1 /*-------------------------------------------------------------------------
4 * This code manages relations that reside on magnetic disk.
6 * Portions Copyright (c) 1996-2008, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
11 * $PostgreSQL: pgsql/src/backend/storage/smgr/md.c,v 1.137 2008/04/18 06:48:38 heikki Exp $
13 *-------------------------------------------------------------------------
21 #include "catalog/catalog.h"
22 #include "miscadmin.h"
23 #include "postmaster/bgwriter.h"
24 #include "storage/fd.h"
25 #include "storage/bufmgr.h"
26 #include "storage/smgr.h"
27 #include "utils/hsearch.h"
28 #include "utils/memutils.h"
31 /* interval for calling AbsorbFsyncRequests in mdsync */
32 #define FSYNCS_PER_ABSORB 10
34 /* special values for the segno arg to RememberFsyncRequest */
35 #define FORGET_RELATION_FSYNC (InvalidBlockNumber)
36 #define FORGET_DATABASE_FSYNC (InvalidBlockNumber-1)
37 #define UNLINK_RELATION_REQUEST (InvalidBlockNumber-2)
40 * On Windows, we have to interpret EACCES as possibly meaning the same as
41 * ENOENT, because if a file is unlinked-but-not-yet-gone on that platform,
42 * that's what you get. Ugh. This code is designed so that we don't
43 * actually believe these cases are okay without further evidence (namely,
44 * a pending fsync request getting revoked ... see mdsync).
47 #define FILE_POSSIBLY_DELETED(err) ((err) == ENOENT)
49 #define FILE_POSSIBLY_DELETED(err) ((err) == ENOENT || (err) == EACCES)
53 * The magnetic disk storage manager keeps track of open file
54 * descriptors in its own descriptor pool. This is done to make it
55 * easier to support relations that are larger than the operating
56 * system's file size limit (often 2GBytes). In order to do that,
57 * we break relations up into "segment" files that are each shorter than
58 * the OS file size limit. The segment size is set by the RELSEG_SIZE
59 * configuration constant in pg_config_manual.h.
61 * On disk, a relation must consist of consecutively numbered segment
62 * files in the pattern
63 * -- Zero or more full segments of exactly RELSEG_SIZE blocks each
64 * -- Exactly one partial segment of size 0 <= size < RELSEG_SIZE blocks
65 * -- Optionally, any number of inactive segments of size 0 blocks.
66 * The full and partial segments are collectively the "active" segments.
67 * Inactive segments are those that once contained data but are currently
68 * not needed because of an mdtruncate() operation. The reason for leaving
69 * them present at size zero, rather than unlinking them, is that other
70 * backends and/or the bgwriter might be holding open file references to
71 * such segments. If the relation expands again after mdtruncate(), such
72 * that a deactivated segment becomes active again, it is important that
73 * such file references still be valid --- else data might get written
74 * out to an unlinked old copy of a segment file that will eventually
77 * The file descriptor pointer (md_fd field) stored in the SMgrRelation
78 * cache is, therefore, just the head of a list of MdfdVec objects, one
79 * per segment. But note the md_fd pointer can be NULL, indicating
82 * Also note that mdfd_chain == NULL does not necessarily mean the relation
83 * doesn't have another segment after this one; we may just not have
84 * opened the next segment yet. (We could not have "all segments are
85 * in the chain" as an invariant anyway, since another backend could
86 * extend the relation when we weren't looking.) We do not make chain
87 * entries for inactive segments, however; as soon as we find a partial
88 * segment, we assume that any subsequent segments are inactive.
90 * All MdfdVec objects are palloc'd in the MdCxt memory context.
92 * On platforms that support large files, USE_SEGMENTED_FILES can be
93 * #undef'd to disable the segmentation logic. In that case each
94 * relation is a single operating-system file.
97 typedef struct _MdfdVec
99 File mdfd_vfd; /* fd number in fd.c's pool */
100 BlockNumber mdfd_segno; /* segment number, from 0 */
101 #ifdef USE_SEGMENTED_FILES
102 struct _MdfdVec *mdfd_chain; /* next segment, or NULL */
106 static MemoryContext MdCxt; /* context for all md.c allocations */
110 * In some contexts (currently, standalone backends and the bgwriter process)
111 * we keep track of pending fsync operations: we need to remember all relation
112 * segments that have been written since the last checkpoint, so that we can
113 * fsync them down to disk before completing the next checkpoint. This hash
114 * table remembers the pending operations. We use a hash table mostly as
115 * a convenient way of eliminating duplicate requests.
117 * We use a similar mechanism to remember no-longer-needed files that can
118 * be deleted after the next checkpoint, but we use a linked list instead of
119 * a hash table, because we don't expect there to be any duplicate requests.
121 * (Regular backends do not track pending operations locally, but forward
122 * them to the bgwriter.)
126 RelFileNode rnode; /* the targeted relation */
127 BlockNumber segno; /* which segment */
128 } PendingOperationTag;
130 typedef uint16 CycleCtr; /* can be any convenient integer size */
134 PendingOperationTag tag; /* hash table key (must be first!) */
135 bool canceled; /* T => request canceled, not yet removed */
136 CycleCtr cycle_ctr; /* mdsync_cycle_ctr when request was made */
137 } PendingOperationEntry;
141 RelFileNode rnode; /* the dead relation to delete */
142 CycleCtr cycle_ctr; /* mdckpt_cycle_ctr when request was made */
143 } PendingUnlinkEntry;
145 static HTAB *pendingOpsTable = NULL;
146 static List *pendingUnlinks = NIL;
148 static CycleCtr mdsync_cycle_ctr = 0;
149 static CycleCtr mdckpt_cycle_ctr = 0;
152 typedef enum /* behavior for mdopen & _mdfd_getseg */
154 EXTENSION_FAIL, /* ereport if segment not present */
155 EXTENSION_RETURN_NULL, /* return NULL if not present */
156 EXTENSION_CREATE /* create new segments as needed */
160 static MdfdVec *mdopen(SMgrRelation reln, ExtensionBehavior behavior);
161 static void register_dirty_segment(SMgrRelation reln, MdfdVec *seg);
162 static void register_unlink(RelFileNode rnode);
163 static MdfdVec *_fdvec_alloc(void);
165 #ifdef USE_SEGMENTED_FILES
166 static MdfdVec *_mdfd_openseg(SMgrRelation reln, BlockNumber segno,
169 static MdfdVec *_mdfd_getseg(SMgrRelation reln, BlockNumber blkno,
170 bool isTemp, ExtensionBehavior behavior);
171 static BlockNumber _mdnblocks(SMgrRelation reln, MdfdVec *seg);
175 * mdinit() -- Initialize private state for magnetic disk storage manager.
180 MdCxt = AllocSetContextCreate(TopMemoryContext,
182 ALLOCSET_DEFAULT_MINSIZE,
183 ALLOCSET_DEFAULT_INITSIZE,
184 ALLOCSET_DEFAULT_MAXSIZE);
187 * Create pending-operations hashtable if we need it. Currently, we need
188 * it if we are standalone (not under a postmaster) OR if we are a
189 * bootstrap-mode subprocess of a postmaster (that is, a startup or
192 if (!IsUnderPostmaster || IsBootstrapProcessingMode())
196 MemSet(&hash_ctl, 0, sizeof(hash_ctl));
197 hash_ctl.keysize = sizeof(PendingOperationTag);
198 hash_ctl.entrysize = sizeof(PendingOperationEntry);
199 hash_ctl.hash = tag_hash;
200 hash_ctl.hcxt = MdCxt;
201 pendingOpsTable = hash_create("Pending Ops Table",
204 HASH_ELEM | HASH_FUNCTION | HASH_CONTEXT);
205 pendingUnlinks = NIL;
210 * mdcreate() -- Create a new relation on magnetic disk.
212 * If isRedo is true, it's okay for the relation to exist already.
215 mdcreate(SMgrRelation reln, bool isRedo)
220 if (isRedo && reln->md_fd != NULL)
221 return; /* created and opened already... */
223 Assert(reln->md_fd == NULL);
225 path = relpath(reln->smgr_rnode);
227 fd = PathNameOpenFile(path, O_RDWR | O_CREAT | O_EXCL | PG_BINARY, 0600);
231 int save_errno = errno;
234 * During bootstrap, there are cases where a system relation will be
235 * accessed (by internal backend processes) before the bootstrap
236 * script nominally creates it. Therefore, allow the file to exist
237 * already, even if isRedo is not set. (See also mdopen)
239 if (isRedo || IsBootstrapProcessingMode())
240 fd = PathNameOpenFile(path, O_RDWR | PG_BINARY, 0600);
244 /* be sure to report the error reported by create, not open */
247 (errcode_for_file_access(),
248 errmsg("could not create relation %u/%u/%u: %m",
249 reln->smgr_rnode.spcNode,
250 reln->smgr_rnode.dbNode,
251 reln->smgr_rnode.relNode)));
257 reln->md_fd = _fdvec_alloc();
259 reln->md_fd->mdfd_vfd = fd;
260 reln->md_fd->mdfd_segno = 0;
261 #ifdef USE_SEGMENTED_FILES
262 reln->md_fd->mdfd_chain = NULL;
267 * mdunlink() -- Unlink a relation.
269 * Note that we're passed a RelFileNode --- by the time this is called,
270 * there won't be an SMgrRelation hashtable entry anymore.
272 * Actually, we don't unlink the first segment file of the relation, but
273 * just truncate it to zero length, and record a request to unlink it after
274 * the next checkpoint. Additional segments can be unlinked immediately,
275 * however. Leaving the empty file in place prevents that relfilenode
276 * number from being reused. The scenario this protects us from is:
277 * 1. We delete a relation (and commit, and actually remove its file).
278 * 2. We create a new relation, which by chance gets the same relfilenode as
279 * the just-deleted one (OIDs must've wrapped around for that to happen).
280 * 3. We crash before another checkpoint occurs.
281 * During replay, we would delete the file and then recreate it, which is fine
282 * if the contents of the file were repopulated by subsequent WAL entries.
283 * But if we didn't WAL-log insertions, but instead relied on fsyncing the
284 * file after populating it (as for instance CLUSTER and CREATE INDEX do),
285 * the contents of the file would be lost forever. By leaving the empty file
286 * until after the next checkpoint, we prevent reassignment of the relfilenode
287 * number until it's safe, because relfilenode assignment skips over any
290 * If isRedo is true, it's okay for the relation to be already gone.
291 * Also, we should remove the file immediately instead of queuing a request
292 * for later, since during redo there's no possibility of creating a
293 * conflicting relation.
295 * Note: any failure should be reported as WARNING not ERROR, because
296 * we are usually not in a transaction anymore when this is called.
299 mdunlink(RelFileNode rnode, bool isRedo)
305 * We have to clean out any pending fsync requests for the doomed
306 * relation, else the next mdsync() will fail.
308 ForgetRelationFsyncRequests(rnode);
310 path = relpath(rnode);
313 * Delete or truncate the first segment, or only segment if not doing
320 /* truncate(2) would be easier here, but Windows hasn't got it */
323 fd = BasicOpenFile(path, O_RDWR | PG_BINARY, 0);
328 ret = ftruncate(fd, 0);
338 if (!isRedo || errno != ENOENT)
340 (errcode_for_file_access(),
341 errmsg("could not remove relation %u/%u/%u: %m",
347 #ifdef USE_SEGMENTED_FILES
348 /* Delete the additional segments, if any */
351 char *segpath = (char *) palloc(strlen(path) + 12);
355 * Note that because we loop until getting ENOENT, we will correctly
356 * remove all inactive segments as well as active ones.
358 for (segno = 1;; segno++)
360 sprintf(segpath, "%s.%u", path, segno);
361 if (unlink(segpath) < 0)
363 /* ENOENT is expected after the last segment... */
366 (errcode_for_file_access(),
367 errmsg("could not remove segment %u of relation %u/%u/%u: %m",
381 /* Register request to unlink first segment later */
383 register_unlink(rnode);
387 * mdextend() -- Add a block to the specified relation.
389 * The semantics are nearly the same as mdwrite(): write at the
390 * specified position. However, this is to be used for the case of
391 * extending a relation (i.e., blocknum is at or beyond the current
392 * EOF). Note that we assume writing a block beyond current EOF
393 * causes intervening file space to become filled with zeroes.
396 mdextend(SMgrRelation reln, BlockNumber blocknum, char *buffer, bool isTemp)
402 /* This assert is too expensive to have on normally ... */
403 #ifdef CHECK_WRITE_VS_EXTEND
404 Assert(blocknum >= mdnblocks(reln));
408 * If a relation manages to grow to 2^32-1 blocks, refuse to extend it any
409 * more --- we mustn't create a block whose number actually is
410 * InvalidBlockNumber.
412 if (blocknum == InvalidBlockNumber)
414 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
415 errmsg("cannot extend relation %u/%u/%u beyond %u blocks",
416 reln->smgr_rnode.spcNode,
417 reln->smgr_rnode.dbNode,
418 reln->smgr_rnode.relNode,
419 InvalidBlockNumber)));
421 v = _mdfd_getseg(reln, blocknum, isTemp, EXTENSION_CREATE);
423 #ifdef USE_SEGMENTED_FILES
424 seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
425 Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
427 seekpos = (off_t) BLCKSZ * blocknum;
431 * Note: because caller usually obtained blocknum by calling mdnblocks,
432 * which did a seek(SEEK_END), this seek is often redundant and will be
433 * optimized away by fd.c. It's not redundant, however, if there is a
434 * partial page at the end of the file. In that case we want to try to
435 * overwrite the partial page with a full page. It's also not redundant
436 * if bufmgr.c had to dump another buffer of the same file to make room
437 * for the new page's buffer.
439 if (FileSeek(v->mdfd_vfd, seekpos, SEEK_SET) != seekpos)
441 (errcode_for_file_access(),
442 errmsg("could not seek to block %u of relation %u/%u/%u: %m",
444 reln->smgr_rnode.spcNode,
445 reln->smgr_rnode.dbNode,
446 reln->smgr_rnode.relNode)));
448 if ((nbytes = FileWrite(v->mdfd_vfd, buffer, BLCKSZ)) != BLCKSZ)
452 (errcode_for_file_access(),
453 errmsg("could not extend relation %u/%u/%u: %m",
454 reln->smgr_rnode.spcNode,
455 reln->smgr_rnode.dbNode,
456 reln->smgr_rnode.relNode),
457 errhint("Check free disk space.")));
458 /* short write: complain appropriately */
460 (errcode(ERRCODE_DISK_FULL),
461 errmsg("could not extend relation %u/%u/%u: wrote only %d of %d bytes at block %u",
462 reln->smgr_rnode.spcNode,
463 reln->smgr_rnode.dbNode,
464 reln->smgr_rnode.relNode,
465 nbytes, BLCKSZ, blocknum),
466 errhint("Check free disk space.")));
470 register_dirty_segment(reln, v);
472 #ifdef USE_SEGMENTED_FILES
473 Assert(_mdnblocks(reln, v) <= ((BlockNumber) RELSEG_SIZE));
478 * mdopen() -- Open the specified relation.
480 * Note we only open the first segment, when there are multiple segments.
482 * If first segment is not present, either ereport or return NULL according
483 * to "behavior". We treat EXTENSION_CREATE the same as EXTENSION_FAIL;
484 * EXTENSION_CREATE means it's OK to extend an existing relation, not to
485 * invent one out of whole cloth.
488 mdopen(SMgrRelation reln, ExtensionBehavior behavior)
494 /* No work if already open */
498 path = relpath(reln->smgr_rnode);
500 fd = PathNameOpenFile(path, O_RDWR | PG_BINARY, 0600);
505 * During bootstrap, there are cases where a system relation will be
506 * accessed (by internal backend processes) before the bootstrap
507 * script nominally creates it. Therefore, accept mdopen() as a
508 * substitute for mdcreate() in bootstrap mode only. (See mdcreate)
510 if (IsBootstrapProcessingMode())
511 fd = PathNameOpenFile(path, O_RDWR | O_CREAT | O_EXCL | PG_BINARY, 0600);
515 if (behavior == EXTENSION_RETURN_NULL &&
516 FILE_POSSIBLY_DELETED(errno))
519 (errcode_for_file_access(),
520 errmsg("could not open relation %u/%u/%u: %m",
521 reln->smgr_rnode.spcNode,
522 reln->smgr_rnode.dbNode,
523 reln->smgr_rnode.relNode)));
529 reln->md_fd = mdfd = _fdvec_alloc();
532 mdfd->mdfd_segno = 0;
533 #ifdef USE_SEGMENTED_FILES
534 mdfd->mdfd_chain = NULL;
535 Assert(_mdnblocks(reln, mdfd) <= ((BlockNumber) RELSEG_SIZE));
542 * mdclose() -- Close the specified relation, if it isn't closed already.
545 mdclose(SMgrRelation reln)
547 MdfdVec *v = reln->md_fd;
549 /* No work if already closed */
553 reln->md_fd = NULL; /* prevent dangling pointer after error */
555 #ifdef USE_SEGMENTED_FILES
560 /* if not closed already */
561 if (v->mdfd_vfd >= 0)
562 FileClose(v->mdfd_vfd);
563 /* Now free vector */
568 if (v->mdfd_vfd >= 0)
569 FileClose(v->mdfd_vfd);
575 * mdread() -- Read the specified block from a relation.
578 mdread(SMgrRelation reln, BlockNumber blocknum, char *buffer)
584 v = _mdfd_getseg(reln, blocknum, false, EXTENSION_FAIL);
586 #ifdef USE_SEGMENTED_FILES
587 seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
588 Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
590 seekpos = (off_t) BLCKSZ * blocknum;
593 if (FileSeek(v->mdfd_vfd, seekpos, SEEK_SET) != seekpos)
595 (errcode_for_file_access(),
596 errmsg("could not seek to block %u of relation %u/%u/%u: %m",
598 reln->smgr_rnode.spcNode,
599 reln->smgr_rnode.dbNode,
600 reln->smgr_rnode.relNode)));
602 if ((nbytes = FileRead(v->mdfd_vfd, buffer, BLCKSZ)) != BLCKSZ)
606 (errcode_for_file_access(),
607 errmsg("could not read block %u of relation %u/%u/%u: %m",
609 reln->smgr_rnode.spcNode,
610 reln->smgr_rnode.dbNode,
611 reln->smgr_rnode.relNode)));
614 * Short read: we are at or past EOF, or we read a partial block at
615 * EOF. Normally this is an error; upper levels should never try to
616 * read a nonexistent block. However, if zero_damaged_pages is ON or
617 * we are InRecovery, we should instead return zeroes without
618 * complaining. This allows, for example, the case of trying to
619 * update a block that was later truncated away.
621 if (zero_damaged_pages || InRecovery)
622 MemSet(buffer, 0, BLCKSZ);
625 (errcode(ERRCODE_DATA_CORRUPTED),
626 errmsg("could not read block %u of relation %u/%u/%u: read only %d of %d bytes",
628 reln->smgr_rnode.spcNode,
629 reln->smgr_rnode.dbNode,
630 reln->smgr_rnode.relNode,
636 * mdwrite() -- Write the supplied block at the appropriate location.
638 * This is to be used only for updating already-existing blocks of a
639 * relation (ie, those before the current EOF). To extend a relation,
643 mdwrite(SMgrRelation reln, BlockNumber blocknum, char *buffer, bool isTemp)
649 /* This assert is too expensive to have on normally ... */
650 #ifdef CHECK_WRITE_VS_EXTEND
651 Assert(blocknum < mdnblocks(reln));
654 v = _mdfd_getseg(reln, blocknum, isTemp, EXTENSION_FAIL);
656 #ifdef USE_SEGMENTED_FILES
657 seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
658 Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
660 seekpos = (off_t) BLCKSZ * blocknum;
663 if (FileSeek(v->mdfd_vfd, seekpos, SEEK_SET) != seekpos)
665 (errcode_for_file_access(),
666 errmsg("could not seek to block %u of relation %u/%u/%u: %m",
668 reln->smgr_rnode.spcNode,
669 reln->smgr_rnode.dbNode,
670 reln->smgr_rnode.relNode)));
672 if ((nbytes = FileWrite(v->mdfd_vfd, buffer, BLCKSZ)) != BLCKSZ)
676 (errcode_for_file_access(),
677 errmsg("could not write block %u of relation %u/%u/%u: %m",
679 reln->smgr_rnode.spcNode,
680 reln->smgr_rnode.dbNode,
681 reln->smgr_rnode.relNode)));
682 /* short write: complain appropriately */
684 (errcode(ERRCODE_DISK_FULL),
685 errmsg("could not write block %u of relation %u/%u/%u: wrote only %d of %d bytes",
687 reln->smgr_rnode.spcNode,
688 reln->smgr_rnode.dbNode,
689 reln->smgr_rnode.relNode,
691 errhint("Check free disk space.")));
695 register_dirty_segment(reln, v);
699 * mdnblocks() -- Get the number of blocks stored in a relation.
701 * Important side effect: all active segments of the relation are opened
702 * and added to the mdfd_chain list. If this routine has not been
703 * called, then only segments up to the last one actually touched
704 * are present in the chain.
707 mdnblocks(SMgrRelation reln)
709 MdfdVec *v = mdopen(reln, EXTENSION_FAIL);
711 #ifdef USE_SEGMENTED_FILES
713 BlockNumber segno = 0;
716 * Skip through any segments that aren't the last one, to avoid redundant
717 * seeks on them. We have previously verified that these segments are
718 * exactly RELSEG_SIZE long, and it's useless to recheck that each time.
720 * NOTE: this assumption could only be wrong if another backend has
721 * truncated the relation. We rely on higher code levels to handle that
722 * scenario by closing and re-opening the md fd, which is handled via
723 * relcache flush. (Since the bgwriter doesn't participate in relcache
724 * flush, it could have segment chain entries for inactive segments;
725 * that's OK because the bgwriter never needs to compute relation size.)
727 while (v->mdfd_chain != NULL)
735 nblocks = _mdnblocks(reln, v);
736 if (nblocks > ((BlockNumber) RELSEG_SIZE))
737 elog(FATAL, "segment too big");
738 if (nblocks < ((BlockNumber) RELSEG_SIZE))
739 return (segno * ((BlockNumber) RELSEG_SIZE)) + nblocks;
742 * If segment is exactly RELSEG_SIZE, advance to next one.
746 if (v->mdfd_chain == NULL)
749 * Because we pass O_CREAT, we will create the next segment (with
750 * zero length) immediately, if the last segment is of length
751 * RELSEG_SIZE. While perhaps not strictly necessary, this keeps
754 v->mdfd_chain = _mdfd_openseg(reln, segno, O_CREAT);
755 if (v->mdfd_chain == NULL)
757 (errcode_for_file_access(),
758 errmsg("could not open segment %u of relation %u/%u/%u: %m",
760 reln->smgr_rnode.spcNode,
761 reln->smgr_rnode.dbNode,
762 reln->smgr_rnode.relNode)));
768 return _mdnblocks(reln, v);
773 * mdtruncate() -- Truncate relation to specified number of blocks.
776 mdtruncate(SMgrRelation reln, BlockNumber nblocks, bool isTemp)
781 #ifdef USE_SEGMENTED_FILES
782 BlockNumber priorblocks;
786 * NOTE: mdnblocks makes sure we have opened all active segments, so that
787 * truncation loop will get them all!
789 curnblk = mdnblocks(reln);
790 if (nblocks > curnblk)
792 /* Bogus request ... but no complaint if InRecovery */
796 (errmsg("could not truncate relation %u/%u/%u to %u blocks: it's only %u blocks now",
797 reln->smgr_rnode.spcNode,
798 reln->smgr_rnode.dbNode,
799 reln->smgr_rnode.relNode,
802 if (nblocks == curnblk)
803 return; /* no work */
805 v = mdopen(reln, EXTENSION_FAIL);
807 #ifdef USE_SEGMENTED_FILES
813 if (priorblocks > nblocks)
816 * This segment is no longer active (and has already been unlinked
817 * from the mdfd_chain). We truncate the file, but do not delete
818 * it, for reasons explained in the header comments.
820 if (FileTruncate(v->mdfd_vfd, 0) < 0)
822 (errcode_for_file_access(),
823 errmsg("could not truncate relation %u/%u/%u to %u blocks: %m",
824 reln->smgr_rnode.spcNode,
825 reln->smgr_rnode.dbNode,
826 reln->smgr_rnode.relNode,
829 register_dirty_segment(reln, v);
831 Assert(ov != reln->md_fd); /* we never drop the 1st segment */
834 else if (priorblocks + ((BlockNumber) RELSEG_SIZE) > nblocks)
837 * This is the last segment we want to keep. Truncate the file to
838 * the right length, and clear chain link that points to any
839 * remaining segments (which we shall zap). NOTE: if nblocks is
840 * exactly a multiple K of RELSEG_SIZE, we will truncate the K+1st
841 * segment to 0 length but keep it. This adheres to the invariant
842 * given in the header comments.
844 BlockNumber lastsegblocks = nblocks - priorblocks;
846 if (FileTruncate(v->mdfd_vfd, (off_t) lastsegblocks * BLCKSZ) < 0)
848 (errcode_for_file_access(),
849 errmsg("could not truncate relation %u/%u/%u to %u blocks: %m",
850 reln->smgr_rnode.spcNode,
851 reln->smgr_rnode.dbNode,
852 reln->smgr_rnode.relNode,
855 register_dirty_segment(reln, v);
857 ov->mdfd_chain = NULL;
862 * We still need this segment and 0 or more blocks beyond it, so
863 * nothing to do here.
867 priorblocks += RELSEG_SIZE;
870 /* For unsegmented files, it's a lot easier */
871 if (FileTruncate(v->mdfd_vfd, (off_t) nblocks * BLCKSZ) < 0)
873 (errcode_for_file_access(),
874 errmsg("could not truncate relation %u/%u/%u to %u blocks: %m",
875 reln->smgr_rnode.spcNode,
876 reln->smgr_rnode.dbNode,
877 reln->smgr_rnode.relNode,
880 register_dirty_segment(reln, v);
885 * mdimmedsync() -- Immediately sync a relation to stable storage.
887 * Note that only writes already issued are synced; this routine knows
888 * nothing of dirty buffers that may exist inside the buffer manager.
891 mdimmedsync(SMgrRelation reln)
897 * NOTE: mdnblocks makes sure we have opened all active segments, so that
898 * fsync loop will get them all!
900 curnblk = mdnblocks(reln);
902 v = mdopen(reln, EXTENSION_FAIL);
904 #ifdef USE_SEGMENTED_FILES
907 if (FileSync(v->mdfd_vfd) < 0)
909 (errcode_for_file_access(),
910 errmsg("could not fsync segment %u of relation %u/%u/%u: %m",
912 reln->smgr_rnode.spcNode,
913 reln->smgr_rnode.dbNode,
914 reln->smgr_rnode.relNode)));
918 if (FileSync(v->mdfd_vfd) < 0)
920 (errcode_for_file_access(),
921 errmsg("could not fsync relation %u/%u/%u: %m",
922 reln->smgr_rnode.spcNode,
923 reln->smgr_rnode.dbNode,
924 reln->smgr_rnode.relNode)));
929 * mdsync() -- Sync previous writes to stable storage.
934 static bool mdsync_in_progress = false;
936 HASH_SEQ_STATUS hstat;
937 PendingOperationEntry *entry;
941 * This is only called during checkpoints, and checkpoints should only
942 * occur in processes that have created a pendingOpsTable.
944 if (!pendingOpsTable)
945 elog(ERROR, "cannot sync without a pendingOpsTable");
948 * If we are in the bgwriter, the sync had better include all fsync
949 * requests that were queued by backends up to this point. The tightest
950 * race condition that could occur is that a buffer that must be written
951 * and fsync'd for the checkpoint could have been dumped by a backend just
952 * before it was visited by BufferSync(). We know the backend will have
953 * queued an fsync request before clearing the buffer's dirtybit, so we
954 * are safe as long as we do an Absorb after completing BufferSync().
956 AbsorbFsyncRequests();
959 * To avoid excess fsync'ing (in the worst case, maybe a never-terminating
960 * checkpoint), we want to ignore fsync requests that are entered into the
961 * hashtable after this point --- they should be processed next time,
962 * instead. We use mdsync_cycle_ctr to tell old entries apart from new
963 * ones: new ones will have cycle_ctr equal to the incremented value of
966 * In normal circumstances, all entries present in the table at this point
967 * will have cycle_ctr exactly equal to the current (about to be old)
968 * value of mdsync_cycle_ctr. However, if we fail partway through the
969 * fsync'ing loop, then older values of cycle_ctr might remain when we
970 * come back here to try again. Repeated checkpoint failures would
971 * eventually wrap the counter around to the point where an old entry
972 * might appear new, causing us to skip it, possibly allowing a checkpoint
973 * to succeed that should not have. To forestall wraparound, any time the
974 * previous mdsync() failed to complete, run through the table and
975 * forcibly set cycle_ctr = mdsync_cycle_ctr.
977 * Think not to merge this loop with the main loop, as the problem is
978 * exactly that that loop may fail before having visited all the entries.
979 * From a performance point of view it doesn't matter anyway, as this path
980 * will never be taken in a system that's functioning normally.
982 if (mdsync_in_progress)
984 /* prior try failed, so update any stale cycle_ctr values */
985 hash_seq_init(&hstat, pendingOpsTable);
986 while ((entry = (PendingOperationEntry *) hash_seq_search(&hstat)) != NULL)
988 entry->cycle_ctr = mdsync_cycle_ctr;
992 /* Advance counter so that new hashtable entries are distinguishable */
995 /* Set flag to detect failure if we don't reach the end of the loop */
996 mdsync_in_progress = true;
998 /* Now scan the hashtable for fsync requests to process */
999 absorb_counter = FSYNCS_PER_ABSORB;
1000 hash_seq_init(&hstat, pendingOpsTable);
1001 while ((entry = (PendingOperationEntry *) hash_seq_search(&hstat)) != NULL)
1004 * If the entry is new then don't process it this time. Note that
1005 * "continue" bypasses the hash-remove call at the bottom of the loop.
1007 if (entry->cycle_ctr == mdsync_cycle_ctr)
1010 /* Else assert we haven't missed it */
1011 Assert((CycleCtr) (entry->cycle_ctr + 1) == mdsync_cycle_ctr);
1014 * If fsync is off then we don't have to bother opening the file at
1015 * all. (We delay checking until this point so that changing fsync on
1016 * the fly behaves sensibly.) Also, if the entry is marked canceled,
1017 * fall through to delete it.
1019 if (enableFsync && !entry->canceled)
1024 * If in bgwriter, we want to absorb pending requests every so
1025 * often to prevent overflow of the fsync request queue. It is
1026 * unspecified whether newly-added entries will be visited by
1027 * hash_seq_search, but we don't care since we don't need to
1028 * process them anyway.
1030 if (--absorb_counter <= 0)
1032 AbsorbFsyncRequests();
1033 absorb_counter = FSYNCS_PER_ABSORB;
1037 * The fsync table could contain requests to fsync segments that
1038 * have been deleted (unlinked) by the time we get to them. Rather
1039 * than just hoping an ENOENT (or EACCES on Windows) error can be
1040 * ignored, what we do on error is absorb pending requests and
1041 * then retry. Since mdunlink() queues a "revoke" message before
1042 * actually unlinking, the fsync request is guaranteed to be
1043 * marked canceled after the absorb if it really was this case.
1044 * DROP DATABASE likewise has to tell us to forget fsync requests
1045 * before it starts deletions.
1047 for (failures = 0;; failures++) /* loop exits at "break" */
1053 * Find or create an smgr hash entry for this relation. This
1054 * may seem a bit unclean -- md calling smgr? But it's really
1055 * the best solution. It ensures that the open file reference
1056 * isn't permanently leaked if we get an error here. (You may
1057 * say "but an unreferenced SMgrRelation is still a leak!" Not
1058 * really, because the only case in which a checkpoint is done
1059 * by a process that isn't about to shut down is in the
1060 * bgwriter, and it will periodically do smgrcloseall(). This
1061 * fact justifies our not closing the reln in the success path
1062 * either, which is a good thing since in non-bgwriter cases
1063 * we couldn't safely do that.) Furthermore, in many cases
1064 * the relation will have been dirtied through this same smgr
1065 * relation, and so we can save a file open/close cycle.
1067 reln = smgropen(entry->tag.rnode);
1070 * It is possible that the relation has been dropped or
1071 * truncated since the fsync request was entered. Therefore,
1072 * allow ENOENT, but only if we didn't fail already on this
1073 * file. This applies both during _mdfd_getseg() and during
1074 * FileSync, since fd.c might have closed the file behind our
1077 seg = _mdfd_getseg(reln,
1078 entry->tag.segno * ((BlockNumber) RELSEG_SIZE),
1079 false, EXTENSION_RETURN_NULL);
1081 FileSync(seg->mdfd_vfd) >= 0)
1082 break; /* success; break out of retry loop */
1085 * XXX is there any point in allowing more than one retry?
1086 * Don't see one at the moment, but easy to change the test
1089 if (!FILE_POSSIBLY_DELETED(errno) ||
1092 (errcode_for_file_access(),
1093 errmsg("could not fsync segment %u of relation %u/%u/%u: %m",
1095 entry->tag.rnode.spcNode,
1096 entry->tag.rnode.dbNode,
1097 entry->tag.rnode.relNode)));
1100 (errcode_for_file_access(),
1101 errmsg("could not fsync segment %u of relation %u/%u/%u, but retrying: %m",
1103 entry->tag.rnode.spcNode,
1104 entry->tag.rnode.dbNode,
1105 entry->tag.rnode.relNode)));
1108 * Absorb incoming requests and check to see if canceled.
1110 AbsorbFsyncRequests();
1111 absorb_counter = FSYNCS_PER_ABSORB; /* might as well... */
1113 if (entry->canceled)
1115 } /* end retry loop */
1119 * If we get here, either we fsync'd successfully, or we don't have to
1120 * because enableFsync is off, or the entry is (now) marked canceled.
1121 * Okay to delete it.
1123 if (hash_search(pendingOpsTable, &entry->tag,
1124 HASH_REMOVE, NULL) == NULL)
1125 elog(ERROR, "pendingOpsTable corrupted");
1126 } /* end loop over hashtable entries */
1128 /* Flag successful completion of mdsync */
1129 mdsync_in_progress = false;
1133 * mdpreckpt() -- Do pre-checkpoint work
1135 * To distinguish unlink requests that arrived before this checkpoint
1136 * started from those that arrived during the checkpoint, we use a cycle
1137 * counter similar to the one we use for fsync requests. That cycle
1138 * counter is incremented here.
1140 * This must be called *before* the checkpoint REDO point is determined.
1141 * That ensures that we won't delete files too soon.
1143 * Note that we can't do anything here that depends on the assumption
1144 * that the checkpoint will be completed.
1152 * In case the prior checkpoint wasn't completed, stamp all entries in the
1153 * list with the current cycle counter. Anything that's in the list at
1154 * the start of checkpoint can surely be deleted after the checkpoint is
1155 * finished, regardless of when the request was made.
1157 foreach(cell, pendingUnlinks)
1159 PendingUnlinkEntry *entry = (PendingUnlinkEntry *) lfirst(cell);
1161 entry->cycle_ctr = mdckpt_cycle_ctr;
1165 * Any unlink requests arriving after this point will be assigned the next
1166 * cycle counter, and won't be unlinked until next checkpoint.
1172 * mdpostckpt() -- Do post-checkpoint work
1174 * Remove any lingering files that can now be safely removed.
1179 while (pendingUnlinks != NIL)
1181 PendingUnlinkEntry *entry = (PendingUnlinkEntry *) linitial(pendingUnlinks);
1185 * New entries are appended to the end, so if the entry is new we've
1186 * reached the end of old entries.
1188 if (entry->cycle_ctr == mdckpt_cycle_ctr)
1191 /* Else assert we haven't missed it */
1192 Assert((CycleCtr) (entry->cycle_ctr + 1) == mdckpt_cycle_ctr);
1194 /* Unlink the file */
1195 path = relpath(entry->rnode);
1196 if (unlink(path) < 0)
1199 * There's a race condition, when the database is dropped at the
1200 * same time that we process the pending unlink requests. If the
1201 * DROP DATABASE deletes the file before we do, we will get ENOENT
1202 * here. rmtree() also has to ignore ENOENT errors, to deal with
1203 * the possibility that we delete the file first.
1205 if (errno != ENOENT)
1207 (errcode_for_file_access(),
1208 errmsg("could not remove relation %u/%u/%u: %m",
1209 entry->rnode.spcNode,
1210 entry->rnode.dbNode,
1211 entry->rnode.relNode)));
1215 pendingUnlinks = list_delete_first(pendingUnlinks);
1221 * register_dirty_segment() -- Mark a relation segment as needing fsync
1223 * If there is a local pending-ops table, just make an entry in it for
1224 * mdsync to process later. Otherwise, try to pass off the fsync request
1225 * to the background writer process. If that fails, just do the fsync
1226 * locally before returning (we expect this will not happen often enough
1227 * to be a performance problem).
1230 register_dirty_segment(SMgrRelation reln, MdfdVec *seg)
1232 if (pendingOpsTable)
1234 /* push it into local pending-ops table */
1235 RememberFsyncRequest(reln->smgr_rnode, seg->mdfd_segno);
1239 if (ForwardFsyncRequest(reln->smgr_rnode, seg->mdfd_segno))
1240 return; /* passed it off successfully */
1242 if (FileSync(seg->mdfd_vfd) < 0)
1244 (errcode_for_file_access(),
1245 errmsg("could not fsync segment %u of relation %u/%u/%u: %m",
1247 reln->smgr_rnode.spcNode,
1248 reln->smgr_rnode.dbNode,
1249 reln->smgr_rnode.relNode)));
1254 * register_unlink() -- Schedule a file to be deleted after next checkpoint
1256 * As with register_dirty_segment, this could involve either a local or
1257 * a remote pending-ops table.
1260 register_unlink(RelFileNode rnode)
1262 if (pendingOpsTable)
1264 /* push it into local pending-ops table */
1265 RememberFsyncRequest(rnode, UNLINK_RELATION_REQUEST);
1270 * Notify the bgwriter about it. If we fail to queue the request
1271 * message, we have to sleep and try again, because we can't simply
1272 * delete the file now. Ugly, but hopefully won't happen often.
1274 * XXX should we just leave the file orphaned instead?
1276 Assert(IsUnderPostmaster);
1277 while (!ForwardFsyncRequest(rnode, UNLINK_RELATION_REQUEST))
1278 pg_usleep(10000L); /* 10 msec seems a good number */
1283 * RememberFsyncRequest() -- callback from bgwriter side of fsync request
1285 * We stuff most fsync requests into the local hash table for execution
1286 * during the bgwriter's next checkpoint. UNLINK requests go into a
1287 * separate linked list, however, because they get processed separately.
1289 * The range of possible segment numbers is way less than the range of
1290 * BlockNumber, so we can reserve high values of segno for special purposes.
1292 * - FORGET_RELATION_FSYNC means to cancel pending fsyncs for a relation
1293 * - FORGET_DATABASE_FSYNC means to cancel pending fsyncs for a whole database
1294 * - UNLINK_RELATION_REQUEST is a request to delete the file after the next
1297 * (Handling the FORGET_* requests is a tad slow because the hash table has
1298 * to be searched linearly, but it doesn't seem worth rethinking the table
1299 * structure for them.)
1302 RememberFsyncRequest(RelFileNode rnode, BlockNumber segno)
1304 Assert(pendingOpsTable);
1306 if (segno == FORGET_RELATION_FSYNC)
1308 /* Remove any pending requests for the entire relation */
1309 HASH_SEQ_STATUS hstat;
1310 PendingOperationEntry *entry;
1312 hash_seq_init(&hstat, pendingOpsTable);
1313 while ((entry = (PendingOperationEntry *) hash_seq_search(&hstat)) != NULL)
1315 if (RelFileNodeEquals(entry->tag.rnode, rnode))
1317 /* Okay, cancel this entry */
1318 entry->canceled = true;
1322 else if (segno == FORGET_DATABASE_FSYNC)
1324 /* Remove any pending requests for the entire database */
1325 HASH_SEQ_STATUS hstat;
1326 PendingOperationEntry *entry;
1331 /* Remove fsync requests */
1332 hash_seq_init(&hstat, pendingOpsTable);
1333 while ((entry = (PendingOperationEntry *) hash_seq_search(&hstat)) != NULL)
1335 if (entry->tag.rnode.dbNode == rnode.dbNode)
1337 /* Okay, cancel this entry */
1338 entry->canceled = true;
1342 /* Remove unlink requests */
1344 for (cell = list_head(pendingUnlinks); cell; cell = next)
1346 PendingUnlinkEntry *entry = (PendingUnlinkEntry *) lfirst(cell);
1349 if (entry->rnode.dbNode == rnode.dbNode)
1351 pendingUnlinks = list_delete_cell(pendingUnlinks, cell, prev);
1358 else if (segno == UNLINK_RELATION_REQUEST)
1360 /* Unlink request: put it in the linked list */
1361 MemoryContext oldcxt = MemoryContextSwitchTo(MdCxt);
1362 PendingUnlinkEntry *entry;
1364 entry = palloc(sizeof(PendingUnlinkEntry));
1365 entry->rnode = rnode;
1366 entry->cycle_ctr = mdckpt_cycle_ctr;
1368 pendingUnlinks = lappend(pendingUnlinks, entry);
1370 MemoryContextSwitchTo(oldcxt);
1374 /* Normal case: enter a request to fsync this segment */
1375 PendingOperationTag key;
1376 PendingOperationEntry *entry;
1379 /* ensure any pad bytes in the hash key are zeroed */
1380 MemSet(&key, 0, sizeof(key));
1384 entry = (PendingOperationEntry *) hash_search(pendingOpsTable,
1388 /* if new or previously canceled entry, initialize it */
1389 if (!found || entry->canceled)
1391 entry->canceled = false;
1392 entry->cycle_ctr = mdsync_cycle_ctr;
1396 * NB: it's intentional that we don't change cycle_ctr if the entry
1397 * already exists. The fsync request must be treated as old, even
1398 * though the new request will be satisfied too by any subsequent
1401 * However, if the entry is present but is marked canceled, we should
1402 * act just as though it wasn't there. The only case where this could
1403 * happen would be if a file had been deleted, we received but did not
1404 * yet act on the cancel request, and the same relfilenode was then
1405 * assigned to a new file. We mustn't lose the new request, but it
1406 * should be considered new not old.
1412 * ForgetRelationFsyncRequests -- forget any fsyncs for a rel
1415 ForgetRelationFsyncRequests(RelFileNode rnode)
1417 if (pendingOpsTable)
1419 /* standalone backend or startup process: fsync state is local */
1420 RememberFsyncRequest(rnode, FORGET_RELATION_FSYNC);
1422 else if (IsUnderPostmaster)
1425 * Notify the bgwriter about it. If we fail to queue the revoke
1426 * message, we have to sleep and try again ... ugly, but hopefully
1427 * won't happen often.
1429 * XXX should we CHECK_FOR_INTERRUPTS in this loop? Escaping with an
1430 * error would leave the no-longer-used file still present on disk,
1431 * which would be bad, so I'm inclined to assume that the bgwriter
1432 * will always empty the queue soon.
1434 while (!ForwardFsyncRequest(rnode, FORGET_RELATION_FSYNC))
1435 pg_usleep(10000L); /* 10 msec seems a good number */
1438 * Note we don't wait for the bgwriter to actually absorb the revoke
1439 * message; see mdsync() for the implications.
1445 * ForgetDatabaseFsyncRequests -- forget any fsyncs and unlinks for a DB
1448 ForgetDatabaseFsyncRequests(Oid dbid)
1452 rnode.dbNode = dbid;
1456 if (pendingOpsTable)
1458 /* standalone backend or startup process: fsync state is local */
1459 RememberFsyncRequest(rnode, FORGET_DATABASE_FSYNC);
1461 else if (IsUnderPostmaster)
1463 /* see notes in ForgetRelationFsyncRequests */
1464 while (!ForwardFsyncRequest(rnode, FORGET_DATABASE_FSYNC))
1465 pg_usleep(10000L); /* 10 msec seems a good number */
1471 * _fdvec_alloc() -- Make a MdfdVec object.
1476 return (MdfdVec *) MemoryContextAlloc(MdCxt, sizeof(MdfdVec));
1479 #ifdef USE_SEGMENTED_FILES
1482 * Open the specified segment of the relation,
1483 * and make a MdfdVec object for it. Returns NULL on failure.
1486 _mdfd_openseg(SMgrRelation reln, BlockNumber segno, int oflags)
1493 path = relpath(reln->smgr_rnode);
1497 /* be sure we have enough space for the '.segno' */
1498 fullpath = (char *) palloc(strlen(path) + 12);
1499 sprintf(fullpath, "%s.%u", path, segno);
1506 fd = PathNameOpenFile(fullpath, O_RDWR | PG_BINARY | oflags, 0600);
1513 /* allocate an mdfdvec entry for it */
1516 /* fill the entry */
1518 v->mdfd_segno = segno;
1519 v->mdfd_chain = NULL;
1520 Assert(_mdnblocks(reln, v) <= ((BlockNumber) RELSEG_SIZE));
1525 #endif /* USE_SEGMENTED_FILES */
1528 * _mdfd_getseg() -- Find the segment of the relation holding the
1531 * If the segment doesn't exist, we ereport, return NULL, or create the
1532 * segment, according to "behavior". Note: isTemp need only be correct
1533 * in the EXTENSION_CREATE case.
1536 _mdfd_getseg(SMgrRelation reln, BlockNumber blkno, bool isTemp,
1537 ExtensionBehavior behavior)
1539 MdfdVec *v = mdopen(reln, behavior);
1541 #ifdef USE_SEGMENTED_FILES
1542 BlockNumber targetseg;
1543 BlockNumber nextsegno;
1546 return NULL; /* only possible if EXTENSION_RETURN_NULL */
1548 targetseg = blkno / ((BlockNumber) RELSEG_SIZE);
1549 for (nextsegno = 1; nextsegno <= targetseg; nextsegno++)
1551 Assert(nextsegno == v->mdfd_segno + 1);
1553 if (v->mdfd_chain == NULL)
1556 * Normally we will create new segments only if authorized by the
1557 * caller (i.e., we are doing mdextend()). But when doing WAL
1558 * recovery, create segments anyway; this allows cases such as
1559 * replaying WAL data that has a write into a high-numbered
1560 * segment of a relation that was later deleted. We want to go
1561 * ahead and create the segments so we can finish out the replay.
1563 * We have to maintain the invariant that segments before the last
1564 * active segment are of size RELSEG_SIZE; therefore, pad them out
1565 * with zeroes if needed. (This only matters if caller is
1566 * extending the relation discontiguously, but that can happen in
1569 if (behavior == EXTENSION_CREATE || InRecovery)
1571 if (_mdnblocks(reln, v) < RELSEG_SIZE)
1573 char *zerobuf = palloc0(BLCKSZ);
1575 mdextend(reln, nextsegno * ((BlockNumber) RELSEG_SIZE) - 1,
1579 v->mdfd_chain = _mdfd_openseg(reln, nextsegno, O_CREAT);
1583 /* We won't create segment if not existent */
1584 v->mdfd_chain = _mdfd_openseg(reln, nextsegno, 0);
1586 if (v->mdfd_chain == NULL)
1588 if (behavior == EXTENSION_RETURN_NULL &&
1589 FILE_POSSIBLY_DELETED(errno))
1592 (errcode_for_file_access(),
1593 errmsg("could not open segment %u of relation %u/%u/%u (target block %u): %m",
1595 reln->smgr_rnode.spcNode,
1596 reln->smgr_rnode.dbNode,
1597 reln->smgr_rnode.relNode,
1609 * Get number of blocks present in a single disk file
1612 _mdnblocks(SMgrRelation reln, MdfdVec *seg)
1616 len = FileSeek(seg->mdfd_vfd, 0L, SEEK_END);
1619 (errcode_for_file_access(),
1620 errmsg("could not seek to end of segment %u of relation %u/%u/%u: %m",
1622 reln->smgr_rnode.spcNode,
1623 reln->smgr_rnode.dbNode,
1624 reln->smgr_rnode.relNode)));
1625 /* note that this calculation will ignore any partial block at EOF */
1626 return (BlockNumber) (len / BLCKSZ);