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 : #include "utils/wait_event.h"
44 :
45 : /*
46 : * The magnetic disk storage manager keeps track of open file
47 : * descriptors in its own descriptor pool. This is done to make it
48 : * easier to support relations that are larger than the operating
49 : * system's file size limit (often 2GBytes). In order to do that,
50 : * we break relations up into "segment" files that are each shorter than
51 : * the OS file size limit. The segment size is set by the RELSEG_SIZE
52 : * configuration constant in pg_config.h.
53 : *
54 : * On disk, a relation must consist of consecutively numbered segment
55 : * files in the pattern
56 : * -- Zero or more full segments of exactly RELSEG_SIZE blocks each
57 : * -- Exactly one partial segment of size 0 <= size < RELSEG_SIZE blocks
58 : * -- Optionally, any number of inactive segments of size 0 blocks.
59 : * The full and partial segments are collectively the "active" segments.
60 : * Inactive segments are those that once contained data but are currently
61 : * not needed because of an mdtruncate() operation. The reason for leaving
62 : * them present at size zero, rather than unlinking them, is that other
63 : * backends and/or the checkpointer might be holding open file references to
64 : * such segments. If the relation expands again after mdtruncate(), such
65 : * that a deactivated segment becomes active again, it is important that
66 : * such file references still be valid --- else data might get written
67 : * out to an unlinked old copy of a segment file that will eventually
68 : * disappear.
69 : *
70 : * RELSEG_SIZE must fit into BlockNumber; but since we expose its value
71 : * as an integer GUC, it actually needs to fit in signed int. It's worth
72 : * having a cross-check for this since configure's --with-segsize options
73 : * could let people select insane values.
74 : */
75 : StaticAssertDecl(RELSEG_SIZE > 0 && RELSEG_SIZE <= INT_MAX,
76 : "RELSEG_SIZE must fit in an integer");
77 :
78 : /*
79 : * File descriptors are stored in the per-fork md_seg_fds arrays inside
80 : * SMgrRelation. The length of these arrays is stored in md_num_open_segs.
81 : * Note that a fork's md_num_open_segs having a specific value does not
82 : * necessarily mean the relation doesn't have additional segments; we may
83 : * just not have opened the next segment yet. (We could not have "all
84 : * segments are in the array" as an invariant anyway, since another backend
85 : * could extend the relation while we aren't looking.) We do not have
86 : * entries for inactive segments, however; as soon as we find a partial
87 : * segment, we assume that any subsequent segments are inactive.
88 : *
89 : * The entire MdfdVec array is palloc'd in the MdCxt memory context.
90 : */
91 :
92 : typedef struct _MdfdVec
93 : {
94 : File mdfd_vfd; /* fd number in fd.c's pool */
95 : BlockNumber mdfd_segno; /* segment number, from 0 */
96 : } MdfdVec;
97 :
98 : static MemoryContext MdCxt; /* context for all MdfdVec objects */
99 :
100 :
101 : /* Populate a file tag describing an md.c segment file. */
102 : #define INIT_MD_FILETAG(a,xx_rlocator,xx_forknum,xx_segno) \
103 : ( \
104 : memset(&(a), 0, sizeof(FileTag)), \
105 : (a).handler = SYNC_HANDLER_MD, \
106 : (a).rlocator = (xx_rlocator), \
107 : (a).forknum = (xx_forknum), \
108 : (a).segno = (xx_segno) \
109 : )
110 :
111 :
112 : /*** behavior for mdopen & _mdfd_getseg ***/
113 : /* ereport if segment not present */
114 : #define EXTENSION_FAIL (1 << 0)
115 : /* return NULL if segment not present */
116 : #define EXTENSION_RETURN_NULL (1 << 1)
117 : /* create new segments as needed */
118 : #define EXTENSION_CREATE (1 << 2)
119 : /* create new segments if needed during recovery */
120 : #define EXTENSION_CREATE_RECOVERY (1 << 3)
121 : /* don't try to open a segment, if not already open */
122 : #define EXTENSION_DONT_OPEN (1 << 5)
123 :
124 :
125 : /*
126 : * Fixed-length string to represent paths to files that need to be built by
127 : * md.c.
128 : *
129 : * The maximum number of segments is MaxBlockNumber / RELSEG_SIZE, where
130 : * RELSEG_SIZE can be set to 1 (for testing only).
131 : */
132 : #define SEGMENT_CHARS OIDCHARS
133 : #define MD_PATH_STR_MAXLEN \
134 : (\
135 : REL_PATH_STR_MAXLEN \
136 : + sizeof((char)'.') \
137 : + SEGMENT_CHARS \
138 : )
139 : typedef struct MdPathStr
140 : {
141 : char str[MD_PATH_STR_MAXLEN + 1];
142 : } MdPathStr;
143 :
144 :
145 : /* local routines */
146 : static void mdunlinkfork(RelFileLocatorBackend rlocator, ForkNumber forknum,
147 : bool isRedo);
148 : static MdfdVec *mdopenfork(SMgrRelation reln, ForkNumber forknum, int behavior);
149 : static void register_dirty_segment(SMgrRelation reln, ForkNumber forknum,
150 : MdfdVec *seg);
151 : static void register_unlink_segment(RelFileLocatorBackend rlocator, ForkNumber forknum,
152 : BlockNumber segno);
153 : static void register_forget_request(RelFileLocatorBackend rlocator, ForkNumber forknum,
154 : BlockNumber segno);
155 : static void _fdvec_resize(SMgrRelation reln,
156 : ForkNumber forknum,
157 : int nseg);
158 : static MdPathStr _mdfd_segpath(SMgrRelation reln, ForkNumber forknum,
159 : BlockNumber segno);
160 : static MdfdVec *_mdfd_openseg(SMgrRelation reln, ForkNumber forknum,
161 : BlockNumber segno, int oflags);
162 : static MdfdVec *_mdfd_getseg(SMgrRelation reln, ForkNumber forknum,
163 : BlockNumber blkno, bool skipFsync, int behavior);
164 : static BlockNumber _mdnblocks(SMgrRelation reln, ForkNumber forknum,
165 : MdfdVec *seg);
166 :
167 : static PgAioResult md_readv_complete(PgAioHandle *ioh, PgAioResult prior_result, uint8 cb_data);
168 : static void md_readv_report(PgAioResult result, const PgAioTargetData *td, int elevel);
169 :
170 : const PgAioHandleCallbacks aio_md_readv_cb = {
171 : .complete_shared = md_readv_complete,
172 : .report = md_readv_report,
173 : };
174 :
175 :
176 : static inline int
177 1767610 : _mdfd_open_flags(void)
178 : {
179 1767610 : int flags = O_RDWR | PG_BINARY;
180 :
181 1767610 : if (io_direct_flags & IO_DIRECT_DATA)
182 341 : flags |= PG_O_DIRECT;
183 :
184 1767610 : return flags;
185 : }
186 :
187 : /*
188 : * mdinit() -- Initialize private state for magnetic disk storage manager.
189 : */
190 : void
191 24173 : mdinit(void)
192 : {
193 24173 : MdCxt = AllocSetContextCreate(TopMemoryContext,
194 : "MdSmgr",
195 : ALLOCSET_DEFAULT_SIZES);
196 24173 : }
197 :
198 : /*
199 : * mdexists() -- Does the physical file exist?
200 : *
201 : * Note: this will return true for lingering files, with pending deletions
202 : */
203 : bool
204 735807 : mdexists(SMgrRelation reln, ForkNumber forknum)
205 : {
206 : /*
207 : * Close it first, to ensure that we notice if the fork has been unlinked
208 : * since we opened it. As an optimization, we can skip that in recovery,
209 : * which already closes relations when dropping them.
210 : */
211 735807 : if (!InRecovery)
212 713981 : mdclose(reln, forknum);
213 :
214 735807 : return (mdopenfork(reln, forknum, EXTENSION_RETURN_NULL) != NULL);
215 : }
216 :
217 : /*
218 : * mdcreate() -- Create a new relation on magnetic disk.
219 : *
220 : * If isRedo is true, it's okay for the relation to exist already.
221 : */
222 : void
223 5837716 : mdcreate(SMgrRelation reln, ForkNumber forknum, bool isRedo)
224 : {
225 : MdfdVec *mdfd;
226 : RelPathStr path;
227 : File fd;
228 :
229 5837716 : if (isRedo && reln->md_num_open_segs[forknum] > 0)
230 5638998 : return; /* created and opened already... */
231 :
232 : Assert(reln->md_num_open_segs[forknum] == 0);
233 :
234 : /*
235 : * We may be using the target table space for the first time in this
236 : * database, so create a per-database subdirectory if needed.
237 : *
238 : * XXX this is a fairly ugly violation of module layering, but this seems
239 : * to be the best place to put the check. Maybe TablespaceCreateDbspace
240 : * should be here and not in commands/tablespace.c? But that would imply
241 : * importing a lot of stuff that smgr.c oughtn't know, either.
242 : */
243 198718 : TablespaceCreateDbspace(reln->smgr_rlocator.locator.spcOid,
244 : reln->smgr_rlocator.locator.dbOid,
245 : isRedo);
246 :
247 198718 : path = relpath(reln->smgr_rlocator, forknum);
248 :
249 198718 : fd = PathNameOpenFile(path.str, _mdfd_open_flags() | O_CREAT | O_EXCL);
250 :
251 198718 : if (fd < 0)
252 : {
253 13877 : int save_errno = errno;
254 :
255 13877 : if (isRedo)
256 13877 : fd = PathNameOpenFile(path.str, _mdfd_open_flags());
257 13877 : if (fd < 0)
258 : {
259 : /* be sure to report the error reported by create, not open */
260 0 : errno = save_errno;
261 0 : ereport(ERROR,
262 : (errcode_for_file_access(),
263 : errmsg("could not create file \"%s\": %m", path.str)));
264 : }
265 : }
266 :
267 198718 : _fdvec_resize(reln, forknum, 1);
268 198718 : mdfd = &reln->md_seg_fds[forknum][0];
269 198718 : mdfd->mdfd_vfd = fd;
270 198718 : mdfd->mdfd_segno = 0;
271 :
272 198718 : if (!SmgrIsTemp(reln))
273 194127 : register_dirty_segment(reln, forknum, mdfd);
274 : }
275 :
276 : /*
277 : * mdunlink() -- Unlink a relation.
278 : *
279 : * Note that we're passed a RelFileLocatorBackend --- by the time this is called,
280 : * there won't be an SMgrRelation hashtable entry anymore.
281 : *
282 : * forknum can be a fork number to delete a specific fork, or InvalidForkNumber
283 : * to delete all forks.
284 : *
285 : * For regular relations, we don't unlink the first segment file of the rel,
286 : * but just truncate it to zero length, and record a request to unlink it after
287 : * the next checkpoint. Additional segments can be unlinked immediately,
288 : * however. Leaving the empty file in place prevents that relfilenumber
289 : * from being reused. The scenario this protects us from is:
290 : * 1. We delete a relation (and commit, and actually remove its file).
291 : * 2. We create a new relation, which by chance gets the same relfilenumber as
292 : * the just-deleted one (OIDs must've wrapped around for that to happen).
293 : * 3. We crash before another checkpoint occurs.
294 : * During replay, we would delete the file and then recreate it, which is fine
295 : * if the contents of the file were repopulated by subsequent WAL entries.
296 : * But if we didn't WAL-log insertions, but instead relied on fsyncing the
297 : * file after populating it (as we do at wal_level=minimal), the contents of
298 : * the file would be lost forever. By leaving the empty file until after the
299 : * next checkpoint, we prevent reassignment of the relfilenumber until it's
300 : * safe, because relfilenumber assignment skips over any existing file.
301 : *
302 : * Additional segments, if any, are truncated and then unlinked. The reason
303 : * for truncating is that other backends may still hold open FDs for these at
304 : * the smgr level, so that the kernel can't remove the file yet. We want to
305 : * reclaim the disk space right away despite that.
306 : *
307 : * We do not need to go through this dance for temp relations, though, because
308 : * we never make WAL entries for temp rels, and so a temp rel poses no threat
309 : * to the health of a regular rel that has taken over its relfilenumber.
310 : * The fact that temp rels and regular rels have different file naming
311 : * patterns provides additional safety. Other backends shouldn't have open
312 : * FDs for them, either.
313 : *
314 : * We also don't do it while performing a binary upgrade. There is no reuse
315 : * hazard in that case, since after a crash or even a simple ERROR, the
316 : * upgrade fails and the whole cluster must be recreated from scratch.
317 : * Furthermore, it is important to remove the files from disk immediately,
318 : * because we may be about to reuse the same relfilenumber.
319 : *
320 : * All the above applies only to the relation's main fork; other forks can
321 : * just be removed immediately, since they are not needed to prevent the
322 : * relfilenumber from being recycled. Also, we do not carefully
323 : * track whether other forks have been created or not, but just attempt to
324 : * unlink them unconditionally; so we should never complain about ENOENT.
325 : *
326 : * If isRedo is true, it's unsurprising for the relation to be already gone.
327 : * Also, we should remove the file immediately instead of queuing a request
328 : * for later, since during redo there's no possibility of creating a
329 : * conflicting relation.
330 : *
331 : * Note: we currently just never warn about ENOENT at all. We could warn in
332 : * the main-fork, non-isRedo case, but it doesn't seem worth the trouble.
333 : *
334 : * Note: any failure should be reported as WARNING not ERROR, because
335 : * we are usually not in a transaction anymore when this is called.
336 : */
337 : void
338 232296 : mdunlink(RelFileLocatorBackend rlocator, ForkNumber forknum, bool isRedo)
339 : {
340 : /* Now do the per-fork work */
341 232296 : if (forknum == InvalidForkNumber)
342 : {
343 0 : for (forknum = 0; forknum <= MAX_FORKNUM; forknum++)
344 0 : mdunlinkfork(rlocator, forknum, isRedo);
345 : }
346 : else
347 232296 : mdunlinkfork(rlocator, forknum, isRedo);
348 232296 : }
349 :
350 : /*
351 : * Truncate a file to release disk space.
352 : */
353 : static int
354 270837 : do_truncate(const char *path)
355 : {
356 : int save_errno;
357 : int ret;
358 :
359 270837 : ret = pg_truncate(path, 0);
360 :
361 : /* Log a warning here to avoid repetition in callers. */
362 270837 : if (ret < 0 && errno != ENOENT)
363 : {
364 0 : save_errno = errno;
365 0 : ereport(WARNING,
366 : (errcode_for_file_access(),
367 : errmsg("could not truncate file \"%s\": %m", path)));
368 0 : errno = save_errno;
369 : }
370 :
371 270837 : return ret;
372 : }
373 :
374 : static void
375 232296 : mdunlinkfork(RelFileLocatorBackend rlocator, ForkNumber forknum, bool isRedo)
376 : {
377 : RelPathStr path;
378 : int ret;
379 : int save_errno;
380 :
381 232296 : path = relpath(rlocator, forknum);
382 :
383 : /*
384 : * Truncate and then unlink the first segment, or just register a request
385 : * to unlink it later, as described in the comments for mdunlink().
386 : */
387 232296 : if (isRedo || IsBinaryUpgrade || forknum != MAIN_FORKNUM ||
388 49824 : RelFileLocatorBackendIsTemp(rlocator))
389 : {
390 186838 : if (!RelFileLocatorBackendIsTemp(rlocator))
391 : {
392 : /* Prevent other backends' fds from holding on to the disk space */
393 169374 : ret = do_truncate(path.str);
394 :
395 : /* Forget any pending sync requests for the first segment */
396 169374 : save_errno = errno;
397 169374 : register_forget_request(rlocator, forknum, 0 /* first seg */ );
398 169374 : errno = save_errno;
399 : }
400 : else
401 17464 : ret = 0;
402 :
403 : /* Next unlink the file, unless it was already found to be missing */
404 186838 : if (ret >= 0 || errno != ENOENT)
405 : {
406 28016 : ret = unlink(path.str);
407 28016 : if (ret < 0 && errno != ENOENT)
408 : {
409 0 : save_errno = errno;
410 0 : ereport(WARNING,
411 : (errcode_for_file_access(),
412 : errmsg("could not remove file \"%s\": %m", path.str)));
413 0 : errno = save_errno;
414 : }
415 : }
416 : }
417 : else
418 : {
419 : /* Prevent other backends' fds from holding on to the disk space */
420 45458 : ret = do_truncate(path.str);
421 :
422 : /* Register request to unlink first segment later */
423 45458 : save_errno = errno;
424 45458 : register_unlink_segment(rlocator, forknum, 0 /* first seg */ );
425 45458 : errno = save_errno;
426 : }
427 :
428 : /*
429 : * Delete any additional segments.
430 : *
431 : * Note that because we loop until getting ENOENT, we will correctly
432 : * remove all inactive segments as well as active ones. Ideally we'd
433 : * continue the loop until getting exactly that errno, but that risks an
434 : * infinite loop if the problem is directory-wide (for instance, if we
435 : * suddenly can't read the data directory itself). We compromise by
436 : * continuing after a non-ENOENT truncate error, but stopping after any
437 : * unlink error. If there is indeed a directory-wide problem, additional
438 : * unlink attempts wouldn't work anyway.
439 : */
440 232296 : if (ret >= 0 || errno != ENOENT)
441 : {
442 : MdPathStr segpath;
443 : BlockNumber segno;
444 :
445 60596 : for (segno = 1;; segno++)
446 : {
447 60596 : sprintf(segpath.str, "%s.%u", path.str, segno);
448 :
449 60596 : if (!RelFileLocatorBackendIsTemp(rlocator))
450 : {
451 : /*
452 : * Prevent other backends' fds from holding on to the disk
453 : * space. We're done if we see ENOENT, though.
454 : */
455 56005 : if (do_truncate(segpath.str) < 0 && errno == ENOENT)
456 56005 : break;
457 :
458 : /*
459 : * Forget any pending sync requests for this segment before we
460 : * try to unlink.
461 : */
462 0 : register_forget_request(rlocator, forknum, segno);
463 : }
464 :
465 4591 : if (unlink(segpath.str) < 0)
466 : {
467 : /* ENOENT is expected after the last segment... */
468 4591 : if (errno != ENOENT)
469 0 : ereport(WARNING,
470 : (errcode_for_file_access(),
471 : errmsg("could not remove file \"%s\": %m", segpath.str)));
472 4591 : break;
473 : }
474 : }
475 : }
476 232296 : }
477 :
478 : /*
479 : * mdextend() -- Add a block to the specified relation.
480 : *
481 : * The semantics are nearly the same as mdwrite(): write at the
482 : * specified position. However, this is to be used for the case of
483 : * extending a relation (i.e., blocknum is at or beyond the current
484 : * EOF). Note that we assume writing a block beyond current EOF
485 : * causes intervening file space to become filled with zeroes.
486 : */
487 : void
488 141002 : mdextend(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
489 : const void *buffer, bool skipFsync)
490 : {
491 : pgoff_t seekpos;
492 : int nbytes;
493 : MdfdVec *v;
494 :
495 : /* If this build supports direct I/O, the buffer must be I/O aligned. */
496 : if (PG_O_DIRECT != 0 && PG_IO_ALIGN_SIZE <= BLCKSZ)
497 : Assert((uintptr_t) buffer == TYPEALIGN(PG_IO_ALIGN_SIZE, buffer));
498 :
499 : /* This assert is too expensive to have on normally ... */
500 : #ifdef CHECK_WRITE_VS_EXTEND
501 : Assert(blocknum >= mdnblocks(reln, forknum));
502 : #endif
503 :
504 : /*
505 : * If a relation manages to grow to 2^32-1 blocks, refuse to extend it any
506 : * more --- we mustn't create a block whose number actually is
507 : * InvalidBlockNumber. (Note that this failure should be unreachable
508 : * because of upstream checks in bufmgr.c.)
509 : */
510 141002 : if (blocknum == InvalidBlockNumber)
511 0 : ereport(ERROR,
512 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
513 : errmsg("cannot extend file \"%s\" beyond %u blocks",
514 : relpath(reln->smgr_rlocator, forknum).str,
515 : InvalidBlockNumber)));
516 :
517 141002 : v = _mdfd_getseg(reln, forknum, blocknum, skipFsync, EXTENSION_CREATE);
518 :
519 141002 : seekpos = (pgoff_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
520 :
521 : Assert(seekpos < (pgoff_t) BLCKSZ * RELSEG_SIZE);
522 :
523 141002 : if ((nbytes = FileWrite(v->mdfd_vfd, buffer, BLCKSZ, seekpos, WAIT_EVENT_DATA_FILE_EXTEND)) != BLCKSZ)
524 : {
525 0 : if (nbytes < 0)
526 0 : ereport(ERROR,
527 : (errcode_for_file_access(),
528 : errmsg("could not extend file \"%s\": %m",
529 : FilePathName(v->mdfd_vfd)),
530 : errhint("Check free disk space.")));
531 : /* short write: complain appropriately */
532 0 : ereport(ERROR,
533 : (errcode(ERRCODE_DISK_FULL),
534 : errmsg("could not extend file \"%s\": wrote only %d of %d bytes at block %u",
535 : FilePathName(v->mdfd_vfd),
536 : nbytes, BLCKSZ, blocknum),
537 : errhint("Check free disk space.")));
538 : }
539 :
540 141002 : if (!skipFsync && !SmgrIsTemp(reln))
541 38 : register_dirty_segment(reln, forknum, v);
542 :
543 : Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
544 141002 : }
545 :
546 : /*
547 : * mdzeroextend() -- Add new zeroed out blocks to the specified relation.
548 : *
549 : * Similar to mdextend(), except the relation can be extended by multiple
550 : * blocks at once and the added blocks will be filled with zeroes.
551 : */
552 : void
553 245431 : mdzeroextend(SMgrRelation reln, ForkNumber forknum,
554 : BlockNumber blocknum, int nblocks, bool skipFsync)
555 : {
556 : MdfdVec *v;
557 245431 : BlockNumber curblocknum = blocknum;
558 245431 : int remblocks = nblocks;
559 :
560 : Assert(nblocks > 0);
561 :
562 : /* This assert is too expensive to have on normally ... */
563 : #ifdef CHECK_WRITE_VS_EXTEND
564 : Assert(blocknum >= mdnblocks(reln, forknum));
565 : #endif
566 :
567 : /*
568 : * If a relation manages to grow to 2^32-1 blocks, refuse to extend it any
569 : * more --- we mustn't create a block whose number actually is
570 : * InvalidBlockNumber or larger.
571 : */
572 245431 : if ((uint64) blocknum + nblocks >= (uint64) InvalidBlockNumber)
573 0 : ereport(ERROR,
574 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
575 : errmsg("cannot extend file \"%s\" beyond %u blocks",
576 : relpath(reln->smgr_rlocator, forknum).str,
577 : InvalidBlockNumber)));
578 :
579 490862 : while (remblocks > 0)
580 : {
581 245431 : BlockNumber segstartblock = curblocknum % ((BlockNumber) RELSEG_SIZE);
582 245431 : pgoff_t seekpos = (pgoff_t) BLCKSZ * segstartblock;
583 : int numblocks;
584 :
585 245431 : if (segstartblock + remblocks > RELSEG_SIZE)
586 0 : numblocks = RELSEG_SIZE - segstartblock;
587 : else
588 245431 : numblocks = remblocks;
589 :
590 245431 : v = _mdfd_getseg(reln, forknum, curblocknum, skipFsync, EXTENSION_CREATE);
591 :
592 : Assert(segstartblock < RELSEG_SIZE);
593 : Assert(segstartblock + numblocks <= RELSEG_SIZE);
594 :
595 : /*
596 : * If available and useful, use posix_fallocate() (via
597 : * FileFallocate()) to extend the relation. That's often more
598 : * efficient than using write(), as it commonly won't cause the kernel
599 : * to allocate page cache space for the extended pages.
600 : *
601 : * However, we don't use FileFallocate() for small extensions, as it
602 : * defeats delayed allocation on some filesystems. Not clear where
603 : * that decision should be made though? For now just use a cutoff of
604 : * 8, anything between 4 and 8 worked OK in some local testing.
605 : */
606 245431 : if (numblocks > 8 &&
607 693 : file_extend_method != FILE_EXTEND_METHOD_WRITE_ZEROS)
608 693 : {
609 693 : int ret = 0;
610 :
611 : #ifdef HAVE_POSIX_FALLOCATE
612 693 : if (file_extend_method == FILE_EXTEND_METHOD_POSIX_FALLOCATE)
613 : {
614 693 : ret = FileFallocate(v->mdfd_vfd,
615 : seekpos, (pgoff_t) BLCKSZ * numblocks,
616 : WAIT_EVENT_DATA_FILE_EXTEND);
617 : }
618 : else
619 : #endif
620 : {
621 0 : elog(ERROR, "unsupported file_extend_method: %d",
622 : file_extend_method);
623 : }
624 693 : if (ret != 0)
625 : {
626 0 : ereport(ERROR,
627 : errcode_for_file_access(),
628 : errmsg("could not extend file \"%s\" with FileFallocate(): %m",
629 : FilePathName(v->mdfd_vfd)),
630 : errhint("Check free disk space."));
631 : }
632 : }
633 : else
634 : {
635 : int ret;
636 :
637 : /*
638 : * Even if we don't want to use fallocate, we can still extend a
639 : * bit more efficiently than writing each 8kB block individually.
640 : * pg_pwrite_zeros() (via FileZero()) uses pg_pwritev_with_retry()
641 : * to avoid multiple writes or needing a zeroed buffer for the
642 : * whole length of the extension.
643 : */
644 244738 : ret = FileZero(v->mdfd_vfd,
645 : seekpos, (pgoff_t) BLCKSZ * numblocks,
646 : WAIT_EVENT_DATA_FILE_EXTEND);
647 244738 : if (ret < 0)
648 0 : ereport(ERROR,
649 : errcode_for_file_access(),
650 : errmsg("could not extend file \"%s\": %m",
651 : FilePathName(v->mdfd_vfd)),
652 : errhint("Check free disk space."));
653 : }
654 :
655 245431 : if (!skipFsync && !SmgrIsTemp(reln))
656 230731 : register_dirty_segment(reln, forknum, v);
657 :
658 : Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
659 :
660 245431 : remblocks -= numblocks;
661 245431 : curblocknum += numblocks;
662 : }
663 245431 : }
664 :
665 : /*
666 : * mdopenfork() -- Open one fork of the specified relation.
667 : *
668 : * Note we only open the first segment, when there are multiple segments.
669 : *
670 : * If first segment is not present, either ereport or return NULL according
671 : * to "behavior". We treat EXTENSION_CREATE the same as EXTENSION_FAIL;
672 : * EXTENSION_CREATE means it's OK to extend an existing relation, not to
673 : * invent one out of whole cloth.
674 : */
675 : static MdfdVec *
676 4951797 : mdopenfork(SMgrRelation reln, ForkNumber forknum, int behavior)
677 : {
678 : MdfdVec *mdfd;
679 : RelPathStr path;
680 : File fd;
681 :
682 : /* No work if already open */
683 4951797 : if (reln->md_num_open_segs[forknum] > 0)
684 3427197 : return &reln->md_seg_fds[forknum][0];
685 :
686 1524600 : path = relpath(reln->smgr_rlocator, forknum);
687 :
688 1524600 : fd = PathNameOpenFile(path.str, _mdfd_open_flags());
689 :
690 1524600 : if (fd < 0)
691 : {
692 484361 : if ((behavior & EXTENSION_RETURN_NULL) &&
693 484339 : FILE_POSSIBLY_DELETED(errno))
694 484339 : return NULL;
695 22 : ereport(ERROR,
696 : (errcode_for_file_access(),
697 : errmsg("could not open file \"%s\": %m", path.str)));
698 : }
699 :
700 1040239 : _fdvec_resize(reln, forknum, 1);
701 1040239 : mdfd = &reln->md_seg_fds[forknum][0];
702 1040239 : mdfd->mdfd_vfd = fd;
703 1040239 : mdfd->mdfd_segno = 0;
704 :
705 : Assert(_mdnblocks(reln, forknum, mdfd) <= ((BlockNumber) RELSEG_SIZE));
706 :
707 1040239 : return mdfd;
708 : }
709 :
710 : /*
711 : * mdopen() -- Initialize newly-opened relation.
712 : */
713 : void
714 1399596 : mdopen(SMgrRelation reln)
715 : {
716 : /* mark it not open */
717 6997980 : for (int forknum = 0; forknum <= MAX_FORKNUM; forknum++)
718 5598384 : reln->md_num_open_segs[forknum] = 0;
719 1399596 : }
720 :
721 : /*
722 : * mdclose() -- Close the specified relation, if it isn't closed already.
723 : */
724 : void
725 4535625 : mdclose(SMgrRelation reln, ForkNumber forknum)
726 : {
727 4535625 : int nopensegs = reln->md_num_open_segs[forknum];
728 :
729 : /* No work if already closed */
730 4535625 : if (nopensegs == 0)
731 3882522 : return;
732 :
733 : /* close segments starting from the end */
734 1306206 : while (nopensegs > 0)
735 : {
736 653103 : MdfdVec *v = &reln->md_seg_fds[forknum][nopensegs - 1];
737 :
738 653103 : FileClose(v->mdfd_vfd);
739 653103 : _fdvec_resize(reln, forknum, nopensegs - 1);
740 653103 : nopensegs--;
741 : }
742 : }
743 :
744 : /*
745 : * mdprefetch() -- Initiate asynchronous read of the specified blocks of a relation
746 : */
747 : bool
748 8801 : mdprefetch(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
749 : int nblocks)
750 : {
751 : #ifdef USE_PREFETCH
752 :
753 : Assert((io_direct_flags & IO_DIRECT_DATA) == 0);
754 :
755 8801 : if ((uint64) blocknum + nblocks > (uint64) MaxBlockNumber + 1)
756 0 : return false;
757 :
758 17602 : while (nblocks > 0)
759 : {
760 : pgoff_t seekpos;
761 : MdfdVec *v;
762 : int nblocks_this_segment;
763 :
764 8801 : v = _mdfd_getseg(reln, forknum, blocknum, false,
765 8801 : InRecovery ? EXTENSION_RETURN_NULL : EXTENSION_FAIL);
766 8801 : if (v == NULL)
767 0 : return false;
768 :
769 8801 : seekpos = (pgoff_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
770 :
771 : Assert(seekpos < (pgoff_t) BLCKSZ * RELSEG_SIZE);
772 :
773 8801 : nblocks_this_segment =
774 8801 : Min(nblocks,
775 : RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
776 :
777 8801 : (void) FilePrefetch(v->mdfd_vfd, seekpos, BLCKSZ * nblocks_this_segment,
778 : WAIT_EVENT_DATA_FILE_PREFETCH);
779 :
780 8801 : blocknum += nblocks_this_segment;
781 8801 : nblocks -= nblocks_this_segment;
782 : }
783 : #endif /* USE_PREFETCH */
784 :
785 8801 : return true;
786 : }
787 :
788 : /*
789 : * Convert an array of buffer address into an array of iovec objects, and
790 : * return the number that were required. 'iov' must have enough space for up
791 : * to 'nblocks' elements, but the number used may be less depending on
792 : * merging. In the case of a run of fully contiguous buffers, a single iovec
793 : * will be populated that can be handled as a plain non-vectored I/O.
794 : */
795 : static int
796 1987043 : buffers_to_iovec(struct iovec *iov, void **buffers, int nblocks)
797 : {
798 : struct iovec *iovp;
799 : int iovcnt;
800 :
801 : Assert(nblocks >= 1);
802 :
803 : /* If this build supports direct I/O, buffers must be I/O aligned. */
804 4159268 : for (int i = 0; i < nblocks; ++i)
805 : {
806 : if (PG_O_DIRECT != 0 && PG_IO_ALIGN_SIZE <= BLCKSZ)
807 : Assert((uintptr_t) buffers[i] ==
808 : TYPEALIGN(PG_IO_ALIGN_SIZE, buffers[i]));
809 : }
810 :
811 : /* Start the first iovec off with the first buffer. */
812 1987043 : iovp = &iov[0];
813 1987043 : iovp->iov_base = buffers[0];
814 1987043 : iovp->iov_len = BLCKSZ;
815 1987043 : iovcnt = 1;
816 :
817 : /* Try to merge the rest. */
818 2172225 : for (int i = 1; i < nblocks; ++i)
819 : {
820 185182 : void *buffer = buffers[i];
821 :
822 185182 : if (((char *) iovp->iov_base + iovp->iov_len) == buffer)
823 : {
824 : /* Contiguous with the last iovec. */
825 184140 : iovp->iov_len += BLCKSZ;
826 : }
827 : else
828 : {
829 : /* Need a new iovec. */
830 1042 : iovp++;
831 1042 : iovp->iov_base = buffer;
832 1042 : iovp->iov_len = BLCKSZ;
833 1042 : iovcnt++;
834 : }
835 : }
836 :
837 1987043 : return iovcnt;
838 : }
839 :
840 : /*
841 : * mdmaxcombine() -- Return the maximum number of total blocks that can be
842 : * combined with an IO starting at blocknum.
843 : */
844 : uint32
845 37431 : mdmaxcombine(SMgrRelation reln, ForkNumber forknum,
846 : BlockNumber blocknum)
847 : {
848 : BlockNumber segoff;
849 :
850 37431 : segoff = blocknum % ((BlockNumber) RELSEG_SIZE);
851 :
852 37431 : return RELSEG_SIZE - segoff;
853 : }
854 :
855 : /*
856 : * mdreadv() -- Read the specified blocks from a relation.
857 : */
858 : void
859 783 : mdreadv(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
860 : void **buffers, BlockNumber nblocks)
861 : {
862 1566 : while (nblocks > 0)
863 : {
864 : struct iovec iov[PG_IOV_MAX];
865 : int iovcnt;
866 : pgoff_t seekpos;
867 : int nbytes;
868 : MdfdVec *v;
869 : BlockNumber nblocks_this_segment;
870 : size_t transferred_this_segment;
871 : size_t size_this_segment;
872 :
873 783 : v = _mdfd_getseg(reln, forknum, blocknum, false,
874 : EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
875 :
876 783 : seekpos = (pgoff_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
877 :
878 : Assert(seekpos < (pgoff_t) BLCKSZ * RELSEG_SIZE);
879 :
880 783 : nblocks_this_segment =
881 783 : Min(nblocks,
882 : RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
883 783 : nblocks_this_segment = Min(nblocks_this_segment, lengthof(iov));
884 :
885 783 : if (nblocks_this_segment != nblocks)
886 0 : elog(ERROR, "read crosses segment boundary");
887 :
888 783 : iovcnt = buffers_to_iovec(iov, buffers, nblocks_this_segment);
889 783 : size_this_segment = nblocks_this_segment * BLCKSZ;
890 783 : transferred_this_segment = 0;
891 :
892 : /*
893 : * Inner loop to continue after a short read. We'll keep going until
894 : * we hit EOF rather than assuming that a short read means we hit the
895 : * end.
896 : */
897 : for (;;)
898 : {
899 0 : TRACE_POSTGRESQL_SMGR_MD_READ_START(forknum, blocknum,
900 : reln->smgr_rlocator.locator.spcOid,
901 : reln->smgr_rlocator.locator.dbOid,
902 : reln->smgr_rlocator.locator.relNumber,
903 : reln->smgr_rlocator.backend);
904 783 : nbytes = FileReadV(v->mdfd_vfd, iov, iovcnt, seekpos,
905 : WAIT_EVENT_DATA_FILE_READ);
906 : TRACE_POSTGRESQL_SMGR_MD_READ_DONE(forknum, blocknum,
907 : reln->smgr_rlocator.locator.spcOid,
908 : reln->smgr_rlocator.locator.dbOid,
909 : reln->smgr_rlocator.locator.relNumber,
910 : reln->smgr_rlocator.backend,
911 : nbytes,
912 : size_this_segment - transferred_this_segment);
913 :
914 : #ifdef SIMULATE_SHORT_READ
915 : nbytes = Min(nbytes, 4096);
916 : #endif
917 :
918 783 : if (nbytes < 0)
919 0 : ereport(ERROR,
920 : (errcode_for_file_access(),
921 : errmsg("could not read blocks %u..%u in file \"%s\": %m",
922 : blocknum,
923 : blocknum + nblocks_this_segment - 1,
924 : FilePathName(v->mdfd_vfd))));
925 :
926 783 : if (nbytes == 0)
927 : {
928 : /*
929 : * We are at or past EOF, or we read a partial block at EOF.
930 : * Normally this is an error; upper levels should never try to
931 : * read a nonexistent block. However, if zero_damaged_pages
932 : * is ON or we are InRecovery, we should instead return zeroes
933 : * without complaining. This allows, for example, the case of
934 : * trying to update a block that was later truncated away.
935 : *
936 : * NB: We think that this codepath is unreachable in recovery
937 : * and incomplete with zero_damaged_pages, as missing segments
938 : * are not created. Putting blocks into the buffer-pool that
939 : * do not exist on disk is rather problematic, as it will not
940 : * be found by scans that rely on smgrnblocks(), as they are
941 : * beyond EOF. It also can cause weird problems with relation
942 : * extension, as relation extension does not expect blocks
943 : * beyond EOF to exist.
944 : *
945 : * Therefore we do not want to copy the logic into
946 : * mdstartreadv(), where it would have to be more complicated
947 : * due to potential differences in the zero_damaged_pages
948 : * setting between the definer and completor of IO.
949 : *
950 : * For PG 18, we are putting an Assert(false) in mdreadv()
951 : * (triggering failures in assertion-enabled builds, but
952 : * continuing to work in production builds). Afterwards we
953 : * plan to remove this code entirely.
954 : */
955 0 : if (zero_damaged_pages || InRecovery)
956 : {
957 : Assert(false); /* see comment above */
958 :
959 0 : for (BlockNumber i = transferred_this_segment / BLCKSZ;
960 0 : i < nblocks_this_segment;
961 0 : ++i)
962 0 : memset(buffers[i], 0, BLCKSZ);
963 0 : break;
964 : }
965 : else
966 0 : ereport(ERROR,
967 : (errcode(ERRCODE_DATA_CORRUPTED),
968 : errmsg("could not read blocks %u..%u in file \"%s\": read only %zu of %zu bytes",
969 : blocknum,
970 : blocknum + nblocks_this_segment - 1,
971 : FilePathName(v->mdfd_vfd),
972 : transferred_this_segment,
973 : size_this_segment)));
974 : }
975 :
976 : /* One loop should usually be enough. */
977 783 : transferred_this_segment += nbytes;
978 : Assert(transferred_this_segment <= size_this_segment);
979 783 : if (transferred_this_segment == size_this_segment)
980 783 : break;
981 :
982 : /* Adjust position and vectors after a short read. */
983 0 : seekpos += nbytes;
984 0 : iovcnt = compute_remaining_iovec(iov, iov, iovcnt, nbytes);
985 : }
986 :
987 783 : nblocks -= nblocks_this_segment;
988 783 : buffers += nblocks_this_segment;
989 783 : blocknum += nblocks_this_segment;
990 : }
991 783 : }
992 :
993 : /*
994 : * mdstartreadv() -- Asynchronous version of mdreadv().
995 : */
996 : void
997 1373710 : mdstartreadv(PgAioHandle *ioh,
998 : SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
999 : void **buffers, BlockNumber nblocks)
1000 : {
1001 : pgoff_t seekpos;
1002 : MdfdVec *v;
1003 : BlockNumber nblocks_this_segment;
1004 : struct iovec *iov;
1005 : int iovcnt;
1006 : int ret;
1007 :
1008 1373710 : v = _mdfd_getseg(reln, forknum, blocknum, false,
1009 : EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
1010 :
1011 1373695 : seekpos = (pgoff_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
1012 :
1013 : Assert(seekpos < (pgoff_t) BLCKSZ * RELSEG_SIZE);
1014 :
1015 1373695 : nblocks_this_segment =
1016 1373695 : Min(nblocks,
1017 : RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
1018 :
1019 1373695 : if (nblocks_this_segment != nblocks)
1020 0 : elog(ERROR, "read crossing segment boundary");
1021 :
1022 1373695 : iovcnt = pgaio_io_get_iovec(ioh, &iov);
1023 :
1024 : Assert(nblocks <= iovcnt);
1025 :
1026 1373695 : iovcnt = buffers_to_iovec(iov, buffers, nblocks_this_segment);
1027 :
1028 : Assert(iovcnt <= nblocks_this_segment);
1029 :
1030 1373695 : if (!(io_direct_flags & IO_DIRECT_DATA))
1031 1372195 : pgaio_io_set_flag(ioh, PGAIO_HF_BUFFERED);
1032 :
1033 1373695 : pgaio_io_set_target_smgr(ioh,
1034 : reln,
1035 : forknum,
1036 : blocknum,
1037 : nblocks,
1038 : false);
1039 1373695 : pgaio_io_register_callbacks(ioh, PGAIO_HCB_MD_READV, 0);
1040 :
1041 1373695 : ret = FileStartReadV(ioh, v->mdfd_vfd, iovcnt, seekpos, WAIT_EVENT_DATA_FILE_READ);
1042 1373695 : if (ret != 0)
1043 0 : ereport(ERROR,
1044 : (errcode_for_file_access(),
1045 : errmsg("could not start reading blocks %u..%u in file \"%s\": %m",
1046 : blocknum,
1047 : blocknum + nblocks_this_segment - 1,
1048 : FilePathName(v->mdfd_vfd))));
1049 :
1050 : /*
1051 : * The error checks corresponding to the post-read checks in mdreadv() are
1052 : * in md_readv_complete().
1053 : *
1054 : * However we chose, at least for now, to not implement the
1055 : * zero_damaged_pages logic present in mdreadv(). As outlined in mdreadv()
1056 : * that logic is rather problematic, and we want to get rid of it. Here
1057 : * equivalent logic would have to be more complicated due to potential
1058 : * differences in the zero_damaged_pages setting between the definer and
1059 : * completor of IO.
1060 : */
1061 1373695 : }
1062 :
1063 : /*
1064 : * mdwritev() -- Write the supplied blocks at the appropriate location.
1065 : *
1066 : * This is to be used only for updating already-existing blocks of a
1067 : * relation (ie, those before the current EOF). To extend a relation,
1068 : * use mdextend().
1069 : */
1070 : void
1071 612565 : mdwritev(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
1072 : const void **buffers, BlockNumber nblocks, bool skipFsync)
1073 : {
1074 : /* This assert is too expensive to have on normally ... */
1075 : #ifdef CHECK_WRITE_VS_EXTEND
1076 : Assert((uint64) blocknum + (uint64) nblocks <= (uint64) mdnblocks(reln, forknum));
1077 : #endif
1078 :
1079 1225130 : while (nblocks > 0)
1080 : {
1081 : struct iovec iov[PG_IOV_MAX];
1082 : int iovcnt;
1083 : pgoff_t seekpos;
1084 : int nbytes;
1085 : MdfdVec *v;
1086 : BlockNumber nblocks_this_segment;
1087 : size_t transferred_this_segment;
1088 : size_t size_this_segment;
1089 :
1090 612565 : v = _mdfd_getseg(reln, forknum, blocknum, skipFsync,
1091 : EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
1092 :
1093 612565 : seekpos = (pgoff_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
1094 :
1095 : Assert(seekpos < (pgoff_t) BLCKSZ * RELSEG_SIZE);
1096 :
1097 612565 : nblocks_this_segment =
1098 612565 : Min(nblocks,
1099 : RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
1100 612565 : nblocks_this_segment = Min(nblocks_this_segment, lengthof(iov));
1101 :
1102 612565 : if (nblocks_this_segment != nblocks)
1103 0 : elog(ERROR, "write crosses segment boundary");
1104 :
1105 612565 : iovcnt = buffers_to_iovec(iov, (void **) buffers, nblocks_this_segment);
1106 612565 : size_this_segment = nblocks_this_segment * BLCKSZ;
1107 612565 : transferred_this_segment = 0;
1108 :
1109 : /*
1110 : * Inner loop to continue after a short write. If the reason is that
1111 : * we're out of disk space, a future attempt should get an ENOSPC
1112 : * error from the kernel.
1113 : */
1114 : for (;;)
1115 : {
1116 0 : TRACE_POSTGRESQL_SMGR_MD_WRITE_START(forknum, blocknum,
1117 : reln->smgr_rlocator.locator.spcOid,
1118 : reln->smgr_rlocator.locator.dbOid,
1119 : reln->smgr_rlocator.locator.relNumber,
1120 : reln->smgr_rlocator.backend);
1121 612565 : nbytes = FileWriteV(v->mdfd_vfd, iov, iovcnt, seekpos,
1122 : WAIT_EVENT_DATA_FILE_WRITE);
1123 : TRACE_POSTGRESQL_SMGR_MD_WRITE_DONE(forknum, blocknum,
1124 : reln->smgr_rlocator.locator.spcOid,
1125 : reln->smgr_rlocator.locator.dbOid,
1126 : reln->smgr_rlocator.locator.relNumber,
1127 : reln->smgr_rlocator.backend,
1128 : nbytes,
1129 : size_this_segment - transferred_this_segment);
1130 :
1131 : #ifdef SIMULATE_SHORT_WRITE
1132 : nbytes = Min(nbytes, 4096);
1133 : #endif
1134 :
1135 612565 : if (nbytes < 0)
1136 : {
1137 0 : bool enospc = errno == ENOSPC;
1138 :
1139 0 : ereport(ERROR,
1140 : (errcode_for_file_access(),
1141 : errmsg("could not write blocks %u..%u in file \"%s\": %m",
1142 : blocknum,
1143 : blocknum + nblocks_this_segment - 1,
1144 : FilePathName(v->mdfd_vfd)),
1145 : enospc ? errhint("Check free disk space.") : 0));
1146 : }
1147 :
1148 : /* One loop should usually be enough. */
1149 612565 : transferred_this_segment += nbytes;
1150 : Assert(transferred_this_segment <= size_this_segment);
1151 612565 : if (transferred_this_segment == size_this_segment)
1152 612565 : break;
1153 :
1154 : /* Adjust position and iovecs after a short write. */
1155 0 : seekpos += nbytes;
1156 0 : iovcnt = compute_remaining_iovec(iov, iov, iovcnt, nbytes);
1157 : }
1158 :
1159 612565 : if (!skipFsync && !SmgrIsTemp(reln))
1160 607584 : register_dirty_segment(reln, forknum, v);
1161 :
1162 612565 : nblocks -= nblocks_this_segment;
1163 612565 : buffers += nblocks_this_segment;
1164 612565 : blocknum += nblocks_this_segment;
1165 : }
1166 612565 : }
1167 :
1168 :
1169 : /*
1170 : * mdwriteback() -- Tell the kernel to write pages back to storage.
1171 : *
1172 : * This accepts a range of blocks because flushing several pages at once is
1173 : * considerably more efficient than doing so individually.
1174 : */
1175 : void
1176 0 : mdwriteback(SMgrRelation reln, ForkNumber forknum,
1177 : BlockNumber blocknum, BlockNumber nblocks)
1178 : {
1179 : Assert((io_direct_flags & IO_DIRECT_DATA) == 0);
1180 :
1181 : /*
1182 : * Issue flush requests in as few requests as possible; have to split at
1183 : * segment boundaries though, since those are actually separate files.
1184 : */
1185 0 : while (nblocks > 0)
1186 : {
1187 0 : BlockNumber nflush = nblocks;
1188 : pgoff_t seekpos;
1189 : MdfdVec *v;
1190 : int segnum_start,
1191 : segnum_end;
1192 :
1193 0 : v = _mdfd_getseg(reln, forknum, blocknum, true /* not used */ ,
1194 : EXTENSION_DONT_OPEN);
1195 :
1196 : /*
1197 : * We might be flushing buffers of already removed relations, that's
1198 : * ok, just ignore that case. If the segment file wasn't open already
1199 : * (ie from a recent mdwrite()), then we don't want to re-open it, to
1200 : * avoid a race with PROCSIGNAL_BARRIER_SMGRRELEASE that might leave
1201 : * us with a descriptor to a file that is about to be unlinked.
1202 : */
1203 0 : if (!v)
1204 0 : return;
1205 :
1206 : /* compute offset inside the current segment */
1207 0 : segnum_start = blocknum / RELSEG_SIZE;
1208 :
1209 : /* compute number of desired writes within the current segment */
1210 0 : segnum_end = (blocknum + nblocks - 1) / RELSEG_SIZE;
1211 0 : if (segnum_start != segnum_end)
1212 0 : nflush = RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE));
1213 :
1214 : Assert(nflush >= 1);
1215 : Assert(nflush <= nblocks);
1216 :
1217 0 : seekpos = (pgoff_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
1218 :
1219 0 : FileWriteback(v->mdfd_vfd, seekpos, (pgoff_t) BLCKSZ * nflush, WAIT_EVENT_DATA_FILE_FLUSH);
1220 :
1221 0 : nblocks -= nflush;
1222 0 : blocknum += nflush;
1223 : }
1224 : }
1225 :
1226 : /*
1227 : * mdnblocks() -- Get the number of blocks stored in a relation.
1228 : *
1229 : * Important side effect: all active segments of the relation are opened
1230 : * and added to the md_seg_fds array. If this routine has not been
1231 : * called, then only segments up to the last one actually touched
1232 : * are present in the array.
1233 : */
1234 : BlockNumber
1235 3454258 : mdnblocks(SMgrRelation reln, ForkNumber forknum)
1236 : {
1237 : MdfdVec *v;
1238 : BlockNumber nblocks;
1239 : BlockNumber segno;
1240 :
1241 3454258 : mdopenfork(reln, forknum, EXTENSION_FAIL);
1242 :
1243 : /* mdopen has opened the first segment */
1244 : Assert(reln->md_num_open_segs[forknum] > 0);
1245 :
1246 : /*
1247 : * Start from the last open segments, to avoid redundant seeks. We have
1248 : * previously verified that these segments are exactly RELSEG_SIZE long,
1249 : * and it's useless to recheck that each time.
1250 : *
1251 : * NOTE: this assumption could only be wrong if another backend has
1252 : * truncated the relation. We rely on higher code levels to handle that
1253 : * scenario by closing and re-opening the md fd, which is handled via
1254 : * relcache flush. (Since the checkpointer doesn't participate in
1255 : * relcache flush, it could have segment entries for inactive segments;
1256 : * that's OK because the checkpointer never needs to compute relation
1257 : * size.)
1258 : */
1259 3454239 : segno = reln->md_num_open_segs[forknum] - 1;
1260 3454239 : v = &reln->md_seg_fds[forknum][segno];
1261 :
1262 : for (;;)
1263 : {
1264 3454239 : nblocks = _mdnblocks(reln, forknum, v);
1265 3454239 : if (nblocks > ((BlockNumber) RELSEG_SIZE))
1266 0 : elog(FATAL, "segment too big");
1267 3454239 : if (nblocks < ((BlockNumber) RELSEG_SIZE))
1268 3454239 : return (segno * ((BlockNumber) RELSEG_SIZE)) + nblocks;
1269 :
1270 : /*
1271 : * If segment is exactly RELSEG_SIZE, advance to next one.
1272 : */
1273 0 : segno++;
1274 :
1275 : /*
1276 : * We used to pass O_CREAT here, but that has the disadvantage that it
1277 : * might create a segment which has vanished through some operating
1278 : * system misadventure. In such a case, creating the segment here
1279 : * undermines _mdfd_getseg's attempts to notice and report an error
1280 : * upon access to a missing segment.
1281 : */
1282 0 : v = _mdfd_openseg(reln, forknum, segno, 0);
1283 0 : if (v == NULL)
1284 0 : return segno * ((BlockNumber) RELSEG_SIZE);
1285 : }
1286 : }
1287 :
1288 : /*
1289 : * mdtruncate() -- Truncate relation to specified number of blocks.
1290 : *
1291 : * Guaranteed not to allocate memory, so it can be used in a critical section.
1292 : * Caller must have called smgrnblocks() to obtain curnblk while holding a
1293 : * sufficient lock to prevent a change in relation size, and not used any smgr
1294 : * functions for this relation or handled interrupts in between. This makes
1295 : * sure we have opened all active segments, so that truncate loop will get
1296 : * them all!
1297 : *
1298 : * If nblocks > curnblk, the request is ignored when we are InRecovery,
1299 : * otherwise, an error is raised.
1300 : */
1301 : void
1302 1151 : mdtruncate(SMgrRelation reln, ForkNumber forknum,
1303 : BlockNumber curnblk, BlockNumber nblocks)
1304 : {
1305 : BlockNumber priorblocks;
1306 : int curopensegs;
1307 :
1308 1151 : if (nblocks > curnblk)
1309 : {
1310 : /* Bogus request ... but no complaint if InRecovery */
1311 0 : if (InRecovery)
1312 0 : return;
1313 0 : ereport(ERROR,
1314 : (errmsg("could not truncate file \"%s\" to %u blocks: it's only %u blocks now",
1315 : relpath(reln->smgr_rlocator, forknum).str,
1316 : nblocks, curnblk)));
1317 : }
1318 1151 : if (nblocks == curnblk)
1319 460 : return; /* no work */
1320 :
1321 : /*
1322 : * Truncate segments, starting at the last one. Starting at the end makes
1323 : * managing the memory for the fd array easier, should there be errors.
1324 : */
1325 691 : curopensegs = reln->md_num_open_segs[forknum];
1326 1382 : while (curopensegs > 0)
1327 : {
1328 : MdfdVec *v;
1329 :
1330 691 : priorblocks = (curopensegs - 1) * RELSEG_SIZE;
1331 :
1332 691 : v = &reln->md_seg_fds[forknum][curopensegs - 1];
1333 :
1334 691 : if (priorblocks > nblocks)
1335 : {
1336 : /*
1337 : * This segment is no longer active. We truncate the file, but do
1338 : * not delete it, for reasons explained in the header comments.
1339 : */
1340 0 : if (FileTruncate(v->mdfd_vfd, 0, WAIT_EVENT_DATA_FILE_TRUNCATE) < 0)
1341 0 : ereport(ERROR,
1342 : (errcode_for_file_access(),
1343 : errmsg("could not truncate file \"%s\": %m",
1344 : FilePathName(v->mdfd_vfd))));
1345 :
1346 0 : if (!SmgrIsTemp(reln))
1347 0 : register_dirty_segment(reln, forknum, v);
1348 :
1349 : /* we never drop the 1st segment */
1350 : Assert(v != &reln->md_seg_fds[forknum][0]);
1351 :
1352 0 : FileClose(v->mdfd_vfd);
1353 0 : _fdvec_resize(reln, forknum, curopensegs - 1);
1354 : }
1355 691 : else if (priorblocks + ((BlockNumber) RELSEG_SIZE) > nblocks)
1356 : {
1357 : /*
1358 : * This is the last segment we want to keep. Truncate the file to
1359 : * the right length. NOTE: if nblocks is exactly a multiple K of
1360 : * RELSEG_SIZE, we will truncate the K+1st segment to 0 length but
1361 : * keep it. This adheres to the invariant given in the header
1362 : * comments.
1363 : */
1364 691 : BlockNumber lastsegblocks = nblocks - priorblocks;
1365 :
1366 691 : if (FileTruncate(v->mdfd_vfd, (pgoff_t) lastsegblocks * BLCKSZ, WAIT_EVENT_DATA_FILE_TRUNCATE) < 0)
1367 0 : ereport(ERROR,
1368 : (errcode_for_file_access(),
1369 : errmsg("could not truncate file \"%s\" to %u blocks: %m",
1370 : FilePathName(v->mdfd_vfd),
1371 : nblocks)));
1372 691 : if (!SmgrIsTemp(reln))
1373 473 : register_dirty_segment(reln, forknum, v);
1374 : }
1375 : else
1376 : {
1377 : /*
1378 : * We still need this segment, so nothing to do for this and any
1379 : * earlier segment.
1380 : */
1381 0 : break;
1382 : }
1383 691 : curopensegs--;
1384 : }
1385 : }
1386 :
1387 : /*
1388 : * mdregistersync() -- Mark whole relation as needing fsync
1389 : */
1390 : void
1391 30362 : mdregistersync(SMgrRelation reln, ForkNumber forknum)
1392 : {
1393 : int segno;
1394 : int min_inactive_seg;
1395 :
1396 : /*
1397 : * NOTE: mdnblocks makes sure we have opened all active segments, so that
1398 : * the loop below will get them all!
1399 : */
1400 30362 : mdnblocks(reln, forknum);
1401 :
1402 30362 : min_inactive_seg = segno = reln->md_num_open_segs[forknum];
1403 :
1404 : /*
1405 : * Temporarily open inactive segments, then close them after sync. There
1406 : * may be some inactive segments left opened after error, but that is
1407 : * harmless. We don't bother to clean them up and take a risk of further
1408 : * trouble. The next mdclose() will soon close them.
1409 : */
1410 30362 : while (_mdfd_openseg(reln, forknum, segno, 0) != NULL)
1411 0 : segno++;
1412 :
1413 60724 : while (segno > 0)
1414 : {
1415 30362 : MdfdVec *v = &reln->md_seg_fds[forknum][segno - 1];
1416 :
1417 30362 : register_dirty_segment(reln, forknum, v);
1418 :
1419 : /* Close inactive segments immediately */
1420 30362 : if (segno > min_inactive_seg)
1421 : {
1422 0 : FileClose(v->mdfd_vfd);
1423 0 : _fdvec_resize(reln, forknum, segno - 1);
1424 : }
1425 :
1426 30362 : segno--;
1427 : }
1428 30362 : }
1429 :
1430 : /*
1431 : * mdimmedsync() -- Immediately sync a relation to stable storage.
1432 : *
1433 : * Note that only writes already issued are synced; this routine knows
1434 : * nothing of dirty buffers that may exist inside the buffer manager. We
1435 : * sync active and inactive segments; smgrDoPendingSyncs() relies on this.
1436 : * Consider a relation skipping WAL. Suppose a checkpoint syncs blocks of
1437 : * some segment, then mdtruncate() renders that segment inactive. If we
1438 : * crash before the next checkpoint syncs the newly-inactive segment, that
1439 : * segment may survive recovery, reintroducing unwanted data into the table.
1440 : */
1441 : void
1442 53 : mdimmedsync(SMgrRelation reln, ForkNumber forknum)
1443 : {
1444 : int segno;
1445 : int min_inactive_seg;
1446 :
1447 : /*
1448 : * NOTE: mdnblocks makes sure we have opened all active segments, so that
1449 : * the loop below will get them all!
1450 : */
1451 53 : mdnblocks(reln, forknum);
1452 :
1453 53 : min_inactive_seg = segno = reln->md_num_open_segs[forknum];
1454 :
1455 : /*
1456 : * Temporarily open inactive segments, then close them after sync. There
1457 : * may be some inactive segments left opened after fsync() error, but that
1458 : * is harmless. We don't bother to clean them up and take a risk of
1459 : * further trouble. The next mdclose() will soon close them.
1460 : */
1461 53 : while (_mdfd_openseg(reln, forknum, segno, 0) != NULL)
1462 0 : segno++;
1463 :
1464 106 : while (segno > 0)
1465 : {
1466 53 : MdfdVec *v = &reln->md_seg_fds[forknum][segno - 1];
1467 :
1468 : /*
1469 : * fsyncs done through mdimmedsync() should be tracked in a separate
1470 : * IOContext than those done through mdsyncfiletag() to differentiate
1471 : * between unavoidable client backend fsyncs (e.g. those done during
1472 : * index build) and those which ideally would have been done by the
1473 : * checkpointer. Since other IO operations bypassing the buffer
1474 : * manager could also be tracked in such an IOContext, wait until
1475 : * these are also tracked to track immediate fsyncs.
1476 : */
1477 53 : if (FileSync(v->mdfd_vfd, WAIT_EVENT_DATA_FILE_IMMEDIATE_SYNC) < 0)
1478 0 : ereport(data_sync_elevel(ERROR),
1479 : (errcode_for_file_access(),
1480 : errmsg("could not fsync file \"%s\": %m",
1481 : FilePathName(v->mdfd_vfd))));
1482 :
1483 : /* Close inactive segments immediately */
1484 53 : if (segno > min_inactive_seg)
1485 : {
1486 0 : FileClose(v->mdfd_vfd);
1487 0 : _fdvec_resize(reln, forknum, segno - 1);
1488 : }
1489 :
1490 53 : segno--;
1491 : }
1492 53 : }
1493 :
1494 : int
1495 502947 : mdfd(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, uint32 *off)
1496 : {
1497 502947 : MdfdVec *v = mdopenfork(reln, forknum, EXTENSION_FAIL);
1498 :
1499 502947 : v = _mdfd_getseg(reln, forknum, blocknum, false,
1500 : EXTENSION_FAIL);
1501 :
1502 502947 : *off = (pgoff_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
1503 :
1504 : Assert(*off < (pgoff_t) BLCKSZ * RELSEG_SIZE);
1505 :
1506 502947 : return FileGetRawDesc(v->mdfd_vfd);
1507 : }
1508 :
1509 : /*
1510 : * register_dirty_segment() -- Mark a relation segment as needing fsync
1511 : *
1512 : * If there is a local pending-ops table, just make an entry in it for
1513 : * ProcessSyncRequests to process later. Otherwise, try to pass off the
1514 : * fsync request to the checkpointer process. If that fails, just do the
1515 : * fsync locally before returning (we hope this will not happen often
1516 : * enough to be a performance problem).
1517 : */
1518 : static void
1519 1063315 : register_dirty_segment(SMgrRelation reln, ForkNumber forknum, MdfdVec *seg)
1520 : {
1521 : FileTag tag;
1522 :
1523 1063315 : INIT_MD_FILETAG(tag, reln->smgr_rlocator.locator, forknum, seg->mdfd_segno);
1524 :
1525 : /* Temp relations should never be fsync'd */
1526 : Assert(!SmgrIsTemp(reln));
1527 :
1528 1063315 : if (!RegisterSyncRequest(&tag, SYNC_REQUEST, false /* retryOnError */ ))
1529 : {
1530 : instr_time io_start;
1531 :
1532 275 : ereport(DEBUG1,
1533 : (errmsg_internal("could not forward fsync request because request queue is full")));
1534 :
1535 275 : io_start = pgstat_prepare_io_time(track_io_timing);
1536 :
1537 275 : if (FileSync(seg->mdfd_vfd, WAIT_EVENT_DATA_FILE_SYNC) < 0)
1538 0 : ereport(data_sync_elevel(ERROR),
1539 : (errcode_for_file_access(),
1540 : errmsg("could not fsync file \"%s\": %m",
1541 : FilePathName(seg->mdfd_vfd))));
1542 :
1543 : /*
1544 : * We have no way of knowing if the current IOContext is
1545 : * IOCONTEXT_NORMAL or IOCONTEXT_[BULKREAD, BULKWRITE, VACUUM] at this
1546 : * point, so count the fsync as being in the IOCONTEXT_NORMAL
1547 : * IOContext. This is probably okay, because the number of backend
1548 : * fsyncs doesn't say anything about the efficacy of the
1549 : * BufferAccessStrategy. And counting both fsyncs done in
1550 : * IOCONTEXT_NORMAL and IOCONTEXT_[BULKREAD, BULKWRITE, VACUUM] under
1551 : * IOCONTEXT_NORMAL is likely clearer when investigating the number of
1552 : * backend fsyncs.
1553 : */
1554 275 : pgstat_count_io_op_time(IOOBJECT_RELATION, IOCONTEXT_NORMAL,
1555 : IOOP_FSYNC, io_start, 1, 0);
1556 : }
1557 1063315 : }
1558 :
1559 : /*
1560 : * register_unlink_segment() -- Schedule a file to be deleted after next checkpoint
1561 : */
1562 : static void
1563 45458 : register_unlink_segment(RelFileLocatorBackend rlocator, ForkNumber forknum,
1564 : BlockNumber segno)
1565 : {
1566 : FileTag tag;
1567 :
1568 45458 : INIT_MD_FILETAG(tag, rlocator.locator, forknum, segno);
1569 :
1570 : /* Should never be used with temp relations */
1571 : Assert(!RelFileLocatorBackendIsTemp(rlocator));
1572 :
1573 45458 : RegisterSyncRequest(&tag, SYNC_UNLINK_REQUEST, true /* retryOnError */ );
1574 45458 : }
1575 :
1576 : /*
1577 : * register_forget_request() -- forget any fsyncs for a relation fork's segment
1578 : */
1579 : static void
1580 169374 : register_forget_request(RelFileLocatorBackend rlocator, ForkNumber forknum,
1581 : BlockNumber segno)
1582 : {
1583 : FileTag tag;
1584 :
1585 169374 : INIT_MD_FILETAG(tag, rlocator.locator, forknum, segno);
1586 :
1587 169374 : RegisterSyncRequest(&tag, SYNC_FORGET_REQUEST, true /* retryOnError */ );
1588 169374 : }
1589 :
1590 : /*
1591 : * ForgetDatabaseSyncRequests -- forget any fsyncs and unlinks for a DB
1592 : */
1593 : void
1594 68 : ForgetDatabaseSyncRequests(Oid dbid)
1595 : {
1596 : FileTag tag;
1597 : RelFileLocator rlocator;
1598 :
1599 68 : rlocator.dbOid = dbid;
1600 68 : rlocator.spcOid = 0;
1601 68 : rlocator.relNumber = 0;
1602 :
1603 68 : INIT_MD_FILETAG(tag, rlocator, InvalidForkNumber, InvalidBlockNumber);
1604 :
1605 68 : RegisterSyncRequest(&tag, SYNC_FILTER_REQUEST, true /* retryOnError */ );
1606 68 : }
1607 :
1608 : /*
1609 : * DropRelationFiles -- drop files of all given relations
1610 : */
1611 : void
1612 2862 : DropRelationFiles(RelFileLocator *delrels, int ndelrels, bool isRedo)
1613 : {
1614 : SMgrRelation *srels;
1615 : int i;
1616 :
1617 2862 : srels = palloc_array(SMgrRelation, ndelrels);
1618 11022 : for (i = 0; i < ndelrels; i++)
1619 : {
1620 8160 : SMgrRelation srel = smgropen(delrels[i], INVALID_PROC_NUMBER);
1621 :
1622 8160 : if (isRedo)
1623 : {
1624 : ForkNumber fork;
1625 :
1626 40610 : for (fork = 0; fork <= MAX_FORKNUM; fork++)
1627 32488 : XLogDropRelation(delrels[i], fork);
1628 : }
1629 8160 : srels[i] = srel;
1630 : }
1631 :
1632 2862 : smgrdounlinkall(srels, ndelrels, isRedo);
1633 :
1634 11022 : for (i = 0; i < ndelrels; i++)
1635 8160 : smgrclose(srels[i]);
1636 2862 : pfree(srels);
1637 2862 : }
1638 :
1639 :
1640 : /*
1641 : * _fdvec_resize() -- Resize the fork's open segments array
1642 : */
1643 : static void
1644 1892060 : _fdvec_resize(SMgrRelation reln,
1645 : ForkNumber forknum,
1646 : int nseg)
1647 : {
1648 1892060 : if (nseg == 0)
1649 : {
1650 653103 : if (reln->md_num_open_segs[forknum] > 0)
1651 : {
1652 653103 : pfree(reln->md_seg_fds[forknum]);
1653 653103 : reln->md_seg_fds[forknum] = NULL;
1654 : }
1655 : }
1656 1238957 : else if (reln->md_num_open_segs[forknum] == 0)
1657 : {
1658 1238957 : reln->md_seg_fds[forknum] =
1659 1238957 : MemoryContextAlloc(MdCxt, sizeof(MdfdVec) * nseg);
1660 : }
1661 0 : else if (nseg > reln->md_num_open_segs[forknum])
1662 : {
1663 : /*
1664 : * It doesn't seem worthwhile complicating the code to amortize
1665 : * repalloc() calls. Those are far faster than PathNameOpenFile() or
1666 : * FileClose(), and the memory context internally will sometimes avoid
1667 : * doing an actual reallocation.
1668 : */
1669 0 : reln->md_seg_fds[forknum] =
1670 0 : repalloc(reln->md_seg_fds[forknum],
1671 : sizeof(MdfdVec) * nseg);
1672 : }
1673 : else
1674 : {
1675 : /*
1676 : * We don't reallocate a smaller array, because we want mdtruncate()
1677 : * to be able to promise that it won't allocate memory, so that it is
1678 : * allowed in a critical section. This means that a bit of space in
1679 : * the array is now wasted, until the next time we add a segment and
1680 : * reallocate.
1681 : */
1682 : }
1683 :
1684 1892060 : reln->md_num_open_segs[forknum] = nseg;
1685 1892060 : }
1686 :
1687 : /*
1688 : * Return the filename for the specified segment of the relation. The
1689 : * returned string is palloc'd.
1690 : */
1691 : static MdPathStr
1692 30427 : _mdfd_segpath(SMgrRelation reln, ForkNumber forknum, BlockNumber segno)
1693 : {
1694 : RelPathStr path;
1695 : MdPathStr fullpath;
1696 :
1697 30427 : path = relpath(reln->smgr_rlocator, forknum);
1698 :
1699 30427 : if (segno > 0)
1700 30427 : sprintf(fullpath.str, "%s.%u", path.str, segno);
1701 : else
1702 0 : strcpy(fullpath.str, path.str);
1703 :
1704 30427 : return fullpath;
1705 : }
1706 :
1707 : /*
1708 : * Open the specified segment of the relation,
1709 : * and make a MdfdVec object for it. Returns NULL on failure.
1710 : */
1711 : static MdfdVec *
1712 30415 : _mdfd_openseg(SMgrRelation reln, ForkNumber forknum, BlockNumber segno,
1713 : int oflags)
1714 : {
1715 : MdfdVec *v;
1716 : File fd;
1717 : MdPathStr fullpath;
1718 :
1719 30415 : fullpath = _mdfd_segpath(reln, forknum, segno);
1720 :
1721 : /* open the file */
1722 30415 : fd = PathNameOpenFile(fullpath.str, _mdfd_open_flags() | oflags);
1723 :
1724 30415 : if (fd < 0)
1725 30415 : return NULL;
1726 :
1727 : /*
1728 : * Segments are always opened in order from lowest to highest, so we must
1729 : * be adding a new one at the end.
1730 : */
1731 : Assert(segno == reln->md_num_open_segs[forknum]);
1732 :
1733 0 : _fdvec_resize(reln, forknum, segno + 1);
1734 :
1735 : /* fill the entry */
1736 0 : v = &reln->md_seg_fds[forknum][segno];
1737 0 : v->mdfd_vfd = fd;
1738 0 : v->mdfd_segno = segno;
1739 :
1740 : Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
1741 :
1742 : /* all done */
1743 0 : return v;
1744 : }
1745 :
1746 : /*
1747 : * _mdfd_getseg() -- Find the segment of the relation holding the
1748 : * specified block.
1749 : *
1750 : * If the segment doesn't exist, we ereport, return NULL, or create the
1751 : * segment, according to "behavior". Note: skipFsync is only used in the
1752 : * EXTENSION_CREATE case.
1753 : */
1754 : static MdfdVec *
1755 2885239 : _mdfd_getseg(SMgrRelation reln, ForkNumber forknum, BlockNumber blkno,
1756 : bool skipFsync, int behavior)
1757 : {
1758 : MdfdVec *v;
1759 : BlockNumber targetseg;
1760 : BlockNumber nextsegno;
1761 :
1762 : /* some way to handle non-existent segments needs to be specified */
1763 : Assert(behavior &
1764 : (EXTENSION_FAIL | EXTENSION_CREATE | EXTENSION_RETURN_NULL |
1765 : EXTENSION_DONT_OPEN));
1766 :
1767 2885239 : targetseg = blkno / ((BlockNumber) RELSEG_SIZE);
1768 :
1769 : /* if an existing and opened segment, we're done */
1770 2885239 : if (targetseg < reln->md_num_open_segs[forknum])
1771 : {
1772 2626442 : v = &reln->md_seg_fds[forknum][targetseg];
1773 2626442 : return v;
1774 : }
1775 :
1776 : /* The caller only wants the segment if we already had it open. */
1777 258797 : if (behavior & EXTENSION_DONT_OPEN)
1778 0 : return NULL;
1779 :
1780 : /*
1781 : * The target segment is not yet open. Iterate over all the segments
1782 : * between the last opened and the target segment. This way missing
1783 : * segments either raise an error, or get created (according to
1784 : * 'behavior'). Start with either the last opened, or the first segment if
1785 : * none was opened before.
1786 : */
1787 258797 : if (reln->md_num_open_segs[forknum] > 0)
1788 12 : v = &reln->md_seg_fds[forknum][reln->md_num_open_segs[forknum] - 1];
1789 : else
1790 : {
1791 258785 : v = mdopenfork(reln, forknum, behavior);
1792 258782 : if (!v)
1793 0 : return NULL; /* if behavior & EXTENSION_RETURN_NULL */
1794 : }
1795 :
1796 258794 : for (nextsegno = reln->md_num_open_segs[forknum];
1797 258794 : nextsegno <= targetseg; nextsegno++)
1798 : {
1799 12 : BlockNumber nblocks = _mdnblocks(reln, forknum, v);
1800 12 : int flags = 0;
1801 :
1802 : Assert(nextsegno == v->mdfd_segno + 1);
1803 :
1804 12 : if (nblocks > ((BlockNumber) RELSEG_SIZE))
1805 0 : elog(FATAL, "segment too big");
1806 :
1807 12 : if ((behavior & EXTENSION_CREATE) ||
1808 12 : (InRecovery && (behavior & EXTENSION_CREATE_RECOVERY)))
1809 : {
1810 : /*
1811 : * Normally we will create new segments only if authorized by the
1812 : * caller (i.e., we are doing mdextend()). But when doing WAL
1813 : * recovery, create segments anyway; this allows cases such as
1814 : * replaying WAL data that has a write into a high-numbered
1815 : * segment of a relation that was later deleted. We want to go
1816 : * ahead and create the segments so we can finish out the replay.
1817 : *
1818 : * We have to maintain the invariant that segments before the last
1819 : * active segment are of size RELSEG_SIZE; therefore, if
1820 : * extending, pad them out with zeroes if needed. (This only
1821 : * matters if in recovery, or if the caller is extending the
1822 : * relation discontiguously, but that can happen in hash indexes.)
1823 : */
1824 0 : if (nblocks < ((BlockNumber) RELSEG_SIZE))
1825 : {
1826 0 : char *zerobuf = palloc_aligned(BLCKSZ, PG_IO_ALIGN_SIZE,
1827 : MCXT_ALLOC_ZERO);
1828 :
1829 0 : mdextend(reln, forknum,
1830 0 : nextsegno * ((BlockNumber) RELSEG_SIZE) - 1,
1831 : zerobuf, skipFsync);
1832 0 : pfree(zerobuf);
1833 : }
1834 0 : flags = O_CREAT;
1835 : }
1836 12 : else if (nblocks < ((BlockNumber) RELSEG_SIZE))
1837 : {
1838 : /*
1839 : * When not extending, only open the next segment if the current
1840 : * one is exactly RELSEG_SIZE. If not (this branch), either
1841 : * return NULL or fail.
1842 : */
1843 12 : if (behavior & EXTENSION_RETURN_NULL)
1844 : {
1845 : /*
1846 : * Some callers discern between reasons for _mdfd_getseg()
1847 : * returning NULL based on errno. As there's no failing
1848 : * syscall involved in this case, explicitly set errno to
1849 : * ENOENT, as that seems the closest interpretation.
1850 : */
1851 0 : errno = ENOENT;
1852 0 : return NULL;
1853 : }
1854 :
1855 12 : ereport(ERROR,
1856 : (errcode_for_file_access(),
1857 : errmsg("could not open file \"%s\" (target block %u): previous segment is only %u blocks",
1858 : _mdfd_segpath(reln, forknum, nextsegno).str,
1859 : blkno, nblocks)));
1860 : }
1861 :
1862 0 : v = _mdfd_openseg(reln, forknum, nextsegno, flags);
1863 :
1864 0 : if (v == NULL)
1865 : {
1866 0 : if ((behavior & EXTENSION_RETURN_NULL) &&
1867 0 : FILE_POSSIBLY_DELETED(errno))
1868 0 : return NULL;
1869 0 : ereport(ERROR,
1870 : (errcode_for_file_access(),
1871 : errmsg("could not open file \"%s\" (target block %u): %m",
1872 : _mdfd_segpath(reln, forknum, nextsegno).str,
1873 : blkno)));
1874 : }
1875 : }
1876 :
1877 258782 : return v;
1878 : }
1879 :
1880 : /*
1881 : * Get number of blocks present in a single disk file
1882 : */
1883 : static BlockNumber
1884 3454251 : _mdnblocks(SMgrRelation reln, ForkNumber forknum, MdfdVec *seg)
1885 : {
1886 : pgoff_t len;
1887 :
1888 3454251 : len = FileSize(seg->mdfd_vfd);
1889 3454251 : if (len < 0)
1890 0 : ereport(ERROR,
1891 : (errcode_for_file_access(),
1892 : errmsg("could not seek to end of file \"%s\": %m",
1893 : FilePathName(seg->mdfd_vfd))));
1894 : /* note that this calculation will ignore any partial block at EOF */
1895 3454251 : return (BlockNumber) (len / BLCKSZ);
1896 : }
1897 :
1898 : /*
1899 : * Sync a file to disk, given a file tag. Write the path into an output
1900 : * buffer so the caller can use it in error messages.
1901 : *
1902 : * Return 0 on success, -1 on failure, with errno set.
1903 : */
1904 : int
1905 0 : mdsyncfiletag(const FileTag *ftag, char *path)
1906 : {
1907 0 : SMgrRelation reln = smgropen(ftag->rlocator, INVALID_PROC_NUMBER);
1908 : File file;
1909 : instr_time io_start;
1910 : bool need_to_close;
1911 : int result,
1912 : save_errno;
1913 :
1914 : /* See if we already have the file open, or need to open it. */
1915 0 : if (ftag->segno < reln->md_num_open_segs[ftag->forknum])
1916 : {
1917 0 : file = reln->md_seg_fds[ftag->forknum][ftag->segno].mdfd_vfd;
1918 0 : strlcpy(path, FilePathName(file), MAXPGPATH);
1919 0 : need_to_close = false;
1920 : }
1921 : else
1922 : {
1923 : MdPathStr p;
1924 :
1925 0 : p = _mdfd_segpath(reln, ftag->forknum, ftag->segno);
1926 0 : strlcpy(path, p.str, MD_PATH_STR_MAXLEN);
1927 :
1928 0 : file = PathNameOpenFile(path, _mdfd_open_flags());
1929 0 : if (file < 0)
1930 0 : return -1;
1931 0 : need_to_close = true;
1932 : }
1933 :
1934 0 : io_start = pgstat_prepare_io_time(track_io_timing);
1935 :
1936 : /* Sync the file. */
1937 0 : result = FileSync(file, WAIT_EVENT_DATA_FILE_SYNC);
1938 0 : save_errno = errno;
1939 :
1940 0 : if (need_to_close)
1941 0 : FileClose(file);
1942 :
1943 0 : pgstat_count_io_op_time(IOOBJECT_RELATION, IOCONTEXT_NORMAL,
1944 : IOOP_FSYNC, io_start, 1, 0);
1945 :
1946 0 : errno = save_errno;
1947 0 : return result;
1948 : }
1949 :
1950 : /*
1951 : * Unlink a file, given a file tag. Write the path into an output
1952 : * buffer so the caller can use it in error messages.
1953 : *
1954 : * Return 0 on success, -1 on failure, with errno set.
1955 : */
1956 : int
1957 35312 : mdunlinkfiletag(const FileTag *ftag, char *path)
1958 : {
1959 : RelPathStr p;
1960 :
1961 : /* Compute the path. */
1962 35312 : p = relpathperm(ftag->rlocator, MAIN_FORKNUM);
1963 35312 : strlcpy(path, p.str, MAXPGPATH);
1964 :
1965 : /* Try to unlink the file. */
1966 35312 : return unlink(path);
1967 : }
1968 :
1969 : /*
1970 : * Check if a given candidate request matches a given tag, when processing
1971 : * a SYNC_FILTER_REQUEST request. This will be called for all pending
1972 : * requests to find out whether to forget them.
1973 : */
1974 : bool
1975 8221 : mdfiletagmatches(const FileTag *ftag, const FileTag *candidate)
1976 : {
1977 : /*
1978 : * For now we only use filter requests as a way to drop all scheduled
1979 : * callbacks relating to a given database, when dropping the database.
1980 : * We'll return true for all candidates that have the same database OID as
1981 : * the ftag from the SYNC_FILTER_REQUEST request, so they're forgotten.
1982 : */
1983 8221 : return ftag->rlocator.dbOid == candidate->rlocator.dbOid;
1984 : }
1985 :
1986 : /*
1987 : * AIO completion callback for mdstartreadv().
1988 : */
1989 : static PgAioResult
1990 1236886 : md_readv_complete(PgAioHandle *ioh, PgAioResult prior_result, uint8 cb_data)
1991 : {
1992 1236886 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
1993 1236886 : PgAioResult result = prior_result;
1994 :
1995 1236886 : if (prior_result.result < 0)
1996 : {
1997 7 : result.status = PGAIO_RS_ERROR;
1998 7 : result.id = PGAIO_HCB_MD_READV;
1999 : /* For "hard" errors, track the error number in error_data */
2000 7 : result.error_data = -prior_result.result;
2001 7 : result.result = 0;
2002 :
2003 : /*
2004 : * Immediately log a message about the IO error, but only to the
2005 : * server log. The reason to do so immediately is that the originator
2006 : * might not process the query result immediately (because it is busy
2007 : * doing another part of query processing) or at all (e.g. if it was
2008 : * cancelled or errored out due to another IO also failing). The
2009 : * definer of the IO will emit an ERROR when processing the IO's
2010 : * results
2011 : */
2012 7 : pgaio_result_report(result, td, LOG_SERVER_ONLY);
2013 :
2014 7 : return result;
2015 : }
2016 :
2017 : /*
2018 : * As explained above smgrstartreadv(), the smgr API operates on the level
2019 : * of blocks, rather than bytes. Convert.
2020 : */
2021 1236879 : result.result /= BLCKSZ;
2022 :
2023 : Assert(result.result <= td->smgr.nblocks);
2024 :
2025 1236879 : if (result.result == 0)
2026 : {
2027 : /* consider 0 blocks read a failure */
2028 2 : result.status = PGAIO_RS_ERROR;
2029 2 : result.id = PGAIO_HCB_MD_READV;
2030 2 : result.error_data = 0;
2031 :
2032 : /* see comment above the "hard error" case */
2033 2 : pgaio_result_report(result, td, LOG_SERVER_ONLY);
2034 :
2035 2 : return result;
2036 : }
2037 :
2038 1236877 : if (result.status != PGAIO_RS_ERROR &&
2039 1236877 : result.result < td->smgr.nblocks)
2040 : {
2041 : /* partial reads should be retried at upper level */
2042 12 : result.status = PGAIO_RS_PARTIAL;
2043 12 : result.id = PGAIO_HCB_MD_READV;
2044 : }
2045 :
2046 1236877 : return result;
2047 : }
2048 :
2049 : /*
2050 : * AIO error reporting callback for mdstartreadv().
2051 : *
2052 : * Errors are encoded as follows:
2053 : * - PgAioResult.error_data != 0 encodes IO that failed with that errno
2054 : * - PgAioResult.error_data == 0 encodes IO that didn't read all data
2055 : */
2056 : static void
2057 28 : md_readv_report(PgAioResult result, const PgAioTargetData *td, int elevel)
2058 : {
2059 : RelPathStr path;
2060 :
2061 28 : path = relpathbackend(td->smgr.rlocator,
2062 : td->smgr.is_temp ? MyProcNumber : INVALID_PROC_NUMBER,
2063 : td->smgr.forkNum);
2064 :
2065 28 : if (result.error_data != 0)
2066 : {
2067 : /* for errcode_for_file_access() and %m */
2068 14 : errno = result.error_data;
2069 :
2070 14 : ereport(elevel,
2071 : errcode_for_file_access(),
2072 : errmsg("could not read blocks %u..%u in file \"%s\": %m",
2073 : td->smgr.blockNum,
2074 : td->smgr.blockNum + td->smgr.nblocks - 1,
2075 : path.str));
2076 : }
2077 : else
2078 : {
2079 : /*
2080 : * NB: This will typically only be output in debug messages, while
2081 : * retrying a partial IO.
2082 : */
2083 14 : ereport(elevel,
2084 : errcode(ERRCODE_DATA_CORRUPTED),
2085 : errmsg("could not read blocks %u..%u in file \"%s\": read only %zu of %zu bytes",
2086 : td->smgr.blockNum,
2087 : td->smgr.blockNum + td->smgr.nblocks - 1,
2088 : path.str,
2089 : result.result * (size_t) BLCKSZ,
2090 : td->smgr.nblocks * (size_t) BLCKSZ));
2091 : }
2092 19 : }
|
