1 /*-------------------------------------------------------------------------
4 * This code manages relations that reside on magnetic disk.
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
13 * Portions Copyright (c) 1996-2016, PostgreSQL Global Development Group
14 * Portions Copyright (c) 1994, Regents of the University of California
18 * src/backend/storage/smgr/md.c
20 *-------------------------------------------------------------------------
28 #include "miscadmin.h"
29 #include "access/xlog.h"
30 #include "catalog/catalog.h"
31 #include "portability/instr_time.h"
32 #include "postmaster/bgwriter.h"
33 #include "storage/fd.h"
34 #include "storage/bufmgr.h"
35 #include "storage/relfilenode.h"
36 #include "storage/smgr.h"
37 #include "utils/hsearch.h"
38 #include "utils/memutils.h"
42 /* intervals for calling AbsorbFsyncRequests in mdsync and mdpostckpt */
43 #define FSYNCS_PER_ABSORB 10
44 #define UNLINKS_PER_ABSORB 10
47 * Special values for the segno arg to RememberFsyncRequest.
49 * Note that CompactCheckpointerRequestQueue assumes that it's OK to remove an
50 * fsync request from the queue if an identical, subsequent request is found.
51 * See comments there before making changes here.
53 #define FORGET_RELATION_FSYNC (InvalidBlockNumber)
54 #define FORGET_DATABASE_FSYNC (InvalidBlockNumber-1)
55 #define UNLINK_RELATION_REQUEST (InvalidBlockNumber-2)
58 * On Windows, we have to interpret EACCES as possibly meaning the same as
59 * ENOENT, because if a file is unlinked-but-not-yet-gone on that platform,
60 * that's what you get. Ugh. This code is designed so that we don't
61 * actually believe these cases are okay without further evidence (namely,
62 * a pending fsync request getting canceled ... see mdsync).
65 #define FILE_POSSIBLY_DELETED(err) ((err) == ENOENT)
67 #define FILE_POSSIBLY_DELETED(err) ((err) == ENOENT || (err) == EACCES)
71 * The magnetic disk storage manager keeps track of open file
72 * descriptors in its own descriptor pool. This is done to make it
73 * easier to support relations that are larger than the operating
74 * system's file size limit (often 2GBytes). In order to do that,
75 * we break relations up into "segment" files that are each shorter than
76 * the OS file size limit. The segment size is set by the RELSEG_SIZE
77 * configuration constant in pg_config.h.
79 * On disk, a relation must consist of consecutively numbered segment
80 * files in the pattern
81 * -- Zero or more full segments of exactly RELSEG_SIZE blocks each
82 * -- Exactly one partial segment of size 0 <= size < RELSEG_SIZE blocks
83 * -- Optionally, any number of inactive segments of size 0 blocks.
84 * The full and partial segments are collectively the "active" segments.
85 * Inactive segments are those that once contained data but are currently
86 * not needed because of an mdtruncate() operation. The reason for leaving
87 * them present at size zero, rather than unlinking them, is that other
88 * backends and/or the checkpointer might be holding open file references to
89 * such segments. If the relation expands again after mdtruncate(), such
90 * that a deactivated segment becomes active again, it is important that
91 * such file references still be valid --- else data might get written
92 * out to an unlinked old copy of a segment file that will eventually
95 * The file descriptor pointer (md_fd field) stored in the SMgrRelation
96 * cache is, therefore, just the head of a list of MdfdVec objects, one
97 * per segment. But note the md_fd pointer can be NULL, indicating
100 * Also note that mdfd_chain == NULL does not necessarily mean the relation
101 * doesn't have another segment after this one; we may just not have
102 * opened the next segment yet. (We could not have "all segments are
103 * in the chain" as an invariant anyway, since another backend could
104 * extend the relation when we weren't looking.) We do not make chain
105 * entries for inactive segments, however; as soon as we find a partial
106 * segment, we assume that any subsequent segments are inactive.
108 * All MdfdVec objects are palloc'd in the MdCxt memory context.
111 typedef struct _MdfdVec
113 File mdfd_vfd; /* fd number in fd.c's pool */
114 BlockNumber mdfd_segno; /* segment number, from 0 */
115 struct _MdfdVec *mdfd_chain; /* next segment, or NULL */
118 static MemoryContext MdCxt; /* context for all MdfdVec objects */
122 * In some contexts (currently, standalone backends and the checkpointer)
123 * we keep track of pending fsync operations: we need to remember all relation
124 * segments that have been written since the last checkpoint, so that we can
125 * fsync them down to disk before completing the next checkpoint. This hash
126 * table remembers the pending operations. We use a hash table mostly as
127 * a convenient way of merging duplicate requests.
129 * We use a similar mechanism to remember no-longer-needed files that can
130 * be deleted after the next checkpoint, but we use a linked list instead of
131 * a hash table, because we don't expect there to be any duplicate requests.
133 * These mechanisms are only used for non-temp relations; we never fsync
134 * temp rels, nor do we need to postpone their deletion (see comments in
137 * (Regular backends do not track pending operations locally, but forward
138 * them to the checkpointer.)
140 typedef uint16 CycleCtr; /* can be any convenient integer size */
144 RelFileNode rnode; /* hash table key (must be first!) */
145 CycleCtr cycle_ctr; /* mdsync_cycle_ctr of oldest request */
146 /* requests[f] has bit n set if we need to fsync segment n of fork f */
147 Bitmapset *requests[MAX_FORKNUM + 1];
148 /* canceled[f] is true if we canceled fsyncs for fork "recently" */
149 bool canceled[MAX_FORKNUM + 1];
150 } PendingOperationEntry;
154 RelFileNode rnode; /* the dead relation to delete */
155 CycleCtr cycle_ctr; /* mdckpt_cycle_ctr when request was made */
156 } PendingUnlinkEntry;
158 static HTAB *pendingOpsTable = NULL;
159 static List *pendingUnlinks = NIL;
160 static MemoryContext pendingOpsCxt; /* context for the above */
162 static CycleCtr mdsync_cycle_ctr = 0;
163 static CycleCtr mdckpt_cycle_ctr = 0;
166 /*** behavior for mdopen & _mdfd_getseg ***/
167 /* ereport if segment not present */
168 #define EXTENSION_FAIL (1 << 0)
169 /* return NULL if segment not present */
170 #define EXTENSION_RETURN_NULL (1 << 1)
171 /* create new segments as needed */
172 #define EXTENSION_CREATE (1 << 2)
173 /* create new segments if needed during recovery */
174 #define EXTENSION_CREATE_RECOVERY (1 << 3)
176 * Allow opening segments which are preceded by segments smaller than
177 * RELSEG_SIZE, e.g. inactive segments (see above). Note that this is breaks
178 * mdnblocks() and related functionality henceforth - which currently is ok,
179 * because this is only required in the checkpointer which never uses
182 #define EXTENSION_DONT_CHECK_SIZE (1 << 4)
186 static void mdunlinkfork(RelFileNodeBackend rnode, ForkNumber forkNum,
188 static MdfdVec *mdopen(SMgrRelation reln, ForkNumber forknum, int behavior);
189 static void register_dirty_segment(SMgrRelation reln, ForkNumber forknum,
191 static void register_unlink(RelFileNodeBackend rnode);
192 static MdfdVec *_fdvec_alloc(void);
193 static char *_mdfd_segpath(SMgrRelation reln, ForkNumber forknum,
195 static MdfdVec *_mdfd_openseg(SMgrRelation reln, ForkNumber forkno,
196 BlockNumber segno, int oflags);
197 static MdfdVec *_mdfd_getseg(SMgrRelation reln, ForkNumber forkno,
198 BlockNumber blkno, bool skipFsync, int behavior);
199 static BlockNumber _mdnblocks(SMgrRelation reln, ForkNumber forknum,
204 * mdinit() -- Initialize private state for magnetic disk storage manager.
209 MdCxt = AllocSetContextCreate(TopMemoryContext,
211 ALLOCSET_DEFAULT_MINSIZE,
212 ALLOCSET_DEFAULT_INITSIZE,
213 ALLOCSET_DEFAULT_MAXSIZE);
216 * Create pending-operations hashtable if we need it. Currently, we need
217 * it if we are standalone (not under a postmaster) or if we are a startup
218 * or checkpointer auxiliary process.
220 if (!IsUnderPostmaster || AmStartupProcess() || AmCheckpointerProcess())
225 * XXX: The checkpointer needs to add entries to the pending ops table
226 * when absorbing fsync requests. That is done within a critical
227 * section, which isn't usually allowed, but we make an exception. It
228 * means that there's a theoretical possibility that you run out of
229 * memory while absorbing fsync requests, which leads to a PANIC.
230 * Fortunately the hash table is small so that's unlikely to happen in
233 pendingOpsCxt = AllocSetContextCreate(MdCxt,
234 "Pending Ops Context",
235 ALLOCSET_DEFAULT_MINSIZE,
236 ALLOCSET_DEFAULT_INITSIZE,
237 ALLOCSET_DEFAULT_MAXSIZE);
238 MemoryContextAllowInCriticalSection(pendingOpsCxt, true);
240 MemSet(&hash_ctl, 0, sizeof(hash_ctl));
241 hash_ctl.keysize = sizeof(RelFileNode);
242 hash_ctl.entrysize = sizeof(PendingOperationEntry);
243 hash_ctl.hcxt = pendingOpsCxt;
244 pendingOpsTable = hash_create("Pending Ops Table",
247 HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
248 pendingUnlinks = NIL;
253 * In archive recovery, we rely on checkpointer to do fsyncs, but we will have
254 * already created the pendingOpsTable during initialization of the startup
255 * process. Calling this function drops the local pendingOpsTable so that
256 * subsequent requests will be forwarded to checkpointer.
259 SetForwardFsyncRequests(void)
261 /* Perform any pending fsyncs we may have queued up, then drop table */
265 hash_destroy(pendingOpsTable);
267 pendingOpsTable = NULL;
270 * We should not have any pending unlink requests, since mdunlink doesn't
271 * queue unlink requests when isRedo.
273 Assert(pendingUnlinks == NIL);
277 * mdexists() -- Does the physical file exist?
279 * Note: this will return true for lingering files, with pending deletions
282 mdexists(SMgrRelation reln, ForkNumber forkNum)
285 * Close it first, to ensure that we notice if the fork has been unlinked
286 * since we opened it.
288 mdclose(reln, forkNum);
290 return (mdopen(reln, forkNum, EXTENSION_RETURN_NULL) != NULL);
294 * mdcreate() -- Create a new relation on magnetic disk.
296 * If isRedo is true, it's okay for the relation to exist already.
299 mdcreate(SMgrRelation reln, ForkNumber forkNum, bool isRedo)
304 if (isRedo && reln->md_fd[forkNum] != NULL)
305 return; /* created and opened already... */
307 Assert(reln->md_fd[forkNum] == NULL);
309 path = relpath(reln->smgr_rnode, forkNum);
311 fd = PathNameOpenFile(path, O_RDWR | O_CREAT | O_EXCL | PG_BINARY, 0600);
315 int save_errno = errno;
318 * During bootstrap, there are cases where a system relation will be
319 * accessed (by internal backend processes) before the bootstrap
320 * script nominally creates it. Therefore, allow the file to exist
321 * already, even if isRedo is not set. (See also mdopen)
323 if (isRedo || IsBootstrapProcessingMode())
324 fd = PathNameOpenFile(path, O_RDWR | PG_BINARY, 0600);
327 /* be sure to report the error reported by create, not open */
330 (errcode_for_file_access(),
331 errmsg("could not create file \"%s\": %m", path)));
337 reln->md_fd[forkNum] = _fdvec_alloc();
339 reln->md_fd[forkNum]->mdfd_vfd = fd;
340 reln->md_fd[forkNum]->mdfd_segno = 0;
341 reln->md_fd[forkNum]->mdfd_chain = NULL;
345 * mdunlink() -- Unlink a relation.
347 * Note that we're passed a RelFileNodeBackend --- by the time this is called,
348 * there won't be an SMgrRelation hashtable entry anymore.
350 * forkNum can be a fork number to delete a specific fork, or InvalidForkNumber
351 * to delete all forks.
353 * For regular relations, we don't unlink the first segment file of the rel,
354 * but just truncate it to zero length, and record a request to unlink it after
355 * the next checkpoint. Additional segments can be unlinked immediately,
356 * however. Leaving the empty file in place prevents that relfilenode
357 * number from being reused. The scenario this protects us from is:
358 * 1. We delete a relation (and commit, and actually remove its file).
359 * 2. We create a new relation, which by chance gets the same relfilenode as
360 * the just-deleted one (OIDs must've wrapped around for that to happen).
361 * 3. We crash before another checkpoint occurs.
362 * During replay, we would delete the file and then recreate it, which is fine
363 * if the contents of the file were repopulated by subsequent WAL entries.
364 * But if we didn't WAL-log insertions, but instead relied on fsyncing the
365 * file after populating it (as for instance CLUSTER and CREATE INDEX do),
366 * the contents of the file would be lost forever. By leaving the empty file
367 * until after the next checkpoint, we prevent reassignment of the relfilenode
368 * number until it's safe, because relfilenode assignment skips over any
371 * We do not need to go through this dance for temp relations, though, because
372 * we never make WAL entries for temp rels, and so a temp rel poses no threat
373 * to the health of a regular rel that has taken over its relfilenode number.
374 * The fact that temp rels and regular rels have different file naming
375 * patterns provides additional safety.
377 * All the above applies only to the relation's main fork; other forks can
378 * just be removed immediately, since they are not needed to prevent the
379 * relfilenode number from being recycled. Also, we do not carefully
380 * track whether other forks have been created or not, but just attempt to
381 * unlink them unconditionally; so we should never complain about ENOENT.
383 * If isRedo is true, it's unsurprising for the relation to be already gone.
384 * Also, we should remove the file immediately instead of queuing a request
385 * for later, since during redo there's no possibility of creating a
386 * conflicting relation.
388 * Note: any failure should be reported as WARNING not ERROR, because
389 * we are usually not in a transaction anymore when this is called.
392 mdunlink(RelFileNodeBackend rnode, ForkNumber forkNum, bool isRedo)
395 * We have to clean out any pending fsync requests for the doomed
396 * relation, else the next mdsync() will fail. There can't be any such
397 * requests for a temp relation, though. We can send just one request
398 * even when deleting multiple forks, since the fsync queuing code accepts
399 * the "InvalidForkNumber = all forks" convention.
401 if (!RelFileNodeBackendIsTemp(rnode))
402 ForgetRelationFsyncRequests(rnode.node, forkNum);
404 /* Now do the per-fork work */
405 if (forkNum == InvalidForkNumber)
407 for (forkNum = 0; forkNum <= MAX_FORKNUM; forkNum++)
408 mdunlinkfork(rnode, forkNum, isRedo);
411 mdunlinkfork(rnode, forkNum, isRedo);
415 mdunlinkfork(RelFileNodeBackend rnode, ForkNumber forkNum, bool isRedo)
420 path = relpath(rnode, forkNum);
423 * Delete or truncate the first segment.
425 if (isRedo || forkNum != MAIN_FORKNUM || RelFileNodeBackendIsTemp(rnode))
428 if (ret < 0 && errno != ENOENT)
430 (errcode_for_file_access(),
431 errmsg("could not remove file \"%s\": %m", path)));
435 /* truncate(2) would be easier here, but Windows hasn't got it */
438 fd = OpenTransientFile(path, O_RDWR | PG_BINARY, 0);
443 ret = ftruncate(fd, 0);
445 CloseTransientFile(fd);
450 if (ret < 0 && errno != ENOENT)
452 (errcode_for_file_access(),
453 errmsg("could not truncate file \"%s\": %m", path)));
455 /* Register request to unlink first segment later */
456 register_unlink(rnode);
460 * Delete any additional segments.
464 char *segpath = (char *) palloc(strlen(path) + 12);
468 * Note that because we loop until getting ENOENT, we will correctly
469 * remove all inactive segments as well as active ones.
471 for (segno = 1;; segno++)
473 sprintf(segpath, "%s.%u", path, segno);
474 if (unlink(segpath) < 0)
476 /* ENOENT is expected after the last segment... */
479 (errcode_for_file_access(),
480 errmsg("could not remove file \"%s\": %m", segpath)));
491 * mdextend() -- Add a block to the specified relation.
493 * The semantics are nearly the same as mdwrite(): write at the
494 * specified position. However, this is to be used for the case of
495 * extending a relation (i.e., blocknum is at or beyond the current
496 * EOF). Note that we assume writing a block beyond current EOF
497 * causes intervening file space to become filled with zeroes.
500 mdextend(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
501 char *buffer, bool skipFsync)
507 /* This assert is too expensive to have on normally ... */
508 #ifdef CHECK_WRITE_VS_EXTEND
509 Assert(blocknum >= mdnblocks(reln, forknum));
513 * If a relation manages to grow to 2^32-1 blocks, refuse to extend it any
514 * more --- we mustn't create a block whose number actually is
515 * InvalidBlockNumber.
517 if (blocknum == InvalidBlockNumber)
519 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
520 errmsg("cannot extend file \"%s\" beyond %u blocks",
521 relpath(reln->smgr_rnode, forknum),
522 InvalidBlockNumber)));
524 v = _mdfd_getseg(reln, forknum, blocknum, skipFsync, EXTENSION_CREATE);
526 seekpos = (off_t) BLCKSZ *(blocknum % ((BlockNumber) RELSEG_SIZE));
528 Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
531 * Note: because caller usually obtained blocknum by calling mdnblocks,
532 * which did a seek(SEEK_END), this seek is often redundant and will be
533 * optimized away by fd.c. It's not redundant, however, if there is a
534 * partial page at the end of the file. In that case we want to try to
535 * overwrite the partial page with a full page. It's also not redundant
536 * if bufmgr.c had to dump another buffer of the same file to make room
537 * for the new page's buffer.
539 if (FileSeek(v->mdfd_vfd, seekpos, SEEK_SET) != seekpos)
541 (errcode_for_file_access(),
542 errmsg("could not seek to block %u in file \"%s\": %m",
543 blocknum, FilePathName(v->mdfd_vfd))));
545 if ((nbytes = FileWrite(v->mdfd_vfd, buffer, BLCKSZ)) != BLCKSZ)
549 (errcode_for_file_access(),
550 errmsg("could not extend file \"%s\": %m",
551 FilePathName(v->mdfd_vfd)),
552 errhint("Check free disk space.")));
553 /* short write: complain appropriately */
555 (errcode(ERRCODE_DISK_FULL),
556 errmsg("could not extend file \"%s\": wrote only %d of %d bytes at block %u",
557 FilePathName(v->mdfd_vfd),
558 nbytes, BLCKSZ, blocknum),
559 errhint("Check free disk space.")));
562 if (!skipFsync && !SmgrIsTemp(reln))
563 register_dirty_segment(reln, forknum, v);
565 Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
569 * mdopen() -- Open the specified relation.
571 * Note we only open the first segment, when there are multiple segments.
573 * If first segment is not present, either ereport or return NULL according
574 * to "behavior". We treat EXTENSION_CREATE the same as EXTENSION_FAIL;
575 * EXTENSION_CREATE means it's OK to extend an existing relation, not to
576 * invent one out of whole cloth.
579 mdopen(SMgrRelation reln, ForkNumber forknum, int behavior)
585 /* No work if already open */
586 if (reln->md_fd[forknum])
587 return reln->md_fd[forknum];
589 path = relpath(reln->smgr_rnode, forknum);
591 fd = PathNameOpenFile(path, O_RDWR | PG_BINARY, 0600);
596 * During bootstrap, there are cases where a system relation will be
597 * accessed (by internal backend processes) before the bootstrap
598 * script nominally creates it. Therefore, accept mdopen() as a
599 * substitute for mdcreate() in bootstrap mode only. (See mdcreate)
601 if (IsBootstrapProcessingMode())
602 fd = PathNameOpenFile(path, O_RDWR | O_CREAT | O_EXCL | PG_BINARY, 0600);
605 if ((behavior & EXTENSION_RETURN_NULL) &&
606 FILE_POSSIBLY_DELETED(errno))
612 (errcode_for_file_access(),
613 errmsg("could not open file \"%s\": %m", path)));
619 reln->md_fd[forknum] = mdfd = _fdvec_alloc();
622 mdfd->mdfd_segno = 0;
623 mdfd->mdfd_chain = NULL;
624 Assert(_mdnblocks(reln, forknum, mdfd) <= ((BlockNumber) RELSEG_SIZE));
630 * mdclose() -- Close the specified relation, if it isn't closed already.
633 mdclose(SMgrRelation reln, ForkNumber forknum)
635 MdfdVec *v = reln->md_fd[forknum];
637 /* No work if already closed */
641 reln->md_fd[forknum] = NULL; /* prevent dangling pointer after error */
647 /* if not closed already */
648 if (v->mdfd_vfd >= 0)
649 FileClose(v->mdfd_vfd);
650 /* Now free vector */
657 * mdprefetch() -- Initiate asynchronous read of the specified block of a relation
660 mdprefetch(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum)
666 v = _mdfd_getseg(reln, forknum, blocknum, false, EXTENSION_FAIL);
668 seekpos = (off_t) BLCKSZ *(blocknum % ((BlockNumber) RELSEG_SIZE));
670 Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
672 (void) FilePrefetch(v->mdfd_vfd, seekpos, BLCKSZ);
673 #endif /* USE_PREFETCH */
677 * mdwriteback() -- Tell the kernel to write pages back to storage.
679 * This accepts a range of blocks because flushing several pages at once is
680 * considerably more efficient than doing so individually.
683 mdwriteback(SMgrRelation reln, ForkNumber forknum,
684 BlockNumber blocknum, BlockNumber nblocks)
687 * Issue flush requests in as few requests as possible; have to split at
688 * segment boundaries though, since those are actually separate files.
692 BlockNumber nflush = nblocks;
698 v = _mdfd_getseg(reln, forknum, blocknum, true /* not used */ ,
699 EXTENSION_RETURN_NULL);
702 * We might be flushing buffers of already removed relations, that's
703 * ok, just ignore that case.
708 /* compute offset inside the current segment */
709 segnum_start = blocknum / RELSEG_SIZE;
711 /* compute number of desired writes within the current segment */
712 segnum_end = (blocknum + nblocks - 1) / RELSEG_SIZE;
713 if (segnum_start != segnum_end)
714 nflush = RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE));
717 Assert(nflush <= nblocks);
719 seekpos = (off_t) BLCKSZ *(blocknum % ((BlockNumber) RELSEG_SIZE));
721 FileWriteback(v->mdfd_vfd, seekpos, (off_t) BLCKSZ * nflush);
729 * mdread() -- Read the specified block from a relation.
732 mdread(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
739 TRACE_POSTGRESQL_SMGR_MD_READ_START(forknum, blocknum,
740 reln->smgr_rnode.node.spcNode,
741 reln->smgr_rnode.node.dbNode,
742 reln->smgr_rnode.node.relNode,
743 reln->smgr_rnode.backend);
745 v = _mdfd_getseg(reln, forknum, blocknum, false,
746 EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
748 seekpos = (off_t) BLCKSZ *(blocknum % ((BlockNumber) RELSEG_SIZE));
750 Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
752 if (FileSeek(v->mdfd_vfd, seekpos, SEEK_SET) != seekpos)
754 (errcode_for_file_access(),
755 errmsg("could not seek to block %u in file \"%s\": %m",
756 blocknum, FilePathName(v->mdfd_vfd))));
758 nbytes = FileRead(v->mdfd_vfd, buffer, BLCKSZ);
760 TRACE_POSTGRESQL_SMGR_MD_READ_DONE(forknum, blocknum,
761 reln->smgr_rnode.node.spcNode,
762 reln->smgr_rnode.node.dbNode,
763 reln->smgr_rnode.node.relNode,
764 reln->smgr_rnode.backend,
768 if (nbytes != BLCKSZ)
772 (errcode_for_file_access(),
773 errmsg("could not read block %u in file \"%s\": %m",
774 blocknum, FilePathName(v->mdfd_vfd))));
777 * Short read: we are at or past EOF, or we read a partial block at
778 * EOF. Normally this is an error; upper levels should never try to
779 * read a nonexistent block. However, if zero_damaged_pages is ON or
780 * we are InRecovery, we should instead return zeroes without
781 * complaining. This allows, for example, the case of trying to
782 * update a block that was later truncated away.
784 if (zero_damaged_pages || InRecovery)
785 MemSet(buffer, 0, BLCKSZ);
788 (errcode(ERRCODE_DATA_CORRUPTED),
789 errmsg("could not read block %u in file \"%s\": read only %d of %d bytes",
790 blocknum, FilePathName(v->mdfd_vfd),
796 * mdwrite() -- Write the supplied block at the appropriate location.
798 * This is to be used only for updating already-existing blocks of a
799 * relation (ie, those before the current EOF). To extend a relation,
803 mdwrite(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
804 char *buffer, bool skipFsync)
810 /* This assert is too expensive to have on normally ... */
811 #ifdef CHECK_WRITE_VS_EXTEND
812 Assert(blocknum < mdnblocks(reln, forknum));
815 TRACE_POSTGRESQL_SMGR_MD_WRITE_START(forknum, blocknum,
816 reln->smgr_rnode.node.spcNode,
817 reln->smgr_rnode.node.dbNode,
818 reln->smgr_rnode.node.relNode,
819 reln->smgr_rnode.backend);
821 v = _mdfd_getseg(reln, forknum, blocknum, skipFsync,
822 EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
824 seekpos = (off_t) BLCKSZ *(blocknum % ((BlockNumber) RELSEG_SIZE));
826 Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
828 if (FileSeek(v->mdfd_vfd, seekpos, SEEK_SET) != seekpos)
830 (errcode_for_file_access(),
831 errmsg("could not seek to block %u in file \"%s\": %m",
832 blocknum, FilePathName(v->mdfd_vfd))));
834 nbytes = FileWrite(v->mdfd_vfd, buffer, BLCKSZ);
836 TRACE_POSTGRESQL_SMGR_MD_WRITE_DONE(forknum, blocknum,
837 reln->smgr_rnode.node.spcNode,
838 reln->smgr_rnode.node.dbNode,
839 reln->smgr_rnode.node.relNode,
840 reln->smgr_rnode.backend,
844 if (nbytes != BLCKSZ)
848 (errcode_for_file_access(),
849 errmsg("could not write block %u in file \"%s\": %m",
850 blocknum, FilePathName(v->mdfd_vfd))));
851 /* short write: complain appropriately */
853 (errcode(ERRCODE_DISK_FULL),
854 errmsg("could not write block %u in file \"%s\": wrote only %d of %d bytes",
856 FilePathName(v->mdfd_vfd),
858 errhint("Check free disk space.")));
861 if (!skipFsync && !SmgrIsTemp(reln))
862 register_dirty_segment(reln, forknum, v);
866 * mdnblocks() -- Get the number of blocks stored in a relation.
868 * Important side effect: all active segments of the relation are opened
869 * and added to the mdfd_chain list. If this routine has not been
870 * called, then only segments up to the last one actually touched
871 * are present in the chain.
874 mdnblocks(SMgrRelation reln, ForkNumber forknum)
876 MdfdVec *v = mdopen(reln, forknum, EXTENSION_FAIL);
878 BlockNumber segno = 0;
881 * Skip through any segments that aren't the last one, to avoid redundant
882 * seeks on them. We have previously verified that these segments are
883 * exactly RELSEG_SIZE long, and it's useless to recheck that each time.
885 * NOTE: this assumption could only be wrong if another backend has
886 * truncated the relation. We rely on higher code levels to handle that
887 * scenario by closing and re-opening the md fd, which is handled via
888 * relcache flush. (Since the checkpointer doesn't participate in
889 * relcache flush, it could have segment chain entries for inactive
890 * segments; that's OK because the checkpointer never needs to compute
893 while (v->mdfd_chain != NULL)
901 nblocks = _mdnblocks(reln, forknum, v);
902 if (nblocks > ((BlockNumber) RELSEG_SIZE))
903 elog(FATAL, "segment too big");
904 if (nblocks < ((BlockNumber) RELSEG_SIZE))
905 return (segno * ((BlockNumber) RELSEG_SIZE)) + nblocks;
908 * If segment is exactly RELSEG_SIZE, advance to next one.
912 if (v->mdfd_chain == NULL)
915 * We used to pass O_CREAT here, but that's has the disadvantage
916 * that it might create a segment which has vanished through some
917 * operating system misadventure. In such a case, creating the
918 * segment here undermines _mdfd_getseg's attempts to notice and
919 * report an error upon access to a missing segment.
921 v->mdfd_chain = _mdfd_openseg(reln, forknum, segno, 0);
922 if (v->mdfd_chain == NULL)
923 return segno * ((BlockNumber) RELSEG_SIZE);
931 * mdtruncate() -- Truncate relation to specified number of blocks.
934 mdtruncate(SMgrRelation reln, ForkNumber forknum, BlockNumber nblocks)
938 BlockNumber priorblocks;
941 * NOTE: mdnblocks makes sure we have opened all active segments, so that
942 * truncation loop will get them all!
944 curnblk = mdnblocks(reln, forknum);
945 if (nblocks > curnblk)
947 /* Bogus request ... but no complaint if InRecovery */
951 (errmsg("could not truncate file \"%s\" to %u blocks: it's only %u blocks now",
952 relpath(reln->smgr_rnode, forknum),
955 if (nblocks == curnblk)
956 return; /* no work */
958 v = mdopen(reln, forknum, EXTENSION_FAIL);
965 if (priorblocks > nblocks)
968 * This segment is no longer active (and has already been unlinked
969 * from the mdfd_chain). We truncate the file, but do not delete
970 * it, for reasons explained in the header comments.
972 if (FileTruncate(v->mdfd_vfd, 0) < 0)
974 (errcode_for_file_access(),
975 errmsg("could not truncate file \"%s\": %m",
976 FilePathName(v->mdfd_vfd))));
978 if (!SmgrIsTemp(reln))
979 register_dirty_segment(reln, forknum, v);
981 Assert(ov != reln->md_fd[forknum]); /* we never drop the 1st
985 else if (priorblocks + ((BlockNumber) RELSEG_SIZE) > nblocks)
988 * This is the last segment we want to keep. Truncate the file to
989 * the right length, and clear chain link that points to any
990 * remaining segments (which we shall zap). NOTE: if nblocks is
991 * exactly a multiple K of RELSEG_SIZE, we will truncate the K+1st
992 * segment to 0 length but keep it. This adheres to the invariant
993 * given in the header comments.
995 BlockNumber lastsegblocks = nblocks - priorblocks;
997 if (FileTruncate(v->mdfd_vfd, (off_t) lastsegblocks * BLCKSZ) < 0)
999 (errcode_for_file_access(),
1000 errmsg("could not truncate file \"%s\" to %u blocks: %m",
1001 FilePathName(v->mdfd_vfd),
1003 if (!SmgrIsTemp(reln))
1004 register_dirty_segment(reln, forknum, v);
1006 ov->mdfd_chain = NULL;
1011 * We still need this segment and 0 or more blocks beyond it, so
1012 * nothing to do here.
1016 priorblocks += RELSEG_SIZE;
1021 * mdimmedsync() -- Immediately sync a relation to stable storage.
1023 * Note that only writes already issued are synced; this routine knows
1024 * nothing of dirty buffers that may exist inside the buffer manager.
1027 mdimmedsync(SMgrRelation reln, ForkNumber forknum)
1032 * NOTE: mdnblocks makes sure we have opened all active segments, so that
1033 * fsync loop will get them all!
1035 mdnblocks(reln, forknum);
1037 v = mdopen(reln, forknum, EXTENSION_FAIL);
1041 if (FileSync(v->mdfd_vfd) < 0)
1043 (errcode_for_file_access(),
1044 errmsg("could not fsync file \"%s\": %m",
1045 FilePathName(v->mdfd_vfd))));
1051 * mdsync() -- Sync previous writes to stable storage.
1056 static bool mdsync_in_progress = false;
1058 HASH_SEQ_STATUS hstat;
1059 PendingOperationEntry *entry;
1062 /* Statistics on sync times */
1064 instr_time sync_start,
1069 uint64 total_elapsed = 0;
1072 * This is only called during checkpoints, and checkpoints should only
1073 * occur in processes that have created a pendingOpsTable.
1075 if (!pendingOpsTable)
1076 elog(ERROR, "cannot sync without a pendingOpsTable");
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().
1087 AbsorbFsyncRequests();
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
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.
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.
1113 if (mdsync_in_progress)
1115 /* prior try failed, so update any stale cycle_ctr values */
1116 hash_seq_init(&hstat, pendingOpsTable);
1117 while ((entry = (PendingOperationEntry *) hash_seq_search(&hstat)) != NULL)
1119 entry->cycle_ctr = mdsync_cycle_ctr;
1123 /* Advance counter so that new hashtable entries are distinguishable */
1126 /* Set flag to detect failure if we don't reach the end of the loop */
1127 mdsync_in_progress = true;
1129 /* Now scan the hashtable for fsync requests to process */
1130 absorb_counter = FSYNCS_PER_ABSORB;
1131 hash_seq_init(&hstat, pendingOpsTable);
1132 while ((entry = (PendingOperationEntry *) hash_seq_search(&hstat)) != NULL)
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.
1141 if (entry->cycle_ctr == mdsync_cycle_ctr)
1144 /* Else assert we haven't missed it */
1145 Assert((CycleCtr) (entry->cycle_ctr + 1) == mdsync_cycle_ctr);
1148 * Scan over the forks and segments represented by the entry.
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.
1158 for (forknum = 0; forknum <= MAX_FORKNUM; forknum++)
1160 Bitmapset *requests = entry->requests[forknum];
1163 entry->requests[forknum] = NULL;
1164 entry->canceled[forknum] = false;
1166 while ((segno = bms_first_member(requests)) >= 0)
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.)
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.
1185 if (--absorb_counter <= 0)
1187 AbsorbFsyncRequests();
1188 absorb_counter = FSYNCS_PER_ABSORB;
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
1203 for (failures = 0;; failures++) /* loop exits at "break" */
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.)
1224 reln = smgropen(entry->rnode, InvalidBackendId);
1226 /* Attempt to open and fsync the target segment */
1227 seg = _mdfd_getseg(reln, forknum,
1228 (BlockNumber) segno * (BlockNumber) RELSEG_SIZE,
1230 EXTENSION_RETURN_NULL
1231 | EXTENSION_DONT_CHECK_SIZE);
1233 INSTR_TIME_SET_CURRENT(sync_start);
1236 FileSync(seg->mdfd_vfd) >= 0)
1238 /* Success; update statistics about sync timing */
1239 INSTR_TIME_SET_CURRENT(sync_end);
1240 sync_diff = sync_end;
1241 INSTR_TIME_SUBTRACT(sync_diff, sync_start);
1242 elapsed = INSTR_TIME_GET_MICROSEC(sync_diff);
1243 if (elapsed > longest)
1245 total_elapsed += elapsed;
1247 if (log_checkpoints)
1248 elog(DEBUG1, "checkpoint sync: number=%d file=%s time=%.3f msec",
1250 FilePathName(seg->mdfd_vfd),
1251 (double) elapsed / 1000);
1253 break; /* out of retry loop */
1256 /* Compute file name for use in message */
1258 path = _mdfd_segpath(reln, forknum, (BlockNumber) segno);
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.
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
1273 if (!FILE_POSSIBLY_DELETED(errno) ||
1276 (errcode_for_file_access(),
1277 errmsg("could not fsync file \"%s\": %m",
1281 (errcode_for_file_access(),
1282 errmsg("could not fsync file \"%s\" but retrying: %m",
1287 * Absorb incoming requests and check to see if a cancel
1288 * arrived for this relation fork.
1290 AbsorbFsyncRequests();
1291 absorb_counter = FSYNCS_PER_ABSORB; /* might as well... */
1293 if (entry->canceled[forknum])
1295 } /* end retry loop */
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.
1306 for (forknum = 0; forknum <= MAX_FORKNUM; forknum++)
1308 if (entry->requests[forknum] != NULL)
1311 if (forknum <= MAX_FORKNUM)
1312 entry->cycle_ctr = mdsync_cycle_ctr;
1315 /* Okay to remove it */
1316 if (hash_search(pendingOpsTable, &entry->rnode,
1317 HASH_REMOVE, NULL) == NULL)
1318 elog(ERROR, "pendingOpsTable corrupted");
1320 } /* end loop over hashtable entries */
1322 /* Return sync performance metrics for report at checkpoint end */
1323 CheckpointStats.ckpt_sync_rels = processed;
1324 CheckpointStats.ckpt_longest_sync = longest;
1325 CheckpointStats.ckpt_agg_sync_time = total_elapsed;
1327 /* Flag successful completion of mdsync */
1328 mdsync_in_progress = false;
1332 * mdpreckpt() -- Do pre-checkpoint work
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.
1339 * This must be called *before* the checkpoint REDO point is determined.
1340 * That ensures that we won't delete files too soon.
1342 * Note that we can't do anything here that depends on the assumption
1343 * that the checkpoint will be completed.
1349 * Any unlink requests arriving after this point will be assigned the next
1350 * cycle counter, and won't be unlinked until next checkpoint.
1356 * mdpostckpt() -- Do post-checkpoint work
1358 * Remove any lingering files that can now be safely removed.
1365 absorb_counter = UNLINKS_PER_ABSORB;
1366 while (pendingUnlinks != NIL)
1368 PendingUnlinkEntry *entry = (PendingUnlinkEntry *) linitial(pendingUnlinks);
1372 * New entries are appended to the end, so if the entry is new we've
1373 * reached the end of old entries.
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.
1380 if (entry->cycle_ctr == mdckpt_cycle_ctr)
1383 /* Unlink the file */
1384 path = relpathperm(entry->rnode, MAIN_FORKNUM);
1385 if (unlink(path) < 0)
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.
1394 if (errno != ENOENT)
1396 (errcode_for_file_access(),
1397 errmsg("could not remove file \"%s\": %m", path)));
1401 /* And remove the list entry */
1402 pendingUnlinks = list_delete_first(pendingUnlinks);
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).
1411 if (--absorb_counter <= 0)
1413 AbsorbFsyncRequests();
1414 absorb_counter = UNLINKS_PER_ABSORB;
1420 * register_dirty_segment() -- Mark a relation segment as needing fsync
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).
1429 register_dirty_segment(SMgrRelation reln, ForkNumber forknum, MdfdVec *seg)
1431 /* Temp relations should never be fsync'd */
1432 Assert(!SmgrIsTemp(reln));
1434 if (pendingOpsTable)
1436 /* push it into local pending-ops table */
1437 RememberFsyncRequest(reln->smgr_rnode.node, forknum, seg->mdfd_segno);
1441 if (ForwardFsyncRequest(reln->smgr_rnode.node, forknum, seg->mdfd_segno))
1442 return; /* passed it off successfully */
1445 (errmsg("could not forward fsync request because request queue is full")));
1447 if (FileSync(seg->mdfd_vfd) < 0)
1449 (errcode_for_file_access(),
1450 errmsg("could not fsync file \"%s\": %m",
1451 FilePathName(seg->mdfd_vfd))));
1456 * register_unlink() -- Schedule a file to be deleted after next checkpoint
1458 * We don't bother passing in the fork number, because this is only used
1461 * As with register_dirty_segment, this could involve either a local or
1462 * a remote pending-ops table.
1465 register_unlink(RelFileNodeBackend rnode)
1467 /* Should never be used with temp relations */
1468 Assert(!RelFileNodeBackendIsTemp(rnode));
1470 if (pendingOpsTable)
1472 /* push it into local pending-ops table */
1473 RememberFsyncRequest(rnode.node, MAIN_FORKNUM,
1474 UNLINK_RELATION_REQUEST);
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.
1483 * XXX should we just leave the file orphaned instead?
1485 Assert(IsUnderPostmaster);
1486 while (!ForwardFsyncRequest(rnode.node, MAIN_FORKNUM,
1487 UNLINK_RELATION_REQUEST))
1488 pg_usleep(10000L); /* 10 msec seems a good number */
1493 * RememberFsyncRequest() -- callback from checkpointer side of fsync request
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.
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.
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
1507 * Note also that we're assuming real segment numbers don't exceed INT_MAX.
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.)
1514 RememberFsyncRequest(RelFileNode rnode, ForkNumber forknum, BlockNumber segno)
1516 Assert(pendingOpsTable);
1518 if (segno == FORGET_RELATION_FSYNC)
1520 /* Remove any pending requests for the relation (one or all forks) */
1521 PendingOperationEntry *entry;
1523 entry = (PendingOperationEntry *) hash_search(pendingOpsTable,
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
1536 if (forknum == InvalidForkNumber)
1538 /* remove requests for all forks */
1539 for (forknum = 0; forknum <= MAX_FORKNUM; forknum++)
1541 bms_free(entry->requests[forknum]);
1542 entry->requests[forknum] = NULL;
1543 entry->canceled[forknum] = true;
1548 /* remove requests for single fork */
1549 bms_free(entry->requests[forknum]);
1550 entry->requests[forknum] = NULL;
1551 entry->canceled[forknum] = true;
1555 else if (segno == FORGET_DATABASE_FSYNC)
1557 /* Remove any pending requests for the entire database */
1558 HASH_SEQ_STATUS hstat;
1559 PendingOperationEntry *entry;
1564 /* Remove fsync requests */
1565 hash_seq_init(&hstat, pendingOpsTable);
1566 while ((entry = (PendingOperationEntry *) hash_seq_search(&hstat)) != NULL)
1568 if (entry->rnode.dbNode == rnode.dbNode)
1570 /* remove requests for all forks */
1571 for (forknum = 0; forknum <= MAX_FORKNUM; forknum++)
1573 bms_free(entry->requests[forknum]);
1574 entry->requests[forknum] = NULL;
1575 entry->canceled[forknum] = true;
1580 /* Remove unlink requests */
1582 for (cell = list_head(pendingUnlinks); cell; cell = next)
1584 PendingUnlinkEntry *entry = (PendingUnlinkEntry *) lfirst(cell);
1587 if (entry->rnode.dbNode == rnode.dbNode)
1589 pendingUnlinks = list_delete_cell(pendingUnlinks, cell, prev);
1596 else if (segno == UNLINK_RELATION_REQUEST)
1598 /* Unlink request: put it in the linked list */
1599 MemoryContext oldcxt = MemoryContextSwitchTo(pendingOpsCxt);
1600 PendingUnlinkEntry *entry;
1602 /* PendingUnlinkEntry doesn't store forknum, since it's always MAIN */
1603 Assert(forknum == MAIN_FORKNUM);
1605 entry = palloc(sizeof(PendingUnlinkEntry));
1606 entry->rnode = rnode;
1607 entry->cycle_ctr = mdckpt_cycle_ctr;
1609 pendingUnlinks = lappend(pendingUnlinks, entry);
1611 MemoryContextSwitchTo(oldcxt);
1615 /* Normal case: enter a request to fsync this segment */
1616 MemoryContext oldcxt = MemoryContextSwitchTo(pendingOpsCxt);
1617 PendingOperationEntry *entry;
1620 entry = (PendingOperationEntry *) hash_search(pendingOpsTable,
1624 /* if new entry, initialize it */
1627 entry->cycle_ctr = mdsync_cycle_ctr;
1628 MemSet(entry->requests, 0, sizeof(entry->requests));
1629 MemSet(entry->canceled, 0, sizeof(entry->canceled));
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.
1638 entry->requests[forknum] = bms_add_member(entry->requests[forknum],
1641 MemoryContextSwitchTo(oldcxt);
1646 * ForgetRelationFsyncRequests -- forget any fsyncs for a relation fork
1648 * forknum == InvalidForkNumber means all forks, although this code doesn't
1649 * actually know that, since it's just forwarding the request elsewhere.
1652 ForgetRelationFsyncRequests(RelFileNode rnode, ForkNumber forknum)
1654 if (pendingOpsTable)
1656 /* standalone backend or startup process: fsync state is local */
1657 RememberFsyncRequest(rnode, forknum, FORGET_RELATION_FSYNC);
1659 else if (IsUnderPostmaster)
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.
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.
1671 while (!ForwardFsyncRequest(rnode, forknum, FORGET_RELATION_FSYNC))
1672 pg_usleep(10000L); /* 10 msec seems a good number */
1675 * Note we don't wait for the checkpointer to actually absorb the
1676 * cancel message; see mdsync() for the implications.
1682 * ForgetDatabaseFsyncRequests -- forget any fsyncs and unlinks for a DB
1685 ForgetDatabaseFsyncRequests(Oid dbid)
1689 rnode.dbNode = dbid;
1693 if (pendingOpsTable)
1695 /* standalone backend or startup process: fsync state is local */
1696 RememberFsyncRequest(rnode, InvalidForkNumber, FORGET_DATABASE_FSYNC);
1698 else if (IsUnderPostmaster)
1700 /* see notes in ForgetRelationFsyncRequests */
1701 while (!ForwardFsyncRequest(rnode, InvalidForkNumber,
1702 FORGET_DATABASE_FSYNC))
1703 pg_usleep(10000L); /* 10 msec seems a good number */
1709 * _fdvec_alloc() -- Make a MdfdVec object.
1714 return (MdfdVec *) MemoryContextAlloc(MdCxt, sizeof(MdfdVec));
1718 * Return the filename for the specified segment of the relation. The
1719 * returned string is palloc'd.
1722 _mdfd_segpath(SMgrRelation reln, ForkNumber forknum, BlockNumber segno)
1727 path = relpath(reln->smgr_rnode, forknum);
1731 fullpath = psprintf("%s.%u", path, segno);
1741 * Open the specified segment of the relation,
1742 * and make a MdfdVec object for it. Returns NULL on failure.
1745 _mdfd_openseg(SMgrRelation reln, ForkNumber forknum, BlockNumber segno,
1752 fullpath = _mdfd_segpath(reln, forknum, segno);
1755 fd = PathNameOpenFile(fullpath, O_RDWR | PG_BINARY | oflags, 0600);
1762 /* allocate an mdfdvec entry for it */
1765 /* fill the entry */
1767 v->mdfd_segno = segno;
1768 v->mdfd_chain = NULL;
1769 Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
1776 * _mdfd_getseg() -- Find the segment of the relation holding the
1779 * If the segment doesn't exist, we ereport, return NULL, or create the
1780 * segment, according to "behavior". Note: skipFsync is only used in the
1781 * EXTENSION_CREATE case.
1784 _mdfd_getseg(SMgrRelation reln, ForkNumber forknum, BlockNumber blkno,
1785 bool skipFsync, int behavior)
1787 MdfdVec *v = mdopen(reln, forknum, behavior);
1788 BlockNumber targetseg;
1789 BlockNumber nextsegno;
1791 /* some way to handle non-existent segments needs to be specified */
1793 (EXTENSION_FAIL | EXTENSION_CREATE | EXTENSION_RETURN_NULL));
1796 return NULL; /* if behavior & EXTENSION_RETURN_NULL */
1798 targetseg = blkno / ((BlockNumber) RELSEG_SIZE);
1799 for (nextsegno = 1; nextsegno <= targetseg; nextsegno++)
1801 Assert(nextsegno == v->mdfd_segno + 1);
1803 if (v->mdfd_chain == NULL)
1805 BlockNumber nblocks = _mdnblocks(reln, forknum, v);
1808 if (nblocks > ((BlockNumber) RELSEG_SIZE))
1809 elog(FATAL, "segment too big");
1811 if ((behavior & EXTENSION_CREATE) ||
1812 (InRecovery && (behavior & EXTENSION_CREATE_RECOVERY)))
1815 * Normally we will create new segments only if authorized by
1816 * the caller (i.e., we are doing mdextend()). But when doing
1817 * WAL recovery, create segments anyway; this allows cases
1818 * such as replaying WAL data that has a write into a
1819 * high-numbered segment of a relation that was later deleted.
1820 * We want to go ahead and create the segments so we can
1821 * finish out the replay. However if the caller has specified
1822 * EXTENSION_REALLY_RETURN_NULL, then extension is not desired
1823 * even in recovery; we won't reach this point in that case.
1825 * We have to maintain the invariant that segments before the
1826 * last active segment are of size RELSEG_SIZE; therefore, if
1827 * extending, pad them out with zeroes if needed. (This only
1828 * matters if in recovery, or if the caller is extending the
1829 * relation discontiguously, but that can happen in hash
1832 if (nblocks < ((BlockNumber) RELSEG_SIZE))
1834 char *zerobuf = palloc0(BLCKSZ);
1836 mdextend(reln, forknum,
1837 nextsegno * ((BlockNumber) RELSEG_SIZE) - 1,
1838 zerobuf, skipFsync);
1843 else if (!(behavior & EXTENSION_DONT_CHECK_SIZE) &&
1844 nblocks < ((BlockNumber) RELSEG_SIZE))
1847 * When not extending (or explicitly including truncated
1848 * segments), only open the next segment if the current one is
1849 * exactly RELSEG_SIZE. If not (this branch), either return
1852 if (behavior & EXTENSION_RETURN_NULL)
1855 * Some callers discern between reasons for _mdfd_getseg()
1856 * returning NULL based on errno. As there's no failing
1857 * syscall involved in this case, explicitly set errno to
1858 * ENOENT, as that seems the closest interpretation.
1865 (errcode_for_file_access(),
1866 errmsg("could not open file \"%s\" (target block %u): previous segment is only %u blocks",
1867 _mdfd_segpath(reln, forknum, nextsegno),
1871 v->mdfd_chain = _mdfd_openseg(reln, forknum, nextsegno, flags);
1873 if (v->mdfd_chain == NULL)
1875 if ((behavior & EXTENSION_RETURN_NULL) &&
1876 FILE_POSSIBLY_DELETED(errno))
1879 (errcode_for_file_access(),
1880 errmsg("could not open file \"%s\" (target block %u): %m",
1881 _mdfd_segpath(reln, forknum, nextsegno),
1891 * Get number of blocks present in a single disk file
1894 _mdnblocks(SMgrRelation reln, ForkNumber forknum, MdfdVec *seg)
1898 len = FileSeek(seg->mdfd_vfd, 0L, SEEK_END);
1901 (errcode_for_file_access(),
1902 errmsg("could not seek to end of file \"%s\": %m",
1903 FilePathName(seg->mdfd_vfd))));
1904 /* note that this calculation will ignore any partial block at EOF */
1905 return (BlockNumber) (len / BLCKSZ);