Age Owner Branch data TLA Line data Source code
1 : : /*-------------------------------------------------------------------------
2 : : *
3 : : * md.c
4 : : * This code manages relations that reside on magnetic disk.
5 : : *
6 : : * Or at least, that was what the Berkeley folk had in mind when they named
7 : : * this file. In reality, what this code provides is an interface from
8 : : * the smgr API to Unix-like filesystem APIs, so it will work with any type
9 : : * of device for which the operating system provides filesystem support.
10 : : * It doesn't matter whether the bits are on spinning rust or some other
11 : : * storage technology.
12 : : *
13 : : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
14 : : * Portions Copyright (c) 1994, Regents of the University of California
15 : : *
16 : : *
17 : : * IDENTIFICATION
18 : : * src/backend/storage/smgr/md.c
19 : : *
20 : : *-------------------------------------------------------------------------
21 : : */
22 : : #include "postgres.h"
23 : :
24 : : #include <unistd.h>
25 : : #include <fcntl.h>
26 : : #include <sys/file.h>
27 : :
28 : : #include "access/xlogutils.h"
29 : : #include "commands/tablespace.h"
30 : : #include "common/file_utils.h"
31 : : #include "miscadmin.h"
32 : : #include "pg_trace.h"
33 : : #include "pgstat.h"
34 : : #include "storage/aio.h"
35 : : #include "storage/bufmgr.h"
36 : : #include "storage/fd.h"
37 : : #include "storage/md.h"
38 : : #include "storage/relfilelocator.h"
39 : : #include "storage/smgr.h"
40 : : #include "storage/sync.h"
41 : : #include "utils/memutils.h"
42 : :
43 : : /*
44 : : * The magnetic disk storage manager keeps track of open file
45 : : * descriptors in its own descriptor pool. This is done to make it
46 : : * easier to support relations that are larger than the operating
47 : : * system's file size limit (often 2GBytes). In order to do that,
48 : : * we break relations up into "segment" files that are each shorter than
49 : : * the OS file size limit. The segment size is set by the RELSEG_SIZE
50 : : * configuration constant in pg_config.h.
51 : : *
52 : : * On disk, a relation must consist of consecutively numbered segment
53 : : * files in the pattern
54 : : * -- Zero or more full segments of exactly RELSEG_SIZE blocks each
55 : : * -- Exactly one partial segment of size 0 <= size < RELSEG_SIZE blocks
56 : : * -- Optionally, any number of inactive segments of size 0 blocks.
57 : : * The full and partial segments are collectively the "active" segments.
58 : : * Inactive segments are those that once contained data but are currently
59 : : * not needed because of an mdtruncate() operation. The reason for leaving
60 : : * them present at size zero, rather than unlinking them, is that other
61 : : * backends and/or the checkpointer might be holding open file references to
62 : : * such segments. If the relation expands again after mdtruncate(), such
63 : : * that a deactivated segment becomes active again, it is important that
64 : : * such file references still be valid --- else data might get written
65 : : * out to an unlinked old copy of a segment file that will eventually
66 : : * disappear.
67 : : *
68 : : * File descriptors are stored in the per-fork md_seg_fds arrays inside
69 : : * SMgrRelation. The length of these arrays is stored in md_num_open_segs.
70 : : * Note that a fork's md_num_open_segs having a specific value does not
71 : : * necessarily mean the relation doesn't have additional segments; we may
72 : : * just not have opened the next segment yet. (We could not have "all
73 : : * segments are in the array" as an invariant anyway, since another backend
74 : : * could extend the relation while we aren't looking.) We do not have
75 : : * entries for inactive segments, however; as soon as we find a partial
76 : : * segment, we assume that any subsequent segments are inactive.
77 : : *
78 : : * The entire MdfdVec array is palloc'd in the MdCxt memory context.
79 : : */
80 : :
81 : : typedef struct _MdfdVec
82 : : {
83 : : File mdfd_vfd; /* fd number in fd.c's pool */
84 : : BlockNumber mdfd_segno; /* segment number, from 0 */
85 : : } MdfdVec;
86 : :
87 : : static MemoryContext MdCxt; /* context for all MdfdVec objects */
88 : :
89 : :
90 : : /* Populate a file tag describing an md.c segment file. */
91 : : #define INIT_MD_FILETAG(a,xx_rlocator,xx_forknum,xx_segno) \
92 : : ( \
93 : : memset(&(a), 0, sizeof(FileTag)), \
94 : : (a).handler = SYNC_HANDLER_MD, \
95 : : (a).rlocator = (xx_rlocator), \
96 : : (a).forknum = (xx_forknum), \
97 : : (a).segno = (xx_segno) \
98 : : )
99 : :
100 : :
101 : : /*** behavior for mdopen & _mdfd_getseg ***/
102 : : /* ereport if segment not present */
103 : : #define EXTENSION_FAIL (1 << 0)
104 : : /* return NULL if segment not present */
105 : : #define EXTENSION_RETURN_NULL (1 << 1)
106 : : /* create new segments as needed */
107 : : #define EXTENSION_CREATE (1 << 2)
108 : : /* create new segments if needed during recovery */
109 : : #define EXTENSION_CREATE_RECOVERY (1 << 3)
110 : : /* don't try to open a segment, if not already open */
111 : : #define EXTENSION_DONT_OPEN (1 << 5)
112 : :
113 : :
114 : : /*
115 : : * Fixed-length string to represent paths to files that need to be built by
116 : : * md.c.
117 : : *
118 : : * The maximum number of segments is MaxBlockNumber / RELSEG_SIZE, where
119 : : * RELSEG_SIZE can be set to 1 (for testing only).
120 : : */
121 : : #define SEGMENT_CHARS OIDCHARS
122 : : #define MD_PATH_STR_MAXLEN \
123 : : (\
124 : : REL_PATH_STR_MAXLEN \
125 : : + sizeof((char)'.') \
126 : : + SEGMENT_CHARS \
127 : : )
128 : : typedef struct MdPathStr
129 : : {
130 : : char str[MD_PATH_STR_MAXLEN + 1];
131 : : } MdPathStr;
132 : :
133 : :
134 : : /* local routines */
135 : : static void mdunlinkfork(RelFileLocatorBackend rlocator, ForkNumber forknum,
136 : : bool isRedo);
137 : : static MdfdVec *mdopenfork(SMgrRelation reln, ForkNumber forknum, int behavior);
138 : : static void register_dirty_segment(SMgrRelation reln, ForkNumber forknum,
139 : : MdfdVec *seg);
140 : : static void register_unlink_segment(RelFileLocatorBackend rlocator, ForkNumber forknum,
141 : : BlockNumber segno);
142 : : static void register_forget_request(RelFileLocatorBackend rlocator, ForkNumber forknum,
143 : : BlockNumber segno);
144 : : static void _fdvec_resize(SMgrRelation reln,
145 : : ForkNumber forknum,
146 : : int nseg);
147 : : static MdPathStr _mdfd_segpath(SMgrRelation reln, ForkNumber forknum,
148 : : BlockNumber segno);
149 : : static MdfdVec *_mdfd_openseg(SMgrRelation reln, ForkNumber forknum,
150 : : BlockNumber segno, int oflags);
151 : : static MdfdVec *_mdfd_getseg(SMgrRelation reln, ForkNumber forknum,
152 : : BlockNumber blkno, bool skipFsync, int behavior);
153 : : static BlockNumber _mdnblocks(SMgrRelation reln, ForkNumber forknum,
154 : : MdfdVec *seg);
155 : :
156 : : static PgAioResult md_readv_complete(PgAioHandle *ioh, PgAioResult prior_result, uint8 cb_data);
157 : : static void md_readv_report(PgAioResult result, const PgAioTargetData *td, int elevel);
158 : :
159 : : const PgAioHandleCallbacks aio_md_readv_cb = {
160 : : .complete_shared = md_readv_complete,
161 : : .report = md_readv_report,
162 : : };
163 : :
164 : :
165 : : static inline int
882 tmunro@postgresql.or 166 :CBC 1248815 : _mdfd_open_flags(void)
167 : : {
168 : 1248815 : int flags = O_RDWR | PG_BINARY;
169 : :
170 [ + + ]: 1248815 : if (io_direct_flags & IO_DIRECT_DATA)
171 : 309 : flags |= PG_O_DIRECT;
172 : :
173 : 1248815 : return flags;
174 : : }
175 : :
176 : : /*
177 : : * mdinit() -- Initialize private state for magnetic disk storage manager.
178 : : */
179 : : void
8837 tgl@sss.pgh.pa.us 180 : 18768 : mdinit(void)
181 : : {
9201 182 : 18768 : MdCxt = AllocSetContextCreate(TopMemoryContext,
183 : : "MdSmgr",
184 : : ALLOCSET_DEFAULT_SIZES);
5917 heikki.linnakangas@i 185 : 18768 : }
186 : :
187 : : /*
188 : : * mdexists() -- Does the physical file exist?
189 : : *
190 : : * Note: this will return true for lingering files, with pending deletions
191 : : */
192 : : bool
1083 pg@bowt.ie 193 : 499257 : mdexists(SMgrRelation reln, ForkNumber forknum)
194 : : {
195 : : /*
196 : : * Close it first, to ensure that we notice if the fork has been unlinked
197 : : * since we opened it. As an optimization, we can skip that in recovery,
198 : : * which already closes relations when dropping them.
199 : : */
1248 tmunro@postgresql.or 200 [ + + ]: 499257 : if (!InRecovery)
1083 pg@bowt.ie 201 : 479071 : mdclose(reln, forknum);
202 : :
203 : 499257 : return (mdopenfork(reln, forknum, EXTENSION_RETURN_NULL) != NULL);
204 : : }
205 : :
206 : : /*
207 : : * mdcreate() -- Create a new relation on magnetic disk.
208 : : *
209 : : * If isRedo is true, it's okay for the relation to exist already.
210 : : */
211 : : void
212 : 5641912 : mdcreate(SMgrRelation reln, ForkNumber forknum, bool isRedo)
213 : : {
214 : : MdfdVec *mdfd;
215 : : RelPathStr path;
216 : : File fd;
217 : :
218 [ + + + + ]: 5641912 : if (isRedo && reln->md_num_open_segs[forknum] > 0)
6821 tgl@sss.pgh.pa.us 219 : 5486047 : return; /* created and opened already... */
220 : :
1083 pg@bowt.ie 221 [ - + ]: 155865 : Assert(reln->md_num_open_segs[forknum] == 0);
222 : :
223 : : /*
224 : : * We may be using the target table space for the first time in this
225 : : * database, so create a per-database subdirectory if needed.
226 : : *
227 : : * XXX this is a fairly ugly violation of module layering, but this seems
228 : : * to be the best place to put the check. Maybe TablespaceCreateDbspace
229 : : * should be here and not in commands/tablespace.c? But that would imply
230 : : * importing a lot of stuff that smgr.c oughtn't know, either.
231 : : */
1158 rhaas@postgresql.org 232 : 155865 : TablespaceCreateDbspace(reln->smgr_rlocator.locator.spcOid,
233 : : reln->smgr_rlocator.locator.dbOid,
234 : : isRedo);
235 : :
1083 pg@bowt.ie 236 : 155865 : path = relpath(reln->smgr_rlocator, forknum);
237 : :
193 andres@anarazel.de 238 : 155865 : fd = PathNameOpenFile(path.str, _mdfd_open_flags() | O_CREAT | O_EXCL);
239 : :
10226 bruce@momjian.us 240 [ + + ]: 155865 : if (fd < 0)
241 : : {
8934 242 : 3983 : int save_errno = errno;
243 : :
2413 akapila@postgresql.o 244 [ + - ]: 3983 : if (isRedo)
193 andres@anarazel.de 245 : 3983 : fd = PathNameOpenFile(path.str, _mdfd_open_flags());
10226 bruce@momjian.us 246 [ - + ]: 3983 : if (fd < 0)
247 : : {
248 : : /* be sure to report the error reported by create, not open */
9210 tgl@sss.pgh.pa.us 249 :UBC 0 : errno = save_errno;
6821 250 [ # # ]: 0 : ereport(ERROR,
251 : : (errcode_for_file_access(),
252 : : errmsg("could not create file \"%s\": %m", path.str)));
253 : : }
254 : : }
255 : :
1083 pg@bowt.ie 256 :CBC 155865 : _fdvec_resize(reln, forknum, 1);
257 : 155865 : mdfd = &reln->md_seg_fds[forknum][0];
3285 andres@anarazel.de 258 : 155865 : mdfd->mdfd_vfd = fd;
259 : 155865 : mdfd->mdfd_segno = 0;
260 : :
795 heikki.linnakangas@i 261 [ + + ]: 155865 : if (!SmgrIsTemp(reln))
262 : 152527 : register_dirty_segment(reln, forknum, mdfd);
263 : : }
264 : :
265 : : /*
266 : : * mdunlink() -- Unlink a relation.
267 : : *
268 : : * Note that we're passed a RelFileLocatorBackend --- by the time this is called,
269 : : * there won't be an SMgrRelation hashtable entry anymore.
270 : : *
271 : : * forknum can be a fork number to delete a specific fork, or InvalidForkNumber
272 : : * to delete all forks.
273 : : *
274 : : * For regular relations, we don't unlink the first segment file of the rel,
275 : : * but just truncate it to zero length, and record a request to unlink it after
276 : : * the next checkpoint. Additional segments can be unlinked immediately,
277 : : * however. Leaving the empty file in place prevents that relfilenumber
278 : : * from being reused. The scenario this protects us from is:
279 : : * 1. We delete a relation (and commit, and actually remove its file).
280 : : * 2. We create a new relation, which by chance gets the same relfilenumber as
281 : : * the just-deleted one (OIDs must've wrapped around for that to happen).
282 : : * 3. We crash before another checkpoint occurs.
283 : : * During replay, we would delete the file and then recreate it, which is fine
284 : : * if the contents of the file were repopulated by subsequent WAL entries.
285 : : * But if we didn't WAL-log insertions, but instead relied on fsyncing the
286 : : * file after populating it (as we do at wal_level=minimal), the contents of
287 : : * the file would be lost forever. By leaving the empty file until after the
288 : : * next checkpoint, we prevent reassignment of the relfilenumber until it's
289 : : * safe, because relfilenumber assignment skips over any existing file.
290 : : *
291 : : * Additional segments, if any, are truncated and then unlinked. The reason
292 : : * for truncating is that other backends may still hold open FDs for these at
293 : : * the smgr level, so that the kernel can't remove the file yet. We want to
294 : : * reclaim the disk space right away despite that.
295 : : *
296 : : * We do not need to go through this dance for temp relations, though, because
297 : : * we never make WAL entries for temp rels, and so a temp rel poses no threat
298 : : * to the health of a regular rel that has taken over its relfilenumber.
299 : : * The fact that temp rels and regular rels have different file naming
300 : : * patterns provides additional safety. Other backends shouldn't have open
301 : : * FDs for them, either.
302 : : *
303 : : * We also don't do it while performing a binary upgrade. There is no reuse
304 : : * hazard in that case, since after a crash or even a simple ERROR, the
305 : : * upgrade fails and the whole cluster must be recreated from scratch.
306 : : * Furthermore, it is important to remove the files from disk immediately,
307 : : * because we may be about to reuse the same relfilenumber.
308 : : *
309 : : * All the above applies only to the relation's main fork; other forks can
310 : : * just be removed immediately, since they are not needed to prevent the
311 : : * relfilenumber from being recycled. Also, we do not carefully
312 : : * track whether other forks have been created or not, but just attempt to
313 : : * unlink them unconditionally; so we should never complain about ENOENT.
314 : : *
315 : : * If isRedo is true, it's unsurprising for the relation to be already gone.
316 : : * Also, we should remove the file immediately instead of queuing a request
317 : : * for later, since during redo there's no possibility of creating a
318 : : * conflicting relation.
319 : : *
320 : : * Note: we currently just never warn about ENOENT at all. We could warn in
321 : : * the main-fork, non-isRedo case, but it doesn't seem worth the trouble.
322 : : *
323 : : * Note: any failure should be reported as WARNING not ERROR, because
324 : : * we are usually not in a transaction anymore when this is called.
325 : : */
326 : : void
1083 pg@bowt.ie 327 : 180784 : mdunlink(RelFileLocatorBackend rlocator, ForkNumber forknum, bool isRedo)
328 : : {
329 : : /* Now do the per-fork work */
330 [ - + ]: 180784 : if (forknum == InvalidForkNumber)
331 : : {
1083 pg@bowt.ie 332 [ # # ]:UBC 0 : for (forknum = 0; forknum <= MAX_FORKNUM; forknum++)
333 : 0 : mdunlinkfork(rlocator, forknum, isRedo);
334 : : }
335 : : else
1083 pg@bowt.ie 336 :CBC 180784 : mdunlinkfork(rlocator, forknum, isRedo);
4797 tgl@sss.pgh.pa.us 337 : 180784 : }
338 : :
339 : : /*
340 : : * Truncate a file to release disk space.
341 : : */
342 : : static int
1740 tmunro@postgresql.or 343 : 212028 : do_truncate(const char *path)
344 : : {
345 : : int save_errno;
346 : : int ret;
347 : :
348 : 212028 : ret = pg_truncate(path, 0);
349 : :
350 : : /* Log a warning here to avoid repetition in callers. */
351 [ + + - + ]: 212028 : if (ret < 0 && errno != ENOENT)
352 : : {
1740 tmunro@postgresql.or 353 :UBC 0 : save_errno = errno;
354 [ # # ]: 0 : ereport(WARNING,
355 : : (errcode_for_file_access(),
356 : : errmsg("could not truncate file \"%s\": %m", path)));
357 : 0 : errno = save_errno;
358 : : }
359 : :
1740 tmunro@postgresql.or 360 :CBC 212028 : return ret;
361 : : }
362 : :
363 : : static void
1083 pg@bowt.ie 364 : 180784 : mdunlinkfork(RelFileLocatorBackend rlocator, ForkNumber forknum, bool isRedo)
365 : : {
366 : : RelPathStr path;
367 : : int ret;
368 : : int save_errno;
369 : :
370 : 180784 : path = relpath(rlocator, forknum);
371 : :
372 : : /*
373 : : * Truncate and then unlink the first segment, or just register a request
374 : : * to unlink it later, as described in the comments for mdunlink().
375 : : */
1032 tgl@sss.pgh.pa.us 376 [ + + + + : 180784 : if (isRedo || IsBinaryUpgrade || forknum != MAIN_FORKNUM ||
+ + ]
377 [ + + ]: 37625 : RelFileLocatorBackendIsTemp(rlocator))
378 : : {
1158 rhaas@postgresql.org 379 [ + + ]: 146321 : if (!RelFileLocatorBackendIsTemp(rlocator))
380 : : {
381 : : /* Prevent other backends' fds from holding on to the disk space */
193 andres@anarazel.de 382 : 133673 : ret = do_truncate(path.str);
383 : :
384 : : /* Forget any pending sync requests for the first segment */
1034 tgl@sss.pgh.pa.us 385 : 133673 : save_errno = errno;
1083 pg@bowt.ie 386 : 133673 : register_forget_request(rlocator, forknum, 0 /* first seg */ );
1034 tgl@sss.pgh.pa.us 387 : 133673 : errno = save_errno;
388 : : }
389 : : else
1740 tmunro@postgresql.or 390 : 12648 : ret = 0;
391 : :
392 : : /* Next unlink the file, unless it was already found to be missing */
1032 tgl@sss.pgh.pa.us 393 [ + + - + ]: 146321 : if (ret >= 0 || errno != ENOENT)
394 : : {
193 andres@anarazel.de 395 : 22081 : ret = unlink(path.str);
1740 tmunro@postgresql.or 396 [ + + - + ]: 22081 : if (ret < 0 && errno != ENOENT)
397 : : {
1032 tgl@sss.pgh.pa.us 398 :UBC 0 : save_errno = errno;
1740 tmunro@postgresql.or 399 [ # # ]: 0 : ereport(WARNING,
400 : : (errcode_for_file_access(),
401 : : errmsg("could not remove file \"%s\": %m", path.str)));
1032 tgl@sss.pgh.pa.us 402 : 0 : errno = save_errno;
403 : : }
404 : : }
405 : : }
406 : : else
407 : : {
408 : : /* Prevent other backends' fds from holding on to the disk space */
193 andres@anarazel.de 409 :CBC 34463 : ret = do_truncate(path.str);
410 : :
411 : : /* Register request to unlink first segment later */
1032 tgl@sss.pgh.pa.us 412 : 34463 : save_errno = errno;
413 : 34463 : register_unlink_segment(rlocator, forknum, 0 /* first seg */ );
414 : 34463 : errno = save_errno;
415 : : }
416 : :
417 : : /*
418 : : * Delete any additional segments.
419 : : *
420 : : * Note that because we loop until getting ENOENT, we will correctly
421 : : * remove all inactive segments as well as active ones. Ideally we'd
422 : : * continue the loop until getting exactly that errno, but that risks an
423 : : * infinite loop if the problem is directory-wide (for instance, if we
424 : : * suddenly can't read the data directory itself). We compromise by
425 : : * continuing after a non-ENOENT truncate error, but stopping after any
426 : : * unlink error. If there is indeed a directory-wide problem, additional
427 : : * unlink attempts wouldn't work anyway.
428 : : */
429 [ + + - + ]: 180784 : if (ret >= 0 || errno != ENOENT)
430 : : {
431 : : MdPathStr segpath;
432 : : BlockNumber segno;
433 : :
434 : 47230 : for (segno = 1;; segno++)
435 : : {
193 andres@anarazel.de 436 : 47230 : sprintf(segpath.str, "%s.%u", path.str, segno);
437 : :
1158 rhaas@postgresql.org 438 [ + + ]: 47230 : if (!RelFileLocatorBackendIsTemp(rlocator))
439 : : {
440 : : /*
441 : : * Prevent other backends' fds from holding on to the disk
442 : : * space. We're done if we see ENOENT, though.
443 : : */
193 andres@anarazel.de 444 [ + - + - ]: 43892 : if (do_truncate(segpath.str) < 0 && errno == ENOENT)
1740 tmunro@postgresql.or 445 : 43892 : break;
446 : :
447 : : /*
448 : : * Forget any pending sync requests for this segment before we
449 : : * try to unlink.
450 : : */
1083 pg@bowt.ie 451 :UBC 0 : register_forget_request(rlocator, forknum, segno);
452 : : }
453 : :
193 andres@anarazel.de 454 [ + - ]:CBC 3338 : if (unlink(segpath.str) < 0)
455 : : {
456 : : /* ENOENT is expected after the last segment... */
9068 tgl@sss.pgh.pa.us 457 [ - + ]: 3338 : if (errno != ENOENT)
6821 tgl@sss.pgh.pa.us 458 [ # # ]:UBC 0 : ereport(WARNING,
459 : : (errcode_for_file_access(),
460 : : errmsg("could not remove file \"%s\": %m", segpath.str)));
9068 tgl@sss.pgh.pa.us 461 :CBC 3338 : break;
462 : : }
463 : : }
464 : : }
10651 scrappy@hub.org 465 : 180784 : }
466 : :
467 : : /*
468 : : * mdextend() -- Add a block to the specified relation.
469 : : *
470 : : * The semantics are nearly the same as mdwrite(): write at the
471 : : * specified position. However, this is to be used for the case of
472 : : * extending a relation (i.e., blocknum is at or beyond the current
473 : : * EOF). Note that we assume writing a block beyond current EOF
474 : : * causes intervening file space to become filled with zeroes.
475 : : */
476 : : void
6235 heikki.linnakangas@i 477 : 117525 : mdextend(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
478 : : const void *buffer, bool skipFsync)
479 : : {
480 : : off_t seekpos;
481 : : int nbytes;
482 : : MdfdVec *v;
483 : :
484 : : /* If this build supports direct I/O, the buffer must be I/O aligned. */
485 : : if (PG_O_DIRECT != 0 && PG_IO_ALIGN_SIZE <= BLCKSZ)
882 tmunro@postgresql.or 486 [ - + ]: 117525 : Assert((uintptr_t) buffer == TYPEALIGN(PG_IO_ALIGN_SIZE, buffer));
487 : :
488 : : /* This assert is too expensive to have on normally ... */
489 : : #ifdef CHECK_WRITE_VS_EXTEND
490 : : Assert(blocknum >= mdnblocks(reln, forknum));
491 : : #endif
492 : :
493 : : /*
494 : : * If a relation manages to grow to 2^32-1 blocks, refuse to extend it any
495 : : * more --- we mustn't create a block whose number actually is
496 : : * InvalidBlockNumber. (Note that this failure should be unreachable
497 : : * because of upstream checks in bufmgr.c.)
498 : : */
6821 tgl@sss.pgh.pa.us 499 [ - + ]: 117525 : if (blocknum == InvalidBlockNumber)
6821 tgl@sss.pgh.pa.us 500 [ # # ]:UBC 0 : ereport(ERROR,
501 : : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
502 : : errmsg("cannot extend file \"%s\" beyond %u blocks",
503 : : relpath(reln->smgr_rlocator, forknum).str,
504 : : InvalidBlockNumber)));
505 : :
5503 rhaas@postgresql.org 506 :CBC 117525 : v = _mdfd_getseg(reln, forknum, blocknum, skipFsync, EXTENSION_CREATE);
507 : :
2999 tgl@sss.pgh.pa.us 508 : 117525 : seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
509 : :
6389 510 [ - + ]: 117525 : Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
511 : :
2495 tmunro@postgresql.or 512 [ - + ]: 117525 : if ((nbytes = FileWrite(v->mdfd_vfd, buffer, BLCKSZ, seekpos, WAIT_EVENT_DATA_FILE_EXTEND)) != BLCKSZ)
513 : : {
6821 tgl@sss.pgh.pa.us 514 [ # # ]:UBC 0 : if (nbytes < 0)
515 [ # # ]: 0 : ereport(ERROR,
516 : : (errcode_for_file_access(),
517 : : errmsg("could not extend file \"%s\": %m",
518 : : FilePathName(v->mdfd_vfd)),
519 : : errhint("Check free disk space.")));
520 : : /* short write: complain appropriately */
521 [ # # ]: 0 : ereport(ERROR,
522 : : (errcode(ERRCODE_DISK_FULL),
523 : : errmsg("could not extend file \"%s\": wrote only %d of %d bytes at block %u",
524 : : FilePathName(v->mdfd_vfd),
525 : : nbytes, BLCKSZ, blocknum),
526 : : errhint("Check free disk space.")));
527 : : }
528 : :
5503 rhaas@postgresql.org 529 [ + + + - ]:CBC 117525 : if (!skipFsync && !SmgrIsTemp(reln))
6235 heikki.linnakangas@i 530 : 30 : register_dirty_segment(reln, forknum, v);
531 : :
532 [ - + ]: 117525 : Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
10651 scrappy@hub.org 533 : 117525 : }
534 : :
535 : : /*
536 : : * mdzeroextend() -- Add new zeroed out blocks to the specified relation.
537 : : *
538 : : * Similar to mdextend(), except the relation can be extended by multiple
539 : : * blocks at once and the added blocks will be filled with zeroes.
540 : : */
541 : : void
885 andres@anarazel.de 542 : 207816 : mdzeroextend(SMgrRelation reln, ForkNumber forknum,
543 : : BlockNumber blocknum, int nblocks, bool skipFsync)
544 : : {
545 : : MdfdVec *v;
546 : 207816 : BlockNumber curblocknum = blocknum;
547 : 207816 : int remblocks = nblocks;
548 : :
549 [ - + ]: 207816 : Assert(nblocks > 0);
550 : :
551 : : /* This assert is too expensive to have on normally ... */
552 : : #ifdef CHECK_WRITE_VS_EXTEND
553 : : Assert(blocknum >= mdnblocks(reln, forknum));
554 : : #endif
555 : :
556 : : /*
557 : : * If a relation manages to grow to 2^32-1 blocks, refuse to extend it any
558 : : * more --- we mustn't create a block whose number actually is
559 : : * InvalidBlockNumber or larger.
560 : : */
561 [ - + ]: 207816 : if ((uint64) blocknum + nblocks >= (uint64) InvalidBlockNumber)
885 andres@anarazel.de 562 [ # # ]:UBC 0 : ereport(ERROR,
563 : : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
564 : : errmsg("cannot extend file \"%s\" beyond %u blocks",
565 : : relpath(reln->smgr_rlocator, forknum).str,
566 : : InvalidBlockNumber)));
567 : :
885 andres@anarazel.de 568 [ + + ]:CBC 415632 : while (remblocks > 0)
569 : : {
841 tgl@sss.pgh.pa.us 570 : 207816 : BlockNumber segstartblock = curblocknum % ((BlockNumber) RELSEG_SIZE);
885 andres@anarazel.de 571 : 207816 : off_t seekpos = (off_t) BLCKSZ * segstartblock;
572 : : int numblocks;
573 : :
574 [ - + ]: 207816 : if (segstartblock + remblocks > RELSEG_SIZE)
885 andres@anarazel.de 575 :UBC 0 : numblocks = RELSEG_SIZE - segstartblock;
576 : : else
885 andres@anarazel.de 577 :CBC 207816 : numblocks = remblocks;
578 : :
579 : 207816 : v = _mdfd_getseg(reln, forknum, curblocknum, skipFsync, EXTENSION_CREATE);
580 : :
581 [ - + ]: 207816 : Assert(segstartblock < RELSEG_SIZE);
582 [ - + ]: 207816 : Assert(segstartblock + numblocks <= RELSEG_SIZE);
583 : :
584 : : /*
585 : : * If available and useful, use posix_fallocate() (via
586 : : * FileFallocate()) to extend the relation. That's often more
587 : : * efficient than using write(), as it commonly won't cause the kernel
588 : : * to allocate page cache space for the extended pages.
589 : : *
590 : : * However, we don't use FileFallocate() for small extensions, as it
591 : : * defeats delayed allocation on some filesystems. Not clear where
592 : : * that decision should be made though? For now just use a cutoff of
593 : : * 8, anything between 4 and 8 worked OK in some local testing.
594 : : */
595 [ + + ]: 207816 : if (numblocks > 8)
596 : : {
597 : : int ret;
598 : :
599 : 526 : ret = FileFallocate(v->mdfd_vfd,
600 : : seekpos, (off_t) BLCKSZ * numblocks,
601 : : WAIT_EVENT_DATA_FILE_EXTEND);
602 [ - + ]: 526 : if (ret != 0)
603 : : {
885 andres@anarazel.de 604 [ # # ]:UBC 0 : ereport(ERROR,
605 : : errcode_for_file_access(),
606 : : errmsg("could not extend file \"%s\" with FileFallocate(): %m",
607 : : FilePathName(v->mdfd_vfd)),
608 : : errhint("Check free disk space."));
609 : : }
610 : : }
611 : : else
612 : : {
613 : : int ret;
614 : :
615 : : /*
616 : : * Even if we don't want to use fallocate, we can still extend a
617 : : * bit more efficiently than writing each 8kB block individually.
618 : : * pg_pwrite_zeros() (via FileZero()) uses pg_pwritev_with_retry()
619 : : * to avoid multiple writes or needing a zeroed buffer for the
620 : : * whole length of the extension.
621 : : */
885 andres@anarazel.de 622 :CBC 207290 : ret = FileZero(v->mdfd_vfd,
623 : : seekpos, (off_t) BLCKSZ * numblocks,
624 : : WAIT_EVENT_DATA_FILE_EXTEND);
625 [ - + ]: 207290 : if (ret < 0)
885 andres@anarazel.de 626 [ # # ]:UBC 0 : ereport(ERROR,
627 : : errcode_for_file_access(),
628 : : errmsg("could not extend file \"%s\": %m",
629 : : FilePathName(v->mdfd_vfd)),
630 : : errhint("Check free disk space."));
631 : : }
632 : :
885 andres@anarazel.de 633 [ + - + + ]:CBC 207816 : if (!skipFsync && !SmgrIsTemp(reln))
634 : 195368 : register_dirty_segment(reln, forknum, v);
635 : :
636 [ - + ]: 207816 : Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
637 : :
638 : 207816 : remblocks -= numblocks;
639 : 207816 : curblocknum += numblocks;
640 : : }
641 : 207816 : }
642 : :
643 : : /*
644 : : * mdopenfork() -- Open one fork of the specified relation.
645 : : *
646 : : * Note we only open the first segment, when there are multiple segments.
647 : : *
648 : : * If first segment is not present, either ereport or return NULL according
649 : : * to "behavior". We treat EXTENSION_CREATE the same as EXTENSION_FAIL;
650 : : * EXTENSION_CREATE means it's OK to extend an existing relation, not to
651 : : * invent one out of whole cloth.
652 : : */
653 : : static MdfdVec *
2243 tmunro@postgresql.or 654 : 3295318 : mdopenfork(SMgrRelation reln, ForkNumber forknum, int behavior)
655 : : {
656 : : MdfdVec *mdfd;
657 : : RelPathStr path;
658 : : File fd;
659 : :
660 : : /* No work if already open */
3285 andres@anarazel.de 661 [ + + ]: 3295318 : if (reln->md_num_open_segs[forknum] > 0)
662 : 2231027 : return &reln->md_seg_fds[forknum][0];
663 : :
1158 rhaas@postgresql.org 664 : 1064291 : path = relpath(reln->smgr_rlocator, forknum);
665 : :
193 andres@anarazel.de 666 : 1064291 : fd = PathNameOpenFile(path.str, _mdfd_open_flags());
667 : :
10226 bruce@momjian.us 668 [ + + ]: 1064291 : if (fd < 0)
669 : : {
2413 akapila@postgresql.o 670 [ + + ]: 351454 : if ((behavior & EXTENSION_RETURN_NULL) &&
671 [ + - ]: 351432 : FILE_POSSIBLY_DELETED(errno))
672 : 351432 : return NULL;
673 [ + - ]: 22 : ereport(ERROR,
674 : : (errcode_for_file_access(),
675 : : errmsg("could not open file \"%s\": %m", path.str)));
676 : : }
677 : :
3285 andres@anarazel.de 678 : 712837 : _fdvec_resize(reln, forknum, 1);
679 : 712837 : mdfd = &reln->md_seg_fds[forknum][0];
7768 tgl@sss.pgh.pa.us 680 : 712837 : mdfd->mdfd_vfd = fd;
681 : 712837 : mdfd->mdfd_segno = 0;
682 : :
6235 heikki.linnakangas@i 683 [ - + ]: 712837 : Assert(_mdnblocks(reln, forknum, mdfd) <= ((BlockNumber) RELSEG_SIZE));
684 : :
7768 tgl@sss.pgh.pa.us 685 : 712837 : return mdfd;
686 : : }
687 : :
688 : : /*
689 : : * mdopen() -- Initialize newly-opened relation.
690 : : */
691 : : void
2243 tmunro@postgresql.or 692 : 947375 : mdopen(SMgrRelation reln)
693 : : {
694 : : /* mark it not open */
695 [ + + ]: 4736875 : for (int forknum = 0; forknum <= MAX_FORKNUM; forknum++)
696 : 3789500 : reln->md_num_open_segs[forknum] = 0;
697 : 947375 : }
698 : :
699 : : /*
700 : : * mdclose() -- Close the specified relation, if it isn't closed already.
701 : : */
702 : : void
6235 heikki.linnakangas@i 703 : 3397071 : mdclose(SMgrRelation reln, ForkNumber forknum)
704 : : {
3285 andres@anarazel.de 705 : 3397071 : int nopensegs = reln->md_num_open_segs[forknum];
706 : :
707 : : /* No work if already closed */
708 [ + + ]: 3397071 : if (nopensegs == 0)
6821 tgl@sss.pgh.pa.us 709 : 2887229 : return;
710 : :
711 : : /* close segments starting from the end */
3285 andres@anarazel.de 712 [ + + ]: 1019684 : while (nopensegs > 0)
713 : : {
714 : 509842 : MdfdVec *v = &reln->md_seg_fds[forknum][nopensegs - 1];
715 : :
2066 noah@leadboat.com 716 : 509842 : FileClose(v->mdfd_vfd);
717 : 509842 : _fdvec_resize(reln, forknum, nopensegs - 1);
3285 andres@anarazel.de 718 : 509842 : nopensegs--;
719 : : }
720 : : }
721 : :
722 : : /*
723 : : * mdprefetch() -- Initiate asynchronous read of the specified blocks of a relation
724 : : */
725 : : bool
630 tmunro@postgresql.or 726 : 8006 : mdprefetch(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
727 : : int nblocks)
728 : : {
729 : : #ifdef USE_PREFETCH
730 : :
882 731 [ - + ]: 8006 : Assert((io_direct_flags & IO_DIRECT_DATA) == 0);
732 : :
630 733 [ - + ]: 8006 : if ((uint64) blocknum + nblocks > (uint64) MaxBlockNumber + 1)
1977 tmunro@postgresql.or 734 :UBC 0 : return false;
735 : :
630 tmunro@postgresql.or 736 [ + + ]:CBC 16012 : while (nblocks > 0)
737 : : {
738 : : off_t seekpos;
739 : : MdfdVec *v;
740 : : int nblocks_this_segment;
741 : :
742 : 8006 : v = _mdfd_getseg(reln, forknum, blocknum, false,
743 [ + + ]: 8006 : InRecovery ? EXTENSION_RETURN_NULL : EXTENSION_FAIL);
744 [ - + ]: 8006 : if (v == NULL)
630 tmunro@postgresql.or 745 :UBC 0 : return false;
746 : :
630 tmunro@postgresql.or 747 :CBC 8006 : seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
748 : :
749 [ - + ]: 8006 : Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
750 : :
751 : 8006 : nblocks_this_segment =
752 : 8006 : Min(nblocks,
753 : : RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
754 : :
755 : 8006 : (void) FilePrefetch(v->mdfd_vfd, seekpos, BLCKSZ * nblocks_this_segment,
756 : : WAIT_EVENT_DATA_FILE_PREFETCH);
757 : :
758 : 8006 : blocknum += nblocks_this_segment;
759 : 8006 : nblocks -= nblocks_this_segment;
760 : : }
761 : : #endif /* USE_PREFETCH */
762 : :
1977 763 : 8006 : return true;
764 : : }
765 : :
766 : : /*
767 : : * Convert an array of buffer address into an array of iovec objects, and
768 : : * return the number that were required. 'iov' must have enough space for up
769 : : * to 'nblocks' elements, but the number used may be less depending on
770 : : * merging. In the case of a run of fully contiguous buffers, a single iovec
771 : : * will be populated that can be handled as a plain non-vectored I/O.
772 : : */
773 : : static int
628 774 : 1803440 : buffers_to_iovec(struct iovec *iov, void **buffers, int nblocks)
775 : : {
776 : : struct iovec *iovp;
777 : : int iovcnt;
778 : :
779 [ - + ]: 1803440 : Assert(nblocks >= 1);
780 : :
781 : : /* If this build supports direct I/O, buffers must be I/O aligned. */
782 [ + + ]: 3766814 : for (int i = 0; i < nblocks; ++i)
783 : : {
784 : : if (PG_O_DIRECT != 0 && PG_IO_ALIGN_SIZE <= BLCKSZ)
785 [ - + ]: 1963374 : Assert((uintptr_t) buffers[i] ==
786 : : TYPEALIGN(PG_IO_ALIGN_SIZE, buffers[i]));
787 : : }
788 : :
789 : : /* Start the first iovec off with the first buffer. */
790 : 1803440 : iovp = &iov[0];
791 : 1803440 : iovp->iov_base = buffers[0];
792 : 1803440 : iovp->iov_len = BLCKSZ;
793 : 1803440 : iovcnt = 1;
794 : :
795 : : /* Try to merge the rest. */
796 [ + + ]: 1963374 : for (int i = 1; i < nblocks; ++i)
797 : : {
798 : 159934 : void *buffer = buffers[i];
799 : :
800 [ + + ]: 159934 : if (((char *) iovp->iov_base + iovp->iov_len) == buffer)
801 : : {
802 : : /* Contiguous with the last iovec. */
803 : 158698 : iovp->iov_len += BLCKSZ;
804 : : }
805 : : else
806 : : {
807 : : /* Need a new iovec. */
808 : 1236 : iovp++;
809 : 1236 : iovp->iov_base = buffer;
810 : 1236 : iovp->iov_len = BLCKSZ;
811 : 1236 : iovcnt++;
812 : : }
813 : : }
814 : :
815 : 1803440 : return iovcnt;
816 : : }
817 : :
818 : : /*
819 : : * mdmaxcombine() -- Return the maximum number of total blocks that can be
820 : : * combined with an IO starting at blocknum.
821 : : */
822 : : uint32
333 andres@anarazel.de 823 : 31935 : mdmaxcombine(SMgrRelation reln, ForkNumber forknum,
824 : : BlockNumber blocknum)
825 : : {
826 : : BlockNumber segoff;
827 : :
828 : 31935 : segoff = blocknum % ((BlockNumber) RELSEG_SIZE);
829 : :
830 : 31935 : return RELSEG_SIZE - segoff;
831 : : }
832 : :
833 : : /*
834 : : * mdreadv() -- Read the specified blocks from a relation.
835 : : */
836 : : void
628 tmunro@postgresql.or 837 : 598 : mdreadv(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
838 : : void **buffers, BlockNumber nblocks)
839 : : {
840 [ + + ]: 1196 : while (nblocks > 0)
841 : : {
842 : : struct iovec iov[PG_IOV_MAX];
843 : : int iovcnt;
844 : : off_t seekpos;
845 : : int nbytes;
846 : : MdfdVec *v;
847 : : BlockNumber nblocks_this_segment;
848 : : size_t transferred_this_segment;
849 : : size_t size_this_segment;
850 : :
851 : 598 : v = _mdfd_getseg(reln, forknum, blocknum, false,
852 : : EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
853 : :
854 : 598 : seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
855 : :
856 [ - + ]: 598 : Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
857 : :
858 : 598 : nblocks_this_segment =
859 : 598 : Min(nblocks,
860 : : RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
861 : 598 : nblocks_this_segment = Min(nblocks_this_segment, lengthof(iov));
862 : :
333 andres@anarazel.de 863 [ - + ]: 598 : if (nblocks_this_segment != nblocks)
333 andres@anarazel.de 864 [ # # ]:UBC 0 : elog(ERROR, "read crosses segment boundary");
865 : :
628 tmunro@postgresql.or 866 :CBC 598 : iovcnt = buffers_to_iovec(iov, buffers, nblocks_this_segment);
867 : 598 : size_this_segment = nblocks_this_segment * BLCKSZ;
868 : 598 : transferred_this_segment = 0;
869 : :
870 : : /*
871 : : * Inner loop to continue after a short read. We'll keep going until
872 : : * we hit EOF rather than assuming that a short read means we hit the
873 : : * end.
874 : : */
875 : : for (;;)
876 : : {
628 tmunro@postgresql.or 877 :UBC 0 : TRACE_POSTGRESQL_SMGR_MD_READ_START(forknum, blocknum,
878 : : reln->smgr_rlocator.locator.spcOid,
879 : : reln->smgr_rlocator.locator.dbOid,
880 : : reln->smgr_rlocator.locator.relNumber,
881 : : reln->smgr_rlocator.backend);
628 tmunro@postgresql.or 882 :CBC 598 : nbytes = FileReadV(v->mdfd_vfd, iov, iovcnt, seekpos,
883 : : WAIT_EVENT_DATA_FILE_READ);
884 : : TRACE_POSTGRESQL_SMGR_MD_READ_DONE(forknum, blocknum,
885 : : reln->smgr_rlocator.locator.spcOid,
886 : : reln->smgr_rlocator.locator.dbOid,
887 : : reln->smgr_rlocator.locator.relNumber,
888 : : reln->smgr_rlocator.backend,
889 : : nbytes,
890 : : size_this_segment - transferred_this_segment);
891 : :
892 : : #ifdef SIMULATE_SHORT_READ
893 : : nbytes = Min(nbytes, 4096);
894 : : #endif
895 : :
896 [ - + ]: 598 : if (nbytes < 0)
628 tmunro@postgresql.or 897 [ # # ]:UBC 0 : ereport(ERROR,
898 : : (errcode_for_file_access(),
899 : : errmsg("could not read blocks %u..%u in file \"%s\": %m",
900 : : blocknum,
901 : : blocknum + nblocks_this_segment - 1,
902 : : FilePathName(v->mdfd_vfd))));
903 : :
628 tmunro@postgresql.or 904 [ - + ]:CBC 598 : if (nbytes == 0)
905 : : {
906 : : /*
907 : : * We are at or past EOF, or we read a partial block at EOF.
908 : : * Normally this is an error; upper levels should never try to
909 : : * read a nonexistent block. However, if zero_damaged_pages
910 : : * is ON or we are InRecovery, we should instead return zeroes
911 : : * without complaining. This allows, for example, the case of
912 : : * trying to update a block that was later truncated away.
913 : : *
914 : : * NB: We think that this codepath is unreachable in recovery
915 : : * and incomplete with zero_damaged_pages, as missing segments
916 : : * are not created. Putting blocks into the buffer-pool that
917 : : * do not exist on disk is rather problematic, as it will not
918 : : * be found by scans that rely on smgrnblocks(), as they are
919 : : * beyond EOF. It also can cause weird problems with relation
920 : : * extension, as relation extension does not expect blocks
921 : : * beyond EOF to exist.
922 : : *
923 : : * Therefore we do not want to copy the logic into
924 : : * mdstartreadv(), where it would have to be more complicated
925 : : * due to potential differences in the zero_damaged_pages
926 : : * setting between the definer and completor of IO.
927 : : *
928 : : * For PG 18, we are putting an Assert(false) in mdreadv()
929 : : * (triggering failures in assertion-enabled builds, but
930 : : * continuing to work in production builds). Afterwards we
931 : : * plan to remove this code entirely.
932 : : */
628 tmunro@postgresql.or 933 [ # # # # ]:UBC 0 : if (zero_damaged_pages || InRecovery)
934 : : {
158 andres@anarazel.de 935 : 0 : Assert(false); /* see comment above */
936 : :
937 : : for (BlockNumber i = transferred_this_segment / BLCKSZ;
938 : : i < nblocks_this_segment;
939 : : ++i)
940 : : memset(buffers[i], 0, BLCKSZ);
941 : : break;
942 : : }
943 : : else
628 tmunro@postgresql.or 944 [ # # ]: 0 : ereport(ERROR,
945 : : (errcode(ERRCODE_DATA_CORRUPTED),
946 : : errmsg("could not read blocks %u..%u in file \"%s\": read only %zu of %zu bytes",
947 : : blocknum,
948 : : blocknum + nblocks_this_segment - 1,
949 : : FilePathName(v->mdfd_vfd),
950 : : transferred_this_segment,
951 : : size_this_segment)));
952 : : }
953 : :
954 : : /* One loop should usually be enough. */
628 tmunro@postgresql.or 955 :CBC 598 : transferred_this_segment += nbytes;
956 [ - + ]: 598 : Assert(transferred_this_segment <= size_this_segment);
957 [ + - ]: 598 : if (transferred_this_segment == size_this_segment)
958 : 598 : break;
959 : :
960 : : /* Adjust position and vectors after a short read. */
628 tmunro@postgresql.or 961 :UBC 0 : seekpos += nbytes;
962 : 0 : iovcnt = compute_remaining_iovec(iov, iov, iovcnt, nbytes);
963 : : }
964 : :
628 tmunro@postgresql.or 965 :CBC 598 : nblocks -= nblocks_this_segment;
966 : 598 : buffers += nblocks_this_segment;
967 : 598 : blocknum += nblocks_this_segment;
968 : : }
10651 scrappy@hub.org 969 : 598 : }
970 : :
971 : : /*
972 : : * mdstartreadv() -- Asynchronous version of mdreadv().
973 : : */
974 : : void
161 andres@anarazel.de 975 : 1245727 : mdstartreadv(PgAioHandle *ioh,
976 : : SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
977 : : void **buffers, BlockNumber nblocks)
978 : : {
979 : : off_t seekpos;
980 : : MdfdVec *v;
981 : : BlockNumber nblocks_this_segment;
982 : : struct iovec *iov;
983 : : int iovcnt;
984 : : int ret;
985 : :
986 : 1245727 : v = _mdfd_getseg(reln, forknum, blocknum, false,
987 : : EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
988 : :
989 : 1245712 : seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
990 : :
991 [ - + ]: 1245712 : Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
992 : :
993 : 1245712 : nblocks_this_segment =
994 : 1245712 : Min(nblocks,
995 : : RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
996 : :
997 [ - + ]: 1245712 : if (nblocks_this_segment != nblocks)
161 andres@anarazel.de 998 [ # # ]:UBC 0 : elog(ERROR, "read crossing segment boundary");
999 : :
161 andres@anarazel.de 1000 :CBC 1245712 : iovcnt = pgaio_io_get_iovec(ioh, &iov);
1001 : :
1002 [ - + ]: 1245712 : Assert(nblocks <= iovcnt);
1003 : :
1004 : 1245712 : iovcnt = buffers_to_iovec(iov, buffers, nblocks_this_segment);
1005 : :
1006 [ - + ]: 1245712 : Assert(iovcnt <= nblocks_this_segment);
1007 : :
1008 [ + + ]: 1245712 : if (!(io_direct_flags & IO_DIRECT_DATA))
1009 : 1244305 : pgaio_io_set_flag(ioh, PGAIO_HF_BUFFERED);
1010 : :
1011 : 1245712 : pgaio_io_set_target_smgr(ioh,
1012 : : reln,
1013 : : forknum,
1014 : : blocknum,
1015 : : nblocks,
1016 : : false);
1017 : 1245712 : pgaio_io_register_callbacks(ioh, PGAIO_HCB_MD_READV, 0);
1018 : :
1019 : 1245712 : ret = FileStartReadV(ioh, v->mdfd_vfd, iovcnt, seekpos, WAIT_EVENT_DATA_FILE_READ);
1020 [ - + ]: 1245712 : if (ret != 0)
161 andres@anarazel.de 1021 [ # # ]:UBC 0 : ereport(ERROR,
1022 : : (errcode_for_file_access(),
1023 : : errmsg("could not start reading blocks %u..%u in file \"%s\": %m",
1024 : : blocknum,
1025 : : blocknum + nblocks_this_segment - 1,
1026 : : FilePathName(v->mdfd_vfd))));
1027 : :
1028 : : /*
1029 : : * The error checks corresponding to the post-read checks in mdreadv() are
1030 : : * in md_readv_complete().
1031 : : *
1032 : : * However we chose, at least for now, to not implement the
1033 : : * zero_damaged_pages logic present in mdreadv(). As outlined in mdreadv()
1034 : : * that logic is rather problematic, and we want to get rid of it. Here
1035 : : * equivalent logic would have to be more complicated due to potential
1036 : : * differences in the zero_damaged_pages setting between the definer and
1037 : : * completor of IO.
1038 : : */
161 andres@anarazel.de 1039 :CBC 1245712 : }
1040 : :
1041 : : /*
1042 : : * mdwritev() -- Write the supplied blocks at the appropriate location.
1043 : : *
1044 : : * This is to be used only for updating already-existing blocks of a
1045 : : * relation (ie, those before the current EOF). To extend a relation,
1046 : : * use mdextend().
1047 : : */
1048 : : void
628 tmunro@postgresql.or 1049 : 557130 : mdwritev(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
1050 : : const void **buffers, BlockNumber nblocks, bool skipFsync)
1051 : : {
1052 : : /* This assert is too expensive to have on normally ... */
1053 : : #ifdef CHECK_WRITE_VS_EXTEND
1054 : : Assert((uint64) blocknum + (uint64) nblocks <= (uint64) mdnblocks(reln, forknum));
1055 : : #endif
1056 : :
1057 [ + + ]: 1114260 : while (nblocks > 0)
1058 : : {
1059 : : struct iovec iov[PG_IOV_MAX];
1060 : : int iovcnt;
1061 : : off_t seekpos;
1062 : : int nbytes;
1063 : : MdfdVec *v;
1064 : : BlockNumber nblocks_this_segment;
1065 : : size_t transferred_this_segment;
1066 : : size_t size_this_segment;
1067 : :
1068 : 557130 : v = _mdfd_getseg(reln, forknum, blocknum, skipFsync,
1069 : : EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
1070 : :
1071 : 557130 : seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
1072 : :
1073 [ - + ]: 557130 : Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
1074 : :
1075 : 557130 : nblocks_this_segment =
1076 : 557130 : Min(nblocks,
1077 : : RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
1078 : 557130 : nblocks_this_segment = Min(nblocks_this_segment, lengthof(iov));
1079 : :
333 andres@anarazel.de 1080 [ - + ]: 557130 : if (nblocks_this_segment != nblocks)
333 andres@anarazel.de 1081 [ # # ]:UBC 0 : elog(ERROR, "write crosses segment boundary");
1082 : :
628 tmunro@postgresql.or 1083 :CBC 557130 : iovcnt = buffers_to_iovec(iov, (void **) buffers, nblocks_this_segment);
1084 : 557130 : size_this_segment = nblocks_this_segment * BLCKSZ;
1085 : 557130 : transferred_this_segment = 0;
1086 : :
1087 : : /*
1088 : : * Inner loop to continue after a short write. If the reason is that
1089 : : * we're out of disk space, a future attempt should get an ENOSPC
1090 : : * error from the kernel.
1091 : : */
1092 : : for (;;)
1093 : : {
628 tmunro@postgresql.or 1094 :UBC 0 : TRACE_POSTGRESQL_SMGR_MD_WRITE_START(forknum, blocknum,
1095 : : reln->smgr_rlocator.locator.spcOid,
1096 : : reln->smgr_rlocator.locator.dbOid,
1097 : : reln->smgr_rlocator.locator.relNumber,
1098 : : reln->smgr_rlocator.backend);
628 tmunro@postgresql.or 1099 :CBC 557130 : nbytes = FileWriteV(v->mdfd_vfd, iov, iovcnt, seekpos,
1100 : : WAIT_EVENT_DATA_FILE_WRITE);
1101 : : TRACE_POSTGRESQL_SMGR_MD_WRITE_DONE(forknum, blocknum,
1102 : : reln->smgr_rlocator.locator.spcOid,
1103 : : reln->smgr_rlocator.locator.dbOid,
1104 : : reln->smgr_rlocator.locator.relNumber,
1105 : : reln->smgr_rlocator.backend,
1106 : : nbytes,
1107 : : size_this_segment - transferred_this_segment);
1108 : :
1109 : : #ifdef SIMULATE_SHORT_WRITE
1110 : : nbytes = Min(nbytes, 4096);
1111 : : #endif
1112 : :
1113 [ - + ]: 557130 : if (nbytes < 0)
1114 : : {
628 tmunro@postgresql.or 1115 :UBC 0 : bool enospc = errno == ENOSPC;
1116 : :
1117 [ # # # # ]: 0 : ereport(ERROR,
1118 : : (errcode_for_file_access(),
1119 : : errmsg("could not write blocks %u..%u in file \"%s\": %m",
1120 : : blocknum,
1121 : : blocknum + nblocks_this_segment - 1,
1122 : : FilePathName(v->mdfd_vfd)),
1123 : : enospc ? errhint("Check free disk space.") : 0));
1124 : : }
1125 : :
1126 : : /* One loop should usually be enough. */
628 tmunro@postgresql.or 1127 :CBC 557130 : transferred_this_segment += nbytes;
1128 [ - + ]: 557130 : Assert(transferred_this_segment <= size_this_segment);
1129 [ + - ]: 557130 : if (transferred_this_segment == size_this_segment)
1130 : 557130 : break;
1131 : :
1132 : : /* Adjust position and iovecs after a short write. */
628 tmunro@postgresql.or 1133 :UBC 0 : seekpos += nbytes;
1134 : 0 : iovcnt = compute_remaining_iovec(iov, iov, iovcnt, nbytes);
1135 : : }
1136 : :
628 tmunro@postgresql.or 1137 [ + + + + ]:CBC 557130 : if (!skipFsync && !SmgrIsTemp(reln))
1138 : 552710 : register_dirty_segment(reln, forknum, v);
1139 : :
1140 : 557130 : nblocks -= nblocks_this_segment;
1141 : 557130 : buffers += nblocks_this_segment;
1142 : 557130 : blocknum += nblocks_this_segment;
1143 : : }
9281 tgl@sss.pgh.pa.us 1144 : 557130 : }
1145 : :
1146 : :
1147 : : /*
1148 : : * mdwriteback() -- Tell the kernel to write pages back to storage.
1149 : : *
1150 : : * This accepts a range of blocks because flushing several pages at once is
1151 : : * considerably more efficient than doing so individually.
1152 : : */
1153 : : void
841 peter@eisentraut.org 1154 :UBC 0 : mdwriteback(SMgrRelation reln, ForkNumber forknum,
1155 : : BlockNumber blocknum, BlockNumber nblocks)
1156 : : {
1157 [ # # ]: 0 : Assert((io_direct_flags & IO_DIRECT_DATA) == 0);
1158 : :
1159 : : /*
1160 : : * Issue flush requests in as few requests as possible; have to split at
1161 : : * segment boundaries though, since those are actually separate files.
1162 : : */
1163 [ # # ]: 0 : while (nblocks > 0)
1164 : : {
1165 : 0 : BlockNumber nflush = nblocks;
1166 : : off_t seekpos;
1167 : : MdfdVec *v;
1168 : : int segnum_start,
1169 : : segnum_end;
1170 : :
1171 : 0 : v = _mdfd_getseg(reln, forknum, blocknum, true /* not used */ ,
1172 : : EXTENSION_DONT_OPEN);
1173 : :
1174 : : /*
1175 : : * We might be flushing buffers of already removed relations, that's
1176 : : * ok, just ignore that case. If the segment file wasn't open already
1177 : : * (ie from a recent mdwrite()), then we don't want to re-open it, to
1178 : : * avoid a race with PROCSIGNAL_BARRIER_SMGRRELEASE that might leave
1179 : : * us with a descriptor to a file that is about to be unlinked.
1180 : : */
1181 [ # # ]: 0 : if (!v)
1182 : 0 : return;
1183 : :
1184 : : /* compute offset inside the current segment */
1185 : 0 : segnum_start = blocknum / RELSEG_SIZE;
1186 : :
1187 : : /* compute number of desired writes within the current segment */
1188 : 0 : segnum_end = (blocknum + nblocks - 1) / RELSEG_SIZE;
1189 [ # # ]: 0 : if (segnum_start != segnum_end)
1190 : 0 : nflush = RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE));
1191 : :
1192 [ # # ]: 0 : Assert(nflush >= 1);
1193 [ # # ]: 0 : Assert(nflush <= nblocks);
1194 : :
1195 : 0 : seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
1196 : :
1197 : 0 : FileWriteback(v->mdfd_vfd, seekpos, (off_t) BLCKSZ * nflush, WAIT_EVENT_DATA_FILE_FLUSH);
1198 : :
1199 : 0 : nblocks -= nflush;
1200 : 0 : blocknum += nflush;
1201 : : }
1202 : : }
1203 : :
1204 : : /*
1205 : : * mdnblocks() -- Get the number of blocks stored in a relation.
1206 : : *
1207 : : * Important side effect: all active segments of the relation are opened
1208 : : * and added to the md_seg_fds array. If this routine has not been
1209 : : * called, then only segments up to the last one actually touched
1210 : : * are present in the array.
1211 : : */
1212 : : BlockNumber
6235 heikki.linnakangas@i 1213 :CBC 2125614 : mdnblocks(SMgrRelation reln, ForkNumber forknum)
1214 : : {
1215 : : MdfdVec *v;
1216 : : BlockNumber nblocks;
1217 : : BlockNumber segno;
1218 : :
1829 bruce@momjian.us 1219 : 2125614 : mdopenfork(reln, forknum, EXTENSION_FAIL);
1220 : :
1221 : : /* mdopen has opened the first segment */
3285 andres@anarazel.de 1222 [ - + ]: 2125595 : Assert(reln->md_num_open_segs[forknum] > 0);
1223 : :
1224 : : /*
1225 : : * Start from the last open segments, to avoid redundant seeks. We have
1226 : : * previously verified that these segments are exactly RELSEG_SIZE long,
1227 : : * and it's useless to recheck that each time.
1228 : : *
1229 : : * NOTE: this assumption could only be wrong if another backend has
1230 : : * truncated the relation. We rely on higher code levels to handle that
1231 : : * scenario by closing and re-opening the md fd, which is handled via
1232 : : * relcache flush. (Since the checkpointer doesn't participate in
1233 : : * relcache flush, it could have segment entries for inactive segments;
1234 : : * that's OK because the checkpointer never needs to compute relation
1235 : : * size.)
1236 : : */
1237 : 2125595 : segno = reln->md_num_open_segs[forknum] - 1;
1238 : 2125595 : v = &reln->md_seg_fds[forknum][segno];
1239 : :
1240 : : for (;;)
1241 : : {
6235 heikki.linnakangas@i 1242 : 2125595 : nblocks = _mdnblocks(reln, forknum, v);
8837 tgl@sss.pgh.pa.us 1243 [ - + ]: 2125595 : if (nblocks > ((BlockNumber) RELSEG_SIZE))
8080 tgl@sss.pgh.pa.us 1244 [ # # ]:UBC 0 : elog(FATAL, "segment too big");
8837 tgl@sss.pgh.pa.us 1245 [ + - ]:CBC 2125595 : if (nblocks < ((BlockNumber) RELSEG_SIZE))
1246 : 2125595 : return (segno * ((BlockNumber) RELSEG_SIZE)) + nblocks;
1247 : :
1248 : : /*
1249 : : * If segment is exactly RELSEG_SIZE, advance to next one.
1250 : : */
8885 tgl@sss.pgh.pa.us 1251 :UBC 0 : segno++;
1252 : :
1253 : : /*
1254 : : * We used to pass O_CREAT here, but that has the disadvantage that it
1255 : : * might create a segment which has vanished through some operating
1256 : : * system misadventure. In such a case, creating the segment here
1257 : : * undermines _mdfd_getseg's attempts to notice and report an error
1258 : : * upon access to a missing segment.
1259 : : */
3285 andres@anarazel.de 1260 : 0 : v = _mdfd_openseg(reln, forknum, segno, 0);
1261 [ # # ]: 0 : if (v == NULL)
1262 : 0 : return segno * ((BlockNumber) RELSEG_SIZE);
1263 : : }
1264 : : }
1265 : :
1266 : : /*
1267 : : * mdtruncate() -- Truncate relation to specified number of blocks.
1268 : : *
1269 : : * Guaranteed not to allocate memory, so it can be used in a critical section.
1270 : : * Caller must have called smgrnblocks() to obtain curnblk while holding a
1271 : : * sufficient lock to prevent a change in relation size, and not used any smgr
1272 : : * functions for this relation or handled interrupts in between. This makes
1273 : : * sure we have opened all active segments, so that truncate loop will get
1274 : : * them all!
1275 : : */
1276 : : void
260 tmunro@postgresql.or 1277 :CBC 907 : mdtruncate(SMgrRelation reln, ForkNumber forknum,
1278 : : BlockNumber curnblk, BlockNumber nblocks)
1279 : : {
1280 : : BlockNumber priorblocks;
1281 : : int curopensegs;
1282 : :
8837 tgl@sss.pgh.pa.us 1283 [ - + ]: 907 : if (nblocks > curnblk)
1284 : : {
1285 : : /* Bogus request ... but no complaint if InRecovery */
6821 tgl@sss.pgh.pa.us 1286 [ # # ]:UBC 0 : if (InRecovery)
1287 : 0 : return;
1288 [ # # ]: 0 : ereport(ERROR,
1289 : : (errmsg("could not truncate file \"%s\" to %u blocks: it's only %u blocks now",
1290 : : relpath(reln->smgr_rlocator, forknum).str,
1291 : : nblocks, curnblk)));
1292 : : }
9501 tgl@sss.pgh.pa.us 1293 [ + + ]:CBC 907 : if (nblocks == curnblk)
6821 1294 : 371 : return; /* no work */
1295 : :
1296 : : /*
1297 : : * Truncate segments, starting at the last one. Starting at the end makes
1298 : : * managing the memory for the fd array easier, should there be errors.
1299 : : */
3285 andres@anarazel.de 1300 : 536 : curopensegs = reln->md_num_open_segs[forknum];
1301 [ + + ]: 1072 : while (curopensegs > 0)
1302 : : {
1303 : : MdfdVec *v;
1304 : :
1305 : 536 : priorblocks = (curopensegs - 1) * RELSEG_SIZE;
1306 : :
1307 : 536 : v = &reln->md_seg_fds[forknum][curopensegs - 1];
1308 : :
9501 tgl@sss.pgh.pa.us 1309 [ - + ]: 536 : if (priorblocks > nblocks)
1310 : : {
1311 : : /*
1312 : : * This segment is no longer active. We truncate the file, but do
1313 : : * not delete it, for reasons explained in the header comments.
1314 : : */
3094 rhaas@postgresql.org 1315 [ # # ]:UBC 0 : if (FileTruncate(v->mdfd_vfd, 0, WAIT_EVENT_DATA_FILE_TRUNCATE) < 0)
6821 tgl@sss.pgh.pa.us 1316 [ # # ]: 0 : ereport(ERROR,
1317 : : (errcode_for_file_access(),
1318 : : errmsg("could not truncate file \"%s\": %m",
1319 : : FilePathName(v->mdfd_vfd))));
1320 : :
5503 rhaas@postgresql.org 1321 [ # # ]: 0 : if (!SmgrIsTemp(reln))
6235 heikki.linnakangas@i 1322 : 0 : register_dirty_segment(reln, forknum, v);
1323 : :
1324 : : /* we never drop the 1st segment */
3285 andres@anarazel.de 1325 [ # # ]: 0 : Assert(v != &reln->md_seg_fds[forknum][0]);
1326 : :
1327 : 0 : FileClose(v->mdfd_vfd);
1328 : 0 : _fdvec_resize(reln, forknum, curopensegs - 1);
1329 : : }
8837 tgl@sss.pgh.pa.us 1330 [ + - ]:CBC 536 : else if (priorblocks + ((BlockNumber) RELSEG_SIZE) > nblocks)
1331 : : {
1332 : : /*
1333 : : * This is the last segment we want to keep. Truncate the file to
1334 : : * the right length. NOTE: if nblocks is exactly a multiple K of
1335 : : * RELSEG_SIZE, we will truncate the K+1st segment to 0 length but
1336 : : * keep it. This adheres to the invariant given in the header
1337 : : * comments.
1338 : : */
8717 bruce@momjian.us 1339 : 536 : BlockNumber lastsegblocks = nblocks - priorblocks;
1340 : :
3094 rhaas@postgresql.org 1341 [ - + ]: 536 : if (FileTruncate(v->mdfd_vfd, (off_t) lastsegblocks * BLCKSZ, WAIT_EVENT_DATA_FILE_TRUNCATE) < 0)
6821 tgl@sss.pgh.pa.us 1342 [ # # ]:UBC 0 : ereport(ERROR,
1343 : : (errcode_for_file_access(),
1344 : : errmsg("could not truncate file \"%s\" to %u blocks: %m",
1345 : : FilePathName(v->mdfd_vfd),
1346 : : nblocks)));
5503 rhaas@postgresql.org 1347 [ + + ]:CBC 536 : if (!SmgrIsTemp(reln))
6235 heikki.linnakangas@i 1348 : 372 : register_dirty_segment(reln, forknum, v);
1349 : : }
1350 : : else
1351 : : {
1352 : : /*
1353 : : * We still need this segment, so nothing to do for this and any
1354 : : * earlier segment.
1355 : : */
3285 andres@anarazel.de 1356 :UBC 0 : break;
1357 : : }
3285 andres@anarazel.de 1358 :CBC 536 : curopensegs--;
1359 : : }
1360 : : }
1361 : :
1362 : : /*
1363 : : * mdregistersync() -- Mark whole relation as needing fsync
1364 : : */
1365 : : void
561 heikki.linnakangas@i 1366 : 24653 : mdregistersync(SMgrRelation reln, ForkNumber forknum)
1367 : : {
1368 : : int segno;
1369 : : int min_inactive_seg;
1370 : :
1371 : : /*
1372 : : * NOTE: mdnblocks makes sure we have opened all active segments, so that
1373 : : * the loop below will get them all!
1374 : : */
1375 : 24653 : mdnblocks(reln, forknum);
1376 : :
1377 : 24653 : min_inactive_seg = segno = reln->md_num_open_segs[forknum];
1378 : :
1379 : : /*
1380 : : * Temporarily open inactive segments, then close them after sync. There
1381 : : * may be some inactive segments left opened after error, but that is
1382 : : * harmless. We don't bother to clean them up and take a risk of further
1383 : : * trouble. The next mdclose() will soon close them.
1384 : : */
1385 [ - + ]: 24653 : while (_mdfd_openseg(reln, forknum, segno, 0) != NULL)
561 heikki.linnakangas@i 1386 :UBC 0 : segno++;
1387 : :
561 heikki.linnakangas@i 1388 [ + + ]:CBC 49306 : while (segno > 0)
1389 : : {
1390 : 24653 : MdfdVec *v = &reln->md_seg_fds[forknum][segno - 1];
1391 : :
1392 : 24653 : register_dirty_segment(reln, forknum, v);
1393 : :
1394 : : /* Close inactive segments immediately */
1395 [ - + ]: 24653 : if (segno > min_inactive_seg)
1396 : : {
561 heikki.linnakangas@i 1397 :UBC 0 : FileClose(v->mdfd_vfd);
1398 : 0 : _fdvec_resize(reln, forknum, segno - 1);
1399 : : }
1400 : :
561 heikki.linnakangas@i 1401 :CBC 24653 : segno--;
1402 : : }
1403 : 24653 : }
1404 : :
1405 : : /*
1406 : : * mdimmedsync() -- Immediately sync a relation to stable storage.
1407 : : *
1408 : : * Note that only writes already issued are synced; this routine knows
1409 : : * nothing of dirty buffers that may exist inside the buffer manager. We
1410 : : * sync active and inactive segments; smgrDoPendingSyncs() relies on this.
1411 : : * Consider a relation skipping WAL. Suppose a checkpoint syncs blocks of
1412 : : * some segment, then mdtruncate() renders that segment inactive. If we
1413 : : * crash before the next checkpoint syncs the newly-inactive segment, that
1414 : : * segment may survive recovery, reintroducing unwanted data into the table.
1415 : : */
1416 : : void
6235 1417 : 23 : mdimmedsync(SMgrRelation reln, ForkNumber forknum)
1418 : : {
1419 : : int segno;
1420 : : int min_inactive_seg;
1421 : :
1422 : : /*
1423 : : * NOTE: mdnblocks makes sure we have opened all active segments, so that
1424 : : * the loop below will get them all!
1425 : : */
5262 peter_e@gmx.net 1426 : 23 : mdnblocks(reln, forknum);
1427 : :
1981 noah@leadboat.com 1428 : 23 : min_inactive_seg = segno = reln->md_num_open_segs[forknum];
1429 : :
1430 : : /*
1431 : : * Temporarily open inactive segments, then close them after sync. There
1432 : : * may be some inactive segments left opened after fsync() error, but that
1433 : : * is harmless. We don't bother to clean them up and take a risk of
1434 : : * further trouble. The next mdclose() will soon close them.
1435 : : */
1436 [ - + ]: 23 : while (_mdfd_openseg(reln, forknum, segno, 0) != NULL)
1981 noah@leadboat.com 1437 :UBC 0 : segno++;
1438 : :
3285 andres@anarazel.de 1439 [ + + ]:CBC 46 : while (segno > 0)
1440 : : {
1441 : 23 : MdfdVec *v = &reln->md_seg_fds[forknum][segno - 1];
1442 : :
1443 : : /*
1444 : : * fsyncs done through mdimmedsync() should be tracked in a separate
1445 : : * IOContext than those done through mdsyncfiletag() to differentiate
1446 : : * between unavoidable client backend fsyncs (e.g. those done during
1447 : : * index build) and those which ideally would have been done by the
1448 : : * checkpointer. Since other IO operations bypassing the buffer
1449 : : * manager could also be tracked in such an IOContext, wait until
1450 : : * these are also tracked to track immediate fsyncs.
1451 : : */
3094 rhaas@postgresql.org 1452 [ - + ]: 23 : if (FileSync(v->mdfd_vfd, WAIT_EVENT_DATA_FILE_IMMEDIATE_SYNC) < 0)
2483 tmunro@postgresql.or 1453 [ # # ]:UBC 0 : ereport(data_sync_elevel(ERROR),
1454 : : (errcode_for_file_access(),
1455 : : errmsg("could not fsync file \"%s\": %m",
1456 : : FilePathName(v->mdfd_vfd))));
1457 : :
1458 : : /* Close inactive segments immediately */
1981 noah@leadboat.com 1459 [ - + ]:CBC 23 : if (segno > min_inactive_seg)
1460 : : {
1981 noah@leadboat.com 1461 :UBC 0 : FileClose(v->mdfd_vfd);
1462 : 0 : _fdvec_resize(reln, forknum, segno - 1);
1463 : : }
1464 : :
3285 andres@anarazel.de 1465 :CBC 23 : segno--;
1466 : : }
7766 tgl@sss.pgh.pa.us 1467 : 23 : }
1468 : :
1469 : : int
161 andres@anarazel.de 1470 : 448402 : mdfd(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, uint32 *off)
1471 : : {
1472 : 448402 : MdfdVec *v = mdopenfork(reln, forknum, EXTENSION_FAIL);
1473 : :
1474 : 448402 : v = _mdfd_getseg(reln, forknum, blocknum, false,
1475 : : EXTENSION_FAIL);
1476 : :
1477 : 448402 : *off = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
1478 : :
1479 [ - + ]: 448402 : Assert(*off < (off_t) BLCKSZ * RELSEG_SIZE);
1480 : :
1481 : 448402 : return FileGetRawDesc(v->mdfd_vfd);
1482 : : }
1483 : :
1484 : : /*
1485 : : * register_dirty_segment() -- Mark a relation segment as needing fsync
1486 : : *
1487 : : * If there is a local pending-ops table, just make an entry in it for
1488 : : * ProcessSyncRequests to process later. Otherwise, try to pass off the
1489 : : * fsync request to the checkpointer process. If that fails, just do the
1490 : : * fsync locally before returning (we hope this will not happen often
1491 : : * enough to be a performance problem).
1492 : : */
1493 : : static void
6235 heikki.linnakangas@i 1494 : 925660 : register_dirty_segment(SMgrRelation reln, ForkNumber forknum, MdfdVec *seg)
1495 : : {
1496 : : FileTag tag;
1497 : :
1158 rhaas@postgresql.org 1498 : 925660 : INIT_MD_FILETAG(tag, reln->smgr_rlocator.locator, forknum, seg->mdfd_segno);
1499 : :
1500 : : /* Temp relations should never be fsync'd */
4799 tgl@sss.pgh.pa.us 1501 [ - + ]: 925660 : Assert(!SmgrIsTemp(reln));
1502 : :
2347 tmunro@postgresql.or 1503 [ + + ]: 925660 : if (!RegisterSyncRequest(&tag, SYNC_REQUEST, false /* retryOnError */ ))
1504 : : {
1505 : : instr_time io_start;
1506 : :
883 andres@anarazel.de 1507 [ - + ]: 243 : ereport(DEBUG1,
1508 : : (errmsg_internal("could not forward fsync request because request queue is full")));
1509 : :
192 michael@paquier.xyz 1510 : 243 : io_start = pgstat_prepare_io_time(track_io_timing);
1511 : :
883 andres@anarazel.de 1512 [ - + ]: 243 : if (FileSync(seg->mdfd_vfd, WAIT_EVENT_DATA_FILE_SYNC) < 0)
883 andres@anarazel.de 1513 [ # # ]:UBC 0 : ereport(data_sync_elevel(ERROR),
1514 : : (errcode_for_file_access(),
1515 : : errmsg("could not fsync file \"%s\": %m",
1516 : : FilePathName(seg->mdfd_vfd))));
1517 : :
1518 : : /*
1519 : : * We have no way of knowing if the current IOContext is
1520 : : * IOCONTEXT_NORMAL or IOCONTEXT_[BULKREAD, BULKWRITE, VACUUM] at this
1521 : : * point, so count the fsync as being in the IOCONTEXT_NORMAL
1522 : : * IOContext. This is probably okay, because the number of backend
1523 : : * fsyncs doesn't say anything about the efficacy of the
1524 : : * BufferAccessStrategy. And counting both fsyncs done in
1525 : : * IOCONTEXT_NORMAL and IOCONTEXT_[BULKREAD, BULKWRITE, VACUUM] under
1526 : : * IOCONTEXT_NORMAL is likely clearer when investigating the number of
1527 : : * backend fsyncs.
1528 : : */
883 andres@anarazel.de 1529 :CBC 243 : pgstat_count_io_op_time(IOOBJECT_RELATION, IOCONTEXT_NORMAL,
1530 : : IOOP_FSYNC, io_start, 1, 0);
1531 : : }
10651 scrappy@hub.org 1532 : 925660 : }
1533 : :
1534 : : /*
1535 : : * register_unlink_segment() -- Schedule a file to be deleted after next checkpoint
1536 : : */
1537 : : static void
1158 rhaas@postgresql.org 1538 : 34463 : register_unlink_segment(RelFileLocatorBackend rlocator, ForkNumber forknum,
1539 : : BlockNumber segno)
1540 : : {
1541 : : FileTag tag;
1542 : :
1543 : 34463 : INIT_MD_FILETAG(tag, rlocator.locator, forknum, segno);
1544 : :
1545 : : /* Should never be used with temp relations */
1546 [ - + ]: 34463 : Assert(!RelFileLocatorBackendIsTemp(rlocator));
1547 : :
2347 tmunro@postgresql.or 1548 : 34463 : RegisterSyncRequest(&tag, SYNC_UNLINK_REQUEST, true /* retryOnError */ );
6505 tgl@sss.pgh.pa.us 1549 : 34463 : }
1550 : :
1551 : : /*
1552 : : * register_forget_request() -- forget any fsyncs for a relation fork's segment
1553 : : */
1554 : : static void
1158 rhaas@postgresql.org 1555 : 133673 : register_forget_request(RelFileLocatorBackend rlocator, ForkNumber forknum,
1556 : : BlockNumber segno)
1557 : : {
1558 : : FileTag tag;
1559 : :
1560 : 133673 : INIT_MD_FILETAG(tag, rlocator.locator, forknum, segno);
1561 : :
2347 tmunro@postgresql.or 1562 : 133673 : RegisterSyncRequest(&tag, SYNC_FORGET_REQUEST, true /* retryOnError */ );
6807 tgl@sss.pgh.pa.us 1563 : 133673 : }
1564 : :
1565 : : /*
1566 : : * ForgetDatabaseSyncRequests -- forget any fsyncs and unlinks for a DB
1567 : : */
1568 : : void
2347 tmunro@postgresql.or 1569 : 59 : ForgetDatabaseSyncRequests(Oid dbid)
1570 : : {
1571 : : FileTag tag;
1572 : : RelFileLocator rlocator;
1573 : :
1158 rhaas@postgresql.org 1574 : 59 : rlocator.dbOid = dbid;
1575 : 59 : rlocator.spcOid = 0;
1576 : 59 : rlocator.relNumber = 0;
1577 : :
1578 : 59 : INIT_MD_FILETAG(tag, rlocator, InvalidForkNumber, InvalidBlockNumber);
1579 : :
2347 tmunro@postgresql.or 1580 : 59 : RegisterSyncRequest(&tag, SYNC_FILTER_REQUEST, true /* retryOnError */ );
9079 vadim4o@yahoo.com 1581 : 59 : }
1582 : :
1583 : : /*
1584 : : * DropRelationFiles -- drop files of all given relations
1585 : : */
1586 : : void
1158 rhaas@postgresql.org 1587 : 2620 : DropRelationFiles(RelFileLocator *delrels, int ndelrels, bool isRedo)
1588 : : {
1589 : : SMgrRelation *srels;
1590 : : int i;
1591 : :
2620 fujii@postgresql.org 1592 : 2620 : srels = palloc(sizeof(SMgrRelation) * ndelrels);
1593 [ + + ]: 10159 : for (i = 0; i < ndelrels; i++)
1594 : : {
552 heikki.linnakangas@i 1595 : 7539 : SMgrRelation srel = smgropen(delrels[i], INVALID_PROC_NUMBER);
1596 : :
2620 fujii@postgresql.org 1597 [ + + ]: 7539 : if (isRedo)
1598 : : {
1599 : : ForkNumber fork;
1600 : :
1601 [ + + ]: 37575 : for (fork = 0; fork <= MAX_FORKNUM; fork++)
1602 : 30060 : XLogDropRelation(delrels[i], fork);
1603 : : }
1604 : 7539 : srels[i] = srel;
1605 : : }
1606 : :
1607 : 2620 : smgrdounlinkall(srels, ndelrels, isRedo);
1608 : :
2355 tomas.vondra@postgre 1609 [ + + ]: 10159 : for (i = 0; i < ndelrels; i++)
2620 fujii@postgresql.org 1610 : 7539 : smgrclose(srels[i]);
1611 : 2620 : pfree(srels);
1612 : 2620 : }
1613 : :
1614 : :
1615 : : /*
1616 : : * _fdvec_resize() -- Resize the fork's open segments array
1617 : : */
1618 : : static void
3285 andres@anarazel.de 1619 : 1378544 : _fdvec_resize(SMgrRelation reln,
1620 : : ForkNumber forknum,
1621 : : int nseg)
1622 : : {
1623 [ + + ]: 1378544 : if (nseg == 0)
1624 : : {
1625 [ + - ]: 509842 : if (reln->md_num_open_segs[forknum] > 0)
1626 : : {
1627 : 509842 : pfree(reln->md_seg_fds[forknum]);
1628 : 509842 : reln->md_seg_fds[forknum] = NULL;
1629 : : }
1630 : : }
1631 [ + - ]: 868702 : else if (reln->md_num_open_segs[forknum] == 0)
1632 : : {
1633 : 868702 : reln->md_seg_fds[forknum] =
1634 : 868702 : MemoryContextAlloc(MdCxt, sizeof(MdfdVec) * nseg);
1635 : : }
260 tmunro@postgresql.or 1636 [ # # ]:UBC 0 : else if (nseg > reln->md_num_open_segs[forknum])
1637 : : {
1638 : : /*
1639 : : * It doesn't seem worthwhile complicating the code to amortize
1640 : : * repalloc() calls. Those are far faster than PathNameOpenFile() or
1641 : : * FileClose(), and the memory context internally will sometimes avoid
1642 : : * doing an actual reallocation.
1643 : : */
3285 andres@anarazel.de 1644 : 0 : reln->md_seg_fds[forknum] =
1645 : 0 : repalloc(reln->md_seg_fds[forknum],
1646 : : sizeof(MdfdVec) * nseg);
1647 : : }
1648 : : else
1649 : : {
1650 : : /*
1651 : : * We don't reallocate a smaller array, because we want mdtruncate()
1652 : : * to be able to promise that it won't allocate memory, so that it is
1653 : : * allowed in a critical section. This means that a bit of space in
1654 : : * the array is now wasted, until the next time we add a segment and
1655 : : * reallocate.
1656 : : */
1657 : : }
1658 : :
3285 andres@anarazel.de 1659 :CBC 1378544 : reln->md_num_open_segs[forknum] = nseg;
10334 vadim4o@yahoo.com 1660 : 1378544 : }
1661 : :
1662 : : /*
1663 : : * Return the filename for the specified segment of the relation. The
1664 : : * returned string is palloc'd.
1665 : : */
1666 : : static MdPathStr
5876 heikki.linnakangas@i 1667 : 24688 : _mdfd_segpath(SMgrRelation reln, ForkNumber forknum, BlockNumber segno)
1668 : : {
1669 : : RelPathStr path;
1670 : : MdPathStr fullpath;
1671 : :
1158 rhaas@postgresql.org 1672 : 24688 : path = relpath(reln->smgr_rlocator, forknum);
1673 : :
10226 bruce@momjian.us 1674 [ + - ]: 24688 : if (segno > 0)
193 andres@anarazel.de 1675 : 24688 : sprintf(fullpath.str, "%s.%u", path.str, segno);
1676 : : else
193 andres@anarazel.de 1677 :UBC 0 : strcpy(fullpath.str, path.str);
1678 : :
5876 heikki.linnakangas@i 1679 :CBC 24688 : return fullpath;
1680 : : }
1681 : :
1682 : : /*
1683 : : * Open the specified segment of the relation,
1684 : : * and make a MdfdVec object for it. Returns NULL on failure.
1685 : : */
1686 : : static MdfdVec *
1687 : 24676 : _mdfd_openseg(SMgrRelation reln, ForkNumber forknum, BlockNumber segno,
1688 : : int oflags)
1689 : : {
1690 : : MdfdVec *v;
1691 : : File fd;
1692 : : MdPathStr fullpath;
1693 : :
1694 : 24676 : fullpath = _mdfd_segpath(reln, forknum, segno);
1695 : :
1696 : : /* open the file */
193 andres@anarazel.de 1697 : 24676 : fd = PathNameOpenFile(fullpath.str, _mdfd_open_flags() | oflags);
1698 : :
10226 bruce@momjian.us 1699 [ + - ]: 24676 : if (fd < 0)
7913 neilc@samurai.com 1700 : 24676 : return NULL;
1701 : :
1702 : : /*
1703 : : * Segments are always opened in order from lowest to highest, so we must
1704 : : * be adding a new one at the end.
1705 : : */
2049 tmunro@postgresql.or 1706 [ # # ]:UBC 0 : Assert(segno == reln->md_num_open_segs[forknum]);
1707 : :
1708 : 0 : _fdvec_resize(reln, forknum, segno + 1);
1709 : :
1710 : : /* fill the entry */
3285 andres@anarazel.de 1711 : 0 : v = &reln->md_seg_fds[forknum][segno];
10226 bruce@momjian.us 1712 : 0 : v->mdfd_vfd = fd;
7768 tgl@sss.pgh.pa.us 1713 : 0 : v->mdfd_segno = segno;
1714 : :
6235 heikki.linnakangas@i 1715 [ # # ]: 0 : Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
1716 : :
1717 : : /* all done */
9867 bruce@momjian.us 1718 : 0 : return v;
1719 : : }
1720 : :
1721 : : /*
1722 : : * _mdfd_getseg() -- Find the segment of the relation holding the
1723 : : * specified block.
1724 : : *
1725 : : * If the segment doesn't exist, we ereport, return NULL, or create the
1726 : : * segment, according to "behavior". Note: skipFsync is only used in the
1727 : : * EXTENSION_CREATE case.
1728 : : */
1729 : : static MdfdVec *
6235 heikki.linnakangas@i 1730 :CBC 2585204 : _mdfd_getseg(SMgrRelation reln, ForkNumber forknum, BlockNumber blkno,
1731 : : bool skipFsync, int behavior)
1732 : : {
1733 : : MdfdVec *v;
1734 : : BlockNumber targetseg;
1735 : : BlockNumber nextsegno;
1736 : :
1737 : : /* some way to handle non-existent segments needs to be specified */
3412 andres@anarazel.de 1738 [ - + ]: 2585204 : Assert(behavior &
1739 : : (EXTENSION_FAIL | EXTENSION_CREATE | EXTENSION_RETURN_NULL |
1740 : : EXTENSION_DONT_OPEN));
1741 : :
6821 tgl@sss.pgh.pa.us 1742 : 2585204 : targetseg = blkno / ((BlockNumber) RELSEG_SIZE);
1743 : :
1744 : : /* if an existing and opened segment, we're done */
3285 andres@anarazel.de 1745 [ + + ]: 2585204 : if (targetseg < reln->md_num_open_segs[forknum])
1746 : : {
1747 : 2363147 : v = &reln->md_seg_fds[forknum][targetseg];
1748 : 2363147 : return v;
1749 : : }
1750 : :
1751 : : /* The caller only wants the segment if we already had it open. */
1218 tmunro@postgresql.or 1752 [ - + ]: 222057 : if (behavior & EXTENSION_DONT_OPEN)
1218 tmunro@postgresql.or 1753 :UBC 0 : return NULL;
1754 : :
1755 : : /*
1756 : : * The target segment is not yet open. Iterate over all the segments
1757 : : * between the last opened and the target segment. This way missing
1758 : : * segments either raise an error, or get created (according to
1759 : : * 'behavior'). Start with either the last opened, or the first segment if
1760 : : * none was opened before.
1761 : : */
3285 andres@anarazel.de 1762 [ + + ]:CBC 222057 : if (reln->md_num_open_segs[forknum] > 0)
1763 : 12 : v = &reln->md_seg_fds[forknum][reln->md_num_open_segs[forknum] - 1];
1764 : : else
1765 : : {
2243 tmunro@postgresql.or 1766 : 222045 : v = mdopenfork(reln, forknum, behavior);
3285 andres@anarazel.de 1767 [ - + ]: 222042 : if (!v)
3285 andres@anarazel.de 1768 :UBC 0 : return NULL; /* if behavior & EXTENSION_RETURN_NULL */
1769 : : }
1770 : :
3285 andres@anarazel.de 1771 :CBC 222054 : for (nextsegno = reln->md_num_open_segs[forknum];
1772 [ + + ]: 222054 : nextsegno <= targetseg; nextsegno++)
1773 : : {
1774 : 12 : BlockNumber nblocks = _mdnblocks(reln, forknum, v);
1775 : 12 : int flags = 0;
1776 : :
1777 [ - + ]: 12 : Assert(nextsegno == v->mdfd_segno + 1);
1778 : :
1779 [ - + ]: 12 : if (nblocks > ((BlockNumber) RELSEG_SIZE))
3285 andres@anarazel.de 1780 [ # # ]:UBC 0 : elog(FATAL, "segment too big");
1781 : :
3285 andres@anarazel.de 1782 [ + - ]:CBC 12 : if ((behavior & EXTENSION_CREATE) ||
1783 [ - + - - ]: 12 : (InRecovery && (behavior & EXTENSION_CREATE_RECOVERY)))
1784 : : {
1785 : : /*
1786 : : * Normally we will create new segments only if authorized by the
1787 : : * caller (i.e., we are doing mdextend()). But when doing WAL
1788 : : * recovery, create segments anyway; this allows cases such as
1789 : : * replaying WAL data that has a write into a high-numbered
1790 : : * segment of a relation that was later deleted. We want to go
1791 : : * ahead and create the segments so we can finish out the replay.
1792 : : *
1793 : : * We have to maintain the invariant that segments before the last
1794 : : * active segment are of size RELSEG_SIZE; therefore, if
1795 : : * extending, pad them out with zeroes if needed. (This only
1796 : : * matters if in recovery, or if the caller is extending the
1797 : : * relation discontiguously, but that can happen in hash indexes.)
1798 : : */
3285 andres@anarazel.de 1799 [ # # ]:UBC 0 : if (nblocks < ((BlockNumber) RELSEG_SIZE))
1800 : : {
882 tmunro@postgresql.or 1801 : 0 : char *zerobuf = palloc_aligned(BLCKSZ, PG_IO_ALIGN_SIZE,
1802 : : MCXT_ALLOC_ZERO);
1803 : :
3285 andres@anarazel.de 1804 : 0 : mdextend(reln, forknum,
1805 : 0 : nextsegno * ((BlockNumber) RELSEG_SIZE) - 1,
1806 : : zerobuf, skipFsync);
1807 : 0 : pfree(zerobuf);
1808 : : }
1809 : 0 : flags = O_CREAT;
1810 : : }
266 tmunro@postgresql.or 1811 [ + - ]:CBC 12 : else if (nblocks < ((BlockNumber) RELSEG_SIZE))
1812 : : {
1813 : : /*
1814 : : * When not extending, only open the next segment if the current
1815 : : * one is exactly RELSEG_SIZE. If not (this branch), either
1816 : : * return NULL or fail.
1817 : : */
3285 andres@anarazel.de 1818 [ - + ]: 12 : if (behavior & EXTENSION_RETURN_NULL)
1819 : : {
1820 : : /*
1821 : : * Some callers discern between reasons for _mdfd_getseg()
1822 : : * returning NULL based on errno. As there's no failing
1823 : : * syscall involved in this case, explicitly set errno to
1824 : : * ENOENT, as that seems the closest interpretation.
1825 : : */
3285 andres@anarazel.de 1826 :UBC 0 : errno = ENOENT;
1827 : 0 : return NULL;
1828 : : }
1829 : :
3285 andres@anarazel.de 1830 [ + - ]:CBC 12 : ereport(ERROR,
1831 : : (errcode_for_file_access(),
1832 : : errmsg("could not open file \"%s\" (target block %u): previous segment is only %u blocks",
1833 : : _mdfd_segpath(reln, forknum, nextsegno).str,
1834 : : blkno, nblocks)));
1835 : : }
1836 : :
3285 andres@anarazel.de 1837 :UBC 0 : v = _mdfd_openseg(reln, forknum, nextsegno, flags);
1838 : :
1839 [ # # ]: 0 : if (v == NULL)
1840 : : {
1841 [ # # ]: 0 : if ((behavior & EXTENSION_RETURN_NULL) &&
1842 [ # # ]: 0 : FILE_POSSIBLY_DELETED(errno))
1843 : 0 : return NULL;
1844 [ # # ]: 0 : ereport(ERROR,
1845 : : (errcode_for_file_access(),
1846 : : errmsg("could not open file \"%s\" (target block %u): %m",
1847 : : _mdfd_segpath(reln, forknum, nextsegno).str,
1848 : : blkno)));
1849 : : }
1850 : : }
1851 : :
9867 bruce@momjian.us 1852 :CBC 222042 : return v;
1853 : : }
1854 : :
1855 : : /*
1856 : : * Get number of blocks present in a single disk file
1857 : : */
1858 : : static BlockNumber
6235 heikki.linnakangas@i 1859 : 3163785 : _mdnblocks(SMgrRelation reln, ForkNumber forknum, MdfdVec *seg)
1860 : : {
1861 : : off_t len;
1862 : :
2495 tmunro@postgresql.or 1863 : 3163785 : len = FileSize(seg->mdfd_vfd);
9278 bruce@momjian.us 1864 [ - + ]: 3163785 : if (len < 0)
6821 tgl@sss.pgh.pa.us 1865 [ # # ]:UBC 0 : ereport(ERROR,
1866 : : (errcode_for_file_access(),
1867 : : errmsg("could not seek to end of file \"%s\": %m",
1868 : : FilePathName(seg->mdfd_vfd))));
1869 : : /* note that this calculation will ignore any partial block at EOF */
6821 tgl@sss.pgh.pa.us 1870 :CBC 3163785 : return (BlockNumber) (len / BLCKSZ);
1871 : : }
1872 : :
1873 : : /*
1874 : : * Sync a file to disk, given a file tag. Write the path into an output
1875 : : * buffer so the caller can use it in error messages.
1876 : : *
1877 : : * Return 0 on success, -1 on failure, with errno set.
1878 : : */
1879 : : int
2347 tmunro@postgresql.or 1880 :UBC 0 : mdsyncfiletag(const FileTag *ftag, char *path)
1881 : : {
552 heikki.linnakangas@i 1882 : 0 : SMgrRelation reln = smgropen(ftag->rlocator, INVALID_PROC_NUMBER);
1883 : : File file;
1884 : : instr_time io_start;
1885 : : bool need_to_close;
1886 : : int result,
1887 : : save_errno;
1888 : :
1889 : : /* See if we already have the file open, or need to open it. */
2093 tmunro@postgresql.or 1890 [ # # ]: 0 : if (ftag->segno < reln->md_num_open_segs[ftag->forknum])
1891 : : {
1892 : 0 : file = reln->md_seg_fds[ftag->forknum][ftag->segno].mdfd_vfd;
1893 : 0 : strlcpy(path, FilePathName(file), MAXPGPATH);
1894 : 0 : need_to_close = false;
1895 : : }
1896 : : else
1897 : : {
1898 : : MdPathStr p;
1899 : :
1900 : 0 : p = _mdfd_segpath(reln, ftag->forknum, ftag->segno);
193 andres@anarazel.de 1901 : 0 : strlcpy(path, p.str, MD_PATH_STR_MAXLEN);
1902 : :
882 tmunro@postgresql.or 1903 : 0 : file = PathNameOpenFile(path, _mdfd_open_flags());
2093 1904 [ # # ]: 0 : if (file < 0)
1905 : 0 : return -1;
1906 : 0 : need_to_close = true;
1907 : : }
1908 : :
192 michael@paquier.xyz 1909 : 0 : io_start = pgstat_prepare_io_time(track_io_timing);
1910 : :
1911 : : /* Sync the file. */
2093 tmunro@postgresql.or 1912 : 0 : result = FileSync(file, WAIT_EVENT_DATA_FILE_SYNC);
1913 : 0 : save_errno = errno;
1914 : :
1915 [ # # ]: 0 : if (need_to_close)
1916 : 0 : FileClose(file);
1917 : :
883 andres@anarazel.de 1918 : 0 : pgstat_count_io_op_time(IOOBJECT_RELATION, IOCONTEXT_NORMAL,
1919 : : IOOP_FSYNC, io_start, 1, 0);
1920 : :
2093 tmunro@postgresql.or 1921 : 0 : errno = save_errno;
1922 : 0 : return result;
1923 : : }
1924 : :
1925 : : /*
1926 : : * Unlink a file, given a file tag. Write the path into an output
1927 : : * buffer so the caller can use it in error messages.
1928 : : *
1929 : : * Return 0 on success, -1 on failure, with errno set.
1930 : : */
1931 : : int
2347 tmunro@postgresql.or 1932 :CBC 32525 : mdunlinkfiletag(const FileTag *ftag, char *path)
1933 : : {
1934 : : RelPathStr p;
1935 : :
1936 : : /* Compute the path. */
1158 rhaas@postgresql.org 1937 : 32525 : p = relpathperm(ftag->rlocator, MAIN_FORKNUM);
193 andres@anarazel.de 1938 : 32525 : strlcpy(path, p.str, MAXPGPATH);
1939 : :
1940 : : /* Try to unlink the file. */
2347 tmunro@postgresql.or 1941 : 32525 : return unlink(path);
1942 : : }
1943 : :
1944 : : /*
1945 : : * Check if a given candidate request matches a given tag, when processing
1946 : : * a SYNC_FILTER_REQUEST request. This will be called for all pending
1947 : : * requests to find out whether to forget them.
1948 : : */
1949 : : bool
1950 : 7804 : mdfiletagmatches(const FileTag *ftag, const FileTag *candidate)
1951 : : {
1952 : : /*
1953 : : * For now we only use filter requests as a way to drop all scheduled
1954 : : * callbacks relating to a given database, when dropping the database.
1955 : : * We'll return true for all candidates that have the same database OID as
1956 : : * the ftag from the SYNC_FILTER_REQUEST request, so they're forgotten.
1957 : : */
1158 rhaas@postgresql.org 1958 : 7804 : return ftag->rlocator.dbOid == candidate->rlocator.dbOid;
1959 : : }
1960 : :
1961 : : /*
1962 : : * AIO completion callback for mdstartreadv().
1963 : : */
1964 : : static PgAioResult
161 andres@anarazel.de 1965 : 1142229 : md_readv_complete(PgAioHandle *ioh, PgAioResult prior_result, uint8 cb_data)
1966 : : {
1967 : 1142229 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
1968 : 1142229 : PgAioResult result = prior_result;
1969 : :
1970 [ - + ]: 1142229 : if (prior_result.result < 0)
1971 : : {
161 andres@anarazel.de 1972 :UBC 0 : result.status = PGAIO_RS_ERROR;
1973 : 0 : result.id = PGAIO_HCB_MD_READV;
1974 : : /* For "hard" errors, track the error number in error_data */
1975 : 0 : result.error_data = -prior_result.result;
1976 : 0 : result.result = 0;
1977 : :
1978 : : /*
1979 : : * Immediately log a message about the IO error, but only to the
1980 : : * server log. The reason to do so immediately is that the originator
1981 : : * might not process the query result immediately (because it is busy
1982 : : * doing another part of query processing) or at all (e.g. if it was
1983 : : * cancelled or errored out due to another IO also failing). The
1984 : : * definer of the IO will emit an ERROR when processing the IO's
1985 : : * results
1986 : : */
1987 : 0 : pgaio_result_report(result, td, LOG_SERVER_ONLY);
1988 : :
1989 : 0 : return result;
1990 : : }
1991 : :
1992 : : /*
1993 : : * As explained above smgrstartreadv(), the smgr API operates on the level
1994 : : * of blocks, rather than bytes. Convert.
1995 : : */
161 andres@anarazel.de 1996 :CBC 1142229 : result.result /= BLCKSZ;
1997 : :
1998 [ - + ]: 1142229 : Assert(result.result <= td->smgr.nblocks);
1999 : :
2000 [ - + ]: 1142229 : if (result.result == 0)
2001 : : {
2002 : : /* consider 0 blocks read a failure */
161 andres@anarazel.de 2003 :UBC 0 : result.status = PGAIO_RS_ERROR;
2004 : 0 : result.id = PGAIO_HCB_MD_READV;
2005 : 0 : result.error_data = 0;
2006 : :
2007 : : /* see comment above the "hard error" case */
2008 : 0 : pgaio_result_report(result, td, LOG_SERVER_ONLY);
2009 : :
2010 : 0 : return result;
2011 : : }
2012 : :
161 andres@anarazel.de 2013 [ + - ]:CBC 1142229 : if (result.status != PGAIO_RS_ERROR &&
2014 [ - + ]: 1142229 : result.result < td->smgr.nblocks)
2015 : : {
2016 : : /* partial reads should be retried at upper level */
161 andres@anarazel.de 2017 :UBC 0 : result.status = PGAIO_RS_PARTIAL;
2018 : 0 : result.id = PGAIO_HCB_MD_READV;
2019 : : }
2020 : :
161 andres@anarazel.de 2021 :CBC 1142229 : return result;
2022 : : }
2023 : :
2024 : : /*
2025 : : * AIO error reporting callback for mdstartreadv().
2026 : : *
2027 : : * Errors are encoded as follows:
2028 : : * - PgAioResult.error_data != 0 encodes IO that failed with that errno
2029 : : * - PgAioResult.error_data == 0 encodes IO that didn't read all data
2030 : : */
2031 : : static void
161 andres@anarazel.de 2032 :UBC 0 : md_readv_report(PgAioResult result, const PgAioTargetData *td, int elevel)
2033 : : {
2034 : : RelPathStr path;
2035 : :
2036 [ # # ]: 0 : path = relpathbackend(td->smgr.rlocator,
2037 : : td->smgr.is_temp ? MyProcNumber : INVALID_PROC_NUMBER,
2038 : : td->smgr.forkNum);
2039 : :
2040 [ # # ]: 0 : if (result.error_data != 0)
2041 : : {
2042 : : /* for errcode_for_file_access() and %m */
2043 : 0 : errno = result.error_data;
2044 : :
2045 [ # # ]: 0 : ereport(elevel,
2046 : : errcode_for_file_access(),
2047 : : errmsg("could not read blocks %u..%u in file \"%s\": %m",
2048 : : td->smgr.blockNum,
2049 : : td->smgr.blockNum + td->smgr.nblocks - 1,
2050 : : path.str));
2051 : : }
2052 : : else
2053 : : {
2054 : : /*
2055 : : * NB: This will typically only be output in debug messages, while
2056 : : * retrying a partial IO.
2057 : : */
2058 [ # # ]: 0 : ereport(elevel,
2059 : : errcode(ERRCODE_DATA_CORRUPTED),
2060 : : errmsg("could not read blocks %u..%u in file \"%s\": read only %zu of %zu bytes",
2061 : : td->smgr.blockNum,
2062 : : td->smgr.blockNum + td->smgr.nblocks - 1,
2063 : : path.str,
2064 : : result.result * (size_t) BLCKSZ,
2065 : : td->smgr.nblocks * (size_t) BLCKSZ));
2066 : : }
2067 : 0 : }
|
