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) 1996-2025, 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 3250844 : _mdfd_open_flags(void)
177 : {
178 3250844 : int flags = O_RDWR | PG_BINARY;
179 :
180 3250844 : if (io_direct_flags & IO_DIRECT_DATA)
181 618 : flags |= PG_O_DIRECT;
182 :
183 3250844 : return flags;
184 : }
185 :
186 : /*
187 : * mdinit() -- Initialize private state for magnetic disk storage manager.
188 : */
189 : void
190 45466 : mdinit(void)
191 : {
192 45466 : MdCxt = AllocSetContextCreate(TopMemoryContext,
193 : "MdSmgr",
194 : ALLOCSET_DEFAULT_SIZES);
195 45466 : }
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 1387386 : 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 1387386 : if (!InRecovery)
211 1345340 : mdclose(reln, forknum);
212 :
213 1387386 : 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 11342462 : mdcreate(SMgrRelation reln, ForkNumber forknum, bool isRedo)
223 : {
224 : MdfdVec *mdfd;
225 : RelPathStr path;
226 : File fd;
227 :
228 11342462 : if (isRedo && reln->md_num_open_segs[forknum] > 0)
229 11003852 : 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 338610 : TablespaceCreateDbspace(reln->smgr_rlocator.locator.spcOid,
243 : reln->smgr_rlocator.locator.dbOid,
244 : isRedo);
245 :
246 338610 : path = relpath(reln->smgr_rlocator, forknum);
247 :
248 338610 : fd = PathNameOpenFile(path.str, _mdfd_open_flags() | O_CREAT | O_EXCL);
249 :
250 338610 : if (fd < 0)
251 : {
252 25194 : int save_errno = errno;
253 :
254 25194 : if (isRedo)
255 25194 : fd = PathNameOpenFile(path.str, _mdfd_open_flags());
256 25194 : 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 338610 : _fdvec_resize(reln, forknum, 1);
267 338610 : mdfd = &reln->md_seg_fds[forknum][0];
268 338610 : mdfd->mdfd_vfd = fd;
269 338610 : mdfd->mdfd_segno = 0;
270 :
271 338610 : if (!SmgrIsTemp(reln))
272 331924 : 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 368216 : mdunlink(RelFileLocatorBackend rlocator, ForkNumber forknum, bool isRedo)
338 : {
339 : /* Now do the per-fork work */
340 368216 : if (forknum == InvalidForkNumber)
341 : {
342 0 : for (forknum = 0; forknum <= MAX_FORKNUM; forknum++)
343 0 : mdunlinkfork(rlocator, forknum, isRedo);
344 : }
345 : else
346 368216 : mdunlinkfork(rlocator, forknum, isRedo);
347 368216 : }
348 :
349 : /*
350 : * Truncate a file to release disk space.
351 : */
352 : static int
353 432342 : do_truncate(const char *path)
354 : {
355 : int save_errno;
356 : int ret;
357 :
358 432342 : ret = pg_truncate(path, 0);
359 :
360 : /* Log a warning here to avoid repetition in callers. */
361 432342 : 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 432342 : return ret;
371 : }
372 :
373 : static void
374 368216 : mdunlinkfork(RelFileLocatorBackend rlocator, ForkNumber forknum, bool isRedo)
375 : {
376 : RelPathStr path;
377 : int ret;
378 : int save_errno;
379 :
380 368216 : 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 368216 : if (isRedo || IsBinaryUpgrade || forknum != MAIN_FORKNUM ||
387 76660 : RelFileLocatorBackendIsTemp(rlocator))
388 : {
389 297904 : if (!RelFileLocatorBackendIsTemp(rlocator))
390 : {
391 : /* Prevent other backends' fds from holding on to the disk space */
392 272512 : ret = do_truncate(path.str);
393 :
394 : /* Forget any pending sync requests for the first segment */
395 272512 : save_errno = errno;
396 272512 : register_forget_request(rlocator, forknum, 0 /* first seg */ );
397 272512 : errno = save_errno;
398 : }
399 : else
400 25392 : ret = 0;
401 :
402 : /* Next unlink the file, unless it was already found to be missing */
403 297904 : if (ret >= 0 || errno != ENOENT)
404 : {
405 44606 : ret = unlink(path.str);
406 44606 : 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 70312 : ret = do_truncate(path.str);
420 :
421 : /* Register request to unlink first segment later */
422 70312 : save_errno = errno;
423 70312 : register_unlink_segment(rlocator, forknum, 0 /* first seg */ );
424 70312 : 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 368216 : if (ret >= 0 || errno != ENOENT)
440 : {
441 : MdPathStr segpath;
442 : BlockNumber segno;
443 :
444 96204 : for (segno = 1;; segno++)
445 : {
446 96204 : sprintf(segpath.str, "%s.%u", path.str, segno);
447 :
448 96204 : 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 89518 : if (do_truncate(segpath.str) < 0 && errno == ENOENT)
455 89518 : 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 6686 : if (unlink(segpath.str) < 0)
465 : {
466 : /* ENOENT is expected after the last segment... */
467 6686 : if (errno != ENOENT)
468 0 : ereport(WARNING,
469 : (errcode_for_file_access(),
470 : errmsg("could not remove file \"%s\": %m", segpath.str)));
471 6686 : break;
472 : }
473 : }
474 : }
475 368216 : }
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 241546 : mdextend(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
488 : const void *buffer, bool skipFsync)
489 : {
490 : off_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 241546 : 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 241546 : v = _mdfd_getseg(reln, forknum, blocknum, skipFsync, EXTENSION_CREATE);
517 :
518 241546 : seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
519 :
520 : Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
521 :
522 241546 : 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 241546 : if (!skipFsync && !SmgrIsTemp(reln))
540 56 : register_dirty_segment(reln, forknum, v);
541 :
542 : Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
543 241546 : }
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 424682 : mdzeroextend(SMgrRelation reln, ForkNumber forknum,
553 : BlockNumber blocknum, int nblocks, bool skipFsync)
554 : {
555 : MdfdVec *v;
556 424682 : BlockNumber curblocknum = blocknum;
557 424682 : 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 424682 : 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 849364 : while (remblocks > 0)
579 : {
580 424682 : BlockNumber segstartblock = curblocknum % ((BlockNumber) RELSEG_SIZE);
581 424682 : off_t seekpos = (off_t) BLCKSZ * segstartblock;
582 : int numblocks;
583 :
584 424682 : if (segstartblock + remblocks > RELSEG_SIZE)
585 0 : numblocks = RELSEG_SIZE - segstartblock;
586 : else
587 424682 : numblocks = remblocks;
588 :
589 424682 : 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 424682 : if (numblocks > 8)
606 : {
607 : int ret;
608 :
609 1014 : ret = FileFallocate(v->mdfd_vfd,
610 : seekpos, (off_t) BLCKSZ * numblocks,
611 : WAIT_EVENT_DATA_FILE_EXTEND);
612 1014 : if (ret != 0)
613 : {
614 0 : ereport(ERROR,
615 : errcode_for_file_access(),
616 : errmsg("could not extend file \"%s\" with FileFallocate(): %m",
617 : FilePathName(v->mdfd_vfd)),
618 : errhint("Check free disk space."));
619 : }
620 : }
621 : else
622 : {
623 : int ret;
624 :
625 : /*
626 : * Even if we don't want to use fallocate, we can still extend a
627 : * bit more efficiently than writing each 8kB block individually.
628 : * pg_pwrite_zeros() (via FileZero()) uses pg_pwritev_with_retry()
629 : * to avoid multiple writes or needing a zeroed buffer for the
630 : * whole length of the extension.
631 : */
632 423668 : ret = FileZero(v->mdfd_vfd,
633 : seekpos, (off_t) BLCKSZ * numblocks,
634 : WAIT_EVENT_DATA_FILE_EXTEND);
635 423668 : if (ret < 0)
636 0 : ereport(ERROR,
637 : errcode_for_file_access(),
638 : errmsg("could not extend file \"%s\": %m",
639 : FilePathName(v->mdfd_vfd)),
640 : errhint("Check free disk space."));
641 : }
642 :
643 424682 : if (!skipFsync && !SmgrIsTemp(reln))
644 401872 : register_dirty_segment(reln, forknum, v);
645 :
646 : Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
647 :
648 424682 : remblocks -= numblocks;
649 424682 : curblocknum += numblocks;
650 : }
651 424682 : }
652 :
653 : /*
654 : * mdopenfork() -- Open one fork of the specified relation.
655 : *
656 : * Note we only open the first segment, when there are multiple segments.
657 : *
658 : * If first segment is not present, either ereport or return NULL according
659 : * to "behavior". We treat EXTENSION_CREATE the same as EXTENSION_FAIL;
660 : * EXTENSION_CREATE means it's OK to extend an existing relation, not to
661 : * invent one out of whole cloth.
662 : */
663 : static MdfdVec *
664 7620250 : mdopenfork(SMgrRelation reln, ForkNumber forknum, int behavior)
665 : {
666 : MdfdVec *mdfd;
667 : RelPathStr path;
668 : File fd;
669 :
670 : /* No work if already open */
671 7620250 : if (reln->md_num_open_segs[forknum] > 0)
672 4783178 : return &reln->md_seg_fds[forknum][0];
673 :
674 2837072 : path = relpath(reln->smgr_rlocator, forknum);
675 :
676 2837072 : fd = PathNameOpenFile(path.str, _mdfd_open_flags());
677 :
678 2837072 : if (fd < 0)
679 : {
680 875634 : if ((behavior & EXTENSION_RETURN_NULL) &&
681 875590 : FILE_POSSIBLY_DELETED(errno))
682 875590 : return NULL;
683 44 : ereport(ERROR,
684 : (errcode_for_file_access(),
685 : errmsg("could not open file \"%s\": %m", path.str)));
686 : }
687 :
688 1961438 : _fdvec_resize(reln, forknum, 1);
689 1961438 : mdfd = &reln->md_seg_fds[forknum][0];
690 1961438 : mdfd->mdfd_vfd = fd;
691 1961438 : mdfd->mdfd_segno = 0;
692 :
693 : Assert(_mdnblocks(reln, forknum, mdfd) <= ((BlockNumber) RELSEG_SIZE));
694 :
695 1961438 : return mdfd;
696 : }
697 :
698 : /*
699 : * mdopen() -- Initialize newly-opened relation.
700 : */
701 : void
702 2599216 : mdopen(SMgrRelation reln)
703 : {
704 : /* mark it not open */
705 12996080 : for (int forknum = 0; forknum <= MAX_FORKNUM; forknum++)
706 10396864 : reln->md_num_open_segs[forknum] = 0;
707 2599216 : }
708 :
709 : /*
710 : * mdclose() -- Close the specified relation, if it isn't closed already.
711 : */
712 : void
713 7610428 : mdclose(SMgrRelation reln, ForkNumber forknum)
714 : {
715 7610428 : int nopensegs = reln->md_num_open_segs[forknum];
716 :
717 : /* No work if already closed */
718 7610428 : if (nopensegs == 0)
719 6485368 : return;
720 :
721 : /* close segments starting from the end */
722 2250120 : while (nopensegs > 0)
723 : {
724 1125060 : MdfdVec *v = &reln->md_seg_fds[forknum][nopensegs - 1];
725 :
726 1125060 : FileClose(v->mdfd_vfd);
727 1125060 : _fdvec_resize(reln, forknum, nopensegs - 1);
728 1125060 : nopensegs--;
729 : }
730 : }
731 :
732 : /*
733 : * mdprefetch() -- Initiate asynchronous read of the specified blocks of a relation
734 : */
735 : bool
736 17138 : mdprefetch(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
737 : int nblocks)
738 : {
739 : #ifdef USE_PREFETCH
740 :
741 : Assert((io_direct_flags & IO_DIRECT_DATA) == 0);
742 :
743 17138 : if ((uint64) blocknum + nblocks > (uint64) MaxBlockNumber + 1)
744 0 : return false;
745 :
746 34276 : while (nblocks > 0)
747 : {
748 : off_t seekpos;
749 : MdfdVec *v;
750 : int nblocks_this_segment;
751 :
752 17138 : v = _mdfd_getseg(reln, forknum, blocknum, false,
753 17138 : InRecovery ? EXTENSION_RETURN_NULL : EXTENSION_FAIL);
754 17138 : if (v == NULL)
755 0 : return false;
756 :
757 17138 : seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
758 :
759 : Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
760 :
761 17138 : nblocks_this_segment =
762 17138 : Min(nblocks,
763 : RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
764 :
765 17138 : (void) FilePrefetch(v->mdfd_vfd, seekpos, BLCKSZ * nblocks_this_segment,
766 : WAIT_EVENT_DATA_FILE_PREFETCH);
767 :
768 17138 : blocknum += nblocks_this_segment;
769 17138 : nblocks -= nblocks_this_segment;
770 : }
771 : #endif /* USE_PREFETCH */
772 :
773 17138 : return true;
774 : }
775 :
776 : /*
777 : * Convert an array of buffer address into an array of iovec objects, and
778 : * return the number that were required. 'iov' must have enough space for up
779 : * to 'nblocks' elements, but the number used may be less depending on
780 : * merging. In the case of a run of fully contiguous buffers, a single iovec
781 : * will be populated that can be handled as a plain non-vectored I/O.
782 : */
783 : static int
784 3643196 : buffers_to_iovec(struct iovec *iov, void **buffers, int nblocks)
785 : {
786 : struct iovec *iovp;
787 : int iovcnt;
788 :
789 : Assert(nblocks >= 1);
790 :
791 : /* If this build supports direct I/O, buffers must be I/O aligned. */
792 7616548 : for (int i = 0; i < nblocks; ++i)
793 : {
794 : if (PG_O_DIRECT != 0 && PG_IO_ALIGN_SIZE <= BLCKSZ)
795 : Assert((uintptr_t) buffers[i] ==
796 : TYPEALIGN(PG_IO_ALIGN_SIZE, buffers[i]));
797 : }
798 :
799 : /* Start the first iovec off with the first buffer. */
800 3643196 : iovp = &iov[0];
801 3643196 : iovp->iov_base = buffers[0];
802 3643196 : iovp->iov_len = BLCKSZ;
803 3643196 : iovcnt = 1;
804 :
805 : /* Try to merge the rest. */
806 3973352 : for (int i = 1; i < nblocks; ++i)
807 : {
808 330156 : void *buffer = buffers[i];
809 :
810 330156 : if (((char *) iovp->iov_base + iovp->iov_len) == buffer)
811 : {
812 : /* Contiguous with the last iovec. */
813 328550 : iovp->iov_len += BLCKSZ;
814 : }
815 : else
816 : {
817 : /* Need a new iovec. */
818 1606 : iovp++;
819 1606 : iovp->iov_base = buffer;
820 1606 : iovp->iov_len = BLCKSZ;
821 1606 : iovcnt++;
822 : }
823 : }
824 :
825 3643196 : return iovcnt;
826 : }
827 :
828 : /*
829 : * mdmaxcombine() -- Return the maximum number of total blocks that can be
830 : * combined with an IO starting at blocknum.
831 : */
832 : uint32
833 67522 : mdmaxcombine(SMgrRelation reln, ForkNumber forknum,
834 : BlockNumber blocknum)
835 : {
836 : BlockNumber segoff;
837 :
838 67522 : segoff = blocknum % ((BlockNumber) RELSEG_SIZE);
839 :
840 67522 : return RELSEG_SIZE - segoff;
841 : }
842 :
843 : /*
844 : * mdreadv() -- Read the specified blocks from a relation.
845 : */
846 : void
847 1196 : mdreadv(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
848 : void **buffers, BlockNumber nblocks)
849 : {
850 2392 : while (nblocks > 0)
851 : {
852 : struct iovec iov[PG_IOV_MAX];
853 : int iovcnt;
854 : off_t seekpos;
855 : int nbytes;
856 : MdfdVec *v;
857 : BlockNumber nblocks_this_segment;
858 : size_t transferred_this_segment;
859 : size_t size_this_segment;
860 :
861 1196 : v = _mdfd_getseg(reln, forknum, blocknum, false,
862 : EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
863 :
864 1196 : seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
865 :
866 : Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
867 :
868 1196 : nblocks_this_segment =
869 1196 : Min(nblocks,
870 : RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
871 1196 : nblocks_this_segment = Min(nblocks_this_segment, lengthof(iov));
872 :
873 1196 : if (nblocks_this_segment != nblocks)
874 0 : elog(ERROR, "read crosses segment boundary");
875 :
876 1196 : iovcnt = buffers_to_iovec(iov, buffers, nblocks_this_segment);
877 1196 : size_this_segment = nblocks_this_segment * BLCKSZ;
878 1196 : transferred_this_segment = 0;
879 :
880 : /*
881 : * Inner loop to continue after a short read. We'll keep going until
882 : * we hit EOF rather than assuming that a short read means we hit the
883 : * end.
884 : */
885 : for (;;)
886 : {
887 0 : TRACE_POSTGRESQL_SMGR_MD_READ_START(forknum, blocknum,
888 : reln->smgr_rlocator.locator.spcOid,
889 : reln->smgr_rlocator.locator.dbOid,
890 : reln->smgr_rlocator.locator.relNumber,
891 : reln->smgr_rlocator.backend);
892 1196 : nbytes = FileReadV(v->mdfd_vfd, iov, iovcnt, seekpos,
893 : WAIT_EVENT_DATA_FILE_READ);
894 : TRACE_POSTGRESQL_SMGR_MD_READ_DONE(forknum, blocknum,
895 : reln->smgr_rlocator.locator.spcOid,
896 : reln->smgr_rlocator.locator.dbOid,
897 : reln->smgr_rlocator.locator.relNumber,
898 : reln->smgr_rlocator.backend,
899 : nbytes,
900 : size_this_segment - transferred_this_segment);
901 :
902 : #ifdef SIMULATE_SHORT_READ
903 : nbytes = Min(nbytes, 4096);
904 : #endif
905 :
906 1196 : if (nbytes < 0)
907 0 : ereport(ERROR,
908 : (errcode_for_file_access(),
909 : errmsg("could not read blocks %u..%u in file \"%s\": %m",
910 : blocknum,
911 : blocknum + nblocks_this_segment - 1,
912 : FilePathName(v->mdfd_vfd))));
913 :
914 1196 : if (nbytes == 0)
915 : {
916 : /*
917 : * We are at or past EOF, or we read a partial block at EOF.
918 : * Normally this is an error; upper levels should never try to
919 : * read a nonexistent block. However, if zero_damaged_pages
920 : * is ON or we are InRecovery, we should instead return zeroes
921 : * without complaining. This allows, for example, the case of
922 : * trying to update a block that was later truncated away.
923 : *
924 : * NB: We think that this codepath is unreachable in recovery
925 : * and incomplete with zero_damaged_pages, as missing segments
926 : * are not created. Putting blocks into the buffer-pool that
927 : * do not exist on disk is rather problematic, as it will not
928 : * be found by scans that rely on smgrnblocks(), as they are
929 : * beyond EOF. It also can cause weird problems with relation
930 : * extension, as relation extension does not expect blocks
931 : * beyond EOF to exist.
932 : *
933 : * Therefore we do not want to copy the logic into
934 : * mdstartreadv(), where it would have to be more complicated
935 : * due to potential differences in the zero_damaged_pages
936 : * setting between the definer and completor of IO.
937 : *
938 : * For PG 18, we are putting an Assert(false) in mdreadv()
939 : * (triggering failures in assertion-enabled builds, but
940 : * continuing to work in production builds). Afterwards we
941 : * plan to remove this code entirely.
942 : */
943 0 : if (zero_damaged_pages || InRecovery)
944 : {
945 : Assert(false); /* see comment above */
946 :
947 0 : for (BlockNumber i = transferred_this_segment / BLCKSZ;
948 0 : i < nblocks_this_segment;
949 0 : ++i)
950 0 : memset(buffers[i], 0, BLCKSZ);
951 0 : break;
952 : }
953 : else
954 0 : ereport(ERROR,
955 : (errcode(ERRCODE_DATA_CORRUPTED),
956 : errmsg("could not read blocks %u..%u in file \"%s\": read only %zu of %zu bytes",
957 : blocknum,
958 : blocknum + nblocks_this_segment - 1,
959 : FilePathName(v->mdfd_vfd),
960 : transferred_this_segment,
961 : size_this_segment)));
962 : }
963 :
964 : /* One loop should usually be enough. */
965 1196 : transferred_this_segment += nbytes;
966 : Assert(transferred_this_segment <= size_this_segment);
967 1196 : if (transferred_this_segment == size_this_segment)
968 1196 : break;
969 :
970 : /* Adjust position and vectors after a short read. */
971 0 : seekpos += nbytes;
972 0 : iovcnt = compute_remaining_iovec(iov, iov, iovcnt, nbytes);
973 : }
974 :
975 1196 : nblocks -= nblocks_this_segment;
976 1196 : buffers += nblocks_this_segment;
977 1196 : blocknum += nblocks_this_segment;
978 : }
979 1196 : }
980 :
981 : /*
982 : * mdstartreadv() -- Asynchronous version of mdreadv().
983 : */
984 : void
985 2508100 : mdstartreadv(PgAioHandle *ioh,
986 : SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
987 : void **buffers, BlockNumber nblocks)
988 : {
989 : off_t seekpos;
990 : MdfdVec *v;
991 : BlockNumber nblocks_this_segment;
992 : struct iovec *iov;
993 : int iovcnt;
994 : int ret;
995 :
996 2508100 : v = _mdfd_getseg(reln, forknum, blocknum, false,
997 : EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
998 :
999 2508070 : seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
1000 :
1001 : Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
1002 :
1003 2508070 : nblocks_this_segment =
1004 2508070 : Min(nblocks,
1005 : RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
1006 :
1007 2508070 : if (nblocks_this_segment != nblocks)
1008 0 : elog(ERROR, "read crossing segment boundary");
1009 :
1010 2508070 : iovcnt = pgaio_io_get_iovec(ioh, &iov);
1011 :
1012 : Assert(nblocks <= iovcnt);
1013 :
1014 2508070 : iovcnt = buffers_to_iovec(iov, buffers, nblocks_this_segment);
1015 :
1016 : Assert(iovcnt <= nblocks_this_segment);
1017 :
1018 2508070 : if (!(io_direct_flags & IO_DIRECT_DATA))
1019 2505264 : pgaio_io_set_flag(ioh, PGAIO_HF_BUFFERED);
1020 :
1021 2508070 : pgaio_io_set_target_smgr(ioh,
1022 : reln,
1023 : forknum,
1024 : blocknum,
1025 : nblocks,
1026 : false);
1027 2508070 : pgaio_io_register_callbacks(ioh, PGAIO_HCB_MD_READV, 0);
1028 :
1029 2508070 : ret = FileStartReadV(ioh, v->mdfd_vfd, iovcnt, seekpos, WAIT_EVENT_DATA_FILE_READ);
1030 2508070 : if (ret != 0)
1031 0 : ereport(ERROR,
1032 : (errcode_for_file_access(),
1033 : errmsg("could not start reading blocks %u..%u in file \"%s\": %m",
1034 : blocknum,
1035 : blocknum + nblocks_this_segment - 1,
1036 : FilePathName(v->mdfd_vfd))));
1037 :
1038 : /*
1039 : * The error checks corresponding to the post-read checks in mdreadv() are
1040 : * in md_readv_complete().
1041 : *
1042 : * However we chose, at least for now, to not implement the
1043 : * zero_damaged_pages logic present in mdreadv(). As outlined in mdreadv()
1044 : * that logic is rather problematic, and we want to get rid of it. Here
1045 : * equivalent logic would have to be more complicated due to potential
1046 : * differences in the zero_damaged_pages setting between the definer and
1047 : * completor of IO.
1048 : */
1049 2508070 : }
1050 :
1051 : /*
1052 : * mdwritev() -- Write the supplied blocks at the appropriate location.
1053 : *
1054 : * This is to be used only for updating already-existing blocks of a
1055 : * relation (ie, those before the current EOF). To extend a relation,
1056 : * use mdextend().
1057 : */
1058 : void
1059 1133930 : mdwritev(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
1060 : const void **buffers, BlockNumber nblocks, bool skipFsync)
1061 : {
1062 : /* This assert is too expensive to have on normally ... */
1063 : #ifdef CHECK_WRITE_VS_EXTEND
1064 : Assert((uint64) blocknum + (uint64) nblocks <= (uint64) mdnblocks(reln, forknum));
1065 : #endif
1066 :
1067 2267860 : while (nblocks > 0)
1068 : {
1069 : struct iovec iov[PG_IOV_MAX];
1070 : int iovcnt;
1071 : off_t seekpos;
1072 : int nbytes;
1073 : MdfdVec *v;
1074 : BlockNumber nblocks_this_segment;
1075 : size_t transferred_this_segment;
1076 : size_t size_this_segment;
1077 :
1078 1133930 : v = _mdfd_getseg(reln, forknum, blocknum, skipFsync,
1079 : EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
1080 :
1081 1133930 : seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
1082 :
1083 : Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
1084 :
1085 1133930 : nblocks_this_segment =
1086 1133930 : Min(nblocks,
1087 : RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
1088 1133930 : nblocks_this_segment = Min(nblocks_this_segment, lengthof(iov));
1089 :
1090 1133930 : if (nblocks_this_segment != nblocks)
1091 0 : elog(ERROR, "write crosses segment boundary");
1092 :
1093 1133930 : iovcnt = buffers_to_iovec(iov, (void **) buffers, nblocks_this_segment);
1094 1133930 : size_this_segment = nblocks_this_segment * BLCKSZ;
1095 1133930 : transferred_this_segment = 0;
1096 :
1097 : /*
1098 : * Inner loop to continue after a short write. If the reason is that
1099 : * we're out of disk space, a future attempt should get an ENOSPC
1100 : * error from the kernel.
1101 : */
1102 : for (;;)
1103 : {
1104 0 : TRACE_POSTGRESQL_SMGR_MD_WRITE_START(forknum, blocknum,
1105 : reln->smgr_rlocator.locator.spcOid,
1106 : reln->smgr_rlocator.locator.dbOid,
1107 : reln->smgr_rlocator.locator.relNumber,
1108 : reln->smgr_rlocator.backend);
1109 1133930 : nbytes = FileWriteV(v->mdfd_vfd, iov, iovcnt, seekpos,
1110 : WAIT_EVENT_DATA_FILE_WRITE);
1111 : TRACE_POSTGRESQL_SMGR_MD_WRITE_DONE(forknum, blocknum,
1112 : reln->smgr_rlocator.locator.spcOid,
1113 : reln->smgr_rlocator.locator.dbOid,
1114 : reln->smgr_rlocator.locator.relNumber,
1115 : reln->smgr_rlocator.backend,
1116 : nbytes,
1117 : size_this_segment - transferred_this_segment);
1118 :
1119 : #ifdef SIMULATE_SHORT_WRITE
1120 : nbytes = Min(nbytes, 4096);
1121 : #endif
1122 :
1123 1133930 : if (nbytes < 0)
1124 : {
1125 0 : bool enospc = errno == ENOSPC;
1126 :
1127 0 : ereport(ERROR,
1128 : (errcode_for_file_access(),
1129 : errmsg("could not write blocks %u..%u in file \"%s\": %m",
1130 : blocknum,
1131 : blocknum + nblocks_this_segment - 1,
1132 : FilePathName(v->mdfd_vfd)),
1133 : enospc ? errhint("Check free disk space.") : 0));
1134 : }
1135 :
1136 : /* One loop should usually be enough. */
1137 1133930 : transferred_this_segment += nbytes;
1138 : Assert(transferred_this_segment <= size_this_segment);
1139 1133930 : if (transferred_this_segment == size_this_segment)
1140 1133930 : break;
1141 :
1142 : /* Adjust position and iovecs after a short write. */
1143 0 : seekpos += nbytes;
1144 0 : iovcnt = compute_remaining_iovec(iov, iov, iovcnt, nbytes);
1145 : }
1146 :
1147 1133930 : if (!skipFsync && !SmgrIsTemp(reln))
1148 1125974 : register_dirty_segment(reln, forknum, v);
1149 :
1150 1133930 : nblocks -= nblocks_this_segment;
1151 1133930 : buffers += nblocks_this_segment;
1152 1133930 : blocknum += nblocks_this_segment;
1153 : }
1154 1133930 : }
1155 :
1156 :
1157 : /*
1158 : * mdwriteback() -- Tell the kernel to write pages back to storage.
1159 : *
1160 : * This accepts a range of blocks because flushing several pages at once is
1161 : * considerably more efficient than doing so individually.
1162 : */
1163 : void
1164 0 : mdwriteback(SMgrRelation reln, ForkNumber forknum,
1165 : BlockNumber blocknum, BlockNumber nblocks)
1166 : {
1167 : Assert((io_direct_flags & IO_DIRECT_DATA) == 0);
1168 :
1169 : /*
1170 : * Issue flush requests in as few requests as possible; have to split at
1171 : * segment boundaries though, since those are actually separate files.
1172 : */
1173 0 : while (nblocks > 0)
1174 : {
1175 0 : BlockNumber nflush = nblocks;
1176 : off_t seekpos;
1177 : MdfdVec *v;
1178 : int segnum_start,
1179 : segnum_end;
1180 :
1181 0 : v = _mdfd_getseg(reln, forknum, blocknum, true /* not used */ ,
1182 : EXTENSION_DONT_OPEN);
1183 :
1184 : /*
1185 : * We might be flushing buffers of already removed relations, that's
1186 : * ok, just ignore that case. If the segment file wasn't open already
1187 : * (ie from a recent mdwrite()), then we don't want to re-open it, to
1188 : * avoid a race with PROCSIGNAL_BARRIER_SMGRRELEASE that might leave
1189 : * us with a descriptor to a file that is about to be unlinked.
1190 : */
1191 0 : if (!v)
1192 0 : return;
1193 :
1194 : /* compute offset inside the current segment */
1195 0 : segnum_start = blocknum / RELSEG_SIZE;
1196 :
1197 : /* compute number of desired writes within the current segment */
1198 0 : segnum_end = (blocknum + nblocks - 1) / RELSEG_SIZE;
1199 0 : if (segnum_start != segnum_end)
1200 0 : nflush = RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE));
1201 :
1202 : Assert(nflush >= 1);
1203 : Assert(nflush <= nblocks);
1204 :
1205 0 : seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
1206 :
1207 0 : FileWriteback(v->mdfd_vfd, seekpos, (off_t) BLCKSZ * nflush, WAIT_EVENT_DATA_FILE_FLUSH);
1208 :
1209 0 : nblocks -= nflush;
1210 0 : blocknum += nflush;
1211 : }
1212 : }
1213 :
1214 : /*
1215 : * mdnblocks() -- Get the number of blocks stored in a relation.
1216 : *
1217 : * Important side effect: all active segments of the relation are opened
1218 : * and added to the md_seg_fds array. If this routine has not been
1219 : * called, then only segments up to the last one actually touched
1220 : * are present in the array.
1221 : */
1222 : BlockNumber
1223 4870078 : mdnblocks(SMgrRelation reln, ForkNumber forknum)
1224 : {
1225 : MdfdVec *v;
1226 : BlockNumber nblocks;
1227 : BlockNumber segno;
1228 :
1229 4870078 : mdopenfork(reln, forknum, EXTENSION_FAIL);
1230 :
1231 : /* mdopen has opened the first segment */
1232 : Assert(reln->md_num_open_segs[forknum] > 0);
1233 :
1234 : /*
1235 : * Start from the last open segments, to avoid redundant seeks. We have
1236 : * previously verified that these segments are exactly RELSEG_SIZE long,
1237 : * and it's useless to recheck that each time.
1238 : *
1239 : * NOTE: this assumption could only be wrong if another backend has
1240 : * truncated the relation. We rely on higher code levels to handle that
1241 : * scenario by closing and re-opening the md fd, which is handled via
1242 : * relcache flush. (Since the checkpointer doesn't participate in
1243 : * relcache flush, it could have segment entries for inactive segments;
1244 : * that's OK because the checkpointer never needs to compute relation
1245 : * size.)
1246 : */
1247 4870040 : segno = reln->md_num_open_segs[forknum] - 1;
1248 4870040 : v = &reln->md_seg_fds[forknum][segno];
1249 :
1250 : for (;;)
1251 : {
1252 4870040 : nblocks = _mdnblocks(reln, forknum, v);
1253 4870040 : if (nblocks > ((BlockNumber) RELSEG_SIZE))
1254 0 : elog(FATAL, "segment too big");
1255 4870040 : if (nblocks < ((BlockNumber) RELSEG_SIZE))
1256 4870040 : return (segno * ((BlockNumber) RELSEG_SIZE)) + nblocks;
1257 :
1258 : /*
1259 : * If segment is exactly RELSEG_SIZE, advance to next one.
1260 : */
1261 0 : segno++;
1262 :
1263 : /*
1264 : * We used to pass O_CREAT here, but that has the disadvantage that it
1265 : * might create a segment which has vanished through some operating
1266 : * system misadventure. In such a case, creating the segment here
1267 : * undermines _mdfd_getseg's attempts to notice and report an error
1268 : * upon access to a missing segment.
1269 : */
1270 0 : v = _mdfd_openseg(reln, forknum, segno, 0);
1271 0 : if (v == NULL)
1272 0 : return segno * ((BlockNumber) RELSEG_SIZE);
1273 : }
1274 : }
1275 :
1276 : /*
1277 : * mdtruncate() -- Truncate relation to specified number of blocks.
1278 : *
1279 : * Guaranteed not to allocate memory, so it can be used in a critical section.
1280 : * Caller must have called smgrnblocks() to obtain curnblk while holding a
1281 : * sufficient lock to prevent a change in relation size, and not used any smgr
1282 : * functions for this relation or handled interrupts in between. This makes
1283 : * sure we have opened all active segments, so that truncate loop will get
1284 : * them all!
1285 : */
1286 : void
1287 1820 : mdtruncate(SMgrRelation reln, ForkNumber forknum,
1288 : BlockNumber curnblk, BlockNumber nblocks)
1289 : {
1290 : BlockNumber priorblocks;
1291 : int curopensegs;
1292 :
1293 1820 : if (nblocks > curnblk)
1294 : {
1295 : /* Bogus request ... but no complaint if InRecovery */
1296 0 : if (InRecovery)
1297 0 : return;
1298 0 : ereport(ERROR,
1299 : (errmsg("could not truncate file \"%s\" to %u blocks: it's only %u blocks now",
1300 : relpath(reln->smgr_rlocator, forknum).str,
1301 : nblocks, curnblk)));
1302 : }
1303 1820 : if (nblocks == curnblk)
1304 732 : return; /* no work */
1305 :
1306 : /*
1307 : * Truncate segments, starting at the last one. Starting at the end makes
1308 : * managing the memory for the fd array easier, should there be errors.
1309 : */
1310 1088 : curopensegs = reln->md_num_open_segs[forknum];
1311 2176 : while (curopensegs > 0)
1312 : {
1313 : MdfdVec *v;
1314 :
1315 1088 : priorblocks = (curopensegs - 1) * RELSEG_SIZE;
1316 :
1317 1088 : v = &reln->md_seg_fds[forknum][curopensegs - 1];
1318 :
1319 1088 : if (priorblocks > nblocks)
1320 : {
1321 : /*
1322 : * This segment is no longer active. We truncate the file, but do
1323 : * not delete it, for reasons explained in the header comments.
1324 : */
1325 0 : if (FileTruncate(v->mdfd_vfd, 0, WAIT_EVENT_DATA_FILE_TRUNCATE) < 0)
1326 0 : ereport(ERROR,
1327 : (errcode_for_file_access(),
1328 : errmsg("could not truncate file \"%s\": %m",
1329 : FilePathName(v->mdfd_vfd))));
1330 :
1331 0 : if (!SmgrIsTemp(reln))
1332 0 : register_dirty_segment(reln, forknum, v);
1333 :
1334 : /* we never drop the 1st segment */
1335 : Assert(v != &reln->md_seg_fds[forknum][0]);
1336 :
1337 0 : FileClose(v->mdfd_vfd);
1338 0 : _fdvec_resize(reln, forknum, curopensegs - 1);
1339 : }
1340 1088 : else if (priorblocks + ((BlockNumber) RELSEG_SIZE) > nblocks)
1341 : {
1342 : /*
1343 : * This is the last segment we want to keep. Truncate the file to
1344 : * the right length. NOTE: if nblocks is exactly a multiple K of
1345 : * RELSEG_SIZE, we will truncate the K+1st segment to 0 length but
1346 : * keep it. This adheres to the invariant given in the header
1347 : * comments.
1348 : */
1349 1088 : BlockNumber lastsegblocks = nblocks - priorblocks;
1350 :
1351 1088 : if (FileTruncate(v->mdfd_vfd, (off_t) lastsegblocks * BLCKSZ, WAIT_EVENT_DATA_FILE_TRUNCATE) < 0)
1352 0 : ereport(ERROR,
1353 : (errcode_for_file_access(),
1354 : errmsg("could not truncate file \"%s\" to %u blocks: %m",
1355 : FilePathName(v->mdfd_vfd),
1356 : nblocks)));
1357 1088 : if (!SmgrIsTemp(reln))
1358 760 : register_dirty_segment(reln, forknum, v);
1359 : }
1360 : else
1361 : {
1362 : /*
1363 : * We still need this segment, so nothing to do for this and any
1364 : * earlier segment.
1365 : */
1366 0 : break;
1367 : }
1368 1088 : curopensegs--;
1369 : }
1370 : }
1371 :
1372 : /*
1373 : * mdregistersync() -- Mark whole relation as needing fsync
1374 : */
1375 : void
1376 49936 : mdregistersync(SMgrRelation reln, ForkNumber forknum)
1377 : {
1378 : int segno;
1379 : int min_inactive_seg;
1380 :
1381 : /*
1382 : * NOTE: mdnblocks makes sure we have opened all active segments, so that
1383 : * the loop below will get them all!
1384 : */
1385 49936 : mdnblocks(reln, forknum);
1386 :
1387 49936 : min_inactive_seg = segno = reln->md_num_open_segs[forknum];
1388 :
1389 : /*
1390 : * Temporarily open inactive segments, then close them after sync. There
1391 : * may be some inactive segments left opened after error, but that is
1392 : * harmless. We don't bother to clean them up and take a risk of further
1393 : * trouble. The next mdclose() will soon close them.
1394 : */
1395 49936 : while (_mdfd_openseg(reln, forknum, segno, 0) != NULL)
1396 0 : segno++;
1397 :
1398 99872 : while (segno > 0)
1399 : {
1400 49936 : MdfdVec *v = &reln->md_seg_fds[forknum][segno - 1];
1401 :
1402 49936 : register_dirty_segment(reln, forknum, v);
1403 :
1404 : /* Close inactive segments immediately */
1405 49936 : if (segno > min_inactive_seg)
1406 : {
1407 0 : FileClose(v->mdfd_vfd);
1408 0 : _fdvec_resize(reln, forknum, segno - 1);
1409 : }
1410 :
1411 49936 : segno--;
1412 : }
1413 49936 : }
1414 :
1415 : /*
1416 : * mdimmedsync() -- Immediately sync a relation to stable storage.
1417 : *
1418 : * Note that only writes already issued are synced; this routine knows
1419 : * nothing of dirty buffers that may exist inside the buffer manager. We
1420 : * sync active and inactive segments; smgrDoPendingSyncs() relies on this.
1421 : * Consider a relation skipping WAL. Suppose a checkpoint syncs blocks of
1422 : * some segment, then mdtruncate() renders that segment inactive. If we
1423 : * crash before the next checkpoint syncs the newly-inactive segment, that
1424 : * segment may survive recovery, reintroducing unwanted data into the table.
1425 : */
1426 : void
1427 32 : mdimmedsync(SMgrRelation reln, ForkNumber forknum)
1428 : {
1429 : int segno;
1430 : int min_inactive_seg;
1431 :
1432 : /*
1433 : * NOTE: mdnblocks makes sure we have opened all active segments, so that
1434 : * the loop below will get them all!
1435 : */
1436 32 : mdnblocks(reln, forknum);
1437 :
1438 32 : min_inactive_seg = segno = reln->md_num_open_segs[forknum];
1439 :
1440 : /*
1441 : * Temporarily open inactive segments, then close them after sync. There
1442 : * may be some inactive segments left opened after fsync() error, but that
1443 : * is harmless. We don't bother to clean them up and take a risk of
1444 : * further trouble. The next mdclose() will soon close them.
1445 : */
1446 32 : while (_mdfd_openseg(reln, forknum, segno, 0) != NULL)
1447 0 : segno++;
1448 :
1449 64 : while (segno > 0)
1450 : {
1451 32 : MdfdVec *v = &reln->md_seg_fds[forknum][segno - 1];
1452 :
1453 : /*
1454 : * fsyncs done through mdimmedsync() should be tracked in a separate
1455 : * IOContext than those done through mdsyncfiletag() to differentiate
1456 : * between unavoidable client backend fsyncs (e.g. those done during
1457 : * index build) and those which ideally would have been done by the
1458 : * checkpointer. Since other IO operations bypassing the buffer
1459 : * manager could also be tracked in such an IOContext, wait until
1460 : * these are also tracked to track immediate fsyncs.
1461 : */
1462 32 : if (FileSync(v->mdfd_vfd, WAIT_EVENT_DATA_FILE_IMMEDIATE_SYNC) < 0)
1463 0 : ereport(data_sync_elevel(ERROR),
1464 : (errcode_for_file_access(),
1465 : errmsg("could not fsync file \"%s\": %m",
1466 : FilePathName(v->mdfd_vfd))));
1467 :
1468 : /* Close inactive segments immediately */
1469 32 : if (segno > min_inactive_seg)
1470 : {
1471 0 : FileClose(v->mdfd_vfd);
1472 0 : _fdvec_resize(reln, forknum, segno - 1);
1473 : }
1474 :
1475 32 : segno--;
1476 : }
1477 32 : }
1478 :
1479 : int
1480 893750 : mdfd(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, uint32 *off)
1481 : {
1482 893750 : MdfdVec *v = mdopenfork(reln, forknum, EXTENSION_FAIL);
1483 :
1484 893750 : v = _mdfd_getseg(reln, forknum, blocknum, false,
1485 : EXTENSION_FAIL);
1486 :
1487 893750 : *off = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
1488 :
1489 : Assert(*off < (off_t) BLCKSZ * RELSEG_SIZE);
1490 :
1491 893750 : return FileGetRawDesc(v->mdfd_vfd);
1492 : }
1493 :
1494 : /*
1495 : * register_dirty_segment() -- Mark a relation segment as needing fsync
1496 : *
1497 : * If there is a local pending-ops table, just make an entry in it for
1498 : * ProcessSyncRequests to process later. Otherwise, try to pass off the
1499 : * fsync request to the checkpointer process. If that fails, just do the
1500 : * fsync locally before returning (we hope this will not happen often
1501 : * enough to be a performance problem).
1502 : */
1503 : static void
1504 1910522 : register_dirty_segment(SMgrRelation reln, ForkNumber forknum, MdfdVec *seg)
1505 : {
1506 : FileTag tag;
1507 :
1508 1910522 : INIT_MD_FILETAG(tag, reln->smgr_rlocator.locator, forknum, seg->mdfd_segno);
1509 :
1510 : /* Temp relations should never be fsync'd */
1511 : Assert(!SmgrIsTemp(reln));
1512 :
1513 1910522 : if (!RegisterSyncRequest(&tag, SYNC_REQUEST, false /* retryOnError */ ))
1514 : {
1515 : instr_time io_start;
1516 :
1517 1448 : ereport(DEBUG1,
1518 : (errmsg_internal("could not forward fsync request because request queue is full")));
1519 :
1520 1448 : io_start = pgstat_prepare_io_time(track_io_timing);
1521 :
1522 1448 : if (FileSync(seg->mdfd_vfd, WAIT_EVENT_DATA_FILE_SYNC) < 0)
1523 0 : ereport(data_sync_elevel(ERROR),
1524 : (errcode_for_file_access(),
1525 : errmsg("could not fsync file \"%s\": %m",
1526 : FilePathName(seg->mdfd_vfd))));
1527 :
1528 : /*
1529 : * We have no way of knowing if the current IOContext is
1530 : * IOCONTEXT_NORMAL or IOCONTEXT_[BULKREAD, BULKWRITE, VACUUM] at this
1531 : * point, so count the fsync as being in the IOCONTEXT_NORMAL
1532 : * IOContext. This is probably okay, because the number of backend
1533 : * fsyncs doesn't say anything about the efficacy of the
1534 : * BufferAccessStrategy. And counting both fsyncs done in
1535 : * IOCONTEXT_NORMAL and IOCONTEXT_[BULKREAD, BULKWRITE, VACUUM] under
1536 : * IOCONTEXT_NORMAL is likely clearer when investigating the number of
1537 : * backend fsyncs.
1538 : */
1539 1448 : pgstat_count_io_op_time(IOOBJECT_RELATION, IOCONTEXT_NORMAL,
1540 : IOOP_FSYNC, io_start, 1, 0);
1541 : }
1542 1910522 : }
1543 :
1544 : /*
1545 : * register_unlink_segment() -- Schedule a file to be deleted after next checkpoint
1546 : */
1547 : static void
1548 70312 : register_unlink_segment(RelFileLocatorBackend rlocator, ForkNumber forknum,
1549 : BlockNumber segno)
1550 : {
1551 : FileTag tag;
1552 :
1553 70312 : INIT_MD_FILETAG(tag, rlocator.locator, forknum, segno);
1554 :
1555 : /* Should never be used with temp relations */
1556 : Assert(!RelFileLocatorBackendIsTemp(rlocator));
1557 :
1558 70312 : RegisterSyncRequest(&tag, SYNC_UNLINK_REQUEST, true /* retryOnError */ );
1559 70312 : }
1560 :
1561 : /*
1562 : * register_forget_request() -- forget any fsyncs for a relation fork's segment
1563 : */
1564 : static void
1565 272512 : register_forget_request(RelFileLocatorBackend rlocator, ForkNumber forknum,
1566 : BlockNumber segno)
1567 : {
1568 : FileTag tag;
1569 :
1570 272512 : INIT_MD_FILETAG(tag, rlocator.locator, forknum, segno);
1571 :
1572 272512 : RegisterSyncRequest(&tag, SYNC_FORGET_REQUEST, true /* retryOnError */ );
1573 272512 : }
1574 :
1575 : /*
1576 : * ForgetDatabaseSyncRequests -- forget any fsyncs and unlinks for a DB
1577 : */
1578 : void
1579 122 : ForgetDatabaseSyncRequests(Oid dbid)
1580 : {
1581 : FileTag tag;
1582 : RelFileLocator rlocator;
1583 :
1584 122 : rlocator.dbOid = dbid;
1585 122 : rlocator.spcOid = 0;
1586 122 : rlocator.relNumber = 0;
1587 :
1588 122 : INIT_MD_FILETAG(tag, rlocator, InvalidForkNumber, InvalidBlockNumber);
1589 :
1590 122 : RegisterSyncRequest(&tag, SYNC_FILTER_REQUEST, true /* retryOnError */ );
1591 122 : }
1592 :
1593 : /*
1594 : * DropRelationFiles -- drop files of all given relations
1595 : */
1596 : void
1597 5310 : DropRelationFiles(RelFileLocator *delrels, int ndelrels, bool isRedo)
1598 : {
1599 : SMgrRelation *srels;
1600 : int i;
1601 :
1602 5310 : srels = palloc(sizeof(SMgrRelation) * ndelrels);
1603 20528 : for (i = 0; i < ndelrels; i++)
1604 : {
1605 15218 : SMgrRelation srel = smgropen(delrels[i], INVALID_PROC_NUMBER);
1606 :
1607 15218 : if (isRedo)
1608 : {
1609 : ForkNumber fork;
1610 :
1611 75850 : for (fork = 0; fork <= MAX_FORKNUM; fork++)
1612 60680 : XLogDropRelation(delrels[i], fork);
1613 : }
1614 15218 : srels[i] = srel;
1615 : }
1616 :
1617 5310 : smgrdounlinkall(srels, ndelrels, isRedo);
1618 :
1619 20528 : for (i = 0; i < ndelrels; i++)
1620 15218 : smgrclose(srels[i]);
1621 5310 : pfree(srels);
1622 5310 : }
1623 :
1624 :
1625 : /*
1626 : * _fdvec_resize() -- Resize the fork's open segments array
1627 : */
1628 : static void
1629 3425108 : _fdvec_resize(SMgrRelation reln,
1630 : ForkNumber forknum,
1631 : int nseg)
1632 : {
1633 3425108 : if (nseg == 0)
1634 : {
1635 1125060 : if (reln->md_num_open_segs[forknum] > 0)
1636 : {
1637 1125060 : pfree(reln->md_seg_fds[forknum]);
1638 1125060 : reln->md_seg_fds[forknum] = NULL;
1639 : }
1640 : }
1641 2300048 : else if (reln->md_num_open_segs[forknum] == 0)
1642 : {
1643 2300048 : reln->md_seg_fds[forknum] =
1644 2300048 : MemoryContextAlloc(MdCxt, sizeof(MdfdVec) * nseg);
1645 : }
1646 0 : else if (nseg > reln->md_num_open_segs[forknum])
1647 : {
1648 : /*
1649 : * It doesn't seem worthwhile complicating the code to amortize
1650 : * repalloc() calls. Those are far faster than PathNameOpenFile() or
1651 : * FileClose(), and the memory context internally will sometimes avoid
1652 : * doing an actual reallocation.
1653 : */
1654 0 : reln->md_seg_fds[forknum] =
1655 0 : repalloc(reln->md_seg_fds[forknum],
1656 : sizeof(MdfdVec) * nseg);
1657 : }
1658 : else
1659 : {
1660 : /*
1661 : * We don't reallocate a smaller array, because we want mdtruncate()
1662 : * to be able to promise that it won't allocate memory, so that it is
1663 : * allowed in a critical section. This means that a bit of space in
1664 : * the array is now wasted, until the next time we add a segment and
1665 : * reallocate.
1666 : */
1667 : }
1668 :
1669 3425108 : reln->md_num_open_segs[forknum] = nseg;
1670 3425108 : }
1671 :
1672 : /*
1673 : * Return the filename for the specified segment of the relation. The
1674 : * returned string is palloc'd.
1675 : */
1676 : static MdPathStr
1677 49992 : _mdfd_segpath(SMgrRelation reln, ForkNumber forknum, BlockNumber segno)
1678 : {
1679 : RelPathStr path;
1680 : MdPathStr fullpath;
1681 :
1682 49992 : path = relpath(reln->smgr_rlocator, forknum);
1683 :
1684 49992 : if (segno > 0)
1685 49992 : sprintf(fullpath.str, "%s.%u", path.str, segno);
1686 : else
1687 0 : strcpy(fullpath.str, path.str);
1688 :
1689 49992 : return fullpath;
1690 : }
1691 :
1692 : /*
1693 : * Open the specified segment of the relation,
1694 : * and make a MdfdVec object for it. Returns NULL on failure.
1695 : */
1696 : static MdfdVec *
1697 49968 : _mdfd_openseg(SMgrRelation reln, ForkNumber forknum, BlockNumber segno,
1698 : int oflags)
1699 : {
1700 : MdfdVec *v;
1701 : File fd;
1702 : MdPathStr fullpath;
1703 :
1704 49968 : fullpath = _mdfd_segpath(reln, forknum, segno);
1705 :
1706 : /* open the file */
1707 49968 : fd = PathNameOpenFile(fullpath.str, _mdfd_open_flags() | oflags);
1708 :
1709 49968 : if (fd < 0)
1710 49968 : return NULL;
1711 :
1712 : /*
1713 : * Segments are always opened in order from lowest to highest, so we must
1714 : * be adding a new one at the end.
1715 : */
1716 : Assert(segno == reln->md_num_open_segs[forknum]);
1717 :
1718 0 : _fdvec_resize(reln, forknum, segno + 1);
1719 :
1720 : /* fill the entry */
1721 0 : v = &reln->md_seg_fds[forknum][segno];
1722 0 : v->mdfd_vfd = fd;
1723 0 : v->mdfd_segno = segno;
1724 :
1725 : Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
1726 :
1727 : /* all done */
1728 0 : return v;
1729 : }
1730 :
1731 : /*
1732 : * _mdfd_getseg() -- Find the segment of the relation holding the
1733 : * specified block.
1734 : *
1735 : * If the segment doesn't exist, we ereport, return NULL, or create the
1736 : * segment, according to "behavior". Note: skipFsync is only used in the
1737 : * EXTENSION_CREATE case.
1738 : */
1739 : static MdfdVec *
1740 5220342 : _mdfd_getseg(SMgrRelation reln, ForkNumber forknum, BlockNumber blkno,
1741 : bool skipFsync, int behavior)
1742 : {
1743 : MdfdVec *v;
1744 : BlockNumber targetseg;
1745 : BlockNumber nextsegno;
1746 :
1747 : /* some way to handle non-existent segments needs to be specified */
1748 : Assert(behavior &
1749 : (EXTENSION_FAIL | EXTENSION_CREATE | EXTENSION_RETURN_NULL |
1750 : EXTENSION_DONT_OPEN));
1751 :
1752 5220342 : targetseg = blkno / ((BlockNumber) RELSEG_SIZE);
1753 :
1754 : /* if an existing and opened segment, we're done */
1755 5220342 : if (targetseg < reln->md_num_open_segs[forknum])
1756 : {
1757 4751282 : v = &reln->md_seg_fds[forknum][targetseg];
1758 4751282 : return v;
1759 : }
1760 :
1761 : /* The caller only wants the segment if we already had it open. */
1762 469060 : if (behavior & EXTENSION_DONT_OPEN)
1763 0 : return NULL;
1764 :
1765 : /*
1766 : * The target segment is not yet open. Iterate over all the segments
1767 : * between the last opened and the target segment. This way missing
1768 : * segments either raise an error, or get created (according to
1769 : * 'behavior'). Start with either the last opened, or the first segment if
1770 : * none was opened before.
1771 : */
1772 469060 : if (reln->md_num_open_segs[forknum] > 0)
1773 24 : v = &reln->md_seg_fds[forknum][reln->md_num_open_segs[forknum] - 1];
1774 : else
1775 : {
1776 469036 : v = mdopenfork(reln, forknum, behavior);
1777 469030 : if (!v)
1778 0 : return NULL; /* if behavior & EXTENSION_RETURN_NULL */
1779 : }
1780 :
1781 469054 : for (nextsegno = reln->md_num_open_segs[forknum];
1782 469054 : nextsegno <= targetseg; nextsegno++)
1783 : {
1784 24 : BlockNumber nblocks = _mdnblocks(reln, forknum, v);
1785 24 : int flags = 0;
1786 :
1787 : Assert(nextsegno == v->mdfd_segno + 1);
1788 :
1789 24 : if (nblocks > ((BlockNumber) RELSEG_SIZE))
1790 0 : elog(FATAL, "segment too big");
1791 :
1792 24 : if ((behavior & EXTENSION_CREATE) ||
1793 24 : (InRecovery && (behavior & EXTENSION_CREATE_RECOVERY)))
1794 : {
1795 : /*
1796 : * Normally we will create new segments only if authorized by the
1797 : * caller (i.e., we are doing mdextend()). But when doing WAL
1798 : * recovery, create segments anyway; this allows cases such as
1799 : * replaying WAL data that has a write into a high-numbered
1800 : * segment of a relation that was later deleted. We want to go
1801 : * ahead and create the segments so we can finish out the replay.
1802 : *
1803 : * We have to maintain the invariant that segments before the last
1804 : * active segment are of size RELSEG_SIZE; therefore, if
1805 : * extending, pad them out with zeroes if needed. (This only
1806 : * matters if in recovery, or if the caller is extending the
1807 : * relation discontiguously, but that can happen in hash indexes.)
1808 : */
1809 0 : if (nblocks < ((BlockNumber) RELSEG_SIZE))
1810 : {
1811 0 : char *zerobuf = palloc_aligned(BLCKSZ, PG_IO_ALIGN_SIZE,
1812 : MCXT_ALLOC_ZERO);
1813 :
1814 0 : mdextend(reln, forknum,
1815 0 : nextsegno * ((BlockNumber) RELSEG_SIZE) - 1,
1816 : zerobuf, skipFsync);
1817 0 : pfree(zerobuf);
1818 : }
1819 0 : flags = O_CREAT;
1820 : }
1821 24 : else if (nblocks < ((BlockNumber) RELSEG_SIZE))
1822 : {
1823 : /*
1824 : * When not extending, only open the next segment if the current
1825 : * one is exactly RELSEG_SIZE. If not (this branch), either
1826 : * return NULL or fail.
1827 : */
1828 24 : if (behavior & EXTENSION_RETURN_NULL)
1829 : {
1830 : /*
1831 : * Some callers discern between reasons for _mdfd_getseg()
1832 : * returning NULL based on errno. As there's no failing
1833 : * syscall involved in this case, explicitly set errno to
1834 : * ENOENT, as that seems the closest interpretation.
1835 : */
1836 0 : errno = ENOENT;
1837 0 : return NULL;
1838 : }
1839 :
1840 24 : ereport(ERROR,
1841 : (errcode_for_file_access(),
1842 : errmsg("could not open file \"%s\" (target block %u): previous segment is only %u blocks",
1843 : _mdfd_segpath(reln, forknum, nextsegno).str,
1844 : blkno, nblocks)));
1845 : }
1846 :
1847 0 : v = _mdfd_openseg(reln, forknum, nextsegno, flags);
1848 :
1849 0 : if (v == NULL)
1850 : {
1851 0 : if ((behavior & EXTENSION_RETURN_NULL) &&
1852 0 : FILE_POSSIBLY_DELETED(errno))
1853 0 : return NULL;
1854 0 : ereport(ERROR,
1855 : (errcode_for_file_access(),
1856 : errmsg("could not open file \"%s\" (target block %u): %m",
1857 : _mdfd_segpath(reln, forknum, nextsegno).str,
1858 : blkno)));
1859 : }
1860 : }
1861 :
1862 469030 : return v;
1863 : }
1864 :
1865 : /*
1866 : * Get number of blocks present in a single disk file
1867 : */
1868 : static BlockNumber
1869 4870064 : _mdnblocks(SMgrRelation reln, ForkNumber forknum, MdfdVec *seg)
1870 : {
1871 : off_t len;
1872 :
1873 4870064 : len = FileSize(seg->mdfd_vfd);
1874 4870064 : if (len < 0)
1875 0 : ereport(ERROR,
1876 : (errcode_for_file_access(),
1877 : errmsg("could not seek to end of file \"%s\": %m",
1878 : FilePathName(seg->mdfd_vfd))));
1879 : /* note that this calculation will ignore any partial block at EOF */
1880 4870064 : return (BlockNumber) (len / BLCKSZ);
1881 : }
1882 :
1883 : /*
1884 : * Sync a file to disk, given a file tag. Write the path into an output
1885 : * buffer so the caller can use it in error messages.
1886 : *
1887 : * Return 0 on success, -1 on failure, with errno set.
1888 : */
1889 : int
1890 0 : mdsyncfiletag(const FileTag *ftag, char *path)
1891 : {
1892 0 : SMgrRelation reln = smgropen(ftag->rlocator, INVALID_PROC_NUMBER);
1893 : File file;
1894 : instr_time io_start;
1895 : bool need_to_close;
1896 : int result,
1897 : save_errno;
1898 :
1899 : /* See if we already have the file open, or need to open it. */
1900 0 : if (ftag->segno < reln->md_num_open_segs[ftag->forknum])
1901 : {
1902 0 : file = reln->md_seg_fds[ftag->forknum][ftag->segno].mdfd_vfd;
1903 0 : strlcpy(path, FilePathName(file), MAXPGPATH);
1904 0 : need_to_close = false;
1905 : }
1906 : else
1907 : {
1908 : MdPathStr p;
1909 :
1910 0 : p = _mdfd_segpath(reln, ftag->forknum, ftag->segno);
1911 0 : strlcpy(path, p.str, MD_PATH_STR_MAXLEN);
1912 :
1913 0 : file = PathNameOpenFile(path, _mdfd_open_flags());
1914 0 : if (file < 0)
1915 0 : return -1;
1916 0 : need_to_close = true;
1917 : }
1918 :
1919 0 : io_start = pgstat_prepare_io_time(track_io_timing);
1920 :
1921 : /* Sync the file. */
1922 0 : result = FileSync(file, WAIT_EVENT_DATA_FILE_SYNC);
1923 0 : save_errno = errno;
1924 :
1925 0 : if (need_to_close)
1926 0 : FileClose(file);
1927 :
1928 0 : pgstat_count_io_op_time(IOOBJECT_RELATION, IOCONTEXT_NORMAL,
1929 : IOOP_FSYNC, io_start, 1, 0);
1930 :
1931 0 : errno = save_errno;
1932 0 : return result;
1933 : }
1934 :
1935 : /*
1936 : * Unlink a file, given a file tag. Write the path into an output
1937 : * buffer so the caller can use it in error messages.
1938 : *
1939 : * Return 0 on success, -1 on failure, with errno set.
1940 : */
1941 : int
1942 66436 : mdunlinkfiletag(const FileTag *ftag, char *path)
1943 : {
1944 : RelPathStr p;
1945 :
1946 : /* Compute the path. */
1947 66436 : p = relpathperm(ftag->rlocator, MAIN_FORKNUM);
1948 66436 : strlcpy(path, p.str, MAXPGPATH);
1949 :
1950 : /* Try to unlink the file. */
1951 66436 : return unlink(path);
1952 : }
1953 :
1954 : /*
1955 : * Check if a given candidate request matches a given tag, when processing
1956 : * a SYNC_FILTER_REQUEST request. This will be called for all pending
1957 : * requests to find out whether to forget them.
1958 : */
1959 : bool
1960 14174 : mdfiletagmatches(const FileTag *ftag, const FileTag *candidate)
1961 : {
1962 : /*
1963 : * For now we only use filter requests as a way to drop all scheduled
1964 : * callbacks relating to a given database, when dropping the database.
1965 : * We'll return true for all candidates that have the same database OID as
1966 : * the ftag from the SYNC_FILTER_REQUEST request, so they're forgotten.
1967 : */
1968 14174 : return ftag->rlocator.dbOid == candidate->rlocator.dbOid;
1969 : }
1970 :
1971 : /*
1972 : * AIO completion callback for mdstartreadv().
1973 : */
1974 : static PgAioResult
1975 2298418 : md_readv_complete(PgAioHandle *ioh, PgAioResult prior_result, uint8 cb_data)
1976 : {
1977 2298418 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
1978 2298418 : PgAioResult result = prior_result;
1979 :
1980 2298418 : if (prior_result.result < 0)
1981 : {
1982 14 : result.status = PGAIO_RS_ERROR;
1983 14 : result.id = PGAIO_HCB_MD_READV;
1984 : /* For "hard" errors, track the error number in error_data */
1985 14 : result.error_data = -prior_result.result;
1986 14 : result.result = 0;
1987 :
1988 : /*
1989 : * Immediately log a message about the IO error, but only to the
1990 : * server log. The reason to do so immediately is that the originator
1991 : * might not process the query result immediately (because it is busy
1992 : * doing another part of query processing) or at all (e.g. if it was
1993 : * cancelled or errored out due to another IO also failing). The
1994 : * definer of the IO will emit an ERROR when processing the IO's
1995 : * results
1996 : */
1997 14 : pgaio_result_report(result, td, LOG_SERVER_ONLY);
1998 :
1999 14 : return result;
2000 : }
2001 :
2002 : /*
2003 : * As explained above smgrstartreadv(), the smgr API operates on the level
2004 : * of blocks, rather than bytes. Convert.
2005 : */
2006 2298404 : result.result /= BLCKSZ;
2007 :
2008 : Assert(result.result <= td->smgr.nblocks);
2009 :
2010 2298404 : if (result.result == 0)
2011 : {
2012 : /* consider 0 blocks read a failure */
2013 4 : result.status = PGAIO_RS_ERROR;
2014 4 : result.id = PGAIO_HCB_MD_READV;
2015 4 : result.error_data = 0;
2016 :
2017 : /* see comment above the "hard error" case */
2018 4 : pgaio_result_report(result, td, LOG_SERVER_ONLY);
2019 :
2020 4 : return result;
2021 : }
2022 :
2023 2298400 : if (result.status != PGAIO_RS_ERROR &&
2024 2298400 : result.result < td->smgr.nblocks)
2025 : {
2026 : /* partial reads should be retried at upper level */
2027 24 : result.status = PGAIO_RS_PARTIAL;
2028 24 : result.id = PGAIO_HCB_MD_READV;
2029 : }
2030 :
2031 2298400 : return result;
2032 : }
2033 :
2034 : /*
2035 : * AIO error reporting callback for mdstartreadv().
2036 : *
2037 : * Errors are encoded as follows:
2038 : * - PgAioResult.error_data != 0 encodes IO that failed with that errno
2039 : * - PgAioResult.error_data == 0 encodes IO that didn't read all data
2040 : */
2041 : static void
2042 56 : md_readv_report(PgAioResult result, const PgAioTargetData *td, int elevel)
2043 : {
2044 : RelPathStr path;
2045 :
2046 56 : path = relpathbackend(td->smgr.rlocator,
2047 : td->smgr.is_temp ? MyProcNumber : INVALID_PROC_NUMBER,
2048 : td->smgr.forkNum);
2049 :
2050 56 : if (result.error_data != 0)
2051 : {
2052 : /* for errcode_for_file_access() and %m */
2053 28 : errno = result.error_data;
2054 :
2055 28 : ereport(elevel,
2056 : errcode_for_file_access(),
2057 : errmsg("could not read blocks %u..%u in file \"%s\": %m",
2058 : td->smgr.blockNum,
2059 : td->smgr.blockNum + td->smgr.nblocks - 1,
2060 : path.str));
2061 : }
2062 : else
2063 : {
2064 : /*
2065 : * NB: This will typically only be output in debug messages, while
2066 : * retrying a partial IO.
2067 : */
2068 28 : ereport(elevel,
2069 : errcode(ERRCODE_DATA_CORRUPTED),
2070 : errmsg("could not read blocks %u..%u in file \"%s\": read only %zu of %zu bytes",
2071 : td->smgr.blockNum,
2072 : td->smgr.blockNum + td->smgr.nblocks - 1,
2073 : path.str,
2074 : result.result * (size_t) BLCKSZ,
2075 : td->smgr.nblocks * (size_t) BLCKSZ));
2076 : }
2077 38 : }
|
