1 /*-------------------------------------------------------------------------
4 * public interface routines to storage manager switch.
6 * All file system operations in POSTGRES dispatch through these
9 * Portions Copyright (c) 1996-2022, PostgreSQL Global Development Group
10 * Portions Copyright (c) 1994, Regents of the University of California
14 * src/backend/storage/smgr/smgr.c
16 *-------------------------------------------------------------------------
20 #include "access/xlogutils.h"
21 #include "lib/ilist.h"
22 #include "storage/bufmgr.h"
23 #include "storage/ipc.h"
24 #include "storage/md.h"
25 #include "storage/smgr.h"
26 #include "utils/hsearch.h"
27 #include "utils/inval.h"
31 * This struct of function pointers defines the API between smgr.c and
32 * any individual storage manager module. Note that smgr subfunctions are
33 * generally expected to report problems via elog(ERROR). An exception is
34 * that smgr_unlink should use elog(WARNING), rather than erroring out,
35 * because we normally unlink relations during post-commit/abort cleanup,
36 * and so it's too late to raise an error. Also, various conditions that
37 * would normally be errors should be allowed during bootstrap and/or WAL
38 * recovery --- see comments in md.c for details.
42 void (*smgr_init
) (void); /* may be NULL */
43 void (*smgr_shutdown
) (void); /* may be NULL */
44 void (*smgr_open
) (SMgrRelation reln
);
45 void (*smgr_close
) (SMgrRelation reln
, ForkNumber forknum
);
46 void (*smgr_create
) (SMgrRelation reln
, ForkNumber forknum
,
48 bool (*smgr_exists
) (SMgrRelation reln
, ForkNumber forknum
);
49 void (*smgr_unlink
) (RelFileNodeBackend rnode
, ForkNumber forknum
,
51 void (*smgr_extend
) (SMgrRelation reln
, ForkNumber forknum
,
52 BlockNumber blocknum
, char *buffer
, bool skipFsync
);
53 bool (*smgr_prefetch
) (SMgrRelation reln
, ForkNumber forknum
,
54 BlockNumber blocknum
);
55 void (*smgr_read
) (SMgrRelation reln
, ForkNumber forknum
,
56 BlockNumber blocknum
, char *buffer
);
57 void (*smgr_write
) (SMgrRelation reln
, ForkNumber forknum
,
58 BlockNumber blocknum
, char *buffer
, bool skipFsync
);
59 void (*smgr_writeback
) (SMgrRelation reln
, ForkNumber forknum
,
60 BlockNumber blocknum
, BlockNumber nblocks
);
61 BlockNumber (*smgr_nblocks
) (SMgrRelation reln
, ForkNumber forknum
);
62 void (*smgr_truncate
) (SMgrRelation reln
, ForkNumber forknum
,
64 void (*smgr_immedsync
) (SMgrRelation reln
, ForkNumber forknum
);
67 static const f_smgr smgrsw
[] = {
71 .smgr_shutdown
= NULL
,
73 .smgr_close
= mdclose
,
74 .smgr_create
= mdcreate
,
75 .smgr_exists
= mdexists
,
76 .smgr_unlink
= mdunlink
,
77 .smgr_extend
= mdextend
,
78 .smgr_prefetch
= mdprefetch
,
80 .smgr_write
= mdwrite
,
81 .smgr_writeback
= mdwriteback
,
82 .smgr_nblocks
= mdnblocks
,
83 .smgr_truncate
= mdtruncate
,
84 .smgr_immedsync
= mdimmedsync
,
88 static const int NSmgr
= lengthof(smgrsw
);
91 * Each backend has a hashtable that stores all extant SMgrRelation objects.
92 * In addition, "unowned" SMgrRelation objects are chained together in a list.
94 static HTAB
*SMgrRelationHash
= NULL
;
96 static dlist_head unowned_relns
;
98 /* local function prototypes */
99 static void smgrshutdown(int code
, Datum arg
);
103 * smgrinit(), smgrshutdown() -- Initialize or shut down storage
106 * Note: smgrinit is called during backend startup (normal or standalone
107 * case), *not* during postmaster start. Therefore, any resources created
108 * here or destroyed in smgrshutdown are backend-local.
115 for (i
= 0; i
< NSmgr
; i
++)
117 if (smgrsw
[i
].smgr_init
)
118 smgrsw
[i
].smgr_init();
121 /* register the shutdown proc */
122 on_proc_exit(smgrshutdown
, 0);
126 * on_proc_exit hook for smgr cleanup during backend shutdown
129 smgrshutdown(int code
, Datum arg
)
133 for (i
= 0; i
< NSmgr
; i
++)
135 if (smgrsw
[i
].smgr_shutdown
)
136 smgrsw
[i
].smgr_shutdown();
141 * smgropen() -- Return an SMgrRelation object, creating it if need be.
143 * This does not attempt to actually open the underlying file.
146 smgropen(RelFileNode rnode
, BackendId backend
)
148 RelFileNodeBackend brnode
;
152 if (SMgrRelationHash
== NULL
)
154 /* First time through: initialize the hash table */
157 ctl
.keysize
= sizeof(RelFileNodeBackend
);
158 ctl
.entrysize
= sizeof(SMgrRelationData
);
159 SMgrRelationHash
= hash_create("smgr relation table", 400,
160 &ctl
, HASH_ELEM
| HASH_BLOBS
);
161 dlist_init(&unowned_relns
);
164 /* Look up or create an entry */
166 brnode
.backend
= backend
;
167 reln
= (SMgrRelation
) hash_search(SMgrRelationHash
,
171 /* Initialize it if not present before */
174 /* hash_search already filled in the lookup key */
175 reln
->smgr_owner
= NULL
;
176 reln
->smgr_targblock
= InvalidBlockNumber
;
177 for (int i
= 0; i
<= MAX_FORKNUM
; ++i
)
178 reln
->smgr_cached_nblocks
[i
] = InvalidBlockNumber
;
179 reln
->smgr_which
= 0; /* we only have md.c at present */
181 /* implementation-specific initialization */
182 smgrsw
[reln
->smgr_which
].smgr_open(reln
);
184 /* it has no owner yet */
185 dlist_push_tail(&unowned_relns
, &reln
->node
);
192 * smgrsetowner() -- Establish a long-lived reference to an SMgrRelation object
194 * There can be only one owner at a time; this is sufficient since currently
195 * the only such owners exist in the relcache.
198 smgrsetowner(SMgrRelation
*owner
, SMgrRelation reln
)
200 /* We don't support "disowning" an SMgrRelation here, use smgrclearowner */
201 Assert(owner
!= NULL
);
204 * First, unhook any old owner. (Normally there shouldn't be any, but it
205 * seems possible that this can happen during swap_relation_files()
206 * depending on the order of processing. It's ok to close the old
207 * relcache entry early in that case.)
209 * If there isn't an old owner, then the reln should be in the unowned
210 * list, and we need to remove it.
212 if (reln
->smgr_owner
)
213 *(reln
->smgr_owner
) = NULL
;
215 dlist_delete(&reln
->node
);
217 /* Now establish the ownership relationship. */
218 reln
->smgr_owner
= owner
;
223 * smgrclearowner() -- Remove long-lived reference to an SMgrRelation object
227 smgrclearowner(SMgrRelation
*owner
, SMgrRelation reln
)
229 /* Do nothing if the SMgrRelation object is not owned by the owner */
230 if (reln
->smgr_owner
!= owner
)
233 /* unset the owner's reference */
236 /* unset our reference to the owner */
237 reln
->smgr_owner
= NULL
;
239 /* add to list of unowned relations */
240 dlist_push_tail(&unowned_relns
, &reln
->node
);
244 * smgrexists() -- Does the underlying file for a fork exist?
247 smgrexists(SMgrRelation reln
, ForkNumber forknum
)
249 return smgrsw
[reln
->smgr_which
].smgr_exists(reln
, forknum
);
253 * smgrclose() -- Close and delete an SMgrRelation object.
256 smgrclose(SMgrRelation reln
)
261 for (forknum
= 0; forknum
<= MAX_FORKNUM
; forknum
++)
262 smgrsw
[reln
->smgr_which
].smgr_close(reln
, forknum
);
264 owner
= reln
->smgr_owner
;
267 dlist_delete(&reln
->node
);
269 if (hash_search(SMgrRelationHash
,
270 (void *) &(reln
->smgr_rnode
),
271 HASH_REMOVE
, NULL
) == NULL
)
272 elog(ERROR
, "SMgrRelation hashtable corrupted");
275 * Unhook the owner pointer, if any. We do this last since in the remote
276 * possibility of failure above, the SMgrRelation object will still exist.
283 * smgrcloseall() -- Close all existing SMgrRelation objects.
288 HASH_SEQ_STATUS status
;
291 /* Nothing to do if hashtable not set up */
292 if (SMgrRelationHash
== NULL
)
295 hash_seq_init(&status
, SMgrRelationHash
);
297 while ((reln
= (SMgrRelation
) hash_seq_search(&status
)) != NULL
)
302 * smgrclosenode() -- Close SMgrRelation object for given RelFileNode,
305 * This has the same effects as smgrclose(smgropen(rnode)), but it avoids
306 * uselessly creating a hashtable entry only to drop it again when no
307 * such entry exists already.
310 smgrclosenode(RelFileNodeBackend rnode
)
314 /* Nothing to do if hashtable not set up */
315 if (SMgrRelationHash
== NULL
)
318 reln
= (SMgrRelation
) hash_search(SMgrRelationHash
,
326 * smgrcreate() -- Create a new relation.
328 * Given an already-created (but presumably unused) SMgrRelation,
329 * cause the underlying disk file or other storage for the fork
333 smgrcreate(SMgrRelation reln
, ForkNumber forknum
, bool isRedo
)
335 smgrsw
[reln
->smgr_which
].smgr_create(reln
, forknum
, isRedo
);
339 * smgrdosyncall() -- Immediately sync all forks of all given relations
341 * All forks of all given relations are synced out to the store.
343 * This is equivalent to FlushRelationBuffers() for each smgr relation,
344 * then calling smgrimmedsync() for all forks of each relation, but it's
345 * significantly quicker so should be preferred when possible.
348 smgrdosyncall(SMgrRelation
*rels
, int nrels
)
356 FlushRelationsAllBuffers(rels
, nrels
);
359 * Sync the physical file(s).
361 for (i
= 0; i
< nrels
; i
++)
363 int which
= rels
[i
]->smgr_which
;
365 for (forknum
= 0; forknum
<= MAX_FORKNUM
; forknum
++)
367 if (smgrsw
[which
].smgr_exists(rels
[i
], forknum
))
368 smgrsw
[which
].smgr_immedsync(rels
[i
], forknum
);
374 * smgrdounlinkall() -- Immediately unlink all forks of all given relations
376 * All forks of all given relations are removed from the store. This
377 * should not be used during transactional operations, since it can't be
380 * If isRedo is true, it is okay for the underlying file(s) to be gone
384 smgrdounlinkall(SMgrRelation
*rels
, int nrels
, bool isRedo
)
387 RelFileNodeBackend
*rnodes
;
394 * Get rid of any remaining buffers for the relations. bufmgr will just
395 * drop them without bothering to write the contents.
397 DropRelFileNodesAllBuffers(rels
, nrels
);
400 * create an array which contains all relations to be dropped, and close
401 * each relation's forks at the smgr level while at it
403 rnodes
= palloc(sizeof(RelFileNodeBackend
) * nrels
);
404 for (i
= 0; i
< nrels
; i
++)
406 RelFileNodeBackend rnode
= rels
[i
]->smgr_rnode
;
407 int which
= rels
[i
]->smgr_which
;
411 /* Close the forks at smgr level */
412 for (forknum
= 0; forknum
<= MAX_FORKNUM
; forknum
++)
413 smgrsw
[which
].smgr_close(rels
[i
], forknum
);
417 * It'd be nice to tell the stats collector to forget them immediately,
418 * too. But we can't because we don't know the OIDs.
422 * Send a shared-inval message to force other backends to close any
423 * dangling smgr references they may have for these rels. We should do
424 * this before starting the actual unlinking, in case we fail partway
425 * through that step. Note that the sinval messages will eventually come
426 * back to this backend, too, and thereby provide a backstop that we
427 * closed our own smgr rel.
429 for (i
= 0; i
< nrels
; i
++)
430 CacheInvalidateSmgr(rnodes
[i
]);
433 * Delete the physical file(s).
435 * Note: smgr_unlink must treat deletion failure as a WARNING, not an
436 * ERROR, because we've already decided to commit or abort the current
440 for (i
= 0; i
< nrels
; i
++)
442 int which
= rels
[i
]->smgr_which
;
444 for (forknum
= 0; forknum
<= MAX_FORKNUM
; forknum
++)
445 smgrsw
[which
].smgr_unlink(rnodes
[i
], forknum
, isRedo
);
453 * smgrextend() -- Add a new block to a file.
455 * The semantics are nearly the same as smgrwrite(): write at the
456 * specified position. However, this is to be used for the case of
457 * extending a relation (i.e., blocknum is at or beyond the current
458 * EOF). Note that we assume writing a block beyond current EOF
459 * causes intervening file space to become filled with zeroes.
462 smgrextend(SMgrRelation reln
, ForkNumber forknum
, BlockNumber blocknum
,
463 char *buffer
, bool skipFsync
)
465 smgrsw
[reln
->smgr_which
].smgr_extend(reln
, forknum
, blocknum
,
469 * Normally we expect this to increase nblocks by one, but if the cached
470 * value isn't as expected, just invalidate it so the next call asks the
473 if (reln
->smgr_cached_nblocks
[forknum
] == blocknum
)
474 reln
->smgr_cached_nblocks
[forknum
] = blocknum
+ 1;
476 reln
->smgr_cached_nblocks
[forknum
] = InvalidBlockNumber
;
480 * smgrprefetch() -- Initiate asynchronous read of the specified block of a relation.
482 * In recovery only, this can return false to indicate that a file
483 * doesn't exist (presumably it has been dropped by a later WAL
487 smgrprefetch(SMgrRelation reln
, ForkNumber forknum
, BlockNumber blocknum
)
489 return smgrsw
[reln
->smgr_which
].smgr_prefetch(reln
, forknum
, blocknum
);
493 * smgrread() -- read a particular block from a relation into the supplied
496 * This routine is called from the buffer manager in order to
497 * instantiate pages in the shared buffer cache. All storage managers
498 * return pages in the format that POSTGRES expects.
501 smgrread(SMgrRelation reln
, ForkNumber forknum
, BlockNumber blocknum
,
504 smgrsw
[reln
->smgr_which
].smgr_read(reln
, forknum
, blocknum
, buffer
);
508 * smgrwrite() -- Write the supplied buffer out.
510 * This is to be used only for updating already-existing blocks of a
511 * relation (ie, those before the current EOF). To extend a relation,
514 * This is not a synchronous write -- the block is not necessarily
515 * on disk at return, only dumped out to the kernel. However,
516 * provisions will be made to fsync the write before the next checkpoint.
518 * skipFsync indicates that the caller will make other provisions to
519 * fsync the relation, so we needn't bother. Temporary relations also
520 * do not require fsync.
523 smgrwrite(SMgrRelation reln
, ForkNumber forknum
, BlockNumber blocknum
,
524 char *buffer
, bool skipFsync
)
526 smgrsw
[reln
->smgr_which
].smgr_write(reln
, forknum
, blocknum
,
532 * smgrwriteback() -- Trigger kernel writeback for the supplied range of
536 smgrwriteback(SMgrRelation reln
, ForkNumber forknum
, BlockNumber blocknum
,
539 smgrsw
[reln
->smgr_which
].smgr_writeback(reln
, forknum
, blocknum
,
544 * smgrnblocks() -- Calculate the number of blocks in the
548 smgrnblocks(SMgrRelation reln
, ForkNumber forknum
)
552 /* Check and return if we get the cached value for the number of blocks. */
553 result
= smgrnblocks_cached(reln
, forknum
);
554 if (result
!= InvalidBlockNumber
)
557 result
= smgrsw
[reln
->smgr_which
].smgr_nblocks(reln
, forknum
);
559 reln
->smgr_cached_nblocks
[forknum
] = result
;
565 * smgrnblocks_cached() -- Get the cached number of blocks in the supplied
568 * Returns an InvalidBlockNumber when not in recovery and when the relation
569 * fork size is not cached.
572 smgrnblocks_cached(SMgrRelation reln
, ForkNumber forknum
)
575 * For now, we only use cached values in recovery due to lack of a shared
576 * invalidation mechanism for changes in file size.
578 if (InRecovery
&& reln
->smgr_cached_nblocks
[forknum
] != InvalidBlockNumber
)
579 return reln
->smgr_cached_nblocks
[forknum
];
581 return InvalidBlockNumber
;
585 * smgrtruncate() -- Truncate the given forks of supplied relation to
586 * each specified numbers of blocks
588 * The truncation is done immediately, so this can't be rolled back.
590 * The caller must hold AccessExclusiveLock on the relation, to ensure that
591 * other backends receive the smgr invalidation event that this function sends
592 * before they access any forks of the relation again.
595 smgrtruncate(SMgrRelation reln
, ForkNumber
*forknum
, int nforks
, BlockNumber
*nblocks
)
600 * Get rid of any buffers for the about-to-be-deleted blocks. bufmgr will
601 * just drop them without bothering to write the contents.
603 DropRelFileNodeBuffers(reln
, forknum
, nforks
, nblocks
);
606 * Send a shared-inval message to force other backends to close any smgr
607 * references they may have for this rel. This is useful because they
608 * might have open file pointers to segments that got removed, and/or
609 * smgr_targblock variables pointing past the new rel end. (The inval
610 * message will come back to our backend, too, causing a
611 * probably-unnecessary local smgr flush. But we don't expect that this
612 * is a performance-critical path.) As in the unlink code, we want to be
613 * sure the message is sent before we start changing things on-disk.
615 CacheInvalidateSmgr(reln
->smgr_rnode
);
617 /* Do the truncation */
618 for (i
= 0; i
< nforks
; i
++)
620 /* Make the cached size is invalid if we encounter an error. */
621 reln
->smgr_cached_nblocks
[forknum
[i
]] = InvalidBlockNumber
;
623 smgrsw
[reln
->smgr_which
].smgr_truncate(reln
, forknum
[i
], nblocks
[i
]);
626 * We might as well update the local smgr_cached_nblocks values. The
627 * smgr cache inval message that this function sent will cause other
628 * backends to invalidate their copies of smgr_fsm_nblocks and
629 * smgr_vm_nblocks, and these ones too at the next command boundary.
630 * But these ensure they aren't outright wrong until then.
632 reln
->smgr_cached_nblocks
[forknum
[i
]] = nblocks
[i
];
637 * smgrimmedsync() -- Force the specified relation to stable storage.
639 * Synchronously force all previous writes to the specified relation
642 * This is useful for building completely new relations (eg, new
643 * indexes). Instead of incrementally WAL-logging the index build
644 * steps, we can just write completed index pages to disk with smgrwrite
645 * or smgrextend, and then fsync the completed index file before
646 * committing the transaction. (This is sufficient for purposes of
647 * crash recovery, since it effectively duplicates forcing a checkpoint
648 * for the completed index. But it is *not* sufficient if one wishes
649 * to use the WAL log for PITR or replication purposes: in that case
650 * we have to make WAL entries as well.)
652 * The preceding writes should specify skipFsync = true to avoid
653 * duplicative fsyncs.
655 * Note that you need to do FlushRelationBuffers() first if there is
656 * any possibility that there are dirty buffers for the relation;
657 * otherwise the sync is not very meaningful.
660 smgrimmedsync(SMgrRelation reln
, ForkNumber forknum
)
662 smgrsw
[reln
->smgr_which
].smgr_immedsync(reln
, forknum
);
668 * This routine is called during transaction commit or abort (it doesn't
669 * particularly care which). All transient SMgrRelation objects are closed.
671 * We do this as a compromise between wanting transient SMgrRelations to
672 * live awhile (to amortize the costs of blind writes of multiple blocks)
673 * and needing them to not live forever (since we're probably holding open
674 * a kernel file descriptor for the underlying file, and we need to ensure
675 * that gets closed reasonably soon if the file gets deleted).
680 dlist_mutable_iter iter
;
683 * Zap all unowned SMgrRelations. We rely on smgrclose() to remove each
686 dlist_foreach_modify(iter
, &unowned_relns
)
688 SMgrRelation rel
= dlist_container(SMgrRelationData
, node
,
691 Assert(rel
->smgr_owner
== NULL
);