Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * multixact.c
4 : * PostgreSQL multi-transaction-log manager
5 : *
6 : * The pg_multixact manager is a pg_xact-like manager that stores an array of
7 : * MultiXactMember for each MultiXactId. It is a fundamental part of the
8 : * shared-row-lock implementation. Each MultiXactMember is comprised of a
9 : * TransactionId and a set of flag bits. The name is a bit historical:
10 : * originally, a MultiXactId consisted of more than one TransactionId (except
11 : * in rare corner cases), hence "multi". Nowadays, however, it's perfectly
12 : * legitimate to have MultiXactIds that only include a single Xid.
13 : *
14 : * The meaning of the flag bits is opaque to this module, but they are mostly
15 : * used in heapam.c to identify lock modes that each of the member transactions
16 : * is holding on any given tuple. This module just contains support to store
17 : * and retrieve the arrays.
18 : *
19 : * We use two SLRU areas, one for storing the offsets at which the data
20 : * starts for each MultiXactId in the other one. This trick allows us to
21 : * store variable length arrays of TransactionIds. (We could alternatively
22 : * use one area containing counts and TransactionIds, with valid MultiXactId
23 : * values pointing at slots containing counts; but that way seems less robust
24 : * since it would get completely confused if someone inquired about a bogus
25 : * MultiXactId that pointed to an intermediate slot containing an XID.)
26 : *
27 : * XLOG interactions: this module generates a record whenever a new OFFSETs or
28 : * MEMBERs page is initialized to zeroes, as well as an
29 : * XLOG_MULTIXACT_CREATE_ID record whenever a new MultiXactId is defined.
30 : * This module ignores the WAL rule "write xlog before data," because it
31 : * suffices that actions recording a MultiXactId in a heap xmax do follow that
32 : * rule. The only way for the MXID to be referenced from any data page is for
33 : * heap_lock_tuple() or heap_update() to have put it there, and each generates
34 : * an XLOG record that must follow ours. The normal LSN interlock between the
35 : * data page and that XLOG record will ensure that our XLOG record reaches
36 : * disk first. If the SLRU members/offsets data reaches disk sooner than the
37 : * XLOG records, we do not care; after recovery, no xmax will refer to it. On
38 : * the flip side, to ensure that all referenced entries _do_ reach disk, this
39 : * module's XLOG records completely rebuild the data entered since the last
40 : * checkpoint. We flush and sync all dirty OFFSETs and MEMBERs pages to disk
41 : * before each checkpoint is considered complete.
42 : *
43 : * Like clog.c, and unlike subtrans.c, we have to preserve state across
44 : * crashes and ensure that MXID and offset numbering increases monotonically
45 : * across a crash. We do this in the same way as it's done for transaction
46 : * IDs: the WAL record is guaranteed to contain evidence of every MXID we
47 : * could need to worry about, and we just make sure that at the end of
48 : * replay, the next-MXID and next-offset counters are at least as large as
49 : * anything we saw during replay.
50 : *
51 : * We are able to remove segments no longer necessary by carefully tracking
52 : * each table's used values: during vacuum, any multixact older than a certain
53 : * value is removed; the cutoff value is stored in pg_class. The minimum value
54 : * across all tables in each database is stored in pg_database, and the global
55 : * minimum across all databases is part of pg_control and is kept in shared
56 : * memory. Whenever that minimum is advanced, the SLRUs are truncated.
57 : *
58 : * When new multixactid values are to be created, care is taken that the
59 : * counter does not fall within the wraparound horizon considering the global
60 : * minimum value.
61 : *
62 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
63 : * Portions Copyright (c) 1994, Regents of the University of California
64 : *
65 : * src/backend/access/transam/multixact.c
66 : *
67 : *-------------------------------------------------------------------------
68 : */
69 : #include "postgres.h"
70 :
71 : #include "access/multixact.h"
72 : #include "access/multixact_internal.h"
73 : #include "access/slru.h"
74 : #include "access/twophase.h"
75 : #include "access/twophase_rmgr.h"
76 : #include "access/xlog.h"
77 : #include "access/xloginsert.h"
78 : #include "access/xlogutils.h"
79 : #include "miscadmin.h"
80 : #include "pg_trace.h"
81 : #include "pgstat.h"
82 : #include "postmaster/autovacuum.h"
83 : #include "storage/pmsignal.h"
84 : #include "storage/proc.h"
85 : #include "storage/procarray.h"
86 : #include "utils/guc_hooks.h"
87 : #include "utils/injection_point.h"
88 : #include "utils/lsyscache.h"
89 : #include "utils/memutils.h"
90 :
91 :
92 : /*
93 : * Thresholds used to keep members disk usage in check when multixids have a
94 : * lot of members. When MULTIXACT_MEMBER_LOW_THRESHOLD is reached, vacuum
95 : * starts freezing multixids more aggressively, even if the normal multixid
96 : * age limits haven't been reached yet.
97 : */
98 : #define MULTIXACT_MEMBER_LOW_THRESHOLD UINT64CONST(2000000000)
99 : #define MULTIXACT_MEMBER_HIGH_THRESHOLD UINT64CONST(4000000000)
100 :
101 : static inline MultiXactId
102 215162 : NextMultiXactId(MultiXactId multi)
103 : {
104 215162 : return multi == MaxMultiXactId ? FirstMultiXactId : multi + 1;
105 : }
106 :
107 : static inline MultiXactId
108 0 : PreviousMultiXactId(MultiXactId multi)
109 : {
110 0 : return multi == FirstMultiXactId ? MaxMultiXactId : multi - 1;
111 : }
112 :
113 : /*
114 : * Links to shared-memory data structures for MultiXact control
115 : */
116 : static SlruCtlData MultiXactOffsetCtlData;
117 : static SlruCtlData MultiXactMemberCtlData;
118 :
119 : #define MultiXactOffsetCtl (&MultiXactOffsetCtlData)
120 : #define MultiXactMemberCtl (&MultiXactMemberCtlData)
121 :
122 : /*
123 : * MultiXact state shared across all backends. All this state is protected
124 : * by MultiXactGenLock. (We also use SLRU bank's lock of MultiXactOffset and
125 : * MultiXactMember to guard accesses to the two sets of SLRU buffers. For
126 : * concurrency's sake, we avoid holding more than one of these locks at a
127 : * time.)
128 : */
129 : typedef struct MultiXactStateData
130 : {
131 : /* next-to-be-assigned MultiXactId */
132 : MultiXactId nextMXact;
133 :
134 : /* next-to-be-assigned offset */
135 : MultiXactOffset nextOffset;
136 :
137 : /* Have we completed multixact startup? */
138 : bool finishedStartup;
139 :
140 : /*
141 : * Oldest multixact that is still potentially referenced by a relation.
142 : * Anything older than this should not be consulted. These values are
143 : * updated by vacuum.
144 : */
145 : MultiXactId oldestMultiXactId;
146 : Oid oldestMultiXactDB;
147 :
148 : /*
149 : * Oldest multixact offset that is potentially referenced by a multixact
150 : * referenced by a relation.
151 : */
152 : MultiXactOffset oldestOffset;
153 :
154 : /* support for anti-wraparound measures */
155 : MultiXactId multiVacLimit;
156 : MultiXactId multiWarnLimit;
157 : MultiXactId multiStopLimit;
158 : MultiXactId multiWrapLimit;
159 :
160 : /*
161 : * Per-backend data starts here. We have two arrays stored in the area
162 : * immediately following the MultiXactStateData struct. Each is indexed by
163 : * ProcNumber.
164 : *
165 : * In both arrays, there's a slot for all normal backends
166 : * (0..MaxBackends-1) followed by a slot for max_prepared_xacts prepared
167 : * transactions.
168 : *
169 : * OldestMemberMXactId[k] is the oldest MultiXactId each backend's current
170 : * transaction(s) could possibly be a member of, or InvalidMultiXactId
171 : * when the backend has no live transaction that could possibly be a
172 : * member of a MultiXact. Each backend sets its entry to the current
173 : * nextMXact counter just before first acquiring a shared lock in a given
174 : * transaction, and clears it at transaction end. (This works because only
175 : * during or after acquiring a shared lock could an XID possibly become a
176 : * member of a MultiXact, and that MultiXact would have to be created
177 : * during or after the lock acquisition.)
178 : *
179 : * OldestVisibleMXactId[k] is the oldest MultiXactId each backend's
180 : * current transaction(s) think is potentially live, or InvalidMultiXactId
181 : * when not in a transaction or not in a transaction that's paid any
182 : * attention to MultiXacts yet. This is computed when first needed in a
183 : * given transaction, and cleared at transaction end. We can compute it
184 : * as the minimum of the valid OldestMemberMXactId[] entries at the time
185 : * we compute it (using nextMXact if none are valid). Each backend is
186 : * required not to attempt to access any SLRU data for MultiXactIds older
187 : * than its own OldestVisibleMXactId[] setting; this is necessary because
188 : * the relevant SLRU data can be concurrently truncated away.
189 : *
190 : * The oldest valid value among all of the OldestMemberMXactId[] and
191 : * OldestVisibleMXactId[] entries is considered by vacuum as the earliest
192 : * possible value still having any live member transaction -- OldestMxact.
193 : * Any value older than that is typically removed from tuple headers, or
194 : * "frozen" via being replaced with a new xmax. VACUUM can sometimes even
195 : * remove an individual MultiXact xmax whose value is >= its OldestMxact
196 : * cutoff, though typically only when no individual member XID is still
197 : * running. See FreezeMultiXactId for full details.
198 : *
199 : * Whenever VACUUM advances relminmxid, then either its OldestMxact cutoff
200 : * or the oldest extant Multi remaining in the table is used as the new
201 : * pg_class.relminmxid value (whichever is earlier). The minimum of all
202 : * relminmxid values in each database is stored in pg_database.datminmxid.
203 : * In turn, the minimum of all of those values is stored in pg_control.
204 : * This is used as the truncation point for pg_multixact when unneeded
205 : * segments get removed by vac_truncate_clog() during vacuuming.
206 : */
207 : MultiXactId perBackendXactIds[FLEXIBLE_ARRAY_MEMBER];
208 : } MultiXactStateData;
209 :
210 : /*
211 : * Size of OldestMemberMXactId and OldestVisibleMXactId arrays.
212 : */
213 : #define MaxOldestSlot (MaxBackends + max_prepared_xacts)
214 :
215 : /* Pointers to the state data in shared memory */
216 : static MultiXactStateData *MultiXactState;
217 : static MultiXactId *OldestMemberMXactId;
218 : static MultiXactId *OldestVisibleMXactId;
219 :
220 :
221 : /*
222 : * Definitions for the backend-local MultiXactId cache.
223 : *
224 : * We use this cache to store known MultiXacts, so we don't need to go to
225 : * SLRU areas every time.
226 : *
227 : * The cache lasts for the duration of a single transaction, the rationale
228 : * for this being that most entries will contain our own TransactionId and
229 : * so they will be uninteresting by the time our next transaction starts.
230 : * (XXX not clear that this is correct --- other members of the MultiXact
231 : * could hang around longer than we did. However, it's not clear what a
232 : * better policy for flushing old cache entries would be.) FIXME actually
233 : * this is plain wrong now that multixact's may contain update Xids.
234 : *
235 : * We allocate the cache entries in a memory context that is deleted at
236 : * transaction end, so we don't need to do retail freeing of entries.
237 : */
238 : typedef struct mXactCacheEnt
239 : {
240 : MultiXactId multi;
241 : int nmembers;
242 : dlist_node node;
243 : MultiXactMember members[FLEXIBLE_ARRAY_MEMBER];
244 : } mXactCacheEnt;
245 :
246 : #define MAX_CACHE_ENTRIES 256
247 : static dclist_head MXactCache = DCLIST_STATIC_INIT(MXactCache);
248 : static MemoryContext MXactContext = NULL;
249 :
250 : #ifdef MULTIXACT_DEBUG
251 : #define debug_elog2(a,b) elog(a,b)
252 : #define debug_elog3(a,b,c) elog(a,b,c)
253 : #define debug_elog4(a,b,c,d) elog(a,b,c,d)
254 : #define debug_elog5(a,b,c,d,e) elog(a,b,c,d,e)
255 : #define debug_elog6(a,b,c,d,e,f) elog(a,b,c,d,e,f)
256 : #else
257 : #define debug_elog2(a,b)
258 : #define debug_elog3(a,b,c)
259 : #define debug_elog4(a,b,c,d)
260 : #define debug_elog5(a,b,c,d,e)
261 : #define debug_elog6(a,b,c,d,e,f)
262 : #endif
263 :
264 : /* internal MultiXactId management */
265 : static void MultiXactIdSetOldestVisible(void);
266 : static void RecordNewMultiXact(MultiXactId multi, MultiXactOffset offset,
267 : int nmembers, MultiXactMember *members);
268 : static MultiXactId GetNewMultiXactId(int nmembers, MultiXactOffset *offset);
269 :
270 : /* MultiXact cache management */
271 : static int mxactMemberComparator(const void *arg1, const void *arg2);
272 : static MultiXactId mXactCacheGetBySet(int nmembers, MultiXactMember *members);
273 : static int mXactCacheGetById(MultiXactId multi, MultiXactMember **members);
274 : static void mXactCachePut(MultiXactId multi, int nmembers,
275 : MultiXactMember *members);
276 :
277 : /* management of SLRU infrastructure */
278 : static bool MultiXactOffsetPagePrecedes(int64 page1, int64 page2);
279 : static bool MultiXactMemberPagePrecedes(int64 page1, int64 page2);
280 : static void ExtendMultiXactOffset(MultiXactId multi);
281 : static void ExtendMultiXactMember(MultiXactOffset offset, int nmembers);
282 : static void SetOldestOffset(void);
283 : static bool find_multixact_start(MultiXactId multi, MultiXactOffset *result);
284 : static void WriteMTruncateXlogRec(Oid oldestMultiDB,
285 : MultiXactId endTruncOff,
286 : MultiXactOffset endTruncMemb);
287 :
288 :
289 : /*
290 : * MultiXactIdCreate
291 : * Construct a MultiXactId representing two TransactionIds.
292 : *
293 : * The two XIDs must be different, or be requesting different statuses.
294 : *
295 : * NB - we don't worry about our local MultiXactId cache here, because that
296 : * is handled by the lower-level routines.
297 : */
298 : MultiXactId
299 2196 : MultiXactIdCreate(TransactionId xid1, MultiXactStatus status1,
300 : TransactionId xid2, MultiXactStatus status2)
301 : {
302 : MultiXactId newMulti;
303 : MultiXactMember members[2];
304 :
305 : Assert(TransactionIdIsValid(xid1));
306 : Assert(TransactionIdIsValid(xid2));
307 :
308 : Assert(!TransactionIdEquals(xid1, xid2) || (status1 != status2));
309 :
310 : /* MultiXactIdSetOldestMember() must have been called already. */
311 : Assert(MultiXactIdIsValid(OldestMemberMXactId[MyProcNumber]));
312 :
313 : /*
314 : * Note: unlike MultiXactIdExpand, we don't bother to check that both XIDs
315 : * are still running. In typical usage, xid2 will be our own XID and the
316 : * caller just did a check on xid1, so it'd be wasted effort.
317 : */
318 :
319 2196 : members[0].xid = xid1;
320 2196 : members[0].status = status1;
321 2196 : members[1].xid = xid2;
322 2196 : members[1].status = status2;
323 :
324 2196 : newMulti = MultiXactIdCreateFromMembers(2, members);
325 :
326 : debug_elog3(DEBUG2, "Create: %s",
327 : mxid_to_string(newMulti, 2, members));
328 :
329 2196 : return newMulti;
330 : }
331 :
332 : /*
333 : * MultiXactIdExpand
334 : * Add a TransactionId to a pre-existing MultiXactId.
335 : *
336 : * If the TransactionId is already a member of the passed MultiXactId with the
337 : * same status, just return it as-is.
338 : *
339 : * Note that we do NOT actually modify the membership of a pre-existing
340 : * MultiXactId; instead we create a new one. This is necessary to avoid
341 : * a race condition against code trying to wait for one MultiXactId to finish;
342 : * see notes in heapam.c.
343 : *
344 : * NB - we don't worry about our local MultiXactId cache here, because that
345 : * is handled by the lower-level routines.
346 : *
347 : * Note: It is critical that MultiXactIds that come from an old cluster (i.e.
348 : * one upgraded by pg_upgrade from a cluster older than this feature) are not
349 : * passed in.
350 : */
351 : MultiXactId
352 151052 : MultiXactIdExpand(MultiXactId multi, TransactionId xid, MultiXactStatus status)
353 : {
354 : MultiXactId newMulti;
355 : MultiXactMember *members;
356 : MultiXactMember *newMembers;
357 : int nmembers;
358 : int i;
359 : int j;
360 :
361 : Assert(MultiXactIdIsValid(multi));
362 : Assert(TransactionIdIsValid(xid));
363 :
364 : /* MultiXactIdSetOldestMember() must have been called already. */
365 : Assert(MultiXactIdIsValid(OldestMemberMXactId[MyProcNumber]));
366 :
367 : debug_elog5(DEBUG2, "Expand: received multi %u, xid %u status %s",
368 : multi, xid, mxstatus_to_string(status));
369 :
370 : /*
371 : * Note: we don't allow for old multis here. The reason is that the only
372 : * caller of this function does a check that the multixact is no longer
373 : * running.
374 : */
375 151052 : nmembers = GetMultiXactIdMembers(multi, &members, false, false);
376 :
377 151052 : if (nmembers < 0)
378 : {
379 : MultiXactMember member;
380 :
381 : /*
382 : * The MultiXactId is obsolete. This can only happen if all the
383 : * MultiXactId members stop running between the caller checking and
384 : * passing it to us. It would be better to return that fact to the
385 : * caller, but it would complicate the API and it's unlikely to happen
386 : * too often, so just deal with it by creating a singleton MultiXact.
387 : */
388 0 : member.xid = xid;
389 0 : member.status = status;
390 0 : newMulti = MultiXactIdCreateFromMembers(1, &member);
391 :
392 : debug_elog4(DEBUG2, "Expand: %u has no members, create singleton %u",
393 : multi, newMulti);
394 0 : return newMulti;
395 : }
396 :
397 : /*
398 : * If the TransactionId is already a member of the MultiXactId with the
399 : * same status, just return the existing MultiXactId.
400 : */
401 2931802 : for (i = 0; i < nmembers; i++)
402 : {
403 2780750 : if (TransactionIdEquals(members[i].xid, xid) &&
404 108 : (members[i].status == status))
405 : {
406 : debug_elog4(DEBUG2, "Expand: %u is already a member of %u",
407 : xid, multi);
408 0 : pfree(members);
409 0 : return multi;
410 : }
411 : }
412 :
413 : /*
414 : * Determine which of the members of the MultiXactId are still of
415 : * interest. This is any running transaction, and also any transaction
416 : * that grabbed something stronger than just a lock and was committed. (An
417 : * update that aborted is of no interest here; and having more than one
418 : * update Xid in a multixact would cause errors elsewhere.)
419 : *
420 : * Removing dead members is not just an optimization: freezing of tuples
421 : * whose Xmax are multis depends on this behavior.
422 : *
423 : * Note we have the same race condition here as above: j could be 0 at the
424 : * end of the loop.
425 : */
426 151052 : newMembers = palloc_array(MultiXactMember, nmembers + 1);
427 :
428 2931802 : for (i = 0, j = 0; i < nmembers; i++)
429 : {
430 2780750 : if (TransactionIdIsInProgress(members[i].xid) ||
431 149350 : (ISUPDATE_from_mxstatus(members[i].status) &&
432 34 : TransactionIdDidCommit(members[i].xid)))
433 : {
434 2631434 : newMembers[j].xid = members[i].xid;
435 2631434 : newMembers[j++].status = members[i].status;
436 : }
437 : }
438 :
439 151052 : newMembers[j].xid = xid;
440 151052 : newMembers[j++].status = status;
441 151052 : newMulti = MultiXactIdCreateFromMembers(j, newMembers);
442 :
443 151052 : pfree(members);
444 151052 : pfree(newMembers);
445 :
446 : debug_elog3(DEBUG2, "Expand: returning new multi %u", newMulti);
447 :
448 151052 : return newMulti;
449 : }
450 :
451 : /*
452 : * MultiXactIdIsRunning
453 : * Returns whether a MultiXactId is "running".
454 : *
455 : * We return true if at least one member of the given MultiXactId is still
456 : * running. Note that a "false" result is certain not to change,
457 : * because it is not legal to add members to an existing MultiXactId.
458 : *
459 : * Caller is expected to have verified that the multixact does not come from
460 : * a pg_upgraded share-locked tuple.
461 : */
462 : bool
463 299596 : MultiXactIdIsRunning(MultiXactId multi, bool isLockOnly)
464 : {
465 : MultiXactMember *members;
466 : int nmembers;
467 : int i;
468 :
469 : debug_elog3(DEBUG2, "IsRunning %u?", multi);
470 :
471 : /*
472 : * "false" here means we assume our callers have checked that the given
473 : * multi cannot possibly come from a pg_upgraded database.
474 : */
475 299596 : nmembers = GetMultiXactIdMembers(multi, &members, false, isLockOnly);
476 :
477 299596 : if (nmembers <= 0)
478 : {
479 : debug_elog2(DEBUG2, "IsRunning: no members");
480 1432 : return false;
481 : }
482 :
483 : /*
484 : * Checking for myself is cheap compared to looking in shared memory;
485 : * return true if any live subtransaction of the current top-level
486 : * transaction is a member.
487 : *
488 : * This is not needed for correctness, it's just a fast path.
489 : */
490 5782836 : for (i = 0; i < nmembers; i++)
491 : {
492 5484984 : if (TransactionIdIsCurrentTransactionId(members[i].xid))
493 : {
494 : debug_elog3(DEBUG2, "IsRunning: I (%d) am running!", i);
495 312 : pfree(members);
496 312 : return true;
497 : }
498 : }
499 :
500 : /*
501 : * This could be made faster by having another entry point in procarray.c,
502 : * walking the PGPROC array only once for all the members. But in most
503 : * cases nmembers should be small enough that it doesn't much matter.
504 : */
505 592234 : for (i = 0; i < nmembers; i++)
506 : {
507 592146 : if (TransactionIdIsInProgress(members[i].xid))
508 : {
509 : debug_elog4(DEBUG2, "IsRunning: member %d (%u) is running",
510 : i, members[i].xid);
511 297764 : pfree(members);
512 297764 : return true;
513 : }
514 : }
515 :
516 88 : pfree(members);
517 :
518 : debug_elog3(DEBUG2, "IsRunning: %u is not running", multi);
519 :
520 88 : return false;
521 : }
522 :
523 : /*
524 : * MultiXactIdSetOldestMember
525 : * Save the oldest MultiXactId this transaction could be a member of.
526 : *
527 : * We set the OldestMemberMXactId for a given transaction the first time it's
528 : * going to do some operation that might require a MultiXactId (tuple lock,
529 : * update or delete). We need to do this even if we end up using a
530 : * TransactionId instead of a MultiXactId, because there is a chance that
531 : * another transaction would add our XID to a MultiXactId.
532 : *
533 : * The value to set is the next-to-be-assigned MultiXactId, so this is meant to
534 : * be called just before doing any such possibly-MultiXactId-able operation.
535 : */
536 : void
537 3986720 : MultiXactIdSetOldestMember(void)
538 : {
539 3986720 : if (!MultiXactIdIsValid(OldestMemberMXactId[MyProcNumber]))
540 : {
541 : MultiXactId nextMXact;
542 :
543 : /*
544 : * You might think we don't need to acquire a lock here, since
545 : * fetching and storing of TransactionIds is probably atomic, but in
546 : * fact we do: suppose we pick up nextMXact and then lose the CPU for
547 : * a long time. Someone else could advance nextMXact, and then
548 : * another someone else could compute an OldestVisibleMXactId that
549 : * would be after the value we are going to store when we get control
550 : * back. Which would be wrong.
551 : *
552 : * Note that a shared lock is sufficient, because it's enough to stop
553 : * someone from advancing nextMXact; and nobody else could be trying
554 : * to write to our OldestMember entry, only reading (and we assume
555 : * storing it is atomic.)
556 : */
557 150412 : LWLockAcquire(MultiXactGenLock, LW_SHARED);
558 :
559 150412 : nextMXact = MultiXactState->nextMXact;
560 :
561 150412 : OldestMemberMXactId[MyProcNumber] = nextMXact;
562 :
563 150412 : LWLockRelease(MultiXactGenLock);
564 :
565 : debug_elog4(DEBUG2, "MultiXact: setting OldestMember[%d] = %u",
566 : MyProcNumber, nextMXact);
567 : }
568 3986720 : }
569 :
570 : /*
571 : * MultiXactIdSetOldestVisible
572 : * Save the oldest MultiXactId this transaction considers possibly live.
573 : *
574 : * We set the OldestVisibleMXactId for a given transaction the first time
575 : * it's going to inspect any MultiXactId. Once we have set this, we are
576 : * guaranteed that SLRU data for MultiXactIds >= our own OldestVisibleMXactId
577 : * won't be truncated away.
578 : *
579 : * The value to set is the oldest of nextMXact and all the valid per-backend
580 : * OldestMemberMXactId[] entries. Because of the locking we do, we can be
581 : * certain that no subsequent call to MultiXactIdSetOldestMember can set
582 : * an OldestMemberMXactId[] entry older than what we compute here. Therefore
583 : * there is no live transaction, now or later, that can be a member of any
584 : * MultiXactId older than the OldestVisibleMXactId we compute here.
585 : */
586 : static void
587 184870 : MultiXactIdSetOldestVisible(void)
588 : {
589 184870 : if (!MultiXactIdIsValid(OldestVisibleMXactId[MyProcNumber]))
590 : {
591 : MultiXactId oldestMXact;
592 : int i;
593 :
594 6358 : LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
595 :
596 6358 : oldestMXact = MultiXactState->nextMXact;
597 811210 : for (i = 0; i < MaxOldestSlot; i++)
598 : {
599 804852 : MultiXactId thisoldest = OldestMemberMXactId[i];
600 :
601 916608 : if (MultiXactIdIsValid(thisoldest) &&
602 111756 : MultiXactIdPrecedes(thisoldest, oldestMXact))
603 11330 : oldestMXact = thisoldest;
604 : }
605 :
606 6358 : OldestVisibleMXactId[MyProcNumber] = oldestMXact;
607 :
608 6358 : LWLockRelease(MultiXactGenLock);
609 :
610 : debug_elog4(DEBUG2, "MultiXact: setting OldestVisible[%d] = %u",
611 : MyProcNumber, oldestMXact);
612 : }
613 184870 : }
614 :
615 : /*
616 : * ReadNextMultiXactId
617 : * Return the next MultiXactId to be assigned, but don't allocate it
618 : */
619 : MultiXactId
620 333964 : ReadNextMultiXactId(void)
621 : {
622 : MultiXactId mxid;
623 :
624 : /* XXX we could presumably do this without a lock. */
625 333964 : LWLockAcquire(MultiXactGenLock, LW_SHARED);
626 333964 : mxid = MultiXactState->nextMXact;
627 333964 : LWLockRelease(MultiXactGenLock);
628 :
629 333964 : return mxid;
630 : }
631 :
632 : /*
633 : * ReadMultiXactIdRange
634 : * Get the range of IDs that may still be referenced by a relation.
635 : */
636 : void
637 2860 : ReadMultiXactIdRange(MultiXactId *oldest, MultiXactId *next)
638 : {
639 2860 : LWLockAcquire(MultiXactGenLock, LW_SHARED);
640 2860 : *oldest = MultiXactState->oldestMultiXactId;
641 2860 : *next = MultiXactState->nextMXact;
642 2860 : LWLockRelease(MultiXactGenLock);
643 2860 : }
644 :
645 :
646 : /*
647 : * MultiXactIdCreateFromMembers
648 : * Make a new MultiXactId from the specified set of members
649 : *
650 : * Make XLOG, SLRU and cache entries for a new MultiXactId, recording the
651 : * given TransactionIds as members. Returns the newly created MultiXactId.
652 : *
653 : * NB: the passed members[] array will be sorted in-place.
654 : */
655 : MultiXactId
656 153250 : MultiXactIdCreateFromMembers(int nmembers, MultiXactMember *members)
657 : {
658 : MultiXactId multi;
659 : MultiXactOffset offset;
660 : xl_multixact_create xlrec;
661 :
662 : debug_elog3(DEBUG2, "Create: %s",
663 : mxid_to_string(InvalidMultiXactId, nmembers, members));
664 :
665 : /*
666 : * See if the same set of members already exists in our cache; if so, just
667 : * re-use that MultiXactId. (Note: it might seem that looking in our
668 : * cache is insufficient, and we ought to search disk to see if a
669 : * duplicate definition already exists. But since we only ever create
670 : * MultiXacts containing our own XID, in most cases any such MultiXacts
671 : * were in fact created by us, and so will be in our cache. There are
672 : * corner cases where someone else added us to a MultiXact without our
673 : * knowledge, but it's not worth checking for.)
674 : */
675 153250 : multi = mXactCacheGetBySet(nmembers, members);
676 153250 : if (MultiXactIdIsValid(multi))
677 : {
678 : debug_elog2(DEBUG2, "Create: in cache!");
679 142680 : return multi;
680 : }
681 :
682 : /* Verify that there is a single update Xid among the given members. */
683 : {
684 : int i;
685 10570 : bool has_update = false;
686 :
687 199944 : for (i = 0; i < nmembers; i++)
688 : {
689 189374 : if (ISUPDATE_from_mxstatus(members[i].status))
690 : {
691 4746 : if (has_update)
692 0 : elog(ERROR, "new multixact has more than one updating member: %s",
693 : mxid_to_string(InvalidMultiXactId, nmembers, members));
694 4746 : has_update = true;
695 : }
696 : }
697 : }
698 :
699 : /* Load the injection point before entering the critical section */
700 10570 : INJECTION_POINT_LOAD("multixact-create-from-members");
701 :
702 : /*
703 : * Assign the MXID and offsets range to use, and make sure there is space
704 : * in the OFFSETs and MEMBERs files. NB: this routine does
705 : * START_CRIT_SECTION().
706 : *
707 : * Note: unlike MultiXactIdCreate and MultiXactIdExpand, we do not check
708 : * that we've called MultiXactIdSetOldestMember here. This is because
709 : * this routine is used in some places to create new MultiXactIds of which
710 : * the current backend is not a member, notably during freezing of multis
711 : * in vacuum. During vacuum, in particular, it would be unacceptable to
712 : * keep OldestMulti set, in case it runs for long.
713 : */
714 10570 : multi = GetNewMultiXactId(nmembers, &offset);
715 :
716 10570 : INJECTION_POINT_CACHED("multixact-create-from-members", NULL);
717 :
718 : /* Make an XLOG entry describing the new MXID. */
719 10570 : xlrec.mid = multi;
720 10570 : xlrec.moff = offset;
721 10570 : xlrec.nmembers = nmembers;
722 :
723 : /*
724 : * XXX Note: there's a lot of padding space in MultiXactMember. We could
725 : * find a more compact representation of this Xlog record -- perhaps all
726 : * the status flags in one XLogRecData, then all the xids in another one?
727 : * Not clear that it's worth the trouble though.
728 : */
729 10570 : XLogBeginInsert();
730 10570 : XLogRegisterData(&xlrec, SizeOfMultiXactCreate);
731 10570 : XLogRegisterData(members, nmembers * sizeof(MultiXactMember));
732 :
733 10570 : (void) XLogInsert(RM_MULTIXACT_ID, XLOG_MULTIXACT_CREATE_ID);
734 :
735 : /* Now enter the information into the OFFSETs and MEMBERs logs */
736 10570 : RecordNewMultiXact(multi, offset, nmembers, members);
737 :
738 : /* Done with critical section */
739 10570 : END_CRIT_SECTION();
740 :
741 : /* Store the new MultiXactId in the local cache, too */
742 10570 : mXactCachePut(multi, nmembers, members);
743 :
744 : debug_elog2(DEBUG2, "Create: all done");
745 :
746 10570 : return multi;
747 : }
748 :
749 : /*
750 : * RecordNewMultiXact
751 : * Write info about a new multixact into the offsets and members files
752 : *
753 : * This is broken out of MultiXactIdCreateFromMembers so that xlog replay can
754 : * use it.
755 : */
756 : static void
757 10578 : RecordNewMultiXact(MultiXactId multi, MultiXactOffset offset,
758 : int nmembers, MultiXactMember *members)
759 : {
760 : int64 pageno;
761 : int64 prev_pageno;
762 : int entryno;
763 : int slotno;
764 : MultiXactOffset *offptr;
765 : MultiXactId next;
766 : int64 next_pageno;
767 : int next_entryno;
768 : MultiXactOffset *next_offptr;
769 : MultiXactOffset next_offset;
770 : LWLock *lock;
771 10578 : LWLock *prevlock = NULL;
772 :
773 : /* position of this multixid in the offsets SLRU area */
774 10578 : pageno = MultiXactIdToOffsetPage(multi);
775 10578 : entryno = MultiXactIdToOffsetEntry(multi);
776 :
777 : /* position of the next multixid */
778 10578 : next = NextMultiXactId(multi);
779 10578 : next_pageno = MultiXactIdToOffsetPage(next);
780 10578 : next_entryno = MultiXactIdToOffsetEntry(next);
781 :
782 : /*
783 : * Set the starting offset of this multixid's members.
784 : *
785 : * In the common case, it was already be set by the previous
786 : * RecordNewMultiXact call, as this was the next multixid of the previous
787 : * multixid. But if multiple backends are generating multixids
788 : * concurrently, we might race ahead and get called before the previous
789 : * multixid.
790 : */
791 10578 : lock = SimpleLruGetBankLock(MultiXactOffsetCtl, pageno);
792 10578 : LWLockAcquire(lock, LW_EXCLUSIVE);
793 :
794 : /*
795 : * Note: we pass the MultiXactId to SimpleLruReadPage as the "transaction"
796 : * to complain about if there's any I/O error. This is kinda bogus, but
797 : * since the errors will always give the full pathname, it should be clear
798 : * enough that a MultiXactId is really involved. Perhaps someday we'll
799 : * take the trouble to generalize the slru.c error reporting code.
800 : */
801 10578 : slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, multi);
802 10578 : offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
803 10578 : offptr += entryno;
804 :
805 10578 : if (*offptr != offset)
806 : {
807 : /* should already be set to the correct value, or not at all */
808 : Assert(*offptr == 0);
809 2 : *offptr = offset;
810 2 : MultiXactOffsetCtl->shared->page_dirty[slotno] = true;
811 : }
812 :
813 : /*
814 : * Set the next multixid's offset to the end of this multixid's members.
815 : */
816 10578 : if (next_pageno == pageno)
817 : {
818 10568 : next_offptr = offptr + 1;
819 : }
820 : else
821 : {
822 : /* must be the first entry on the page */
823 : Assert(next_entryno == 0 || next == FirstMultiXactId);
824 :
825 : /* Swap the lock for a lock on the next page */
826 10 : LWLockRelease(lock);
827 10 : lock = SimpleLruGetBankLock(MultiXactOffsetCtl, next_pageno);
828 10 : LWLockAcquire(lock, LW_EXCLUSIVE);
829 :
830 10 : slotno = SimpleLruReadPage(MultiXactOffsetCtl, next_pageno, true, next);
831 10 : next_offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
832 10 : next_offptr += next_entryno;
833 : }
834 :
835 : /* Like in GetNewMultiXactId(), skip over offset 0 */
836 10578 : next_offset = offset + nmembers;
837 10578 : if (next_offset == 0)
838 0 : next_offset = 1;
839 10578 : if (*next_offptr != next_offset)
840 : {
841 : /* should already be set to the correct value, or not at all */
842 : Assert(*next_offptr == 0);
843 10578 : *next_offptr = next_offset;
844 10578 : MultiXactOffsetCtl->shared->page_dirty[slotno] = true;
845 : }
846 :
847 : /* Release MultiXactOffset SLRU lock. */
848 10578 : LWLockRelease(lock);
849 :
850 10578 : prev_pageno = -1;
851 :
852 199968 : for (int i = 0; i < nmembers; i++, offset++)
853 : {
854 : TransactionId *memberptr;
855 : uint32 *flagsptr;
856 : uint32 flagsval;
857 : int bshift;
858 : int flagsoff;
859 : int memberoff;
860 :
861 : Assert(members[i].status <= MultiXactStatusUpdate);
862 :
863 189390 : pageno = MXOffsetToMemberPage(offset);
864 189390 : memberoff = MXOffsetToMemberOffset(offset);
865 189390 : flagsoff = MXOffsetToFlagsOffset(offset);
866 189390 : bshift = MXOffsetToFlagsBitShift(offset);
867 :
868 189390 : if (pageno != prev_pageno)
869 : {
870 : /*
871 : * MultiXactMember SLRU page is changed so check if this new page
872 : * fall into the different SLRU bank then release the old bank's
873 : * lock and acquire lock on the new bank.
874 : */
875 10686 : lock = SimpleLruGetBankLock(MultiXactMemberCtl, pageno);
876 10686 : if (lock != prevlock)
877 : {
878 10686 : if (prevlock != NULL)
879 108 : LWLockRelease(prevlock);
880 :
881 10686 : LWLockAcquire(lock, LW_EXCLUSIVE);
882 10686 : prevlock = lock;
883 : }
884 10686 : slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, multi);
885 10686 : prev_pageno = pageno;
886 : }
887 :
888 189390 : memberptr = (TransactionId *)
889 189390 : (MultiXactMemberCtl->shared->page_buffer[slotno] + memberoff);
890 :
891 189390 : *memberptr = members[i].xid;
892 :
893 189390 : flagsptr = (uint32 *)
894 189390 : (MultiXactMemberCtl->shared->page_buffer[slotno] + flagsoff);
895 :
896 189390 : flagsval = *flagsptr;
897 189390 : flagsval &= ~(((1 << MXACT_MEMBER_BITS_PER_XACT) - 1) << bshift);
898 189390 : flagsval |= (members[i].status << bshift);
899 189390 : *flagsptr = flagsval;
900 :
901 189390 : MultiXactMemberCtl->shared->page_dirty[slotno] = true;
902 : }
903 :
904 10578 : if (prevlock != NULL)
905 10578 : LWLockRelease(prevlock);
906 10578 : }
907 :
908 : /*
909 : * GetNewMultiXactId
910 : * Get the next MultiXactId.
911 : *
912 : * Also, reserve the needed amount of space in the "members" area. The
913 : * starting offset of the reserved space is returned in *offset.
914 : *
915 : * This may generate XLOG records for expansion of the offsets and/or members
916 : * files. Unfortunately, we have to do that while holding MultiXactGenLock
917 : * to avoid race conditions --- the XLOG record for zeroing a page must appear
918 : * before any backend can possibly try to store data in that page!
919 : *
920 : * We start a critical section before advancing the shared counters. The
921 : * caller must end the critical section after writing SLRU data.
922 : */
923 : static MultiXactId
924 10570 : GetNewMultiXactId(int nmembers, MultiXactOffset *offset)
925 : {
926 : MultiXactId result;
927 : MultiXactOffset nextOffset;
928 :
929 : debug_elog3(DEBUG2, "GetNew: for %d xids", nmembers);
930 :
931 : /* safety check, we should never get this far in a HS standby */
932 10570 : if (RecoveryInProgress())
933 0 : elog(ERROR, "cannot assign MultiXactIds during recovery");
934 :
935 10570 : LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
936 :
937 : /* Assign the MXID */
938 10570 : result = MultiXactState->nextMXact;
939 :
940 : /*----------
941 : * Check to see if it's safe to assign another MultiXactId. This protects
942 : * against catastrophic data loss due to multixact wraparound. The basic
943 : * rules are:
944 : *
945 : * If we're past multiVacLimit or the safe threshold for member storage
946 : * space, or we don't know what the safe threshold for member storage is,
947 : * start trying to force autovacuum cycles.
948 : * If we're past multiWarnLimit, start issuing warnings.
949 : * If we're past multiStopLimit, refuse to create new MultiXactIds.
950 : *
951 : * Note these are pretty much the same protections in GetNewTransactionId.
952 : *----------
953 : */
954 10570 : if (!MultiXactIdPrecedes(result, MultiXactState->multiVacLimit))
955 : {
956 : /*
957 : * For safety's sake, we release MultiXactGenLock while sending
958 : * signals, warnings, etc. This is not so much because we care about
959 : * preserving concurrency in this situation, as to avoid any
960 : * possibility of deadlock while doing get_database_name(). First,
961 : * copy all the shared values we'll need in this path.
962 : */
963 0 : MultiXactId multiWarnLimit = MultiXactState->multiWarnLimit;
964 0 : MultiXactId multiStopLimit = MultiXactState->multiStopLimit;
965 0 : MultiXactId multiWrapLimit = MultiXactState->multiWrapLimit;
966 0 : Oid oldest_datoid = MultiXactState->oldestMultiXactDB;
967 :
968 0 : LWLockRelease(MultiXactGenLock);
969 :
970 0 : if (IsUnderPostmaster &&
971 0 : !MultiXactIdPrecedes(result, multiStopLimit))
972 : {
973 0 : char *oldest_datname = get_database_name(oldest_datoid);
974 :
975 : /*
976 : * Immediately kick autovacuum into action as we're already in
977 : * ERROR territory.
978 : */
979 0 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
980 :
981 : /* complain even if that DB has disappeared */
982 0 : if (oldest_datname)
983 0 : ereport(ERROR,
984 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
985 : errmsg("database is not accepting commands that assign new MultiXactIds to avoid wraparound data loss in database \"%s\"",
986 : oldest_datname),
987 : errhint("Execute a database-wide VACUUM in that database.\n"
988 : "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
989 : else
990 0 : ereport(ERROR,
991 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
992 : errmsg("database is not accepting commands that assign new MultiXactIds to avoid wraparound data loss in database with OID %u",
993 : oldest_datoid),
994 : errhint("Execute a database-wide VACUUM in that database.\n"
995 : "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
996 : }
997 :
998 : /*
999 : * To avoid swamping the postmaster with signals, we issue the autovac
1000 : * request only once per 64K multis generated. This still gives
1001 : * plenty of chances before we get into real trouble.
1002 : */
1003 0 : if (IsUnderPostmaster && ((result % 65536) == 0 || result == FirstMultiXactId))
1004 0 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
1005 :
1006 0 : if (!MultiXactIdPrecedes(result, multiWarnLimit))
1007 : {
1008 0 : char *oldest_datname = get_database_name(oldest_datoid);
1009 :
1010 : /* complain even if that DB has disappeared */
1011 0 : if (oldest_datname)
1012 0 : ereport(WARNING,
1013 : (errmsg_plural("database \"%s\" must be vacuumed before %u more MultiXactId is used",
1014 : "database \"%s\" must be vacuumed before %u more MultiXactIds are used",
1015 : multiWrapLimit - result,
1016 : oldest_datname,
1017 : multiWrapLimit - result),
1018 : errhint("Execute a database-wide VACUUM in that database.\n"
1019 : "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
1020 : else
1021 0 : ereport(WARNING,
1022 : (errmsg_plural("database with OID %u must be vacuumed before %u more MultiXactId is used",
1023 : "database with OID %u must be vacuumed before %u more MultiXactIds are used",
1024 : multiWrapLimit - result,
1025 : oldest_datoid,
1026 : multiWrapLimit - result),
1027 : errhint("Execute a database-wide VACUUM in that database.\n"
1028 : "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
1029 : }
1030 :
1031 : /* Re-acquire lock and start over */
1032 0 : LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
1033 0 : result = MultiXactState->nextMXact;
1034 : }
1035 :
1036 : /*
1037 : * Make sure there is room for the next MXID in the file. Assigning this
1038 : * MXID sets the next MXID's offset already.
1039 : */
1040 10570 : ExtendMultiXactOffset(NextMultiXactId(result));
1041 :
1042 : /*
1043 : * Reserve the members space, similarly to above.
1044 : */
1045 10570 : nextOffset = MultiXactState->nextOffset;
1046 :
1047 : /*
1048 : * Offsets are 64-bit integers and will never wrap around. Firstly, it
1049 : * would take an unrealistic amount of time and resources to consume 2^64
1050 : * offsets. Secondly, multixid creation is WAL-logged, so you would run
1051 : * out of LSNs before reaching offset wraparound. Nevertheless, check for
1052 : * wraparound as a sanity check.
1053 : */
1054 10570 : if (nextOffset + nmembers < nextOffset)
1055 0 : ereport(ERROR,
1056 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1057 : errmsg("MultiXact members would wrap around")));
1058 10570 : *offset = nextOffset;
1059 :
1060 10570 : ExtendMultiXactMember(nextOffset, nmembers);
1061 :
1062 : /*
1063 : * Critical section from here until caller has written the data into the
1064 : * just-reserved SLRU space; we don't want to error out with a partly
1065 : * written MultiXact structure. (In particular, failing to write our
1066 : * start offset after advancing nextMXact would effectively corrupt the
1067 : * previous MultiXact.)
1068 : */
1069 10570 : START_CRIT_SECTION();
1070 :
1071 : /*
1072 : * Advance counters. As in GetNewTransactionId(), this must not happen
1073 : * until after file extension has succeeded!
1074 : */
1075 10570 : MultiXactState->nextMXact = NextMultiXactId(result);
1076 10570 : MultiXactState->nextOffset += nmembers;
1077 :
1078 10570 : LWLockRelease(MultiXactGenLock);
1079 :
1080 : debug_elog4(DEBUG2, "GetNew: returning %u offset %" PRIu64,
1081 : result, *offset);
1082 10570 : return result;
1083 : }
1084 :
1085 : /*
1086 : * GetMultiXactIdMembers
1087 : * Return the set of MultiXactMembers that make up a MultiXactId
1088 : *
1089 : * Return value is the number of members found, or -1 if there are none,
1090 : * and *members is set to a newly palloc'ed array of members. It's the
1091 : * caller's responsibility to free it when done with it.
1092 : *
1093 : * from_pgupgrade must be passed as true if and only if only the multixact
1094 : * corresponds to a value from a tuple that was locked in a 9.2-or-older
1095 : * installation and later pg_upgrade'd (that is, the infomask is
1096 : * HEAP_LOCKED_UPGRADED). In this case, we know for certain that no members
1097 : * can still be running, so we return -1 just like for an empty multixact
1098 : * without any further checking. It would be wrong to try to resolve such a
1099 : * multixact: either the multixact is within the current valid multixact
1100 : * range, in which case the returned result would be bogus, or outside that
1101 : * range, in which case an error would be raised.
1102 : *
1103 : * In all other cases, the passed multixact must be within the known valid
1104 : * range, that is, greater than or equal to oldestMultiXactId, and less than
1105 : * nextMXact. Otherwise, an error is raised.
1106 : *
1107 : * isLockOnly must be set to true if caller is certain that the given multi
1108 : * is used only to lock tuples; can be false without loss of correctness,
1109 : * but passing a true means we can return quickly without checking for
1110 : * old updates.
1111 : */
1112 : int
1113 1099618 : GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
1114 : bool from_pgupgrade, bool isLockOnly)
1115 : {
1116 : int64 pageno;
1117 : int64 prev_pageno;
1118 : int entryno;
1119 : int slotno;
1120 : MultiXactOffset *offptr;
1121 : MultiXactOffset offset;
1122 : MultiXactOffset nextMXOffset;
1123 : int length;
1124 : MultiXactId oldestMXact;
1125 : MultiXactId nextMXact;
1126 : MultiXactMember *ptr;
1127 : LWLock *lock;
1128 :
1129 : debug_elog3(DEBUG2, "GetMembers: asked for %u", multi);
1130 :
1131 1099618 : if (!MultiXactIdIsValid(multi) || from_pgupgrade)
1132 : {
1133 0 : *members = NULL;
1134 0 : return -1;
1135 : }
1136 :
1137 : /* See if the MultiXactId is in the local cache */
1138 1099618 : length = mXactCacheGetById(multi, members);
1139 1099618 : if (length >= 0)
1140 : {
1141 : debug_elog3(DEBUG2, "GetMembers: found %s in the cache",
1142 : mxid_to_string(multi, length, *members));
1143 914748 : return length;
1144 : }
1145 :
1146 : /* Set our OldestVisibleMXactId[] entry if we didn't already */
1147 184870 : MultiXactIdSetOldestVisible();
1148 :
1149 : /*
1150 : * If we know the multi is used only for locking and not for updates, then
1151 : * we can skip checking if the value is older than our oldest visible
1152 : * multi. It cannot possibly still be running.
1153 : */
1154 192224 : if (isLockOnly &&
1155 7354 : MultiXactIdPrecedes(multi, OldestVisibleMXactId[MyProcNumber]))
1156 : {
1157 : debug_elog2(DEBUG2, "GetMembers: a locker-only multi is too old");
1158 1434 : *members = NULL;
1159 1434 : return -1;
1160 : }
1161 :
1162 : /*
1163 : * We check known limits on MultiXact before resorting to the SLRU area.
1164 : *
1165 : * An ID older than MultiXactState->oldestMultiXactId cannot possibly be
1166 : * useful; it has already been removed, or will be removed shortly, by
1167 : * truncation. If one is passed, an error is raised.
1168 : *
1169 : * Also, an ID >= nextMXact shouldn't ever be seen here; if it is seen, it
1170 : * implies undetected ID wraparound has occurred. This raises a hard
1171 : * error.
1172 : *
1173 : * Shared lock is enough here since we aren't modifying any global state.
1174 : * Acquire it just long enough to grab the current counter values.
1175 : */
1176 183436 : LWLockAcquire(MultiXactGenLock, LW_SHARED);
1177 :
1178 183436 : oldestMXact = MultiXactState->oldestMultiXactId;
1179 183436 : nextMXact = MultiXactState->nextMXact;
1180 :
1181 183436 : LWLockRelease(MultiXactGenLock);
1182 :
1183 183436 : if (MultiXactIdPrecedes(multi, oldestMXact))
1184 0 : ereport(ERROR,
1185 : (errcode(ERRCODE_INTERNAL_ERROR),
1186 : errmsg("MultiXactId %u does no longer exist -- apparent wraparound",
1187 : multi)));
1188 :
1189 183436 : if (!MultiXactIdPrecedes(multi, nextMXact))
1190 0 : ereport(ERROR,
1191 : (errcode(ERRCODE_INTERNAL_ERROR),
1192 : errmsg("MultiXactId %u has not been created yet -- apparent wraparound",
1193 : multi)));
1194 :
1195 : /*
1196 : * Find out the offset at which we need to start reading MultiXactMembers
1197 : * and the number of members in the multixact. We determine the latter as
1198 : * the difference between this multixact's starting offset and the next
1199 : * one's.
1200 : */
1201 183436 : pageno = MultiXactIdToOffsetPage(multi);
1202 183436 : entryno = MultiXactIdToOffsetEntry(multi);
1203 :
1204 : /* Acquire the bank lock for the page we need. */
1205 183436 : lock = SimpleLruGetBankLock(MultiXactOffsetCtl, pageno);
1206 183436 : LWLockAcquire(lock, LW_EXCLUSIVE);
1207 :
1208 : /* read this multi's offset */
1209 183436 : slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, multi);
1210 183436 : offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
1211 183436 : offptr += entryno;
1212 183436 : offset = *offptr;
1213 :
1214 183436 : if (offset == 0)
1215 0 : ereport(ERROR,
1216 : (errcode(ERRCODE_DATA_CORRUPTED),
1217 : errmsg("MultiXact %u has invalid offset", multi)));
1218 :
1219 : /* read next multi's offset */
1220 : {
1221 : MultiXactId tmpMXact;
1222 :
1223 : /* handle wraparound if needed */
1224 183436 : tmpMXact = NextMultiXactId(multi);
1225 :
1226 183436 : prev_pageno = pageno;
1227 :
1228 183436 : pageno = MultiXactIdToOffsetPage(tmpMXact);
1229 183436 : entryno = MultiXactIdToOffsetEntry(tmpMXact);
1230 :
1231 183436 : if (pageno != prev_pageno)
1232 : {
1233 : LWLock *newlock;
1234 :
1235 : /*
1236 : * Since we're going to access a different SLRU page, if this page
1237 : * falls under a different bank, release the old bank's lock and
1238 : * acquire the lock of the new bank.
1239 : */
1240 26 : newlock = SimpleLruGetBankLock(MultiXactOffsetCtl, pageno);
1241 26 : if (newlock != lock)
1242 : {
1243 0 : LWLockRelease(lock);
1244 0 : LWLockAcquire(newlock, LW_EXCLUSIVE);
1245 0 : lock = newlock;
1246 : }
1247 26 : slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, tmpMXact);
1248 : }
1249 :
1250 183436 : offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
1251 183436 : offptr += entryno;
1252 183436 : nextMXOffset = *offptr;
1253 : }
1254 :
1255 183436 : LWLockRelease(lock);
1256 183436 : lock = NULL;
1257 :
1258 : /* Sanity check the next offset */
1259 183436 : if (nextMXOffset == 0)
1260 0 : ereport(ERROR,
1261 : (errcode(ERRCODE_DATA_CORRUPTED),
1262 : errmsg("MultiXact %u has invalid next offset", multi)));
1263 183436 : if (nextMXOffset == offset)
1264 0 : ereport(ERROR,
1265 : (errcode(ERRCODE_DATA_CORRUPTED),
1266 : errmsg("MultiXact %u with offset (%" PRIu64 ") has zero members",
1267 : multi, offset)));
1268 183436 : if (nextMXOffset < offset)
1269 0 : ereport(ERROR,
1270 : (errcode(ERRCODE_DATA_CORRUPTED),
1271 : errmsg("MultiXact %u has offset (%" PRIu64 ") greater than its next offset (%" PRIu64 ")",
1272 : multi, offset, nextMXOffset)));
1273 183436 : if (nextMXOffset - offset > INT32_MAX)
1274 0 : ereport(ERROR,
1275 : (errcode(ERRCODE_DATA_CORRUPTED),
1276 : errmsg("MultiXact %u has too many members (%" PRIu64 ")",
1277 : multi, nextMXOffset - offset)));
1278 183436 : length = nextMXOffset - offset;
1279 :
1280 : /* read the members */
1281 183436 : ptr = (MultiXactMember *) palloc(length * sizeof(MultiXactMember));
1282 183436 : prev_pageno = -1;
1283 3655690 : for (int i = 0; i < length; i++, offset++)
1284 : {
1285 : TransactionId *xactptr;
1286 : uint32 *flagsptr;
1287 : int flagsoff;
1288 : int bshift;
1289 : int memberoff;
1290 :
1291 3472254 : pageno = MXOffsetToMemberPage(offset);
1292 3472254 : memberoff = MXOffsetToMemberOffset(offset);
1293 :
1294 3472254 : if (pageno != prev_pageno)
1295 : {
1296 : LWLock *newlock;
1297 :
1298 : /*
1299 : * Since we're going to access a different SLRU page, if this page
1300 : * falls under a different bank, release the old bank's lock and
1301 : * acquire the lock of the new bank.
1302 : */
1303 183760 : newlock = SimpleLruGetBankLock(MultiXactMemberCtl, pageno);
1304 183760 : if (newlock != lock)
1305 : {
1306 183760 : if (lock)
1307 324 : LWLockRelease(lock);
1308 183760 : LWLockAcquire(newlock, LW_EXCLUSIVE);
1309 183760 : lock = newlock;
1310 : }
1311 :
1312 183760 : slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, multi);
1313 183760 : prev_pageno = pageno;
1314 : }
1315 :
1316 3472254 : xactptr = (TransactionId *)
1317 3472254 : (MultiXactMemberCtl->shared->page_buffer[slotno] + memberoff);
1318 : Assert(TransactionIdIsValid(*xactptr));
1319 :
1320 3472254 : flagsoff = MXOffsetToFlagsOffset(offset);
1321 3472254 : bshift = MXOffsetToFlagsBitShift(offset);
1322 3472254 : flagsptr = (uint32 *) (MultiXactMemberCtl->shared->page_buffer[slotno] + flagsoff);
1323 :
1324 3472254 : ptr[i].xid = *xactptr;
1325 3472254 : ptr[i].status = (*flagsptr >> bshift) & MXACT_MEMBER_XACT_BITMASK;
1326 : }
1327 :
1328 183436 : LWLockRelease(lock);
1329 :
1330 : /*
1331 : * Copy the result into the local cache.
1332 : */
1333 183436 : mXactCachePut(multi, length, ptr);
1334 :
1335 : debug_elog3(DEBUG2, "GetMembers: no cache for %s",
1336 : mxid_to_string(multi, length, ptr));
1337 183436 : *members = ptr;
1338 183436 : return length;
1339 : }
1340 :
1341 : /*
1342 : * mxactMemberComparator
1343 : * qsort comparison function for MultiXactMember
1344 : *
1345 : * We can't use wraparound comparison for XIDs because that does not respect
1346 : * the triangle inequality! Any old sort order will do.
1347 : */
1348 : static int
1349 6101280 : mxactMemberComparator(const void *arg1, const void *arg2)
1350 : {
1351 6101280 : MultiXactMember member1 = *(const MultiXactMember *) arg1;
1352 6101280 : MultiXactMember member2 = *(const MultiXactMember *) arg2;
1353 :
1354 6101280 : if (member1.xid > member2.xid)
1355 54 : return 1;
1356 6101226 : if (member1.xid < member2.xid)
1357 6100800 : return -1;
1358 426 : if (member1.status > member2.status)
1359 32 : return 1;
1360 394 : if (member1.status < member2.status)
1361 394 : return -1;
1362 0 : return 0;
1363 : }
1364 :
1365 : /*
1366 : * mXactCacheGetBySet
1367 : * returns a MultiXactId from the cache based on the set of
1368 : * TransactionIds that compose it, or InvalidMultiXactId if
1369 : * none matches.
1370 : *
1371 : * This is helpful, for example, if two transactions want to lock a huge
1372 : * table. By using the cache, the second will use the same MultiXactId
1373 : * for the majority of tuples, thus keeping MultiXactId usage low (saving
1374 : * both I/O and wraparound issues).
1375 : *
1376 : * NB: the passed members array will be sorted in-place.
1377 : */
1378 : static MultiXactId
1379 153250 : mXactCacheGetBySet(int nmembers, MultiXactMember *members)
1380 : {
1381 : dlist_iter iter;
1382 :
1383 : debug_elog3(DEBUG2, "CacheGet: looking for %s",
1384 : mxid_to_string(InvalidMultiXactId, nmembers, members));
1385 :
1386 : /* sort the array so comparison is easy */
1387 153250 : qsort(members, nmembers, sizeof(MultiXactMember), mxactMemberComparator);
1388 :
1389 616598 : dclist_foreach(iter, &MXactCache)
1390 : {
1391 606028 : mXactCacheEnt *entry = dclist_container(mXactCacheEnt, node,
1392 : iter.cur);
1393 :
1394 606028 : if (entry->nmembers != nmembers)
1395 170656 : continue;
1396 :
1397 : /*
1398 : * We assume the cache entries are sorted, and that the unused bits in
1399 : * "status" are zeroed.
1400 : */
1401 435372 : if (memcmp(members, entry->members, nmembers * sizeof(MultiXactMember)) == 0)
1402 : {
1403 : debug_elog3(DEBUG2, "CacheGet: found %u", entry->multi);
1404 142680 : dclist_move_head(&MXactCache, iter.cur);
1405 142680 : return entry->multi;
1406 : }
1407 : }
1408 :
1409 : debug_elog2(DEBUG2, "CacheGet: not found :-(");
1410 10570 : return InvalidMultiXactId;
1411 : }
1412 :
1413 : /*
1414 : * mXactCacheGetById
1415 : * returns the composing MultiXactMember set from the cache for a
1416 : * given MultiXactId, if present.
1417 : *
1418 : * If successful, *xids is set to the address of a palloc'd copy of the
1419 : * MultiXactMember set. Return value is number of members, or -1 on failure.
1420 : */
1421 : static int
1422 1099618 : mXactCacheGetById(MultiXactId multi, MultiXactMember **members)
1423 : {
1424 : dlist_iter iter;
1425 :
1426 : debug_elog3(DEBUG2, "CacheGet: looking for %u", multi);
1427 :
1428 9811202 : dclist_foreach(iter, &MXactCache)
1429 : {
1430 9626332 : mXactCacheEnt *entry = dclist_container(mXactCacheEnt, node,
1431 : iter.cur);
1432 :
1433 9626332 : if (entry->multi == multi)
1434 : {
1435 : MultiXactMember *ptr;
1436 : Size size;
1437 :
1438 914748 : size = sizeof(MultiXactMember) * entry->nmembers;
1439 914748 : ptr = (MultiXactMember *) palloc(size);
1440 :
1441 914748 : memcpy(ptr, entry->members, size);
1442 :
1443 : debug_elog3(DEBUG2, "CacheGet: found %s",
1444 : mxid_to_string(multi,
1445 : entry->nmembers,
1446 : entry->members));
1447 :
1448 : /*
1449 : * Note we modify the list while not using a modifiable iterator.
1450 : * This is acceptable only because we exit the iteration
1451 : * immediately afterwards.
1452 : */
1453 914748 : dclist_move_head(&MXactCache, iter.cur);
1454 :
1455 914748 : *members = ptr;
1456 914748 : return entry->nmembers;
1457 : }
1458 : }
1459 :
1460 : debug_elog2(DEBUG2, "CacheGet: not found");
1461 184870 : return -1;
1462 : }
1463 :
1464 : /*
1465 : * mXactCachePut
1466 : * Add a new MultiXactId and its composing set into the local cache.
1467 : */
1468 : static void
1469 194006 : mXactCachePut(MultiXactId multi, int nmembers, MultiXactMember *members)
1470 : {
1471 : mXactCacheEnt *entry;
1472 :
1473 : debug_elog3(DEBUG2, "CachePut: storing %s",
1474 : mxid_to_string(multi, nmembers, members));
1475 :
1476 194006 : if (MXactContext == NULL)
1477 : {
1478 : /* The cache only lives as long as the current transaction */
1479 : debug_elog2(DEBUG2, "CachePut: initializing memory context");
1480 6634 : MXactContext = AllocSetContextCreate(TopTransactionContext,
1481 : "MultiXact cache context",
1482 : ALLOCSET_SMALL_SIZES);
1483 : }
1484 :
1485 : entry = (mXactCacheEnt *)
1486 194006 : MemoryContextAlloc(MXactContext,
1487 194006 : offsetof(mXactCacheEnt, members) +
1488 : nmembers * sizeof(MultiXactMember));
1489 :
1490 194006 : entry->multi = multi;
1491 194006 : entry->nmembers = nmembers;
1492 194006 : memcpy(entry->members, members, nmembers * sizeof(MultiXactMember));
1493 :
1494 : /* mXactCacheGetBySet assumes the entries are sorted, so sort them */
1495 194006 : qsort(entry->members, nmembers, sizeof(MultiXactMember), mxactMemberComparator);
1496 :
1497 194006 : dclist_push_head(&MXactCache, &entry->node);
1498 194006 : if (dclist_count(&MXactCache) > MAX_CACHE_ENTRIES)
1499 : {
1500 : dlist_node *node;
1501 :
1502 18956 : node = dclist_tail_node(&MXactCache);
1503 18956 : dclist_delete_from(&MXactCache, node);
1504 :
1505 18956 : entry = dclist_container(mXactCacheEnt, node, node);
1506 : debug_elog3(DEBUG2, "CachePut: pruning cached multi %u",
1507 : entry->multi);
1508 :
1509 18956 : pfree(entry);
1510 : }
1511 194006 : }
1512 :
1513 : char *
1514 376176 : mxstatus_to_string(MultiXactStatus status)
1515 : {
1516 376176 : switch (status)
1517 : {
1518 367268 : case MultiXactStatusForKeyShare:
1519 367268 : return "keysh";
1520 0 : case MultiXactStatusForShare:
1521 0 : return "sh";
1522 0 : case MultiXactStatusForNoKeyUpdate:
1523 0 : return "fornokeyupd";
1524 0 : case MultiXactStatusForUpdate:
1525 0 : return "forupd";
1526 8908 : case MultiXactStatusNoKeyUpdate:
1527 8908 : return "nokeyupd";
1528 0 : case MultiXactStatusUpdate:
1529 0 : return "upd";
1530 0 : default:
1531 0 : elog(ERROR, "unrecognized multixact status %d", status);
1532 : return "";
1533 : }
1534 : }
1535 :
1536 : char *
1537 0 : mxid_to_string(MultiXactId multi, int nmembers, MultiXactMember *members)
1538 : {
1539 : static char *str = NULL;
1540 : StringInfoData buf;
1541 : int i;
1542 :
1543 0 : if (str != NULL)
1544 0 : pfree(str);
1545 :
1546 0 : initStringInfo(&buf);
1547 :
1548 0 : appendStringInfo(&buf, "%u %d[%u (%s)", multi, nmembers, members[0].xid,
1549 : mxstatus_to_string(members[0].status));
1550 :
1551 0 : for (i = 1; i < nmembers; i++)
1552 0 : appendStringInfo(&buf, ", %u (%s)", members[i].xid,
1553 0 : mxstatus_to_string(members[i].status));
1554 :
1555 0 : appendStringInfoChar(&buf, ']');
1556 0 : str = MemoryContextStrdup(TopMemoryContext, buf.data);
1557 0 : pfree(buf.data);
1558 0 : return str;
1559 : }
1560 :
1561 : /*
1562 : * AtEOXact_MultiXact
1563 : * Handle transaction end for MultiXact
1564 : *
1565 : * This is called at top transaction commit or abort (we don't care which).
1566 : */
1567 : void
1568 999658 : AtEOXact_MultiXact(void)
1569 : {
1570 : /*
1571 : * Reset our OldestMemberMXactId and OldestVisibleMXactId values, both of
1572 : * which should only be valid while within a transaction.
1573 : *
1574 : * We assume that storing a MultiXactId is atomic and so we need not take
1575 : * MultiXactGenLock to do this.
1576 : */
1577 999658 : OldestMemberMXactId[MyProcNumber] = InvalidMultiXactId;
1578 999658 : OldestVisibleMXactId[MyProcNumber] = InvalidMultiXactId;
1579 :
1580 : /*
1581 : * Discard the local MultiXactId cache. Since MXactContext was created as
1582 : * a child of TopTransactionContext, we needn't delete it explicitly.
1583 : */
1584 999658 : MXactContext = NULL;
1585 999658 : dclist_init(&MXactCache);
1586 999658 : }
1587 :
1588 : /*
1589 : * AtPrepare_MultiXact
1590 : * Save multixact state at 2PC transaction prepare
1591 : *
1592 : * In this phase, we only store our OldestMemberMXactId value in the two-phase
1593 : * state file.
1594 : */
1595 : void
1596 634 : AtPrepare_MultiXact(void)
1597 : {
1598 634 : MultiXactId myOldestMember = OldestMemberMXactId[MyProcNumber];
1599 :
1600 634 : if (MultiXactIdIsValid(myOldestMember))
1601 126 : RegisterTwoPhaseRecord(TWOPHASE_RM_MULTIXACT_ID, 0,
1602 : &myOldestMember, sizeof(MultiXactId));
1603 634 : }
1604 :
1605 : /*
1606 : * PostPrepare_MultiXact
1607 : * Clean up after successful PREPARE TRANSACTION
1608 : */
1609 : void
1610 634 : PostPrepare_MultiXact(FullTransactionId fxid)
1611 : {
1612 : MultiXactId myOldestMember;
1613 :
1614 : /*
1615 : * Transfer our OldestMemberMXactId value to the slot reserved for the
1616 : * prepared transaction.
1617 : */
1618 634 : myOldestMember = OldestMemberMXactId[MyProcNumber];
1619 634 : if (MultiXactIdIsValid(myOldestMember))
1620 : {
1621 126 : ProcNumber dummyProcNumber = TwoPhaseGetDummyProcNumber(fxid, false);
1622 :
1623 : /*
1624 : * Even though storing MultiXactId is atomic, acquire lock to make
1625 : * sure others see both changes, not just the reset of the slot of the
1626 : * current backend. Using a volatile pointer might suffice, but this
1627 : * isn't a hot spot.
1628 : */
1629 126 : LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
1630 :
1631 126 : OldestMemberMXactId[dummyProcNumber] = myOldestMember;
1632 126 : OldestMemberMXactId[MyProcNumber] = InvalidMultiXactId;
1633 :
1634 126 : LWLockRelease(MultiXactGenLock);
1635 : }
1636 :
1637 : /*
1638 : * We don't need to transfer OldestVisibleMXactId value, because the
1639 : * transaction is not going to be looking at any more multixacts once it's
1640 : * prepared.
1641 : *
1642 : * We assume that storing a MultiXactId is atomic and so we need not take
1643 : * MultiXactGenLock to do this.
1644 : */
1645 634 : OldestVisibleMXactId[MyProcNumber] = InvalidMultiXactId;
1646 :
1647 : /*
1648 : * Discard the local MultiXactId cache like in AtEOXact_MultiXact.
1649 : */
1650 634 : MXactContext = NULL;
1651 634 : dclist_init(&MXactCache);
1652 634 : }
1653 :
1654 : /*
1655 : * multixact_twophase_recover
1656 : * Recover the state of a prepared transaction at startup
1657 : */
1658 : void
1659 16 : multixact_twophase_recover(FullTransactionId fxid, uint16 info,
1660 : void *recdata, uint32 len)
1661 : {
1662 16 : ProcNumber dummyProcNumber = TwoPhaseGetDummyProcNumber(fxid, false);
1663 : MultiXactId oldestMember;
1664 :
1665 : /*
1666 : * Get the oldest member XID from the state file record, and set it in the
1667 : * OldestMemberMXactId slot reserved for this prepared transaction.
1668 : */
1669 : Assert(len == sizeof(MultiXactId));
1670 16 : oldestMember = *((MultiXactId *) recdata);
1671 :
1672 16 : OldestMemberMXactId[dummyProcNumber] = oldestMember;
1673 16 : }
1674 :
1675 : /*
1676 : * multixact_twophase_postcommit
1677 : * Similar to AtEOXact_MultiXact but for COMMIT PREPARED
1678 : */
1679 : void
1680 136 : multixact_twophase_postcommit(FullTransactionId fxid, uint16 info,
1681 : void *recdata, uint32 len)
1682 : {
1683 136 : ProcNumber dummyProcNumber = TwoPhaseGetDummyProcNumber(fxid, true);
1684 :
1685 : Assert(len == sizeof(MultiXactId));
1686 :
1687 136 : OldestMemberMXactId[dummyProcNumber] = InvalidMultiXactId;
1688 136 : }
1689 :
1690 : /*
1691 : * multixact_twophase_postabort
1692 : * This is actually just the same as the COMMIT case.
1693 : */
1694 : void
1695 52 : multixact_twophase_postabort(FullTransactionId fxid, uint16 info,
1696 : void *recdata, uint32 len)
1697 : {
1698 52 : multixact_twophase_postcommit(fxid, info, recdata, len);
1699 52 : }
1700 :
1701 : /*
1702 : * Initialization of shared memory for MultiXact. We use two SLRU areas,
1703 : * thus double memory. Also, reserve space for the shared MultiXactState
1704 : * struct and the per-backend MultiXactId arrays (two of those, too).
1705 : */
1706 : Size
1707 4254 : MultiXactShmemSize(void)
1708 : {
1709 : Size size;
1710 :
1711 : /* We need 2*MaxOldestSlot perBackendXactIds[] entries */
1712 : #define SHARED_MULTIXACT_STATE_SIZE \
1713 : add_size(offsetof(MultiXactStateData, perBackendXactIds), \
1714 : mul_size(sizeof(MultiXactId) * 2, MaxOldestSlot))
1715 :
1716 4254 : size = SHARED_MULTIXACT_STATE_SIZE;
1717 4254 : size = add_size(size, SimpleLruShmemSize(multixact_offset_buffers, 0));
1718 4254 : size = add_size(size, SimpleLruShmemSize(multixact_member_buffers, 0));
1719 :
1720 4254 : return size;
1721 : }
1722 :
1723 : void
1724 2280 : MultiXactShmemInit(void)
1725 : {
1726 : bool found;
1727 :
1728 : debug_elog2(DEBUG2, "Shared Memory Init for MultiXact");
1729 :
1730 2280 : MultiXactOffsetCtl->PagePrecedes = MultiXactOffsetPagePrecedes;
1731 2280 : MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
1732 :
1733 2280 : SimpleLruInit(MultiXactOffsetCtl,
1734 : "multixact_offset", multixact_offset_buffers, 0,
1735 : "pg_multixact/offsets", LWTRANCHE_MULTIXACTOFFSET_BUFFER,
1736 : LWTRANCHE_MULTIXACTOFFSET_SLRU,
1737 : SYNC_HANDLER_MULTIXACT_OFFSET,
1738 : false);
1739 : SlruPagePrecedesUnitTests(MultiXactOffsetCtl, MULTIXACT_OFFSETS_PER_PAGE);
1740 2280 : SimpleLruInit(MultiXactMemberCtl,
1741 : "multixact_member", multixact_member_buffers, 0,
1742 : "pg_multixact/members", LWTRANCHE_MULTIXACTMEMBER_BUFFER,
1743 : LWTRANCHE_MULTIXACTMEMBER_SLRU,
1744 : SYNC_HANDLER_MULTIXACT_MEMBER,
1745 : true);
1746 : /* doesn't call SimpleLruTruncate() or meet criteria for unit tests */
1747 :
1748 : /* Initialize our shared state struct */
1749 2280 : MultiXactState = ShmemInitStruct("Shared MultiXact State",
1750 2280 : SHARED_MULTIXACT_STATE_SIZE,
1751 : &found);
1752 2280 : if (!IsUnderPostmaster)
1753 : {
1754 : Assert(!found);
1755 :
1756 : /* Make sure we zero out the per-backend state */
1757 39038 : MemSet(MultiXactState, 0, SHARED_MULTIXACT_STATE_SIZE);
1758 : }
1759 : else
1760 : Assert(found);
1761 :
1762 : /*
1763 : * Set up array pointers.
1764 : */
1765 2280 : OldestMemberMXactId = MultiXactState->perBackendXactIds;
1766 2280 : OldestVisibleMXactId = OldestMemberMXactId + MaxOldestSlot;
1767 2280 : }
1768 :
1769 : /*
1770 : * GUC check_hook for multixact_offset_buffers
1771 : */
1772 : bool
1773 2356 : check_multixact_offset_buffers(int *newval, void **extra, GucSource source)
1774 : {
1775 2356 : return check_slru_buffers("multixact_offset_buffers", newval);
1776 : }
1777 :
1778 : /*
1779 : * GUC check_hook for multixact_member_buffers
1780 : */
1781 : bool
1782 2356 : check_multixact_member_buffers(int *newval, void **extra, GucSource source)
1783 : {
1784 2356 : return check_slru_buffers("multixact_member_buffers", newval);
1785 : }
1786 :
1787 : /*
1788 : * This func must be called ONCE on system install. It creates the initial
1789 : * MultiXact segments. (The MultiXacts directories are assumed to have been
1790 : * created by initdb, and MultiXactShmemInit must have been called already.)
1791 : */
1792 : void
1793 102 : BootStrapMultiXact(void)
1794 : {
1795 : /* Zero the initial pages and flush them to disk */
1796 102 : SimpleLruZeroAndWritePage(MultiXactOffsetCtl, 0);
1797 102 : SimpleLruZeroAndWritePage(MultiXactMemberCtl, 0);
1798 102 : }
1799 :
1800 : /*
1801 : * This must be called ONCE during postmaster or standalone-backend startup.
1802 : *
1803 : * StartupXLOG has already established nextMXact/nextOffset by calling
1804 : * MultiXactSetNextMXact and/or MultiXactAdvanceNextMXact, and the oldestMulti
1805 : * info from pg_control and/or MultiXactAdvanceOldest, but we haven't yet
1806 : * replayed WAL.
1807 : */
1808 : void
1809 1982 : StartupMultiXact(void)
1810 : {
1811 1982 : MultiXactId multi = MultiXactState->nextMXact;
1812 1982 : MultiXactOffset offset = MultiXactState->nextOffset;
1813 : int64 pageno;
1814 :
1815 : /*
1816 : * Initialize offset's idea of the latest page number.
1817 : */
1818 1982 : pageno = MultiXactIdToOffsetPage(multi);
1819 1982 : pg_atomic_write_u64(&MultiXactOffsetCtl->shared->latest_page_number,
1820 : pageno);
1821 :
1822 : /*
1823 : * Initialize member's idea of the latest page number.
1824 : */
1825 1982 : pageno = MXOffsetToMemberPage(offset);
1826 1982 : pg_atomic_write_u64(&MultiXactMemberCtl->shared->latest_page_number,
1827 : pageno);
1828 1982 : }
1829 :
1830 : /*
1831 : * This must be called ONCE at the end of startup/recovery.
1832 : */
1833 : void
1834 1860 : TrimMultiXact(void)
1835 : {
1836 : MultiXactId nextMXact;
1837 : MultiXactOffset offset;
1838 : MultiXactId oldestMXact;
1839 : Oid oldestMXactDB;
1840 : int64 pageno;
1841 : int entryno;
1842 : int flagsoff;
1843 :
1844 1860 : LWLockAcquire(MultiXactGenLock, LW_SHARED);
1845 1860 : nextMXact = MultiXactState->nextMXact;
1846 1860 : offset = MultiXactState->nextOffset;
1847 1860 : oldestMXact = MultiXactState->oldestMultiXactId;
1848 1860 : oldestMXactDB = MultiXactState->oldestMultiXactDB;
1849 1860 : LWLockRelease(MultiXactGenLock);
1850 :
1851 : /* Clean up offsets state */
1852 :
1853 : /*
1854 : * (Re-)Initialize our idea of the latest page number for offsets.
1855 : */
1856 1860 : pageno = MultiXactIdToOffsetPage(nextMXact);
1857 1860 : pg_atomic_write_u64(&MultiXactOffsetCtl->shared->latest_page_number,
1858 : pageno);
1859 :
1860 : /*
1861 : * Set the offset of nextMXact on the offsets page. This is normally done
1862 : * in RecordNewMultiXact() of the previous multixact, but let's be sure
1863 : * the next page exists, if the nextMXact was reset with pg_resetwal for
1864 : * example.
1865 : *
1866 : * Zero out the remainder of the page. See notes in TrimCLOG() for
1867 : * background. Unlike CLOG, some WAL record covers every pg_multixact
1868 : * SLRU mutation. Since, also unlike CLOG, we ignore the WAL rule "write
1869 : * xlog before data," nextMXact successors may carry obsolete, nonzero
1870 : * offset values.
1871 : */
1872 1860 : entryno = MultiXactIdToOffsetEntry(nextMXact);
1873 : {
1874 : int slotno;
1875 : MultiXactOffset *offptr;
1876 1860 : LWLock *lock = SimpleLruGetBankLock(MultiXactOffsetCtl, pageno);
1877 :
1878 1860 : LWLockAcquire(lock, LW_EXCLUSIVE);
1879 1860 : if (entryno == 0 || nextMXact == FirstMultiXactId)
1880 1824 : slotno = SimpleLruZeroPage(MultiXactOffsetCtl, pageno);
1881 : else
1882 36 : slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, nextMXact);
1883 1860 : offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
1884 1860 : offptr += entryno;
1885 :
1886 1860 : *offptr = offset;
1887 1860 : if (entryno != 0 && (entryno + 1) * sizeof(MultiXactOffset) != BLCKSZ)
1888 3102 : MemSet(offptr + 1, 0, BLCKSZ - (entryno + 1) * sizeof(MultiXactOffset));
1889 :
1890 1860 : MultiXactOffsetCtl->shared->page_dirty[slotno] = true;
1891 1860 : LWLockRelease(lock);
1892 : }
1893 :
1894 : /*
1895 : * And the same for members.
1896 : *
1897 : * (Re-)Initialize our idea of the latest page number for members.
1898 : */
1899 1860 : pageno = MXOffsetToMemberPage(offset);
1900 1860 : pg_atomic_write_u64(&MultiXactMemberCtl->shared->latest_page_number,
1901 : pageno);
1902 :
1903 : /*
1904 : * Zero out the remainder of the current members page. See notes in
1905 : * TrimCLOG() for motivation.
1906 : */
1907 1860 : flagsoff = MXOffsetToFlagsOffset(offset);
1908 1860 : if (flagsoff != 0)
1909 : {
1910 : int slotno;
1911 : TransactionId *xidptr;
1912 : int memberoff;
1913 34 : LWLock *lock = SimpleLruGetBankLock(MultiXactMemberCtl, pageno);
1914 :
1915 34 : LWLockAcquire(lock, LW_EXCLUSIVE);
1916 34 : memberoff = MXOffsetToMemberOffset(offset);
1917 34 : slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, offset);
1918 34 : xidptr = (TransactionId *)
1919 34 : (MultiXactMemberCtl->shared->page_buffer[slotno] + memberoff);
1920 :
1921 34 : MemSet(xidptr, 0, BLCKSZ - memberoff);
1922 :
1923 : /*
1924 : * Note: we don't need to zero out the flag bits in the remaining
1925 : * members of the current group, because they are always reset before
1926 : * writing.
1927 : */
1928 :
1929 34 : MultiXactMemberCtl->shared->page_dirty[slotno] = true;
1930 34 : LWLockRelease(lock);
1931 : }
1932 :
1933 : /* signal that we're officially up */
1934 1860 : LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
1935 1860 : MultiXactState->finishedStartup = true;
1936 1860 : LWLockRelease(MultiXactGenLock);
1937 :
1938 : /* Now compute how far away the next multixid wraparound is. */
1939 1860 : SetMultiXactIdLimit(oldestMXact, oldestMXactDB);
1940 1860 : }
1941 :
1942 : /*
1943 : * Get the MultiXact data to save in a checkpoint record
1944 : */
1945 : void
1946 3184 : MultiXactGetCheckptMulti(bool is_shutdown,
1947 : MultiXactId *nextMulti,
1948 : MultiXactOffset *nextMultiOffset,
1949 : MultiXactId *oldestMulti,
1950 : Oid *oldestMultiDB)
1951 : {
1952 3184 : LWLockAcquire(MultiXactGenLock, LW_SHARED);
1953 3184 : *nextMulti = MultiXactState->nextMXact;
1954 3184 : *nextMultiOffset = MultiXactState->nextOffset;
1955 3184 : *oldestMulti = MultiXactState->oldestMultiXactId;
1956 3184 : *oldestMultiDB = MultiXactState->oldestMultiXactDB;
1957 3184 : LWLockRelease(MultiXactGenLock);
1958 :
1959 : debug_elog6(DEBUG2,
1960 : "MultiXact: checkpoint is nextMulti %u, nextOffset %" PRIu64 ", oldestMulti %u in DB %u",
1961 : *nextMulti, *nextMultiOffset, *oldestMulti, *oldestMultiDB);
1962 3184 : }
1963 :
1964 : /*
1965 : * Perform a checkpoint --- either during shutdown, or on-the-fly
1966 : */
1967 : void
1968 3568 : CheckPointMultiXact(void)
1969 : {
1970 : TRACE_POSTGRESQL_MULTIXACT_CHECKPOINT_START(true);
1971 :
1972 : /*
1973 : * Write dirty MultiXact pages to disk. This may result in sync requests
1974 : * queued for later handling by ProcessSyncRequests(), as part of the
1975 : * checkpoint.
1976 : */
1977 3568 : SimpleLruWriteAll(MultiXactOffsetCtl, true);
1978 3568 : SimpleLruWriteAll(MultiXactMemberCtl, true);
1979 :
1980 : TRACE_POSTGRESQL_MULTIXACT_CHECKPOINT_DONE(true);
1981 3568 : }
1982 :
1983 : /*
1984 : * Set the next-to-be-assigned MultiXactId and offset
1985 : *
1986 : * This is used when we can determine the correct next ID/offset exactly
1987 : * from a checkpoint record. Although this is only called during bootstrap
1988 : * and XLog replay, we take the lock in case any hot-standby backends are
1989 : * examining the values.
1990 : */
1991 : void
1992 2166 : MultiXactSetNextMXact(MultiXactId nextMulti,
1993 : MultiXactOffset nextMultiOffset)
1994 : {
1995 : Assert(MultiXactIdIsValid(nextMulti));
1996 : debug_elog4(DEBUG2, "MultiXact: setting next multi to %u offset %" PRIu64,
1997 : nextMulti, nextMultiOffset);
1998 :
1999 2166 : LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2000 2166 : MultiXactState->nextMXact = nextMulti;
2001 2166 : MultiXactState->nextOffset = nextMultiOffset;
2002 2166 : LWLockRelease(MultiXactGenLock);
2003 2166 : }
2004 :
2005 : /*
2006 : * Determine the last safe MultiXactId to allocate given the currently oldest
2007 : * datminmxid (ie, the oldest MultiXactId that might exist in any database
2008 : * of our cluster), and the OID of the (or a) database with that value.
2009 : *
2010 : * This also updates MultiXactState->oldestOffset, by looking up the offset of
2011 : * MultiXactState->oldestMultiXactId.
2012 : */
2013 : void
2014 5796 : SetMultiXactIdLimit(MultiXactId oldest_datminmxid, Oid oldest_datoid)
2015 : {
2016 : MultiXactId multiVacLimit;
2017 : MultiXactId multiWarnLimit;
2018 : MultiXactId multiStopLimit;
2019 : MultiXactId multiWrapLimit;
2020 : MultiXactId curMulti;
2021 :
2022 : Assert(MultiXactIdIsValid(oldest_datminmxid));
2023 :
2024 : /*
2025 : * We pretend that a wrap will happen halfway through the multixact ID
2026 : * space, but that's not really true, because multixacts wrap differently
2027 : * from transaction IDs.
2028 : */
2029 5796 : multiWrapLimit = oldest_datminmxid + (MaxMultiXactId >> 1);
2030 5796 : if (multiWrapLimit < FirstMultiXactId)
2031 0 : multiWrapLimit += FirstMultiXactId;
2032 :
2033 : /*
2034 : * We'll refuse to continue assigning MultiXactIds once we get within 3M
2035 : * multi of data loss. See SetTransactionIdLimit.
2036 : */
2037 5796 : multiStopLimit = multiWrapLimit - 3000000;
2038 5796 : if (multiStopLimit < FirstMultiXactId)
2039 0 : multiStopLimit -= FirstMultiXactId;
2040 :
2041 : /*
2042 : * We'll start complaining loudly when we get within 40M multis of data
2043 : * loss. This is kind of arbitrary, but if you let your gas gauge get
2044 : * down to 2% of full, would you be looking for the next gas station? We
2045 : * need to be fairly liberal about this number because there are lots of
2046 : * scenarios where most transactions are done by automatic clients that
2047 : * won't pay attention to warnings. (No, we're not gonna make this
2048 : * configurable. If you know enough to configure it, you know enough to
2049 : * not get in this kind of trouble in the first place.)
2050 : */
2051 5796 : multiWarnLimit = multiWrapLimit - 40000000;
2052 5796 : if (multiWarnLimit < FirstMultiXactId)
2053 0 : multiWarnLimit -= FirstMultiXactId;
2054 :
2055 : /*
2056 : * We'll start trying to force autovacuums when oldest_datminmxid gets to
2057 : * be more than autovacuum_multixact_freeze_max_age mxids old.
2058 : *
2059 : * Note: autovacuum_multixact_freeze_max_age is a PGC_POSTMASTER parameter
2060 : * so that we don't have to worry about dealing with on-the-fly changes in
2061 : * its value. See SetTransactionIdLimit.
2062 : */
2063 5796 : multiVacLimit = oldest_datminmxid + autovacuum_multixact_freeze_max_age;
2064 5796 : if (multiVacLimit < FirstMultiXactId)
2065 0 : multiVacLimit += FirstMultiXactId;
2066 :
2067 : /* Grab lock for just long enough to set the new limit values */
2068 5796 : LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2069 5796 : MultiXactState->oldestMultiXactId = oldest_datminmxid;
2070 5796 : MultiXactState->oldestMultiXactDB = oldest_datoid;
2071 5796 : MultiXactState->multiVacLimit = multiVacLimit;
2072 5796 : MultiXactState->multiWarnLimit = multiWarnLimit;
2073 5796 : MultiXactState->multiStopLimit = multiStopLimit;
2074 5796 : MultiXactState->multiWrapLimit = multiWrapLimit;
2075 5796 : curMulti = MultiXactState->nextMXact;
2076 5796 : LWLockRelease(MultiXactGenLock);
2077 :
2078 : /* Log the info */
2079 5796 : ereport(DEBUG1,
2080 : (errmsg_internal("MultiXactId wrap limit is %u, limited by database with OID %u",
2081 : multiWrapLimit, oldest_datoid)));
2082 :
2083 : /*
2084 : * Computing the actual limits is only possible once the data directory is
2085 : * in a consistent state. There's no need to compute the limits while
2086 : * still replaying WAL - no decisions about new multis are made even
2087 : * though multixact creations might be replayed. So we'll only do further
2088 : * checks after TrimMultiXact() has been called.
2089 : */
2090 5796 : if (!MultiXactState->finishedStartup)
2091 2088 : return;
2092 :
2093 : Assert(!InRecovery);
2094 :
2095 : /*
2096 : * Offsets are 64-bits wide and never wrap around, so we don't need to
2097 : * consider them for emergency autovacuum purposes. But now that we're in
2098 : * a consistent state, determine MultiXactState->oldestOffset. It will be
2099 : * used to adjust the freezing cutoff, to keep the offsets disk usage in
2100 : * check.
2101 : */
2102 3708 : SetOldestOffset();
2103 :
2104 : /*
2105 : * If past the autovacuum force point, immediately signal an autovac
2106 : * request. The reason for this is that autovac only processes one
2107 : * database per invocation. Once it's finished cleaning up the oldest
2108 : * database, it'll call here, and we'll signal the postmaster to start
2109 : * another iteration immediately if there are still any old databases.
2110 : */
2111 3708 : if (MultiXactIdPrecedes(multiVacLimit, curMulti) && IsUnderPostmaster)
2112 0 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
2113 :
2114 : /* Give an immediate warning if past the wrap warn point */
2115 3708 : if (MultiXactIdPrecedes(multiWarnLimit, curMulti))
2116 : {
2117 : char *oldest_datname;
2118 :
2119 : /*
2120 : * We can be called when not inside a transaction, for example during
2121 : * StartupXLOG(). In such a case we cannot do database access, so we
2122 : * must just report the oldest DB's OID.
2123 : *
2124 : * Note: it's also possible that get_database_name fails and returns
2125 : * NULL, for example because the database just got dropped. We'll
2126 : * still warn, even though the warning might now be unnecessary.
2127 : */
2128 0 : if (IsTransactionState())
2129 0 : oldest_datname = get_database_name(oldest_datoid);
2130 : else
2131 0 : oldest_datname = NULL;
2132 :
2133 0 : if (oldest_datname)
2134 0 : ereport(WARNING,
2135 : (errmsg_plural("database \"%s\" must be vacuumed before %u more MultiXactId is used",
2136 : "database \"%s\" must be vacuumed before %u more MultiXactIds are used",
2137 : multiWrapLimit - curMulti,
2138 : oldest_datname,
2139 : multiWrapLimit - curMulti),
2140 : errhint("To avoid MultiXactId assignment failures, execute a database-wide VACUUM in that database.\n"
2141 : "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
2142 : else
2143 0 : ereport(WARNING,
2144 : (errmsg_plural("database with OID %u must be vacuumed before %u more MultiXactId is used",
2145 : "database with OID %u must be vacuumed before %u more MultiXactIds are used",
2146 : multiWrapLimit - curMulti,
2147 : oldest_datoid,
2148 : multiWrapLimit - curMulti),
2149 : errhint("To avoid MultiXactId assignment failures, execute a database-wide VACUUM in that database.\n"
2150 : "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
2151 : }
2152 : }
2153 :
2154 : /*
2155 : * Ensure the next-to-be-assigned MultiXactId is at least minMulti,
2156 : * and similarly nextOffset is at least minMultiOffset.
2157 : *
2158 : * This is used when we can determine minimum safe values from an XLog
2159 : * record (either an on-line checkpoint or an mxact creation log entry).
2160 : * Although this is only called during XLog replay, we take the lock in case
2161 : * any hot-standby backends are examining the values.
2162 : */
2163 : void
2164 1392 : MultiXactAdvanceNextMXact(MultiXactId minMulti,
2165 : MultiXactOffset minMultiOffset)
2166 : {
2167 : Assert(MultiXactIdIsValid(minMulti));
2168 :
2169 1392 : LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2170 1392 : if (MultiXactIdPrecedes(MultiXactState->nextMXact, minMulti))
2171 : {
2172 : debug_elog3(DEBUG2, "MultiXact: setting next multi to %u", minMulti);
2173 8 : MultiXactState->nextMXact = minMulti;
2174 : }
2175 1392 : if (MultiXactState->nextOffset < minMultiOffset)
2176 : {
2177 : debug_elog3(DEBUG2, "MultiXact: setting next offset to %" PRIu64,
2178 : minMultiOffset);
2179 8 : MultiXactState->nextOffset = minMultiOffset;
2180 : }
2181 1392 : LWLockRelease(MultiXactGenLock);
2182 1392 : }
2183 :
2184 : /*
2185 : * Update our oldestMultiXactId value, but only if it's more recent than what
2186 : * we had.
2187 : *
2188 : * This may only be called during WAL replay.
2189 : */
2190 : void
2191 1462 : MultiXactAdvanceOldest(MultiXactId oldestMulti, Oid oldestMultiDB)
2192 : {
2193 : Assert(InRecovery);
2194 :
2195 1462 : if (MultiXactIdPrecedes(MultiXactState->oldestMultiXactId, oldestMulti))
2196 0 : SetMultiXactIdLimit(oldestMulti, oldestMultiDB);
2197 1462 : }
2198 :
2199 : /*
2200 : * Make sure that MultiXactOffset has room for a newly-allocated MultiXactId.
2201 : *
2202 : * NB: this is called while holding MultiXactGenLock. We want it to be very
2203 : * fast most of the time; even when it's not so fast, no actual I/O need
2204 : * happen unless we're forced to write out a dirty log or xlog page to make
2205 : * room in shared memory.
2206 : */
2207 : static void
2208 10570 : ExtendMultiXactOffset(MultiXactId multi)
2209 : {
2210 : int64 pageno;
2211 : LWLock *lock;
2212 :
2213 : /*
2214 : * No work except at first MultiXactId of a page. But beware: just after
2215 : * wraparound, the first MultiXactId of page zero is FirstMultiXactId.
2216 : */
2217 10570 : if (MultiXactIdToOffsetEntry(multi) != 0 &&
2218 : multi != FirstMultiXactId)
2219 10560 : return;
2220 :
2221 10 : pageno = MultiXactIdToOffsetPage(multi);
2222 10 : lock = SimpleLruGetBankLock(MultiXactOffsetCtl, pageno);
2223 :
2224 10 : LWLockAcquire(lock, LW_EXCLUSIVE);
2225 :
2226 : /* Zero the page and make a WAL entry about it */
2227 10 : SimpleLruZeroPage(MultiXactOffsetCtl, pageno);
2228 10 : XLogSimpleInsertInt64(RM_MULTIXACT_ID, XLOG_MULTIXACT_ZERO_OFF_PAGE,
2229 : pageno);
2230 :
2231 10 : LWLockRelease(lock);
2232 : }
2233 :
2234 : /*
2235 : * Make sure that MultiXactMember has room for the members of a newly-
2236 : * allocated MultiXactId.
2237 : *
2238 : * Like the above routine, this is called while holding MultiXactGenLock;
2239 : * same comments apply.
2240 : */
2241 : static void
2242 10570 : ExtendMultiXactMember(MultiXactOffset offset, int nmembers)
2243 : {
2244 : /*
2245 : * It's possible that the members span more than one page of the members
2246 : * file, so we loop to ensure we consider each page. The coding is not
2247 : * optimal if the members span several pages, but that seems unusual
2248 : * enough to not worry much about.
2249 : */
2250 21248 : while (nmembers > 0)
2251 : {
2252 : int flagsoff;
2253 : int flagsbit;
2254 : uint32 difference;
2255 :
2256 : /*
2257 : * Only zero when at first entry of a page.
2258 : */
2259 10678 : flagsoff = MXOffsetToFlagsOffset(offset);
2260 10678 : flagsbit = MXOffsetToFlagsBitShift(offset);
2261 10678 : if (flagsoff == 0 && flagsbit == 0)
2262 : {
2263 : int64 pageno;
2264 : LWLock *lock;
2265 :
2266 114 : pageno = MXOffsetToMemberPage(offset);
2267 114 : lock = SimpleLruGetBankLock(MultiXactMemberCtl, pageno);
2268 :
2269 114 : LWLockAcquire(lock, LW_EXCLUSIVE);
2270 :
2271 : /* Zero the page and make a WAL entry about it */
2272 114 : SimpleLruZeroPage(MultiXactMemberCtl, pageno);
2273 114 : XLogSimpleInsertInt64(RM_MULTIXACT_ID,
2274 : XLOG_MULTIXACT_ZERO_MEM_PAGE, pageno);
2275 :
2276 114 : LWLockRelease(lock);
2277 : }
2278 :
2279 : /* Compute the number of items till end of current page. */
2280 10678 : difference = MULTIXACT_MEMBERS_PER_PAGE - offset % MULTIXACT_MEMBERS_PER_PAGE;
2281 :
2282 : /*
2283 : * Advance to next page. OK if nmembers goes negative.
2284 : */
2285 10678 : nmembers -= difference;
2286 10678 : offset += difference;
2287 : }
2288 10570 : }
2289 :
2290 : /*
2291 : * GetOldestMultiXactId
2292 : *
2293 : * Return the oldest MultiXactId that's still possibly still seen as live by
2294 : * any running transaction. Older ones might still exist on disk, but they no
2295 : * longer have any running member transaction.
2296 : *
2297 : * It's not safe to truncate MultiXact SLRU segments on the value returned by
2298 : * this function; however, it can be set as the new relminmxid for any table
2299 : * that VACUUM knows has no remaining MXIDs < the same value. It is only safe
2300 : * to truncate SLRUs when no table can possibly still have a referencing MXID.
2301 : */
2302 : MultiXactId
2303 256916 : GetOldestMultiXactId(void)
2304 : {
2305 : MultiXactId oldestMXact;
2306 : int i;
2307 :
2308 : /*
2309 : * This is the oldest valid value among all the OldestMemberMXactId[] and
2310 : * OldestVisibleMXactId[] entries, or nextMXact if none are valid.
2311 : */
2312 256916 : LWLockAcquire(MultiXactGenLock, LW_SHARED);
2313 256916 : oldestMXact = MultiXactState->nextMXact;
2314 31909982 : for (i = 0; i < MaxOldestSlot; i++)
2315 : {
2316 : MultiXactId thisoldest;
2317 :
2318 31653066 : thisoldest = OldestMemberMXactId[i];
2319 31725086 : if (MultiXactIdIsValid(thisoldest) &&
2320 72020 : MultiXactIdPrecedes(thisoldest, oldestMXact))
2321 24 : oldestMXact = thisoldest;
2322 31653066 : thisoldest = OldestVisibleMXactId[i];
2323 31653370 : if (MultiXactIdIsValid(thisoldest) &&
2324 304 : MultiXactIdPrecedes(thisoldest, oldestMXact))
2325 4 : oldestMXact = thisoldest;
2326 : }
2327 :
2328 256916 : LWLockRelease(MultiXactGenLock);
2329 :
2330 256916 : return oldestMXact;
2331 : }
2332 :
2333 : /*
2334 : * Calculate the oldest member offset and install it in MultiXactState, where
2335 : * it can be used to adjust multixid freezing cutoffs.
2336 : */
2337 : static void
2338 3708 : SetOldestOffset(void)
2339 : {
2340 : MultiXactId oldestMultiXactId;
2341 : MultiXactId nextMXact;
2342 3708 : MultiXactOffset oldestOffset = 0; /* placate compiler */
2343 : MultiXactOffset nextOffset;
2344 3708 : bool oldestOffsetKnown = false;
2345 :
2346 : /*
2347 : * NB: Have to prevent concurrent truncation, we might otherwise try to
2348 : * lookup an oldestMulti that's concurrently getting truncated away.
2349 : */
2350 3708 : LWLockAcquire(MultiXactTruncationLock, LW_SHARED);
2351 :
2352 : /* Read relevant fields from shared memory. */
2353 3708 : LWLockAcquire(MultiXactGenLock, LW_SHARED);
2354 3708 : oldestMultiXactId = MultiXactState->oldestMultiXactId;
2355 3708 : nextMXact = MultiXactState->nextMXact;
2356 3708 : nextOffset = MultiXactState->nextOffset;
2357 : Assert(MultiXactState->finishedStartup);
2358 3708 : LWLockRelease(MultiXactGenLock);
2359 :
2360 : /*
2361 : * Determine the offset of the oldest multixact. Normally, we can read
2362 : * the offset from the multixact itself, but there's an important special
2363 : * case: if there are no multixacts in existence at all, oldestMXact
2364 : * obviously can't point to one. It will instead point to the multixact
2365 : * ID that will be assigned the next time one is needed.
2366 : */
2367 3708 : if (oldestMultiXactId == nextMXact)
2368 : {
2369 : /*
2370 : * When the next multixact gets created, it will be stored at the next
2371 : * offset.
2372 : */
2373 3668 : oldestOffset = nextOffset;
2374 3668 : oldestOffsetKnown = true;
2375 : }
2376 : else
2377 : {
2378 : /*
2379 : * Look up the offset at which the oldest existing multixact's members
2380 : * are stored. If we cannot find it, be careful not to fail, and
2381 : * leave oldestOffset unchanged. oldestOffset is initialized to zero
2382 : * at system startup, which prevents truncating members until a proper
2383 : * value is calculated.
2384 : *
2385 : * (We had bugs in early releases of PostgreSQL 9.3.X and 9.4.X where
2386 : * the supposedly-earliest multixact might not really exist. Those
2387 : * should be long gone by now, so this should not fail, but let's
2388 : * still be defensive.)
2389 : */
2390 : oldestOffsetKnown =
2391 40 : find_multixact_start(oldestMultiXactId, &oldestOffset);
2392 :
2393 40 : if (oldestOffsetKnown)
2394 40 : ereport(DEBUG1,
2395 : (errmsg_internal("oldest MultiXactId member is at offset %" PRIu64,
2396 : oldestOffset)));
2397 : else
2398 0 : ereport(LOG,
2399 : (errmsg("MultiXact member truncation is disabled because oldest checkpointed MultiXact %u does not exist on disk",
2400 : oldestMultiXactId)));
2401 : }
2402 :
2403 3708 : LWLockRelease(MultiXactTruncationLock);
2404 :
2405 : /* Install the computed value */
2406 3708 : if (oldestOffsetKnown)
2407 : {
2408 3708 : LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2409 3708 : MultiXactState->oldestOffset = oldestOffset;
2410 3708 : LWLockRelease(MultiXactGenLock);
2411 : }
2412 3708 : }
2413 :
2414 : /*
2415 : * Find the starting offset of the given MultiXactId.
2416 : *
2417 : * Returns false if the file containing the multi does not exist on disk.
2418 : * Otherwise, returns true and sets *result to the starting member offset.
2419 : *
2420 : * This function does not prevent concurrent truncation, so if that's
2421 : * required, the caller has to protect against that.
2422 : */
2423 : static bool
2424 40 : find_multixact_start(MultiXactId multi, MultiXactOffset *result)
2425 : {
2426 : MultiXactOffset offset;
2427 : int64 pageno;
2428 : int entryno;
2429 : int slotno;
2430 : MultiXactOffset *offptr;
2431 :
2432 : Assert(MultiXactState->finishedStartup);
2433 :
2434 40 : pageno = MultiXactIdToOffsetPage(multi);
2435 40 : entryno = MultiXactIdToOffsetEntry(multi);
2436 :
2437 : /*
2438 : * Write out dirty data, so PhysicalPageExists can work correctly.
2439 : */
2440 40 : SimpleLruWriteAll(MultiXactOffsetCtl, true);
2441 40 : SimpleLruWriteAll(MultiXactMemberCtl, true);
2442 :
2443 40 : if (!SimpleLruDoesPhysicalPageExist(MultiXactOffsetCtl, pageno))
2444 0 : return false;
2445 :
2446 : /* lock is acquired by SimpleLruReadPage_ReadOnly */
2447 40 : slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno, multi);
2448 40 : offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
2449 40 : offptr += entryno;
2450 40 : offset = *offptr;
2451 40 : LWLockRelease(SimpleLruGetBankLock(MultiXactOffsetCtl, pageno));
2452 :
2453 40 : *result = offset;
2454 40 : return true;
2455 : }
2456 :
2457 : /*
2458 : * GetMultiXactInfo
2459 : *
2460 : * Returns information about the current MultiXact state, as of:
2461 : * multixacts: Number of MultiXacts (nextMultiXactId - oldestMultiXactId)
2462 : * nextOffset: Next-to-be-assigned offset
2463 : * oldestMultiXactId: Oldest MultiXact ID still in use
2464 : * oldestOffset: Oldest offset still in use
2465 : */
2466 : void
2467 190136 : GetMultiXactInfo(uint32 *multixacts, MultiXactOffset *nextOffset,
2468 : MultiXactId *oldestMultiXactId, MultiXactOffset *oldestOffset)
2469 : {
2470 : MultiXactId nextMultiXactId;
2471 :
2472 190136 : LWLockAcquire(MultiXactGenLock, LW_SHARED);
2473 190136 : *nextOffset = MultiXactState->nextOffset;
2474 190136 : *oldestMultiXactId = MultiXactState->oldestMultiXactId;
2475 190136 : nextMultiXactId = MultiXactState->nextMXact;
2476 190136 : *oldestOffset = MultiXactState->oldestOffset;
2477 190136 : LWLockRelease(MultiXactGenLock);
2478 :
2479 190136 : *multixacts = nextMultiXactId - *oldestMultiXactId;
2480 190136 : }
2481 :
2482 : /*
2483 : * Multixact members can be removed once the multixacts that refer to them
2484 : * are older than every datminmxid. autovacuum_multixact_freeze_max_age and
2485 : * vacuum_multixact_freeze_table_age work together to make sure we never have
2486 : * too many multixacts; we hope that, at least under normal circumstances,
2487 : * this will also be sufficient to keep us from using too many offsets.
2488 : * However, if the average multixact has many members, we might accumulate a
2489 : * large amount of members, consuming disk space, while still using few enough
2490 : * multixids that the multixid limits fail to trigger relminmxid advancement
2491 : * by VACUUM.
2492 : *
2493 : * To prevent that, if the members space usage exceeds a threshold
2494 : * (MULTIXACT_MEMBER_LOW_THRESHOLD), we effectively reduce
2495 : * autovacuum_multixact_freeze_max_age to a value just less than the number of
2496 : * multixacts in use. We hope that this will quickly trigger autovacuuming on
2497 : * the table or tables with the oldest relminmxid, thus allowing datminmxid
2498 : * values to advance and removing some members.
2499 : *
2500 : * As the amount of the member space in use grows, we become more aggressive
2501 : * in clamping this value. That not only causes autovacuum to ramp up, but
2502 : * also makes any manual vacuums the user issues more aggressive. This
2503 : * happens because vacuum_get_cutoffs() will clamp the freeze table and the
2504 : * minimum freeze age cutoffs based on the effective
2505 : * autovacuum_multixact_freeze_max_age this function returns. At the extreme,
2506 : * when the members usage reaches MULTIXACT_MEMBER_HIGH_THRESHOLD, we clamp
2507 : * freeze_max_age to zero, and every vacuum of any table will freeze every
2508 : * multixact.
2509 : */
2510 : int
2511 190112 : MultiXactMemberFreezeThreshold(void)
2512 : {
2513 : uint32 multixacts;
2514 : uint32 victim_multixacts;
2515 : double fraction;
2516 : int result;
2517 : MultiXactId oldestMultiXactId;
2518 : MultiXactOffset oldestOffset;
2519 : MultiXactOffset nextOffset;
2520 : uint64 members;
2521 :
2522 : /* Read the current offsets and multixact usage. */
2523 190112 : GetMultiXactInfo(&multixacts, &nextOffset, &oldestMultiXactId, &oldestOffset);
2524 190112 : members = nextOffset - oldestOffset;
2525 :
2526 : /* If member space utilization is low, no special action is required. */
2527 190112 : if (members <= MULTIXACT_MEMBER_LOW_THRESHOLD)
2528 190112 : return autovacuum_multixact_freeze_max_age;
2529 :
2530 : /*
2531 : * Compute a target for relminmxid advancement. The number of multixacts
2532 : * we try to eliminate from the system is based on how far we are past
2533 : * MULTIXACT_MEMBER_LOW_THRESHOLD.
2534 : *
2535 : * The way this formula works is that when members is exactly at the low
2536 : * threshold, fraction = 0.0, and we set freeze_max_age equal to
2537 : * mxid_age(oldestMultiXactId). As members grows further, towards the
2538 : * high threshold, fraction grows linearly from 0.0 to 1.0, and the result
2539 : * shrinks from mxid_age(oldestMultiXactId) to 0. Beyond the high
2540 : * threshold, fraction > 1.0 and the result is clamped to 0.
2541 : */
2542 0 : fraction = (double) (members - MULTIXACT_MEMBER_LOW_THRESHOLD) /
2543 : (MULTIXACT_MEMBER_HIGH_THRESHOLD - MULTIXACT_MEMBER_LOW_THRESHOLD);
2544 :
2545 : /* fraction could be > 1.0, but lowest possible freeze age is zero */
2546 0 : if (fraction >= 1.0)
2547 0 : return 0;
2548 :
2549 0 : victim_multixacts = multixacts * fraction;
2550 0 : result = multixacts - victim_multixacts;
2551 :
2552 : /*
2553 : * Clamp to autovacuum_multixact_freeze_max_age, so that we never make
2554 : * autovacuum less aggressive than it would otherwise be.
2555 : */
2556 0 : return Min(result, autovacuum_multixact_freeze_max_age);
2557 : }
2558 :
2559 :
2560 : /*
2561 : * Delete members segments older than newOldestOffset
2562 : */
2563 : static void
2564 0 : PerformMembersTruncation(MultiXactOffset newOldestOffset)
2565 : {
2566 0 : SimpleLruTruncate(MultiXactMemberCtl,
2567 : MXOffsetToMemberPage(newOldestOffset));
2568 0 : }
2569 :
2570 : /*
2571 : * Delete offsets segments older than newOldestMulti
2572 : */
2573 : static void
2574 0 : PerformOffsetsTruncation(MultiXactId newOldestMulti)
2575 : {
2576 : /*
2577 : * We step back one multixact to avoid passing a cutoff page that hasn't
2578 : * been created yet in the rare case that oldestMulti would be the first
2579 : * item on a page and oldestMulti == nextMulti. In that case, if we
2580 : * didn't subtract one, we'd trigger SimpleLruTruncate's wraparound
2581 : * detection.
2582 : */
2583 0 : SimpleLruTruncate(MultiXactOffsetCtl,
2584 : MultiXactIdToOffsetPage(PreviousMultiXactId(newOldestMulti)));
2585 0 : }
2586 :
2587 : /*
2588 : * Remove all MultiXactOffset and MultiXactMember segments before the oldest
2589 : * ones still of interest.
2590 : *
2591 : * This is only called on a primary as part of vacuum (via
2592 : * vac_truncate_clog()). During recovery truncation is done by replaying
2593 : * truncation WAL records logged here.
2594 : *
2595 : * newOldestMulti is the oldest currently required multixact, newOldestMultiDB
2596 : * is one of the databases preventing newOldestMulti from increasing.
2597 : */
2598 : void
2599 1848 : TruncateMultiXact(MultiXactId newOldestMulti, Oid newOldestMultiDB)
2600 : {
2601 : MultiXactId oldestMulti;
2602 : MultiXactId nextMulti;
2603 : MultiXactOffset newOldestOffset;
2604 : MultiXactOffset nextOffset;
2605 :
2606 : Assert(!RecoveryInProgress());
2607 : Assert(MultiXactState->finishedStartup);
2608 : Assert(MultiXactIdIsValid(newOldestMulti));
2609 :
2610 : /*
2611 : * We can only allow one truncation to happen at once. Otherwise parts of
2612 : * members might vanish while we're doing lookups or similar. There's no
2613 : * need to have an interlock with creating new multis or such, since those
2614 : * are constrained by the limits (which only grow, never shrink).
2615 : */
2616 1848 : LWLockAcquire(MultiXactTruncationLock, LW_EXCLUSIVE);
2617 :
2618 1848 : LWLockAcquire(MultiXactGenLock, LW_SHARED);
2619 1848 : nextMulti = MultiXactState->nextMXact;
2620 1848 : nextOffset = MultiXactState->nextOffset;
2621 1848 : oldestMulti = MultiXactState->oldestMultiXactId;
2622 1848 : LWLockRelease(MultiXactGenLock);
2623 :
2624 : /*
2625 : * Make sure to only attempt truncation if there's values to truncate
2626 : * away. In normal processing values shouldn't go backwards, but there's
2627 : * some corner cases (due to bugs) where that's possible.
2628 : */
2629 1848 : if (MultiXactIdPrecedesOrEquals(newOldestMulti, oldestMulti))
2630 : {
2631 1848 : LWLockRelease(MultiXactTruncationLock);
2632 1848 : return;
2633 : }
2634 :
2635 : /*
2636 : * Compute up to where to truncate MultiXactMember. Lookup the
2637 : * corresponding member offset for newOldestMulti for that.
2638 : */
2639 0 : if (newOldestMulti == nextMulti)
2640 : {
2641 : /* there are NO MultiXacts */
2642 0 : newOldestOffset = nextOffset;
2643 : }
2644 0 : else if (!find_multixact_start(newOldestMulti, &newOldestOffset))
2645 : {
2646 0 : ereport(LOG,
2647 : (errmsg("cannot truncate up to MultiXact %u because it does not exist on disk, skipping truncation",
2648 : newOldestMulti)));
2649 0 : LWLockRelease(MultiXactTruncationLock);
2650 0 : return;
2651 : }
2652 :
2653 : /*
2654 : * On crash, MultiXactIdCreateFromMembers() can leave behind multixids
2655 : * that were not yet written out and hence have zero offset on disk. If
2656 : * such a multixid becomes oldestMulti, we won't be able to look up its
2657 : * offset. That should be rare, so we don't try to do anything smart about
2658 : * it. Just skip the truncation, and hope that by the next truncation
2659 : * attempt, oldestMulti has advanced to a valid multixid.
2660 : */
2661 0 : if (newOldestOffset == 0)
2662 : {
2663 0 : ereport(LOG,
2664 : (errmsg("cannot truncate up to MultiXact %u because it has invalid offset, skipping truncation",
2665 : newOldestMulti)));
2666 0 : LWLockRelease(MultiXactTruncationLock);
2667 0 : return;
2668 : }
2669 :
2670 0 : elog(DEBUG1, "performing multixact truncation: "
2671 : "oldestMulti %u (offsets segment %" PRIx64 "), "
2672 : "oldestOffset %" PRIu64 " (members segment %" PRIx64 ")",
2673 : newOldestMulti,
2674 : MultiXactIdToOffsetSegment(newOldestMulti),
2675 : newOldestOffset,
2676 : MXOffsetToMemberSegment(newOldestOffset));
2677 :
2678 : /*
2679 : * Do truncation, and the WAL logging of the truncation, in a critical
2680 : * section. That way offsets/members cannot get out of sync anymore, i.e.
2681 : * once consistent the newOldestMulti will always exist in members, even
2682 : * if we crashed in the wrong moment.
2683 : */
2684 0 : START_CRIT_SECTION();
2685 :
2686 : /*
2687 : * Prevent checkpoints from being scheduled concurrently. This is critical
2688 : * because otherwise a truncation record might not be replayed after a
2689 : * crash/basebackup, even though the state of the data directory would
2690 : * require it.
2691 : */
2692 : Assert((MyProc->delayChkptFlags & DELAY_CHKPT_START) == 0);
2693 0 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
2694 :
2695 : /* WAL log truncation */
2696 0 : WriteMTruncateXlogRec(newOldestMultiDB, newOldestMulti, newOldestOffset);
2697 :
2698 : /*
2699 : * Update in-memory limits before performing the truncation, while inside
2700 : * the critical section: Have to do it before truncation, to prevent
2701 : * concurrent lookups of those values. Has to be inside the critical
2702 : * section as otherwise a future call to this function would error out,
2703 : * while looking up the oldest member in offsets, if our caller crashes
2704 : * before updating the limits.
2705 : */
2706 0 : LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2707 0 : MultiXactState->oldestMultiXactId = newOldestMulti;
2708 0 : MultiXactState->oldestMultiXactDB = newOldestMultiDB;
2709 0 : MultiXactState->oldestOffset = newOldestOffset;
2710 0 : LWLockRelease(MultiXactGenLock);
2711 :
2712 : /* First truncate members */
2713 0 : PerformMembersTruncation(newOldestOffset);
2714 :
2715 : /* Then offsets */
2716 0 : PerformOffsetsTruncation(newOldestMulti);
2717 :
2718 0 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
2719 :
2720 0 : END_CRIT_SECTION();
2721 0 : LWLockRelease(MultiXactTruncationLock);
2722 : }
2723 :
2724 : /*
2725 : * Decide whether a MultiXactOffset page number is "older" for truncation
2726 : * purposes. Analogous to CLOGPagePrecedes().
2727 : *
2728 : * Offsetting the values is optional, because MultiXactIdPrecedes() has
2729 : * translational symmetry.
2730 : */
2731 : static bool
2732 0 : MultiXactOffsetPagePrecedes(int64 page1, int64 page2)
2733 : {
2734 : MultiXactId multi1;
2735 : MultiXactId multi2;
2736 :
2737 0 : multi1 = ((MultiXactId) page1) * MULTIXACT_OFFSETS_PER_PAGE;
2738 0 : multi1 += FirstMultiXactId + 1;
2739 0 : multi2 = ((MultiXactId) page2) * MULTIXACT_OFFSETS_PER_PAGE;
2740 0 : multi2 += FirstMultiXactId + 1;
2741 :
2742 0 : return (MultiXactIdPrecedes(multi1, multi2) &&
2743 0 : MultiXactIdPrecedes(multi1,
2744 : multi2 + MULTIXACT_OFFSETS_PER_PAGE - 1));
2745 : }
2746 :
2747 : /*
2748 : * Decide whether a MultiXactMember page number is "older" for truncation
2749 : * purposes. There is no "invalid offset number" and members never wrap
2750 : * around, so use the numbers verbatim.
2751 : */
2752 : static bool
2753 0 : MultiXactMemberPagePrecedes(int64 page1, int64 page2)
2754 : {
2755 0 : return page1 < page2;
2756 : }
2757 :
2758 : /*
2759 : * Decide which of two MultiXactIds is earlier.
2760 : *
2761 : * XXX do we need to do something special for InvalidMultiXactId?
2762 : * (Doesn't look like it.)
2763 : */
2764 : bool
2765 2716342 : MultiXactIdPrecedes(MultiXactId multi1, MultiXactId multi2)
2766 : {
2767 2716342 : int32 diff = (int32) (multi1 - multi2);
2768 :
2769 2716342 : return (diff < 0);
2770 : }
2771 :
2772 : /*
2773 : * MultiXactIdPrecedesOrEquals -- is multi1 logically <= multi2?
2774 : *
2775 : * XXX do we need to do something special for InvalidMultiXactId?
2776 : * (Doesn't look like it.)
2777 : */
2778 : bool
2779 13568 : MultiXactIdPrecedesOrEquals(MultiXactId multi1, MultiXactId multi2)
2780 : {
2781 13568 : int32 diff = (int32) (multi1 - multi2);
2782 :
2783 13568 : return (diff <= 0);
2784 : }
2785 :
2786 :
2787 : /*
2788 : * Write a TRUNCATE xlog record
2789 : *
2790 : * We must flush the xlog record to disk before returning --- see notes in
2791 : * TruncateCLOG().
2792 : */
2793 : static void
2794 0 : WriteMTruncateXlogRec(Oid oldestMultiDB,
2795 : MultiXactId oldestMulti,
2796 : MultiXactOffset oldestOffset)
2797 : {
2798 : XLogRecPtr recptr;
2799 : xl_multixact_truncate xlrec;
2800 :
2801 0 : xlrec.oldestMultiDB = oldestMultiDB;
2802 0 : xlrec.oldestMulti = oldestMulti;
2803 0 : xlrec.oldestOffset = oldestOffset;
2804 :
2805 0 : XLogBeginInsert();
2806 0 : XLogRegisterData(&xlrec, SizeOfMultiXactTruncate);
2807 0 : recptr = XLogInsert(RM_MULTIXACT_ID, XLOG_MULTIXACT_TRUNCATE_ID);
2808 0 : XLogFlush(recptr);
2809 0 : }
2810 :
2811 : /*
2812 : * MULTIXACT resource manager's routines
2813 : */
2814 : void
2815 8 : multixact_redo(XLogReaderState *record)
2816 : {
2817 8 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
2818 :
2819 : /* Backup blocks are not used in multixact records */
2820 : Assert(!XLogRecHasAnyBlockRefs(record));
2821 :
2822 8 : if (info == XLOG_MULTIXACT_ZERO_OFF_PAGE)
2823 : {
2824 : int64 pageno;
2825 :
2826 0 : memcpy(&pageno, XLogRecGetData(record), sizeof(pageno));
2827 0 : SimpleLruZeroAndWritePage(MultiXactOffsetCtl, pageno);
2828 : }
2829 8 : else if (info == XLOG_MULTIXACT_ZERO_MEM_PAGE)
2830 : {
2831 : int64 pageno;
2832 :
2833 0 : memcpy(&pageno, XLogRecGetData(record), sizeof(pageno));
2834 0 : SimpleLruZeroAndWritePage(MultiXactMemberCtl, pageno);
2835 : }
2836 8 : else if (info == XLOG_MULTIXACT_CREATE_ID)
2837 : {
2838 8 : xl_multixact_create *xlrec =
2839 8 : (xl_multixact_create *) XLogRecGetData(record);
2840 : TransactionId max_xid;
2841 : int i;
2842 :
2843 : /* Store the data back into the SLRU files */
2844 8 : RecordNewMultiXact(xlrec->mid, xlrec->moff, xlrec->nmembers,
2845 8 : xlrec->members);
2846 :
2847 : /* Make sure nextMXact/nextOffset are beyond what this record has */
2848 8 : MultiXactAdvanceNextMXact(NextMultiXactId(xlrec->mid),
2849 8 : xlrec->moff + xlrec->nmembers);
2850 :
2851 : /*
2852 : * Make sure nextXid is beyond any XID mentioned in the record. This
2853 : * should be unnecessary, since any XID found here ought to have other
2854 : * evidence in the XLOG, but let's be safe.
2855 : */
2856 8 : max_xid = XLogRecGetXid(record);
2857 24 : for (i = 0; i < xlrec->nmembers; i++)
2858 : {
2859 16 : if (TransactionIdPrecedes(max_xid, xlrec->members[i].xid))
2860 0 : max_xid = xlrec->members[i].xid;
2861 : }
2862 :
2863 8 : AdvanceNextFullTransactionIdPastXid(max_xid);
2864 : }
2865 0 : else if (info == XLOG_MULTIXACT_TRUNCATE_ID)
2866 : {
2867 : xl_multixact_truncate xlrec;
2868 : int64 pageno;
2869 :
2870 0 : memcpy(&xlrec, XLogRecGetData(record),
2871 : SizeOfMultiXactTruncate);
2872 :
2873 0 : elog(DEBUG1, "replaying multixact truncation: "
2874 : "oldestMulti %u (offsets segment %" PRIx64 "), "
2875 : "oldestOffset %" PRIu64 " (members segment %" PRIx64 ")",
2876 : xlrec.oldestMulti,
2877 : MultiXactIdToOffsetSegment(xlrec.oldestMulti),
2878 : xlrec.oldestOffset,
2879 : MXOffsetToMemberSegment(xlrec.oldestOffset));
2880 :
2881 : /* should not be required, but more than cheap enough */
2882 0 : LWLockAcquire(MultiXactTruncationLock, LW_EXCLUSIVE);
2883 :
2884 : /*
2885 : * Advance the horizon values, so they're current at the end of
2886 : * recovery.
2887 : */
2888 0 : SetMultiXactIdLimit(xlrec.oldestMulti, xlrec.oldestMultiDB);
2889 :
2890 0 : PerformMembersTruncation(xlrec.oldestOffset);
2891 :
2892 : /*
2893 : * During XLOG replay, latest_page_number isn't necessarily set up
2894 : * yet; insert a suitable value to bypass the sanity test in
2895 : * SimpleLruTruncate.
2896 : */
2897 0 : pageno = MultiXactIdToOffsetPage(xlrec.oldestMulti);
2898 0 : pg_atomic_write_u64(&MultiXactOffsetCtl->shared->latest_page_number,
2899 : pageno);
2900 0 : PerformOffsetsTruncation(xlrec.oldestMulti);
2901 :
2902 0 : LWLockRelease(MultiXactTruncationLock);
2903 : }
2904 : else
2905 0 : elog(PANIC, "multixact_redo: unknown op code %u", info);
2906 8 : }
2907 :
2908 : /*
2909 : * Entrypoint for sync.c to sync offsets files.
2910 : */
2911 : int
2912 0 : multixactoffsetssyncfiletag(const FileTag *ftag, char *path)
2913 : {
2914 0 : return SlruSyncFileTag(MultiXactOffsetCtl, ftag, path);
2915 : }
2916 :
2917 : /*
2918 : * Entrypoint for sync.c to sync members files.
2919 : */
2920 : int
2921 0 : multixactmemberssyncfiletag(const FileTag *ftag, char *path)
2922 : {
2923 0 : return SlruSyncFileTag(MultiXactMemberCtl, ftag, path);
2924 : }
|
