec *
675 4016243 : mdopenfork(SMgrRelation reln, ForkNumber forknum, int behavior)
676 : {
677 : MdfdVec *mdfd;
678 : RelPathStr path;
679 : File fd;
680 :
681 : /* No work if already open */
682 4016243 : if (reln->md_num_open_segs[forknum] > 0)
683 2651397 : return &reln->md_seg_fds[forknum][0];
684 :
685 1364846 : path = relpath(reln->smgr_rlocator, forknum);
686 :
687 1364846 : fd = PathNameOpenFile(path.str, _mdfd_open_flags());
688 :
689 1364846 : if (fd < 0)
690 : {
691 403929 : if ((behavior & EXTENSION_RETURN_NULL) &&
692 403907 : FILE_POSSIBLY_DELETED(errno))
693 403907 : return NULL;
694 22 : ereport(ERROR,
695 : (errcode_for_file_access(),
696 : errmsg("could not open file \"%s\": %m", path.str)));
697 : }
698 :
699 960917 : _fdvec_resize(reln, forknum, 1);
700 960917 : mdfd = &reln->md_seg_fds[forknum][0];
701 960917 : mdfd->mdfd_vfd = fd;
702 960917 : mdfd->mdfd_segno = 0;
703 :
704 : Assert(_mdnblocks(reln, forknum, mdfd) <= ((BlockNumber) RELSEG_SIZE));
705 :
706 960917 : return mdfd;
707 : }
708 :
709 : /*
710 : * mdopen() -- Initialize newly-opened relation.
711 : */
712 : void
713 1312950 : mdopen(SMgrRelation reln)
714 : {
715 : /* mark it not open */
716 6564750 : for (int forknum = 0; forknum <= MAX_FORKNUM; forknum++)
717 5251800 : reln->md_num_open_segs[forknum] = 0;
718 1312950 : }
719 :
720 : /*
721 : * mdclose() -- Close the specified relation, if it isn't closed already.
722 : */
723 : void
724 3865903 : mdclose(SMgrRelation reln, ForkNumber forknum)
725 : {
726 3865903 : int nopensegs = reln->md_num_open_segs[forknum];
727 :
728 : /* No work if already closed */
729 3865903 : if (nopensegs == 0)
730 3295104 : return;
731 :
732 : /* close segments starting from the end */
733 1141598 : while (nopensegs > 0)
734 : {
735 570799 : MdfdVec *v = &reln->md_seg_fds[forknum][nopensegs - 1];
736 :
737 570799 : FileClose(v->mdfd_vfd);
738 570799 : _fdvec_resize(reln, forknum, nopensegs - 1);
739 570799 : nopensegs--;
740 : }
741 : }
742 :
743 : /*
744 : * mdprefetch() -- Initiate asynchronous read of the specified blocks of a relation
745 : */
746 : bool
747 8818 : mdprefetch(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
748 : int nblocks)
749 : {
750 : #ifdef USE_PREFETCH
751 :
752 : Assert((io_direct_flags & IO_DIRECT_DATA) == 0);
753 :
754 8818 : if ((uint64) blocknum + nblocks > (uint64) MaxBlockNumber + 1)
755 0 : return false;
756 :
757 17636 : while (nblocks > 0)
758 : {
759 : pgoff_t seekpos;
760 : MdfdVec *v;
761 : int nblocks_this_segment;
762 :
763 8818 : v = _mdfd_getseg(reln, forknum, blocknum, false,
764 8818 : InRecovery ? EXTENSION_RETURN_NULL : EXTENSION_FAIL);
765 8818 : if (v == NULL)
766 0 : return false;
767 :
768 8818 : seekpos = (pgoff_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
769 :
770 : Assert(seekpos < (pgoff_t) BLCKSZ * RELSEG_SIZE);
771 :
772 8818 : nblocks_this_segment =
773 8818 : Min(nblocks,
774 : RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
775 :
776 8818 : (void) FilePrefetch(v->mdfd_vfd, seekpos, BLCKSZ * nblocks_this_segment,
777 : WAIT_EVENT_DATA_FILE_PREFETCH);
778 :
779 8818 : blocknum += nblocks_this_segment;
780 8818 : nblocks -= nblocks_this_segment;
781 : }
782 : #endif /* USE_PREFETCH */
783 :
784 8818 : return true;
785 : }
786 :
787 : /*
788 : * Convert an array of buffer address into an array of iovec objects, and
789 : * return the number that were required. 'iov' must have enough space for up
790 : * to 'nblocks' elements, but the number used may be less depending on
791 : * merging. In the case of a run of fully contiguous buffers, a single iovec
792 : * will be populated that can be handled as a plain non-vectored I/O.
793 : */
794 : static int
795 1893605 : buffers_to_iovec(struct iovec *iov, void **buffers, int nblocks)
796 : {
797 : struct iovec *iovp;
798 : int iovcnt;
799 :
800 : Assert(nblocks >= 1);
801 :
802 : /* If this build supports direct I/O, buffers must be I/O aligned. */
803 3960781 : for (int i = 0; i < nblocks; ++i)
804 : {
805 : if (PG_O_DIRECT != 0 && PG_IO_ALIGN_SIZE <= BLCKSZ)
806 : Assert((uintptr_t) buffers[i] ==
807 : TYPEALIGN(PG_IO_ALIGN_SIZE, buffers[i]));
808 : }
809 :
810 : /* Start the first iovec off with the first buffer. */
811 1893605 : iovp = &iov[0];
812 1893605 : iovp->iov_base = buffers[0];
813 1893605 : iovp->iov_len = BLCKSZ;
814 1893605 : iovcnt = 1;
815 :
816 : /* Try to merge the rest. */
817 2067176 : for (int i = 1; i < nblocks; ++i)
818 : {
819 173571 : void *buffer = buffers[i];
820 :
821 173571 : if (((char *) iovp->iov_base + iovp->iov_len) == buffer)
822 : {
823 : /* Contiguous with the last iovec. */
824 172836 : iovp->iov_len += BLCKSZ;
825 : }
826 : else
827 : {
828 : /* Need a new iovec. */
829 735 : iovp++;
830 735 : iovp->iov_base = buffer;
831 735 : iovp->iov_len = BLCKSZ;
832 735 : iovcnt++;
833 : }
834 : }
835 :
836 1893605 : return iovcnt;
837 : }
838 :
839 : /*
840 : * mdmaxcombine() -- Return the maximum number of total blocks that can be
841 : * combined with an IO starting at blocknum.
842 : */
843 : uint32
844 35108 : mdmaxcombine(SMgrRelation reln, ForkNumber forknum,
845 : BlockNumber blocknum)
846 : {
847 : BlockNumber segoff;
848 :
849 35108 : segoff = blocknum % ((BlockNumber) RELSEG_SIZE);
850 :
851 35108 : return RELSEG_SIZE - segoff;
852 : }
853 :
854 : /*
855 : * mdreadv() -- Read the specified blocks from a relation.
856 : */
857 : void
858 598 : mdreadv(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
859 : void **buffers, BlockNumber nblocks)
860 : {
861 1196 : while (nblocks > 0)
862 : {
863 : struct iovec iov[PG_IOV_MAX];
864 : int iovcnt;
865 : pgoff_t seekpos;
866 : int nbytes;
867 : MdfdVec *v;
868 : BlockNumber nblocks_this_segment;
869 : size_t transferred_this_segment;
870 : size_t size_this_segment;
871 :
872 598 : v = _mdfd_getseg(reln, forknum, blocknum, false,
873 : EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
874 :
875 598 : seekpos = (pgoff_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
876 :
877 : Assert(seekpos < (pgoff_t) BLCKSZ * RELSEG_SIZE);
878 :
879 598 : nblocks_this_segment =
880 598 : Min(nblocks,
881 : RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
882 598 : nblocks_this_segment = Min(nblocks_this_segment, lengthof(iov));
883 :
884 598 : if (nblocks_this_segment != nblocks)
885 0 : elog(ERROR, "read crosses segment boundary");
886 :
887 598 : iovcnt = buffers_to_iovec(iov, buffers, nblocks_this_segment);
888 598 : size_this_segment = nblocks_this_segment * BLCKSZ;
889 598 : transferred_this_segment = 0;
890 :
891 : /*
892 : * Inner loop to continue after a short read. We'll keep going until
893 : * we hit EOF rather than assuming that a short read means we hit the
894 : * end.
895 : */
896 : for (;;)
897 : {
898 0 : TRACE_POSTGRESQL_SMGR_MD_READ_START(forknum, blocknum,
899 : reln->smgr_rlocator.locator.spcOid,
900 : reln->smgr_rlocator.locator.dbOid,
901 : reln->smgr_rlocator.locator.relNumber,
902 : reln->smgr_rlocator.backend);
903 598 : nbytes = FileReadV(v->mdfd_vfd, iov, iovcnt, seekpos,
904 : WAIT_EVENT_DATA_FILE_READ);
905 : TRACE_POSTGRESQL_SMGR_MD_READ_DONE(forknum, blocknum,
906 : reln->smgr_rlocator.locator.spcOid,
907 : reln->smgr_rlocator.locator.dbOid,
908 : reln->smgr_rlocator.locator.relNumber,
909 : reln->smgr_rlocator.backend,
910 : nbytes,
911 : size_this_segment - transferred_this_segment);
912 :
913 : #ifdef SIMULATE_SHORT_READ
914 : nbytes = Min(nbytes, 4096);
915 : #endif
916 :
917 598 : if (nbytes < 0)
918 0 : ereport(ERROR,
919 : (errcode_for_file_access(),
920 : errmsg("could not read blocks %u..%u in file \"%s\": %m",
921 : blocknum,
922 : blocknum + nblocks_this_segment - 1,
923 : FilePathName(v->mdfd_vfd))));
924 :
925 598 : if (nbytes == 0)
926 : {
927 : /*
928 : * We are at or past EOF, or we read a partial block at EOF.
929 : * Normally this is an error; upper levels should never try to
930 : * read a nonexistent block. However, if zero_damaged_pages
931 : * is ON or we are InRecovery, we should instead return zeroes
932 : * without complaining. This allows, for example, the case of
933 : * trying to update a block that was later truncated away.
934 : *
935 : * NB: We think that this codepath is unreachable in recovery
936 : * and incomplete with zero_damaged_pages, as missing segments
937 : * are not created. Putting blocks into the buffer-pool that
938 : * do not exist on disk is rather problematic, as it will not
939 : * be found by scans that rely on smgrnblocks(), as they are
940 : * beyond EOF. It also can cause weird problems with relation
941 : * extension, as relation extension does not expect blocks
942 : * beyond EOF to exist.
943 : *
944 : * Therefore we do not want to copy the logic into
945 : * mdstartreadv(), where it would have to be more complicated
946 : * due to potential differences in the zero_damaged_pages
947 : * setting between the definer and completor of IO.
948 : *
949 : * For PG 18, we are putting an Assert(false) in mdreadv()
950 : * (triggering failures in assertion-enabled builds, but
951 : * continuing to work in production builds). Afterwards we
952 : * plan to remove this code entirely.
953 : */
954 0 : if (zero_damaged_pages || InRecovery)
955 : {
956 : Assert(false); /* see comment above */
957 :
958 0 : for (BlockNumber i = transferred_this_segment / BLCKSZ;
959 0 : i < nblocks_this_segment;
960 0 : ++i)
961 0 : memset(buffers[i], 0, BLCKSZ);
962 0 : break;
963 : }
964 : else
965 0 : ereport(ERROR,
966 : (errcode(ERRCODE_DATA_CORRUPTED),
967 : errmsg("could not read blocks %u..%u in file \"%s\": read only %zu of %zu bytes",
968 : blocknum,
969 : blocknum + nblocks_this_segment - 1,
970 : FilePathName(v->mdfd_vfd),
971 : transferred_this_segment,
972 : size_this_segment)));
973 : }
974 :
975 : /* One loop should usually be enough. */
976 598 : transferred_this_segment += nbytes;
977 : Assert(transferred_this_segment <= size_this_segment);
978 598 : if (transferred_this_segment == size_this_segment)
979 598 : break;
980 :
981 : /* Adjust position and vectors after a short read. */
982 0 : seekpos += nbytes;
983 0 : iovcnt = compute_remaining_iovec(iov, iov, iovcnt, nbytes);
984 : }
985 :
986 598 : nblocks -= nblocks_this_segment;
987 598 : buffers += nblocks_this_segment;
988 598 : blocknum += nblocks_this_segment;
989 : }
990 598 : }
991 :
992 : /*
993 : * mdstartreadv() -- Asynchronous version of mdreadv().
994 : */
995 : void
996 1303966 : mdstartreadv(PgAioHandle *ioh,
997 : SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
998 : void **buffers, BlockNumber nblocks)
999 : {
1000 : pgoff_t seekpos;
1001 : MdfdVec *v;
1002 : BlockNumber nblocks_this_segment;
1003 : struct iovec *iov;
1004 : int iovcnt;
1005 : int ret;
1006 :
1007 1303966 : v = _mdfd_getseg(reln, forknum, blocknum, false,
1008 : EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
1009 :
1010 1303951 : seekpos = (pgoff_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
1011 :
1012 : Assert(seekpos < (pgoff_t) BLCKSZ * RELSEG_SIZE);
1013 :
1014 1303951 : nblocks_this_segment =
1015 1303951 : Min(nblocks,
1016 : RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
1017 :
1018 1303951 : if (nblocks_this_segment != nblocks)
1019 0 : elog(ERROR, "read crossing segment boundary");
1020 :
1021 1303951 : iovcnt = pgaio_io_get_iovec(ioh, &iov);
1022 :
1023 : Assert(nblocks <= iovcnt);
1024 :
1025 1303951 : iovcnt = buffers_to_iovec(iov, buffers, nblocks_this_segment);
1026 :
1027 : Assert(iovcnt <= nblocks_this_segment);
1028 :
1029 1303951 : if (!(io_direct_flags & IO_DIRECT_DATA))
1030 1302552 : pgaio_io_set_flag(ioh, PGAIO_HF_BUFFERED);
1031 :
1032 1303951 : pgaio_io_set_target_smgr(ioh,
1033 : reln,
1034 : forknum,
1035 : blocknum,
1036 : nblocks,
1037 : false);
1038 1303951 : pgaio_io_register_callbacks(ioh, PGAIO_HCB_MD_READV, 0);
1039 :
1040 1303951 : ret = FileStartReadV(ioh, v->mdfd_vfd, iovcnt, seekpos, WAIT_EVENT_DATA_FILE_READ);
1041 1303951 : if (ret != 0)
1042 0 : ereport(ERROR,
1043 : (errcode_for_file_access(),
1044 : errmsg("could not start reading blocks %u..%u in file \"%s\": %m",
1045 : blocknum,
1046 : blocknum + nblocks_this_segment - 1,
1047 : FilePathName(v->mdfd_vfd))));
1048 :
1049 : /*
1050 : * The error checks corresponding to the post-read checks in mdreadv() are
1051 : * in md_readv_complete().
1052 : *
1053 : * However we chose, at least for now, to not implement the
1054 : * zero_damaged_pages logic present in mdreadv(). As outlined in mdreadv()
1055 : * that logic is rather problematic, and we want to get rid of it. Here
1056 : * equivalent logic would have to be more complicated due to potential
1057 : * differences in the zero_damaged_pages setting between the definer and
1058 : * completor of IO.
1059 : */
1060 1303951 : }
1061 :
1062 : /*
1063 : * mdwritev() -- Write the supplied blocks at the appropriate location.
1064 : *
1065 : * This is to be used only for updating already-existing blocks of a
1066 : * relation (ie, those before the current EOF). To extend a relation,
1067 : * use mdextend().
1068 : */
1069 : void
1070 589056 : mdwritev(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
1071 : const void **buffers, BlockNumber nblocks, bool skipFsync)
1072 : {
1073 : /* This assert is too expensive to have on normally ... */
1074 : #ifdef CHECK_WRITE_VS_EXTEND
1075 : Assert((uint64) blocknum + (uint64) nblocks <= (uint64) mdnblocks(reln, forknum));
1076 : #endif
1077 :
1078 1178112 : while (nblocks > 0)
1079 : {
1080 : struct iovec iov[PG_IOV_MAX];
1081 : int iovcnt;
1082 : pgoff_t seekpos;
1083 : int nbytes;
1084 : MdfdVec *v;
1085 : BlockNumber nblocks_this_segment;
1086 : size_t transferred_this_segment;
1087 : size_t size_this_segment;
1088 :
1089 589056 : v = _mdfd_getseg(reln, forknum, blocknum, skipFsync,
1090 : EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
1091 :
1092 589056 : seekpos = (pgoff_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
1093 :
1094 : Assert(seekpos < (pgoff_t) BLCKSZ * RELSEG_SIZE);
1095 :
1096 589056 : nblocks_this_segment =
1097 589056 : Min(nblocks,
1098 : RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
1099 589056 : nblocks_this_segment = Min(nblocks_this_segment, lengthof(iov));
1100 :
1101 589056 : if (nblocks_this_segment != nblocks)
1102 0 : elog(ERROR, "write crosses segment boundary");
1103 :
1104 589056 : iovcnt = buffers_to_iovec(iov, (void **) buffers, nblocks_this_segment);
1105 589056 : size_this_segment = nblocks_this_segment * BLCKSZ;
1106 589056 : transferred_this_segment = 0;
1107 :
1108 : /*
1109 : * Inner loop to continue after a short write. If the reason is that
1110 : * we're out of disk space, a future attempt should get an ENOSPC
1111 : * error from the kernel.
1112 : */
1113 : for (;;)
1114 : {
1115 0 : TRACE_POSTGRESQL_SMGR_MD_WRITE_START(forknum, blocknum,
1116 : reln->smgr_rlocator.locator.spcOid,
1117 : reln->smgr_rlocator.locator.dbOid,
1118 : reln->smgr_rlocator.locator.relNumber,
1119 : reln->smgr_rlocator.backend);
1120 589056 : nbytes = FileWriteV(v->mdfd_vfd, iov, iovcnt, seekpos,
1121 : WAIT_EVENT_DATA_FILE_WRITE);
1122 : TRACE_POSTGRESQL_SMGR_MD_WRITE_DONE(forknum, blocknum,
1123 : reln->smgr_rlocator.locator.spcOid,
1124 : reln->smgr_rlocator.locator.dbOid,
1125 : reln->smgr_rlocator.locator.relNumber,
1126 : reln->smgr_rlocator.backend,
1127 : nbytes,
1128 : size_this_segment - transferred_this_segment);
1129 :
1130 : #ifdef SIMULATE_SHORT_WRITE
1131 : nbytes = Min(nbytes, 4096);
1132 : #endif
1133 :
1134 589056 : if (nbytes < 0)
1135 : {
1136 0 : bool enospc = errno == ENOSPC;
1137 :
1138 0 : ereport(ERROR,
1139 : (errcode_for_file_access(),
1140 : errmsg("could not write blocks %u..%u in file \"%s\": %m",
1141 : blocknum,
1142 : blocknum + nblocks_this_segment - 1,
1143 : FilePathName(v->mdfd_vfd)),
1144 : enospc ? errhint("Check free disk space.") : 0));
1145 : }
1146 :
1147 : /* One loop should usually be enough. */
1148 589056 : transferred_this_segment += nbytes;
1149 : Assert(transferred_this_segment <= size_this_segment);
1150 589056 : if (transferred_this_segment == size_this_segment)
1151 589056 : break;
1152 :
1153 : /* Adjust position and iovecs after a short write. */
1154 0 : seekpos += nbytes;
1155 0 : iovcnt = compute_remaining_iovec(iov, iov, iovcnt, nbytes);
1156 : }
1157 :
1158 589056 : if (!skipFsync && !SmgrIsTemp(reln))
1159 585077 : register_dirty_segment(reln, forknum, v);
1160 :
1161 589056 : nblocks -= nblocks_this_segment;
1162 589056 : buffers += nblocks_this_segment;
1163 589056 : blocknum += nblocks_this_segment;
1164 : }
1165 589056 : }
1166 :
1167 :
1168 : /*
1169 : * mdwriteback() -- Tell the kernel to write pages back to storage.
1170 : *
1171 : * This accepts a range of blocks because flushing several pages at once is
1172 : * considerably more efficient than doing so individually.
1173 : */
1174 : void
1175 0 : mdwriteback(SMgrRelation reln, ForkNumber forknum,
1176 : BlockNumber blocknum, BlockNumber nblocks)
1177 : {
1178 : Assert((io_direct_flags & IO_DIRECT_DATA) == 0);
1179 :
1180 : /*
1181 : * Issue flush requests in as few requests as possible; have to split at
1182 : * segment boundaries though, since those are actually separate files.
1183 : */
1184 0 : while (nblocks > 0)
1185 : {
1186 0 : BlockNumber nflush = nblocks;
1187 : pgoff_t seekpos;
1188 : MdfdVec *v;
1189 : int segnum_start,
1190 : segnum_end;
1191 :
1192 0 : v = _mdfd_getseg(reln, forknum, blocknum, true /* not used */ ,
1193 : EXTENSION_DONT_OPEN);
1194 :
1195 : /*
1196 : * We might be flushing buffers of already removed relations, that's
1197 : * ok, just ignore that case. If the segment file wasn't open already
1198 : * (ie from a recent mdwrite()), then we don't want to re-open it, to
1199 : * avoid a race with PROCSIGNAL_BARRIER_SMGRRELEASE that might leave
1200 : * us with a descriptor to a file that is about to be unlinked.
1201 : */
1202 0 : if (!v)
1203 0 : return;
1204 :
1205 : /* compute offset inside the current segment */
1206 0 : segnum_start = blocknum / RELSEG_SIZE;
1207 :
1208 : /* compute number of desired writes within the current segment */
1209 0 : segnum_end = (blocknum + nblocks - 1) / RELSEG_SIZE;
1210 0 : if (segnum_start != segnum_end)
1211 0 : nflush = RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE));
1212 :
1213 : Assert(nflush >= 1);
1214 : Assert(nflush <= nblocks);
1215 :
1216 0 : seekpos = (pgoff_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
1217 :
1218 0 : FileWriteback(v->mdfd_vfd, seekpos, (pgoff_t) BLCKSZ * nflush, WAIT_EVENT_DATA_FILE_FLUSH);
1219 :
1220 0 : nblocks -= nflush;
1221 0 : blocknum += nflush;
1222 : }
1223 : }
1224 :
1225 : /*
1226 : * mdnblocks() -- Get the number of blocks stored in a relation.
1227 : *
1228 : * Important side effect: all active segments of the relation are opened
1229 : * and added to the md_seg_fds array. If this routine has not been
1230 : * called, then only segments up to the last one actually touched
1231 : * are present in the array.
1232 : */
1233 : BlockNumber
1234 2673118 : mdnblocks(SMgrRelation reln, ForkNumber forknum)
1235 : {
1236 : MdfdVec *v;
1237 : BlockNumber nblocks;
1238 : BlockNumber segno;
1239 :
1240 2673118 : mdopenfork(reln, forknum, EXTENSION_FAIL);
1241 :
1242 : /* mdopen has opened the first segment */
1243 : Assert(reln->md_num_open_segs[forknum] > 0);
1244 :
1245 : /*
1246 : * Start from the last open segments, to avoid redundant seeks. We have
1247 : * previously verified that these segments are exactly RELSEG_SIZE long,
1248 : * and it's useless to recheck that each time.
1249 : *
1250 : * NOTE: this assumption could only be wrong if another backend has
1251 : * truncated the relation. We rely on higher code levels to handle that
1252 : * scenario by closing and re-opening the md fd, which is handled via
1253 : * relcache flush. (Since the checkpointer doesn't participate in
1254 : * relcache flush, it could have segment entries for inactive segments;
1255 : * that's OK because the checkpointer never needs to compute relation
1256 : * size.)
1257 : */
1258 2673099 : segno = reln->md_num_open_segs[forknum] - 1;
1259 2673099 : v = &reln->md_seg_fds[forknum][segno];
1260 :
1261 : for (;;)
1262 : {
1263 2673099 : nblocks = _mdnblocks(reln, forknum, v);
1264 2673099 : if (nblocks > ((BlockNumber) RELSEG_SIZE))
1265 0 : elog(FATAL, "segment too big");
1266 2673099 : if (nblocks < ((BlockNumber) RELSEG_SIZE))
1267 2673099 : return (segno * ((BlockNumber) RELSEG_SIZE)) + nblocks;
1268 :
1269 : /*
1270 : * If segment is exactly RELSEG_SIZE, advance to next one.
1271 : */
1272 0 : segno++;
1273 :
1274 : /*
1275 : * We used to pass O_CREAT here, but that has the disadvantage that it
1276 : * might create a segment which has vanished through some operating
1277 : * system misadventure. In such a case, creating the segment here
1278 : * undermines _mdfd_getseg's attempts to notice and report an error
1279 : * upon access to a missing segment.
1280 : */
1281 0 : v = _mdfd_openseg(reln, forknum, segno, 0);
1282 0 : if (v == NULL)
1283 0 : return segno * ((BlockNumber) RELSEG_SIZE);
1284 : }
1285 : }
1286 :
1287 : /*
1288 : * mdtruncate() -- Truncate relation to specified number of blocks.
1289 : *
1290 : * Guaranteed not to allocate memory, so it can be used in a critical section.
1291 : * Caller must have called smgrnblocks() to obtain curnblk while holding a
1292 : * sufficient lock to prevent a change in relation size, and not used any smgr
1293 : * functions for this relation or handled interrupts in between. This makes
1294 : * sure we have opened all active segments, so that truncate loop will get
1295 : * them all!
1296 : *
1297 : * If nblocks > curnblk, the request is ignored when we are InRecovery,
1298 : * otherwise, an error is raised.
1299 : */
1300 : void
1301 914 : mdtruncate(SMgrRelation reln, ForkNumber forknum,
1302 : BlockNumber curnblk, BlockNumber nblocks)
1303 : {
1304 : BlockNumber priorblocks;
1305 : int curopensegs;
1306 :
1307 914 : if (nblocks > curnblk)
1308 : {
1309 : /* Bogus request ... but no complaint if InRecovery */
1310 0 : if (InRecovery)
1311 0 : return;
1312 0 : ereport(ERROR,
1313 : (errmsg("could not truncate file \"%s\" to %u blocks: it's only %u blocks now",
1314 : relpath(reln->smgr_rlocator, forknum).str,
1315 : nblocks, curnblk)));
1316 : }
1317 914 : if (nblocks == curnblk)
1318 357 : return; /* no work */
1319 :
1320 : /*
1321 : * Truncate segments, starting at the last one. Starting at the end makes
1322 : * managing the memory for the fd array easier, should there be errors.
1323 : */
1324 557 : curopensegs = reln->md_num_open_segs[forknum];
1325 1114 : while (curopensegs > 0)
1326 : {
1327 : MdfdVec *v;
1328 :
1329 557 : priorblocks = (curopensegs - 1) * RELSEG_SIZE;
1330 :
1331 557 : v = &reln->md_seg_fds[forknum][curopensegs - 1];
1332 :
1333 557 : if (priorblocks > nblocks)
1334 : {
1335 : /*
1336 : * This segment is no longer active. We truncate the file, but do
1337 : * not delete it, for reasons explained in the header comments.
1338 : */
1339 0 : if (FileTruncate(v->mdfd_vfd, 0, WAIT_EVENT_DATA_FILE_TRUNCATE) < 0)
1340 0 : ereport(ERROR,
1341 : (errcode_for_file_access(),
1342 : errmsg("could not truncate file \"%s\": %m",
1343 : FilePathName(v->mdfd_vfd))));
1344 :
1345 0 : if (!SmgrIsTemp(reln))
1346 0 : register_dirty_segment(reln, forknum, v);
1347 :
1348 : /* we never drop the 1st segment */
1349 : Assert(v != &reln->md_seg_fds[forknum][0]);
1350 :
1351 0 : FileClose(v->mdfd_vfd);
1352 0 : _fdvec_resize(reln, forknum, curopensegs - 1);
1353 : }
1354 557 : else if (priorblocks + ((BlockNumber) RELSEG_SIZE) > nblocks)
1355 : {
1356 : /*
1357 : * This is the last segment we want to keep. Truncate the file to
1358 : * the right length. NOTE: if nblocks is exactly a multiple K of
1359 : * RELSEG_SIZE, we will truncate the K+1st segment to 0 length but
1360 : * keep it. This adheres to the invariant given in the header
1361 : * comments.
1362 : */
1363 557 : BlockNumber lastsegblocks = nblocks - priorblocks;
1364 :
1365 557 : if (FileTruncate(v->mdfd_vfd, (pgoff_t) lastsegblocks * BLCKSZ, WAIT_EVENT_DATA_FILE_TRUNCATE) < 0)
1366 0 : ereport(ERROR,
1367 : (errcode_for_file_access(),
1368 : errmsg("could not truncate file \"%s\" to %u blocks: %m",
1369 : FilePathName(v->mdfd_vfd),
1370 : nblocks)));
1371 557 : if (!SmgrIsTemp(reln))
1372 392 : register_dirty_segment(reln, forknum, v);
1373 : }
1374 : else
1375 : {
1376 : /*
1377 : * We still need this segment, so nothing to do for this and any
1378 : * earlier segment.
1379 : */
1380 0 : break;
1381 : }
1382 557 : curopensegs--;
1383 : }
1384 : }
1385 :
1386 : /*
1387 : * mdregistersync() -- Mark whole relation as needing fsync
1388 : */
1389 : void
1390 25741 : mdregistersync(SMgrRelation reln, ForkNumber forknum)
1391 : {
1392 : int segno;
1393 : int min_inactive_seg;
1394 :
1395 : /*
1396 : * NOTE: mdnblocks makes sure we have opened all active segments, so that
1397 : * the loop below will get them all!
1398 : */
1399 25741 : mdnblocks(reln, forknum);
1400 :
1401 25741 : min_inactive_seg = segno = reln->md_num_open_segs[forknum];
1402 :
1403 : /*
1404 : * Temporarily open inactive segments, then close them after sync. There
1405 : * may be some inactive segments left opened after error, but that is
1406 : * harmless. We don't bother to clean them up and take a risk of further
1407 : * trouble. The next mdclose() will soon close them.
1408 : */
1409 25741 : while (_mdfd_openseg(reln, forknum, segno, 0) != NULL)
1410 0 : segno++;
1411 :
1412 51482 : while (segno > 0)
1413 : {
1414 25741 : MdfdVec *v = &reln->md_seg_fds[forknum][segno - 1];
1415 :
1416 25741 : register_dirty_segment(reln, forknum, v);
1417 :
1418 : /* Close inactive segments immediately */
1419 25741 : if (segno > min_inactive_seg)
1420 : {
1421 0 : FileClose(v->mdfd_vfd);
1422 0 : _fdvec_resize(reln, forknum, segno - 1);
1423 : }
1424 :
1425 25741 : segno--;
1426 : }
1427 25741 : }
1428 :
1429 : /*
1430 : * mdimmedsync() -- Immediately sync a relation to stable storage.
1431 : *
1432 : * Note that only writes already issued are synced; this routine knows
1433 : * nothing of dirty buffers that may exist inside the buffer manager. We
1434 : * sync active and inactive segments; smgrDoPendingSyncs() relies on this.
1435 : * Consider a relation skipping WAL. Suppose a checkpoint syncs blocks of
1436 : * some segment, then mdtruncate() renders that segment inactive. If we
1437 : * crash before the next checkpoint syncs the newly-inactive segment, that
1438 : * segment may survive recovery, reintroducing unwanted data into the table.
1439 : */
1440 : void
1441 14 : mdimmedsync(SMgrRelation reln, ForkNumber forknum)
1442 : {
1443 : int segno;
1444 : int min_inactive_seg;
1445 :
1446 : /*
1447 : * NOTE: mdnblocks makes sure we have opened all active segments, so that
1448 : * the loop below will get them all!
1449 : */
1450 14 : mdnblocks(reln, forknum);
1451 :
1452 14 : min_inactive_seg = segno = reln->md_num_open_segs[forknum];
1453 :
1454 : /*
1455 : * Temporarily open inactive segments, then close them after sync. There
1456 : * may be some inactive segments left opened after fsync() error, but that
1457 : * is harmless. We don't bother to clean them up and take a risk of
1458 : * further trouble. The next mdclose() will soon close them.
1459 : */
1460 14 : while (_mdfd_openseg(reln, forknum, segno, 0) != NULL)
1461 0 : segno++;
1462 :
1463 28 : while (segno > 0)
1464 : {
1465 14 : MdfdVec *v = &reln->md_seg_fds[forknum][segno - 1];
1466 :
1467 : /*
1468 : * fsyncs done through mdimmedsync() should be tracked in a separate
1469 : * IOContext than those done through mdsyncfiletag() to differentiate
1470 : * between unavoidable client backend fsyncs (e.g. those done during
1471 : * index build) and those which ideally would have been done by the
1472 : * checkpointer. Since other IO operations bypassing the buffer
1473 : * manager could also be tracked in such an IOContext, wait until
1474 : * these are also tracked to track immediate fsyncs.
1475 : */
1476 14 : if (FileSync(v->mdfd_vfd, WAIT_EVENT_DATA_FILE_IMMEDIATE_SYNC) < 0)
1477 0 : ereport(data_sync_elevel(ERROR),
1478 : (errcode_for_file_access(),
1479 : errmsg("could not fsync file \"%s\": %m",
1480 : FilePathName(v->mdfd_vfd))));
1481 :
1482 : /* Close inactive segments immediately */
1483 14 : if (segno > min_inactive_seg)
1484 : {
1485 0 : FileClose(v->mdfd_vfd);
1486 0 : _fdvec_resize(reln, forknum, segno - 1);
1487 : }
1488 :
1489 14 : segno--;
1490 : }
1491 14 : }
1492 :
1493 : int
1494 464681 : mdfd(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, uint32 *off)
1495 : {
1496 464681 : MdfdVec *v = mdopenfork(reln, forknum, EXTENSION_FAIL);
1497 :
1498 464681 : v = _mdfd_getseg(reln, forknum, blocknum, false,
1499 : EXTENSION_FAIL);
1500 :
1501 464681 : *off = (pgoff_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
1502 :
1503 : Assert(*off < (pgoff_t) BLCKSZ * RELSEG_SIZE);
1504 :
1505 464681 : return FileGetRawDesc(v->mdfd_vfd);
1506 : }
1507 :
1508 : /*
1509 : * register_dirty_segment() -- Mark a relation segment as needing fsync
1510 : *
1511 : * If there is a local pending-ops table, just make an entry in it for
1512 : * ProcessSyncRequests to process later. Otherwise, try to pass off the
1513 : * fsync request to the checkpointer process. If that fails, just do the
1514 : * fsync locally before returning (we hope this will not happen often
1515 : * enough to be a performance problem).
1516 : */
1517 : static void
1518 986142 : register_dirty_segment(SMgrRelation reln, ForkNumber forknum, MdfdVec *seg)
1519 : {
1520 : FileTag tag;
1521 :
1522 986142 : INIT_MD_FILETAG(tag, reln->smgr_rlocator.locator, forknum, seg->mdfd_segno);
1523 :
1524 : /* Temp relations should never be fsync'd */
1525 : Assert(!SmgrIsTemp(reln));
1526 :
1527 986142 : if (!RegisterSyncRequest(&tag, SYNC_REQUEST, false /* retryOnError */ ))
1528 : {
1529 : instr_time io_start;
1530 :
1531 563 : ereport(DEBUG1,
1532 : (errmsg_internal("could not forward fsync request because request queue is full")));
1533 :
1534 563 : io_start = pgstat_prepare_io_time(track_io_timing);
1535 :
1536 563 : if (FileSync(seg->mdfd_vfd, WAIT_EVENT_DATA_FILE_SYNC) < 0)
1537 0 : ereport(data_sync_elevel(ERROR),
1538 : (errcode_for_file_access(),
1539 : errmsg("could not fsync file \"%s\": %m",
1540 : FilePathName(seg->mdfd_vfd))));
1541 :
1542 : /*
1543 : * We have no way of knowing if the current IOContext is
1544 : * IOCONTEXT_NORMAL or IOCONTEXT_[BULKREAD, BULKWRITE, VACUUM] at this
1545 : * point, so count the fsync as being in the IOCONTEXT_NORMAL
1546 : * IOContext. This is probably okay, because the number of backend
1547 : * fsyncs doesn't say anything about the efficacy of the
1548 : * BufferAccessStrategy. And counting both fsyncs done in
1549 : * IOCONTEXT_NORMAL and IOCONTEXT_[BULKREAD, BULKWRITE, VACUUM] under
1550 : * IOCONTEXT_NORMAL is likely clearer when investigating the number of
1551 : * backend fsyncs.
1552 : */
1553 563 : pgstat_count_io_op_time(IOOBJECT_RELATION, IOCONTEXT_NORMAL,
1554 : IOOP_FSYNC, io_start, 1, 0);
1555 : }
1556 986142 : }
1557 :
1558 : /*
1559 : * register_unlink_segment() -- Schedule a file to be deleted after next checkpoint
1560 : */
1561 : static void
1562 36940 : register_unlink_segment(RelFileLocatorBackend rlocator, ForkNumber forknum,
1563 : BlockNumber segno)
1564 : {
1565 : FileTag tag;
1566 :
1567 36940 : INIT_MD_FILETAG(tag, rlocator.locator, forknum, segno);
1568 :
1569 : /* Should never be used with temp relations */
1570 : Assert(!RelFileLocatorBackendIsTemp(rlocator));
1571 :
1572 36940 : RegisterSyncRequest(&tag, SYNC_UNLINK_REQUEST, true /* retryOnError */ );
1573 36940 : }
1574 :
1575 : /*
1576 : * register_forget_request() -- forget any fsyncs for a relation fork's segment
1577 : */
1578 : static void
1579 143408 : register_forget_request(RelFileLocatorBackend rlocator, ForkNumber forknum,
1580 : BlockNumber segno)
1581 : {
1582 : FileTag tag;
1583 :
1584 143408 : INIT_MD_FILETAG(tag, rlocator.locator, forknum, segno);
1585 :
1586 143408 : RegisterSyncRequest(&tag, SYNC_FORGET_REQUEST, true /* retryOnError */ );
1587 143408 : }
1588 :
1589 : /*
1590 : * ForgetDatabaseSyncRequests -- forget any fsyncs and unlinks for a DB
1591 : */
1592 : void
1593 63 : ForgetDatabaseSyncRequests(Oid dbid)
1594 : {
1595 : FileTag tag;
1596 : RelFileLocator rlocator;
1597 :
1598 63 : rlocator.dbOid = dbid;
1599 63 : rlocator.spcOid = 0;
1600 63 : rlocator.relNumber = 0;
1601 :
1602 63 : INIT_MD_FILETAG(tag, rlocator, InvalidForkNumber, InvalidBlockNumber);
1603 :
1604 63 : RegisterSyncRequest(&tag, SYNC_FILTER_REQUEST, true /* retryOnError */ );
1605 63 : }
1606 :
1607 : /*
1608 : * DropRelationFiles -- drop files of all given relations
1609 : */
1610 : void
1611 2816 : DropRelationFiles(RelFileLocator *delrels, int ndelrels, bool isRedo)
1612 : {
1613 : SMgrRelation *srels;
1614 : int i;
1615 :
1616 2816 : srels = palloc_array(SMgrRelation, ndelrels);
1617 10881 : for (i = 0; i < ndelrels; i++)
1618 : {
1619 8065 : SMgrRelation srel = smgropen(delrels[i], INVALID_PROC_NUMBER);
1620 :
1621 8065 : if (isRedo)
1622 : {
1623 : ForkNumber fork;
1624 :
1625 40135 : for (fork = 0; fork <= MAX_FORKNUM; fork++)
1626 32108 : XLogDropRelation(delrels[i], fork);
1627 : }
1628 8065 : srels[i] = srel;
1629 : }
1630 :
1631 2816 : smgrdounlinkall(srels, ndelrels, isRedo);
1632 :
1633 10881 : for (i = 0; i < ndelrels; i++)
1634 8065 : smgrclose(srels[i]);
1635 2816 : pfree(srels);
1636 2816 : }
1637 :
1638 :
1639 : /*
1640 : * _fdvec_resize() -- Resize the fork's open segments array
1641 : */
1642 : static void
1643 1706044 : _fdvec_resize(SMgrRelation reln,
1644 : ForkNumber forknum,
1645 : int nseg)
1646 : {
1647 1706044 : if (nseg == 0)
1648 : {
1649 570799 : if (reln->md_num_open_segs[forknum] > 0)
1650 : {
1651 570799 : pfree(reln->md_seg_fds[forknum]);
1652 570799 : reln->md_seg_fds[forknum] = NULL;
1653 : }
1654 : }
1655 1135245 : else if (reln->md_num_open_segs[forknum] == 0)
1656 : {
1657 1135245 : reln->md_seg_fds[forknum] =
1658 1135245 : MemoryContextAlloc(MdCxt, sizeof(MdfdVec) * nseg);
1659 : }
1660 0 : else if (nseg > reln->md_num_open_segs[forknum])
1661 : {
1662 : /*
1663 : * It doesn't seem worthwhile complicating the code to amortize
1664 : * repalloc() calls. Those are far faster than PathNameOpenFile() or
1665 : * FileClose(), and the memory context internally will sometimes avoid
1666 : * doing an actual reallocation.
1667 : */
1668 0 : reln->md_seg_fds[forknum] =
1669 0 : repalloc(reln->md_seg_fds[forknum],
1670 : sizeof(MdfdVec) * nseg);
1671 : }
1672 : else
1673 : {
1674 : /*
1675 : * We don't reallocate a smaller array, because we want mdtruncate()
1676 : * to be able to promise that it won't allocate memory, so that it is
1677 : * allowed in a critical section. This means that a bit of space in
1678 : * the array is now wasted, until the next time we add a segment and
1679 : * reallocate.
1680 : */
1681 : }
1682 :
1683 1706044 : reln->md_num_open_segs[forknum] = nseg;
1684 1706044 : }
1685 :
1686 : /*
1687 : * Return the filename for the specified segment of the relation. The
1688 : * returned string is palloc'd.
1689 : */
1690 : static MdPathStr
1691 25767 : _mdfd_segpath(SMgrRelation reln, ForkNumber forknum, BlockNumber segno)
1692 : {
1693 : RelPathStr path;
1694 : MdPathStr fullpath;
1695 :
1696 25767 : path = relpath(reln->smgr_rlocator, forknum);
1697 :
1698 25767 : if (segno > 0)
1699 25767 : sprintf(fullpath.str, "%s.%u", path.str, segno);
1700 : else
1701 0 : strcpy(fullpath.str, path.str);
1702 :
1703 25767 : return fullpath;
1704 : }
1705 :
1706 : /*
1707 : * Open the specified segment of the relation,
1708 : * and make a MdfdVec object for it. Returns NULL on failure.
1709 : */
1710 : static MdfdVec *
1711 25755 : _mdfd_openseg(SMgrRelation reln, ForkNumber forknum, BlockNumber segno,
1712 : int oflags)
1713 : {
1714 : MdfdVec *v;
1715 : File fd;
1716 : MdPathStr fullpath;
1717 :
1718 25755 : fullpath = _mdfd_segpath(reln, forknum, segno);
1719 :
1720 : /* open the file */
1721 25755 : fd = PathNameOpenFile(fullpath.str, _mdfd_open_flags() | oflags);
1722 :
1723 25755 : if (fd < 0)
1724 25755 : return NULL;
1725 :
1726 : /*
1727 : * Segments are always opened in order from lowest to highest, so we must
1728 : * be adding a new one at the end.
1729 : */
1730 : Assert(segno == reln->md_num_open_segs[forknum]);
1731 :
1732 0 : _fdvec_resize(reln, forknum, segno + 1);
1733 :
1734 : /* fill the entry */
1735 0 : v = &reln->md_seg_fds[forknum][segno];
1736 0 : v->mdfd_vfd = fd;
1737 0 : v->mdfd_segno = segno;
1738 :
1739 : Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
1740 :
1741 : /* all done */
1742 0 : return v;
1743 : }
1744 :
1745 : /*
1746 : * _mdfd_getseg() -- Find the segment of the relation holding the
1747 : * specified block.
1748 : *
1749 : * If the segment doesn't exist, we ereport, return NULL, or create the
1750 : * segment, according to "behavior". Note: skipFsync is only used in the
1751 : * EXTENSION_CREATE case.
1752 : */
1753 : static MdfdVec *
1754 2705892 : _mdfd_getseg(SMgrRelation reln, ForkNumber forknum, BlockNumber blkno,
1755 : bool skipFsync, int behavior)
1756 : {
1757 : MdfdVec *v;
1758 : BlockNumber targetseg;
1759 : BlockNumber nextsegno;
1760 :
1761 : /* some way to handle non-existent segments needs to be specified */
1762 : Assert(behavior &
1763 : (EXTENSION_FAIL | EXTENSION_CREATE | EXTENSION_RETURN_NULL |
1764 : EXTENSION_DONT_OPEN));
1765 :
1766 2705892 : targetseg = blkno / ((BlockNumber) RELSEG_SIZE);
1767 :
1768 : /* if an existing and opened segment, we're done */
1769 2705892 : if (targetseg < reln->md_num_open_segs[forknum])
1770 : {
1771 2462034 : v = &reln->md_seg_fds[forknum][targetseg];
1772 2462034 : return v;
1773 : }
1774 :
1775 : /* The caller only wants the segment if we already had it open. */
1776 243858 : if (behavior & EXTENSION_DONT_OPEN)
1777 0 : return NULL;
1778 :
1779 : /*
1780 : * The target segment is not yet open. Iterate over all the segments
1781 : * between the last opened and the target segment. This way missing
1782 : * segments either raise an error, or get created (according to
1783 : * 'behavior'). Start with either the last opened, or the first segment if
1784 : * none was opened before.
1785 : */
1786 243858 : if (reln->md_num_open_segs[forknum] > 0)
1787 12 : v = &reln->md_seg_fds[forknum][reln->md_num_open_segs[forknum] - 1];
1788 : else
1789 : {
1790 243846 : v = mdopenfork(reln, forknum, behavior);
1791 243843 : if (!v)
1792 0 : return NULL; /* if behavior & EXTENSION_RETURN_NULL */
1793 : }
1794 :
1795 243855 : for (nextsegno = reln->md_num_open_segs[forknum];
1796 243855 : nextsegno <= targetseg; nextsegno++)
1797 : {
1798 12 : BlockNumber nblocks = _mdnblocks(reln, forknum, v);
1799 12 : int flags = 0;
1800 :
1801 : Assert(nextsegno == v->mdfd_segno + 1);
1802 :
1803 12 : if (nblocks > ((BlockNumber) RELSEG_SIZE))
1804 0 : elog(FATAL, "segment too big");
1805 :
1806 12 : if ((behavior & EXTENSION_CREATE) ||
1807 12 : (InRecovery && (behavior & EXTENSION_CREATE_RECOVERY)))
1808 : {
1809 : /*
1810 : * Normally we will create new segments only if authorized by the
1811 : * caller (i.e., we are doing mdextend()). But when doing WAL
1812 : * recovery, create segments anyway; this allows cases such as
1813 : * replaying WAL data that has a write into a high-numbered
1814 : * segment of a relation that was later deleted. We want to go
1815 : * ahead and create the segments so we can finish out the replay.
1816 : *
1817 : * We have to maintain the invariant that segments before the last
1818 : * active segment are of size RELSEG_SIZE; therefore, if
1819 : * extending, pad them out with zeroes if needed. (This only
1820 : * matters if in recovery, or if the caller is extending the
1821 : * relation discontiguously, but that can happen in hash indexes.)
1822 : */
1823 0 : if (nblocks < ((BlockNumber) RELSEG_SIZE))
1824 : {
1825 0 : char *zerobuf = palloc_aligned(BLCKSZ, PG_IO_ALIGN_SIZE,
1826 : MCXT_ALLOC_ZERO);
1827 :
1828 0 : mdextend(reln, forknum,
1829 0 : nextsegno * ((BlockNumber) RELSEG_SIZE) - 1,
1830 : zerobuf, skipFsync);
1831 0 : pfree(zerobuf);
1832 : }
1833 0 : flags = O_CREAT;
1834 : }
1835 12 : else if (nblocks < ((BlockNumber) RELSEG_SIZE))
1836 : {
1837 : /*
1838 : * When not extending, only open the next segment if the current
1839 : * one is exactly RELSEG_SIZE. If not (this branch), either
1840 : * return NULL or fail.
1841 : */
1842 12 : if (behavior & EXTENSION_RETURN_NULL)
1843 : {
1844 : /*
1845 : * Some callers discern between reasons for _mdfd_getseg()
1846 : * returning NULL based on errno. As there's no failing
1847 : * syscall involved in this case, explicitly set errno to
1848 : * ENOENT, as that seems the closest interpretation.
1849 : */
1850 0 : errno = ENOENT;
1851 0 : return NULL;
1852 : }
1853 :
1854 12 : ereport(ERROR,
1855 : (errcode_for_file_access(),
1856 : errmsg("could not open file \"%s\" (target block %u): previous segment is only %u blocks",
1857 : _mdfd_segpath(reln, forknum, nextsegno).str,
1858 : blkno, nblocks)));
1859 : }
1860 :
1861 0 : v = _mdfd_openseg(reln, forknum, nextsegno, flags);
1862 :
1863 0 : if (v == NULL)
1864 : {
1865 0 : if ((behavior & EXTENSION_RETURN_NULL) &&
1866 0 : FILE_POSSIBLY_DELETED(errno))
1867 0 : return NULL;
1868 0 : ereport(ERROR,
1869 : (errcode_for_file_access(),
1870 : errmsg("could not open file \"%s\" (target block %u): %m",
1871 : _mdfd_segpath(reln, forknum, nextsegno).str,
1872 : blkno)));
1873 : }
1874 : }
1875 :
1876 243843 : return v;
1877 : }
1878 :
1879 : /*
1880 : * Get number of blocks present in a single disk file
1881 : */
1882 : static BlockNumber
1883 2673111 : _mdnblocks(SMgrRelation reln, ForkNumber forknum, MdfdVec *seg)
1884 : {
1885 : pgoff_t len;
1886 :
1887 2673111 : len = FileSize(seg->mdfd_vfd);
1888 2673111 : if (len < 0)
1889 0 : ereport(ERROR,
1890 : (errcode_for_file_access(),
1891 : errmsg("could not seek to end of file \"%s\": %m",
1892 : FilePathName(seg->mdfd_vfd))));
1893 : /* note that this calculation will ignore any partial block at EOF */
1894 2673111 : return (BlockNumber) (len / BLCKSZ);
1895 : }
1896 :
1897 : /*
1898 : * Sync a file to disk, given a file tag. Write the path into an output
1899 : * buffer so the caller can use it in error messages.
1900 : *
1901 : * Return 0 on success, -1 on failure, with errno set.
1902 : */
1903 : int
1904 0 : mdsyncfiletag(const FileTag *ftag, char *path)
1905 : {
1906 0 : SMgrRelation reln = smgropen(ftag->rlocator, INVALID_PROC_NUMBER);
1907 : File file;
1908 : instr_time io_start;
1909 : bool need_to_close;
1910 : int result,
1911 : save_errno;
1912 :
1913 : /* See if we already have the file open, or need to open it. */
1914 0 : if (ftag->segno < reln->md_num_open_segs[ftag->forknum])
1915 : {
1916 0 : file = reln->md_seg_fds[ftag->forknum][ftag->segno].mdfd_vfd;
1917 0 : strlcpy(path, FilePathName(file), MAXPGPATH);
1918 0 : need_to_close = false;
1919 : }
1920 : else
1921 : {
1922 : MdPathStr p;
1923 :
1924 0 : p = _mdfd_segpath(reln, ftag->forknum, ftag->segno);
1925 0 : strlcpy(path, p.str, MD_PATH_STR_MAXLEN);
1926 :
1927 0 : file = PathNameOpenFile(path, _mdfd_open_flags());
1928 0 : if (file < 0)
1929 0 : return -1;
1930 0 : need_to_close = true;
1931 : }
1932 :
1933 0 : io_start = pgstat_prepare_io_time(track_io_timing);
1934 :
1935 : /* Sync the file. */
1936 0 : result = FileSync(file, WAIT_EVENT_DATA_FILE_SYNC);
1937 0 : save_errno = errno;
1938 :
1939 0 : if (need_to_close)
1940 0 : FileClose(file);
1941 :
1942 0 : pgstat_count_io_op_time(IOOBJECT_RELATION, IOCONTEXT_NORMAL,
1943 : IOOP_FSYNC, io_start, 1, 0);
1944 :
1945 0 : errno = save_errno;
1946 0 : return result;
1947 : }
1948 :
1949 : /*
1950 : * Unlink a file, given a file tag. Write the path into an output
1951 : * buffer so the caller can use it in error messages.
1952 : *
1953 : * Return 0 on success, -1 on failure, with errno set.
1954 : */
1955 : int
1956 34972 : mdunlinkfiletag(const FileTag *ftag, char *path)
1957 : {
1958 : RelPathStr p;
1959 :
1960 : /* Compute the path. */
1961 34972 : p = relpathperm(ftag->rlocator, MAIN_FORKNUM);
1962 34972 : strlcpy(path, p.str, MAXPGPATH);
1963 :
1964 : /* Try to unlink the file. */
1965 34972 : return unlink(path);
1966 : }
1967 :
1968 : /*
1969 : * Check if a given candidate request matches a given tag, when processing
1970 : * a SYNC_FILTER_REQUEST request. This will be called for all pending
1971 : * requests to find out whether to forget them.
1972 : */
1973 : bool
1974 7244 : mdfiletagmatches(const FileTag *ftag, const FileTag *candidate)
1975 : {
1976 : /*
1977 : * For now we only use filter requests as a way to drop all scheduled
1978 : * callbacks relating to a given database, when dropping the database.
1979 : * We'll return true for all candidates that have the same database OID as
1980 : * the ftag from the SYNC_FILTER_REQUEST request, so they're forgotten.
1981 : */
1982 7244 : return ftag->rlocator.dbOid == candidate->rlocator.dbOid;
1983 : }
1984 :
1985 : /*
1986 : * AIO completion callback for mdstartreadv().
1987 : */
1988 : static PgAioResult
1989 1198512 : md_readv_complete(PgAioHandle *ioh, PgAioResult prior_result, uint8 cb_data)
1990 : {
1991 1198512 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
1992 1198512 : PgAioResult result = prior_result;
1993 :
1994 1198512 : if (prior_result.result < 0)
1995 : {
1996 7 : result.status = PGAIO_RS_ERROR;
1997 7 : result.id = PGAIO_HCB_MD_READV;
1998 : /* For "hard" errors, track the error number in error_data */
1999 7 : result.error_data = -prior_result.result;
2000 7 : result.result = 0;
2001 :
2002 : /*
2003 : * Immediately log a message about the IO error, but only to the
2004 : * server log. The reason to do so immediately is that the originator
2005 : * might not process the query result immediately (because it is busy
2006 : * doing another part of query processing) or at all (e.g. if it was
2007 : * cancelled or errored out due to another IO also failing). The
2008 : * definer of the IO will emit an ERROR when processing the IO's
2009 : * results
2010 : */
2011 7 : pgaio_result_report(result, td, LOG_SERVER_ONLY);
2012 :
2013 7 : return result;
2014 : }
2015 :
2016 : /*
2017 : * As explained above smgrstartreadv(), the smgr API operates on the level
2018 : * of blocks, rather than bytes. Convert.
2019 : */
2020 1198505 : result.result /= BLCKSZ;
2021 :
2022 : Assert(result.result <= td->smgr.nblocks);
2023 :
2024 1198505 : if (result.result == 0)
2025 : {
2026 : /* consider 0 blocks read a failure */
2027 2 : result.status = PGAIO_RS_ERROR;
2028 2 : result.id = PGAIO_HCB_MD_READV;
2029 2 : result.error_data = 0;
2030 :
2031 : /* see comment above the "hard error" case */
2032 2 : pgaio_result_report(result, td, LOG_SERVER_ONLY);
2033 :
2034 2 : return result;
2035 : }
2036 :
2037 1198503 : if (result.status != PGAIO_RS_ERROR &&
2038 1198503 : result.result < td->smgr.nblocks)
2039 : {
2040 : /* partial reads should be retried at upper level */
2041 12 : result.status = PGAIO_RS_PARTIAL;
2042 12 : result.id = PGAIO_HCB_MD_READV;
2043 : }
2044 :
2045 1198503 : return result;
2046 : }
2047 :
2048 : /*
2049 : * AIO error reporting callback for mdstartreadv().
2050 : *
2051 : * Errors are encoded as follows:
2052 : * - PgAioResult.error_data != 0 encodes IO that failed with that errno
2053 : * - PgAioResult.error_data == 0 encodes IO that didn't read all data
2054 : */
2055 : static void
2056 28 : md_readv_report(PgAioResult result, const PgAioTargetData *td, int elevel)
2057 : {
2058 : RelPathStr path;
2059 :
2060 28 : path = relpathbackend(td->smgr.rlocator,
2061 : td->smgr.is_temp ? MyProcNumber : INVALID_PROC_NUMBER,
2062 : td->smgr.forkNum);
2063 :
2064 28 : if (result.error_data != 0)
2065 : {
2066 : /* for errcode_for_file_access() and %m */
2067 14 : errno = result.error_data;
2068 :
2069 14 : ereport(elevel,
2070 : errcode_for_file_access(),
2071 : errmsg("could not read blocks %u..%u in file \"%s\": %m",
2072 : td->smgr.blockNum,
2073 : td->smgr.blockNum + td->smgr.nblocks - 1,
2074 : path.str));
2075 : }
2076 : else
2077 : {
2078 : /*
2079 : * NB: This will typically only be output in debug messages, while
2080 : * retrying a partial IO.
2081 : */
2082 14 : ereport(elevel,
2083 : errcode(ERRCODE_DATA_CORRUPTED),
2084 : errmsg("could not read blocks %u..%u in file \"%s\": read only %zu of %zu bytes",
2085 : td->smgr.blockNum,
2086 : td->smgr.blockNum + td->smgr.nblocks - 1,
2087 : path.str,
2088 : result.result * (size_t) BLCKSZ,
2089 : td->smgr.nblocks * (size_t) BLCKSZ));
2090 : }
2091 19 : }
|
