Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * inval.c
4 : * POSTGRES cache invalidation dispatcher code.
5 : *
6 : * This is subtle stuff, so pay attention:
7 : *
8 : * When a tuple is updated or deleted, our standard visibility rules
9 : * consider that it is *still valid* so long as we are in the same command,
10 : * ie, until the next CommandCounterIncrement() or transaction commit.
11 : * (See access/heap/heapam_visibility.c, and note that system catalogs are
12 : * generally scanned under the most current snapshot available, rather than
13 : * the transaction snapshot.) At the command boundary, the old tuple stops
14 : * being valid and the new version, if any, becomes valid. Therefore,
15 : * we cannot simply flush a tuple from the system caches during heap_update()
16 : * or heap_delete(). The tuple is still good at that point; what's more,
17 : * even if we did flush it, it might be reloaded into the caches by a later
18 : * request in the same command. So the correct behavior is to keep a list
19 : * of outdated (updated/deleted) tuples and then do the required cache
20 : * flushes at the next command boundary. We must also keep track of
21 : * inserted tuples so that we can flush "negative" cache entries that match
22 : * the new tuples; again, that mustn't happen until end of command.
23 : *
24 : * Once we have finished the command, we still need to remember inserted
25 : * tuples (including new versions of updated tuples), so that we can flush
26 : * them from the caches if we abort the transaction. Similarly, we'd better
27 : * be able to flush "negative" cache entries that may have been loaded in
28 : * place of deleted tuples, so we still need the deleted ones too.
29 : *
30 : * If we successfully complete the transaction, we have to broadcast all
31 : * these invalidation events to other backends (via the SI message queue)
32 : * so that they can flush obsolete entries from their caches. Note we have
33 : * to record the transaction commit before sending SI messages, otherwise
34 : * the other backends won't see our updated tuples as good.
35 : *
36 : * When a subtransaction aborts, we can process and discard any events
37 : * it has queued. When a subtransaction commits, we just add its events
38 : * to the pending lists of the parent transaction.
39 : *
40 : * In short, we need to remember until xact end every insert or delete
41 : * of a tuple that might be in the system caches. Updates are treated as
42 : * two events, delete + insert, for simplicity. (If the update doesn't
43 : * change the tuple hash value, catcache.c optimizes this into one event.)
44 : *
45 : * We do not need to register EVERY tuple operation in this way, just those
46 : * on tuples in relations that have associated catcaches. We do, however,
47 : * have to register every operation on every tuple that *could* be in a
48 : * catcache, whether or not it currently is in our cache. Also, if the
49 : * tuple is in a relation that has multiple catcaches, we need to register
50 : * an invalidation message for each such catcache. catcache.c's
51 : * PrepareToInvalidateCacheTuple() routine provides the knowledge of which
52 : * catcaches may need invalidation for a given tuple.
53 : *
54 : * Also, whenever we see an operation on a pg_class, pg_attribute, or
55 : * pg_index tuple, we register a relcache flush operation for the relation
56 : * described by that tuple (as specified in CacheInvalidateHeapTuple()).
57 : * Likewise for pg_constraint tuples for foreign keys on relations.
58 : *
59 : * We keep the relcache flush requests in lists separate from the catcache
60 : * tuple flush requests. This allows us to issue all the pending catcache
61 : * flushes before we issue relcache flushes, which saves us from loading
62 : * a catcache tuple during relcache load only to flush it again right away.
63 : * Also, we avoid queuing multiple relcache flush requests for the same
64 : * relation, since a relcache flush is relatively expensive to do.
65 : * (XXX is it worth testing likewise for duplicate catcache flush entries?
66 : * Probably not.)
67 : *
68 : * Many subsystems own higher-level caches that depend on relcache and/or
69 : * catcache, and they register callbacks here to invalidate their caches.
70 : * While building a higher-level cache entry, a backend may receive a
71 : * callback for the being-built entry or one of its dependencies. This
72 : * implies the new higher-level entry would be born stale, and it might
73 : * remain stale for the life of the backend. Many caches do not prevent
74 : * that. They rely on DDL for can't-miss catalog changes taking
75 : * AccessExclusiveLock on suitable objects. (For a change made with less
76 : * locking, backends might never read the change.) The relation cache,
77 : * however, needs to reflect changes from CREATE INDEX CONCURRENTLY no later
78 : * than the beginning of the next transaction. Hence, when a relevant
79 : * invalidation callback arrives during a build, relcache.c reattempts that
80 : * build. Caches with similar needs could do likewise.
81 : *
82 : * If a relcache flush is issued for a system relation that we preload
83 : * from the relcache init file, we must also delete the init file so that
84 : * it will be rebuilt during the next backend restart. The actual work of
85 : * manipulating the init file is in relcache.c, but we keep track of the
86 : * need for it here.
87 : *
88 : * Currently, inval messages are sent without regard for the possibility
89 : * that the object described by the catalog tuple might be a session-local
90 : * object such as a temporary table. This is because (1) this code has
91 : * no practical way to tell the difference, and (2) it is not certain that
92 : * other backends don't have catalog cache or even relcache entries for
93 : * such tables, anyway; there is nothing that prevents that. It might be
94 : * worth trying to avoid sending such inval traffic in the future, if those
95 : * problems can be overcome cheaply.
96 : *
97 : * When wal_level=logical, write invalidations into WAL at each command end to
98 : * support the decoding of the in-progress transactions. See
99 : * CommandEndInvalidationMessages.
100 : *
101 : * Portions Copyright (c) 1996-2023, PostgreSQL Global Development Group
102 : * Portions Copyright (c) 1994, Regents of the University of California
103 : *
104 : * IDENTIFICATION
105 : * src/backend/utils/cache/inval.c
106 : *
107 : *-------------------------------------------------------------------------
108 : */
109 : #include "postgres.h"
110 :
111 : #include <limits.h>
112 :
113 : #include "access/htup_details.h"
114 : #include "access/xact.h"
115 : #include "access/xloginsert.h"
116 : #include "catalog/catalog.h"
117 : #include "catalog/pg_constraint.h"
118 : #include "miscadmin.h"
119 : #include "storage/sinval.h"
120 : #include "storage/smgr.h"
121 : #include "utils/catcache.h"
122 : #include "utils/guc.h"
123 : #include "utils/inval.h"
124 : #include "utils/memdebug.h"
125 : #include "utils/memutils.h"
126 : #include "utils/rel.h"
127 : #include "utils/relmapper.h"
128 : #include "utils/snapmgr.h"
129 : #include "utils/syscache.h"
130 :
131 :
132 : /*
133 : * Pending requests are stored as ready-to-send SharedInvalidationMessages.
134 : * We keep the messages themselves in arrays in TopTransactionContext
135 : * (there are separate arrays for catcache and relcache messages). Control
136 : * information is kept in a chain of TransInvalidationInfo structs, also
137 : * allocated in TopTransactionContext. (We could keep a subtransaction's
138 : * TransInvalidationInfo in its CurTransactionContext; but that's more
139 : * wasteful not less so, since in very many scenarios it'd be the only
140 : * allocation in the subtransaction's CurTransactionContext.)
141 : *
142 : * We can store the message arrays densely, and yet avoid moving data around
143 : * within an array, because within any one subtransaction we need only
144 : * distinguish between messages emitted by prior commands and those emitted
145 : * by the current command. Once a command completes and we've done local
146 : * processing on its messages, we can fold those into the prior-commands
147 : * messages just by changing array indexes in the TransInvalidationInfo
148 : * struct. Similarly, we need distinguish messages of prior subtransactions
149 : * from those of the current subtransaction only until the subtransaction
150 : * completes, after which we adjust the array indexes in the parent's
151 : * TransInvalidationInfo to include the subtransaction's messages.
152 : *
153 : * The ordering of the individual messages within a command's or
154 : * subtransaction's output is not considered significant, although this
155 : * implementation happens to preserve the order in which they were queued.
156 : * (Previous versions of this code did not preserve it.)
157 : *
158 : * For notational convenience, control information is kept in two-element
159 : * arrays, the first for catcache messages and the second for relcache
160 : * messages.
161 : */
162 : #define CatCacheMsgs 0
163 : #define RelCacheMsgs 1
164 :
165 : /* Pointers to main arrays in TopTransactionContext */
166 : typedef struct InvalMessageArray
167 : {
168 : SharedInvalidationMessage *msgs; /* palloc'd array (can be expanded) */
169 : int maxmsgs; /* current allocated size of array */
170 : } InvalMessageArray;
171 :
172 : static InvalMessageArray InvalMessageArrays[2];
173 :
174 : /* Control information for one logical group of messages */
175 : typedef struct InvalidationMsgsGroup
176 : {
177 : int firstmsg[2]; /* first index in relevant array */
178 : int nextmsg[2]; /* last+1 index */
179 : } InvalidationMsgsGroup;
180 :
181 : /* Macros to help preserve InvalidationMsgsGroup abstraction */
182 : #define SetSubGroupToFollow(targetgroup, priorgroup, subgroup) \
183 : do { \
184 : (targetgroup)->firstmsg[subgroup] = \
185 : (targetgroup)->nextmsg[subgroup] = \
186 : (priorgroup)->nextmsg[subgroup]; \
187 : } while (0)
188 :
189 : #define SetGroupToFollow(targetgroup, priorgroup) \
190 : do { \
191 : SetSubGroupToFollow(targetgroup, priorgroup, CatCacheMsgs); \
192 : SetSubGroupToFollow(targetgroup, priorgroup, RelCacheMsgs); \
193 : } while (0)
194 :
195 : #define NumMessagesInSubGroup(group, subgroup) \
196 : ((group)->nextmsg[subgroup] - (group)->firstmsg[subgroup])
197 :
198 : #define NumMessagesInGroup(group) \
199 : (NumMessagesInSubGroup(group, CatCacheMsgs) + \
200 : NumMessagesInSubGroup(group, RelCacheMsgs))
201 :
202 :
203 : /*----------------
204 : * Invalidation messages are divided into two groups:
205 : * 1) events so far in current command, not yet reflected to caches.
206 : * 2) events in previous commands of current transaction; these have
207 : * been reflected to local caches, and must be either broadcast to
208 : * other backends or rolled back from local cache when we commit
209 : * or abort the transaction.
210 : * Actually, we need such groups for each level of nested transaction,
211 : * so that we can discard events from an aborted subtransaction. When
212 : * a subtransaction commits, we append its events to the parent's groups.
213 : *
214 : * The relcache-file-invalidated flag can just be a simple boolean,
215 : * since we only act on it at transaction commit; we don't care which
216 : * command of the transaction set it.
217 : *----------------
218 : */
219 :
220 : typedef struct TransInvalidationInfo
221 : {
222 : /* Back link to parent transaction's info */
223 : struct TransInvalidationInfo *parent;
224 :
225 : /* Subtransaction nesting depth */
226 : int my_level;
227 :
228 : /* Events emitted by current command */
229 : InvalidationMsgsGroup CurrentCmdInvalidMsgs;
230 :
231 : /* Events emitted by previous commands of this (sub)transaction */
232 : InvalidationMsgsGroup PriorCmdInvalidMsgs;
233 :
234 : /* init file must be invalidated? */
235 : bool RelcacheInitFileInval;
236 : } TransInvalidationInfo;
237 :
238 : static TransInvalidationInfo *transInvalInfo = NULL;
239 :
240 : /* GUC storage */
241 : int debug_discard_caches = 0;
242 :
243 : /*
244 : * Dynamically-registered callback functions. Current implementation
245 : * assumes there won't be enough of these to justify a dynamically resizable
246 : * array; it'd be easy to improve that if needed.
247 : *
248 : * To avoid searching in CallSyscacheCallbacks, all callbacks for a given
249 : * syscache are linked into a list pointed to by syscache_callback_links[id].
250 : * The link values are syscache_callback_list[] index plus 1, or 0 for none.
251 : */
252 :
253 : #define MAX_SYSCACHE_CALLBACKS 64
254 : #define MAX_RELCACHE_CALLBACKS 10
255 :
256 : static struct SYSCACHECALLBACK
257 : {
258 : int16 id; /* cache number */
259 : int16 link; /* next callback index+1 for same cache */
260 : SyscacheCallbackFunction function;
261 : Datum arg;
262 : } syscache_callback_list[MAX_SYSCACHE_CALLBACKS];
263 :
264 : static int16 syscache_callback_links[SysCacheSize];
265 :
266 : static int syscache_callback_count = 0;
267 :
268 : static struct RELCACHECALLBACK
269 : {
270 : RelcacheCallbackFunction function;
271 : Datum arg;
272 : } relcache_callback_list[MAX_RELCACHE_CALLBACKS];
273 :
274 : static int relcache_callback_count = 0;
275 :
276 : /* ----------------------------------------------------------------
277 : * Invalidation subgroup support functions
278 : * ----------------------------------------------------------------
279 : */
280 :
281 : /*
282 : * AddInvalidationMessage
283 : * Add an invalidation message to a (sub)group.
284 : *
285 : * The group must be the last active one, since we assume we can add to the
286 : * end of the relevant InvalMessageArray.
287 : *
288 : * subgroup must be CatCacheMsgs or RelCacheMsgs.
289 : */
290 : static void
291 5155228 : AddInvalidationMessage(InvalidationMsgsGroup *group, int subgroup,
292 : const SharedInvalidationMessage *msg)
293 : {
294 5155228 : InvalMessageArray *ima = &InvalMessageArrays[subgroup];
295 5155228 : int nextindex = group->nextmsg[subgroup];
296 :
297 5155228 : if (nextindex >= ima->maxmsgs)
298 : {
299 347794 : if (ima->msgs == NULL)
300 : {
301 : /* Create new storage array in TopTransactionContext */
302 301188 : int reqsize = 32; /* arbitrary */
303 :
304 301188 : ima->msgs = (SharedInvalidationMessage *)
305 301188 : MemoryContextAlloc(TopTransactionContext,
306 : reqsize * sizeof(SharedInvalidationMessage));
307 301188 : ima->maxmsgs = reqsize;
308 : Assert(nextindex == 0);
309 : }
310 : else
311 : {
312 : /* Enlarge storage array */
313 46606 : int reqsize = 2 * ima->maxmsgs;
314 :
315 46606 : ima->msgs = (SharedInvalidationMessage *)
316 46606 : repalloc(ima->msgs,
317 : reqsize * sizeof(SharedInvalidationMessage));
318 46606 : ima->maxmsgs = reqsize;
319 : }
320 : }
321 : /* Okay, add message to current group */
322 5155228 : ima->msgs[nextindex] = *msg;
323 5155228 : group->nextmsg[subgroup]++;
324 5155228 : }
325 :
326 : /*
327 : * Append one subgroup of invalidation messages to another, resetting
328 : * the source subgroup to empty.
329 : */
330 : static void
331 1447160 : AppendInvalidationMessageSubGroup(InvalidationMsgsGroup *dest,
332 : InvalidationMsgsGroup *src,
333 : int subgroup)
334 : {
335 : /* Messages must be adjacent in main array */
336 : Assert(dest->nextmsg[subgroup] == src->firstmsg[subgroup]);
337 :
338 : /* ... which makes this easy: */
339 1447160 : dest->nextmsg[subgroup] = src->nextmsg[subgroup];
340 :
341 : /*
342 : * This is handy for some callers and irrelevant for others. But we do it
343 : * always, reasoning that it's bad to leave different groups pointing at
344 : * the same fragment of the message array.
345 : */
346 1447160 : SetSubGroupToFollow(src, dest, subgroup);
347 1447160 : }
348 :
349 : /*
350 : * Process a subgroup of invalidation messages.
351 : *
352 : * This is a macro that executes the given code fragment for each message in
353 : * a message subgroup. The fragment should refer to the message as *msg.
354 : */
355 : #define ProcessMessageSubGroup(group, subgroup, codeFragment) \
356 : do { \
357 : int _msgindex = (group)->firstmsg[subgroup]; \
358 : int _endmsg = (group)->nextmsg[subgroup]; \
359 : for (; _msgindex < _endmsg; _msgindex++) \
360 : { \
361 : SharedInvalidationMessage *msg = \
362 : &InvalMessageArrays[subgroup].msgs[_msgindex]; \
363 : codeFragment; \
364 : } \
365 : } while (0)
366 :
367 : /*
368 : * Process a subgroup of invalidation messages as an array.
369 : *
370 : * As above, but the code fragment can handle an array of messages.
371 : * The fragment should refer to the messages as msgs[], with n entries.
372 : */
373 : #define ProcessMessageSubGroupMulti(group, subgroup, codeFragment) \
374 : do { \
375 : int n = NumMessagesInSubGroup(group, subgroup); \
376 : if (n > 0) { \
377 : SharedInvalidationMessage *msgs = \
378 : &InvalMessageArrays[subgroup].msgs[(group)->firstmsg[subgroup]]; \
379 : codeFragment; \
380 : } \
381 : } while (0)
382 :
383 :
384 : /* ----------------------------------------------------------------
385 : * Invalidation group support functions
386 : *
387 : * These routines understand about the division of a logical invalidation
388 : * group into separate physical arrays for catcache and relcache entries.
389 : * ----------------------------------------------------------------
390 : */
391 :
392 : /*
393 : * Add a catcache inval entry
394 : */
395 : static void
396 4264430 : AddCatcacheInvalidationMessage(InvalidationMsgsGroup *group,
397 : int id, uint32 hashValue, Oid dbId)
398 : {
399 : SharedInvalidationMessage msg;
400 :
401 : Assert(id < CHAR_MAX);
402 4264430 : msg.cc.id = (int8) id;
403 4264430 : msg.cc.dbId = dbId;
404 4264430 : msg.cc.hashValue = hashValue;
405 :
406 : /*
407 : * Define padding bytes in SharedInvalidationMessage structs to be
408 : * defined. Otherwise the sinvaladt.c ringbuffer, which is accessed by
409 : * multiple processes, will cause spurious valgrind warnings about
410 : * undefined memory being used. That's because valgrind remembers the
411 : * undefined bytes from the last local process's store, not realizing that
412 : * another process has written since, filling the previously uninitialized
413 : * bytes
414 : */
415 : VALGRIND_MAKE_MEM_DEFINED(&msg, sizeof(msg));
416 :
417 4264430 : AddInvalidationMessage(group, CatCacheMsgs, &msg);
418 4264430 : }
419 :
420 : /*
421 : * Add a whole-catalog inval entry
422 : */
423 : static void
424 202 : AddCatalogInvalidationMessage(InvalidationMsgsGroup *group,
425 : Oid dbId, Oid catId)
426 : {
427 : SharedInvalidationMessage msg;
428 :
429 202 : msg.cat.id = SHAREDINVALCATALOG_ID;
430 202 : msg.cat.dbId = dbId;
431 202 : msg.cat.catId = catId;
432 : /* check AddCatcacheInvalidationMessage() for an explanation */
433 : VALGRIND_MAKE_MEM_DEFINED(&msg, sizeof(msg));
434 :
435 202 : AddInvalidationMessage(group, CatCacheMsgs, &msg);
436 202 : }
437 :
438 : /*
439 : * Add a relcache inval entry
440 : */
441 : static void
442 1462626 : AddRelcacheInvalidationMessage(InvalidationMsgsGroup *group,
443 : Oid dbId, Oid relId)
444 : {
445 : SharedInvalidationMessage msg;
446 :
447 : /*
448 : * Don't add a duplicate item. We assume dbId need not be checked because
449 : * it will never change. InvalidOid for relId means all relations so we
450 : * don't need to add individual ones when it is present.
451 : */
452 4876186 : ProcessMessageSubGroup(group, RelCacheMsgs,
453 : if (msg->rc.id == SHAREDINVALRELCACHE_ID &&
454 : (msg->rc.relId == relId ||
455 : msg->rc.relId == InvalidOid))
456 : return);
457 :
458 : /* OK, add the item */
459 483792 : msg.rc.id = SHAREDINVALRELCACHE_ID;
460 483792 : msg.rc.dbId = dbId;
461 483792 : msg.rc.relId = relId;
462 : /* check AddCatcacheInvalidationMessage() for an explanation */
463 : VALGRIND_MAKE_MEM_DEFINED(&msg, sizeof(msg));
464 :
465 483792 : AddInvalidationMessage(group, RelCacheMsgs, &msg);
466 : }
467 :
468 : /*
469 : * Add a snapshot inval entry
470 : *
471 : * We put these into the relcache subgroup for simplicity.
472 : */
473 : static void
474 801092 : AddSnapshotInvalidationMessage(InvalidationMsgsGroup *group,
475 : Oid dbId, Oid relId)
476 : {
477 : SharedInvalidationMessage msg;
478 :
479 : /* Don't add a duplicate item */
480 : /* We assume dbId need not be checked because it will never change */
481 1203946 : ProcessMessageSubGroup(group, RelCacheMsgs,
482 : if (msg->sn.id == SHAREDINVALSNAPSHOT_ID &&
483 : msg->sn.relId == relId)
484 : return);
485 :
486 : /* OK, add the item */
487 406804 : msg.sn.id = SHAREDINVALSNAPSHOT_ID;
488 406804 : msg.sn.dbId = dbId;
489 406804 : msg.sn.relId = relId;
490 : /* check AddCatcacheInvalidationMessage() for an explanation */
491 : VALGRIND_MAKE_MEM_DEFINED(&msg, sizeof(msg));
492 :
493 406804 : AddInvalidationMessage(group, RelCacheMsgs, &msg);
494 : }
495 :
496 : /*
497 : * Append one group of invalidation messages to another, resetting
498 : * the source group to empty.
499 : */
500 : static void
501 723580 : AppendInvalidationMessages(InvalidationMsgsGroup *dest,
502 : InvalidationMsgsGroup *src)
503 : {
504 723580 : AppendInvalidationMessageSubGroup(dest, src, CatCacheMsgs);
505 723580 : AppendInvalidationMessageSubGroup(dest, src, RelCacheMsgs);
506 723580 : }
507 :
508 : /*
509 : * Execute the given function for all the messages in an invalidation group.
510 : * The group is not altered.
511 : *
512 : * catcache entries are processed first, for reasons mentioned above.
513 : */
514 : static void
515 562080 : ProcessInvalidationMessages(InvalidationMsgsGroup *group,
516 : void (*func) (SharedInvalidationMessage *msg))
517 : {
518 4801240 : ProcessMessageSubGroup(group, CatCacheMsgs, func(msg));
519 1435902 : ProcessMessageSubGroup(group, RelCacheMsgs, func(msg));
520 562074 : }
521 :
522 : /*
523 : * As above, but the function is able to process an array of messages
524 : * rather than just one at a time.
525 : */
526 : static void
527 165320 : ProcessInvalidationMessagesMulti(InvalidationMsgsGroup *group,
528 : void (*func) (const SharedInvalidationMessage *msgs, int n))
529 : {
530 165320 : ProcessMessageSubGroupMulti(group, CatCacheMsgs, func(msgs, n));
531 165320 : ProcessMessageSubGroupMulti(group, RelCacheMsgs, func(msgs, n));
532 165320 : }
533 :
534 : /* ----------------------------------------------------------------
535 : * private support functions
536 : * ----------------------------------------------------------------
537 : */
538 :
539 : /*
540 : * RegisterCatcacheInvalidation
541 : *
542 : * Register an invalidation event for a catcache tuple entry.
543 : */
544 : static void
545 4264430 : RegisterCatcacheInvalidation(int cacheId,
546 : uint32 hashValue,
547 : Oid dbId)
548 : {
549 4264430 : AddCatcacheInvalidationMessage(&transInvalInfo->CurrentCmdInvalidMsgs,
550 : cacheId, hashValue, dbId);
551 4264430 : }
552 :
553 : /*
554 : * RegisterCatalogInvalidation
555 : *
556 : * Register an invalidation event for all catcache entries from a catalog.
557 : */
558 : static void
559 202 : RegisterCatalogInvalidation(Oid dbId, Oid catId)
560 : {
561 202 : AddCatalogInvalidationMessage(&transInvalInfo->CurrentCmdInvalidMsgs,
562 : dbId, catId);
563 202 : }
564 :
565 : /*
566 : * RegisterRelcacheInvalidation
567 : *
568 : * As above, but register a relcache invalidation event.
569 : */
570 : static void
571 1462626 : RegisterRelcacheInvalidation(Oid dbId, Oid relId)
572 : {
573 1462626 : AddRelcacheInvalidationMessage(&transInvalInfo->CurrentCmdInvalidMsgs,
574 : dbId, relId);
575 :
576 : /*
577 : * Most of the time, relcache invalidation is associated with system
578 : * catalog updates, but there are a few cases where it isn't. Quick hack
579 : * to ensure that the next CommandCounterIncrement() will think that we
580 : * need to do CommandEndInvalidationMessages().
581 : */
582 1462626 : (void) GetCurrentCommandId(true);
583 :
584 : /*
585 : * If the relation being invalidated is one of those cached in a relcache
586 : * init file, mark that we need to zap that file at commit. For simplicity
587 : * invalidations for a specific database always invalidate the shared file
588 : * as well. Also zap when we are invalidating whole relcache.
589 : */
590 1462626 : if (relId == InvalidOid || RelationIdIsInInitFile(relId))
591 33434 : transInvalInfo->RelcacheInitFileInval = true;
592 1462626 : }
593 :
594 : /*
595 : * RegisterSnapshotInvalidation
596 : *
597 : * Register an invalidation event for MVCC scans against a given catalog.
598 : * Only needed for catalogs that don't have catcaches.
599 : */
600 : static void
601 801092 : RegisterSnapshotInvalidation(Oid dbId, Oid relId)
602 : {
603 801092 : AddSnapshotInvalidationMessage(&transInvalInfo->CurrentCmdInvalidMsgs,
604 : dbId, relId);
605 801092 : }
606 :
607 : /*
608 : * PrepareInvalidationState
609 : * Initialize inval data for the current (sub)transaction.
610 : */
611 : static void
612 3321894 : PrepareInvalidationState(void)
613 : {
614 : TransInvalidationInfo *myInfo;
615 :
616 6474748 : if (transInvalInfo != NULL &&
617 3152854 : transInvalInfo->my_level == GetCurrentTransactionNestLevel())
618 3152712 : return;
619 :
620 : myInfo = (TransInvalidationInfo *)
621 169182 : MemoryContextAllocZero(TopTransactionContext,
622 : sizeof(TransInvalidationInfo));
623 169182 : myInfo->parent = transInvalInfo;
624 169182 : myInfo->my_level = GetCurrentTransactionNestLevel();
625 :
626 : /* Now, do we have a previous stack entry? */
627 169182 : if (transInvalInfo != NULL)
628 : {
629 : /* Yes; this one should be for a deeper nesting level. */
630 : Assert(myInfo->my_level > transInvalInfo->my_level);
631 :
632 : /*
633 : * The parent (sub)transaction must not have any current (i.e.,
634 : * not-yet-locally-processed) messages. If it did, we'd have a
635 : * semantic problem: the new subtransaction presumably ought not be
636 : * able to see those events yet, but since the CommandCounter is
637 : * linear, that can't work once the subtransaction advances the
638 : * counter. This is a convenient place to check for that, as well as
639 : * being important to keep management of the message arrays simple.
640 : */
641 142 : if (NumMessagesInGroup(&transInvalInfo->CurrentCmdInvalidMsgs) != 0)
642 0 : elog(ERROR, "cannot start a subtransaction when there are unprocessed inval messages");
643 :
644 : /*
645 : * MemoryContextAllocZero set firstmsg = nextmsg = 0 in each group,
646 : * which is fine for the first (sub)transaction, but otherwise we need
647 : * to update them to follow whatever is already in the arrays.
648 : */
649 142 : SetGroupToFollow(&myInfo->PriorCmdInvalidMsgs,
650 : &transInvalInfo->CurrentCmdInvalidMsgs);
651 142 : SetGroupToFollow(&myInfo->CurrentCmdInvalidMsgs,
652 : &myInfo->PriorCmdInvalidMsgs);
653 : }
654 : else
655 : {
656 : /*
657 : * Here, we need only clear any array pointers left over from a prior
658 : * transaction.
659 : */
660 169040 : InvalMessageArrays[CatCacheMsgs].msgs = NULL;
661 169040 : InvalMessageArrays[CatCacheMsgs].maxmsgs = 0;
662 169040 : InvalMessageArrays[RelCacheMsgs].msgs = NULL;
663 169040 : InvalMessageArrays[RelCacheMsgs].maxmsgs = 0;
664 : }
665 :
666 169182 : transInvalInfo = myInfo;
667 : }
668 :
669 : /* ----------------------------------------------------------------
670 : * public functions
671 : * ----------------------------------------------------------------
672 : */
673 :
674 : void
675 3862 : InvalidateSystemCachesExtended(bool debug_discard)
676 : {
677 : int i;
678 :
679 3862 : InvalidateCatalogSnapshot();
680 3862 : ResetCatalogCaches();
681 3862 : RelationCacheInvalidate(debug_discard); /* gets smgr and relmap too */
682 :
683 58254 : for (i = 0; i < syscache_callback_count; i++)
684 : {
685 54392 : struct SYSCACHECALLBACK *ccitem = syscache_callback_list + i;
686 :
687 54392 : ccitem->function(ccitem->arg, ccitem->id, 0);
688 : }
689 :
690 8804 : for (i = 0; i < relcache_callback_count; i++)
691 : {
692 4942 : struct RELCACHECALLBACK *ccitem = relcache_callback_list + i;
693 :
694 4942 : ccitem->function(ccitem->arg, InvalidOid);
695 : }
696 3862 : }
697 :
698 : /*
699 : * LocalExecuteInvalidationMessage
700 : *
701 : * Process a single invalidation message (which could be of any type).
702 : * Only the local caches are flushed; this does not transmit the message
703 : * to other backends.
704 : */
705 : void
706 28440218 : LocalExecuteInvalidationMessage(SharedInvalidationMessage *msg)
707 : {
708 28440218 : if (msg->id >= 0)
709 : {
710 23235984 : if (msg->cc.dbId == MyDatabaseId || msg->cc.dbId == InvalidOid)
711 : {
712 16450260 : InvalidateCatalogSnapshot();
713 :
714 16450260 : SysCacheInvalidate(msg->cc.id, msg->cc.hashValue);
715 :
716 16450260 : CallSyscacheCallbacks(msg->cc.id, msg->cc.hashValue);
717 : }
718 : }
719 5204234 : else if (msg->id == SHAREDINVALCATALOG_ID)
720 : {
721 772 : if (msg->cat.dbId == MyDatabaseId || msg->cat.dbId == InvalidOid)
722 : {
723 664 : InvalidateCatalogSnapshot();
724 :
725 664 : CatalogCacheFlushCatalog(msg->cat.catId);
726 :
727 : /* CatalogCacheFlushCatalog calls CallSyscacheCallbacks as needed */
728 : }
729 : }
730 5203462 : else if (msg->id == SHAREDINVALRELCACHE_ID)
731 : {
732 2621028 : if (msg->rc.dbId == MyDatabaseId || msg->rc.dbId == InvalidOid)
733 : {
734 : int i;
735 :
736 1891714 : if (msg->rc.relId == InvalidOid)
737 234 : RelationCacheInvalidate(false);
738 : else
739 1891480 : RelationCacheInvalidateEntry(msg->rc.relId);
740 :
741 5231882 : for (i = 0; i < relcache_callback_count; i++)
742 : {
743 3340174 : struct RELCACHECALLBACK *ccitem = relcache_callback_list + i;
744 :
745 3340174 : ccitem->function(ccitem->arg, msg->rc.relId);
746 : }
747 : }
748 : }
749 2582434 : else if (msg->id == SHAREDINVALSMGR_ID)
750 : {
751 : /*
752 : * We could have smgr entries for relations of other databases, so no
753 : * short-circuit test is possible here.
754 : */
755 : RelFileLocatorBackend rlocator;
756 :
757 360984 : rlocator.locator = msg->sm.rlocator;
758 360984 : rlocator.backend = (msg->sm.backend_hi << 16) | (int) msg->sm.backend_lo;
759 360984 : smgrcloserellocator(rlocator);
760 : }
761 2221450 : else if (msg->id == SHAREDINVALRELMAP_ID)
762 : {
763 : /* We only care about our own database and shared catalogs */
764 330 : if (msg->rm.dbId == InvalidOid)
765 164 : RelationMapInvalidate(true);
766 166 : else if (msg->rm.dbId == MyDatabaseId)
767 100 : RelationMapInvalidate(false);
768 : }
769 2221120 : else if (msg->id == SHAREDINVALSNAPSHOT_ID)
770 : {
771 : /* We only care about our own database and shared catalogs */
772 2221120 : if (msg->sn.dbId == InvalidOid)
773 74690 : InvalidateCatalogSnapshot();
774 2146430 : else if (msg->sn.dbId == MyDatabaseId)
775 1572326 : InvalidateCatalogSnapshot();
776 : }
777 : else
778 0 : elog(FATAL, "unrecognized SI message ID: %d", msg->id);
779 28440212 : }
780 :
781 : /*
782 : * InvalidateSystemCaches
783 : *
784 : * This blows away all tuples in the system catalog caches and
785 : * all the cached relation descriptors and smgr cache entries.
786 : * Relation descriptors that have positive refcounts are then rebuilt.
787 : *
788 : * We call this when we see a shared-inval-queue overflow signal,
789 : * since that tells us we've lost some shared-inval messages and hence
790 : * don't know what needs to be invalidated.
791 : */
792 : void
793 3862 : InvalidateSystemCaches(void)
794 : {
795 3862 : InvalidateSystemCachesExtended(false);
796 3862 : }
797 :
798 : /*
799 : * AcceptInvalidationMessages
800 : * Read and process invalidation messages from the shared invalidation
801 : * message queue.
802 : *
803 : * Note:
804 : * This should be called as the first step in processing a transaction.
805 : */
806 : void
807 24905844 : AcceptInvalidationMessages(void)
808 : {
809 24905844 : ReceiveSharedInvalidMessages(LocalExecuteInvalidationMessage,
810 : InvalidateSystemCaches);
811 :
812 : /*----------
813 : * Test code to force cache flushes anytime a flush could happen.
814 : *
815 : * This helps detect intermittent faults caused by code that reads a cache
816 : * entry and then performs an action that could invalidate the entry, but
817 : * rarely actually does so. This can spot issues that would otherwise
818 : * only arise with badly timed concurrent DDL, for example.
819 : *
820 : * The default debug_discard_caches = 0 does no forced cache flushes.
821 : *
822 : * If used with CLOBBER_FREED_MEMORY,
823 : * debug_discard_caches = 1 (formerly known as CLOBBER_CACHE_ALWAYS)
824 : * provides a fairly thorough test that the system contains no cache-flush
825 : * hazards. However, it also makes the system unbelievably slow --- the
826 : * regression tests take about 100 times longer than normal.
827 : *
828 : * If you're a glutton for punishment, try
829 : * debug_discard_caches = 3 (formerly known as CLOBBER_CACHE_RECURSIVELY).
830 : * This slows things by at least a factor of 10000, so I wouldn't suggest
831 : * trying to run the entire regression tests that way. It's useful to try
832 : * a few simple tests, to make sure that cache reload isn't subject to
833 : * internal cache-flush hazards, but after you've done a few thousand
834 : * recursive reloads it's unlikely you'll learn more.
835 : *----------
836 : */
837 : #ifdef DISCARD_CACHES_ENABLED
838 : {
839 : static int recursion_depth = 0;
840 :
841 : if (recursion_depth < debug_discard_caches)
842 : {
843 : recursion_depth++;
844 : InvalidateSystemCachesExtended(true);
845 : recursion_depth--;
846 : }
847 : }
848 : #endif
849 24905844 : }
850 :
851 : /*
852 : * PostPrepare_Inval
853 : * Clean up after successful PREPARE.
854 : *
855 : * Here, we want to act as though the transaction aborted, so that we will
856 : * undo any syscache changes it made, thereby bringing us into sync with the
857 : * outside world, which doesn't believe the transaction committed yet.
858 : *
859 : * If the prepared transaction is later aborted, there is nothing more to
860 : * do; if it commits, we will receive the consequent inval messages just
861 : * like everyone else.
862 : */
863 : void
864 748 : PostPrepare_Inval(void)
865 : {
866 748 : AtEOXact_Inval(false);
867 748 : }
868 :
869 : /*
870 : * xactGetCommittedInvalidationMessages() is called by
871 : * RecordTransactionCommit() to collect invalidation messages to add to the
872 : * commit record. This applies only to commit message types, never to
873 : * abort records. Must always run before AtEOXact_Inval(), since that
874 : * removes the data we need to see.
875 : *
876 : * Remember that this runs before we have officially committed, so we
877 : * must not do anything here to change what might occur *if* we should
878 : * fail between here and the actual commit.
879 : *
880 : * see also xact_redo_commit() and xact_desc_commit()
881 : */
882 : int
883 314816 : xactGetCommittedInvalidationMessages(SharedInvalidationMessage **msgs,
884 : bool *RelcacheInitFileInval)
885 : {
886 : SharedInvalidationMessage *msgarray;
887 : int nummsgs;
888 : int nmsgs;
889 :
890 : /* Quick exit if we haven't done anything with invalidation messages. */
891 314816 : if (transInvalInfo == NULL)
892 : {
893 191518 : *RelcacheInitFileInval = false;
894 191518 : *msgs = NULL;
895 191518 : return 0;
896 : }
897 :
898 : /* Must be at top of stack */
899 : Assert(transInvalInfo->my_level == 1 && transInvalInfo->parent == NULL);
900 :
901 : /*
902 : * Relcache init file invalidation requires processing both before and
903 : * after we send the SI messages. However, we need not do anything unless
904 : * we committed.
905 : */
906 123298 : *RelcacheInitFileInval = transInvalInfo->RelcacheInitFileInval;
907 :
908 : /*
909 : * Collect all the pending messages into a single contiguous array of
910 : * invalidation messages, to simplify what needs to happen while building
911 : * the commit WAL message. Maintain the order that they would be
912 : * processed in by AtEOXact_Inval(), to ensure emulated behaviour in redo
913 : * is as similar as possible to original. We want the same bugs, if any,
914 : * not new ones.
915 : */
916 123298 : nummsgs = NumMessagesInGroup(&transInvalInfo->PriorCmdInvalidMsgs) +
917 123298 : NumMessagesInGroup(&transInvalInfo->CurrentCmdInvalidMsgs);
918 :
919 123298 : *msgs = msgarray = (SharedInvalidationMessage *)
920 123298 : MemoryContextAlloc(CurTransactionContext,
921 : nummsgs * sizeof(SharedInvalidationMessage));
922 :
923 123298 : nmsgs = 0;
924 123298 : ProcessMessageSubGroupMulti(&transInvalInfo->PriorCmdInvalidMsgs,
925 : CatCacheMsgs,
926 : (memcpy(msgarray + nmsgs,
927 : msgs,
928 : n * sizeof(SharedInvalidationMessage)),
929 : nmsgs += n));
930 123298 : ProcessMessageSubGroupMulti(&transInvalInfo->CurrentCmdInvalidMsgs,
931 : CatCacheMsgs,
932 : (memcpy(msgarray + nmsgs,
933 : msgs,
934 : n * sizeof(SharedInvalidationMessage)),
935 : nmsgs += n));
936 123298 : ProcessMessageSubGroupMulti(&transInvalInfo->PriorCmdInvalidMsgs,
937 : RelCacheMsgs,
938 : (memcpy(msgarray + nmsgs,
939 : msgs,
940 : n * sizeof(SharedInvalidationMessage)),
941 : nmsgs += n));
942 123298 : ProcessMessageSubGroupMulti(&transInvalInfo->CurrentCmdInvalidMsgs,
943 : RelCacheMsgs,
944 : (memcpy(msgarray + nmsgs,
945 : msgs,
946 : n * sizeof(SharedInvalidationMessage)),
947 : nmsgs += n));
948 : Assert(nmsgs == nummsgs);
949 :
950 123298 : return nmsgs;
951 : }
952 :
953 : /*
954 : * ProcessCommittedInvalidationMessages is executed by xact_redo_commit() or
955 : * standby_redo() to process invalidation messages. Currently that happens
956 : * only at end-of-xact.
957 : *
958 : * Relcache init file invalidation requires processing both
959 : * before and after we send the SI messages. See AtEOXact_Inval()
960 : */
961 : void
962 35056 : ProcessCommittedInvalidationMessages(SharedInvalidationMessage *msgs,
963 : int nmsgs, bool RelcacheInitFileInval,
964 : Oid dbid, Oid tsid)
965 : {
966 35056 : if (nmsgs <= 0)
967 8902 : return;
968 :
969 26154 : elog(DEBUG4, "replaying commit with %d messages%s", nmsgs,
970 : (RelcacheInitFileInval ? " and relcache file invalidation" : ""));
971 :
972 26154 : if (RelcacheInitFileInval)
973 : {
974 240 : elog(DEBUG4, "removing relcache init files for database %u", dbid);
975 :
976 : /*
977 : * RelationCacheInitFilePreInvalidate, when the invalidation message
978 : * is for a specific database, requires DatabasePath to be set, but we
979 : * should not use SetDatabasePath during recovery, since it is
980 : * intended to be used only once by normal backends. Hence, a quick
981 : * hack: set DatabasePath directly then unset after use.
982 : */
983 240 : if (OidIsValid(dbid))
984 240 : DatabasePath = GetDatabasePath(dbid, tsid);
985 :
986 240 : RelationCacheInitFilePreInvalidate();
987 :
988 240 : if (OidIsValid(dbid))
989 : {
990 240 : pfree(DatabasePath);
991 240 : DatabasePath = NULL;
992 : }
993 : }
994 :
995 26154 : SendSharedInvalidMessages(msgs, nmsgs);
996 :
997 26154 : if (RelcacheInitFileInval)
998 240 : RelationCacheInitFilePostInvalidate();
999 : }
1000 :
1001 : /*
1002 : * AtEOXact_Inval
1003 : * Process queued-up invalidation messages at end of main transaction.
1004 : *
1005 : * If isCommit, we must send out the messages in our PriorCmdInvalidMsgs list
1006 : * to the shared invalidation message queue. Note that these will be read
1007 : * not only by other backends, but also by our own backend at the next
1008 : * transaction start (via AcceptInvalidationMessages). This means that
1009 : * we can skip immediate local processing of anything that's still in
1010 : * CurrentCmdInvalidMsgs, and just send that list out too.
1011 : *
1012 : * If not isCommit, we are aborting, and must locally process the messages
1013 : * in PriorCmdInvalidMsgs. No messages need be sent to other backends,
1014 : * since they'll not have seen our changed tuples anyway. We can forget
1015 : * about CurrentCmdInvalidMsgs too, since those changes haven't touched
1016 : * the caches yet.
1017 : *
1018 : * In any case, reset our state to empty. We need not physically
1019 : * free memory here, since TopTransactionContext is about to be emptied
1020 : * anyway.
1021 : *
1022 : * Note:
1023 : * This should be called as the last step in processing a transaction.
1024 : */
1025 : void
1026 516616 : AtEOXact_Inval(bool isCommit)
1027 : {
1028 : /* Quick exit if no messages */
1029 516616 : if (transInvalInfo == NULL)
1030 347640 : return;
1031 :
1032 : /* Must be at top of stack */
1033 : Assert(transInvalInfo->my_level == 1 && transInvalInfo->parent == NULL);
1034 :
1035 168976 : if (isCommit)
1036 : {
1037 : /*
1038 : * Relcache init file invalidation requires processing both before and
1039 : * after we send the SI messages. However, we need not do anything
1040 : * unless we committed.
1041 : */
1042 165320 : if (transInvalInfo->RelcacheInitFileInval)
1043 11796 : RelationCacheInitFilePreInvalidate();
1044 :
1045 165320 : AppendInvalidationMessages(&transInvalInfo->PriorCmdInvalidMsgs,
1046 165320 : &transInvalInfo->CurrentCmdInvalidMsgs);
1047 :
1048 165320 : ProcessInvalidationMessagesMulti(&transInvalInfo->PriorCmdInvalidMsgs,
1049 : SendSharedInvalidMessages);
1050 :
1051 165320 : if (transInvalInfo->RelcacheInitFileInval)
1052 11796 : RelationCacheInitFilePostInvalidate();
1053 : }
1054 : else
1055 : {
1056 3656 : ProcessInvalidationMessages(&transInvalInfo->PriorCmdInvalidMsgs,
1057 : LocalExecuteInvalidationMessage);
1058 : }
1059 :
1060 : /* Need not free anything explicitly */
1061 168976 : transInvalInfo = NULL;
1062 : }
1063 :
1064 : /*
1065 : * AtEOSubXact_Inval
1066 : * Process queued-up invalidation messages at end of subtransaction.
1067 : *
1068 : * If isCommit, process CurrentCmdInvalidMsgs if any (there probably aren't),
1069 : * and then attach both CurrentCmdInvalidMsgs and PriorCmdInvalidMsgs to the
1070 : * parent's PriorCmdInvalidMsgs list.
1071 : *
1072 : * If not isCommit, we are aborting, and must locally process the messages
1073 : * in PriorCmdInvalidMsgs. No messages need be sent to other backends.
1074 : * We can forget about CurrentCmdInvalidMsgs too, since those changes haven't
1075 : * touched the caches yet.
1076 : *
1077 : * In any case, pop the transaction stack. We need not physically free memory
1078 : * here, since CurTransactionContext is about to be emptied anyway
1079 : * (if aborting). Beware of the possibility of aborting the same nesting
1080 : * level twice, though.
1081 : */
1082 : void
1083 18048 : AtEOSubXact_Inval(bool isCommit)
1084 : {
1085 : int my_level;
1086 18048 : TransInvalidationInfo *myInfo = transInvalInfo;
1087 :
1088 : /* Quick exit if no messages. */
1089 18048 : if (myInfo == NULL)
1090 16480 : return;
1091 :
1092 : /* Also bail out quickly if messages are not for this level. */
1093 1568 : my_level = GetCurrentTransactionNestLevel();
1094 1568 : if (myInfo->my_level != my_level)
1095 : {
1096 : Assert(myInfo->my_level < my_level);
1097 1294 : return;
1098 : }
1099 :
1100 274 : if (isCommit)
1101 : {
1102 : /* If CurrentCmdInvalidMsgs still has anything, fix it */
1103 92 : CommandEndInvalidationMessages();
1104 :
1105 : /*
1106 : * We create invalidation stack entries lazily, so the parent might
1107 : * not have one. Instead of creating one, moving all the data over,
1108 : * and then freeing our own, we can just adjust the level of our own
1109 : * entry.
1110 : */
1111 92 : if (myInfo->parent == NULL || myInfo->parent->my_level < my_level - 1)
1112 : {
1113 68 : myInfo->my_level--;
1114 68 : return;
1115 : }
1116 :
1117 : /*
1118 : * Pass up my inval messages to parent. Notice that we stick them in
1119 : * PriorCmdInvalidMsgs, not CurrentCmdInvalidMsgs, since they've
1120 : * already been locally processed. (This would trigger the Assert in
1121 : * AppendInvalidationMessageSubGroup if the parent's
1122 : * CurrentCmdInvalidMsgs isn't empty; but we already checked that in
1123 : * PrepareInvalidationState.)
1124 : */
1125 24 : AppendInvalidationMessages(&myInfo->parent->PriorCmdInvalidMsgs,
1126 : &myInfo->PriorCmdInvalidMsgs);
1127 :
1128 : /* Must readjust parent's CurrentCmdInvalidMsgs indexes now */
1129 24 : SetGroupToFollow(&myInfo->parent->CurrentCmdInvalidMsgs,
1130 : &myInfo->parent->PriorCmdInvalidMsgs);
1131 :
1132 : /* Pending relcache inval becomes parent's problem too */
1133 24 : if (myInfo->RelcacheInitFileInval)
1134 0 : myInfo->parent->RelcacheInitFileInval = true;
1135 :
1136 : /* Pop the transaction state stack */
1137 24 : transInvalInfo = myInfo->parent;
1138 :
1139 : /* Need not free anything else explicitly */
1140 24 : pfree(myInfo);
1141 : }
1142 : else
1143 : {
1144 182 : ProcessInvalidationMessages(&myInfo->PriorCmdInvalidMsgs,
1145 : LocalExecuteInvalidationMessage);
1146 :
1147 : /* Pop the transaction state stack */
1148 182 : transInvalInfo = myInfo->parent;
1149 :
1150 : /* Need not free anything else explicitly */
1151 182 : pfree(myInfo);
1152 : }
1153 : }
1154 :
1155 : /*
1156 : * CommandEndInvalidationMessages
1157 : * Process queued-up invalidation messages at end of one command
1158 : * in a transaction.
1159 : *
1160 : * Here, we send no messages to the shared queue, since we don't know yet if
1161 : * we will commit. We do need to locally process the CurrentCmdInvalidMsgs
1162 : * list, so as to flush our caches of any entries we have outdated in the
1163 : * current command. We then move the current-cmd list over to become part
1164 : * of the prior-cmds list.
1165 : *
1166 : * Note:
1167 : * This should be called during CommandCounterIncrement(),
1168 : * after we have advanced the command ID.
1169 : */
1170 : void
1171 917932 : CommandEndInvalidationMessages(void)
1172 : {
1173 : /*
1174 : * You might think this shouldn't be called outside any transaction, but
1175 : * bootstrap does it, and also ABORT issued when not in a transaction. So
1176 : * just quietly return if no state to work on.
1177 : */
1178 917932 : if (transInvalInfo == NULL)
1179 359690 : return;
1180 :
1181 558242 : ProcessInvalidationMessages(&transInvalInfo->CurrentCmdInvalidMsgs,
1182 : LocalExecuteInvalidationMessage);
1183 :
1184 : /* WAL Log per-command invalidation messages for wal_level=logical */
1185 558236 : if (XLogLogicalInfoActive())
1186 7114 : LogLogicalInvalidations();
1187 :
1188 558236 : AppendInvalidationMessages(&transInvalInfo->PriorCmdInvalidMsgs,
1189 558236 : &transInvalInfo->CurrentCmdInvalidMsgs);
1190 : }
1191 :
1192 :
1193 : /*
1194 : * CacheInvalidateHeapTuple
1195 : * Register the given tuple for invalidation at end of command
1196 : * (ie, current command is creating or outdating this tuple).
1197 : * Also, detect whether a relcache invalidation is implied.
1198 : *
1199 : * For an insert or delete, tuple is the target tuple and newtuple is NULL.
1200 : * For an update, we are called just once, with tuple being the old tuple
1201 : * version and newtuple the new version. This allows avoidance of duplicate
1202 : * effort during an update.
1203 : */
1204 : void
1205 19775050 : CacheInvalidateHeapTuple(Relation relation,
1206 : HeapTuple tuple,
1207 : HeapTuple newtuple)
1208 : {
1209 : Oid tupleRelId;
1210 : Oid databaseId;
1211 : Oid relationId;
1212 :
1213 : /* Do nothing during bootstrap */
1214 19775050 : if (IsBootstrapProcessingMode())
1215 801856 : return;
1216 :
1217 : /*
1218 : * We only need to worry about invalidation for tuples that are in system
1219 : * catalogs; user-relation tuples are never in catcaches and can't affect
1220 : * the relcache either.
1221 : */
1222 18973194 : if (!IsCatalogRelation(relation))
1223 15814760 : return;
1224 :
1225 : /*
1226 : * IsCatalogRelation() will return true for TOAST tables of system
1227 : * catalogs, but we don't care about those, either.
1228 : */
1229 3158434 : if (IsToastRelation(relation))
1230 21140 : return;
1231 :
1232 : /*
1233 : * If we're not prepared to queue invalidation messages for this
1234 : * subtransaction level, get ready now.
1235 : */
1236 3137294 : PrepareInvalidationState();
1237 :
1238 : /*
1239 : * First let the catcache do its thing
1240 : */
1241 3137294 : tupleRelId = RelationGetRelid(relation);
1242 3137294 : if (RelationInvalidatesSnapshotsOnly(tupleRelId))
1243 : {
1244 801092 : databaseId = IsSharedRelation(tupleRelId) ? InvalidOid : MyDatabaseId;
1245 801092 : RegisterSnapshotInvalidation(databaseId, tupleRelId);
1246 : }
1247 : else
1248 2336202 : PrepareToInvalidateCacheTuple(relation, tuple, newtuple,
1249 : RegisterCatcacheInvalidation);
1250 :
1251 : /*
1252 : * Now, is this tuple one of the primary definers of a relcache entry? See
1253 : * comments in file header for deeper explanation.
1254 : *
1255 : * Note we ignore newtuple here; we assume an update cannot move a tuple
1256 : * from being part of one relcache entry to being part of another.
1257 : */
1258 3137294 : if (tupleRelId == RelationRelationId)
1259 : {
1260 311860 : Form_pg_class classtup = (Form_pg_class) GETSTRUCT(tuple);
1261 :
1262 311860 : relationId = classtup->oid;
1263 311860 : if (classtup->relisshared)
1264 4424 : databaseId = InvalidOid;
1265 : else
1266 307436 : databaseId = MyDatabaseId;
1267 : }
1268 2825434 : else if (tupleRelId == AttributeRelationId)
1269 : {
1270 931362 : Form_pg_attribute atttup = (Form_pg_attribute) GETSTRUCT(tuple);
1271 :
1272 931362 : relationId = atttup->attrelid;
1273 :
1274 : /*
1275 : * KLUGE ALERT: we always send the relcache event with MyDatabaseId,
1276 : * even if the rel in question is shared (which we can't easily tell).
1277 : * This essentially means that only backends in this same database
1278 : * will react to the relcache flush request. This is in fact
1279 : * appropriate, since only those backends could see our pg_attribute
1280 : * change anyway. It looks a bit ugly though. (In practice, shared
1281 : * relations can't have schema changes after bootstrap, so we should
1282 : * never come here for a shared rel anyway.)
1283 : */
1284 931362 : databaseId = MyDatabaseId;
1285 : }
1286 1894072 : else if (tupleRelId == IndexRelationId)
1287 : {
1288 55566 : Form_pg_index indextup = (Form_pg_index) GETSTRUCT(tuple);
1289 :
1290 : /*
1291 : * When a pg_index row is updated, we should send out a relcache inval
1292 : * for the index relation. As above, we don't know the shared status
1293 : * of the index, but in practice it doesn't matter since indexes of
1294 : * shared catalogs can't have such updates.
1295 : */
1296 55566 : relationId = indextup->indexrelid;
1297 55566 : databaseId = MyDatabaseId;
1298 : }
1299 1838506 : else if (tupleRelId == ConstraintRelationId)
1300 : {
1301 49592 : Form_pg_constraint constrtup = (Form_pg_constraint) GETSTRUCT(tuple);
1302 :
1303 : /*
1304 : * Foreign keys are part of relcache entries, too, so send out an
1305 : * inval for the table that the FK applies to.
1306 : */
1307 49592 : if (constrtup->contype == CONSTRAINT_FOREIGN &&
1308 6292 : OidIsValid(constrtup->conrelid))
1309 : {
1310 6292 : relationId = constrtup->conrelid;
1311 6292 : databaseId = MyDatabaseId;
1312 : }
1313 : else
1314 43300 : return;
1315 : }
1316 : else
1317 1788914 : return;
1318 :
1319 : /*
1320 : * Yes. We need to register a relcache invalidation event.
1321 : */
1322 1305080 : RegisterRelcacheInvalidation(databaseId, relationId);
1323 : }
1324 :
1325 : /*
1326 : * CacheInvalidateCatalog
1327 : * Register invalidation of the whole content of a system catalog.
1328 : *
1329 : * This is normally used in VACUUM FULL/CLUSTER, where we haven't so much
1330 : * changed any tuples as moved them around. Some uses of catcache entries
1331 : * expect their TIDs to be correct, so we have to blow away the entries.
1332 : *
1333 : * Note: we expect caller to verify that the rel actually is a system
1334 : * catalog. If it isn't, no great harm is done, just a wasted sinval message.
1335 : */
1336 : void
1337 202 : CacheInvalidateCatalog(Oid catalogId)
1338 : {
1339 : Oid databaseId;
1340 :
1341 202 : PrepareInvalidationState();
1342 :
1343 202 : if (IsSharedRelation(catalogId))
1344 36 : databaseId = InvalidOid;
1345 : else
1346 166 : databaseId = MyDatabaseId;
1347 :
1348 202 : RegisterCatalogInvalidation(databaseId, catalogId);
1349 202 : }
1350 :
1351 : /*
1352 : * CacheInvalidateRelcache
1353 : * Register invalidation of the specified relation's relcache entry
1354 : * at end of command.
1355 : *
1356 : * This is used in places that need to force relcache rebuild but aren't
1357 : * changing any of the tuples recognized as contributors to the relcache
1358 : * entry by CacheInvalidateHeapTuple. (An example is dropping an index.)
1359 : */
1360 : void
1361 98478 : CacheInvalidateRelcache(Relation relation)
1362 : {
1363 : Oid databaseId;
1364 : Oid relationId;
1365 :
1366 98478 : PrepareInvalidationState();
1367 :
1368 98478 : relationId = RelationGetRelid(relation);
1369 98478 : if (relation->rd_rel->relisshared)
1370 2772 : databaseId = InvalidOid;
1371 : else
1372 95706 : databaseId = MyDatabaseId;
1373 :
1374 98478 : RegisterRelcacheInvalidation(databaseId, relationId);
1375 98478 : }
1376 :
1377 : /*
1378 : * CacheInvalidateRelcacheAll
1379 : * Register invalidation of the whole relcache at the end of command.
1380 : *
1381 : * This is used by alter publication as changes in publications may affect
1382 : * large number of tables.
1383 : */
1384 : void
1385 82 : CacheInvalidateRelcacheAll(void)
1386 : {
1387 82 : PrepareInvalidationState();
1388 :
1389 82 : RegisterRelcacheInvalidation(InvalidOid, InvalidOid);
1390 82 : }
1391 :
1392 : /*
1393 : * CacheInvalidateRelcacheByTuple
1394 : * As above, but relation is identified by passing its pg_class tuple.
1395 : */
1396 : void
1397 58986 : CacheInvalidateRelcacheByTuple(HeapTuple classTuple)
1398 : {
1399 58986 : Form_pg_class classtup = (Form_pg_class) GETSTRUCT(classTuple);
1400 : Oid databaseId;
1401 : Oid relationId;
1402 :
1403 58986 : PrepareInvalidationState();
1404 :
1405 58986 : relationId = classtup->oid;
1406 58986 : if (classtup->relisshared)
1407 1082 : databaseId = InvalidOid;
1408 : else
1409 57904 : databaseId = MyDatabaseId;
1410 58986 : RegisterRelcacheInvalidation(databaseId, relationId);
1411 58986 : }
1412 :
1413 : /*
1414 : * CacheInvalidateRelcacheByRelid
1415 : * As above, but relation is identified by passing its OID.
1416 : * This is the least efficient of the three options; use one of
1417 : * the above routines if you have a Relation or pg_class tuple.
1418 : */
1419 : void
1420 26852 : CacheInvalidateRelcacheByRelid(Oid relid)
1421 : {
1422 : HeapTuple tup;
1423 :
1424 26852 : PrepareInvalidationState();
1425 :
1426 26852 : tup = SearchSysCache1(RELOID, ObjectIdGetDatum(relid));
1427 26852 : if (!HeapTupleIsValid(tup))
1428 0 : elog(ERROR, "cache lookup failed for relation %u", relid);
1429 26852 : CacheInvalidateRelcacheByTuple(tup);
1430 26852 : ReleaseSysCache(tup);
1431 26852 : }
1432 :
1433 :
1434 : /*
1435 : * CacheInvalidateSmgr
1436 : * Register invalidation of smgr references to a physical relation.
1437 : *
1438 : * Sending this type of invalidation msg forces other backends to close open
1439 : * smgr entries for the rel. This should be done to flush dangling open-file
1440 : * references when the physical rel is being dropped or truncated. Because
1441 : * these are nontransactional (i.e., not-rollback-able) operations, we just
1442 : * send the inval message immediately without any queuing.
1443 : *
1444 : * Note: in most cases there will have been a relcache flush issued against
1445 : * the rel at the logical level. We need a separate smgr-level flush because
1446 : * it is possible for backends to have open smgr entries for rels they don't
1447 : * have a relcache entry for, e.g. because the only thing they ever did with
1448 : * the rel is write out dirty shared buffers.
1449 : *
1450 : * Note: because these messages are nontransactional, they won't be captured
1451 : * in commit/abort WAL entries. Instead, calls to CacheInvalidateSmgr()
1452 : * should happen in low-level smgr.c routines, which are executed while
1453 : * replaying WAL as well as when creating it.
1454 : *
1455 : * Note: In order to avoid bloating SharedInvalidationMessage, we store only
1456 : * three bytes of the backend ID using what would otherwise be padding space.
1457 : * Thus, the maximum possible backend ID is 2^23-1.
1458 : */
1459 : void
1460 80842 : CacheInvalidateSmgr(RelFileLocatorBackend rlocator)
1461 : {
1462 : SharedInvalidationMessage msg;
1463 :
1464 80842 : msg.sm.id = SHAREDINVALSMGR_ID;
1465 80842 : msg.sm.backend_hi = rlocator.backend >> 16;
1466 80842 : msg.sm.backend_lo = rlocator.backend & 0xffff;
1467 80842 : msg.sm.rlocator = rlocator.locator;
1468 : /* check AddCatcacheInvalidationMessage() for an explanation */
1469 : VALGRIND_MAKE_MEM_DEFINED(&msg, sizeof(msg));
1470 :
1471 80842 : SendSharedInvalidMessages(&msg, 1);
1472 80842 : }
1473 :
1474 : /*
1475 : * CacheInvalidateRelmap
1476 : * Register invalidation of the relation mapping for a database,
1477 : * or for the shared catalogs if databaseId is zero.
1478 : *
1479 : * Sending this type of invalidation msg forces other backends to re-read
1480 : * the indicated relation mapping file. It is also necessary to send a
1481 : * relcache inval for the specific relations whose mapping has been altered,
1482 : * else the relcache won't get updated with the new filenode data.
1483 : *
1484 : * Note: because these messages are nontransactional, they won't be captured
1485 : * in commit/abort WAL entries. Instead, calls to CacheInvalidateRelmap()
1486 : * should happen in low-level relmapper.c routines, which are executed while
1487 : * replaying WAL as well as when creating it.
1488 : */
1489 : void
1490 208 : CacheInvalidateRelmap(Oid databaseId)
1491 : {
1492 : SharedInvalidationMessage msg;
1493 :
1494 208 : msg.rm.id = SHAREDINVALRELMAP_ID;
1495 208 : msg.rm.dbId = databaseId;
1496 : /* check AddCatcacheInvalidationMessage() for an explanation */
1497 : VALGRIND_MAKE_MEM_DEFINED(&msg, sizeof(msg));
1498 :
1499 208 : SendSharedInvalidMessages(&msg, 1);
1500 208 : }
1501 :
1502 :
1503 : /*
1504 : * CacheRegisterSyscacheCallback
1505 : * Register the specified function to be called for all future
1506 : * invalidation events in the specified cache. The cache ID and the
1507 : * hash value of the tuple being invalidated will be passed to the
1508 : * function.
1509 : *
1510 : * NOTE: Hash value zero will be passed if a cache reset request is received.
1511 : * In this case the called routines should flush all cached state.
1512 : * Yes, there's a possibility of a false match to zero, but it doesn't seem
1513 : * worth troubling over, especially since most of the current callees just
1514 : * flush all cached state anyway.
1515 : */
1516 : void
1517 355860 : CacheRegisterSyscacheCallback(int cacheid,
1518 : SyscacheCallbackFunction func,
1519 : Datum arg)
1520 : {
1521 355860 : if (cacheid < 0 || cacheid >= SysCacheSize)
1522 0 : elog(FATAL, "invalid cache ID: %d", cacheid);
1523 355860 : if (syscache_callback_count >= MAX_SYSCACHE_CALLBACKS)
1524 0 : elog(FATAL, "out of syscache_callback_list slots");
1525 :
1526 355860 : if (syscache_callback_links[cacheid] == 0)
1527 : {
1528 : /* first callback for this cache */
1529 282038 : syscache_callback_links[cacheid] = syscache_callback_count + 1;
1530 : }
1531 : else
1532 : {
1533 : /* add to end of chain, so that older callbacks are called first */
1534 73822 : int i = syscache_callback_links[cacheid] - 1;
1535 :
1536 97830 : while (syscache_callback_list[i].link > 0)
1537 24008 : i = syscache_callback_list[i].link - 1;
1538 73822 : syscache_callback_list[i].link = syscache_callback_count + 1;
1539 : }
1540 :
1541 355860 : syscache_callback_list[syscache_callback_count].id = cacheid;
1542 355860 : syscache_callback_list[syscache_callback_count].link = 0;
1543 355860 : syscache_callback_list[syscache_callback_count].function = func;
1544 355860 : syscache_callback_list[syscache_callback_count].arg = arg;
1545 :
1546 355860 : ++syscache_callback_count;
1547 355860 : }
1548 :
1549 : /*
1550 : * CacheRegisterRelcacheCallback
1551 : * Register the specified function to be called for all future
1552 : * relcache invalidation events. The OID of the relation being
1553 : * invalidated will be passed to the function.
1554 : *
1555 : * NOTE: InvalidOid will be passed if a cache reset request is received.
1556 : * In this case the called routines should flush all cached state.
1557 : */
1558 : void
1559 31692 : CacheRegisterRelcacheCallback(RelcacheCallbackFunction func,
1560 : Datum arg)
1561 : {
1562 31692 : if (relcache_callback_count >= MAX_RELCACHE_CALLBACKS)
1563 0 : elog(FATAL, "out of relcache_callback_list slots");
1564 :
1565 31692 : relcache_callback_list[relcache_callback_count].function = func;
1566 31692 : relcache_callback_list[relcache_callback_count].arg = arg;
1567 :
1568 31692 : ++relcache_callback_count;
1569 31692 : }
1570 :
1571 : /*
1572 : * CallSyscacheCallbacks
1573 : *
1574 : * This is exported so that CatalogCacheFlushCatalog can call it, saving
1575 : * this module from knowing which catcache IDs correspond to which catalogs.
1576 : */
1577 : void
1578 16451100 : CallSyscacheCallbacks(int cacheid, uint32 hashvalue)
1579 : {
1580 : int i;
1581 :
1582 16451100 : if (cacheid < 0 || cacheid >= SysCacheSize)
1583 0 : elog(ERROR, "invalid cache ID: %d", cacheid);
1584 :
1585 16451100 : i = syscache_callback_links[cacheid] - 1;
1586 18738356 : while (i >= 0)
1587 : {
1588 2287256 : struct SYSCACHECALLBACK *ccitem = syscache_callback_list + i;
1589 :
1590 : Assert(ccitem->id == cacheid);
1591 2287256 : ccitem->function(ccitem->arg, cacheid, hashvalue);
1592 2287256 : i = ccitem->link - 1;
1593 : }
1594 16451100 : }
1595 :
1596 : /*
1597 : * LogLogicalInvalidations
1598 : *
1599 : * Emit WAL for invalidations caused by the current command.
1600 : *
1601 : * This is currently only used for logging invalidations at the command end
1602 : * or at commit time if any invalidations are pending.
1603 : */
1604 : void
1605 26918 : LogLogicalInvalidations(void)
1606 : {
1607 : xl_xact_invals xlrec;
1608 : InvalidationMsgsGroup *group;
1609 : int nmsgs;
1610 :
1611 : /* Quick exit if we haven't done anything with invalidation messages. */
1612 26918 : if (transInvalInfo == NULL)
1613 16766 : return;
1614 :
1615 10152 : group = &transInvalInfo->CurrentCmdInvalidMsgs;
1616 10152 : nmsgs = NumMessagesInGroup(group);
1617 :
1618 10152 : if (nmsgs > 0)
1619 : {
1620 : /* prepare record */
1621 8268 : memset(&xlrec, 0, MinSizeOfXactInvals);
1622 8268 : xlrec.nmsgs = nmsgs;
1623 :
1624 : /* perform insertion */
1625 8268 : XLogBeginInsert();
1626 8268 : XLogRegisterData((char *) (&xlrec), MinSizeOfXactInvals);
1627 8268 : ProcessMessageSubGroupMulti(group, CatCacheMsgs,
1628 : XLogRegisterData((char *) msgs,
1629 : n * sizeof(SharedInvalidationMessage)));
1630 8268 : ProcessMessageSubGroupMulti(group, RelCacheMsgs,
1631 : XLogRegisterData((char *) msgs,
1632 : n * sizeof(SharedInvalidationMessage)));
1633 8268 : XLogInsert(RM_XACT_ID, XLOG_XACT_INVALIDATIONS);
1634 : }
1635 : }
|