1 /*-------------------------------------------------------------------------
4 * This code manages relations that reside on magnetic disk.
6 * Portions Copyright (c) 1996-2007, 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 1.130 2007/11/15 20:36:40 tgl 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 * Defining LET_OS_MANAGE_FILESIZE disables the segmentation logic,
93 * for use on machines that support large files. Beware that that
94 * code has not been tested in a long time and is probably bit-rotted.
97 typedef struct _MdfdVec
99 File mdfd_vfd; /* fd number in fd.c's pool */
100 BlockNumber mdfd_segno; /* segment number, from 0 */
101 #ifndef LET_OS_MANAGE_FILESIZE /* for large relations */
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 #ifndef LET_OS_MANAGE_FILESIZE
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 #ifndef LET_OS_MANAGE_FILESIZE
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 relation,
306 * 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
319 ret = truncate(path, 0);
322 if (!isRedo || errno != ENOENT)
324 (errcode_for_file_access(),
325 errmsg("could not remove relation %u/%u/%u: %m",
331 #ifndef LET_OS_MANAGE_FILESIZE
332 /* Delete the additional segments, if any */
335 char *segpath = (char *) palloc(strlen(path) + 12);
339 * Note that because we loop until getting ENOENT, we will
340 * correctly remove all inactive segments as well as active ones.
342 for (segno = 1;; segno++)
344 sprintf(segpath, "%s.%u", path, segno);
345 if (unlink(segpath) < 0)
347 /* ENOENT is expected after the last segment... */
350 (errcode_for_file_access(),
351 errmsg("could not remove segment %u of relation %u/%u/%u: %m",
365 /* Register request to unlink first segment later */
367 register_unlink(rnode);
371 * mdextend() -- Add a block to the specified relation.
373 * The semantics are nearly the same as mdwrite(): write at the
374 * specified position. However, this is to be used for the case of
375 * extending a relation (i.e., blocknum is at or beyond the current
376 * EOF). Note that we assume writing a block beyond current EOF
377 * causes intervening file space to become filled with zeroes.
380 mdextend(SMgrRelation reln, BlockNumber blocknum, char *buffer, bool isTemp)
386 /* This assert is too expensive to have on normally ... */
387 #ifdef CHECK_WRITE_VS_EXTEND
388 Assert(blocknum >= mdnblocks(reln));
392 * If a relation manages to grow to 2^32-1 blocks, refuse to extend it
393 * any more --- we mustn't create a block whose number
394 * actually is InvalidBlockNumber.
396 if (blocknum == InvalidBlockNumber)
398 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
399 errmsg("cannot extend relation %u/%u/%u beyond %u blocks",
400 reln->smgr_rnode.spcNode,
401 reln->smgr_rnode.dbNode,
402 reln->smgr_rnode.relNode,
403 InvalidBlockNumber)));
405 v = _mdfd_getseg(reln, blocknum, isTemp, EXTENSION_CREATE);
407 #ifndef LET_OS_MANAGE_FILESIZE
408 seekpos = (long) (BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE)));
409 Assert(seekpos < BLCKSZ * RELSEG_SIZE);
411 seekpos = (long) (BLCKSZ * (blocknum));
415 * Note: because caller usually obtained blocknum by calling mdnblocks,
416 * which did a seek(SEEK_END), this seek is often redundant and will be
417 * optimized away by fd.c. It's not redundant, however, if there is a
418 * partial page at the end of the file. In that case we want to try to
419 * overwrite the partial page with a full page. It's also not redundant
420 * if bufmgr.c had to dump another buffer of the same file to make room
421 * for the new page's buffer.
423 if (FileSeek(v->mdfd_vfd, seekpos, SEEK_SET) != seekpos)
425 (errcode_for_file_access(),
426 errmsg("could not seek to block %u of relation %u/%u/%u: %m",
428 reln->smgr_rnode.spcNode,
429 reln->smgr_rnode.dbNode,
430 reln->smgr_rnode.relNode)));
432 if ((nbytes = FileWrite(v->mdfd_vfd, buffer, BLCKSZ)) != BLCKSZ)
436 (errcode_for_file_access(),
437 errmsg("could not extend relation %u/%u/%u: %m",
438 reln->smgr_rnode.spcNode,
439 reln->smgr_rnode.dbNode,
440 reln->smgr_rnode.relNode),
441 errhint("Check free disk space.")));
442 /* short write: complain appropriately */
444 (errcode(ERRCODE_DISK_FULL),
445 errmsg("could not extend relation %u/%u/%u: wrote only %d of %d bytes at block %u",
446 reln->smgr_rnode.spcNode,
447 reln->smgr_rnode.dbNode,
448 reln->smgr_rnode.relNode,
449 nbytes, BLCKSZ, blocknum),
450 errhint("Check free disk space.")));
454 register_dirty_segment(reln, v);
456 #ifndef LET_OS_MANAGE_FILESIZE
457 Assert(_mdnblocks(reln, v) <= ((BlockNumber) RELSEG_SIZE));
462 * mdopen() -- Open the specified relation.
464 * Note we only open the first segment, when there are multiple segments.
466 * If first segment is not present, either ereport or return NULL according
467 * to "behavior". We treat EXTENSION_CREATE the same as EXTENSION_FAIL;
468 * EXTENSION_CREATE means it's OK to extend an existing relation, not to
469 * invent one out of whole cloth.
472 mdopen(SMgrRelation reln, ExtensionBehavior behavior)
478 /* No work if already open */
482 path = relpath(reln->smgr_rnode);
484 fd = PathNameOpenFile(path, O_RDWR | PG_BINARY, 0600);
489 * During bootstrap, there are cases where a system relation will be
490 * accessed (by internal backend processes) before the bootstrap
491 * script nominally creates it. Therefore, accept mdopen() as a
492 * substitute for mdcreate() in bootstrap mode only. (See mdcreate)
494 if (IsBootstrapProcessingMode())
495 fd = PathNameOpenFile(path, O_RDWR | O_CREAT | O_EXCL | PG_BINARY, 0600);
499 if (behavior == EXTENSION_RETURN_NULL &&
500 FILE_POSSIBLY_DELETED(errno))
503 (errcode_for_file_access(),
504 errmsg("could not open relation %u/%u/%u: %m",
505 reln->smgr_rnode.spcNode,
506 reln->smgr_rnode.dbNode,
507 reln->smgr_rnode.relNode)));
513 reln->md_fd = mdfd = _fdvec_alloc();
516 mdfd->mdfd_segno = 0;
517 #ifndef LET_OS_MANAGE_FILESIZE
518 mdfd->mdfd_chain = NULL;
519 Assert(_mdnblocks(reln, mdfd) <= ((BlockNumber) RELSEG_SIZE));
526 * mdclose() -- Close the specified relation, if it isn't closed already.
529 mdclose(SMgrRelation reln)
531 MdfdVec *v = reln->md_fd;
533 /* No work if already closed */
537 reln->md_fd = NULL; /* prevent dangling pointer after error */
539 #ifndef LET_OS_MANAGE_FILESIZE
544 /* if not closed already */
545 if (v->mdfd_vfd >= 0)
546 FileClose(v->mdfd_vfd);
547 /* Now free vector */
552 if (v->mdfd_vfd >= 0)
553 FileClose(v->mdfd_vfd);
559 * mdread() -- Read the specified block from a relation.
562 mdread(SMgrRelation reln, BlockNumber blocknum, char *buffer)
568 v = _mdfd_getseg(reln, blocknum, false, EXTENSION_FAIL);
570 #ifndef LET_OS_MANAGE_FILESIZE
571 seekpos = (long) (BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE)));
572 Assert(seekpos < BLCKSZ * RELSEG_SIZE);
574 seekpos = (long) (BLCKSZ * (blocknum));
577 if (FileSeek(v->mdfd_vfd, seekpos, SEEK_SET) != seekpos)
579 (errcode_for_file_access(),
580 errmsg("could not seek to block %u of relation %u/%u/%u: %m",
582 reln->smgr_rnode.spcNode,
583 reln->smgr_rnode.dbNode,
584 reln->smgr_rnode.relNode)));
586 if ((nbytes = FileRead(v->mdfd_vfd, buffer, BLCKSZ)) != BLCKSZ)
590 (errcode_for_file_access(),
591 errmsg("could not read block %u of relation %u/%u/%u: %m",
593 reln->smgr_rnode.spcNode,
594 reln->smgr_rnode.dbNode,
595 reln->smgr_rnode.relNode)));
597 * Short read: we are at or past EOF, or we read a partial block at
598 * EOF. Normally this is an error; upper levels should never try to
599 * read a nonexistent block. However, if zero_damaged_pages is ON
600 * or we are InRecovery, we should instead return zeroes without
601 * complaining. This allows, for example, the case of trying to
602 * update a block that was later truncated away.
604 if (zero_damaged_pages || InRecovery)
605 MemSet(buffer, 0, BLCKSZ);
608 (errcode(ERRCODE_DATA_CORRUPTED),
609 errmsg("could not read block %u of relation %u/%u/%u: read only %d of %d bytes",
611 reln->smgr_rnode.spcNode,
612 reln->smgr_rnode.dbNode,
613 reln->smgr_rnode.relNode,
619 * mdwrite() -- Write the supplied block at the appropriate location.
621 * This is to be used only for updating already-existing blocks of a
622 * relation (ie, those before the current EOF). To extend a relation,
626 mdwrite(SMgrRelation reln, BlockNumber blocknum, char *buffer, bool isTemp)
632 /* This assert is too expensive to have on normally ... */
633 #ifdef CHECK_WRITE_VS_EXTEND
634 Assert(blocknum < mdnblocks(reln));
637 v = _mdfd_getseg(reln, blocknum, isTemp, EXTENSION_FAIL);
639 #ifndef LET_OS_MANAGE_FILESIZE
640 seekpos = (long) (BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE)));
641 Assert(seekpos < BLCKSZ * RELSEG_SIZE);
643 seekpos = (long) (BLCKSZ * (blocknum));
646 if (FileSeek(v->mdfd_vfd, seekpos, SEEK_SET) != seekpos)
648 (errcode_for_file_access(),
649 errmsg("could not seek to block %u of relation %u/%u/%u: %m",
651 reln->smgr_rnode.spcNode,
652 reln->smgr_rnode.dbNode,
653 reln->smgr_rnode.relNode)));
655 if ((nbytes = FileWrite(v->mdfd_vfd, buffer, BLCKSZ)) != BLCKSZ)
659 (errcode_for_file_access(),
660 errmsg("could not write block %u of relation %u/%u/%u: %m",
662 reln->smgr_rnode.spcNode,
663 reln->smgr_rnode.dbNode,
664 reln->smgr_rnode.relNode)));
665 /* short write: complain appropriately */
667 (errcode(ERRCODE_DISK_FULL),
668 errmsg("could not write block %u of relation %u/%u/%u: wrote only %d of %d bytes",
670 reln->smgr_rnode.spcNode,
671 reln->smgr_rnode.dbNode,
672 reln->smgr_rnode.relNode,
674 errhint("Check free disk space.")));
678 register_dirty_segment(reln, v);
682 * mdnblocks() -- Get the number of blocks stored in a relation.
684 * Important side effect: all active segments of the relation are opened
685 * and added to the mdfd_chain list. If this routine has not been
686 * called, then only segments up to the last one actually touched
687 * are present in the chain.
690 mdnblocks(SMgrRelation reln)
692 MdfdVec *v = mdopen(reln, EXTENSION_FAIL);
694 #ifndef LET_OS_MANAGE_FILESIZE
696 BlockNumber segno = 0;
699 * Skip through any segments that aren't the last one, to avoid redundant
700 * seeks on them. We have previously verified that these segments are
701 * exactly RELSEG_SIZE long, and it's useless to recheck that each time.
703 * NOTE: this assumption could only be wrong if another backend has
704 * truncated the relation. We rely on higher code levels to handle that
705 * scenario by closing and re-opening the md fd, which is handled via
706 * relcache flush. (Since the bgwriter doesn't participate in relcache
707 * flush, it could have segment chain entries for inactive segments;
708 * that's OK because the bgwriter never needs to compute relation size.)
710 while (v->mdfd_chain != NULL)
718 nblocks = _mdnblocks(reln, v);
719 if (nblocks > ((BlockNumber) RELSEG_SIZE))
720 elog(FATAL, "segment too big");
721 if (nblocks < ((BlockNumber) RELSEG_SIZE))
722 return (segno * ((BlockNumber) RELSEG_SIZE)) + nblocks;
725 * If segment is exactly RELSEG_SIZE, advance to next one.
729 if (v->mdfd_chain == NULL)
732 * Because we pass O_CREAT, we will create the next segment (with
733 * zero length) immediately, if the last segment is of length
734 * RELSEG_SIZE. While perhaps not strictly necessary, this keeps
737 v->mdfd_chain = _mdfd_openseg(reln, segno, O_CREAT);
738 if (v->mdfd_chain == NULL)
740 (errcode_for_file_access(),
741 errmsg("could not open segment %u of relation %u/%u/%u: %m",
743 reln->smgr_rnode.spcNode,
744 reln->smgr_rnode.dbNode,
745 reln->smgr_rnode.relNode)));
751 return _mdnblocks(reln, v);
756 * mdtruncate() -- Truncate relation to specified number of blocks.
759 mdtruncate(SMgrRelation reln, BlockNumber nblocks, bool isTemp)
764 #ifndef LET_OS_MANAGE_FILESIZE
765 BlockNumber priorblocks;
769 * NOTE: mdnblocks makes sure we have opened all active segments, so
770 * that truncation loop will get them all!
772 curnblk = mdnblocks(reln);
773 if (nblocks > curnblk)
775 /* Bogus request ... but no complaint if InRecovery */
779 (errmsg("could not truncate relation %u/%u/%u to %u blocks: it's only %u blocks now",
780 reln->smgr_rnode.spcNode,
781 reln->smgr_rnode.dbNode,
782 reln->smgr_rnode.relNode,
785 if (nblocks == curnblk)
786 return; /* no work */
788 v = mdopen(reln, EXTENSION_FAIL);
790 #ifndef LET_OS_MANAGE_FILESIZE
796 if (priorblocks > nblocks)
799 * This segment is no longer active (and has already been
800 * unlinked from the mdfd_chain). We truncate the file, but do
801 * not delete it, for reasons explained in the header comments.
803 if (FileTruncate(v->mdfd_vfd, 0) < 0)
805 (errcode_for_file_access(),
806 errmsg("could not truncate relation %u/%u/%u to %u blocks: %m",
807 reln->smgr_rnode.spcNode,
808 reln->smgr_rnode.dbNode,
809 reln->smgr_rnode.relNode,
812 register_dirty_segment(reln, v);
814 Assert(ov != reln->md_fd); /* we never drop the 1st segment */
817 else if (priorblocks + ((BlockNumber) RELSEG_SIZE) > nblocks)
820 * This is the last segment we want to keep. Truncate the file to
821 * the right length, and clear chain link that points to any
822 * remaining segments (which we shall zap). NOTE: if nblocks is
823 * exactly a multiple K of RELSEG_SIZE, we will truncate the K+1st
824 * segment to 0 length but keep it. This adheres to the invariant
825 * given in the header comments.
827 BlockNumber lastsegblocks = nblocks - priorblocks;
829 if (FileTruncate(v->mdfd_vfd, lastsegblocks * BLCKSZ) < 0)
831 (errcode_for_file_access(),
832 errmsg("could not truncate relation %u/%u/%u to %u blocks: %m",
833 reln->smgr_rnode.spcNode,
834 reln->smgr_rnode.dbNode,
835 reln->smgr_rnode.relNode,
838 register_dirty_segment(reln, v);
840 ov->mdfd_chain = NULL;
845 * We still need this segment and 0 or more blocks beyond it, so
846 * nothing to do here.
850 priorblocks += RELSEG_SIZE;
853 if (FileTruncate(v->mdfd_vfd, nblocks * BLCKSZ) < 0)
855 (errcode_for_file_access(),
856 errmsg("could not truncate relation %u/%u/%u to %u blocks: %m",
857 reln->smgr_rnode.spcNode,
858 reln->smgr_rnode.dbNode,
859 reln->smgr_rnode.relNode,
862 register_dirty_segment(reln, v);
867 * mdimmedsync() -- Immediately sync a relation to stable storage.
869 * Note that only writes already issued are synced; this routine knows
870 * nothing of dirty buffers that may exist inside the buffer manager.
873 mdimmedsync(SMgrRelation reln)
879 * NOTE: mdnblocks makes sure we have opened all active segments, so
880 * that fsync loop will get them all!
882 curnblk = mdnblocks(reln);
884 v = mdopen(reln, EXTENSION_FAIL);
886 #ifndef LET_OS_MANAGE_FILESIZE
889 if (FileSync(v->mdfd_vfd) < 0)
891 (errcode_for_file_access(),
892 errmsg("could not fsync segment %u of relation %u/%u/%u: %m",
894 reln->smgr_rnode.spcNode,
895 reln->smgr_rnode.dbNode,
896 reln->smgr_rnode.relNode)));
900 if (FileSync(v->mdfd_vfd) < 0)
902 (errcode_for_file_access(),
903 errmsg("could not fsync segment %u of relation %u/%u/%u: %m",
905 reln->smgr_rnode.spcNode,
906 reln->smgr_rnode.dbNode,
907 reln->smgr_rnode.relNode)));
912 * mdsync() -- Sync previous writes to stable storage.
917 static bool mdsync_in_progress = false;
919 HASH_SEQ_STATUS hstat;
920 PendingOperationEntry *entry;
924 * This is only called during checkpoints, and checkpoints should only
925 * occur in processes that have created a pendingOpsTable.
927 if (!pendingOpsTable)
928 elog(ERROR, "cannot sync without a pendingOpsTable");
931 * If we are in the bgwriter, the sync had better include all fsync
932 * requests that were queued by backends up to this point. The tightest
933 * race condition that could occur is that a buffer that must be written
934 * and fsync'd for the checkpoint could have been dumped by a backend
935 * just before it was visited by BufferSync(). We know the backend will
936 * have queued an fsync request before clearing the buffer's dirtybit,
937 * so we are safe as long as we do an Absorb after completing BufferSync().
939 AbsorbFsyncRequests();
942 * To avoid excess fsync'ing (in the worst case, maybe a never-terminating
943 * checkpoint), we want to ignore fsync requests that are entered into the
944 * hashtable after this point --- they should be processed next time,
945 * instead. We use mdsync_cycle_ctr to tell old entries apart from new
946 * ones: new ones will have cycle_ctr equal to the incremented value of
949 * In normal circumstances, all entries present in the table at this
950 * point will have cycle_ctr exactly equal to the current (about to be old)
951 * value of mdsync_cycle_ctr. However, if we fail partway through the
952 * fsync'ing loop, then older values of cycle_ctr might remain when we
953 * come back here to try again. Repeated checkpoint failures would
954 * eventually wrap the counter around to the point where an old entry
955 * might appear new, causing us to skip it, possibly allowing a checkpoint
956 * to succeed that should not have. To forestall wraparound, any time
957 * the previous mdsync() failed to complete, run through the table and
958 * forcibly set cycle_ctr = mdsync_cycle_ctr.
960 * Think not to merge this loop with the main loop, as the problem is
961 * exactly that that loop may fail before having visited all the entries.
962 * From a performance point of view it doesn't matter anyway, as this
963 * path will never be taken in a system that's functioning normally.
965 if (mdsync_in_progress)
967 /* prior try failed, so update any stale cycle_ctr values */
968 hash_seq_init(&hstat, pendingOpsTable);
969 while ((entry = (PendingOperationEntry *) hash_seq_search(&hstat)) != NULL)
971 entry->cycle_ctr = mdsync_cycle_ctr;
975 /* Advance counter so that new hashtable entries are distinguishable */
978 /* Set flag to detect failure if we don't reach the end of the loop */
979 mdsync_in_progress = true;
981 /* Now scan the hashtable for fsync requests to process */
982 absorb_counter = FSYNCS_PER_ABSORB;
983 hash_seq_init(&hstat, pendingOpsTable);
984 while ((entry = (PendingOperationEntry *) hash_seq_search(&hstat)) != NULL)
987 * If the entry is new then don't process it this time. Note that
988 * "continue" bypasses the hash-remove call at the bottom of the loop.
990 if (entry->cycle_ctr == mdsync_cycle_ctr)
993 /* Else assert we haven't missed it */
994 Assert((CycleCtr) (entry->cycle_ctr + 1) == mdsync_cycle_ctr);
997 * If fsync is off then we don't have to bother opening the file
998 * at all. (We delay checking until this point so that changing
999 * fsync on the fly behaves sensibly.) Also, if the entry is
1000 * marked canceled, fall through to delete it.
1002 if (enableFsync && !entry->canceled)
1007 * If in bgwriter, we want to absorb pending requests every so
1008 * often to prevent overflow of the fsync request queue. It is
1009 * unspecified whether newly-added entries will be visited by
1010 * hash_seq_search, but we don't care since we don't need to
1011 * process them anyway.
1013 if (--absorb_counter <= 0)
1015 AbsorbFsyncRequests();
1016 absorb_counter = FSYNCS_PER_ABSORB;
1020 * The fsync table could contain requests to fsync segments that
1021 * have been deleted (unlinked) by the time we get to them.
1022 * Rather than just hoping an ENOENT (or EACCES on Windows) error
1023 * can be ignored, what we do on error is absorb pending requests
1024 * and then retry. Since mdunlink() queues a "revoke" message
1025 * before actually unlinking, the fsync request is guaranteed to
1026 * be marked canceled after the absorb if it really was this case.
1027 * DROP DATABASE likewise has to tell us to forget fsync requests
1028 * before it starts deletions.
1030 for (failures = 0; ; failures++) /* loop exits at "break" */
1036 * Find or create an smgr hash entry for this relation. This
1037 * may seem a bit unclean -- md calling smgr? But it's really
1038 * the best solution. It ensures that the open file reference
1039 * isn't permanently leaked if we get an error here. (You may
1040 * say "but an unreferenced SMgrRelation is still a leak!" Not
1041 * really, because the only case in which a checkpoint is done
1042 * by a process that isn't about to shut down is in the
1043 * bgwriter, and it will periodically do smgrcloseall(). This
1044 * fact justifies our not closing the reln in the success path
1045 * either, which is a good thing since in non-bgwriter cases
1046 * we couldn't safely do that.) Furthermore, in many cases
1047 * the relation will have been dirtied through this same smgr
1048 * relation, and so we can save a file open/close cycle.
1050 reln = smgropen(entry->tag.rnode);
1053 * It is possible that the relation has been dropped or
1054 * truncated since the fsync request was entered. Therefore,
1055 * allow ENOENT, but only if we didn't fail already on
1056 * this file. This applies both during _mdfd_getseg() and
1057 * during FileSync, since fd.c might have closed the file
1060 seg = _mdfd_getseg(reln,
1061 entry->tag.segno * ((BlockNumber) RELSEG_SIZE),
1062 false, EXTENSION_RETURN_NULL);
1064 FileSync(seg->mdfd_vfd) >= 0)
1065 break; /* success; break out of retry loop */
1068 * XXX is there any point in allowing more than one retry?
1069 * Don't see one at the moment, but easy to change the
1072 if (!FILE_POSSIBLY_DELETED(errno) ||
1075 (errcode_for_file_access(),
1076 errmsg("could not fsync segment %u of relation %u/%u/%u: %m",
1078 entry->tag.rnode.spcNode,
1079 entry->tag.rnode.dbNode,
1080 entry->tag.rnode.relNode)));
1083 (errcode_for_file_access(),
1084 errmsg("could not fsync segment %u of relation %u/%u/%u, but retrying: %m",
1086 entry->tag.rnode.spcNode,
1087 entry->tag.rnode.dbNode,
1088 entry->tag.rnode.relNode)));
1091 * Absorb incoming requests and check to see if canceled.
1093 AbsorbFsyncRequests();
1094 absorb_counter = FSYNCS_PER_ABSORB; /* might as well... */
1096 if (entry->canceled)
1098 } /* end retry loop */
1102 * If we get here, either we fsync'd successfully, or we don't have
1103 * to because enableFsync is off, or the entry is (now) marked
1104 * canceled. Okay to delete it.
1106 if (hash_search(pendingOpsTable, &entry->tag,
1107 HASH_REMOVE, NULL) == NULL)
1108 elog(ERROR, "pendingOpsTable corrupted");
1109 } /* end loop over hashtable entries */
1111 /* Flag successful completion of mdsync */
1112 mdsync_in_progress = false;
1116 * mdpreckpt() -- Do pre-checkpoint work
1118 * To distinguish unlink requests that arrived before this checkpoint
1119 * started from those that arrived during the checkpoint, we use a cycle
1120 * counter similar to the one we use for fsync requests. That cycle
1121 * counter is incremented here.
1123 * This must be called *before* the checkpoint REDO point is determined.
1124 * That ensures that we won't delete files too soon.
1126 * Note that we can't do anything here that depends on the assumption
1127 * that the checkpoint will be completed.
1135 * In case the prior checkpoint wasn't completed, stamp all entries in
1136 * the list with the current cycle counter. Anything that's in the
1137 * list at the start of checkpoint can surely be deleted after the
1138 * checkpoint is finished, regardless of when the request was made.
1140 foreach(cell, pendingUnlinks)
1142 PendingUnlinkEntry *entry = (PendingUnlinkEntry *) lfirst(cell);
1144 entry->cycle_ctr = mdckpt_cycle_ctr;
1148 * Any unlink requests arriving after this point will be assigned the
1149 * next cycle counter, and won't be unlinked until next checkpoint.
1155 * mdpostckpt() -- Do post-checkpoint work
1157 * Remove any lingering files that can now be safely removed.
1162 while (pendingUnlinks != NIL)
1164 PendingUnlinkEntry *entry = (PendingUnlinkEntry *) linitial(pendingUnlinks);
1168 * New entries are appended to the end, so if the entry is new
1169 * we've reached the end of old entries.
1171 if (entry->cycle_ctr == mdsync_cycle_ctr)
1174 /* Else assert we haven't missed it */
1175 Assert((CycleCtr) (entry->cycle_ctr + 1) == mdckpt_cycle_ctr);
1177 /* Unlink the file */
1178 path = relpath(entry->rnode);
1179 if (unlink(path) < 0)
1182 * ENOENT shouldn't happen either, but it doesn't really matter
1183 * because we would've deleted it now anyway.
1185 if (errno != ENOENT)
1187 (errcode_for_file_access(),
1188 errmsg("could not remove relation %u/%u/%u: %m",
1189 entry->rnode.spcNode,
1190 entry->rnode.dbNode,
1191 entry->rnode.relNode)));
1195 pendingUnlinks = list_delete_first(pendingUnlinks);
1201 * register_dirty_segment() -- Mark a relation segment as needing fsync
1203 * If there is a local pending-ops table, just make an entry in it for
1204 * mdsync to process later. Otherwise, try to pass off the fsync request
1205 * to the background writer process. If that fails, just do the fsync
1206 * locally before returning (we expect this will not happen often enough
1207 * to be a performance problem).
1210 register_dirty_segment(SMgrRelation reln, MdfdVec *seg)
1212 if (pendingOpsTable)
1214 /* push it into local pending-ops table */
1215 RememberFsyncRequest(reln->smgr_rnode, seg->mdfd_segno);
1219 if (ForwardFsyncRequest(reln->smgr_rnode, seg->mdfd_segno))
1220 return; /* passed it off successfully */
1222 if (FileSync(seg->mdfd_vfd) < 0)
1224 (errcode_for_file_access(),
1225 errmsg("could not fsync segment %u of relation %u/%u/%u: %m",
1227 reln->smgr_rnode.spcNode,
1228 reln->smgr_rnode.dbNode,
1229 reln->smgr_rnode.relNode)));
1234 * register_unlink() -- Schedule a file to be deleted after next checkpoint
1236 * As with register_dirty_segment, this could involve either a local or
1237 * a remote pending-ops table.
1240 register_unlink(RelFileNode rnode)
1242 if (pendingOpsTable)
1244 /* push it into local pending-ops table */
1245 RememberFsyncRequest(rnode, UNLINK_RELATION_REQUEST);
1250 * Notify the bgwriter about it. If we fail to queue the request
1251 * message, we have to sleep and try again, because we can't simply
1252 * delete the file now. Ugly, but hopefully won't happen often.
1254 * XXX should we just leave the file orphaned instead?
1256 Assert(IsUnderPostmaster);
1257 while (!ForwardFsyncRequest(rnode, UNLINK_RELATION_REQUEST))
1258 pg_usleep(10000L); /* 10 msec seems a good number */
1263 * RememberFsyncRequest() -- callback from bgwriter side of fsync request
1265 * We stuff most fsync requests into the local hash table for execution
1266 * during the bgwriter's next checkpoint. UNLINK requests go into a
1267 * separate linked list, however, because they get processed separately.
1269 * The range of possible segment numbers is way less than the range of
1270 * BlockNumber, so we can reserve high values of segno for special purposes.
1272 * - FORGET_RELATION_FSYNC means to cancel pending fsyncs for a relation
1273 * - FORGET_DATABASE_FSYNC means to cancel pending fsyncs for a whole database
1274 * - UNLINK_RELATION_REQUEST is a request to delete the file after the next
1277 * (Handling the FORGET_* requests is a tad slow because the hash table has
1278 * to be searched linearly, but it doesn't seem worth rethinking the table
1279 * structure for them.)
1282 RememberFsyncRequest(RelFileNode rnode, BlockNumber segno)
1284 Assert(pendingOpsTable);
1286 if (segno == FORGET_RELATION_FSYNC)
1288 /* Remove any pending requests for the entire relation */
1289 HASH_SEQ_STATUS hstat;
1290 PendingOperationEntry *entry;
1292 hash_seq_init(&hstat, pendingOpsTable);
1293 while ((entry = (PendingOperationEntry *) hash_seq_search(&hstat)) != NULL)
1295 if (RelFileNodeEquals(entry->tag.rnode, rnode))
1297 /* Okay, cancel this entry */
1298 entry->canceled = true;
1302 else if (segno == FORGET_DATABASE_FSYNC)
1304 /* Remove any pending requests for the entire database */
1305 HASH_SEQ_STATUS hstat;
1306 PendingOperationEntry *entry;
1308 hash_seq_init(&hstat, pendingOpsTable);
1309 while ((entry = (PendingOperationEntry *) hash_seq_search(&hstat)) != NULL)
1311 if (entry->tag.rnode.dbNode == rnode.dbNode)
1313 /* Okay, cancel this entry */
1314 entry->canceled = true;
1318 else if (segno == UNLINK_RELATION_REQUEST)
1320 /* Unlink request: put it in the linked list */
1321 MemoryContext oldcxt = MemoryContextSwitchTo(MdCxt);
1322 PendingUnlinkEntry *entry;
1324 entry = palloc(sizeof(PendingUnlinkEntry));
1325 entry->rnode = rnode;
1326 entry->cycle_ctr = mdckpt_cycle_ctr;
1328 pendingUnlinks = lappend(pendingUnlinks, entry);
1330 MemoryContextSwitchTo(oldcxt);
1334 /* Normal case: enter a request to fsync this segment */
1335 PendingOperationTag key;
1336 PendingOperationEntry *entry;
1339 /* ensure any pad bytes in the hash key are zeroed */
1340 MemSet(&key, 0, sizeof(key));
1344 entry = (PendingOperationEntry *) hash_search(pendingOpsTable,
1348 /* if new or previously canceled entry, initialize it */
1349 if (!found || entry->canceled)
1351 entry->canceled = false;
1352 entry->cycle_ctr = mdsync_cycle_ctr;
1355 * NB: it's intentional that we don't change cycle_ctr if the entry
1356 * already exists. The fsync request must be treated as old, even
1357 * though the new request will be satisfied too by any subsequent
1360 * However, if the entry is present but is marked canceled, we should
1361 * act just as though it wasn't there. The only case where this could
1362 * happen would be if a file had been deleted, we received but did not
1363 * yet act on the cancel request, and the same relfilenode was then
1364 * assigned to a new file. We mustn't lose the new request, but
1365 * it should be considered new not old.
1371 * ForgetRelationFsyncRequests -- ensure any fsyncs for a rel are forgotten
1374 ForgetRelationFsyncRequests(RelFileNode rnode)
1376 if (pendingOpsTable)
1378 /* standalone backend or startup process: fsync state is local */
1379 RememberFsyncRequest(rnode, FORGET_RELATION_FSYNC);
1381 else if (IsUnderPostmaster)
1384 * Notify the bgwriter about it. If we fail to queue the revoke
1385 * message, we have to sleep and try again ... ugly, but hopefully
1386 * won't happen often.
1388 * XXX should we CHECK_FOR_INTERRUPTS in this loop? Escaping with
1389 * an error would leave the no-longer-used file still present on
1390 * disk, which would be bad, so I'm inclined to assume that the
1391 * bgwriter will always empty the queue soon.
1393 while (!ForwardFsyncRequest(rnode, FORGET_RELATION_FSYNC))
1394 pg_usleep(10000L); /* 10 msec seems a good number */
1396 * Note we don't wait for the bgwriter to actually absorb the
1397 * revoke message; see mdsync() for the implications.
1403 * ForgetDatabaseFsyncRequests -- ensure any fsyncs for a DB are forgotten
1406 ForgetDatabaseFsyncRequests(Oid dbid)
1410 rnode.dbNode = dbid;
1414 if (pendingOpsTable)
1416 /* standalone backend or startup process: fsync state is local */
1417 RememberFsyncRequest(rnode, FORGET_DATABASE_FSYNC);
1419 else if (IsUnderPostmaster)
1421 /* see notes in ForgetRelationFsyncRequests */
1422 while (!ForwardFsyncRequest(rnode, FORGET_DATABASE_FSYNC))
1423 pg_usleep(10000L); /* 10 msec seems a good number */
1429 * _fdvec_alloc() -- Make a MdfdVec object.
1434 return (MdfdVec *) MemoryContextAlloc(MdCxt, sizeof(MdfdVec));
1437 #ifndef LET_OS_MANAGE_FILESIZE
1440 * Open the specified segment of the relation,
1441 * and make a MdfdVec object for it. Returns NULL on failure.
1444 _mdfd_openseg(SMgrRelation reln, BlockNumber segno, int oflags)
1451 path = relpath(reln->smgr_rnode);
1455 /* be sure we have enough space for the '.segno' */
1456 fullpath = (char *) palloc(strlen(path) + 12);
1457 sprintf(fullpath, "%s.%u", path, segno);
1464 fd = PathNameOpenFile(fullpath, O_RDWR | PG_BINARY | oflags, 0600);
1471 /* allocate an mdfdvec entry for it */
1474 /* fill the entry */
1476 v->mdfd_segno = segno;
1477 v->mdfd_chain = NULL;
1478 Assert(_mdnblocks(reln, v) <= ((BlockNumber) RELSEG_SIZE));
1483 #endif /* LET_OS_MANAGE_FILESIZE */
1486 * _mdfd_getseg() -- Find the segment of the relation holding the
1489 * If the segment doesn't exist, we ereport, return NULL, or create the
1490 * segment, according to "behavior". Note: isTemp need only be correct
1491 * in the EXTENSION_CREATE case.
1494 _mdfd_getseg(SMgrRelation reln, BlockNumber blkno, bool isTemp,
1495 ExtensionBehavior behavior)
1497 MdfdVec *v = mdopen(reln, behavior);
1499 #ifndef LET_OS_MANAGE_FILESIZE
1500 BlockNumber targetseg;
1501 BlockNumber nextsegno;
1504 return NULL; /* only possible if EXTENSION_RETURN_NULL */
1506 targetseg = blkno / ((BlockNumber) RELSEG_SIZE);
1507 for (nextsegno = 1; nextsegno <= targetseg; nextsegno++)
1509 Assert(nextsegno == v->mdfd_segno + 1);
1511 if (v->mdfd_chain == NULL)
1514 * Normally we will create new segments only if authorized by
1515 * the caller (i.e., we are doing mdextend()). But when doing
1516 * WAL recovery, create segments anyway; this allows cases such as
1517 * replaying WAL data that has a write into a high-numbered
1518 * segment of a relation that was later deleted. We want to go
1519 * ahead and create the segments so we can finish out the replay.
1521 * We have to maintain the invariant that segments before the
1522 * last active segment are of size RELSEG_SIZE; therefore, pad
1523 * them out with zeroes if needed. (This only matters if caller
1524 * is extending the relation discontiguously, but that can happen
1527 if (behavior == EXTENSION_CREATE || InRecovery)
1529 if (_mdnblocks(reln, v) < RELSEG_SIZE)
1531 char *zerobuf = palloc0(BLCKSZ);
1533 mdextend(reln, nextsegno * ((BlockNumber) RELSEG_SIZE) - 1,
1537 v->mdfd_chain = _mdfd_openseg(reln, nextsegno, O_CREAT);
1541 /* We won't create segment if not existent */
1542 v->mdfd_chain = _mdfd_openseg(reln, nextsegno, 0);
1544 if (v->mdfd_chain == NULL)
1546 if (behavior == EXTENSION_RETURN_NULL &&
1547 FILE_POSSIBLY_DELETED(errno))
1550 (errcode_for_file_access(),
1551 errmsg("could not open segment %u of relation %u/%u/%u (target block %u): %m",
1553 reln->smgr_rnode.spcNode,
1554 reln->smgr_rnode.dbNode,
1555 reln->smgr_rnode.relNode,
1567 * Get number of blocks present in a single disk file
1570 _mdnblocks(SMgrRelation reln, MdfdVec *seg)
1574 len = FileSeek(seg->mdfd_vfd, 0L, SEEK_END);
1577 (errcode_for_file_access(),
1578 errmsg("could not seek to end of segment %u of relation %u/%u/%u: %m",
1580 reln->smgr_rnode.spcNode,
1581 reln->smgr_rnode.dbNode,
1582 reln->smgr_rnode.relNode)));
1583 /* note that this calculation will ignore any partial block at EOF */
1584 return (BlockNumber) (len / BLCKSZ);