Branch data Line data Source code
1 : : /*-------------------------------------------------------------------------
2 : : *
3 : : * md.c
4 : : * This code manages relations that reside on magnetic disk.
5 : : *
6 : : * Or at least, that was what the Berkeley folk had in mind when they named
7 : : * this file. In reality, what this code provides is an interface from
8 : : * the smgr API to Unix-like filesystem APIs, so it will work with any type
9 : : * of device for which the operating system provides filesystem support.
10 : : * It doesn't matter whether the bits are on spinning rust or some other
11 : : * storage technology.
12 : : *
13 : : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
14 : : * Portions Copyright (c) 1994, Regents of the University of California
15 : : *
16 : : *
17 : : * IDENTIFICATION
18 : : * src/backend/storage/smgr/md.c
19 : : *
20 : : *-------------------------------------------------------------------------
21 : : */
22 : : #include "postgres.h"
23 : :
24 : : #include <limits.h>
25 : : #include <unistd.h>
26 : : #include <fcntl.h>
27 : : #include <sys/file.h>
28 : :
29 : : #include "access/xlogutils.h"
30 : : #include "commands/tablespace.h"
31 : : #include "common/file_utils.h"
32 : : #include "miscadmin.h"
33 : : #include "pg_trace.h"
34 : : #include "pgstat.h"
35 : : #include "storage/aio.h"
36 : : #include "storage/bufmgr.h"
37 : : #include "storage/fd.h"
38 : : #include "storage/md.h"
39 : : #include "storage/relfilelocator.h"
40 : : #include "storage/smgr.h"
41 : : #include "storage/sync.h"
42 : : #include "utils/memutils.h"
43 : :
44 : : /*
45 : : * The magnetic disk storage manager keeps track of open file
46 : : * descriptors in its own descriptor pool. This is done to make it
47 : : * easier to support relations that are larger than the operating
48 : : * system's file size limit (often 2GBytes). In order to do that,
49 : : * we break relations up into "segment" files that are each shorter than
50 : : * the OS file size limit. The segment size is set by the RELSEG_SIZE
51 : : * configuration constant in pg_config.h.
52 : : *
53 : : * On disk, a relation must consist of consecutively numbered segment
54 : : * files in the pattern
55 : : * -- Zero or more full segments of exactly RELSEG_SIZE blocks each
56 : : * -- Exactly one partial segment of size 0 <= size < RELSEG_SIZE blocks
57 : : * -- Optionally, any number of inactive segments of size 0 blocks.
58 : : * The full and partial segments are collectively the "active" segments.
59 : : * Inactive segments are those that once contained data but are currently
60 : : * not needed because of an mdtruncate() operation. The reason for leaving
61 : : * them present at size zero, rather than unlinking them, is that other
62 : : * backends and/or the checkpointer might be holding open file references to
63 : : * such segments. If the relation expands again after mdtruncate(), such
64 : : * that a deactivated segment becomes active again, it is important that
65 : : * such file references still be valid --- else data might get written
66 : : * out to an unlinked old copy of a segment file that will eventually
67 : : * disappear.
68 : : *
69 : : * RELSEG_SIZE must fit into BlockNumber; but since we expose its value
70 : : * as an integer GUC, it actually needs to fit in signed int. It's worth
71 : : * having a cross-check for this since configure's --with-segsize options
72 : : * could let people select insane values.
73 : : */
74 : : StaticAssertDecl(RELSEG_SIZE > 0 && RELSEG_SIZE <= INT_MAX,
75 : : "RELSEG_SIZE must fit in an integer");
76 : :
77 : : /*
78 : : * File descriptors are stored in the per-fork md_seg_fds arrays inside
79 : : * SMgrRelation. The length of these arrays is stored in md_num_open_segs.
80 : : * Note that a fork's md_num_open_segs having a specific value does not
81 : : * necessarily mean the relation doesn't have additional segments; we may
82 : : * just not have opened the next segment yet. (We could not have "all
83 : : * segments are in the array" as an invariant anyway, since another backend
84 : : * could extend the relation while we aren't looking.) We do not have
85 : : * entries for inactive segments, however; as soon as we find a partial
86 : : * segment, we assume that any subsequent segments are inactive.
87 : : *
88 : : * The entire MdfdVec array is palloc'd in the MdCxt memory context.
89 : : */
90 : :
91 : : typedef struct _MdfdVec
92 : : {
93 : : File mdfd_vfd; /* fd number in fd.c's pool */
94 : : BlockNumber mdfd_segno; /* segment number, from 0 */
95 : : } MdfdVec;
96 : :
97 : : static MemoryContext MdCxt; /* context for all MdfdVec objects */
98 : :
99 : :
100 : : /* Populate a file tag describing an md.c segment file. */
101 : : #define INIT_MD_FILETAG(a,xx_rlocator,xx_forknum,xx_segno) \
102 : : ( \
103 : : memset(&(a), 0, sizeof(FileTag)), \
104 : : (a).handler = SYNC_HANDLER_MD, \
105 : : (a).rlocator = (xx_rlocator), \
106 : : (a).forknum = (xx_forknum), \
107 : : (a).segno = (xx_segno) \
108 : : )
109 : :
110 : :
111 : : /*** behavior for mdopen & _mdfd_getseg ***/
112 : : /* ereport if segment not present */
113 : : #define EXTENSION_FAIL (1 << 0)
114 : : /* return NULL if segment not present */
115 : : #define EXTENSION_RETURN_NULL (1 << 1)
116 : : /* create new segments as needed */
117 : : #define EXTENSION_CREATE (1 << 2)
118 : : /* create new segments if needed during recovery */
119 : : #define EXTENSION_CREATE_RECOVERY (1 << 3)
120 : : /* don't try to open a segment, if not already open */
121 : : #define EXTENSION_DONT_OPEN (1 << 5)
122 : :
123 : :
124 : : /*
125 : : * Fixed-length string to represent paths to files that need to be built by
126 : : * md.c.
127 : : *
128 : : * The maximum number of segments is MaxBlockNumber / RELSEG_SIZE, where
129 : : * RELSEG_SIZE can be set to 1 (for testing only).
130 : : */
131 : : #define SEGMENT_CHARS OIDCHARS
132 : : #define MD_PATH_STR_MAXLEN \
133 : : (\
134 : : REL_PATH_STR_MAXLEN \
135 : : + sizeof((char)'.') \
136 : : + SEGMENT_CHARS \
137 : : )
138 : : typedef struct MdPathStr
139 : : {
140 : : char str[MD_PATH_STR_MAXLEN + 1];
141 : : } MdPathStr;
142 : :
143 : :
144 : : /* local routines */
145 : : static void mdunlinkfork(RelFileLocatorBackend rlocator, ForkNumber forknum,
146 : : bool isRedo);
147 : : static MdfdVec *mdopenfork(SMgrRelation reln, ForkNumber forknum, int behavior);
148 : : static void register_dirty_segment(SMgrRelation reln, ForkNumber forknum,
149 : : MdfdVec *seg);
150 : : static void register_unlink_segment(RelFileLocatorBackend rlocator, ForkNumber forknum,
151 : : BlockNumber segno);
152 : : static void register_forget_request(RelFileLocatorBackend rlocator, ForkNumber forknum,
153 : : BlockNumber segno);
154 : : static void _fdvec_resize(SMgrRelation reln,
155 : : ForkNumber forknum,
156 : : int nseg);
157 : : static MdPathStr _mdfd_segpath(SMgrRelation reln, ForkNumber forknum,
158 : : BlockNumber segno);
159 : : static MdfdVec *_mdfd_openseg(SMgrRelation reln, ForkNumber forknum,
160 : : BlockNumber segno, int oflags);
161 : : static MdfdVec *_mdfd_getseg(SMgrRelation reln, ForkNumber forknum,
162 : : BlockNumber blkno, bool skipFsync, int behavior);
163 : : static BlockNumber _mdnblocks(SMgrRelation reln, ForkNumber forknum,
164 : : MdfdVec *seg);
165 : :
166 : : static PgAioResult md_readv_complete(PgAioHandle *ioh, PgAioResult prior_result, uint8 cb_data);
167 : : static void md_readv_report(PgAioResult result, const PgAioTargetData *td, int elevel);
168 : :
169 : : const PgAioHandleCallbacks aio_md_readv_cb = {
170 : : .complete_shared = md_readv_complete,
171 : : .report = md_readv_report,
172 : : };
173 : :
174 : :
175 : : static inline int
176 : 99981 : _mdfd_open_flags(void)
177 : : {
178 : 99981 : int flags = O_RDWR | PG_BINARY;
179 : :
180 [ + - ]: 99981 : if (io_direct_flags & IO_DIRECT_DATA)
181 : 0 : flags |= PG_O_DIRECT;
182 : :
183 : 199962 : return flags;
184 : 99981 : }
185 : :
186 : : /*
187 : : * mdinit() -- Initialize private state for magnetic disk storage manager.
188 : : */
189 : : void
190 : 806 : mdinit(void)
191 : : {
192 : 806 : MdCxt = AllocSetContextCreate(TopMemoryContext,
193 : : "MdSmgr",
194 : : ALLOCSET_DEFAULT_SIZES);
195 : 806 : }
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 : 58572 : 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 [ - + ]: 58572 : if (!InRecovery)
211 : 58572 : mdclose(reln, forknum);
212 : :
213 : 58572 : 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 : 12496 : mdcreate(SMgrRelation reln, ForkNumber forknum, bool isRedo)
223 : : {
224 : 12496 : MdfdVec *mdfd;
225 : 12496 : RelPathStr path;
226 : 12496 : File fd;
227 : :
228 [ - + # # ]: 12496 : if (isRedo && reln->md_num_open_segs[forknum] > 0)
229 : 0 : return; /* created and opened already... */
230 : :
231 [ + - ]: 12496 : 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 : 24992 : TablespaceCreateDbspace(reln->smgr_rlocator.locator.spcOid,
243 : 12496 : reln->smgr_rlocator.locator.dbOid,
244 : 12496 : isRedo);
245 : :
246 : 12496 : path = relpath(reln->smgr_rlocator, forknum);
247 : :
248 : 12496 : fd = PathNameOpenFile(path.str, _mdfd_open_flags() | O_CREAT | O_EXCL);
249 : :
250 [ + - ]: 12496 : if (fd < 0)
251 : : {
252 : 0 : int save_errno = errno;
253 : :
254 [ # # ]: 0 : if (isRedo)
255 : 0 : fd = PathNameOpenFile(path.str, _mdfd_open_flags());
256 [ # # ]: 0 : 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 : 0 : }
264 : 0 : }
265 : :
266 : 12496 : _fdvec_resize(reln, forknum, 1);
267 : 12496 : mdfd = &reln->md_seg_fds[forknum][0];
268 : 12496 : mdfd->mdfd_vfd = fd;
269 : 12496 : mdfd->mdfd_segno = 0;
270 : :
271 [ + + ]: 12496 : if (!SmgrIsTemp(reln))
272 : 11440 : register_dirty_segment(reln, forknum, mdfd);
273 [ - + ]: 12496 : }
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 : 35696 : mdunlink(RelFileLocatorBackend rlocator, ForkNumber forknum, bool isRedo)
338 : : {
339 : : /* Now do the per-fork work */
340 [ - + ]: 35696 : if (forknum == InvalidForkNumber)
341 : : {
342 [ # # ]: 0 : for (forknum = 0; forknum <= MAX_FORKNUM; forknum++)
343 : 0 : mdunlinkfork(rlocator, forknum, isRedo);
344 : 0 : }
345 : : else
346 : 35696 : mdunlinkfork(rlocator, forknum, isRedo);
347 : 35696 : }
348 : :
349 : : /*
350 : : * Truncate a file to release disk space.
351 : : */
352 : : static int
353 : 39870 : do_truncate(const char *path)
354 : : {
355 : 39870 : int save_errno;
356 : 39870 : int ret;
357 : :
358 : 39870 : ret = pg_truncate(path, 0);
359 : :
360 : : /* Log a warning here to avoid repetition in callers. */
361 [ + + + - ]: 39870 : 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 : 0 : }
369 : :
370 : 79740 : return ret;
371 : 39870 : }
372 : :
373 : : static void
374 : 35696 : mdunlinkfork(RelFileLocatorBackend rlocator, ForkNumber forknum, bool isRedo)
375 : : {
376 : 35696 : RelPathStr path;
377 : 35696 : int ret;
378 : 35696 : int save_errno;
379 : :
380 : 35696 : 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 [ + - + - : 35696 : if (isRedo || IsBinaryUpgrade || forknum != MAIN_FORKNUM ||
+ + + + ]
387 : 8924 : RelFileLocatorBackendIsTemp(rlocator))
388 : : {
389 [ + + ]: 27784 : if (!RelFileLocatorBackendIsTemp(rlocator))
390 : : {
391 : : /* Prevent other backends' fds from holding on to the disk space */
392 : 23736 : ret = do_truncate(path.str);
393 : :
394 : : /* Forget any pending sync requests for the first segment */
395 : 23736 : save_errno = errno;
396 : 23736 : register_forget_request(rlocator, forknum, 0 /* first seg */ );
397 : 23736 : errno = save_errno;
398 : 23736 : }
399 : : else
400 : 4048 : ret = 0;
401 : :
402 : : /* Next unlink the file, unless it was already found to be missing */
403 [ + + - + ]: 27784 : if (ret >= 0 || errno != ENOENT)
404 : : {
405 : 4359 : ret = unlink(path.str);
406 [ + + + - ]: 4359 : 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 : 0 : }
414 : 4359 : }
415 : 27784 : }
416 : : else
417 : : {
418 : : /* Prevent other backends' fds from holding on to the disk space */
419 : 7912 : ret = do_truncate(path.str);
420 : :
421 : : /* Register request to unlink first segment later */
422 : 7912 : save_errno = errno;
423 : 7912 : register_unlink_segment(rlocator, forknum, 0 /* first seg */ );
424 : 7912 : 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 [ + + - + ]: 35696 : if (ret >= 0 || errno != ENOENT)
440 : : {
441 : 9278 : MdPathStr segpath;
442 : 9278 : BlockNumber segno;
443 : :
444 : 9278 : for (segno = 1;; segno++)
445 : : {
446 : 9278 : sprintf(segpath.str, "%s.%u", path.str, segno);
447 : :
448 [ + + ]: 9278 : 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 [ + - + - ]: 8222 : if (do_truncate(segpath.str) < 0 && errno == ENOENT)
455 : 8222 : 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 : 0 : }
463 : :
464 [ - + ]: 1056 : if (unlink(segpath.str) < 0)
465 : : {
466 : : /* ENOENT is expected after the last segment... */
467 [ + - ]: 1056 : if (errno != ENOENT)
468 [ # # # # ]: 0 : ereport(WARNING,
469 : : (errcode_for_file_access(),
470 : : errmsg("could not remove file \"%s\": %m", segpath.str)));
471 : 1056 : break;
472 : : }
473 : 0 : }
474 : 9278 : }
475 : 35696 : }
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 : 10470 : mdextend(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
488 : : const void *buffer, bool skipFsync)
489 : : {
490 : 10470 : pgoff_t seekpos;
491 : 10470 : int nbytes;
492 : 10470 : 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 [ + - ]: 10470 : 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 [ + - ]: 10470 : 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 : 10470 : v = _mdfd_getseg(reln, forknum, blocknum, skipFsync, EXTENSION_CREATE);
517 : :
518 : 10470 : seekpos = (pgoff_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
519 : :
520 [ + - ]: 10470 : Assert(seekpos < (pgoff_t) BLCKSZ * RELSEG_SIZE);
521 : :
522 [ + - ]: 10470 : 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 : 0 : }
538 : :
539 [ + + - + ]: 10470 : if (!skipFsync && !SmgrIsTemp(reln))
540 : 9 : register_dirty_segment(reln, forknum, v);
541 : :
542 [ + - ]: 10470 : Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
543 : 10470 : }
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 : 27454 : mdzeroextend(SMgrRelation reln, ForkNumber forknum,
553 : : BlockNumber blocknum, int nblocks, bool skipFsync)
554 : : {
555 : 27454 : MdfdVec *v;
556 : 27454 : BlockNumber curblocknum = blocknum;
557 : 27454 : int remblocks = nblocks;
558 : :
559 [ + - ]: 27454 : 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 [ + - ]: 27454 : 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 [ + + ]: 54908 : while (remblocks > 0)
579 : : {
580 : 27454 : BlockNumber segstartblock = curblocknum % ((BlockNumber) RELSEG_SIZE);
581 : 27454 : pgoff_t seekpos = (pgoff_t) BLCKSZ * segstartblock;
582 : 27454 : int numblocks;
583 : :
584 [ - + ]: 27454 : if (segstartblock + remblocks > RELSEG_SIZE)
585 : 0 : numblocks = RELSEG_SIZE - segstartblock;
586 : : else
587 : 27454 : numblocks = remblocks;
588 : :
589 : 27454 : v = _mdfd_getseg(reln, forknum, curblocknum, skipFsync, EXTENSION_CREATE);
590 : :
591 [ - + ]: 27454 : Assert(segstartblock < RELSEG_SIZE);
592 [ - + ]: 27454 : 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 [ + + ]: 27454 : if (numblocks > 8)
606 : : {
607 : 131 : int ret;
608 : :
609 : 262 : ret = FileFallocate(v->mdfd_vfd,
610 : 131 : seekpos, (pgoff_t) BLCKSZ * numblocks,
611 : : WAIT_EVENT_DATA_FILE_EXTEND);
612 [ + - ]: 131 : 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 : 0 : }
620 : 131 : }
621 : : else
622 : : {
623 : 27323 : 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 : 54646 : ret = FileZero(v->mdfd_vfd,
633 : 27323 : seekpos, (pgoff_t) BLCKSZ * numblocks,
634 : : WAIT_EVENT_DATA_FILE_EXTEND);
635 [ + - ]: 27323 : 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 : 27323 : }
642 : :
643 [ + - + + ]: 27454 : if (!skipFsync && !SmgrIsTemp(reln))
644 : 24322 : register_dirty_segment(reln, forknum, v);
645 : :
646 [ - + ]: 27454 : Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
647 : :
648 : 27454 : remblocks -= numblocks;
649 : 27454 : curblocknum += numblocks;
650 : 27454 : }
651 : 27454 : }
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 : 598324 : mdopenfork(SMgrRelation reln, ForkNumber forknum, int behavior)
665 : : {
666 : 598324 : MdfdVec *mdfd;
667 : 598324 : RelPathStr path;
668 : 598324 : File fd;
669 : :
670 : : /* No work if already open */
671 [ + + ]: 598324 : if (reln->md_num_open_segs[forknum] > 0)
672 : 514345 : return &reln->md_seg_fds[forknum][0];
673 : :
674 : 83979 : path = relpath(reln->smgr_rlocator, forknum);
675 : :
676 : 83979 : fd = PathNameOpenFile(path.str, _mdfd_open_flags());
677 : :
678 [ + + ]: 83979 : if (fd < 0)
679 : : {
680 [ + - ]: 42362 : if ((behavior & EXTENSION_RETURN_NULL) &&
681 : 42362 : FILE_POSSIBLY_DELETED(errno))
682 : 42362 : return NULL;
683 [ # # # # ]: 0 : ereport(ERROR,
684 : : (errcode_for_file_access(),
685 : : errmsg("could not open file \"%s\": %m", path.str)));
686 : 0 : }
687 : :
688 : 41617 : _fdvec_resize(reln, forknum, 1);
689 : 41617 : mdfd = &reln->md_seg_fds[forknum][0];
690 : 41617 : mdfd->mdfd_vfd = fd;
691 : 41617 : mdfd->mdfd_segno = 0;
692 : :
693 [ + - ]: 41617 : Assert(_mdnblocks(reln, forknum, mdfd) <= ((BlockNumber) RELSEG_SIZE));
694 : :
695 : 41617 : return mdfd;
696 : 598324 : }
697 : :
698 : : /*
699 : : * mdopen() -- Initialize newly-opened relation.
700 : : */
701 : : void
702 : 65565 : mdopen(SMgrRelation reln)
703 : : {
704 : : /* mark it not open */
705 [ + + ]: 327825 : for (int forknum = 0; forknum <= MAX_FORKNUM; forknum++)
706 : 262260 : reln->md_num_open_segs[forknum] = 0;
707 : 65565 : }
708 : :
709 : : /*
710 : : * mdclose() -- Close the specified relation, if it isn't closed already.
711 : : */
712 : : void
713 : 447416 : mdclose(SMgrRelation reln, ForkNumber forknum)
714 : : {
715 : 447416 : int nopensegs = reln->md_num_open_segs[forknum];
716 : :
717 : : /* No work if already closed */
718 [ + + ]: 447416 : if (nopensegs == 0)
719 : 405401 : return;
720 : :
721 : : /* close segments starting from the end */
722 [ + + ]: 84030 : while (nopensegs > 0)
723 : : {
724 : 42015 : MdfdVec *v = &reln->md_seg_fds[forknum][nopensegs - 1];
725 : :
726 : 42015 : FileClose(v->mdfd_vfd);
727 : 42015 : _fdvec_resize(reln, forknum, nopensegs - 1);
728 : 42015 : nopensegs--;
729 : 42015 : }
730 [ - + ]: 447416 : }
731 : :
732 : : /*
733 : : * mdprefetch() -- Initiate asynchronous read of the specified blocks of a relation
734 : : */
735 : : bool
736 : 0 : mdprefetch(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
737 : : int nblocks)
738 : : {
739 : : #ifdef USE_PREFETCH
740 : :
741 [ # # ]: 0 : Assert((io_direct_flags & IO_DIRECT_DATA) == 0);
742 : :
743 [ # # ]: 0 : if ((uint64) blocknum + nblocks > (uint64) MaxBlockNumber + 1)
744 : 0 : return false;
745 : :
746 [ # # ]: 0 : while (nblocks > 0)
747 : : {
748 : 0 : pgoff_t seekpos;
749 : 0 : MdfdVec *v;
750 : 0 : int nblocks_this_segment;
751 : :
752 : 0 : v = _mdfd_getseg(reln, forknum, blocknum, false,
753 : 0 : InRecovery ? EXTENSION_RETURN_NULL : EXTENSION_FAIL);
754 [ # # ]: 0 : if (v == NULL)
755 : 0 : return false;
756 : :
757 : 0 : seekpos = (pgoff_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
758 : :
759 [ # # ]: 0 : Assert(seekpos < (pgoff_t) BLCKSZ * RELSEG_SIZE);
760 : :
761 : 0 : nblocks_this_segment =
762 [ # # ]: 0 : Min(nblocks,
763 : : RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
764 : :
765 : 0 : (void) FilePrefetch(v->mdfd_vfd, seekpos, BLCKSZ * nblocks_this_segment,
766 : : WAIT_EVENT_DATA_FILE_PREFETCH);
767 : :
768 : 0 : blocknum += nblocks_this_segment;
769 : 0 : nblocks -= nblocks_this_segment;
770 [ # # # ]: 0 : }
771 : : #endif /* USE_PREFETCH */
772 : :
773 : 0 : return true;
774 : 0 : }
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 : 15230 : buffers_to_iovec(struct iovec *iov, void **buffers, int nblocks)
785 : : {
786 : 15230 : struct iovec *iovp;
787 : 15230 : int iovcnt;
788 : :
789 [ + - ]: 15230 : Assert(nblocks >= 1);
790 : :
791 : : /* If this build supports direct I/O, buffers must be I/O aligned. */
792 [ + + ]: 34558 : for (int i = 0; i < nblocks; ++i)
793 : : {
794 : : if (PG_O_DIRECT != 0 && PG_IO_ALIGN_SIZE <= BLCKSZ)
795 [ + - ]: 19328 : Assert((uintptr_t) buffers[i] ==
796 : : TYPEALIGN(PG_IO_ALIGN_SIZE, buffers[i]));
797 : 19328 : }
798 : :
799 : : /* Start the first iovec off with the first buffer. */
800 : 15230 : iovp = &iov[0];
801 : 15230 : iovp->iov_base = buffers[0];
802 : 15230 : iovp->iov_len = BLCKSZ;
803 : 15230 : iovcnt = 1;
804 : :
805 : : /* Try to merge the rest. */
806 [ + + ]: 19328 : for (int i = 1; i < nblocks; ++i)
807 : : {
808 : 4098 : void *buffer = buffers[i];
809 : :
810 [ + + ]: 4098 : if (((char *) iovp->iov_base + iovp->iov_len) == buffer)
811 : : {
812 : : /* Contiguous with the last iovec. */
813 : 3890 : iovp->iov_len += BLCKSZ;
814 : 3890 : }
815 : : else
816 : : {
817 : : /* Need a new iovec. */
818 : 208 : iovp++;
819 : 208 : iovp->iov_base = buffer;
820 : 208 : iovp->iov_len = BLCKSZ;
821 : 208 : iovcnt++;
822 : : }
823 : 4098 : }
824 : :
825 : 30460 : return iovcnt;
826 : 15230 : }
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 : 639 : mdmaxcombine(SMgrRelation reln, ForkNumber forknum,
834 : : BlockNumber blocknum)
835 : : {
836 : 639 : BlockNumber segoff;
837 : :
838 : 639 : segoff = blocknum % ((BlockNumber) RELSEG_SIZE);
839 : :
840 : 1278 : return RELSEG_SIZE - segoff;
841 : 639 : }
842 : :
843 : : /*
844 : : * mdreadv() -- Read the specified blocks from a relation.
845 : : */
846 : : void
847 : 177 : mdreadv(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
848 : : void **buffers, BlockNumber nblocks)
849 : : {
850 [ + + ]: 354 : while (nblocks > 0)
851 : : {
852 : 177 : struct iovec iov[PG_IOV_MAX];
853 : 177 : int iovcnt;
854 : 177 : pgoff_t seekpos;
855 : 177 : int nbytes;
856 : 177 : MdfdVec *v;
857 : 177 : BlockNumber nblocks_this_segment;
858 : 177 : size_t transferred_this_segment;
859 : 177 : size_t size_this_segment;
860 : :
861 : 177 : v = _mdfd_getseg(reln, forknum, blocknum, false,
862 : : EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
863 : :
864 : 177 : seekpos = (pgoff_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
865 : :
866 [ + - ]: 177 : Assert(seekpos < (pgoff_t) BLCKSZ * RELSEG_SIZE);
867 : :
868 : 177 : nblocks_this_segment =
869 [ + - ]: 177 : Min(nblocks,
870 : : RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
871 [ + - ]: 177 : nblocks_this_segment = Min(nblocks_this_segment, lengthof(iov));
872 : :
873 [ + - ]: 177 : if (nblocks_this_segment != nblocks)
874 [ # # # # ]: 0 : elog(ERROR, "read crosses segment boundary");
875 : :
876 : 177 : iovcnt = buffers_to_iovec(iov, buffers, nblocks_this_segment);
877 : 177 : size_this_segment = nblocks_this_segment * BLCKSZ;
878 : 177 : 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 : 177 : for (;;)
886 : : {
887 : 177 : 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 : 177 : nbytes = FileReadV(v->mdfd_vfd, iov, iovcnt, seekpos,
893 : : WAIT_EVENT_DATA_FILE_READ);
894 : 177 : 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 [ + - ]: 177 : 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 [ + - ]: 177 : 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 : 0 : 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 : 0 : }
963 : :
964 : : /* One loop should usually be enough. */
965 : 177 : transferred_this_segment += nbytes;
966 [ + - ]: 177 : Assert(transferred_this_segment <= size_this_segment);
967 [ - + ]: 177 : if (transferred_this_segment == size_this_segment)
968 : 177 : 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 : 177 : nblocks -= nblocks_this_segment;
976 : 177 : buffers += nblocks_this_segment;
977 : 177 : blocknum += nblocks_this_segment;
978 : 177 : }
979 : 177 : }
980 : :
981 : : /*
982 : : * mdstartreadv() -- Asynchronous version of mdreadv().
983 : : */
984 : : void
985 : 7052 : mdstartreadv(PgAioHandle *ioh,
986 : : SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
987 : : void **buffers, BlockNumber nblocks)
988 : : {
989 : 7052 : pgoff_t seekpos;
990 : 7052 : MdfdVec *v;
991 : 7052 : BlockNumber nblocks_this_segment;
992 : 7052 : struct iovec *iov;
993 : 7052 : int iovcnt;
994 : 7052 : int ret;
995 : :
996 : 7052 : v = _mdfd_getseg(reln, forknum, blocknum, false,
997 : : EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
998 : :
999 : 7052 : seekpos = (pgoff_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
1000 : :
1001 [ + - ]: 7052 : Assert(seekpos < (pgoff_t) BLCKSZ * RELSEG_SIZE);
1002 : :
1003 : 7052 : nblocks_this_segment =
1004 [ + - ]: 7052 : Min(nblocks,
1005 : : RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
1006 : :
1007 [ + - ]: 7052 : if (nblocks_this_segment != nblocks)
1008 [ # # # # ]: 0 : elog(ERROR, "read crossing segment boundary");
1009 : :
1010 : 7052 : iovcnt = pgaio_io_get_iovec(ioh, &iov);
1011 : :
1012 [ + - ]: 7052 : Assert(nblocks <= iovcnt);
1013 : :
1014 : 7052 : iovcnt = buffers_to_iovec(iov, buffers, nblocks_this_segment);
1015 : :
1016 [ + - ]: 7052 : Assert(iovcnt <= nblocks_this_segment);
1017 : :
1018 [ - + ]: 7052 : if (!(io_direct_flags & IO_DIRECT_DATA))
1019 : 7052 : pgaio_io_set_flag(ioh, PGAIO_HF_BUFFERED);
1020 : :
1021 : 14104 : pgaio_io_set_target_smgr(ioh,
1022 : 7052 : reln,
1023 : 7052 : forknum,
1024 : 7052 : blocknum,
1025 : 7052 : nblocks,
1026 : : false);
1027 : 7052 : pgaio_io_register_callbacks(ioh, PGAIO_HCB_MD_READV, 0);
1028 : :
1029 : 7052 : ret = FileStartReadV(ioh, v->mdfd_vfd, iovcnt, seekpos, WAIT_EVENT_DATA_FILE_READ);
1030 [ + - ]: 7052 : 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 : 7052 : }
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 : 8001 : 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 [ + + ]: 16002 : while (nblocks > 0)
1068 : : {
1069 : 8001 : struct iovec iov[PG_IOV_MAX];
1070 : 8001 : int iovcnt;
1071 : 8001 : pgoff_t seekpos;
1072 : 8001 : int nbytes;
1073 : 8001 : MdfdVec *v;
1074 : 8001 : BlockNumber nblocks_this_segment;
1075 : 8001 : size_t transferred_this_segment;
1076 : 8001 : size_t size_this_segment;
1077 : :
1078 : 8001 : v = _mdfd_getseg(reln, forknum, blocknum, skipFsync,
1079 : : EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
1080 : :
1081 : 8001 : seekpos = (pgoff_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
1082 : :
1083 [ - + ]: 8001 : Assert(seekpos < (pgoff_t) BLCKSZ * RELSEG_SIZE);
1084 : :
1085 : 8001 : nblocks_this_segment =
1086 [ + - ]: 8001 : Min(nblocks,
1087 : : RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
1088 [ + - ]: 8001 : nblocks_this_segment = Min(nblocks_this_segment, lengthof(iov));
1089 : :
1090 [ + - ]: 8001 : if (nblocks_this_segment != nblocks)
1091 [ # # # # ]: 0 : elog(ERROR, "write crosses segment boundary");
1092 : :
1093 : 8001 : iovcnt = buffers_to_iovec(iov, (void **) buffers, nblocks_this_segment);
1094 : 8001 : size_this_segment = nblocks_this_segment * BLCKSZ;
1095 : 8001 : 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 : 8001 : for (;;)
1103 : : {
1104 : 8001 : 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 : 8001 : nbytes = FileWriteV(v->mdfd_vfd, iov, iovcnt, seekpos,
1110 : : WAIT_EVENT_DATA_FILE_WRITE);
1111 : 8001 : 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 [ + - ]: 8001 : 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 : 0 : }
1135 : :
1136 : : /* One loop should usually be enough. */
1137 : 8001 : transferred_this_segment += nbytes;
1138 [ + - ]: 8001 : Assert(transferred_this_segment <= size_this_segment);
1139 [ - + ]: 8001 : if (transferred_this_segment == size_this_segment)
1140 : 8001 : 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 [ + + + + ]: 8001 : if (!skipFsync && !SmgrIsTemp(reln))
1148 : 6998 : register_dirty_segment(reln, forknum, v);
1149 : :
1150 : 8001 : nblocks -= nblocks_this_segment;
1151 : 8001 : buffers += nblocks_this_segment;
1152 : 8001 : blocknum += nblocks_this_segment;
1153 : 8001 : }
1154 : 8001 : }
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 [ # # ]: 0 : 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 : 0 : pgoff_t seekpos;
1177 : 0 : MdfdVec *v;
1178 : 0 : 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 [ # # ]: 0 : Assert(nflush >= 1);
1203 [ # # ]: 0 : Assert(nflush <= nblocks);
1204 : :
1205 : 0 : seekpos = (pgoff_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
1206 : :
1207 : 0 : FileWriteback(v->mdfd_vfd, seekpos, (pgoff_t) BLCKSZ * nflush, WAIT_EVENT_DATA_FILE_FLUSH);
1208 : :
1209 : 0 : nblocks -= nflush;
1210 : 0 : blocknum += nflush;
1211 [ # # # ]: 0 : }
1212 : 0 : }
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 : 538585 : mdnblocks(SMgrRelation reln, ForkNumber forknum)
1224 : : {
1225 : 538585 : MdfdVec *v;
1226 : 538585 : BlockNumber nblocks;
1227 : 538585 : BlockNumber segno;
1228 : :
1229 : 538585 : mdopenfork(reln, forknum, EXTENSION_FAIL);
1230 : :
1231 : : /* mdopen has opened the first segment */
1232 [ + - ]: 538585 : 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 : 538585 : segno = reln->md_num_open_segs[forknum] - 1;
1248 : 538585 : v = &reln->md_seg_fds[forknum][segno];
1249 : :
1250 : 538585 : for (;;)
1251 : : {
1252 : 538585 : nblocks = _mdnblocks(reln, forknum, v);
1253 [ + - ]: 538585 : if (nblocks > ((BlockNumber) RELSEG_SIZE))
1254 [ # # # # ]: 0 : elog(FATAL, "segment too big");
1255 [ + - ]: 538585 : if (nblocks < ((BlockNumber) RELSEG_SIZE))
1256 : 538585 : 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 : 538585 : }
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 : : * If nblocks > curnblk, the request is ignored when we are InRecovery,
1287 : : * otherwise, an error is raised.
1288 : : */
1289 : : void
1290 : 200 : mdtruncate(SMgrRelation reln, ForkNumber forknum,
1291 : : BlockNumber curnblk, BlockNumber nblocks)
1292 : : {
1293 : 200 : BlockNumber priorblocks;
1294 : 200 : int curopensegs;
1295 : :
1296 [ + - ]: 200 : if (nblocks > curnblk)
1297 : : {
1298 : : /* Bogus request ... but no complaint if InRecovery */
1299 [ # # ]: 0 : if (InRecovery)
1300 : 0 : return;
1301 [ # # # # ]: 0 : ereport(ERROR,
1302 : : (errmsg("could not truncate file \"%s\" to %u blocks: it's only %u blocks now",
1303 : : relpath(reln->smgr_rlocator, forknum).str,
1304 : : nblocks, curnblk)));
1305 : 0 : }
1306 [ + + ]: 200 : if (nblocks == curnblk)
1307 : 98 : return; /* no work */
1308 : :
1309 : : /*
1310 : : * Truncate segments, starting at the last one. Starting at the end makes
1311 : : * managing the memory for the fd array easier, should there be errors.
1312 : : */
1313 : 102 : curopensegs = reln->md_num_open_segs[forknum];
1314 [ + + ]: 204 : while (curopensegs > 0)
1315 : : {
1316 : 102 : MdfdVec *v;
1317 : :
1318 : 102 : priorblocks = (curopensegs - 1) * RELSEG_SIZE;
1319 : :
1320 : 102 : v = &reln->md_seg_fds[forknum][curopensegs - 1];
1321 : :
1322 [ - + ]: 102 : if (priorblocks > nblocks)
1323 : : {
1324 : : /*
1325 : : * This segment is no longer active. We truncate the file, but do
1326 : : * not delete it, for reasons explained in the header comments.
1327 : : */
1328 [ # # ]: 0 : if (FileTruncate(v->mdfd_vfd, 0, WAIT_EVENT_DATA_FILE_TRUNCATE) < 0)
1329 [ # # # # ]: 0 : ereport(ERROR,
1330 : : (errcode_for_file_access(),
1331 : : errmsg("could not truncate file \"%s\": %m",
1332 : : FilePathName(v->mdfd_vfd))));
1333 : :
1334 [ # # ]: 0 : if (!SmgrIsTemp(reln))
1335 : 0 : register_dirty_segment(reln, forknum, v);
1336 : :
1337 : : /* we never drop the 1st segment */
1338 [ # # ]: 0 : Assert(v != &reln->md_seg_fds[forknum][0]);
1339 : :
1340 : 0 : FileClose(v->mdfd_vfd);
1341 : 0 : _fdvec_resize(reln, forknum, curopensegs - 1);
1342 : 0 : }
1343 [ + - ]: 102 : else if (priorblocks + ((BlockNumber) RELSEG_SIZE) > nblocks)
1344 : : {
1345 : : /*
1346 : : * This is the last segment we want to keep. Truncate the file to
1347 : : * the right length. NOTE: if nblocks is exactly a multiple K of
1348 : : * RELSEG_SIZE, we will truncate the K+1st segment to 0 length but
1349 : : * keep it. This adheres to the invariant given in the header
1350 : : * comments.
1351 : : */
1352 : 102 : BlockNumber lastsegblocks = nblocks - priorblocks;
1353 : :
1354 [ - + ]: 102 : if (FileTruncate(v->mdfd_vfd, (pgoff_t) lastsegblocks * BLCKSZ, WAIT_EVENT_DATA_FILE_TRUNCATE) < 0)
1355 [ # # # # ]: 0 : ereport(ERROR,
1356 : : (errcode_for_file_access(),
1357 : : errmsg("could not truncate file \"%s\" to %u blocks: %m",
1358 : : FilePathName(v->mdfd_vfd),
1359 : : nblocks)));
1360 [ + + ]: 102 : if (!SmgrIsTemp(reln))
1361 : 49 : register_dirty_segment(reln, forknum, v);
1362 : 102 : }
1363 : : else
1364 : : {
1365 : : /*
1366 : : * We still need this segment, so nothing to do for this and any
1367 : : * earlier segment.
1368 : : */
1369 : 0 : break;
1370 : : }
1371 : 102 : curopensegs--;
1372 [ - + ]: 102 : }
1373 : 200 : }
1374 : :
1375 : : /*
1376 : : * mdregistersync() -- Mark whole relation as needing fsync
1377 : : */
1378 : : void
1379 : 3498 : mdregistersync(SMgrRelation reln, ForkNumber forknum)
1380 : : {
1381 : 3498 : int segno;
1382 : 3498 : int min_inactive_seg;
1383 : :
1384 : : /*
1385 : : * NOTE: mdnblocks makes sure we have opened all active segments, so that
1386 : : * the loop below will get them all!
1387 : : */
1388 : 3498 : mdnblocks(reln, forknum);
1389 : :
1390 : 3498 : min_inactive_seg = segno = reln->md_num_open_segs[forknum];
1391 : :
1392 : : /*
1393 : : * Temporarily open inactive segments, then close them after sync. There
1394 : : * may be some inactive segments left opened after error, but that is
1395 : : * harmless. We don't bother to clean them up and take a risk of further
1396 : : * trouble. The next mdclose() will soon close them.
1397 : : */
1398 [ - + ]: 3498 : while (_mdfd_openseg(reln, forknum, segno, 0) != NULL)
1399 : 0 : segno++;
1400 : :
1401 [ + + ]: 6996 : while (segno > 0)
1402 : : {
1403 : 3498 : MdfdVec *v = &reln->md_seg_fds[forknum][segno - 1];
1404 : :
1405 : 3498 : register_dirty_segment(reln, forknum, v);
1406 : :
1407 : : /* Close inactive segments immediately */
1408 [ + - ]: 3498 : if (segno > min_inactive_seg)
1409 : : {
1410 : 0 : FileClose(v->mdfd_vfd);
1411 : 0 : _fdvec_resize(reln, forknum, segno - 1);
1412 : 0 : }
1413 : :
1414 : 3498 : segno--;
1415 : 3498 : }
1416 : 3498 : }
1417 : :
1418 : : /*
1419 : : * mdimmedsync() -- Immediately sync a relation to stable storage.
1420 : : *
1421 : : * Note that only writes already issued are synced; this routine knows
1422 : : * nothing of dirty buffers that may exist inside the buffer manager. We
1423 : : * sync active and inactive segments; smgrDoPendingSyncs() relies on this.
1424 : : * Consider a relation skipping WAL. Suppose a checkpoint syncs blocks of
1425 : : * some segment, then mdtruncate() renders that segment inactive. If we
1426 : : * crash before the next checkpoint syncs the newly-inactive segment, that
1427 : : * segment may survive recovery, reintroducing unwanted data into the table.
1428 : : */
1429 : : void
1430 : 8 : mdimmedsync(SMgrRelation reln, ForkNumber forknum)
1431 : : {
1432 : 8 : int segno;
1433 : 8 : int min_inactive_seg;
1434 : :
1435 : : /*
1436 : : * NOTE: mdnblocks makes sure we have opened all active segments, so that
1437 : : * the loop below will get them all!
1438 : : */
1439 : 8 : mdnblocks(reln, forknum);
1440 : :
1441 : 8 : min_inactive_seg = segno = reln->md_num_open_segs[forknum];
1442 : :
1443 : : /*
1444 : : * Temporarily open inactive segments, then close them after sync. There
1445 : : * may be some inactive segments left opened after fsync() error, but that
1446 : : * is harmless. We don't bother to clean them up and take a risk of
1447 : : * further trouble. The next mdclose() will soon close them.
1448 : : */
1449 [ - + ]: 8 : while (_mdfd_openseg(reln, forknum, segno, 0) != NULL)
1450 : 0 : segno++;
1451 : :
1452 [ + + ]: 16 : while (segno > 0)
1453 : : {
1454 : 8 : MdfdVec *v = &reln->md_seg_fds[forknum][segno - 1];
1455 : :
1456 : : /*
1457 : : * fsyncs done through mdimmedsync() should be tracked in a separate
1458 : : * IOContext than those done through mdsyncfiletag() to differentiate
1459 : : * between unavoidable client backend fsyncs (e.g. those done during
1460 : : * index build) and those which ideally would have been done by the
1461 : : * checkpointer. Since other IO operations bypassing the buffer
1462 : : * manager could also be tracked in such an IOContext, wait until
1463 : : * these are also tracked to track immediate fsyncs.
1464 : : */
1465 [ + - ]: 8 : if (FileSync(v->mdfd_vfd, WAIT_EVENT_DATA_FILE_IMMEDIATE_SYNC) < 0)
1466 [ # # # # : 0 : ereport(data_sync_elevel(ERROR),
# # ]
1467 : : (errcode_for_file_access(),
1468 : : errmsg("could not fsync file \"%s\": %m",
1469 : : FilePathName(v->mdfd_vfd))));
1470 : :
1471 : : /* Close inactive segments immediately */
1472 [ + - ]: 8 : if (segno > min_inactive_seg)
1473 : : {
1474 : 0 : FileClose(v->mdfd_vfd);
1475 : 0 : _fdvec_resize(reln, forknum, segno - 1);
1476 : 0 : }
1477 : :
1478 : 8 : segno--;
1479 : 8 : }
1480 : 8 : }
1481 : :
1482 : : int
1483 : 287 : mdfd(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, uint32 *off)
1484 : : {
1485 : 287 : MdfdVec *v = mdopenfork(reln, forknum, EXTENSION_FAIL);
1486 : :
1487 : 287 : v = _mdfd_getseg(reln, forknum, blocknum, false,
1488 : : EXTENSION_FAIL);
1489 : :
1490 : 287 : *off = (pgoff_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
1491 : :
1492 [ + - ]: 287 : Assert(*off < (pgoff_t) BLCKSZ * RELSEG_SIZE);
1493 : :
1494 : 574 : return FileGetRawDesc(v->mdfd_vfd);
1495 : 287 : }
1496 : :
1497 : : /*
1498 : : * register_dirty_segment() -- Mark a relation segment as needing fsync
1499 : : *
1500 : : * If there is a local pending-ops table, just make an entry in it for
1501 : : * ProcessSyncRequests to process later. Otherwise, try to pass off the
1502 : : * fsync request to the checkpointer process. If that fails, just do the
1503 : : * fsync locally before returning (we hope this will not happen often
1504 : : * enough to be a performance problem).
1505 : : */
1506 : : static void
1507 : 46316 : register_dirty_segment(SMgrRelation reln, ForkNumber forknum, MdfdVec *seg)
1508 : : {
1509 : 46316 : FileTag tag;
1510 : :
1511 : 46316 : INIT_MD_FILETAG(tag, reln->smgr_rlocator.locator, forknum, seg->mdfd_segno);
1512 : :
1513 : : /* Temp relations should never be fsync'd */
1514 [ + - ]: 46316 : Assert(!SmgrIsTemp(reln));
1515 : :
1516 [ + - ]: 46316 : if (!RegisterSyncRequest(&tag, SYNC_REQUEST, false /* retryOnError */ ))
1517 : : {
1518 : 0 : instr_time io_start;
1519 : :
1520 [ # # # # ]: 0 : ereport(DEBUG1,
1521 : : (errmsg_internal("could not forward fsync request because request queue is full")));
1522 : :
1523 : 0 : io_start = pgstat_prepare_io_time(track_io_timing);
1524 : :
1525 [ # # ]: 0 : if (FileSync(seg->mdfd_vfd, WAIT_EVENT_DATA_FILE_SYNC) < 0)
1526 [ # # # # : 0 : ereport(data_sync_elevel(ERROR),
# # ]
1527 : : (errcode_for_file_access(),
1528 : : errmsg("could not fsync file \"%s\": %m",
1529 : : FilePathName(seg->mdfd_vfd))));
1530 : :
1531 : : /*
1532 : : * We have no way of knowing if the current IOContext is
1533 : : * IOCONTEXT_NORMAL or IOCONTEXT_[BULKREAD, BULKWRITE, VACUUM] at this
1534 : : * point, so count the fsync as being in the IOCONTEXT_NORMAL
1535 : : * IOContext. This is probably okay, because the number of backend
1536 : : * fsyncs doesn't say anything about the efficacy of the
1537 : : * BufferAccessStrategy. And counting both fsyncs done in
1538 : : * IOCONTEXT_NORMAL and IOCONTEXT_[BULKREAD, BULKWRITE, VACUUM] under
1539 : : * IOCONTEXT_NORMAL is likely clearer when investigating the number of
1540 : : * backend fsyncs.
1541 : : */
1542 : 0 : pgstat_count_io_op_time(IOOBJECT_RELATION, IOCONTEXT_NORMAL,
1543 : : IOOP_FSYNC, io_start, 1, 0);
1544 : 0 : }
1545 : 46316 : }
1546 : :
1547 : : /*
1548 : : * register_unlink_segment() -- Schedule a file to be deleted after next checkpoint
1549 : : */
1550 : : static void
1551 : 7912 : register_unlink_segment(RelFileLocatorBackend rlocator, ForkNumber forknum,
1552 : : BlockNumber segno)
1553 : : {
1554 : 7912 : FileTag tag;
1555 : :
1556 : 7912 : INIT_MD_FILETAG(tag, rlocator.locator, forknum, segno);
1557 : :
1558 : : /* Should never be used with temp relations */
1559 [ + - ]: 7912 : Assert(!RelFileLocatorBackendIsTemp(rlocator));
1560 : :
1561 : 7912 : RegisterSyncRequest(&tag, SYNC_UNLINK_REQUEST, true /* retryOnError */ );
1562 : 7912 : }
1563 : :
1564 : : /*
1565 : : * register_forget_request() -- forget any fsyncs for a relation fork's segment
1566 : : */
1567 : : static void
1568 : 23736 : register_forget_request(RelFileLocatorBackend rlocator, ForkNumber forknum,
1569 : : BlockNumber segno)
1570 : : {
1571 : 23736 : FileTag tag;
1572 : :
1573 : 23736 : INIT_MD_FILETAG(tag, rlocator.locator, forknum, segno);
1574 : :
1575 : 23736 : RegisterSyncRequest(&tag, SYNC_FORGET_REQUEST, true /* retryOnError */ );
1576 : 23736 : }
1577 : :
1578 : : /*
1579 : : * ForgetDatabaseSyncRequests -- forget any fsyncs and unlinks for a DB
1580 : : */
1581 : : void
1582 : 1 : ForgetDatabaseSyncRequests(Oid dbid)
1583 : : {
1584 : 1 : FileTag tag;
1585 : 1 : RelFileLocator rlocator;
1586 : :
1587 : 1 : rlocator.dbOid = dbid;
1588 : 1 : rlocator.spcOid = 0;
1589 : 1 : rlocator.relNumber = 0;
1590 : :
1591 : 1 : INIT_MD_FILETAG(tag, rlocator, InvalidForkNumber, InvalidBlockNumber);
1592 : :
1593 : 1 : RegisterSyncRequest(&tag, SYNC_FILTER_REQUEST, true /* retryOnError */ );
1594 : 1 : }
1595 : :
1596 : : /*
1597 : : * DropRelationFiles -- drop files of all given relations
1598 : : */
1599 : : void
1600 : 0 : DropRelationFiles(RelFileLocator *delrels, int ndelrels, bool isRedo)
1601 : : {
1602 : 0 : SMgrRelation *srels;
1603 : 0 : int i;
1604 : :
1605 : 0 : srels = palloc_array(SMgrRelation, ndelrels);
1606 [ # # ]: 0 : for (i = 0; i < ndelrels; i++)
1607 : : {
1608 : 0 : SMgrRelation srel = smgropen(delrels[i], INVALID_PROC_NUMBER);
1609 : :
1610 [ # # ]: 0 : if (isRedo)
1611 : : {
1612 : 0 : ForkNumber fork;
1613 : :
1614 [ # # ]: 0 : for (fork = 0; fork <= MAX_FORKNUM; fork++)
1615 : 0 : XLogDropRelation(delrels[i], fork);
1616 : 0 : }
1617 : 0 : srels[i] = srel;
1618 : 0 : }
1619 : :
1620 : 0 : smgrdounlinkall(srels, ndelrels, isRedo);
1621 : :
1622 [ # # ]: 0 : for (i = 0; i < ndelrels; i++)
1623 : 0 : smgrclose(srels[i]);
1624 : 0 : pfree(srels);
1625 : 0 : }
1626 : :
1627 : :
1628 : : /*
1629 : : * _fdvec_resize() -- Resize the fork's open segments array
1630 : : */
1631 : : static void
1632 : 96128 : _fdvec_resize(SMgrRelation reln,
1633 : : ForkNumber forknum,
1634 : : int nseg)
1635 : : {
1636 [ + + ]: 96128 : if (nseg == 0)
1637 : : {
1638 [ - + ]: 42015 : if (reln->md_num_open_segs[forknum] > 0)
1639 : : {
1640 : 42015 : pfree(reln->md_seg_fds[forknum]);
1641 : 42015 : reln->md_seg_fds[forknum] = NULL;
1642 : 42015 : }
1643 : 42015 : }
1644 [ - + ]: 54113 : else if (reln->md_num_open_segs[forknum] == 0)
1645 : : {
1646 : 54113 : reln->md_seg_fds[forknum] =
1647 : 54113 : MemoryContextAlloc(MdCxt, sizeof(MdfdVec) * nseg);
1648 : 54113 : }
1649 [ # # ]: 0 : else if (nseg > reln->md_num_open_segs[forknum])
1650 : : {
1651 : : /*
1652 : : * It doesn't seem worthwhile complicating the code to amortize
1653 : : * repalloc() calls. Those are far faster than PathNameOpenFile() or
1654 : : * FileClose(), and the memory context internally will sometimes avoid
1655 : : * doing an actual reallocation.
1656 : : */
1657 : 0 : reln->md_seg_fds[forknum] =
1658 : 0 : repalloc(reln->md_seg_fds[forknum],
1659 : 0 : sizeof(MdfdVec) * nseg);
1660 : 0 : }
1661 : : else
1662 : : {
1663 : : /*
1664 : : * We don't reallocate a smaller array, because we want mdtruncate()
1665 : : * to be able to promise that it won't allocate memory, so that it is
1666 : : * allowed in a critical section. This means that a bit of space in
1667 : : * the array is now wasted, until the next time we add a segment and
1668 : : * reallocate.
1669 : : */
1670 : : }
1671 : :
1672 : 96128 : reln->md_num_open_segs[forknum] = nseg;
1673 : 96128 : }
1674 : :
1675 : : /*
1676 : : * Return the filename for the specified segment of the relation. The
1677 : : * returned string is palloc'd.
1678 : : */
1679 : : static MdPathStr
1680 : 3506 : _mdfd_segpath(SMgrRelation reln, ForkNumber forknum, BlockNumber segno)
1681 : : {
1682 : 3506 : RelPathStr path;
1683 : : MdPathStr fullpath;
1684 : :
1685 : 3506 : path = relpath(reln->smgr_rlocator, forknum);
1686 : :
1687 [ + - ]: 3506 : if (segno > 0)
1688 : 3506 : sprintf(fullpath.str, "%s.%u", path.str, segno);
1689 : : else
1690 : 0 : strcpy(fullpath.str, path.str);
1691 : :
1692 : : return fullpath;
1693 : 3506 : }
1694 : :
1695 : : /*
1696 : : * Open the specified segment of the relation,
1697 : : * and make a MdfdVec object for it. Returns NULL on failure.
1698 : : */
1699 : : static MdfdVec *
1700 : 3506 : _mdfd_openseg(SMgrRelation reln, ForkNumber forknum, BlockNumber segno,
1701 : : int oflags)
1702 : : {
1703 : 3506 : MdfdVec *v;
1704 : 3506 : File fd;
1705 : 3506 : MdPathStr fullpath;
1706 : :
1707 : 3506 : fullpath = _mdfd_segpath(reln, forknum, segno);
1708 : :
1709 : : /* open the file */
1710 : 3506 : fd = PathNameOpenFile(fullpath.str, _mdfd_open_flags() | oflags);
1711 : :
1712 [ - + ]: 3506 : if (fd < 0)
1713 : 3506 : return NULL;
1714 : :
1715 : : /*
1716 : : * Segments are always opened in order from lowest to highest, so we must
1717 : : * be adding a new one at the end.
1718 : : */
1719 [ # # ]: 0 : Assert(segno == reln->md_num_open_segs[forknum]);
1720 : :
1721 : 0 : _fdvec_resize(reln, forknum, segno + 1);
1722 : :
1723 : : /* fill the entry */
1724 : 0 : v = &reln->md_seg_fds[forknum][segno];
1725 : 0 : v->mdfd_vfd = fd;
1726 : 0 : v->mdfd_segno = segno;
1727 : :
1728 [ # # ]: 0 : Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
1729 : :
1730 : : /* all done */
1731 : 0 : return v;
1732 : 3506 : }
1733 : :
1734 : : /*
1735 : : * _mdfd_getseg() -- Find the segment of the relation holding the
1736 : : * specified block.
1737 : : *
1738 : : * If the segment doesn't exist, we ereport, return NULL, or create the
1739 : : * segment, according to "behavior". Note: skipFsync is only used in the
1740 : : * EXTENSION_CREATE case.
1741 : : */
1742 : : static MdfdVec *
1743 : 53441 : _mdfd_getseg(SMgrRelation reln, ForkNumber forknum, BlockNumber blkno,
1744 : : bool skipFsync, int behavior)
1745 : : {
1746 : 53441 : MdfdVec *v;
1747 : 53441 : BlockNumber targetseg;
1748 : 53441 : BlockNumber nextsegno;
1749 : :
1750 : : /* some way to handle non-existent segments needs to be specified */
1751 [ + - ]: 53441 : Assert(behavior &
1752 : : (EXTENSION_FAIL | EXTENSION_CREATE | EXTENSION_RETURN_NULL |
1753 : : EXTENSION_DONT_OPEN));
1754 : :
1755 : 53441 : targetseg = blkno / ((BlockNumber) RELSEG_SIZE);
1756 : :
1757 : : /* if an existing and opened segment, we're done */
1758 [ + + ]: 53441 : if (targetseg < reln->md_num_open_segs[forknum])
1759 : : {
1760 : 52561 : v = &reln->md_seg_fds[forknum][targetseg];
1761 : 52561 : return v;
1762 : : }
1763 : :
1764 : : /* The caller only wants the segment if we already had it open. */
1765 [ - + ]: 880 : if (behavior & EXTENSION_DONT_OPEN)
1766 : 0 : return NULL;
1767 : :
1768 : : /*
1769 : : * The target segment is not yet open. Iterate over all the segments
1770 : : * between the last opened and the target segment. This way missing
1771 : : * segments either raise an error, or get created (according to
1772 : : * 'behavior'). Start with either the last opened, or the first segment if
1773 : : * none was opened before.
1774 : : */
1775 [ - + ]: 880 : if (reln->md_num_open_segs[forknum] > 0)
1776 : 0 : v = &reln->md_seg_fds[forknum][reln->md_num_open_segs[forknum] - 1];
1777 : : else
1778 : : {
1779 : 880 : v = mdopenfork(reln, forknum, behavior);
1780 [ + - ]: 880 : if (!v)
1781 : 0 : return NULL; /* if behavior & EXTENSION_RETURN_NULL */
1782 : : }
1783 : :
1784 [ - + ]: 880 : for (nextsegno = reln->md_num_open_segs[forknum];
1785 : 880 : nextsegno <= targetseg; nextsegno++)
1786 : : {
1787 : 0 : BlockNumber nblocks = _mdnblocks(reln, forknum, v);
1788 : 0 : int flags = 0;
1789 : :
1790 [ # # ]: 0 : Assert(nextsegno == v->mdfd_segno + 1);
1791 : :
1792 [ # # ]: 0 : if (nblocks > ((BlockNumber) RELSEG_SIZE))
1793 [ # # # # ]: 0 : elog(FATAL, "segment too big");
1794 : :
1795 [ # # # # ]: 0 : if ((behavior & EXTENSION_CREATE) ||
1796 [ # # ]: 0 : (InRecovery && (behavior & EXTENSION_CREATE_RECOVERY)))
1797 : : {
1798 : : /*
1799 : : * Normally we will create new segments only if authorized by the
1800 : : * caller (i.e., we are doing mdextend()). But when doing WAL
1801 : : * recovery, create segments anyway; this allows cases such as
1802 : : * replaying WAL data that has a write into a high-numbered
1803 : : * segment of a relation that was later deleted. We want to go
1804 : : * ahead and create the segments so we can finish out the replay.
1805 : : *
1806 : : * We have to maintain the invariant that segments before the last
1807 : : * active segment are of size RELSEG_SIZE; therefore, if
1808 : : * extending, pad them out with zeroes if needed. (This only
1809 : : * matters if in recovery, or if the caller is extending the
1810 : : * relation discontiguously, but that can happen in hash indexes.)
1811 : : */
1812 [ # # ]: 0 : if (nblocks < ((BlockNumber) RELSEG_SIZE))
1813 : : {
1814 : 0 : char *zerobuf = palloc_aligned(BLCKSZ, PG_IO_ALIGN_SIZE,
1815 : : MCXT_ALLOC_ZERO);
1816 : :
1817 : 0 : mdextend(reln, forknum,
1818 : 0 : nextsegno * ((BlockNumber) RELSEG_SIZE) - 1,
1819 : 0 : zerobuf, skipFsync);
1820 : 0 : pfree(zerobuf);
1821 : 0 : }
1822 : 0 : flags = O_CREAT;
1823 : 0 : }
1824 [ # # ]: 0 : else if (nblocks < ((BlockNumber) RELSEG_SIZE))
1825 : : {
1826 : : /*
1827 : : * When not extending, only open the next segment if the current
1828 : : * one is exactly RELSEG_SIZE. If not (this branch), either
1829 : : * return NULL or fail.
1830 : : */
1831 [ # # ]: 0 : if (behavior & EXTENSION_RETURN_NULL)
1832 : : {
1833 : : /*
1834 : : * Some callers discern between reasons for _mdfd_getseg()
1835 : : * returning NULL based on errno. As there's no failing
1836 : : * syscall involved in this case, explicitly set errno to
1837 : : * ENOENT, as that seems the closest interpretation.
1838 : : */
1839 : 0 : errno = ENOENT;
1840 : 0 : return NULL;
1841 : : }
1842 : :
1843 [ # # # # ]: 0 : ereport(ERROR,
1844 : : (errcode_for_file_access(),
1845 : : errmsg("could not open file \"%s\" (target block %u): previous segment is only %u blocks",
1846 : : _mdfd_segpath(reln, forknum, nextsegno).str,
1847 : : blkno, nblocks)));
1848 : 0 : }
1849 : :
1850 : 0 : v = _mdfd_openseg(reln, forknum, nextsegno, flags);
1851 : :
1852 [ # # ]: 0 : if (v == NULL)
1853 : : {
1854 [ # # ]: 0 : if ((behavior & EXTENSION_RETURN_NULL) &&
1855 : 0 : FILE_POSSIBLY_DELETED(errno))
1856 : 0 : return NULL;
1857 [ # # # # ]: 0 : ereport(ERROR,
1858 : : (errcode_for_file_access(),
1859 : : errmsg("could not open file \"%s\" (target block %u): %m",
1860 : : _mdfd_segpath(reln, forknum, nextsegno).str,
1861 : : blkno)));
1862 : 0 : }
1863 [ # # ]: 0 : }
1864 : :
1865 : 880 : return v;
1866 : 53441 : }
1867 : :
1868 : : /*
1869 : : * Get number of blocks present in a single disk file
1870 : : */
1871 : : static BlockNumber
1872 : 618126 : _mdnblocks(SMgrRelation reln, ForkNumber forknum, MdfdVec *seg)
1873 : : {
1874 : 618126 : pgoff_t len;
1875 : :
1876 : 618126 : len = FileSize(seg->mdfd_vfd);
1877 [ + - ]: 618126 : if (len < 0)
1878 [ # # # # ]: 0 : ereport(ERROR,
1879 : : (errcode_for_file_access(),
1880 : : errmsg("could not seek to end of file \"%s\": %m",
1881 : : FilePathName(seg->mdfd_vfd))));
1882 : : /* note that this calculation will ignore any partial block at EOF */
1883 : 1236252 : return (BlockNumber) (len / BLCKSZ);
1884 : 618126 : }
1885 : :
1886 : : /*
1887 : : * Sync a file to disk, given a file tag. Write the path into an output
1888 : : * buffer so the caller can use it in error messages.
1889 : : *
1890 : : * Return 0 on success, -1 on failure, with errno set.
1891 : : */
1892 : : int
1893 : 0 : mdsyncfiletag(const FileTag *ftag, char *path)
1894 : : {
1895 : 0 : SMgrRelation reln = smgropen(ftag->rlocator, INVALID_PROC_NUMBER);
1896 : 0 : File file;
1897 : 0 : instr_time io_start;
1898 : 0 : bool need_to_close;
1899 : 0 : int result,
1900 : : save_errno;
1901 : :
1902 : : /* See if we already have the file open, or need to open it. */
1903 [ # # ]: 0 : if (ftag->segno < reln->md_num_open_segs[ftag->forknum])
1904 : : {
1905 : 0 : file = reln->md_seg_fds[ftag->forknum][ftag->segno].mdfd_vfd;
1906 : 0 : strlcpy(path, FilePathName(file), MAXPGPATH);
1907 : 0 : need_to_close = false;
1908 : 0 : }
1909 : : else
1910 : : {
1911 : 0 : MdPathStr p;
1912 : :
1913 : 0 : p = _mdfd_segpath(reln, ftag->forknum, ftag->segno);
1914 : 0 : strlcpy(path, p.str, MD_PATH_STR_MAXLEN);
1915 : :
1916 : 0 : file = PathNameOpenFile(path, _mdfd_open_flags());
1917 [ # # ]: 0 : if (file < 0)
1918 : 0 : return -1;
1919 : 0 : need_to_close = true;
1920 [ # # ]: 0 : }
1921 : :
1922 : 0 : io_start = pgstat_prepare_io_time(track_io_timing);
1923 : :
1924 : : /* Sync the file. */
1925 : 0 : result = FileSync(file, WAIT_EVENT_DATA_FILE_SYNC);
1926 : 0 : save_errno = errno;
1927 : :
1928 [ # # ]: 0 : if (need_to_close)
1929 : 0 : FileClose(file);
1930 : :
1931 : 0 : pgstat_count_io_op_time(IOOBJECT_RELATION, IOCONTEXT_NORMAL,
1932 : : IOOP_FSYNC, io_start, 1, 0);
1933 : :
1934 : 0 : errno = save_errno;
1935 : 0 : return result;
1936 : 0 : }
1937 : :
1938 : : /*
1939 : : * Unlink a file, given a file tag. Write the path into an output
1940 : : * buffer so the caller can use it in error messages.
1941 : : *
1942 : : * Return 0 on success, -1 on failure, with errno set.
1943 : : */
1944 : : int
1945 : 0 : mdunlinkfiletag(const FileTag *ftag, char *path)
1946 : : {
1947 : 0 : RelPathStr p;
1948 : :
1949 : : /* Compute the path. */
1950 : 0 : p = relpathperm(ftag->rlocator, MAIN_FORKNUM);
1951 : 0 : strlcpy(path, p.str, MAXPGPATH);
1952 : :
1953 : : /* Try to unlink the file. */
1954 : 0 : return unlink(path);
1955 : 0 : }
1956 : :
1957 : : /*
1958 : : * Check if a given candidate request matches a given tag, when processing
1959 : : * a SYNC_FILTER_REQUEST request. This will be called for all pending
1960 : : * requests to find out whether to forget them.
1961 : : */
1962 : : bool
1963 : 0 : mdfiletagmatches(const FileTag *ftag, const FileTag *candidate)
1964 : : {
1965 : : /*
1966 : : * For now we only use filter requests as a way to drop all scheduled
1967 : : * callbacks relating to a given database, when dropping the database.
1968 : : * We'll return true for all candidates that have the same database OID as
1969 : : * the ftag from the SYNC_FILTER_REQUEST request, so they're forgotten.
1970 : : */
1971 : 0 : return ftag->rlocator.dbOid == candidate->rlocator.dbOid;
1972 : : }
1973 : :
1974 : : /*
1975 : : * AIO completion callback for mdstartreadv().
1976 : : */
1977 : : static PgAioResult
1978 : 6681 : md_readv_complete(PgAioHandle *ioh, PgAioResult prior_result, uint8 cb_data)
1979 : : {
1980 : 6681 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
1981 : 6681 : PgAioResult result = prior_result;
1982 : :
1983 [ + - ]: 6681 : if (prior_result.result < 0)
1984 : : {
1985 : 0 : result.status = PGAIO_RS_ERROR;
1986 : 0 : result.id = PGAIO_HCB_MD_READV;
1987 : : /* For "hard" errors, track the error number in error_data */
1988 : 0 : result.error_data = -prior_result.result;
1989 : 0 : result.result = 0;
1990 : :
1991 : : /*
1992 : : * Immediately log a message about the IO error, but only to the
1993 : : * server log. The reason to do so immediately is that the originator
1994 : : * might not process the query result immediately (because it is busy
1995 : : * doing another part of query processing) or at all (e.g. if it was
1996 : : * cancelled or errored out due to another IO also failing). The
1997 : : * definer of the IO will emit an ERROR when processing the IO's
1998 : : * results
1999 : : */
2000 : 0 : pgaio_result_report(result, td, LOG_SERVER_ONLY);
2001 : :
2002 : 0 : return result;
2003 : : }
2004 : :
2005 : : /*
2006 : : * As explained above smgrstartreadv(), the smgr API operates on the level
2007 : : * of blocks, rather than bytes. Convert.
2008 : : */
2009 : 6681 : result.result /= BLCKSZ;
2010 : :
2011 [ + - ]: 6681 : Assert(result.result <= td->smgr.nblocks);
2012 : :
2013 [ + - ]: 6681 : if (result.result == 0)
2014 : : {
2015 : : /* consider 0 blocks read a failure */
2016 : 0 : result.status = PGAIO_RS_ERROR;
2017 : 0 : result.id = PGAIO_HCB_MD_READV;
2018 : 0 : result.error_data = 0;
2019 : :
2020 : : /* see comment above the "hard error" case */
2021 : 0 : pgaio_result_report(result, td, LOG_SERVER_ONLY);
2022 : :
2023 : 0 : return result;
2024 : : }
2025 : :
2026 [ + - + - ]: 6681 : if (result.status != PGAIO_RS_ERROR &&
2027 : 6681 : result.result < td->smgr.nblocks)
2028 : : {
2029 : : /* partial reads should be retried at upper level */
2030 : 0 : result.status = PGAIO_RS_PARTIAL;
2031 : 0 : result.id = PGAIO_HCB_MD_READV;
2032 : 0 : }
2033 : :
2034 : 6681 : return result;
2035 : 6681 : }
2036 : :
2037 : : /*
2038 : : * AIO error reporting callback for mdstartreadv().
2039 : : *
2040 : : * Errors are encoded as follows:
2041 : : * - PgAioResult.error_data != 0 encodes IO that failed with that errno
2042 : : * - PgAioResult.error_data == 0 encodes IO that didn't read all data
2043 : : */
2044 : : static void
2045 : 0 : md_readv_report(PgAioResult result, const PgAioTargetData *td, int elevel)
2046 : : {
2047 : 0 : RelPathStr path;
2048 : :
2049 [ # # ]: 0 : path = relpathbackend(td->smgr.rlocator,
2050 : : td->smgr.is_temp ? MyProcNumber : INVALID_PROC_NUMBER,
2051 : : td->smgr.forkNum);
2052 : :
2053 [ # # ]: 0 : if (result.error_data != 0)
2054 : : {
2055 : : /* for errcode_for_file_access() and %m */
2056 : 0 : errno = result.error_data;
2057 : :
2058 [ # # # # : 0 : ereport(elevel,
# # # # #
# ]
2059 : : errcode_for_file_access(),
2060 : : errmsg("could not read blocks %u..%u in file \"%s\": %m",
2061 : : td->smgr.blockNum,
2062 : : td->smgr.blockNum + td->smgr.nblocks - 1,
2063 : : path.str));
2064 : 0 : }
2065 : : else
2066 : : {
2067 : : /*
2068 : : * NB: This will typically only be output in debug messages, while
2069 : : * retrying a partial IO.
2070 : : */
2071 [ # # # # : 0 : ereport(elevel,
# # # # #
# ]
2072 : : errcode(ERRCODE_DATA_CORRUPTED),
2073 : : errmsg("could not read blocks %u..%u in file \"%s\": read only %zu of %zu bytes",
2074 : : td->smgr.blockNum,
2075 : : td->smgr.blockNum + td->smgr.nblocks - 1,
2076 : : path.str,
2077 : : result.result * (size_t) BLCKSZ,
2078 : : td->smgr.nblocks * (size_t) BLCKSZ));
2079 : : }
2080 : 0 : }
|
