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) 1996-2017, 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 "miscadmin.h"
29 : #include "access/xlog.h"
30 : #include "catalog/catalog.h"
31 : #include "pgstat.h"
32 : #include "portability/instr_time.h"
33 : #include "postmaster/bgwriter.h"
34 : #include "storage/fd.h"
35 : #include "storage/bufmgr.h"
36 : #include "storage/relfilenode.h"
37 : #include "storage/smgr.h"
38 : #include "utils/hsearch.h"
39 : #include "utils/memutils.h"
40 : #include "pg_trace.h"
41 :
42 :
43 : /* intervals for calling AbsorbFsyncRequests in mdsync and mdpostckpt */
44 : #define FSYNCS_PER_ABSORB 10
45 : #define UNLINKS_PER_ABSORB 10
46 :
47 : /*
48 : * Special values for the segno arg to RememberFsyncRequest.
49 : *
50 : * Note that CompactCheckpointerRequestQueue assumes that it's OK to remove an
51 : * fsync request from the queue if an identical, subsequent request is found.
52 : * See comments there before making changes here.
53 : */
54 : #define FORGET_RELATION_FSYNC (InvalidBlockNumber)
55 : #define FORGET_DATABASE_FSYNC (InvalidBlockNumber-1)
56 : #define UNLINK_RELATION_REQUEST (InvalidBlockNumber-2)
57 :
58 : /*
59 : * On Windows, we have to interpret EACCES as possibly meaning the same as
60 : * ENOENT, because if a file is unlinked-but-not-yet-gone on that platform,
61 : * that's what you get. Ugh. This code is designed so that we don't
62 : * actually believe these cases are okay without further evidence (namely,
63 : * a pending fsync request getting canceled ... see mdsync).
64 : */
65 : #ifndef WIN32
66 : #define FILE_POSSIBLY_DELETED(err) ((err) == ENOENT)
67 : #else
68 : #define FILE_POSSIBLY_DELETED(err) ((err) == ENOENT || (err) == EACCES)
69 : #endif
70 :
71 : /*
72 : * The magnetic disk storage manager keeps track of open file
73 : * descriptors in its own descriptor pool. This is done to make it
74 : * easier to support relations that are larger than the operating
75 : * system's file size limit (often 2GBytes). In order to do that,
76 : * we break relations up into "segment" files that are each shorter than
77 : * the OS file size limit. The segment size is set by the RELSEG_SIZE
78 : * configuration constant in pg_config.h.
79 : *
80 : * On disk, a relation must consist of consecutively numbered segment
81 : * files in the pattern
82 : * -- Zero or more full segments of exactly RELSEG_SIZE blocks each
83 : * -- Exactly one partial segment of size 0 <= size < RELSEG_SIZE blocks
84 : * -- Optionally, any number of inactive segments of size 0 blocks.
85 : * The full and partial segments are collectively the "active" segments.
86 : * Inactive segments are those that once contained data but are currently
87 : * not needed because of an mdtruncate() operation. The reason for leaving
88 : * them present at size zero, rather than unlinking them, is that other
89 : * backends and/or the checkpointer might be holding open file references to
90 : * such segments. If the relation expands again after mdtruncate(), such
91 : * that a deactivated segment becomes active again, it is important that
92 : * such file references still be valid --- else data might get written
93 : * out to an unlinked old copy of a segment file that will eventually
94 : * disappear.
95 : *
96 : * File descriptors are stored in the per-fork md_seg_fds arrays inside
97 : * SMgrRelation. The length of these arrays is stored in md_num_open_segs.
98 : * Note that a fork's md_num_open_segs having a specific value does not
99 : * necessarily mean the relation doesn't have additional segments; we may
100 : * just not have opened the next segment yet. (We could not have "all
101 : * segments are in the array" as an invariant anyway, since another backend
102 : * could extend the relation while we aren't looking.) We do not have
103 : * entries for inactive segments, however; as soon as we find a partial
104 : * segment, we assume that any subsequent segments are inactive.
105 : *
106 : * The entire MdfdVec array is palloc'd in the MdCxt memory context.
107 : */
108 :
109 : typedef struct _MdfdVec
110 : {
111 : File mdfd_vfd; /* fd number in fd.c's pool */
112 : BlockNumber mdfd_segno; /* segment number, from 0 */
113 : } MdfdVec;
114 :
115 : static MemoryContext MdCxt; /* context for all MdfdVec objects */
116 :
117 :
118 : /*
119 : * In some contexts (currently, standalone backends and the checkpointer)
120 : * we keep track of pending fsync operations: we need to remember all relation
121 : * segments that have been written since the last checkpoint, so that we can
122 : * fsync them down to disk before completing the next checkpoint. This hash
123 : * table remembers the pending operations. We use a hash table mostly as
124 : * a convenient way of merging duplicate requests.
125 : *
126 : * We use a similar mechanism to remember no-longer-needed files that can
127 : * be deleted after the next checkpoint, but we use a linked list instead of
128 : * a hash table, because we don't expect there to be any duplicate requests.
129 : *
130 : * These mechanisms are only used for non-temp relations; we never fsync
131 : * temp rels, nor do we need to postpone their deletion (see comments in
132 : * mdunlink).
133 : *
134 : * (Regular backends do not track pending operations locally, but forward
135 : * them to the checkpointer.)
136 : */
137 : typedef uint16 CycleCtr; /* can be any convenient integer size */
138 :
139 : typedef struct
140 : {
141 : RelFileNode rnode; /* hash table key (must be first!) */
142 : CycleCtr cycle_ctr; /* mdsync_cycle_ctr of oldest request */
143 : /* requests[f] has bit n set if we need to fsync segment n of fork f */
144 : Bitmapset *requests[MAX_FORKNUM + 1];
145 : /* canceled[f] is true if we canceled fsyncs for fork "recently" */
146 : bool canceled[MAX_FORKNUM + 1];
147 : } PendingOperationEntry;
148 :
149 : typedef struct
150 : {
151 : RelFileNode rnode; /* the dead relation to delete */
152 : CycleCtr cycle_ctr; /* mdckpt_cycle_ctr when request was made */
153 : } PendingUnlinkEntry;
154 :
155 : static HTAB *pendingOpsTable = NULL;
156 : static List *pendingUnlinks = NIL;
157 : static MemoryContext pendingOpsCxt; /* context for the above */
158 :
159 : static CycleCtr mdsync_cycle_ctr = 0;
160 : static CycleCtr mdckpt_cycle_ctr = 0;
161 :
162 :
163 : /*** behavior for mdopen & _mdfd_getseg ***/
164 : /* ereport if segment not present */
165 : #define EXTENSION_FAIL (1 << 0)
166 : /* return NULL if segment not present */
167 : #define EXTENSION_RETURN_NULL (1 << 1)
168 : /* create new segments as needed */
169 : #define EXTENSION_CREATE (1 << 2)
170 : /* create new segments if needed during recovery */
171 : #define EXTENSION_CREATE_RECOVERY (1 << 3)
172 : /*
173 : * Allow opening segments which are preceded by segments smaller than
174 : * RELSEG_SIZE, e.g. inactive segments (see above). Note that this is breaks
175 : * mdnblocks() and related functionality henceforth - which currently is ok,
176 : * because this is only required in the checkpointer which never uses
177 : * mdnblocks().
178 : */
179 : #define EXTENSION_DONT_CHECK_SIZE (1 << 4)
180 :
181 :
182 : /* local routines */
183 : static void mdunlinkfork(RelFileNodeBackend rnode, ForkNumber forkNum,
184 : bool isRedo);
185 : static MdfdVec *mdopen(SMgrRelation reln, ForkNumber forknum, int behavior);
186 : static void register_dirty_segment(SMgrRelation reln, ForkNumber forknum,
187 : MdfdVec *seg);
188 : static void register_unlink(RelFileNodeBackend rnode);
189 : static void _fdvec_resize(SMgrRelation reln,
190 : ForkNumber forknum,
191 : int nseg);
192 : static char *_mdfd_segpath(SMgrRelation reln, ForkNumber forknum,
193 : BlockNumber segno);
194 : static MdfdVec *_mdfd_openseg(SMgrRelation reln, ForkNumber forkno,
195 : BlockNumber segno, int oflags);
196 : static MdfdVec *_mdfd_getseg(SMgrRelation reln, ForkNumber forkno,
197 : BlockNumber blkno, bool skipFsync, int behavior);
198 : static BlockNumber _mdnblocks(SMgrRelation reln, ForkNumber forknum,
199 : MdfdVec *seg);
200 :
201 :
202 : /*
203 : * mdinit() -- Initialize private state for magnetic disk storage manager.
204 : */
205 : void
206 344 : mdinit(void)
207 : {
208 344 : MdCxt = AllocSetContextCreate(TopMemoryContext,
209 : "MdSmgr",
210 : ALLOCSET_DEFAULT_SIZES);
211 :
212 : /*
213 : * Create pending-operations hashtable if we need it. Currently, we need
214 : * it if we are standalone (not under a postmaster) or if we are a startup
215 : * or checkpointer auxiliary process.
216 : */
217 344 : if (!IsUnderPostmaster || AmStartupProcess() || AmCheckpointerProcess())
218 : {
219 : HASHCTL hash_ctl;
220 :
221 : /*
222 : * XXX: The checkpointer needs to add entries to the pending ops table
223 : * when absorbing fsync requests. That is done within a critical
224 : * section, which isn't usually allowed, but we make an exception. It
225 : * means that there's a theoretical possibility that you run out of
226 : * memory while absorbing fsync requests, which leads to a PANIC.
227 : * Fortunately the hash table is small so that's unlikely to happen in
228 : * practice.
229 : */
230 6 : pendingOpsCxt = AllocSetContextCreate(MdCxt,
231 : "Pending ops context",
232 : ALLOCSET_DEFAULT_SIZES);
233 6 : MemoryContextAllowInCriticalSection(pendingOpsCxt, true);
234 :
235 6 : MemSet(&hash_ctl, 0, sizeof(hash_ctl));
236 6 : hash_ctl.keysize = sizeof(RelFileNode);
237 6 : hash_ctl.entrysize = sizeof(PendingOperationEntry);
238 6 : hash_ctl.hcxt = pendingOpsCxt;
239 6 : pendingOpsTable = hash_create("Pending Ops Table",
240 : 100L,
241 : &hash_ctl,
242 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
243 6 : pendingUnlinks = NIL;
244 : }
245 344 : }
246 :
247 : /*
248 : * In archive recovery, we rely on checkpointer to do fsyncs, but we will have
249 : * already created the pendingOpsTable during initialization of the startup
250 : * process. Calling this function drops the local pendingOpsTable so that
251 : * subsequent requests will be forwarded to checkpointer.
252 : */
253 : void
254 0 : SetForwardFsyncRequests(void)
255 : {
256 : /* Perform any pending fsyncs we may have queued up, then drop table */
257 0 : if (pendingOpsTable)
258 : {
259 0 : mdsync();
260 0 : hash_destroy(pendingOpsTable);
261 : }
262 0 : pendingOpsTable = NULL;
263 :
264 : /*
265 : * We should not have any pending unlink requests, since mdunlink doesn't
266 : * queue unlink requests when isRedo.
267 : */
268 0 : Assert(pendingUnlinks == NIL);
269 0 : }
270 :
271 : /*
272 : * mdexists() -- Does the physical file exist?
273 : *
274 : * Note: this will return true for lingering files, with pending deletions
275 : */
276 : bool
277 10547 : mdexists(SMgrRelation reln, ForkNumber forkNum)
278 : {
279 : /*
280 : * Close it first, to ensure that we notice if the fork has been unlinked
281 : * since we opened it.
282 : */
283 10547 : mdclose(reln, forkNum);
284 :
285 10547 : return (mdopen(reln, forkNum, EXTENSION_RETURN_NULL) != NULL);
286 : }
287 :
288 : /*
289 : * mdcreate() -- Create a new relation on magnetic disk.
290 : *
291 : * If isRedo is true, it's okay for the relation to exist already.
292 : */
293 : void
294 4479 : mdcreate(SMgrRelation reln, ForkNumber forkNum, bool isRedo)
295 : {
296 : MdfdVec *mdfd;
297 : char *path;
298 : File fd;
299 :
300 4479 : if (isRedo && reln->md_num_open_segs[forkNum] > 0)
301 4479 : return; /* created and opened already... */
302 :
303 4479 : Assert(reln->md_num_open_segs[forkNum] == 0);
304 :
305 4479 : path = relpath(reln->smgr_rnode, forkNum);
306 :
307 4479 : fd = PathNameOpenFile(path, O_RDWR | O_CREAT | O_EXCL | PG_BINARY, 0600);
308 :
309 4479 : if (fd < 0)
310 : {
311 0 : int save_errno = errno;
312 :
313 : /*
314 : * During bootstrap, there are cases where a system relation will be
315 : * accessed (by internal backend processes) before the bootstrap
316 : * script nominally creates it. Therefore, allow the file to exist
317 : * already, even if isRedo is not set. (See also mdopen)
318 : */
319 0 : if (isRedo || IsBootstrapProcessingMode())
320 0 : fd = PathNameOpenFile(path, O_RDWR | PG_BINARY, 0600);
321 0 : if (fd < 0)
322 : {
323 : /* be sure to report the error reported by create, not open */
324 0 : errno = save_errno;
325 0 : ereport(ERROR,
326 : (errcode_for_file_access(),
327 : errmsg("could not create file \"%s\": %m", path)));
328 : }
329 : }
330 :
331 4479 : pfree(path);
332 :
333 4479 : _fdvec_resize(reln, forkNum, 1);
334 4479 : mdfd = &reln->md_seg_fds[forkNum][0];
335 4479 : mdfd->mdfd_vfd = fd;
336 4479 : mdfd->mdfd_segno = 0;
337 : }
338 :
339 : /*
340 : * mdunlink() -- Unlink a relation.
341 : *
342 : * Note that we're passed a RelFileNodeBackend --- by the time this is called,
343 : * there won't be an SMgrRelation hashtable entry anymore.
344 : *
345 : * forkNum can be a fork number to delete a specific fork, or InvalidForkNumber
346 : * to delete all forks.
347 : *
348 : * For regular relations, we don't unlink the first segment file of the rel,
349 : * but just truncate it to zero length, and record a request to unlink it after
350 : * the next checkpoint. Additional segments can be unlinked immediately,
351 : * however. Leaving the empty file in place prevents that relfilenode
352 : * number from being reused. The scenario this protects us from is:
353 : * 1. We delete a relation (and commit, and actually remove its file).
354 : * 2. We create a new relation, which by chance gets the same relfilenode as
355 : * the just-deleted one (OIDs must've wrapped around for that to happen).
356 : * 3. We crash before another checkpoint occurs.
357 : * During replay, we would delete the file and then recreate it, which is fine
358 : * if the contents of the file were repopulated by subsequent WAL entries.
359 : * But if we didn't WAL-log insertions, but instead relied on fsyncing the
360 : * file after populating it (as for instance CLUSTER and CREATE INDEX do),
361 : * the contents of the file would be lost forever. By leaving the empty file
362 : * until after the next checkpoint, we prevent reassignment of the relfilenode
363 : * number until it's safe, because relfilenode assignment skips over any
364 : * existing file.
365 : *
366 : * We do not need to go through this dance for temp relations, though, because
367 : * we never make WAL entries for temp rels, and so a temp rel poses no threat
368 : * to the health of a regular rel that has taken over its relfilenode number.
369 : * The fact that temp rels and regular rels have different file naming
370 : * patterns provides additional safety.
371 : *
372 : * All the above applies only to the relation's main fork; other forks can
373 : * just be removed immediately, since they are not needed to prevent the
374 : * relfilenode number from being recycled. Also, we do not carefully
375 : * track whether other forks have been created or not, but just attempt to
376 : * unlink them unconditionally; so we should never complain about ENOENT.
377 : *
378 : * If isRedo is true, it's unsurprising for the relation to be already gone.
379 : * Also, we should remove the file immediately instead of queuing a request
380 : * for later, since during redo there's no possibility of creating a
381 : * conflicting relation.
382 : *
383 : * Note: any failure should be reported as WARNING not ERROR, because
384 : * we are usually not in a transaction anymore when this is called.
385 : */
386 : void
387 12721 : mdunlink(RelFileNodeBackend rnode, ForkNumber forkNum, bool isRedo)
388 : {
389 : /*
390 : * We have to clean out any pending fsync requests for the doomed
391 : * relation, else the next mdsync() will fail. There can't be any such
392 : * requests for a temp relation, though. We can send just one request
393 : * even when deleting multiple forks, since the fsync queuing code accepts
394 : * the "InvalidForkNumber = all forks" convention.
395 : */
396 12721 : if (!RelFileNodeBackendIsTemp(rnode))
397 10301 : ForgetRelationFsyncRequests(rnode.node, forkNum);
398 :
399 : /* Now do the per-fork work */
400 12721 : if (forkNum == InvalidForkNumber)
401 : {
402 5 : for (forkNum = 0; forkNum <= MAX_FORKNUM; forkNum++)
403 4 : mdunlinkfork(rnode, forkNum, isRedo);
404 : }
405 : else
406 12720 : mdunlinkfork(rnode, forkNum, isRedo);
407 12721 : }
408 :
409 : static void
410 12724 : mdunlinkfork(RelFileNodeBackend rnode, ForkNumber forkNum, bool isRedo)
411 : {
412 : char *path;
413 : int ret;
414 :
415 12724 : path = relpath(rnode, forkNum);
416 :
417 : /*
418 : * Delete or truncate the first segment.
419 : */
420 12724 : if (isRedo || forkNum != MAIN_FORKNUM || RelFileNodeBackendIsTemp(rnode))
421 : {
422 10148 : ret = unlink(path);
423 20296 : if (ret < 0 && errno != ENOENT)
424 0 : ereport(WARNING,
425 : (errcode_for_file_access(),
426 : errmsg("could not remove file \"%s\": %m", path)));
427 : }
428 : else
429 : {
430 : /* truncate(2) would be easier here, but Windows hasn't got it */
431 : int fd;
432 :
433 2576 : fd = OpenTransientFile(path, O_RDWR | PG_BINARY, 0);
434 2576 : if (fd >= 0)
435 : {
436 : int save_errno;
437 :
438 2576 : ret = ftruncate(fd, 0);
439 2576 : save_errno = errno;
440 2576 : CloseTransientFile(fd);
441 2576 : errno = save_errno;
442 : }
443 : else
444 0 : ret = -1;
445 2576 : if (ret < 0 && errno != ENOENT)
446 0 : ereport(WARNING,
447 : (errcode_for_file_access(),
448 : errmsg("could not truncate file \"%s\": %m", path)));
449 :
450 : /* Register request to unlink first segment later */
451 2576 : register_unlink(rnode);
452 : }
453 :
454 : /*
455 : * Delete any additional segments.
456 : */
457 12724 : if (ret >= 0)
458 : {
459 3260 : char *segpath = (char *) palloc(strlen(path) + 12);
460 : BlockNumber segno;
461 :
462 : /*
463 : * Note that because we loop until getting ENOENT, we will correctly
464 : * remove all inactive segments as well as active ones.
465 : */
466 3260 : for (segno = 1;; segno++)
467 : {
468 3260 : sprintf(segpath, "%s.%u", path, segno);
469 3260 : if (unlink(segpath) < 0)
470 : {
471 : /* ENOENT is expected after the last segment... */
472 3260 : if (errno != ENOENT)
473 0 : ereport(WARNING,
474 : (errcode_for_file_access(),
475 : errmsg("could not remove file \"%s\": %m", segpath)));
476 3260 : break;
477 : }
478 0 : }
479 3260 : pfree(segpath);
480 : }
481 :
482 12724 : pfree(path);
483 12724 : }
484 :
485 : /*
486 : * mdextend() -- Add a block to the specified relation.
487 : *
488 : * The semantics are nearly the same as mdwrite(): write at the
489 : * specified position. However, this is to be used for the case of
490 : * extending a relation (i.e., blocknum is at or beyond the current
491 : * EOF). Note that we assume writing a block beyond current EOF
492 : * causes intervening file space to become filled with zeroes.
493 : */
494 : void
495 18586 : mdextend(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
496 : char *buffer, bool skipFsync)
497 : {
498 : off_t seekpos;
499 : int nbytes;
500 : MdfdVec *v;
501 :
502 : /* This assert is too expensive to have on normally ... */
503 : #ifdef CHECK_WRITE_VS_EXTEND
504 : Assert(blocknum >= mdnblocks(reln, forknum));
505 : #endif
506 :
507 : /*
508 : * If a relation manages to grow to 2^32-1 blocks, refuse to extend it any
509 : * more --- we mustn't create a block whose number actually is
510 : * InvalidBlockNumber.
511 : */
512 18586 : if (blocknum == InvalidBlockNumber)
513 0 : ereport(ERROR,
514 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
515 : errmsg("cannot extend file \"%s\" beyond %u blocks",
516 : relpath(reln->smgr_rnode, forknum),
517 : InvalidBlockNumber)));
518 :
519 18586 : v = _mdfd_getseg(reln, forknum, blocknum, skipFsync, EXTENSION_CREATE);
520 :
521 18586 : seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
522 :
523 18586 : Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
524 :
525 : /*
526 : * Note: because caller usually obtained blocknum by calling mdnblocks,
527 : * which did a seek(SEEK_END), this seek is often redundant and will be
528 : * optimized away by fd.c. It's not redundant, however, if there is a
529 : * partial page at the end of the file. In that case we want to try to
530 : * overwrite the partial page with a full page. It's also not redundant
531 : * if bufmgr.c had to dump another buffer of the same file to make room
532 : * for the new page's buffer.
533 : */
534 18586 : if (FileSeek(v->mdfd_vfd, seekpos, SEEK_SET) != seekpos)
535 0 : ereport(ERROR,
536 : (errcode_for_file_access(),
537 : errmsg("could not seek to block %u in file \"%s\": %m",
538 : blocknum, FilePathName(v->mdfd_vfd))));
539 :
540 18586 : if ((nbytes = FileWrite(v->mdfd_vfd, buffer, BLCKSZ, WAIT_EVENT_DATA_FILE_EXTEND)) != BLCKSZ)
541 : {
542 0 : if (nbytes < 0)
543 0 : ereport(ERROR,
544 : (errcode_for_file_access(),
545 : errmsg("could not extend file \"%s\": %m",
546 : FilePathName(v->mdfd_vfd)),
547 : errhint("Check free disk space.")));
548 : /* short write: complain appropriately */
549 0 : ereport(ERROR,
550 : (errcode(ERRCODE_DISK_FULL),
551 : errmsg("could not extend file \"%s\": wrote only %d of %d bytes at block %u",
552 : FilePathName(v->mdfd_vfd),
553 : nbytes, BLCKSZ, blocknum),
554 : errhint("Check free disk space.")));
555 : }
556 :
557 18586 : if (!skipFsync && !SmgrIsTemp(reln))
558 13668 : register_dirty_segment(reln, forknum, v);
559 :
560 18586 : Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
561 18586 : }
562 :
563 : /*
564 : * mdopen() -- Open the specified relation.
565 : *
566 : * Note we only open the first segment, when there are multiple segments.
567 : *
568 : * If first segment is not present, either ereport or return NULL according
569 : * to "behavior". We treat EXTENSION_CREATE the same as EXTENSION_FAIL;
570 : * EXTENSION_CREATE means it's OK to extend an existing relation, not to
571 : * invent one out of whole cloth.
572 : */
573 : static MdfdVec *
574 97554 : mdopen(SMgrRelation reln, ForkNumber forknum, int behavior)
575 : {
576 : MdfdVec *mdfd;
577 : char *path;
578 : File fd;
579 :
580 : /* No work if already open */
581 97554 : if (reln->md_num_open_segs[forknum] > 0)
582 75597 : return &reln->md_seg_fds[forknum][0];
583 :
584 21957 : path = relpath(reln->smgr_rnode, forknum);
585 :
586 21957 : fd = PathNameOpenFile(path, O_RDWR | PG_BINARY, 0600);
587 :
588 21957 : if (fd < 0)
589 : {
590 : /*
591 : * During bootstrap, there are cases where a system relation will be
592 : * accessed (by internal backend processes) before the bootstrap
593 : * script nominally creates it. Therefore, accept mdopen() as a
594 : * substitute for mdcreate() in bootstrap mode only. (See mdcreate)
595 : */
596 7405 : if (IsBootstrapProcessingMode())
597 100 : fd = PathNameOpenFile(path, O_RDWR | O_CREAT | O_EXCL | PG_BINARY, 0600);
598 7405 : if (fd < 0)
599 : {
600 14610 : if ((behavior & EXTENSION_RETURN_NULL) &&
601 7305 : FILE_POSSIBLY_DELETED(errno))
602 : {
603 7305 : pfree(path);
604 7305 : return NULL;
605 : }
606 0 : ereport(ERROR,
607 : (errcode_for_file_access(),
608 : errmsg("could not open file \"%s\": %m", path)));
609 : }
610 : }
611 :
612 14652 : pfree(path);
613 :
614 14652 : _fdvec_resize(reln, forknum, 1);
615 14652 : mdfd = &reln->md_seg_fds[forknum][0];
616 14652 : mdfd->mdfd_vfd = fd;
617 14652 : mdfd->mdfd_segno = 0;
618 :
619 14652 : Assert(_mdnblocks(reln, forknum, mdfd) <= ((BlockNumber) RELSEG_SIZE));
620 :
621 14652 : return mdfd;
622 : }
623 :
624 : /*
625 : * mdclose() -- Close the specified relation, if it isn't closed already.
626 : */
627 : void
628 90479 : mdclose(SMgrRelation reln, ForkNumber forknum)
629 : {
630 90479 : int nopensegs = reln->md_num_open_segs[forknum];
631 :
632 : /* No work if already closed */
633 90479 : if (nopensegs == 0)
634 169023 : return;
635 :
636 : /* close segments starting from the end */
637 35805 : while (nopensegs > 0)
638 : {
639 11935 : MdfdVec *v = &reln->md_seg_fds[forknum][nopensegs - 1];
640 :
641 : /* if not closed already */
642 11935 : if (v->mdfd_vfd >= 0)
643 : {
644 11935 : FileClose(v->mdfd_vfd);
645 11935 : v->mdfd_vfd = -1;
646 : }
647 :
648 11935 : nopensegs--;
649 : }
650 :
651 : /* resize just once, avoids pointless reallocations */
652 11935 : _fdvec_resize(reln, forknum, 0);
653 : }
654 :
655 : /*
656 : * mdprefetch() -- Initiate asynchronous read of the specified block of a relation
657 : */
658 : void
659 0 : mdprefetch(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum)
660 : {
661 : #ifdef USE_PREFETCH
662 : off_t seekpos;
663 : MdfdVec *v;
664 :
665 0 : v = _mdfd_getseg(reln, forknum, blocknum, false, EXTENSION_FAIL);
666 :
667 0 : seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
668 :
669 0 : Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
670 :
671 0 : (void) FilePrefetch(v->mdfd_vfd, seekpos, BLCKSZ, WAIT_EVENT_DATA_FILE_PREFETCH);
672 : #endif /* USE_PREFETCH */
673 0 : }
674 :
675 : /*
676 : * mdwriteback() -- Tell the kernel to write pages back to storage.
677 : *
678 : * This accepts a range of blocks because flushing several pages at once is
679 : * considerably more efficient than doing so individually.
680 : */
681 : void
682 1477 : mdwriteback(SMgrRelation reln, ForkNumber forknum,
683 : BlockNumber blocknum, BlockNumber nblocks)
684 : {
685 : /*
686 : * Issue flush requests in as few requests as possible; have to split at
687 : * segment boundaries though, since those are actually separate files.
688 : */
689 4431 : while (nblocks > 0)
690 : {
691 1477 : BlockNumber nflush = nblocks;
692 : off_t seekpos;
693 : MdfdVec *v;
694 : int segnum_start,
695 : segnum_end;
696 :
697 1477 : v = _mdfd_getseg(reln, forknum, blocknum, true /* not used */ ,
698 : EXTENSION_RETURN_NULL);
699 :
700 : /*
701 : * We might be flushing buffers of already removed relations, that's
702 : * ok, just ignore that case.
703 : */
704 1477 : if (!v)
705 1477 : return;
706 :
707 : /* compute offset inside the current segment */
708 1477 : segnum_start = blocknum / RELSEG_SIZE;
709 :
710 : /* compute number of desired writes within the current segment */
711 1477 : segnum_end = (blocknum + nblocks - 1) / RELSEG_SIZE;
712 1477 : if (segnum_start != segnum_end)
713 0 : nflush = RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE));
714 :
715 1477 : Assert(nflush >= 1);
716 1477 : Assert(nflush <= nblocks);
717 :
718 1477 : seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
719 :
720 1477 : FileWriteback(v->mdfd_vfd, seekpos, (off_t) BLCKSZ * nflush, WAIT_EVENT_DATA_FILE_FLUSH);
721 :
722 1477 : nblocks -= nflush;
723 1477 : blocknum += nflush;
724 : }
725 : }
726 :
727 : /*
728 : * mdread() -- Read the specified block from a relation.
729 : */
730 : void
731 4439 : mdread(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
732 : char *buffer)
733 : {
734 : off_t seekpos;
735 : int nbytes;
736 : MdfdVec *v;
737 :
738 : TRACE_POSTGRESQL_SMGR_MD_READ_START(forknum, blocknum,
739 : reln->smgr_rnode.node.spcNode,
740 : reln->smgr_rnode.node.dbNode,
741 : reln->smgr_rnode.node.relNode,
742 : reln->smgr_rnode.backend);
743 :
744 4439 : v = _mdfd_getseg(reln, forknum, blocknum, false,
745 : EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
746 :
747 4439 : seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
748 :
749 4439 : Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
750 :
751 4439 : if (FileSeek(v->mdfd_vfd, seekpos, SEEK_SET) != seekpos)
752 0 : ereport(ERROR,
753 : (errcode_for_file_access(),
754 : errmsg("could not seek to block %u in file \"%s\": %m",
755 : blocknum, FilePathName(v->mdfd_vfd))));
756 :
757 4439 : nbytes = FileRead(v->mdfd_vfd, buffer, BLCKSZ, WAIT_EVENT_DATA_FILE_READ);
758 :
759 : TRACE_POSTGRESQL_SMGR_MD_READ_DONE(forknum, blocknum,
760 : reln->smgr_rnode.node.spcNode,
761 : reln->smgr_rnode.node.dbNode,
762 : reln->smgr_rnode.node.relNode,
763 : reln->smgr_rnode.backend,
764 : nbytes,
765 : BLCKSZ);
766 :
767 4439 : if (nbytes != BLCKSZ)
768 : {
769 0 : if (nbytes < 0)
770 0 : ereport(ERROR,
771 : (errcode_for_file_access(),
772 : errmsg("could not read block %u in file \"%s\": %m",
773 : blocknum, FilePathName(v->mdfd_vfd))));
774 :
775 : /*
776 : * Short read: we are at or past EOF, or we read a partial block at
777 : * EOF. Normally this is an error; upper levels should never try to
778 : * read a nonexistent block. However, if zero_damaged_pages is ON or
779 : * we are InRecovery, we should instead return zeroes without
780 : * complaining. This allows, for example, the case of trying to
781 : * update a block that was later truncated away.
782 : */
783 0 : if (zero_damaged_pages || InRecovery)
784 0 : MemSet(buffer, 0, BLCKSZ);
785 : else
786 0 : ereport(ERROR,
787 : (errcode(ERRCODE_DATA_CORRUPTED),
788 : errmsg("could not read block %u in file \"%s\": read only %d of %d bytes",
789 : blocknum, FilePathName(v->mdfd_vfd),
790 : nbytes, BLCKSZ)));
791 : }
792 4439 : }
793 :
794 : /*
795 : * mdwrite() -- Write the supplied block at the appropriate location.
796 : *
797 : * This is to be used only for updating already-existing blocks of a
798 : * relation (ie, those before the current EOF). To extend a relation,
799 : * use mdextend().
800 : */
801 : void
802 8838 : mdwrite(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
803 : char *buffer, bool skipFsync)
804 : {
805 : off_t seekpos;
806 : int nbytes;
807 : MdfdVec *v;
808 :
809 : /* This assert is too expensive to have on normally ... */
810 : #ifdef CHECK_WRITE_VS_EXTEND
811 : Assert(blocknum < mdnblocks(reln, forknum));
812 : #endif
813 :
814 : TRACE_POSTGRESQL_SMGR_MD_WRITE_START(forknum, blocknum,
815 : reln->smgr_rnode.node.spcNode,
816 : reln->smgr_rnode.node.dbNode,
817 : reln->smgr_rnode.node.relNode,
818 : reln->smgr_rnode.backend);
819 :
820 8838 : v = _mdfd_getseg(reln, forknum, blocknum, skipFsync,
821 : EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
822 :
823 8838 : seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
824 :
825 8838 : Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
826 :
827 8838 : if (FileSeek(v->mdfd_vfd, seekpos, SEEK_SET) != seekpos)
828 0 : ereport(ERROR,
829 : (errcode_for_file_access(),
830 : errmsg("could not seek to block %u in file \"%s\": %m",
831 : blocknum, FilePathName(v->mdfd_vfd))));
832 :
833 8838 : nbytes = FileWrite(v->mdfd_vfd, buffer, BLCKSZ, WAIT_EVENT_DATA_FILE_WRITE);
834 :
835 : TRACE_POSTGRESQL_SMGR_MD_WRITE_DONE(forknum, blocknum,
836 : reln->smgr_rnode.node.spcNode,
837 : reln->smgr_rnode.node.dbNode,
838 : reln->smgr_rnode.node.relNode,
839 : reln->smgr_rnode.backend,
840 : nbytes,
841 : BLCKSZ);
842 :
843 8838 : if (nbytes != BLCKSZ)
844 : {
845 0 : if (nbytes < 0)
846 0 : ereport(ERROR,
847 : (errcode_for_file_access(),
848 : errmsg("could not write block %u in file \"%s\": %m",
849 : blocknum, FilePathName(v->mdfd_vfd))));
850 : /* short write: complain appropriately */
851 0 : ereport(ERROR,
852 : (errcode(ERRCODE_DISK_FULL),
853 : errmsg("could not write block %u in file \"%s\": wrote only %d of %d bytes",
854 : blocknum,
855 : FilePathName(v->mdfd_vfd),
856 : nbytes, BLCKSZ),
857 : errhint("Check free disk space.")));
858 : }
859 :
860 8838 : if (!skipFsync && !SmgrIsTemp(reln))
861 8542 : register_dirty_segment(reln, forknum, v);
862 8838 : }
863 :
864 : /*
865 : * mdnblocks() -- Get the number of blocks stored in a relation.
866 : *
867 : * Important side effect: all active segments of the relation are opened
868 : * and added to the mdfd_seg_fds array. If this routine has not been
869 : * called, then only segments up to the last one actually touched
870 : * are present in the array.
871 : */
872 : BlockNumber
873 85461 : mdnblocks(SMgrRelation reln, ForkNumber forknum)
874 : {
875 85461 : MdfdVec *v = mdopen(reln, forknum, EXTENSION_FAIL);
876 : BlockNumber nblocks;
877 85461 : BlockNumber segno = 0;
878 :
879 : /* mdopen has opened the first segment */
880 85461 : Assert(reln->md_num_open_segs[forknum] > 0);
881 :
882 : /*
883 : * Start from the last open segments, to avoid redundant seeks. We have
884 : * previously verified that these segments are exactly RELSEG_SIZE long,
885 : * and it's useless to recheck that each time.
886 : *
887 : * NOTE: this assumption could only be wrong if another backend has
888 : * truncated the relation. We rely on higher code levels to handle that
889 : * scenario by closing and re-opening the md fd, which is handled via
890 : * relcache flush. (Since the checkpointer doesn't participate in
891 : * relcache flush, it could have segment entries for inactive segments;
892 : * that's OK because the checkpointer never needs to compute relation
893 : * size.)
894 : */
895 85461 : segno = reln->md_num_open_segs[forknum] - 1;
896 85461 : v = &reln->md_seg_fds[forknum][segno];
897 :
898 : for (;;)
899 : {
900 85461 : nblocks = _mdnblocks(reln, forknum, v);
901 85461 : if (nblocks > ((BlockNumber) RELSEG_SIZE))
902 0 : elog(FATAL, "segment too big");
903 85461 : if (nblocks < ((BlockNumber) RELSEG_SIZE))
904 85461 : return (segno * ((BlockNumber) RELSEG_SIZE)) + nblocks;
905 :
906 : /*
907 : * If segment is exactly RELSEG_SIZE, advance to next one.
908 : */
909 0 : segno++;
910 :
911 : /*
912 : * We used to pass O_CREAT here, but that's has the disadvantage that
913 : * it might create a segment which has vanished through some operating
914 : * system misadventure. In such a case, creating the segment here
915 : * undermines _mdfd_getseg's attempts to notice and report an error
916 : * upon access to a missing segment.
917 : */
918 0 : v = _mdfd_openseg(reln, forknum, segno, 0);
919 0 : if (v == NULL)
920 0 : return segno * ((BlockNumber) RELSEG_SIZE);
921 0 : }
922 : }
923 :
924 : /*
925 : * mdtruncate() -- Truncate relation to specified number of blocks.
926 : */
927 : void
928 46 : mdtruncate(SMgrRelation reln, ForkNumber forknum, BlockNumber nblocks)
929 : {
930 : BlockNumber curnblk;
931 : BlockNumber priorblocks;
932 : int curopensegs;
933 :
934 : /*
935 : * NOTE: mdnblocks makes sure we have opened all active segments, so that
936 : * truncation loop will get them all!
937 : */
938 46 : curnblk = mdnblocks(reln, forknum);
939 46 : if (nblocks > curnblk)
940 : {
941 : /* Bogus request ... but no complaint if InRecovery */
942 0 : if (InRecovery)
943 0 : return;
944 0 : ereport(ERROR,
945 : (errmsg("could not truncate file \"%s\" to %u blocks: it's only %u blocks now",
946 : relpath(reln->smgr_rnode, forknum),
947 : nblocks, curnblk)));
948 : }
949 46 : if (nblocks == curnblk)
950 21 : return; /* no work */
951 :
952 : /*
953 : * Truncate segments, starting at the last one. Starting at the end makes
954 : * managing the memory for the fd array easier, should there be errors.
955 : */
956 25 : curopensegs = reln->md_num_open_segs[forknum];
957 75 : while (curopensegs > 0)
958 : {
959 : MdfdVec *v;
960 :
961 25 : priorblocks = (curopensegs - 1) * RELSEG_SIZE;
962 :
963 25 : v = &reln->md_seg_fds[forknum][curopensegs - 1];
964 :
965 25 : if (priorblocks > nblocks)
966 : {
967 : /*
968 : * This segment is no longer active. We truncate the file, but do
969 : * not delete it, for reasons explained in the header comments.
970 : */
971 0 : if (FileTruncate(v->mdfd_vfd, 0, WAIT_EVENT_DATA_FILE_TRUNCATE) < 0)
972 0 : ereport(ERROR,
973 : (errcode_for_file_access(),
974 : errmsg("could not truncate file \"%s\": %m",
975 : FilePathName(v->mdfd_vfd))));
976 :
977 0 : if (!SmgrIsTemp(reln))
978 0 : register_dirty_segment(reln, forknum, v);
979 :
980 : /* we never drop the 1st segment */
981 0 : Assert(v != &reln->md_seg_fds[forknum][0]);
982 :
983 0 : FileClose(v->mdfd_vfd);
984 0 : _fdvec_resize(reln, forknum, curopensegs - 1);
985 : }
986 25 : else if (priorblocks + ((BlockNumber) RELSEG_SIZE) > nblocks)
987 : {
988 : /*
989 : * This is the last segment we want to keep. Truncate the file to
990 : * the right length. NOTE: if nblocks is exactly a multiple K of
991 : * RELSEG_SIZE, we will truncate the K+1st segment to 0 length but
992 : * keep it. This adheres to the invariant given in the header
993 : * comments.
994 : */
995 25 : BlockNumber lastsegblocks = nblocks - priorblocks;
996 :
997 25 : if (FileTruncate(v->mdfd_vfd, (off_t) lastsegblocks * BLCKSZ, WAIT_EVENT_DATA_FILE_TRUNCATE) < 0)
998 0 : ereport(ERROR,
999 : (errcode_for_file_access(),
1000 : errmsg("could not truncate file \"%s\" to %u blocks: %m",
1001 : FilePathName(v->mdfd_vfd),
1002 : nblocks)));
1003 25 : if (!SmgrIsTemp(reln))
1004 20 : register_dirty_segment(reln, forknum, v);
1005 : }
1006 : else
1007 : {
1008 : /*
1009 : * We still need this segment, so nothing to do for this and any
1010 : * earlier segment.
1011 : */
1012 0 : break;
1013 : }
1014 25 : curopensegs--;
1015 : }
1016 : }
1017 :
1018 : /*
1019 : * mdimmedsync() -- Immediately sync a relation to stable storage.
1020 : *
1021 : * Note that only writes already issued are synced; this routine knows
1022 : * nothing of dirty buffers that may exist inside the buffer manager.
1023 : */
1024 : void
1025 1264 : mdimmedsync(SMgrRelation reln, ForkNumber forknum)
1026 : {
1027 : int segno;
1028 :
1029 : /*
1030 : * NOTE: mdnblocks makes sure we have opened all active segments, so that
1031 : * fsync loop will get them all!
1032 : */
1033 1264 : mdnblocks(reln, forknum);
1034 :
1035 1264 : segno = reln->md_num_open_segs[forknum];
1036 :
1037 3792 : while (segno > 0)
1038 : {
1039 1264 : MdfdVec *v = &reln->md_seg_fds[forknum][segno - 1];
1040 :
1041 1264 : if (FileSync(v->mdfd_vfd, WAIT_EVENT_DATA_FILE_IMMEDIATE_SYNC) < 0)
1042 0 : ereport(ERROR,
1043 : (errcode_for_file_access(),
1044 : errmsg("could not fsync file \"%s\": %m",
1045 : FilePathName(v->mdfd_vfd))));
1046 1264 : segno--;
1047 : }
1048 1264 : }
1049 :
1050 : /*
1051 : * mdsync() -- Sync previous writes to stable storage.
1052 : */
1053 : void
1054 11 : mdsync(void)
1055 : {
1056 : static bool mdsync_in_progress = false;
1057 :
1058 : HASH_SEQ_STATUS hstat;
1059 : PendingOperationEntry *entry;
1060 : int absorb_counter;
1061 :
1062 : /* Statistics on sync times */
1063 11 : int processed = 0;
1064 : instr_time sync_start,
1065 : sync_end,
1066 : sync_diff;
1067 : uint64 elapsed;
1068 11 : uint64 longest = 0;
1069 11 : uint64 total_elapsed = 0;
1070 :
1071 : /*
1072 : * This is only called during checkpoints, and checkpoints should only
1073 : * occur in processes that have created a pendingOpsTable.
1074 : */
1075 11 : if (!pendingOpsTable)
1076 0 : elog(ERROR, "cannot sync without a pendingOpsTable");
1077 :
1078 : /*
1079 : * If we are in the checkpointer, the sync had better include all fsync
1080 : * requests that were queued by backends up to this point. The tightest
1081 : * race condition that could occur is that a buffer that must be written
1082 : * and fsync'd for the checkpoint could have been dumped by a backend just
1083 : * before it was visited by BufferSync(). We know the backend will have
1084 : * queued an fsync request before clearing the buffer's dirtybit, so we
1085 : * are safe as long as we do an Absorb after completing BufferSync().
1086 : */
1087 11 : AbsorbFsyncRequests();
1088 :
1089 : /*
1090 : * To avoid excess fsync'ing (in the worst case, maybe a never-terminating
1091 : * checkpoint), we want to ignore fsync requests that are entered into the
1092 : * hashtable after this point --- they should be processed next time,
1093 : * instead. We use mdsync_cycle_ctr to tell old entries apart from new
1094 : * ones: new ones will have cycle_ctr equal to the incremented value of
1095 : * mdsync_cycle_ctr.
1096 : *
1097 : * In normal circumstances, all entries present in the table at this point
1098 : * will have cycle_ctr exactly equal to the current (about to be old)
1099 : * value of mdsync_cycle_ctr. However, if we fail partway through the
1100 : * fsync'ing loop, then older values of cycle_ctr might remain when we
1101 : * come back here to try again. Repeated checkpoint failures would
1102 : * eventually wrap the counter around to the point where an old entry
1103 : * might appear new, causing us to skip it, possibly allowing a checkpoint
1104 : * to succeed that should not have. To forestall wraparound, any time the
1105 : * previous mdsync() failed to complete, run through the table and
1106 : * forcibly set cycle_ctr = mdsync_cycle_ctr.
1107 : *
1108 : * Think not to merge this loop with the main loop, as the problem is
1109 : * exactly that that loop may fail before having visited all the entries.
1110 : * From a performance point of view it doesn't matter anyway, as this path
1111 : * will never be taken in a system that's functioning normally.
1112 : */
1113 11 : if (mdsync_in_progress)
1114 : {
1115 : /* prior try failed, so update any stale cycle_ctr values */
1116 0 : hash_seq_init(&hstat, pendingOpsTable);
1117 0 : while ((entry = (PendingOperationEntry *) hash_seq_search(&hstat)) != NULL)
1118 : {
1119 0 : entry->cycle_ctr = mdsync_cycle_ctr;
1120 : }
1121 : }
1122 :
1123 : /* Advance counter so that new hashtable entries are distinguishable */
1124 11 : mdsync_cycle_ctr++;
1125 :
1126 : /* Set flag to detect failure if we don't reach the end of the loop */
1127 11 : mdsync_in_progress = true;
1128 :
1129 : /* Now scan the hashtable for fsync requests to process */
1130 11 : absorb_counter = FSYNCS_PER_ABSORB;
1131 11 : hash_seq_init(&hstat, pendingOpsTable);
1132 11 : while ((entry = (PendingOperationEntry *) hash_seq_search(&hstat)) != NULL)
1133 : {
1134 : ForkNumber forknum;
1135 :
1136 : /*
1137 : * If the entry is new then don't process it this time; it might
1138 : * contain multiple fsync-request bits, but they are all new. Note
1139 : * "continue" bypasses the hash-remove call at the bottom of the loop.
1140 : */
1141 1630 : if (entry->cycle_ctr == mdsync_cycle_ctr)
1142 0 : continue;
1143 :
1144 : /* Else assert we haven't missed it */
1145 1630 : Assert((CycleCtr) (entry->cycle_ctr + 1) == mdsync_cycle_ctr);
1146 :
1147 : /*
1148 : * Scan over the forks and segments represented by the entry.
1149 : *
1150 : * The bitmap manipulations are slightly tricky, because we can call
1151 : * AbsorbFsyncRequests() inside the loop and that could result in
1152 : * bms_add_member() modifying and even re-palloc'ing the bitmapsets.
1153 : * This is okay because we unlink each bitmapset from the hashtable
1154 : * entry before scanning it. That means that any incoming fsync
1155 : * requests will be processed now if they reach the table before we
1156 : * begin to scan their fork.
1157 : */
1158 8150 : for (forknum = 0; forknum <= MAX_FORKNUM; forknum++)
1159 : {
1160 6520 : Bitmapset *requests = entry->requests[forknum];
1161 : int segno;
1162 :
1163 6520 : entry->requests[forknum] = NULL;
1164 6520 : entry->canceled[forknum] = false;
1165 :
1166 14214 : while ((segno = bms_first_member(requests)) >= 0)
1167 : {
1168 : int failures;
1169 :
1170 : /*
1171 : * If fsync is off then we don't have to bother opening the
1172 : * file at all. (We delay checking until this point so that
1173 : * changing fsync on the fly behaves sensibly.)
1174 : */
1175 1174 : if (!enableFsync)
1176 1174 : continue;
1177 :
1178 : /*
1179 : * If in checkpointer, we want to absorb pending requests
1180 : * every so often to prevent overflow of the fsync request
1181 : * queue. It is unspecified whether newly-added entries will
1182 : * be visited by hash_seq_search, but we don't care since we
1183 : * don't need to process them anyway.
1184 : */
1185 0 : if (--absorb_counter <= 0)
1186 : {
1187 0 : AbsorbFsyncRequests();
1188 0 : absorb_counter = FSYNCS_PER_ABSORB;
1189 : }
1190 :
1191 : /*
1192 : * The fsync table could contain requests to fsync segments
1193 : * that have been deleted (unlinked) by the time we get to
1194 : * them. Rather than just hoping an ENOENT (or EACCES on
1195 : * Windows) error can be ignored, what we do on error is
1196 : * absorb pending requests and then retry. Since mdunlink()
1197 : * queues a "cancel" message before actually unlinking, the
1198 : * fsync request is guaranteed to be marked canceled after the
1199 : * absorb if it really was this case. DROP DATABASE likewise
1200 : * has to tell us to forget fsync requests before it starts
1201 : * deletions.
1202 : */
1203 0 : for (failures = 0;; failures++) /* loop exits at "break" */
1204 : {
1205 : SMgrRelation reln;
1206 : MdfdVec *seg;
1207 : char *path;
1208 : int save_errno;
1209 :
1210 : /*
1211 : * Find or create an smgr hash entry for this relation.
1212 : * This may seem a bit unclean -- md calling smgr? But
1213 : * it's really the best solution. It ensures that the
1214 : * open file reference isn't permanently leaked if we get
1215 : * an error here. (You may say "but an unreferenced
1216 : * SMgrRelation is still a leak!" Not really, because the
1217 : * only case in which a checkpoint is done by a process
1218 : * that isn't about to shut down is in the checkpointer,
1219 : * and it will periodically do smgrcloseall(). This fact
1220 : * justifies our not closing the reln in the success path
1221 : * either, which is a good thing since in non-checkpointer
1222 : * cases we couldn't safely do that.)
1223 : */
1224 0 : reln = smgropen(entry->rnode, InvalidBackendId);
1225 :
1226 : /* Attempt to open and fsync the target segment */
1227 0 : seg = _mdfd_getseg(reln, forknum,
1228 0 : (BlockNumber) segno * (BlockNumber) RELSEG_SIZE,
1229 : false,
1230 : EXTENSION_RETURN_NULL
1231 : | EXTENSION_DONT_CHECK_SIZE);
1232 :
1233 0 : INSTR_TIME_SET_CURRENT(sync_start);
1234 :
1235 0 : if (seg != NULL &&
1236 0 : FileSync(seg->mdfd_vfd, WAIT_EVENT_DATA_FILE_SYNC) >= 0)
1237 : {
1238 : /* Success; update statistics about sync timing */
1239 0 : INSTR_TIME_SET_CURRENT(sync_end);
1240 0 : sync_diff = sync_end;
1241 0 : INSTR_TIME_SUBTRACT(sync_diff, sync_start);
1242 0 : elapsed = INSTR_TIME_GET_MICROSEC(sync_diff);
1243 0 : if (elapsed > longest)
1244 0 : longest = elapsed;
1245 0 : total_elapsed += elapsed;
1246 0 : processed++;
1247 0 : if (log_checkpoints)
1248 0 : elog(DEBUG1, "checkpoint sync: number=%d file=%s time=%.3f msec",
1249 : processed,
1250 : FilePathName(seg->mdfd_vfd),
1251 : (double) elapsed / 1000);
1252 :
1253 0 : break; /* out of retry loop */
1254 : }
1255 :
1256 : /* Compute file name for use in message */
1257 0 : save_errno = errno;
1258 0 : path = _mdfd_segpath(reln, forknum, (BlockNumber) segno);
1259 0 : errno = save_errno;
1260 :
1261 : /*
1262 : * It is possible that the relation has been dropped or
1263 : * truncated since the fsync request was entered.
1264 : * Therefore, allow ENOENT, but only if we didn't fail
1265 : * already on this file. This applies both for
1266 : * _mdfd_getseg() and for FileSync, since fd.c might have
1267 : * closed the file behind our back.
1268 : *
1269 : * XXX is there any point in allowing more than one retry?
1270 : * Don't see one at the moment, but easy to change the
1271 : * test here if so.
1272 : */
1273 0 : if (!FILE_POSSIBLY_DELETED(errno) ||
1274 : failures > 0)
1275 0 : ereport(ERROR,
1276 : (errcode_for_file_access(),
1277 : errmsg("could not fsync file \"%s\": %m",
1278 : path)));
1279 : else
1280 0 : ereport(DEBUG1,
1281 : (errcode_for_file_access(),
1282 : errmsg("could not fsync file \"%s\" but retrying: %m",
1283 : path)));
1284 0 : pfree(path);
1285 :
1286 : /*
1287 : * Absorb incoming requests and check to see if a cancel
1288 : * arrived for this relation fork.
1289 : */
1290 0 : AbsorbFsyncRequests();
1291 0 : absorb_counter = FSYNCS_PER_ABSORB; /* might as well... */
1292 :
1293 0 : if (entry->canceled[forknum])
1294 0 : break;
1295 0 : } /* end retry loop */
1296 : }
1297 6520 : bms_free(requests);
1298 : }
1299 :
1300 : /*
1301 : * We've finished everything that was requested before we started to
1302 : * scan the entry. If no new requests have been inserted meanwhile,
1303 : * remove the entry. Otherwise, update its cycle counter, as all the
1304 : * requests now in it must have arrived during this cycle.
1305 : */
1306 8150 : for (forknum = 0; forknum <= MAX_FORKNUM; forknum++)
1307 : {
1308 6520 : if (entry->requests[forknum] != NULL)
1309 0 : break;
1310 : }
1311 1630 : if (forknum <= MAX_FORKNUM)
1312 0 : entry->cycle_ctr = mdsync_cycle_ctr;
1313 : else
1314 : {
1315 : /* Okay to remove it */
1316 1630 : if (hash_search(pendingOpsTable, &entry->rnode,
1317 : HASH_REMOVE, NULL) == NULL)
1318 0 : elog(ERROR, "pendingOpsTable corrupted");
1319 : }
1320 : } /* end loop over hashtable entries */
1321 :
1322 : /* Return sync performance metrics for report at checkpoint end */
1323 11 : CheckpointStats.ckpt_sync_rels = processed;
1324 11 : CheckpointStats.ckpt_longest_sync = longest;
1325 11 : CheckpointStats.ckpt_agg_sync_time = total_elapsed;
1326 :
1327 : /* Flag successful completion of mdsync */
1328 11 : mdsync_in_progress = false;
1329 11 : }
1330 :
1331 : /*
1332 : * mdpreckpt() -- Do pre-checkpoint work
1333 : *
1334 : * To distinguish unlink requests that arrived before this checkpoint
1335 : * started from those that arrived during the checkpoint, we use a cycle
1336 : * counter similar to the one we use for fsync requests. That cycle
1337 : * counter is incremented here.
1338 : *
1339 : * This must be called *before* the checkpoint REDO point is determined.
1340 : * That ensures that we won't delete files too soon.
1341 : *
1342 : * Note that we can't do anything here that depends on the assumption
1343 : * that the checkpoint will be completed.
1344 : */
1345 : void
1346 11 : mdpreckpt(void)
1347 : {
1348 : /*
1349 : * Any unlink requests arriving after this point will be assigned the next
1350 : * cycle counter, and won't be unlinked until next checkpoint.
1351 : */
1352 11 : mdckpt_cycle_ctr++;
1353 11 : }
1354 :
1355 : /*
1356 : * mdpostckpt() -- Do post-checkpoint work
1357 : *
1358 : * Remove any lingering files that can now be safely removed.
1359 : */
1360 : void
1361 11 : mdpostckpt(void)
1362 : {
1363 : int absorb_counter;
1364 :
1365 11 : absorb_counter = UNLINKS_PER_ABSORB;
1366 2598 : while (pendingUnlinks != NIL)
1367 : {
1368 2576 : PendingUnlinkEntry *entry = (PendingUnlinkEntry *) linitial(pendingUnlinks);
1369 : char *path;
1370 :
1371 : /*
1372 : * New entries are appended to the end, so if the entry is new we've
1373 : * reached the end of old entries.
1374 : *
1375 : * Note: if just the right number of consecutive checkpoints fail, we
1376 : * could be fooled here by cycle_ctr wraparound. However, the only
1377 : * consequence is that we'd delay unlinking for one more checkpoint,
1378 : * which is perfectly tolerable.
1379 : */
1380 2576 : if (entry->cycle_ctr == mdckpt_cycle_ctr)
1381 0 : break;
1382 :
1383 : /* Unlink the file */
1384 2576 : path = relpathperm(entry->rnode, MAIN_FORKNUM);
1385 2576 : if (unlink(path) < 0)
1386 : {
1387 : /*
1388 : * There's a race condition, when the database is dropped at the
1389 : * same time that we process the pending unlink requests. If the
1390 : * DROP DATABASE deletes the file before we do, we will get ENOENT
1391 : * here. rmtree() also has to ignore ENOENT errors, to deal with
1392 : * the possibility that we delete the file first.
1393 : */
1394 0 : if (errno != ENOENT)
1395 0 : ereport(WARNING,
1396 : (errcode_for_file_access(),
1397 : errmsg("could not remove file \"%s\": %m", path)));
1398 : }
1399 2576 : pfree(path);
1400 :
1401 : /* And remove the list entry */
1402 2576 : pendingUnlinks = list_delete_first(pendingUnlinks);
1403 2576 : pfree(entry);
1404 :
1405 : /*
1406 : * As in mdsync, we don't want to stop absorbing fsync requests for a
1407 : * long time when there are many deletions to be done. We can safely
1408 : * call AbsorbFsyncRequests() at this point in the loop (note it might
1409 : * try to delete list entries).
1410 : */
1411 2576 : if (--absorb_counter <= 0)
1412 : {
1413 256 : AbsorbFsyncRequests();
1414 256 : absorb_counter = UNLINKS_PER_ABSORB;
1415 : }
1416 : }
1417 11 : }
1418 :
1419 : /*
1420 : * register_dirty_segment() -- Mark a relation segment as needing fsync
1421 : *
1422 : * If there is a local pending-ops table, just make an entry in it for
1423 : * mdsync to process later. Otherwise, try to pass off the fsync request
1424 : * to the checkpointer process. If that fails, just do the fsync
1425 : * locally before returning (we hope this will not happen often enough
1426 : * to be a performance problem).
1427 : */
1428 : static void
1429 22230 : register_dirty_segment(SMgrRelation reln, ForkNumber forknum, MdfdVec *seg)
1430 : {
1431 : /* Temp relations should never be fsync'd */
1432 22230 : Assert(!SmgrIsTemp(reln));
1433 :
1434 22230 : if (pendingOpsTable)
1435 : {
1436 : /* push it into local pending-ops table */
1437 9205 : RememberFsyncRequest(reln->smgr_rnode.node, forknum, seg->mdfd_segno);
1438 : }
1439 : else
1440 : {
1441 13025 : if (ForwardFsyncRequest(reln->smgr_rnode.node, forknum, seg->mdfd_segno))
1442 35255 : return; /* passed it off successfully */
1443 :
1444 0 : ereport(DEBUG1,
1445 : (errmsg("could not forward fsync request because request queue is full")));
1446 :
1447 0 : if (FileSync(seg->mdfd_vfd, WAIT_EVENT_DATA_FILE_SYNC) < 0)
1448 0 : ereport(ERROR,
1449 : (errcode_for_file_access(),
1450 : errmsg("could not fsync file \"%s\": %m",
1451 : FilePathName(seg->mdfd_vfd))));
1452 : }
1453 : }
1454 :
1455 : /*
1456 : * register_unlink() -- Schedule a file to be deleted after next checkpoint
1457 : *
1458 : * We don't bother passing in the fork number, because this is only used
1459 : * with main forks.
1460 : *
1461 : * As with register_dirty_segment, this could involve either a local or
1462 : * a remote pending-ops table.
1463 : */
1464 : static void
1465 2576 : register_unlink(RelFileNodeBackend rnode)
1466 : {
1467 : /* Should never be used with temp relations */
1468 2576 : Assert(!RelFileNodeBackendIsTemp(rnode));
1469 :
1470 2576 : if (pendingOpsTable)
1471 : {
1472 : /* push it into local pending-ops table */
1473 0 : RememberFsyncRequest(rnode.node, MAIN_FORKNUM,
1474 : UNLINK_RELATION_REQUEST);
1475 : }
1476 : else
1477 : {
1478 : /*
1479 : * Notify the checkpointer about it. If we fail to queue the request
1480 : * message, we have to sleep and try again, because we can't simply
1481 : * delete the file now. Ugly, but hopefully won't happen often.
1482 : *
1483 : * XXX should we just leave the file orphaned instead?
1484 : */
1485 2576 : Assert(IsUnderPostmaster);
1486 5152 : while (!ForwardFsyncRequest(rnode.node, MAIN_FORKNUM,
1487 : UNLINK_RELATION_REQUEST))
1488 0 : pg_usleep(10000L); /* 10 msec seems a good number */
1489 : }
1490 2576 : }
1491 :
1492 : /*
1493 : * RememberFsyncRequest() -- callback from checkpointer side of fsync request
1494 : *
1495 : * We stuff fsync requests into the local hash table for execution
1496 : * during the checkpointer's next checkpoint. UNLINK requests go into a
1497 : * separate linked list, however, because they get processed separately.
1498 : *
1499 : * The range of possible segment numbers is way less than the range of
1500 : * BlockNumber, so we can reserve high values of segno for special purposes.
1501 : * We define three:
1502 : * - FORGET_RELATION_FSYNC means to cancel pending fsyncs for a relation,
1503 : * either for one fork, or all forks if forknum is InvalidForkNumber
1504 : * - FORGET_DATABASE_FSYNC means to cancel pending fsyncs for a whole database
1505 : * - UNLINK_RELATION_REQUEST is a request to delete the file after the next
1506 : * checkpoint.
1507 : * Note also that we're assuming real segment numbers don't exceed INT_MAX.
1508 : *
1509 : * (Handling FORGET_DATABASE_FSYNC requests is a tad slow because the hash
1510 : * table has to be searched linearly, but dropping a database is a pretty
1511 : * heavyweight operation anyhow, so we'll live with it.)
1512 : */
1513 : void
1514 35107 : RememberFsyncRequest(RelFileNode rnode, ForkNumber forknum, BlockNumber segno)
1515 : {
1516 35107 : Assert(pendingOpsTable);
1517 :
1518 35107 : if (segno == FORGET_RELATION_FSYNC)
1519 : {
1520 : /* Remove any pending requests for the relation (one or all forks) */
1521 : PendingOperationEntry *entry;
1522 :
1523 10301 : entry = (PendingOperationEntry *) hash_search(pendingOpsTable,
1524 : &rnode,
1525 : HASH_FIND,
1526 : NULL);
1527 10301 : if (entry)
1528 : {
1529 : /*
1530 : * We can't just delete the entry since mdsync could have an
1531 : * active hashtable scan. Instead we delete the bitmapsets; this
1532 : * is safe because of the way mdsync is coded. We also set the
1533 : * "canceled" flags so that mdsync can tell that a cancel arrived
1534 : * for the fork(s).
1535 : */
1536 3588 : if (forknum == InvalidForkNumber)
1537 : {
1538 : /* remove requests for all forks */
1539 0 : for (forknum = 0; forknum <= MAX_FORKNUM; forknum++)
1540 : {
1541 0 : bms_free(entry->requests[forknum]);
1542 0 : entry->requests[forknum] = NULL;
1543 0 : entry->canceled[forknum] = true;
1544 : }
1545 : }
1546 : else
1547 : {
1548 : /* remove requests for single fork */
1549 3588 : bms_free(entry->requests[forknum]);
1550 3588 : entry->requests[forknum] = NULL;
1551 3588 : entry->canceled[forknum] = true;
1552 : }
1553 : }
1554 : }
1555 24806 : else if (segno == FORGET_DATABASE_FSYNC)
1556 : {
1557 : /* Remove any pending requests for the entire database */
1558 : HASH_SEQ_STATUS hstat;
1559 : PendingOperationEntry *entry;
1560 : ListCell *cell,
1561 : *prev,
1562 : *next;
1563 :
1564 : /* Remove fsync requests */
1565 0 : hash_seq_init(&hstat, pendingOpsTable);
1566 0 : while ((entry = (PendingOperationEntry *) hash_seq_search(&hstat)) != NULL)
1567 : {
1568 0 : if (entry->rnode.dbNode == rnode.dbNode)
1569 : {
1570 : /* remove requests for all forks */
1571 0 : for (forknum = 0; forknum <= MAX_FORKNUM; forknum++)
1572 : {
1573 0 : bms_free(entry->requests[forknum]);
1574 0 : entry->requests[forknum] = NULL;
1575 0 : entry->canceled[forknum] = true;
1576 : }
1577 : }
1578 : }
1579 :
1580 : /* Remove unlink requests */
1581 0 : prev = NULL;
1582 0 : for (cell = list_head(pendingUnlinks); cell; cell = next)
1583 : {
1584 0 : PendingUnlinkEntry *entry = (PendingUnlinkEntry *) lfirst(cell);
1585 :
1586 0 : next = lnext(cell);
1587 0 : if (entry->rnode.dbNode == rnode.dbNode)
1588 : {
1589 0 : pendingUnlinks = list_delete_cell(pendingUnlinks, cell, prev);
1590 0 : pfree(entry);
1591 : }
1592 : else
1593 0 : prev = cell;
1594 : }
1595 : }
1596 24806 : else if (segno == UNLINK_RELATION_REQUEST)
1597 : {
1598 : /* Unlink request: put it in the linked list */
1599 2576 : MemoryContext oldcxt = MemoryContextSwitchTo(pendingOpsCxt);
1600 : PendingUnlinkEntry *entry;
1601 :
1602 : /* PendingUnlinkEntry doesn't store forknum, since it's always MAIN */
1603 2576 : Assert(forknum == MAIN_FORKNUM);
1604 :
1605 2576 : entry = palloc(sizeof(PendingUnlinkEntry));
1606 2576 : entry->rnode = rnode;
1607 2576 : entry->cycle_ctr = mdckpt_cycle_ctr;
1608 :
1609 2576 : pendingUnlinks = lappend(pendingUnlinks, entry);
1610 :
1611 2576 : MemoryContextSwitchTo(oldcxt);
1612 : }
1613 : else
1614 : {
1615 : /* Normal case: enter a request to fsync this segment */
1616 22230 : MemoryContext oldcxt = MemoryContextSwitchTo(pendingOpsCxt);
1617 : PendingOperationEntry *entry;
1618 : bool found;
1619 :
1620 22230 : entry = (PendingOperationEntry *) hash_search(pendingOpsTable,
1621 : &rnode,
1622 : HASH_ENTER,
1623 : &found);
1624 : /* if new entry, initialize it */
1625 22230 : if (!found)
1626 : {
1627 1630 : entry->cycle_ctr = mdsync_cycle_ctr;
1628 1630 : MemSet(entry->requests, 0, sizeof(entry->requests));
1629 1630 : MemSet(entry->canceled, 0, sizeof(entry->canceled));
1630 : }
1631 :
1632 : /*
1633 : * NB: it's intentional that we don't change cycle_ctr if the entry
1634 : * already exists. The cycle_ctr must represent the oldest fsync
1635 : * request that could be in the entry.
1636 : */
1637 :
1638 22230 : entry->requests[forknum] = bms_add_member(entry->requests[forknum],
1639 : (int) segno);
1640 :
1641 22230 : MemoryContextSwitchTo(oldcxt);
1642 : }
1643 35107 : }
1644 :
1645 : /*
1646 : * ForgetRelationFsyncRequests -- forget any fsyncs for a relation fork
1647 : *
1648 : * forknum == InvalidForkNumber means all forks, although this code doesn't
1649 : * actually know that, since it's just forwarding the request elsewhere.
1650 : */
1651 : void
1652 10301 : ForgetRelationFsyncRequests(RelFileNode rnode, ForkNumber forknum)
1653 : {
1654 10301 : if (pendingOpsTable)
1655 : {
1656 : /* standalone backend or startup process: fsync state is local */
1657 0 : RememberFsyncRequest(rnode, forknum, FORGET_RELATION_FSYNC);
1658 : }
1659 10301 : else if (IsUnderPostmaster)
1660 : {
1661 : /*
1662 : * Notify the checkpointer about it. If we fail to queue the cancel
1663 : * message, we have to sleep and try again ... ugly, but hopefully
1664 : * won't happen often.
1665 : *
1666 : * XXX should we CHECK_FOR_INTERRUPTS in this loop? Escaping with an
1667 : * error would leave the no-longer-used file still present on disk,
1668 : * which would be bad, so I'm inclined to assume that the checkpointer
1669 : * will always empty the queue soon.
1670 : */
1671 20602 : while (!ForwardFsyncRequest(rnode, forknum, FORGET_RELATION_FSYNC))
1672 0 : pg_usleep(10000L); /* 10 msec seems a good number */
1673 :
1674 : /*
1675 : * Note we don't wait for the checkpointer to actually absorb the
1676 : * cancel message; see mdsync() for the implications.
1677 : */
1678 : }
1679 10301 : }
1680 :
1681 : /*
1682 : * ForgetDatabaseFsyncRequests -- forget any fsyncs and unlinks for a DB
1683 : */
1684 : void
1685 0 : ForgetDatabaseFsyncRequests(Oid dbid)
1686 : {
1687 : RelFileNode rnode;
1688 :
1689 0 : rnode.dbNode = dbid;
1690 0 : rnode.spcNode = 0;
1691 0 : rnode.relNode = 0;
1692 :
1693 0 : if (pendingOpsTable)
1694 : {
1695 : /* standalone backend or startup process: fsync state is local */
1696 0 : RememberFsyncRequest(rnode, InvalidForkNumber, FORGET_DATABASE_FSYNC);
1697 : }
1698 0 : else if (IsUnderPostmaster)
1699 : {
1700 : /* see notes in ForgetRelationFsyncRequests */
1701 0 : while (!ForwardFsyncRequest(rnode, InvalidForkNumber,
1702 : FORGET_DATABASE_FSYNC))
1703 0 : pg_usleep(10000L); /* 10 msec seems a good number */
1704 : }
1705 0 : }
1706 :
1707 :
1708 : /*
1709 : * _fdvec_resize() -- Resize the fork's open segments array
1710 : */
1711 : static void
1712 31066 : _fdvec_resize(SMgrRelation reln,
1713 : ForkNumber forknum,
1714 : int nseg)
1715 : {
1716 31066 : if (nseg == 0)
1717 : {
1718 11935 : if (reln->md_num_open_segs[forknum] > 0)
1719 : {
1720 11935 : pfree(reln->md_seg_fds[forknum]);
1721 11935 : reln->md_seg_fds[forknum] = NULL;
1722 : }
1723 : }
1724 19131 : else if (reln->md_num_open_segs[forknum] == 0)
1725 : {
1726 19131 : reln->md_seg_fds[forknum] =
1727 19131 : MemoryContextAlloc(MdCxt, sizeof(MdfdVec) * nseg);
1728 : }
1729 : else
1730 : {
1731 : /*
1732 : * It doesn't seem worthwhile complicating the code by having a more
1733 : * aggressive growth strategy here; the number of segments doesn't
1734 : * grow that fast, and the memory context internally will sometimes
1735 : * avoid doing an actual reallocation.
1736 : */
1737 0 : reln->md_seg_fds[forknum] =
1738 0 : repalloc(reln->md_seg_fds[forknum],
1739 : sizeof(MdfdVec) * nseg);
1740 : }
1741 :
1742 31066 : reln->md_num_open_segs[forknum] = nseg;
1743 31066 : }
1744 :
1745 : /*
1746 : * Return the filename for the specified segment of the relation. The
1747 : * returned string is palloc'd.
1748 : */
1749 : static char *
1750 0 : _mdfd_segpath(SMgrRelation reln, ForkNumber forknum, BlockNumber segno)
1751 : {
1752 : char *path,
1753 : *fullpath;
1754 :
1755 0 : path = relpath(reln->smgr_rnode, forknum);
1756 :
1757 0 : if (segno > 0)
1758 : {
1759 0 : fullpath = psprintf("%s.%u", path, segno);
1760 0 : pfree(path);
1761 : }
1762 : else
1763 0 : fullpath = path;
1764 :
1765 0 : return fullpath;
1766 : }
1767 :
1768 : /*
1769 : * Open the specified segment of the relation,
1770 : * and make a MdfdVec object for it. Returns NULL on failure.
1771 : */
1772 : static MdfdVec *
1773 0 : _mdfd_openseg(SMgrRelation reln, ForkNumber forknum, BlockNumber segno,
1774 : int oflags)
1775 : {
1776 : MdfdVec *v;
1777 : int fd;
1778 : char *fullpath;
1779 :
1780 0 : fullpath = _mdfd_segpath(reln, forknum, segno);
1781 :
1782 : /* open the file */
1783 0 : fd = PathNameOpenFile(fullpath, O_RDWR | PG_BINARY | oflags, 0600);
1784 :
1785 0 : pfree(fullpath);
1786 :
1787 0 : if (fd < 0)
1788 0 : return NULL;
1789 :
1790 0 : if (segno <= reln->md_num_open_segs[forknum])
1791 0 : _fdvec_resize(reln, forknum, segno + 1);
1792 :
1793 : /* fill the entry */
1794 0 : v = &reln->md_seg_fds[forknum][segno];
1795 0 : v->mdfd_vfd = fd;
1796 0 : v->mdfd_segno = segno;
1797 :
1798 0 : Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
1799 :
1800 : /* all done */
1801 0 : return v;
1802 : }
1803 :
1804 : /*
1805 : * _mdfd_getseg() -- Find the segment of the relation holding the
1806 : * specified block.
1807 : *
1808 : * If the segment doesn't exist, we ereport, return NULL, or create the
1809 : * segment, according to "behavior". Note: skipFsync is only used in the
1810 : * EXTENSION_CREATE case.
1811 : */
1812 : static MdfdVec *
1813 33340 : _mdfd_getseg(SMgrRelation reln, ForkNumber forknum, BlockNumber blkno,
1814 : bool skipFsync, int behavior)
1815 : {
1816 : MdfdVec *v;
1817 : BlockNumber targetseg;
1818 : BlockNumber nextsegno;
1819 :
1820 : /* some way to handle non-existent segments needs to be specified */
1821 33340 : Assert(behavior &
1822 : (EXTENSION_FAIL | EXTENSION_CREATE | EXTENSION_RETURN_NULL));
1823 :
1824 33340 : targetseg = blkno / ((BlockNumber) RELSEG_SIZE);
1825 :
1826 : /* if an existing and opened segment, we're done */
1827 33340 : if (targetseg < reln->md_num_open_segs[forknum])
1828 : {
1829 31794 : v = &reln->md_seg_fds[forknum][targetseg];
1830 31794 : return v;
1831 : }
1832 :
1833 : /*
1834 : * The target segment is not yet open. Iterate over all the segments
1835 : * between the last opened and the target segment. This way missing
1836 : * segments either raise an error, or get created (according to
1837 : * 'behavior'). Start with either the last opened, or the first segment if
1838 : * none was opened before.
1839 : */
1840 1546 : if (reln->md_num_open_segs[forknum] > 0)
1841 0 : v = &reln->md_seg_fds[forknum][reln->md_num_open_segs[forknum] - 1];
1842 : else
1843 : {
1844 1546 : v = mdopen(reln, forknum, behavior);
1845 1546 : if (!v)
1846 0 : return NULL; /* if behavior & EXTENSION_RETURN_NULL */
1847 : }
1848 :
1849 3092 : for (nextsegno = reln->md_num_open_segs[forknum];
1850 0 : nextsegno <= targetseg; nextsegno++)
1851 : {
1852 0 : BlockNumber nblocks = _mdnblocks(reln, forknum, v);
1853 0 : int flags = 0;
1854 :
1855 0 : Assert(nextsegno == v->mdfd_segno + 1);
1856 :
1857 0 : if (nblocks > ((BlockNumber) RELSEG_SIZE))
1858 0 : elog(FATAL, "segment too big");
1859 :
1860 0 : if ((behavior & EXTENSION_CREATE) ||
1861 0 : (InRecovery && (behavior & EXTENSION_CREATE_RECOVERY)))
1862 : {
1863 : /*
1864 : * Normally we will create new segments only if authorized by the
1865 : * caller (i.e., we are doing mdextend()). But when doing WAL
1866 : * recovery, create segments anyway; this allows cases such as
1867 : * replaying WAL data that has a write into a high-numbered
1868 : * segment of a relation that was later deleted. We want to go
1869 : * ahead and create the segments so we can finish out the replay.
1870 : * However if the caller has specified
1871 : * EXTENSION_REALLY_RETURN_NULL, then extension is not desired
1872 : * even in recovery; we won't reach this point in that case.
1873 : *
1874 : * We have to maintain the invariant that segments before the last
1875 : * active segment are of size RELSEG_SIZE; therefore, if
1876 : * extending, pad them out with zeroes if needed. (This only
1877 : * matters if in recovery, or if the caller is extending the
1878 : * relation discontiguously, but that can happen in hash indexes.)
1879 : */
1880 0 : if (nblocks < ((BlockNumber) RELSEG_SIZE))
1881 : {
1882 0 : char *zerobuf = palloc0(BLCKSZ);
1883 :
1884 0 : mdextend(reln, forknum,
1885 0 : nextsegno * ((BlockNumber) RELSEG_SIZE) - 1,
1886 : zerobuf, skipFsync);
1887 0 : pfree(zerobuf);
1888 : }
1889 0 : flags = O_CREAT;
1890 : }
1891 0 : else if (!(behavior & EXTENSION_DONT_CHECK_SIZE) &&
1892 : nblocks < ((BlockNumber) RELSEG_SIZE))
1893 : {
1894 : /*
1895 : * When not extending (or explicitly including truncated
1896 : * segments), only open the next segment if the current one is
1897 : * exactly RELSEG_SIZE. If not (this branch), either return NULL
1898 : * or fail.
1899 : */
1900 0 : if (behavior & EXTENSION_RETURN_NULL)
1901 : {
1902 : /*
1903 : * Some callers discern between reasons for _mdfd_getseg()
1904 : * returning NULL based on errno. As there's no failing
1905 : * syscall involved in this case, explicitly set errno to
1906 : * ENOENT, as that seems the closest interpretation.
1907 : */
1908 0 : errno = ENOENT;
1909 0 : return NULL;
1910 : }
1911 :
1912 0 : ereport(ERROR,
1913 : (errcode_for_file_access(),
1914 : errmsg("could not open file \"%s\" (target block %u): previous segment is only %u blocks",
1915 : _mdfd_segpath(reln, forknum, nextsegno),
1916 : blkno, nblocks)));
1917 : }
1918 :
1919 0 : v = _mdfd_openseg(reln, forknum, nextsegno, flags);
1920 :
1921 0 : if (v == NULL)
1922 : {
1923 0 : if ((behavior & EXTENSION_RETURN_NULL) &&
1924 0 : FILE_POSSIBLY_DELETED(errno))
1925 0 : return NULL;
1926 0 : ereport(ERROR,
1927 : (errcode_for_file_access(),
1928 : errmsg("could not open file \"%s\" (target block %u): %m",
1929 : _mdfd_segpath(reln, forknum, nextsegno),
1930 : blkno)));
1931 : }
1932 : }
1933 :
1934 1546 : return v;
1935 : }
1936 :
1937 : /*
1938 : * Get number of blocks present in a single disk file
1939 : */
1940 : static BlockNumber
1941 118699 : _mdnblocks(SMgrRelation reln, ForkNumber forknum, MdfdVec *seg)
1942 : {
1943 : off_t len;
1944 :
1945 118699 : len = FileSeek(seg->mdfd_vfd, 0L, SEEK_END);
1946 118699 : if (len < 0)
1947 0 : ereport(ERROR,
1948 : (errcode_for_file_access(),
1949 : errmsg("could not seek to end of file \"%s\": %m",
1950 : FilePathName(seg->mdfd_vfd))));
1951 : /* note that this calculation will ignore any partial block at EOF */
1952 118699 : return (BlockNumber) (len / BLCKSZ);
1953 : }
|
