Debian Codegraph @

src/backend/access/transam/multixact.c

raw
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) 1996-2025, 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/slru.h"
73 #include "access/transam.h"
74 #include "access/twophase.h"
75 #include "access/twophase_rmgr.h"
76 #include "access/xact.h"
77 #include "access/xlog.h"
78 #include "access/xloginsert.h"
79 #include "access/xlogutils.h"
80 #include "commands/dbcommands.h"
81 #include "funcapi.h"
82 #include "lib/ilist.h"
83 #include "miscadmin.h"
84 #include "pg_trace.h"
85 #include "pgstat.h"
86 #include "postmaster/autovacuum.h"
87 #include "storage/pmsignal.h"
88 #include "storage/proc.h"
89 #include "storage/procarray.h"
90 #include "utils/fmgrprotos.h"
91 #include "utils/guc_hooks.h"
92 #include "utils/injection_point.h"
93 #include "utils/memutils.h"
94
95
96 /*
97 * Defines for MultiXactOffset page sizes. A page is the same BLCKSZ as is
98 * used everywhere else in Postgres.
99 *
100 * Note: because MultiXactOffsets are 32 bits and wrap around at 0xFFFFFFFF,
101 * MultiXact page numbering also wraps around at
102 * 0xFFFFFFFF/MULTIXACT_OFFSETS_PER_PAGE, and segment numbering at
103 * 0xFFFFFFFF/MULTIXACT_OFFSETS_PER_PAGE/SLRU_PAGES_PER_SEGMENT. We need
104 * take no explicit notice of that fact in this module, except when comparing
105 * segment and page numbers in TruncateMultiXact (see
106 * MultiXactOffsetPagePrecedes).
107 */
108
109 /* We need four bytes per offset */
110 #define MULTIXACT_OFFSETS_PER_PAGE (BLCKSZ / sizeof(MultiXactOffset))
111
112 static inline int64
113 MultiXactIdToOffsetPage(MultiXactId multi)
114 {
115 return multi / MULTIXACT_OFFSETS_PER_PAGE;
116 }
117
118 static inline int
119 MultiXactIdToOffsetEntry(MultiXactId multi)
120 {
121 return multi % MULTIXACT_OFFSETS_PER_PAGE;
122 }
123
124 static inline int64
125 MultiXactIdToOffsetSegment(MultiXactId multi)
126 {
127 return MultiXactIdToOffsetPage(multi) / SLRU_PAGES_PER_SEGMENT;
128 }
129
130 /*
131 * The situation for members is a bit more complex: we store one byte of
132 * additional flag bits for each TransactionId. To do this without getting
133 * into alignment issues, we store four bytes of flags, and then the
134 * corresponding 4 Xids. Each such 5-word (20-byte) set we call a "group", and
135 * are stored as a whole in pages. Thus, with 8kB BLCKSZ, we keep 409 groups
136 * per page. This wastes 12 bytes per page, but that's OK -- simplicity (and
137 * performance) trumps space efficiency here.
138 *
139 * Note that the "offset" macros work with byte offset, not array indexes, so
140 * arithmetic must be done using "char *" pointers.
141 */
142 /* We need eight bits per xact, so one xact fits in a byte */
143 #define MXACT_MEMBER_BITS_PER_XACT 8
144 #define MXACT_MEMBER_FLAGS_PER_BYTE 1
145 #define MXACT_MEMBER_XACT_BITMASK ((1 << MXACT_MEMBER_BITS_PER_XACT) - 1)
146
147 /* how many full bytes of flags are there in a group? */
148 #define MULTIXACT_FLAGBYTES_PER_GROUP 4
149 #define MULTIXACT_MEMBERS_PER_MEMBERGROUP \
150 (MULTIXACT_FLAGBYTES_PER_GROUP * MXACT_MEMBER_FLAGS_PER_BYTE)
151 /* size in bytes of a complete group */
152 #define MULTIXACT_MEMBERGROUP_SIZE \
153 (sizeof(TransactionId) * MULTIXACT_MEMBERS_PER_MEMBERGROUP + MULTIXACT_FLAGBYTES_PER_GROUP)
154 #define MULTIXACT_MEMBERGROUPS_PER_PAGE (BLCKSZ / MULTIXACT_MEMBERGROUP_SIZE)
155 #define MULTIXACT_MEMBERS_PER_PAGE \
156 (MULTIXACT_MEMBERGROUPS_PER_PAGE * MULTIXACT_MEMBERS_PER_MEMBERGROUP)
157
158 /*
159 * Because the number of items per page is not a divisor of the last item
160 * number (member 0xFFFFFFFF), the last segment does not use the maximum number
161 * of pages, and moreover the last used page therein does not use the same
162 * number of items as previous pages. (Another way to say it is that the
163 * 0xFFFFFFFF member is somewhere in the middle of the last page, so the page
164 * has some empty space after that item.)
165 *
166 * This constant is the number of members in the last page of the last segment.
167 */
168 #define MAX_MEMBERS_IN_LAST_MEMBERS_PAGE \
169 ((uint32) ((0xFFFFFFFF % MULTIXACT_MEMBERS_PER_PAGE) + 1))
170
171 /* page in which a member is to be found */
172 static inline int64
173 MXOffsetToMemberPage(MultiXactOffset offset)
174 {
175 return offset / MULTIXACT_MEMBERS_PER_PAGE;
176 }
177
178 static inline int64
179 MXOffsetToMemberSegment(MultiXactOffset offset)
180 {
181 return MXOffsetToMemberPage(offset) / SLRU_PAGES_PER_SEGMENT;
182 }
183
184 /* Location (byte offset within page) of flag word for a given member */
185 static inline int
186 MXOffsetToFlagsOffset(MultiXactOffset offset)
187 {
188 MultiXactOffset group = offset / MULTIXACT_MEMBERS_PER_MEMBERGROUP;
189 int grouponpg = group % MULTIXACT_MEMBERGROUPS_PER_PAGE;
190 int byteoff = grouponpg * MULTIXACT_MEMBERGROUP_SIZE;
191
192 return byteoff;
193 }
194
195 static inline int
196 MXOffsetToFlagsBitShift(MultiXactOffset offset)
197 {
198 int member_in_group = offset % MULTIXACT_MEMBERS_PER_MEMBERGROUP;
199 int bshift = member_in_group * MXACT_MEMBER_BITS_PER_XACT;
200
201 return bshift;
202 }
203
204 /* Location (byte offset within page) of TransactionId of given member */
205 static inline int
206 MXOffsetToMemberOffset(MultiXactOffset offset)
207 {
208 int member_in_group = offset % MULTIXACT_MEMBERS_PER_MEMBERGROUP;
209
210 return MXOffsetToFlagsOffset(offset) +
211 MULTIXACT_FLAGBYTES_PER_GROUP +
212 member_in_group * sizeof(TransactionId);
213 }
214
215 /* Multixact members wraparound thresholds. */
216 #define MULTIXACT_MEMBER_SAFE_THRESHOLD (MaxMultiXactOffset / 2)
217 #define MULTIXACT_MEMBER_DANGER_THRESHOLD \
218 (MaxMultiXactOffset - MaxMultiXactOffset / 4)
219
220 static inline MultiXactId
221 PreviousMultiXactId(MultiXactId multi)
222 {
223 return multi == FirstMultiXactId ? MaxMultiXactId : multi - 1;
224 }
225
226 /*
227 * Links to shared-memory data structures for MultiXact control
228 */
229 static SlruCtlData MultiXactOffsetCtlData;
230 static SlruCtlData MultiXactMemberCtlData;
231
232 #define MultiXactOffsetCtl (&MultiXactOffsetCtlData)
233 #define MultiXactMemberCtl (&MultiXactMemberCtlData)
234
235 /*
236 * MultiXact state shared across all backends. All this state is protected
237 * by MultiXactGenLock. (We also use SLRU bank's lock of MultiXactOffset and
238 * MultiXactMember to guard accesses to the two sets of SLRU buffers. For
239 * concurrency's sake, we avoid holding more than one of these locks at a
240 * time.)
241 */
242 typedef struct MultiXactStateData
243 {
244 /* next-to-be-assigned MultiXactId */
245 MultiXactId nextMXact;
246
247 /* next-to-be-assigned offset */
248 MultiXactOffset nextOffset;
249
250 /* Have we completed multixact startup? */
251 bool finishedStartup;
252
253 /*
254 * Oldest multixact that is still potentially referenced by a relation.
255 * Anything older than this should not be consulted. These values are
256 * updated by vacuum.
257 */
258 MultiXactId oldestMultiXactId;
259 Oid oldestMultiXactDB;
260
261 /*
262 * Oldest multixact offset that is potentially referenced by a multixact
263 * referenced by a relation. We don't always know this value, so there's
264 * a flag here to indicate whether or not we currently do.
265 */
266 MultiXactOffset oldestOffset;
267 bool oldestOffsetKnown;
268
269 /* support for anti-wraparound measures */
270 MultiXactId multiVacLimit;
271 MultiXactId multiWarnLimit;
272 MultiXactId multiStopLimit;
273 MultiXactId multiWrapLimit;
274
275 /* support for members anti-wraparound measures */
276 MultiXactOffset offsetStopLimit; /* known if oldestOffsetKnown */
277
278 /*
279 * Per-backend data starts here. We have two arrays stored in the area
280 * immediately following the MultiXactStateData struct:
281 *
282 * OldestMemberMXactId[k] is the oldest MultiXactId each backend's current
283 * transaction(s) could possibly be a member of, or InvalidMultiXactId
284 * when the backend has no live transaction that could possibly be a
285 * member of a MultiXact. Each backend sets its entry to the current
286 * nextMXact counter just before first acquiring a shared lock in a given
287 * transaction, and clears it at transaction end. (This works because only
288 * during or after acquiring a shared lock could an XID possibly become a
289 * member of a MultiXact, and that MultiXact would have to be created
290 * during or after the lock acquisition.)
291 *
292 * In the OldestMemberMXactId array, there's a slot for all normal
293 * backends (0..MaxBackends-1) followed by a slot for max_prepared_xacts
294 * prepared transactions.
295 *
296 * OldestVisibleMXactId[k] is the oldest MultiXactId each backend's
297 * current transaction(s) think is potentially live, or InvalidMultiXactId
298 * when not in a transaction or not in a transaction that's paid any
299 * attention to MultiXacts yet. This is computed when first needed in a
300 * given transaction, and cleared at transaction end. We can compute it
301 * as the minimum of the valid OldestMemberMXactId[] entries at the time
302 * we compute it (using nextMXact if none are valid). Each backend is
303 * required not to attempt to access any SLRU data for MultiXactIds older
304 * than its own OldestVisibleMXactId[] setting; this is necessary because
305 * the relevant SLRU data can be concurrently truncated away.
306 *
307 * In the OldestVisibleMXactId array, there's a slot for all normal
308 * backends (0..MaxBackends-1) only. No slots for prepared transactions.
309 *
310 * The oldest valid value among all of the OldestMemberMXactId[] and
311 * OldestVisibleMXactId[] entries is considered by vacuum as the earliest
312 * possible value still having any live member transaction -- OldestMxact.
313 * Any value older than that is typically removed from tuple headers, or
314 * "frozen" via being replaced with a new xmax. VACUUM can sometimes even
315 * remove an individual MultiXact xmax whose value is >= its OldestMxact
316 * cutoff, though typically only when no individual member XID is still
317 * running. See FreezeMultiXactId for full details.
318 *
319 * Whenever VACUUM advances relminmxid, then either its OldestMxact cutoff
320 * or the oldest extant Multi remaining in the table is used as the new
321 * pg_class.relminmxid value (whichever is earlier). The minimum of all
322 * relminmxid values in each database is stored in pg_database.datminmxid.
323 * In turn, the minimum of all of those values is stored in pg_control.
324 * This is used as the truncation point for pg_multixact when unneeded
325 * segments get removed by vac_truncate_clog() during vacuuming.
326 */
327 MultiXactId perBackendXactIds[FLEXIBLE_ARRAY_MEMBER];
328 } MultiXactStateData;
329
330 /*
331 * Sizes of OldestMemberMXactId and OldestVisibleMXactId arrays.
332 */
333 #define NumMemberSlots (MaxBackends + max_prepared_xacts)
334 #define NumVisibleSlots MaxBackends
335
336 /* Pointers to the state data in shared memory */
337 static MultiXactStateData *MultiXactState;
338 static MultiXactId *OldestMemberMXactId;
339 static MultiXactId *OldestVisibleMXactId;
340
341
342 static inline MultiXactId *
343 MyOldestMemberMXactIdSlot(void)
344 {
345 /*
346 * The first MaxBackends entries in the OldestMemberMXactId array are
347 * reserved for regular backends. MyProcNumber should index into one of
348 * them.
349 */
350 Assert(MyProcNumber >= 0 && MyProcNumber < MaxBackends);
351 return &OldestMemberMXactId[MyProcNumber];
352 }
353
354 static inline MultiXactId *
355 PreparedXactOldestMemberMXactIdSlot(ProcNumber procno)
356 {
357 int prepared_xact_idx;
358
359 Assert(procno >= FIRST_PREPARED_XACT_PROC_NUMBER);
360 prepared_xact_idx = procno - FIRST_PREPARED_XACT_PROC_NUMBER;
361
362 /*
363 * The first MaxBackends entries in the OldestMemberMXactId array are
364 * reserved for regular backends. Prepared xacts come after them.
365 */
366 Assert(MaxBackends + prepared_xact_idx < NumMemberSlots);
367 return &OldestMemberMXactId[MaxBackends + prepared_xact_idx];
368 }
369
370 static inline MultiXactId *
371 MyOldestVisibleMXactIdSlot(void)
372 {
373 Assert(MyProcNumber >= 0 && MyProcNumber < NumVisibleSlots);
374 return &OldestVisibleMXactId[MyProcNumber];
375 }
376
377 /*
378 * Definitions for the backend-local MultiXactId cache.
379 *
380 * We use this cache to store known MultiXacts, so we don't need to go to
381 * SLRU areas every time.
382 *
383 * The cache lasts for the duration of a single transaction, the rationale
384 * for this being that most entries will contain our own TransactionId and
385 * so they will be uninteresting by the time our next transaction starts.
386 * (XXX not clear that this is correct --- other members of the MultiXact
387 * could hang around longer than we did. However, it's not clear what a
388 * better policy for flushing old cache entries would be.) FIXME actually
389 * this is plain wrong now that multixact's may contain update Xids.
390 *
391 * We allocate the cache entries in a memory context that is deleted at
392 * transaction end, so we don't need to do retail freeing of entries.
393 */
394 typedef struct mXactCacheEnt
395 {
396 MultiXactId multi;
397 int nmembers;
398 dlist_node node;
399 MultiXactMember members[FLEXIBLE_ARRAY_MEMBER];
400 } mXactCacheEnt;
401
402 #define MAX_CACHE_ENTRIES 256
403 static dclist_head MXactCache = DCLIST_STATIC_INIT(MXactCache);
404 static MemoryContext MXactContext = NULL;
405
406 #ifdef MULTIXACT_DEBUG
407 #define debug_elog2(a,b) elog(a,b)
408 #define debug_elog3(a,b,c) elog(a,b,c)
409 #define debug_elog4(a,b,c,d) elog(a,b,c,d)
410 #define debug_elog5(a,b,c,d,e) elog(a,b,c,d,e)
411 #define debug_elog6(a,b,c,d,e,f) elog(a,b,c,d,e,f)
412 #else
413 #define debug_elog2(a,b)
414 #define debug_elog3(a,b,c)
415 #define debug_elog4(a,b,c,d)
416 #define debug_elog5(a,b,c,d,e)
417 #define debug_elog6(a,b,c,d,e,f)
418 #endif
419
420 /*
421 * Hack to deal with WAL generated with older minor versions.
422 *
423 * last_initialized_offsets_page is the XLOG_MULTIXACT_ZERO_OFF_PAGE record
424 * that we saw during WAL replay, or -1 if we haven't seen any yet.
425 *
426 * pre_initialized_offsets_page is the last page that was implicitly
427 * initialized by replaying a XLOG_MULTIXACT_CREATE_ID record, when we had not
428 * seen a XLOG_MULTIXACT_ZERO_OFF_PAGE record for the page yet.
429 */
430 static int64 last_initialized_offsets_page = -1;
431 static int64 pre_initialized_offsets_page = -1;
432
433 /* internal MultiXactId management */
434 static void MultiXactIdSetOldestVisible(void);
435 static void RecordNewMultiXact(MultiXactId multi, MultiXactOffset offset,
436 int nmembers, MultiXactMember *members);
437 static MultiXactId GetNewMultiXactId(int nmembers, MultiXactOffset *offset);
438
439 /* MultiXact cache management */
440 static int mxactMemberComparator(const void *arg1, const void *arg2);
441 static MultiXactId mXactCacheGetBySet(int nmembers, MultiXactMember *members);
442 static int mXactCacheGetById(MultiXactId multi, MultiXactMember **members);
443 static void mXactCachePut(MultiXactId multi, int nmembers,
444 MultiXactMember *members);
445
446 static char *mxstatus_to_string(MultiXactStatus status);
447
448 /* management of SLRU infrastructure */
449 static int ZeroMultiXactOffsetPage(int64 pageno, bool writeXlog);
450 static int ZeroMultiXactMemberPage(int64 pageno, bool writeXlog);
451 static bool MultiXactOffsetPagePrecedes(int64 page1, int64 page2);
452 static bool MultiXactMemberPagePrecedes(int64 page1, int64 page2);
453 static bool MultiXactOffsetPrecedes(MultiXactOffset offset1,
454 MultiXactOffset offset2);
455 static void ExtendMultiXactOffset(MultiXactId multi);
456 static void ExtendMultiXactMember(MultiXactOffset offset, int nmembers);
457 static bool MultiXactOffsetWouldWrap(MultiXactOffset boundary,
458 MultiXactOffset start, uint32 distance);
459 static bool SetOffsetVacuumLimit(bool is_startup);
460 static bool find_multixact_start(MultiXactId multi, MultiXactOffset *result);
461 static void WriteMZeroPageXlogRec(int64 pageno, uint8 info);
462 static void WriteMTruncateXlogRec(Oid oldestMultiDB,
463 MultiXactId startTruncOff,
464 MultiXactId endTruncOff,
465 MultiXactOffset startTruncMemb,
466 MultiXactOffset endTruncMemb);
467
468
469 /*
470 * MultiXactIdCreate
471 * Construct a MultiXactId representing two TransactionIds.
472 *
473 * The two XIDs must be different, or be requesting different statuses.
474 *
475 * NB - we don't worry about our local MultiXactId cache here, because that
476 * is handled by the lower-level routines.
477 */
478 MultiXactId
479 MultiXactIdCreate(TransactionId xid1, MultiXactStatus status1,
480 TransactionId xid2, MultiXactStatus status2)
481 {
482 MultiXactId newMulti;
483 MultiXactMember members[2];
484
485 Assert(TransactionIdIsValid(xid1));
486 Assert(TransactionIdIsValid(xid2));
487
488 Assert(!TransactionIdEquals(xid1, xid2) || (status1 != status2));
489
490 /* MultiXactIdSetOldestMember() must have been called already. */
491 Assert(MultiXactIdIsValid(*MyOldestMemberMXactIdSlot()));
492
493 /*
494 * Note: unlike MultiXactIdExpand, we don't bother to check that both XIDs
495 * are still running. In typical usage, xid2 will be our own XID and the
496 * caller just did a check on xid1, so it'd be wasted effort.
497 */
498
499 members[0].xid = xid1;
500 members[0].status = status1;
501 members[1].xid = xid2;
502 members[1].status = status2;
503
504 newMulti = MultiXactIdCreateFromMembers(2, members);
505
506 debug_elog3(DEBUG2, "Create: %s",
507 mxid_to_string(newMulti, 2, members));
508
509 return newMulti;
510 }
511
512 /*
513 * MultiXactIdExpand
514 * Add a TransactionId to a pre-existing MultiXactId.
515 *
516 * If the TransactionId is already a member of the passed MultiXactId with the
517 * same status, just return it as-is.
518 *
519 * Note that we do NOT actually modify the membership of a pre-existing
520 * MultiXactId; instead we create a new one. This is necessary to avoid
521 * a race condition against code trying to wait for one MultiXactId to finish;
522 * see notes in heapam.c.
523 *
524 * NB - we don't worry about our local MultiXactId cache here, because that
525 * is handled by the lower-level routines.
526 *
527 * Note: It is critical that MultiXactIds that come from an old cluster (i.e.
528 * one upgraded by pg_upgrade from a cluster older than this feature) are not
529 * passed in.
530 */
531 MultiXactId
532 MultiXactIdExpand(MultiXactId multi, TransactionId xid, MultiXactStatus status)
533 {
534 MultiXactId newMulti;
535 MultiXactMember *members;
536 MultiXactMember *newMembers;
537 int nmembers;
538 int i;
539 int j;
540
541 Assert(MultiXactIdIsValid(multi));
542 Assert(TransactionIdIsValid(xid));
543
544 /* MultiXactIdSetOldestMember() must have been called already. */
545 Assert(MultiXactIdIsValid(*MyOldestMemberMXactIdSlot()));
546
547 debug_elog5(DEBUG2, "Expand: received multi %u, xid %u status %s",
548 multi, xid, mxstatus_to_string(status));
549
550 /*
551 * Note: we don't allow for old multis here. The reason is that the only
552 * caller of this function does a check that the multixact is no longer
553 * running.
554 */
555 nmembers = GetMultiXactIdMembers(multi, &members, false, false);
556
557 if (nmembers < 0)
558 {
559 MultiXactMember member;
560
561 /*
562 * The MultiXactId is obsolete. This can only happen if all the
563 * MultiXactId members stop running between the caller checking and
564 * passing it to us. It would be better to return that fact to the
565 * caller, but it would complicate the API and it's unlikely to happen
566 * too often, so just deal with it by creating a singleton MultiXact.
567 */
568 member.xid = xid;
569 member.status = status;
570 newMulti = MultiXactIdCreateFromMembers(1, &member);
571
572 debug_elog4(DEBUG2, "Expand: %u has no members, create singleton %u",
573 multi, newMulti);
574 return newMulti;
575 }
576
577 /*
578 * If the TransactionId is already a member of the MultiXactId with the
579 * same status, just return the existing MultiXactId.
580 */
581 for (i = 0; i < nmembers; i++)
582 {
583 if (TransactionIdEquals(members[i].xid, xid) &&
584 (members[i].status == status))
585 {
586 debug_elog4(DEBUG2, "Expand: %u is already a member of %u",
587 xid, multi);
588 pfree(members);
589 return multi;
590 }
591 }
592
593 /*
594 * Determine which of the members of the MultiXactId are still of
595 * interest. This is any running transaction, and also any transaction
596 * that grabbed something stronger than just a lock and was committed. (An
597 * update that aborted is of no interest here; and having more than one
598 * update Xid in a multixact would cause errors elsewhere.)
599 *
600 * Removing dead members is not just an optimization: freezing of tuples
601 * whose Xmax are multis depends on this behavior.
602 *
603 * Note we have the same race condition here as above: j could be 0 at the
604 * end of the loop.
605 */
606 newMembers = (MultiXactMember *)
607 palloc(sizeof(MultiXactMember) * (nmembers + 1));
608
609 for (i = 0, j = 0; i < nmembers; i++)
610 {
611 if (TransactionIdIsInProgress(members[i].xid) ||
612 (ISUPDATE_from_mxstatus(members[i].status) &&
613 TransactionIdDidCommit(members[i].xid)))
614 {
615 newMembers[j].xid = members[i].xid;
616 newMembers[j++].status = members[i].status;
617 }
618 }
619
620 newMembers[j].xid = xid;
621 newMembers[j++].status = status;
622 newMulti = MultiXactIdCreateFromMembers(j, newMembers);
623
624 pfree(members);
625 pfree(newMembers);
626
627 debug_elog3(DEBUG2, "Expand: returning new multi %u", newMulti);
628
629 return newMulti;
630 }
631
632 /*
633 * MultiXactIdIsRunning
634 * Returns whether a MultiXactId is "running".
635 *
636 * We return true if at least one member of the given MultiXactId is still
637 * running. Note that a "false" result is certain not to change,
638 * because it is not legal to add members to an existing MultiXactId.
639 *
640 * Caller is expected to have verified that the multixact does not come from
641 * a pg_upgraded share-locked tuple.
642 */
643 bool
644 MultiXactIdIsRunning(MultiXactId multi, bool isLockOnly)
645 {
646 MultiXactMember *members;
647 int nmembers;
648 int i;
649
650 debug_elog3(DEBUG2, "IsRunning %u?", multi);
651
652 /*
653 * "false" here means we assume our callers have checked that the given
654 * multi cannot possibly come from a pg_upgraded database.
655 */
656 nmembers = GetMultiXactIdMembers(multi, &members, false, isLockOnly);
657
658 if (nmembers <= 0)
659 {
660 debug_elog2(DEBUG2, "IsRunning: no members");
661 return false;
662 }
663
664 /*
665 * Checking for myself is cheap compared to looking in shared memory;
666 * return true if any live subtransaction of the current top-level
667 * transaction is a member.
668 *
669 * This is not needed for correctness, it's just a fast path.
670 */
671 for (i = 0; i < nmembers; i++)
672 {
673 if (TransactionIdIsCurrentTransactionId(members[i].xid))
674 {
675 debug_elog3(DEBUG2, "IsRunning: I (%d) am running!", i);
676 pfree(members);
677 return true;
678 }
679 }
680
681 /*
682 * This could be made faster by having another entry point in procarray.c,
683 * walking the PGPROC array only once for all the members. But in most
684 * cases nmembers should be small enough that it doesn't much matter.
685 */
686 for (i = 0; i < nmembers; i++)
687 {
688 if (TransactionIdIsInProgress(members[i].xid))
689 {
690 debug_elog4(DEBUG2, "IsRunning: member %d (%u) is running",
691 i, members[i].xid);
692 pfree(members);
693 return true;
694 }
695 }
696
697 pfree(members);
698
699 debug_elog3(DEBUG2, "IsRunning: %u is not running", multi);
700
701 return false;
702 }
703
704 /*
705 * MultiXactIdSetOldestMember
706 * Save the oldest MultiXactId this transaction could be a member of.
707 *
708 * We set the OldestMemberMXactId for a given transaction the first time it's
709 * going to do some operation that might require a MultiXactId (tuple lock,
710 * update or delete). We need to do this even if we end up using a
711 * TransactionId instead of a MultiXactId, because there is a chance that
712 * another transaction would add our XID to a MultiXactId.
713 *
714 * The value to set is the next-to-be-assigned MultiXactId, so this is meant to
715 * be called just before doing any such possibly-MultiXactId-able operation.
716 */
717 void
718 MultiXactIdSetOldestMember(void)
719 {
720 if (!MultiXactIdIsValid(*MyOldestMemberMXactIdSlot()))
721 {
722 MultiXactId nextMXact;
723
724 /*
725 * You might think we don't need to acquire a lock here, since
726 * fetching and storing of TransactionIds is probably atomic, but in
727 * fact we do: suppose we pick up nextMXact and then lose the CPU for
728 * a long time. Someone else could advance nextMXact, and then
729 * another someone else could compute an OldestVisibleMXactId that
730 * would be after the value we are going to store when we get control
731 * back. Which would be wrong.
732 *
733 * Note that a shared lock is sufficient, because it's enough to stop
734 * someone from advancing nextMXact; and nobody else could be trying
735 * to write to our OldestMember entry, only reading (and we assume
736 * storing it is atomic.)
737 */
738 LWLockAcquire(MultiXactGenLock, LW_SHARED);
739
740 /*
741 * We have to beware of the possibility that nextMXact is in the
742 * wrapped-around state. We don't fix the counter itself here, but we
743 * must be sure to store a valid value in our array entry.
744 */
745 nextMXact = MultiXactState->nextMXact;
746 if (nextMXact < FirstMultiXactId)
747 nextMXact = FirstMultiXactId;
748
749 *MyOldestMemberMXactIdSlot() = nextMXact;
750
751 LWLockRelease(MultiXactGenLock);
752
753 debug_elog4(DEBUG2, "MultiXact: setting OldestMember[%d] = %u",
754 MyProcNumber, nextMXact);
755 }
756 }
757
758 /*
759 * MultiXactIdSetOldestVisible
760 * Save the oldest MultiXactId this transaction considers possibly live.
761 *
762 * We set the OldestVisibleMXactId for a given transaction the first time
763 * it's going to inspect any MultiXactId. Once we have set this, we are
764 * guaranteed that SLRU data for MultiXactIds >= our own OldestVisibleMXactId
765 * won't be truncated away.
766 *
767 * The value to set is the oldest of nextMXact and all the valid per-backend
768 * OldestMemberMXactId[] entries. Because of the locking we do, we can be
769 * certain that no subsequent call to MultiXactIdSetOldestMember can set
770 * an OldestMemberMXactId[] entry older than what we compute here. Therefore
771 * there is no live transaction, now or later, that can be a member of any
772 * MultiXactId older than the OldestVisibleMXactId we compute here.
773 */
774 static void
775 MultiXactIdSetOldestVisible(void)
776 {
777 if (!MultiXactIdIsValid(*MyOldestVisibleMXactIdSlot()))
778 {
779 MultiXactId oldestMXact;
780 int i;
781
782 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
783
784 /*
785 * We have to beware of the possibility that nextMXact is in the
786 * wrapped-around state. We don't fix the counter itself here, but we
787 * must be sure to store a valid value in our array entry.
788 */
789 oldestMXact = MultiXactState->nextMXact;
790 if (oldestMXact < FirstMultiXactId)
791 oldestMXact = FirstMultiXactId;
792
793 for (i = 0; i < NumMemberSlots; i++)
794 {
795 MultiXactId thisoldest = OldestMemberMXactId[i];
796
797 if (MultiXactIdIsValid(thisoldest) &&
798 MultiXactIdPrecedes(thisoldest, oldestMXact))
799 oldestMXact = thisoldest;
800 }
801
802 *MyOldestVisibleMXactIdSlot() = oldestMXact;
803
804 LWLockRelease(MultiXactGenLock);
805
806 debug_elog4(DEBUG2, "MultiXact: setting OldestVisible[%d] = %u",
807 MyProcNumber, oldestMXact);
808 }
809 }
810
811 /*
812 * ReadNextMultiXactId
813 * Return the next MultiXactId to be assigned, but don't allocate it
814 */
815 MultiXactId
816 ReadNextMultiXactId(void)
817 {
818 MultiXactId mxid;
819
820 /* XXX we could presumably do this without a lock. */
821 LWLockAcquire(MultiXactGenLock, LW_SHARED);
822 mxid = MultiXactState->nextMXact;
823 LWLockRelease(MultiXactGenLock);
824
825 if (mxid < FirstMultiXactId)
826 mxid = FirstMultiXactId;
827
828 return mxid;
829 }
830
831 /*
832 * ReadMultiXactIdRange
833 * Get the range of IDs that may still be referenced by a relation.
834 */
835 void
836 ReadMultiXactIdRange(MultiXactId *oldest, MultiXactId *next)
837 {
838 LWLockAcquire(MultiXactGenLock, LW_SHARED);
839 *oldest = MultiXactState->oldestMultiXactId;
840 *next = MultiXactState->nextMXact;
841 LWLockRelease(MultiXactGenLock);
842
843 if (*oldest < FirstMultiXactId)
844 *oldest = FirstMultiXactId;
845 if (*next < FirstMultiXactId)
846 *next = FirstMultiXactId;
847 }
848
849
850 /*
851 * MultiXactIdCreateFromMembers
852 * Make a new MultiXactId from the specified set of members
853 *
854 * Make XLOG, SLRU and cache entries for a new MultiXactId, recording the
855 * given TransactionIds as members. Returns the newly created MultiXactId.
856 *
857 * NB: the passed members[] array will be sorted in-place.
858 */
859 MultiXactId
860 MultiXactIdCreateFromMembers(int nmembers, MultiXactMember *members)
861 {
862 MultiXactId multi;
863 MultiXactOffset offset;
864 xl_multixact_create xlrec;
865
866 debug_elog3(DEBUG2, "Create: %s",
867 mxid_to_string(InvalidMultiXactId, nmembers, members));
868
869 /*
870 * See if the same set of members already exists in our cache; if so, just
871 * re-use that MultiXactId. (Note: it might seem that looking in our
872 * cache is insufficient, and we ought to search disk to see if a
873 * duplicate definition already exists. But since we only ever create
874 * MultiXacts containing our own XID, in most cases any such MultiXacts
875 * were in fact created by us, and so will be in our cache. There are
876 * corner cases where someone else added us to a MultiXact without our
877 * knowledge, but it's not worth checking for.)
878 */
879 multi = mXactCacheGetBySet(nmembers, members);
880 if (MultiXactIdIsValid(multi))
881 {
882 debug_elog2(DEBUG2, "Create: in cache!");
883 return multi;
884 }
885
886 /* Verify that there is a single update Xid among the given members. */
887 {
888 int i;
889 bool has_update = false;
890
891 for (i = 0; i < nmembers; i++)
892 {
893 if (ISUPDATE_from_mxstatus(members[i].status))
894 {
895 if (has_update)
896 elog(ERROR, "new multixact has more than one updating member: %s",
897 mxid_to_string(InvalidMultiXactId, nmembers, members));
898 has_update = true;
899 }
900 }
901 }
902
903 /* Load the injection point before entering the critical section */
904 INJECTION_POINT_LOAD("multixact-create-from-members");
905
906 /*
907 * Assign the MXID and offsets range to use, and make sure there is space
908 * in the OFFSETs and MEMBERs files. NB: this routine does
909 * START_CRIT_SECTION().
910 *
911 * Note: unlike MultiXactIdCreate and MultiXactIdExpand, we do not check
912 * that we've called MultiXactIdSetOldestMember here. This is because
913 * this routine is used in some places to create new MultiXactIds of which
914 * the current backend is not a member, notably during freezing of multis
915 * in vacuum. During vacuum, in particular, it would be unacceptable to
916 * keep OldestMulti set, in case it runs for long.
917 */
918 multi = GetNewMultiXactId(nmembers, &offset);
919
920 INJECTION_POINT_CACHED("multixact-create-from-members", NULL);
921
922 /* Make an XLOG entry describing the new MXID. */
923 xlrec.mid = multi;
924 xlrec.moff = offset;
925 xlrec.nmembers = nmembers;
926
927 /*
928 * XXX Note: there's a lot of padding space in MultiXactMember. We could
929 * find a more compact representation of this Xlog record -- perhaps all
930 * the status flags in one XLogRecData, then all the xids in another one?
931 * Not clear that it's worth the trouble though.
932 */
933 XLogBeginInsert();
934 XLogRegisterData(&xlrec, SizeOfMultiXactCreate);
935 XLogRegisterData(members, nmembers * sizeof(MultiXactMember));
936
937 (void) XLogInsert(RM_MULTIXACT_ID, XLOG_MULTIXACT_CREATE_ID);
938
939 /* Now enter the information into the OFFSETs and MEMBERs logs */
940 RecordNewMultiXact(multi, offset, nmembers, members);
941
942 /* Done with critical section */
943 END_CRIT_SECTION();
944
945 /* Store the new MultiXactId in the local cache, too */
946 mXactCachePut(multi, nmembers, members);
947
948 debug_elog2(DEBUG2, "Create: all done");
949
950 return multi;
951 }
952
953 /*
954 * RecordNewMultiXact
955 * Write info about a new multixact into the offsets and members files
956 *
957 * This is broken out of MultiXactIdCreateFromMembers so that xlog replay can
958 * use it.
959 */
960 static void
961 RecordNewMultiXact(MultiXactId multi, MultiXactOffset offset,
962 int nmembers, MultiXactMember *members)
963 {
964 int64 pageno;
965 int64 prev_pageno;
966 int entryno;
967 int slotno;
968 MultiXactOffset *offptr;
969 MultiXactId next;
970 int64 next_pageno;
971 int next_entryno;
972 MultiXactOffset *next_offptr;
973 MultiXactOffset next_offset;
974 LWLock *lock;
975 LWLock *prevlock = NULL;
976
977 /* position of this multixid in the offsets SLRU area */
978 pageno = MultiXactIdToOffsetPage(multi);
979 entryno = MultiXactIdToOffsetEntry(multi);
980
981 /* position of the next multixid */
982 next = multi + 1;
983 if (next < FirstMultiXactId)
984 next = FirstMultiXactId;
985 next_pageno = MultiXactIdToOffsetPage(next);
986 next_entryno = MultiXactIdToOffsetEntry(next);
987
988 /*
989 * Older minor versions didn't set the next multixid's offset in this
990 * function, and therefore didn't initialize the next page until the next
991 * multixid was assigned. If we're replaying WAL that was generated by
992 * such a version, the next page might not be initialized yet. Initialize
993 * it now.
994 */
995 if (InRecovery && next_pageno != pageno)
996 {
997 bool init_needed;
998
999 /*----------
1000 * Check if the page exists, and if not, initialize it now.
1001 *
1002 * The straightforward way to check if the page exists is to call
1003 * SimpleLruDoesPhysicalPageExist(). However, there two problems with
1004 * that:
1005 *
1006 * 1. It's somewhat expensive to call on every page switch.
1007 *
1008 * 2. It does not take into account pages that have been initialized
1009 * in the SLRU buffer cache but not yet flushed to disk. For such
1010 * pages, it will incorrectly return false.
1011 *
1012 * To fix both of those problems, if we have replayed any
1013 * XLOG_MULTIXACT_ZERO_OFF_PAGE records, we assume that the last page
1014 * that was zeroed by XLOG_MULTIXACT_ZERO_OFF_PAGE is the last page
1015 * that exists. This works because the XLOG_MULTIXACT_ZERO_OFF_PAGE
1016 * records must appear in the WAL in order, unlike CREATE_ID records.
1017 * We only resort to SimpleLruDoesPhysicalPageExist() if we haven't
1018 * seen any XLOG_MULTIXACT_ZERO_OFF_PAGE records yet, which should
1019 * happen at most once after starting WAL recovery.
1020 *
1021 * As an extra safety measure, if we do resort to
1022 * SimpleLruDoesPhysicalPageExist(), flush the SLRU buffers first so
1023 * that it will return an accurate result.
1024 *----------
1025 */
1026 if (last_initialized_offsets_page == -1)
1027 {
1028 SimpleLruWriteAll(MultiXactOffsetCtl, false);
1029 init_needed = !SimpleLruDoesPhysicalPageExist(MultiXactOffsetCtl, next_pageno);
1030 }
1031 else
1032 init_needed = (last_initialized_offsets_page == pageno);
1033
1034 if (init_needed)
1035 {
1036 elog(DEBUG1, "next offsets page is not initialized, initializing it now");
1037
1038 lock = SimpleLruGetBankLock(MultiXactOffsetCtl, next_pageno);
1039 LWLockAcquire(lock, LW_EXCLUSIVE);
1040
1041 /* Create and zero the page */
1042 slotno = SimpleLruZeroPage(MultiXactOffsetCtl, next_pageno);
1043
1044 /* Make sure it's written out */
1045 SimpleLruWritePage(MultiXactOffsetCtl, slotno);
1046 Assert(!MultiXactOffsetCtl->shared->page_dirty[slotno]);
1047
1048 LWLockRelease(lock);
1049
1050 /*
1051 * Remember that we initialized the page, so that we don't zero it
1052 * again at the XLOG_MULTIXACT_ZERO_OFF_PAGE record.
1053 */
1054 pre_initialized_offsets_page = next_pageno;
1055 last_initialized_offsets_page = next_pageno;
1056 }
1057 }
1058
1059 /*
1060 * Set the starting offset of this multixid's members.
1061 *
1062 * In the common case, it was already be set by the previous
1063 * RecordNewMultiXact call, as this was the next multixid of the previous
1064 * multixid. But if multiple backends are generating multixids
1065 * concurrently, we might race ahead and get called before the previous
1066 * multixid.
1067 */
1068 lock = SimpleLruGetBankLock(MultiXactOffsetCtl, pageno);
1069 LWLockAcquire(lock, LW_EXCLUSIVE);
1070
1071 /*
1072 * Note: we pass the MultiXactId to SimpleLruReadPage as the "transaction"
1073 * to complain about if there's any I/O error. This is kinda bogus, but
1074 * since the errors will always give the full pathname, it should be clear
1075 * enough that a MultiXactId is really involved. Perhaps someday we'll
1076 * take the trouble to generalize the slru.c error reporting code.
1077 */
1078 slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, multi);
1079 offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
1080 offptr += entryno;
1081
1082 if (*offptr != offset)
1083 {
1084 /* should already be set to the correct value, or not at all */
1085 Assert(*offptr == 0);
1086 *offptr = offset;
1087 MultiXactOffsetCtl->shared->page_dirty[slotno] = true;
1088 }
1089
1090 /*
1091 * Set the next multixid's offset to the end of this multixid's members.
1092 */
1093 if (next_pageno == pageno)
1094 {
1095 next_offptr = offptr + 1;
1096 }
1097 else
1098 {
1099 /* must be the first entry on the page */
1100 Assert(next_entryno == 0 || next == FirstMultiXactId);
1101
1102 /* Swap the lock for a lock on the next page */
1103 LWLockRelease(lock);
1104 lock = SimpleLruGetBankLock(MultiXactOffsetCtl, next_pageno);
1105 LWLockAcquire(lock, LW_EXCLUSIVE);
1106
1107 slotno = SimpleLruReadPage(MultiXactOffsetCtl, next_pageno, true, next);
1108 next_offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
1109 next_offptr += next_entryno;
1110 }
1111
1112 /* Like in GetNewMultiXactId(), skip over offset 0 */
1113 next_offset = offset + nmembers;
1114 if (next_offset == 0)
1115 next_offset = 1;
1116 if (*next_offptr != next_offset)
1117 {
1118 /* should already be set to the correct value, or not at all */
1119 Assert(*next_offptr == 0);
1120 *next_offptr = next_offset;
1121 MultiXactOffsetCtl->shared->page_dirty[slotno] = true;
1122 }
1123
1124 /* Release MultiXactOffset SLRU lock. */
1125 LWLockRelease(lock);
1126
1127 prev_pageno = -1;
1128
1129 for (int i = 0; i < nmembers; i++, offset++)
1130 {
1131 TransactionId *memberptr;
1132 uint32 *flagsptr;
1133 uint32 flagsval;
1134 int bshift;
1135 int flagsoff;
1136 int memberoff;
1137
1138 Assert(members[i].status <= MultiXactStatusUpdate);
1139
1140 pageno = MXOffsetToMemberPage(offset);
1141 memberoff = MXOffsetToMemberOffset(offset);
1142 flagsoff = MXOffsetToFlagsOffset(offset);
1143 bshift = MXOffsetToFlagsBitShift(offset);
1144
1145 if (pageno != prev_pageno)
1146 {
1147 /*
1148 * MultiXactMember SLRU page is changed so check if this new page
1149 * fall into the different SLRU bank then release the old bank's
1150 * lock and acquire lock on the new bank.
1151 */
1152 lock = SimpleLruGetBankLock(MultiXactMemberCtl, pageno);
1153 if (lock != prevlock)
1154 {
1155 if (prevlock != NULL)
1156 LWLockRelease(prevlock);
1157
1158 LWLockAcquire(lock, LW_EXCLUSIVE);
1159 prevlock = lock;
1160 }
1161 slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, multi);
1162 prev_pageno = pageno;
1163 }
1164
1165 memberptr = (TransactionId *)
1166 (MultiXactMemberCtl->shared->page_buffer[slotno] + memberoff);
1167
1168 *memberptr = members[i].xid;
1169
1170 flagsptr = (uint32 *)
1171 (MultiXactMemberCtl->shared->page_buffer[slotno] + flagsoff);
1172
1173 flagsval = *flagsptr;
1174 flagsval &= ~(((1 << MXACT_MEMBER_BITS_PER_XACT) - 1) << bshift);
1175 flagsval |= (members[i].status << bshift);
1176 *flagsptr = flagsval;
1177
1178 MultiXactMemberCtl->shared->page_dirty[slotno] = true;
1179 }
1180
1181 if (prevlock != NULL)
1182 LWLockRelease(prevlock);
1183 }
1184
1185 /*
1186 * GetNewMultiXactId
1187 * Get the next MultiXactId.
1188 *
1189 * Also, reserve the needed amount of space in the "members" area. The
1190 * starting offset of the reserved space is returned in *offset.
1191 *
1192 * This may generate XLOG records for expansion of the offsets and/or members
1193 * files. Unfortunately, we have to do that while holding MultiXactGenLock
1194 * to avoid race conditions --- the XLOG record for zeroing a page must appear
1195 * before any backend can possibly try to store data in that page!
1196 *
1197 * We start a critical section before advancing the shared counters. The
1198 * caller must end the critical section after writing SLRU data.
1199 */
1200 static MultiXactId
1201 GetNewMultiXactId(int nmembers, MultiXactOffset *offset)
1202 {
1203 MultiXactId result;
1204 MultiXactOffset nextOffset;
1205
1206 debug_elog3(DEBUG2, "GetNew: for %d xids", nmembers);
1207
1208 /* safety check, we should never get this far in a HS standby */
1209 if (RecoveryInProgress())
1210 elog(ERROR, "cannot assign MultiXactIds during recovery");
1211
1212 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
1213
1214 /* Handle wraparound of the nextMXact counter */
1215 if (MultiXactState->nextMXact < FirstMultiXactId)
1216 MultiXactState->nextMXact = FirstMultiXactId;
1217
1218 /* Assign the MXID */
1219 result = MultiXactState->nextMXact;
1220
1221 /*----------
1222 * Check to see if it's safe to assign another MultiXactId. This protects
1223 * against catastrophic data loss due to multixact wraparound. The basic
1224 * rules are:
1225 *
1226 * If we're past multiVacLimit or the safe threshold for member storage
1227 * space, or we don't know what the safe threshold for member storage is,
1228 * start trying to force autovacuum cycles.
1229 * If we're past multiWarnLimit, start issuing warnings.
1230 * If we're past multiStopLimit, refuse to create new MultiXactIds.
1231 *
1232 * Note these are pretty much the same protections in GetNewTransactionId.
1233 *----------
1234 */
1235 if (!MultiXactIdPrecedes(result, MultiXactState->multiVacLimit))
1236 {
1237 /*
1238 * For safety's sake, we release MultiXactGenLock while sending
1239 * signals, warnings, etc. This is not so much because we care about
1240 * preserving concurrency in this situation, as to avoid any
1241 * possibility of deadlock while doing get_database_name(). First,
1242 * copy all the shared values we'll need in this path.
1243 */
1244 MultiXactId multiWarnLimit = MultiXactState->multiWarnLimit;
1245 MultiXactId multiStopLimit = MultiXactState->multiStopLimit;
1246 MultiXactId multiWrapLimit = MultiXactState->multiWrapLimit;
1247 Oid oldest_datoid = MultiXactState->oldestMultiXactDB;
1248
1249 LWLockRelease(MultiXactGenLock);
1250
1251 if (IsUnderPostmaster &&
1252 !MultiXactIdPrecedes(result, multiStopLimit))
1253 {
1254 char *oldest_datname = get_database_name(oldest_datoid);
1255
1256 /*
1257 * Immediately kick autovacuum into action as we're already in
1258 * ERROR territory.
1259 */
1260 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
1261
1262 /* complain even if that DB has disappeared */
1263 if (oldest_datname)
1264 ereport(ERROR,
1265 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1266 errmsg("database is not accepting commands that assign new MultiXactIds to avoid wraparound data loss in database \"%s\"",
1267 oldest_datname),
1268 errhint("Execute a database-wide VACUUM in that database.\n"
1269 "You might also need to commit or roll back old prepared transactions.")));
1270 else
1271 ereport(ERROR,
1272 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1273 errmsg("database is not accepting commands that assign new MultiXactIds to avoid wraparound data loss in database with OID %u",
1274 oldest_datoid),
1275 errhint("Execute a database-wide VACUUM in that database.\n"
1276 "You might also need to commit or roll back old prepared transactions.")));
1277 }
1278
1279 /*
1280 * To avoid swamping the postmaster with signals, we issue the autovac
1281 * request only once per 64K multis generated. This still gives
1282 * plenty of chances before we get into real trouble.
1283 */
1284 if (IsUnderPostmaster && (result % 65536) == 0)
1285 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
1286
1287 if (!MultiXactIdPrecedes(result, multiWarnLimit))
1288 {
1289 char *oldest_datname = get_database_name(oldest_datoid);
1290
1291 /* complain even if that DB has disappeared */
1292 if (oldest_datname)
1293 ereport(WARNING,
1294 (errmsg_plural("database \"%s\" must be vacuumed before %u more MultiXactId is used",
1295 "database \"%s\" must be vacuumed before %u more MultiXactIds are used",
1296 multiWrapLimit - result,
1297 oldest_datname,
1298 multiWrapLimit - result),
1299 errhint("Execute a database-wide VACUUM in that database.\n"
1300 "You might also need to commit or roll back old prepared transactions.")));
1301 else
1302 ereport(WARNING,
1303 (errmsg_plural("database with OID %u must be vacuumed before %u more MultiXactId is used",
1304 "database with OID %u must be vacuumed before %u more MultiXactIds are used",
1305 multiWrapLimit - result,
1306 oldest_datoid,
1307 multiWrapLimit - result),
1308 errhint("Execute a database-wide VACUUM in that database.\n"
1309 "You might also need to commit or roll back old prepared transactions.")));
1310 }
1311
1312 /* Re-acquire lock and start over */
1313 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
1314 result = MultiXactState->nextMXact;
1315 if (result < FirstMultiXactId)
1316 result = FirstMultiXactId;
1317 }
1318
1319 /*
1320 * Make sure there is room for the next MXID in the file. Assigning this
1321 * MXID sets the next MXID's offset already.
1322 */
1323 ExtendMultiXactOffset(result + 1);
1324
1325 /*
1326 * Reserve the members space, similarly to above. Also, be careful not to
1327 * return zero as the starting offset for any multixact. See
1328 * GetMultiXactIdMembers() for motivation.
1329 */
1330 nextOffset = MultiXactState->nextOffset;
1331 if (nextOffset == 0)
1332 {
1333 *offset = 1;
1334 nmembers++; /* allocate member slot 0 too */
1335 }
1336 else
1337 *offset = nextOffset;
1338
1339 /*----------
1340 * Protect against overrun of the members space as well, with the
1341 * following rules:
1342 *
1343 * If we're past offsetStopLimit, refuse to generate more multis.
1344 * If we're close to offsetStopLimit, emit a warning.
1345 *
1346 * Arbitrarily, we start emitting warnings when we're 20 segments or less
1347 * from offsetStopLimit.
1348 *
1349 * Note we haven't updated the shared state yet, so if we fail at this
1350 * point, the multixact ID we grabbed can still be used by the next guy.
1351 *
1352 * Note that there is no point in forcing autovacuum runs here: the
1353 * multixact freeze settings would have to be reduced for that to have any
1354 * effect.
1355 *----------
1356 */
1357 #define OFFSET_WARN_SEGMENTS 20
1358 if (MultiXactState->oldestOffsetKnown &&
1359 MultiXactOffsetWouldWrap(MultiXactState->offsetStopLimit, nextOffset,
1360 nmembers))
1361 {
1362 /* see comment in the corresponding offsets wraparound case */
1363 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
1364
1365 ereport(ERROR,
1366 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1367 errmsg("multixact \"members\" limit exceeded"),
1368 errdetail_plural("This command would create a multixact with %u members, but the remaining space is only enough for %u member.",
1369 "This command would create a multixact with %u members, but the remaining space is only enough for %u members.",
1370 MultiXactState->offsetStopLimit - nextOffset - 1,
1371 nmembers,
1372 MultiXactState->offsetStopLimit - nextOffset - 1),
1373 errhint("Execute a database-wide VACUUM in database with OID %u with reduced \"vacuum_multixact_freeze_min_age\" and \"vacuum_multixact_freeze_table_age\" settings.",
1374 MultiXactState->oldestMultiXactDB)));
1375 }
1376
1377 /*
1378 * Check whether we should kick autovacuum into action, to prevent members
1379 * wraparound. NB we use a much larger window to trigger autovacuum than
1380 * just the warning limit. The warning is just a measure of last resort -
1381 * this is in line with GetNewTransactionId's behaviour.
1382 */
1383 if (!MultiXactState->oldestOffsetKnown ||
1384 (MultiXactState->nextOffset - MultiXactState->oldestOffset
1385 > MULTIXACT_MEMBER_SAFE_THRESHOLD))
1386 {
1387 /*
1388 * To avoid swamping the postmaster with signals, we issue the autovac
1389 * request only when crossing a segment boundary. With default
1390 * compilation settings that's roughly after 50k members. This still
1391 * gives plenty of chances before we get into real trouble.
1392 */
1393 if ((MXOffsetToMemberPage(nextOffset) / SLRU_PAGES_PER_SEGMENT) !=
1394 (MXOffsetToMemberPage(nextOffset + nmembers) / SLRU_PAGES_PER_SEGMENT))
1395 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
1396 }
1397
1398 if (MultiXactState->oldestOffsetKnown &&
1399 MultiXactOffsetWouldWrap(MultiXactState->offsetStopLimit,
1400 nextOffset,
1401 nmembers + MULTIXACT_MEMBERS_PER_PAGE * SLRU_PAGES_PER_SEGMENT * OFFSET_WARN_SEGMENTS))
1402 ereport(WARNING,
1403 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1404 errmsg_plural("database with OID %u must be vacuumed before %d more multixact member is used",
1405 "database with OID %u must be vacuumed before %d more multixact members are used",
1406 MultiXactState->offsetStopLimit - nextOffset + nmembers,
1407 MultiXactState->oldestMultiXactDB,
1408 MultiXactState->offsetStopLimit - nextOffset + nmembers),
1409 errhint("Execute a database-wide VACUUM in that database with reduced \"vacuum_multixact_freeze_min_age\" and \"vacuum_multixact_freeze_table_age\" settings.")));
1410
1411 ExtendMultiXactMember(nextOffset, nmembers);
1412
1413 /*
1414 * Critical section from here until caller has written the data into the
1415 * just-reserved SLRU space; we don't want to error out with a partly
1416 * written MultiXact structure. (In particular, failing to write our
1417 * start offset after advancing nextMXact would effectively corrupt the
1418 * previous MultiXact.)
1419 */
1420 START_CRIT_SECTION();
1421
1422 /*
1423 * Advance counters. As in GetNewTransactionId(), this must not happen
1424 * until after file extension has succeeded!
1425 *
1426 * We don't care about MultiXactId wraparound here; it will be handled by
1427 * the next iteration. But note that nextMXact may be InvalidMultiXactId
1428 * or the first value on a segment-beginning page after this routine
1429 * exits, so anyone else looking at the variable must be prepared to deal
1430 * with either case. Similarly, nextOffset may be zero, but we won't use
1431 * that as the actual start offset of the next multixact.
1432 */
1433 (MultiXactState->nextMXact)++;
1434
1435 MultiXactState->nextOffset += nmembers;
1436
1437 LWLockRelease(MultiXactGenLock);
1438
1439 debug_elog4(DEBUG2, "GetNew: returning %u offset %u", result, *offset);
1440 return result;
1441 }
1442
1443 /*
1444 * GetMultiXactIdMembers
1445 * Return the set of MultiXactMembers that make up a MultiXactId
1446 *
1447 * Return value is the number of members found, or -1 if there are none,
1448 * and *members is set to a newly palloc'ed array of members. It's the
1449 * caller's responsibility to free it when done with it.
1450 *
1451 * from_pgupgrade must be passed as true if and only if only the multixact
1452 * corresponds to a value from a tuple that was locked in a 9.2-or-older
1453 * installation and later pg_upgrade'd (that is, the infomask is
1454 * HEAP_LOCKED_UPGRADED). In this case, we know for certain that no members
1455 * can still be running, so we return -1 just like for an empty multixact
1456 * without any further checking. It would be wrong to try to resolve such a
1457 * multixact: either the multixact is within the current valid multixact
1458 * range, in which case the returned result would be bogus, or outside that
1459 * range, in which case an error would be raised.
1460 *
1461 * In all other cases, the passed multixact must be within the known valid
1462 * range, that is, greater than or equal to oldestMultiXactId, and less than
1463 * nextMXact. Otherwise, an error is raised.
1464 *
1465 * isLockOnly must be set to true if caller is certain that the given multi
1466 * is used only to lock tuples; can be false without loss of correctness,
1467 * but passing a true means we can return quickly without checking for
1468 * old updates.
1469 */
1470 int
1471 GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
1472 bool from_pgupgrade, bool isLockOnly)
1473 {
1474 int64 pageno;
1475 int64 prev_pageno;
1476 int entryno;
1477 int slotno;
1478 MultiXactOffset *offptr;
1479 MultiXactOffset offset;
1480 int length;
1481 int truelength;
1482 MultiXactId oldestMXact;
1483 MultiXactId nextMXact;
1484 MultiXactId tmpMXact;
1485 MultiXactOffset nextOffset;
1486 MultiXactMember *ptr;
1487 LWLock *lock;
1488
1489 debug_elog3(DEBUG2, "GetMembers: asked for %u", multi);
1490
1491 if (!MultiXactIdIsValid(multi) || from_pgupgrade)
1492 {
1493 *members = NULL;
1494 return -1;
1495 }
1496
1497 /* See if the MultiXactId is in the local cache */
1498 length = mXactCacheGetById(multi, members);
1499 if (length >= 0)
1500 {
1501 debug_elog3(DEBUG2, "GetMembers: found %s in the cache",
1502 mxid_to_string(multi, length, *members));
1503 return length;
1504 }
1505
1506 /* Set our OldestVisibleMXactId[] entry if we didn't already */
1507 MultiXactIdSetOldestVisible();
1508
1509 /*
1510 * If we know the multi is used only for locking and not for updates, then
1511 * we can skip checking if the value is older than our oldest visible
1512 * multi. It cannot possibly still be running.
1513 */
1514 if (isLockOnly &&
1515 MultiXactIdPrecedes(multi, *MyOldestVisibleMXactIdSlot()))
1516 {
1517 debug_elog2(DEBUG2, "GetMembers: a locker-only multi is too old");
1518 *members = NULL;
1519 return -1;
1520 }
1521
1522 /*
1523 * We check known limits on MultiXact before resorting to the SLRU area.
1524 *
1525 * An ID older than MultiXactState->oldestMultiXactId cannot possibly be
1526 * useful; it has already been removed, or will be removed shortly, by
1527 * truncation. If one is passed, an error is raised.
1528 *
1529 * Also, an ID >= nextMXact shouldn't ever be seen here; if it is seen, it
1530 * implies undetected ID wraparound has occurred. This raises a hard
1531 * error.
1532 *
1533 * Shared lock is enough here since we aren't modifying any global state.
1534 * Acquire it just long enough to grab the current counter values. We may
1535 * need both nextMXact and nextOffset; see below.
1536 */
1537 LWLockAcquire(MultiXactGenLock, LW_SHARED);
1538
1539 oldestMXact = MultiXactState->oldestMultiXactId;
1540 nextMXact = MultiXactState->nextMXact;
1541 nextOffset = MultiXactState->nextOffset;
1542
1543 LWLockRelease(MultiXactGenLock);
1544
1545 if (MultiXactIdPrecedes(multi, oldestMXact))
1546 ereport(ERROR,
1547 (errcode(ERRCODE_INTERNAL_ERROR),
1548 errmsg("MultiXactId %u does no longer exist -- apparent wraparound",
1549 multi)));
1550
1551 if (!MultiXactIdPrecedes(multi, nextMXact))
1552 ereport(ERROR,
1553 (errcode(ERRCODE_INTERNAL_ERROR),
1554 errmsg("MultiXactId %u has not been created yet -- apparent wraparound",
1555 multi)));
1556
1557 /*
1558 * Find out the offset at which we need to start reading MultiXactMembers
1559 * and the number of members in the multixact. We determine the latter as
1560 * the difference between this multixact's starting offset and the next
1561 * one's. However, there are some corner cases to worry about:
1562 *
1563 * 1. This multixact may be the latest one created, in which case there is
1564 * no next one to look at. The next multixact's offset should be set
1565 * already, as we set it in RecordNewMultiXact(), but we used to not do
1566 * that in older minor versions. To cope with that case, if this
1567 * multixact is the latest one created, use the nextOffset value we read
1568 * above as the endpoint.
1569 *
1570 * 2. Because GetNewMultiXactId skips over offset zero, to reserve zero
1571 * for to mean "unset", there is an ambiguity near the point of offset
1572 * wraparound. If we see next multixact's offset is one, is that our
1573 * multixact's actual endpoint, or did it end at zero with a subsequent
1574 * increment? We handle this using the knowledge that if the zero'th
1575 * member slot wasn't filled, it'll contain zero, and zero isn't a valid
1576 * transaction ID so it can't be a multixact member. Therefore, if we
1577 * read a zero from the members array, just ignore it.
1578 *
1579 * This is all pretty messy, but the mess occurs only in infrequent corner
1580 * cases, so it seems better than holding the MultiXactGenLock for a long
1581 * time on every multixact creation.
1582 */
1583 pageno = MultiXactIdToOffsetPage(multi);
1584 entryno = MultiXactIdToOffsetEntry(multi);
1585
1586 /* Acquire the bank lock for the page we need. */
1587 lock = SimpleLruGetBankLock(MultiXactOffsetCtl, pageno);
1588 LWLockAcquire(lock, LW_EXCLUSIVE);
1589
1590 slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, multi);
1591 offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
1592 offptr += entryno;
1593 offset = *offptr;
1594
1595 Assert(offset != 0);
1596
1597 /*
1598 * Use the same increment rule as GetNewMultiXactId(), that is, don't
1599 * handle wraparound explicitly until needed.
1600 */
1601 tmpMXact = multi + 1;
1602
1603 if (nextMXact == tmpMXact)
1604 {
1605 /* Corner case 1: there is no next multixact */
1606 length = nextOffset - offset;
1607 }
1608 else
1609 {
1610 MultiXactOffset nextMXOffset;
1611
1612 /* handle wraparound if needed */
1613 if (tmpMXact < FirstMultiXactId)
1614 tmpMXact = FirstMultiXactId;
1615
1616 prev_pageno = pageno;
1617
1618 pageno = MultiXactIdToOffsetPage(tmpMXact);
1619 entryno = MultiXactIdToOffsetEntry(tmpMXact);
1620
1621 if (pageno != prev_pageno)
1622 {
1623 LWLock *newlock;
1624
1625 /*
1626 * Since we're going to access a different SLRU page, if this page
1627 * falls under a different bank, release the old bank's lock and
1628 * acquire the lock of the new bank.
1629 */
1630 newlock = SimpleLruGetBankLock(MultiXactOffsetCtl, pageno);
1631 if (newlock != lock)
1632 {
1633 LWLockRelease(lock);
1634 LWLockAcquire(newlock, LW_EXCLUSIVE);
1635 lock = newlock;
1636 }
1637 slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, tmpMXact);
1638 }
1639
1640 offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
1641 offptr += entryno;
1642 nextMXOffset = *offptr;
1643
1644 if (nextMXOffset == 0)
1645 ereport(ERROR,
1646 (errcode(ERRCODE_DATA_CORRUPTED),
1647 errmsg("MultiXact %u has invalid next offset",
1648 multi)));
1649
1650 length = nextMXOffset - offset;
1651 }
1652
1653 LWLockRelease(lock);
1654 lock = NULL;
1655
1656 ptr = (MultiXactMember *) palloc(length * sizeof(MultiXactMember));
1657
1658 truelength = 0;
1659 prev_pageno = -1;
1660 for (int i = 0; i < length; i++, offset++)
1661 {
1662 TransactionId *xactptr;
1663 uint32 *flagsptr;
1664 int flagsoff;
1665 int bshift;
1666 int memberoff;
1667
1668 pageno = MXOffsetToMemberPage(offset);
1669 memberoff = MXOffsetToMemberOffset(offset);
1670
1671 if (pageno != prev_pageno)
1672 {
1673 LWLock *newlock;
1674
1675 /*
1676 * Since we're going to access a different SLRU page, if this page
1677 * falls under a different bank, release the old bank's lock and
1678 * acquire the lock of the new bank.
1679 */
1680 newlock = SimpleLruGetBankLock(MultiXactMemberCtl, pageno);
1681 if (newlock != lock)
1682 {
1683 if (lock)
1684 LWLockRelease(lock);
1685 LWLockAcquire(newlock, LW_EXCLUSIVE);
1686 lock = newlock;
1687 }
1688
1689 slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, multi);
1690 prev_pageno = pageno;
1691 }
1692
1693 xactptr = (TransactionId *)
1694 (MultiXactMemberCtl->shared->page_buffer[slotno] + memberoff);
1695
1696 if (!TransactionIdIsValid(*xactptr))
1697 {
1698 /*
1699 * Corner case 2: offset must have wrapped around to unused slot
1700 * zero.
1701 */
1702 Assert(offset == 0);
1703 continue;
1704 }
1705
1706 flagsoff = MXOffsetToFlagsOffset(offset);
1707 bshift = MXOffsetToFlagsBitShift(offset);
1708 flagsptr = (uint32 *) (MultiXactMemberCtl->shared->page_buffer[slotno] + flagsoff);
1709
1710 ptr[truelength].xid = *xactptr;
1711 ptr[truelength].status = (*flagsptr >> bshift) & MXACT_MEMBER_XACT_BITMASK;
1712 truelength++;
1713 }
1714
1715 LWLockRelease(lock);
1716
1717 /* A multixid with zero members should not happen */
1718 Assert(truelength > 0);
1719
1720 /*
1721 * Copy the result into the local cache.
1722 */
1723 mXactCachePut(multi, truelength, ptr);
1724
1725 debug_elog3(DEBUG2, "GetMembers: no cache for %s",
1726 mxid_to_string(multi, truelength, ptr));
1727 *members = ptr;
1728 return truelength;
1729 }
1730
1731 /*
1732 * mxactMemberComparator
1733 * qsort comparison function for MultiXactMember
1734 *
1735 * We can't use wraparound comparison for XIDs because that does not respect
1736 * the triangle inequality! Any old sort order will do.
1737 */
1738 static int
1739 mxactMemberComparator(const void *arg1, const void *arg2)
1740 {
1741 MultiXactMember member1 = *(const MultiXactMember *) arg1;
1742 MultiXactMember member2 = *(const MultiXactMember *) arg2;
1743
1744 if (member1.xid > member2.xid)
1745 return 1;
1746 if (member1.xid < member2.xid)
1747 return -1;
1748 if (member1.status > member2.status)
1749 return 1;
1750 if (member1.status < member2.status)
1751 return -1;
1752 return 0;
1753 }
1754
1755 /*
1756 * mXactCacheGetBySet
1757 * returns a MultiXactId from the cache based on the set of
1758 * TransactionIds that compose it, or InvalidMultiXactId if
1759 * none matches.
1760 *
1761 * This is helpful, for example, if two transactions want to lock a huge
1762 * table. By using the cache, the second will use the same MultiXactId
1763 * for the majority of tuples, thus keeping MultiXactId usage low (saving
1764 * both I/O and wraparound issues).
1765 *
1766 * NB: the passed members array will be sorted in-place.
1767 */
1768 static MultiXactId
1769 mXactCacheGetBySet(int nmembers, MultiXactMember *members)
1770 {
1771 dlist_iter iter;
1772
1773 debug_elog3(DEBUG2, "CacheGet: looking for %s",
1774 mxid_to_string(InvalidMultiXactId, nmembers, members));
1775
1776 /* sort the array so comparison is easy */
1777 qsort(members, nmembers, sizeof(MultiXactMember), mxactMemberComparator);
1778
1779 dclist_foreach(iter, &MXactCache)
1780 {
1781 mXactCacheEnt *entry = dclist_container(mXactCacheEnt, node,
1782 iter.cur);
1783
1784 if (entry->nmembers != nmembers)
1785 continue;
1786
1787 /*
1788 * We assume the cache entries are sorted, and that the unused bits in
1789 * "status" are zeroed.
1790 */
1791 if (memcmp(members, entry->members, nmembers * sizeof(MultiXactMember)) == 0)
1792 {
1793 debug_elog3(DEBUG2, "CacheGet: found %u", entry->multi);
1794 dclist_move_head(&MXactCache, iter.cur);
1795 return entry->multi;
1796 }
1797 }
1798
1799 debug_elog2(DEBUG2, "CacheGet: not found :-(");
1800 return InvalidMultiXactId;
1801 }
1802
1803 /*
1804 * mXactCacheGetById
1805 * returns the composing MultiXactMember set from the cache for a
1806 * given MultiXactId, if present.
1807 *
1808 * If successful, *xids is set to the address of a palloc'd copy of the
1809 * MultiXactMember set. Return value is number of members, or -1 on failure.
1810 */
1811 static int
1812 mXactCacheGetById(MultiXactId multi, MultiXactMember **members)
1813 {
1814 dlist_iter iter;
1815
1816 debug_elog3(DEBUG2, "CacheGet: looking for %u", multi);
1817
1818 dclist_foreach(iter, &MXactCache)
1819 {
1820 mXactCacheEnt *entry = dclist_container(mXactCacheEnt, node,
1821 iter.cur);
1822
1823 if (entry->multi == multi)
1824 {
1825 MultiXactMember *ptr;
1826 Size size;
1827
1828 size = sizeof(MultiXactMember) * entry->nmembers;
1829 ptr = (MultiXactMember *) palloc(size);
1830
1831 memcpy(ptr, entry->members, size);
1832
1833 debug_elog3(DEBUG2, "CacheGet: found %s",
1834 mxid_to_string(multi,
1835 entry->nmembers,
1836 entry->members));
1837
1838 /*
1839 * Note we modify the list while not using a modifiable iterator.
1840 * This is acceptable only because we exit the iteration
1841 * immediately afterwards.
1842 */
1843 dclist_move_head(&MXactCache, iter.cur);
1844
1845 *members = ptr;
1846 return entry->nmembers;
1847 }
1848 }
1849
1850 debug_elog2(DEBUG2, "CacheGet: not found");
1851 return -1;
1852 }
1853
1854 /*
1855 * mXactCachePut
1856 * Add a new MultiXactId and its composing set into the local cache.
1857 */
1858 static void
1859 mXactCachePut(MultiXactId multi, int nmembers, MultiXactMember *members)
1860 {
1861 mXactCacheEnt *entry;
1862
1863 debug_elog3(DEBUG2, "CachePut: storing %s",
1864 mxid_to_string(multi, nmembers, members));
1865
1866 if (MXactContext == NULL)
1867 {
1868 /* The cache only lives as long as the current transaction */
1869 debug_elog2(DEBUG2, "CachePut: initializing memory context");
1870 MXactContext = AllocSetContextCreate(TopTransactionContext,
1871 "MultiXact cache context",
1872 ALLOCSET_SMALL_SIZES);
1873 }
1874
1875 entry = (mXactCacheEnt *)
1876 MemoryContextAlloc(MXactContext,
1877 offsetof(mXactCacheEnt, members) +
1878 nmembers * sizeof(MultiXactMember));
1879
1880 entry->multi = multi;
1881 entry->nmembers = nmembers;
1882 memcpy(entry->members, members, nmembers * sizeof(MultiXactMember));
1883
1884 /* mXactCacheGetBySet assumes the entries are sorted, so sort them */
1885 qsort(entry->members, nmembers, sizeof(MultiXactMember), mxactMemberComparator);
1886
1887 dclist_push_head(&MXactCache, &entry->node);
1888 if (dclist_count(&MXactCache) > MAX_CACHE_ENTRIES)
1889 {
1890 dlist_node *node;
1891
1892 node = dclist_tail_node(&MXactCache);
1893 dclist_delete_from(&MXactCache, node);
1894
1895 entry = dclist_container(mXactCacheEnt, node, node);
1896 debug_elog3(DEBUG2, "CachePut: pruning cached multi %u",
1897 entry->multi);
1898
1899 pfree(entry);
1900 }
1901 }
1902
1903 static char *
1904 mxstatus_to_string(MultiXactStatus status)
1905 {
1906 switch (status)
1907 {
1908 case MultiXactStatusForKeyShare:
1909 return "keysh";
1910 case MultiXactStatusForShare:
1911 return "sh";
1912 case MultiXactStatusForNoKeyUpdate:
1913 return "fornokeyupd";
1914 case MultiXactStatusForUpdate:
1915 return "forupd";
1916 case MultiXactStatusNoKeyUpdate:
1917 return "nokeyupd";
1918 case MultiXactStatusUpdate:
1919 return "upd";
1920 default:
1921 elog(ERROR, "unrecognized multixact status %d", status);
1922 return "";
1923 }
1924 }
1925
1926 char *
1927 mxid_to_string(MultiXactId multi, int nmembers, MultiXactMember *members)
1928 {
1929 static char *str = NULL;
1930 StringInfoData buf;
1931 int i;
1932
1933 if (str != NULL)
1934 pfree(str);
1935
1936 initStringInfo(&buf);
1937
1938 appendStringInfo(&buf, "%u %d[%u (%s)", multi, nmembers, members[0].xid,
1939 mxstatus_to_string(members[0].status));
1940
1941 for (i = 1; i < nmembers; i++)
1942 appendStringInfo(&buf, ", %u (%s)", members[i].xid,
1943 mxstatus_to_string(members[i].status));
1944
1945 appendStringInfoChar(&buf, ']');
1946 str = MemoryContextStrdup(TopMemoryContext, buf.data);
1947 pfree(buf.data);
1948 return str;
1949 }
1950
1951 /*
1952 * AtEOXact_MultiXact
1953 * Handle transaction end for MultiXact
1954 *
1955 * This is called at top transaction commit or abort (we don't care which).
1956 */
1957 void
1958 AtEOXact_MultiXact(void)
1959 {
1960 /*
1961 * Reset our OldestMemberMXactId and OldestVisibleMXactId values, both of
1962 * which should only be valid while within a transaction.
1963 *
1964 * We assume that storing a MultiXactId is atomic and so we need not take
1965 * MultiXactGenLock to do this.
1966 */
1967 *MyOldestMemberMXactIdSlot() = InvalidMultiXactId;
1968 *MyOldestVisibleMXactIdSlot() = InvalidMultiXactId;
1969
1970 /*
1971 * Discard the local MultiXactId cache. Since MXactContext was created as
1972 * a child of TopTransactionContext, we needn't delete it explicitly.
1973 */
1974 MXactContext = NULL;
1975 dclist_init(&MXactCache);
1976 }
1977
1978 /*
1979 * AtPrepare_MultiXact
1980 * Save multixact state at 2PC transaction prepare
1981 *
1982 * In this phase, we only store our OldestMemberMXactId value in the two-phase
1983 * state file.
1984 */
1985 void
1986 AtPrepare_MultiXact(void)
1987 {
1988 MultiXactId myOldestMember = *MyOldestMemberMXactIdSlot();
1989
1990 if (MultiXactIdIsValid(myOldestMember))
1991 RegisterTwoPhaseRecord(TWOPHASE_RM_MULTIXACT_ID, 0,
1992 &myOldestMember, sizeof(MultiXactId));
1993 }
1994
1995 /*
1996 * PostPrepare_MultiXact
1997 * Clean up after successful PREPARE TRANSACTION
1998 */
1999 void
2000 PostPrepare_MultiXact(TransactionId xid)
2001 {
2002 MultiXactId myOldestMember;
2003
2004 /*
2005 * Transfer our OldestMemberMXactId value to the slot reserved for the
2006 * prepared transaction.
2007 */
2008 myOldestMember = *MyOldestMemberMXactIdSlot();
2009 if (MultiXactIdIsValid(myOldestMember))
2010 {
2011 ProcNumber dummyProcNumber = TwoPhaseGetDummyProcNumber(xid, false);
2012
2013 /*
2014 * Even though storing MultiXactId is atomic, acquire lock to make
2015 * sure others see both changes, not just the reset of the slot of the
2016 * current backend. Using a volatile pointer might suffice, but this
2017 * isn't a hot spot.
2018 */
2019 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2020
2021 *PreparedXactOldestMemberMXactIdSlot(dummyProcNumber) = myOldestMember;
2022 *MyOldestMemberMXactIdSlot() = InvalidMultiXactId;
2023
2024 LWLockRelease(MultiXactGenLock);
2025 }
2026
2027 /*
2028 * We don't need to transfer OldestVisibleMXactId value, because the
2029 * transaction is not going to be looking at any more multixacts once it's
2030 * prepared.
2031 *
2032 * We assume that storing a MultiXactId is atomic and so we need not take
2033 * MultiXactGenLock to do this.
2034 */
2035 *MyOldestVisibleMXactIdSlot() = InvalidMultiXactId;
2036
2037 /*
2038 * Discard the local MultiXactId cache like in AtEOXact_MultiXact.
2039 */
2040 MXactContext = NULL;
2041 dclist_init(&MXactCache);
2042 }
2043
2044 /*
2045 * multixact_twophase_recover
2046 * Recover the state of a prepared transaction at startup
2047 */
2048 void
2049 multixact_twophase_recover(TransactionId xid, uint16 info,
2050 void *recdata, uint32 len)
2051 {
2052 ProcNumber dummyProcNumber = TwoPhaseGetDummyProcNumber(xid, false);
2053 MultiXactId oldestMember;
2054
2055 /*
2056 * Get the oldest member XID from the state file record, and set it in the
2057 * OldestMemberMXactId slot reserved for this prepared transaction.
2058 */
2059 Assert(len == sizeof(MultiXactId));
2060 oldestMember = *((MultiXactId *) recdata);
2061
2062 *PreparedXactOldestMemberMXactIdSlot(dummyProcNumber) = oldestMember;
2063 }
2064
2065 /*
2066 * multixact_twophase_postcommit
2067 * Similar to AtEOXact_MultiXact but for COMMIT PREPARED
2068 */
2069 void
2070 multixact_twophase_postcommit(TransactionId xid, uint16 info,
2071 void *recdata, uint32 len)
2072 {
2073 ProcNumber dummyProcNumber = TwoPhaseGetDummyProcNumber(xid, true);
2074
2075 Assert(len == sizeof(MultiXactId));
2076
2077 *PreparedXactOldestMemberMXactIdSlot(dummyProcNumber) = InvalidMultiXactId;
2078 }
2079
2080 /*
2081 * multixact_twophase_postabort
2082 * This is actually just the same as the COMMIT case.
2083 */
2084 void
2085 multixact_twophase_postabort(TransactionId xid, uint16 info,
2086 void *recdata, uint32 len)
2087 {
2088 multixact_twophase_postcommit(xid, info, recdata, len);
2089 }
2090
2091 /*
2092 * Initialization of shared memory for MultiXact.
2093 *
2094 * MultiXactSharedStateShmemSize() calculates the size of the MultiXactState
2095 * struct, and the two per-backend MultiXactId arrays. They are carved out of
2096 * the same allocation. MultiXactShmemSize() additionally includes the memory
2097 * needed for the two SLRU areas.
2098 */
2099 static Size
2100 MultiXactSharedStateShmemSize(void)
2101 {
2102 Size size;
2103
2104 size = offsetof(MultiXactStateData, perBackendXactIds);
2105 size = add_size(size,
2106 mul_size(sizeof(MultiXactId), NumMemberSlots));
2107 size = add_size(size,
2108 mul_size(sizeof(MultiXactId), NumVisibleSlots));
2109 return size;
2110 }
2111
2112 Size
2113 MultiXactShmemSize(void)
2114 {
2115 Size size;
2116
2117 size = MultiXactSharedStateShmemSize();
2118 size = add_size(size, SimpleLruShmemSize(multixact_offset_buffers, 0));
2119 size = add_size(size, SimpleLruShmemSize(multixact_member_buffers, 0));
2120
2121 return size;
2122 }
2123
2124 void
2125 MultiXactShmemInit(void)
2126 {
2127 bool found;
2128
2129 debug_elog2(DEBUG2, "Shared Memory Init for MultiXact");
2130
2131 MultiXactOffsetCtl->PagePrecedes = MultiXactOffsetPagePrecedes;
2132 MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
2133
2134 SimpleLruInit(MultiXactOffsetCtl,
2135 "multixact_offset", multixact_offset_buffers, 0,
2136 "pg_multixact/offsets", LWTRANCHE_MULTIXACTOFFSET_BUFFER,
2137 LWTRANCHE_MULTIXACTOFFSET_SLRU,
2138 SYNC_HANDLER_MULTIXACT_OFFSET,
2139 false);
2140 SlruPagePrecedesUnitTests(MultiXactOffsetCtl, MULTIXACT_OFFSETS_PER_PAGE);
2141 SimpleLruInit(MultiXactMemberCtl,
2142 "multixact_member", multixact_member_buffers, 0,
2143 "pg_multixact/members", LWTRANCHE_MULTIXACTMEMBER_BUFFER,
2144 LWTRANCHE_MULTIXACTMEMBER_SLRU,
2145 SYNC_HANDLER_MULTIXACT_MEMBER,
2146 false);
2147 /* doesn't call SimpleLruTruncate() or meet criteria for unit tests */
2148
2149 /* Initialize our shared state struct */
2150 MultiXactState = ShmemInitStruct("Shared MultiXact State",
2151 MultiXactSharedStateShmemSize(),
2152 &found);
2153 if (!IsUnderPostmaster)
2154 {
2155 Assert(!found);
2156
2157 /* Make sure we zero out the per-backend state */
2158 MemSet(MultiXactState, 0, MultiXactSharedStateShmemSize());
2159 }
2160 else
2161 Assert(found);
2162
2163 /*
2164 * Set up array pointers.
2165 */
2166 OldestMemberMXactId = MultiXactState->perBackendXactIds;
2167 OldestVisibleMXactId = OldestMemberMXactId + NumMemberSlots;
2168 }
2169
2170 /*
2171 * GUC check_hook for multixact_offset_buffers
2172 */
2173 bool
2174 check_multixact_offset_buffers(int *newval, void **extra, GucSource source)
2175 {
2176 return check_slru_buffers("multixact_offset_buffers", newval);
2177 }
2178
2179 /*
2180 * GUC check_hook for multixact_member_buffers
2181 */
2182 bool
2183 check_multixact_member_buffers(int *newval, void **extra, GucSource source)
2184 {
2185 return check_slru_buffers("multixact_member_buffers", newval);
2186 }
2187
2188 /*
2189 * This func must be called ONCE on system install. It creates the initial
2190 * MultiXact segments. (The MultiXacts directories are assumed to have been
2191 * created by initdb, and MultiXactShmemInit must have been called already.)
2192 */
2193 void
2194 BootStrapMultiXact(void)
2195 {
2196 int slotno;
2197 LWLock *lock;
2198
2199 lock = SimpleLruGetBankLock(MultiXactOffsetCtl, 0);
2200 LWLockAcquire(lock, LW_EXCLUSIVE);
2201
2202 /* Create and zero the first page of the offsets log */
2203 slotno = ZeroMultiXactOffsetPage(0, false);
2204
2205 /* Make sure it's written out */
2206 SimpleLruWritePage(MultiXactOffsetCtl, slotno);
2207 Assert(!MultiXactOffsetCtl->shared->page_dirty[slotno]);
2208
2209 LWLockRelease(lock);
2210
2211 lock = SimpleLruGetBankLock(MultiXactMemberCtl, 0);
2212 LWLockAcquire(lock, LW_EXCLUSIVE);
2213
2214 /* Create and zero the first page of the members log */
2215 slotno = ZeroMultiXactMemberPage(0, false);
2216
2217 /* Make sure it's written out */
2218 SimpleLruWritePage(MultiXactMemberCtl, slotno);
2219 Assert(!MultiXactMemberCtl->shared->page_dirty[slotno]);
2220
2221 LWLockRelease(lock);
2222 }
2223
2224 /*
2225 * Initialize (or reinitialize) a page of MultiXactOffset to zeroes.
2226 * If writeXlog is true, also emit an XLOG record saying we did this.
2227 *
2228 * The page is not actually written, just set up in shared memory.
2229 * The slot number of the new page is returned.
2230 *
2231 * Control lock must be held at entry, and will be held at exit.
2232 */
2233 static int
2234 ZeroMultiXactOffsetPage(int64 pageno, bool writeXlog)
2235 {
2236 int slotno;
2237
2238 slotno = SimpleLruZeroPage(MultiXactOffsetCtl, pageno);
2239
2240 if (writeXlog)
2241 WriteMZeroPageXlogRec(pageno, XLOG_MULTIXACT_ZERO_OFF_PAGE);
2242
2243 return slotno;
2244 }
2245
2246 /*
2247 * Ditto, for MultiXactMember
2248 */
2249 static int
2250 ZeroMultiXactMemberPage(int64 pageno, bool writeXlog)
2251 {
2252 int slotno;
2253
2254 slotno = SimpleLruZeroPage(MultiXactMemberCtl, pageno);
2255
2256 if (writeXlog)
2257 WriteMZeroPageXlogRec(pageno, XLOG_MULTIXACT_ZERO_MEM_PAGE);
2258
2259 return slotno;
2260 }
2261
2262 /*
2263 * MaybeExtendOffsetSlru
2264 * Extend the offsets SLRU area, if necessary
2265 *
2266 * After a binary upgrade from <= 9.2, the pg_multixact/offsets SLRU area might
2267 * contain files that are shorter than necessary; this would occur if the old
2268 * installation had used multixacts beyond the first page (files cannot be
2269 * copied, because the on-disk representation is different). pg_upgrade would
2270 * update pg_control to set the next offset value to be at that position, so
2271 * that tuples marked as locked by such MultiXacts would be seen as visible
2272 * without having to consult multixact. However, trying to create and use a
2273 * new MultiXactId would result in an error because the page on which the new
2274 * value would reside does not exist. This routine is in charge of creating
2275 * such pages.
2276 */
2277 static void
2278 MaybeExtendOffsetSlru(void)
2279 {
2280 int64 pageno;
2281 LWLock *lock;
2282
2283 pageno = MultiXactIdToOffsetPage(MultiXactState->nextMXact);
2284 lock = SimpleLruGetBankLock(MultiXactOffsetCtl, pageno);
2285
2286 LWLockAcquire(lock, LW_EXCLUSIVE);
2287
2288 if (!SimpleLruDoesPhysicalPageExist(MultiXactOffsetCtl, pageno))
2289 {
2290 int slotno;
2291
2292 /*
2293 * Fortunately for us, SimpleLruWritePage is already prepared to deal
2294 * with creating a new segment file even if the page we're writing is
2295 * not the first in it, so this is enough.
2296 */
2297 slotno = ZeroMultiXactOffsetPage(pageno, false);
2298 SimpleLruWritePage(MultiXactOffsetCtl, slotno);
2299 }
2300
2301 LWLockRelease(lock);
2302 }
2303
2304 /*
2305 * This must be called ONCE during postmaster or standalone-backend startup.
2306 *
2307 * StartupXLOG has already established nextMXact/nextOffset by calling
2308 * MultiXactSetNextMXact and/or MultiXactAdvanceNextMXact, and the oldestMulti
2309 * info from pg_control and/or MultiXactAdvanceOldest, but we haven't yet
2310 * replayed WAL.
2311 */
2312 void
2313 StartupMultiXact(void)
2314 {
2315 MultiXactId multi = MultiXactState->nextMXact;
2316 MultiXactOffset offset = MultiXactState->nextOffset;
2317 int64 pageno;
2318
2319 /*
2320 * Initialize offset's idea of the latest page number.
2321 */
2322 pageno = MultiXactIdToOffsetPage(multi);
2323 pg_atomic_write_u64(&MultiXactOffsetCtl->shared->latest_page_number,
2324 pageno);
2325
2326 /*
2327 * Initialize member's idea of the latest page number.
2328 */
2329 pageno = MXOffsetToMemberPage(offset);
2330 pg_atomic_write_u64(&MultiXactMemberCtl->shared->latest_page_number,
2331 pageno);
2332 }
2333
2334 /*
2335 * This must be called ONCE at the end of startup/recovery.
2336 */
2337 void
2338 TrimMultiXact(void)
2339 {
2340 MultiXactId nextMXact;
2341 MultiXactOffset offset;
2342 MultiXactId oldestMXact;
2343 Oid oldestMXactDB;
2344 int64 pageno;
2345 int entryno;
2346 int flagsoff;
2347
2348 LWLockAcquire(MultiXactGenLock, LW_SHARED);
2349 nextMXact = MultiXactState->nextMXact;
2350 offset = MultiXactState->nextOffset;
2351 oldestMXact = MultiXactState->oldestMultiXactId;
2352 oldestMXactDB = MultiXactState->oldestMultiXactDB;
2353 LWLockRelease(MultiXactGenLock);
2354
2355 /* Clean up offsets state */
2356
2357 /*
2358 * (Re-)Initialize our idea of the latest page number for offsets.
2359 */
2360 pageno = MultiXactIdToOffsetPage(nextMXact);
2361 pg_atomic_write_u64(&MultiXactOffsetCtl->shared->latest_page_number,
2362 pageno);
2363
2364 /*
2365 * Set the offset of nextMXact on the offsets page. This is normally done
2366 * in RecordNewMultiXact() of the previous multixact, but we used to not
2367 * do that in older minor versions. To ensure that the next offset is set
2368 * if the binary was just upgraded from an older minor version, do it now.
2369 *
2370 * Zero out the remainder of the page. See notes in TrimCLOG() for
2371 * background. Unlike CLOG, some WAL record covers every pg_multixact
2372 * SLRU mutation. Since, also unlike CLOG, we ignore the WAL rule "write
2373 * xlog before data," nextMXact successors may carry obsolete, nonzero
2374 * offset values.
2375 */
2376 entryno = MultiXactIdToOffsetEntry(nextMXact);
2377 {
2378 int slotno;
2379 MultiXactOffset *offptr;
2380 LWLock *lock = SimpleLruGetBankLock(MultiXactOffsetCtl, pageno);
2381
2382 LWLockAcquire(lock, LW_EXCLUSIVE);
2383 if (entryno == 0)
2384 slotno = SimpleLruZeroPage(MultiXactOffsetCtl, pageno);
2385 else
2386 slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, nextMXact);
2387 offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
2388 offptr += entryno;
2389
2390 *offptr = offset;
2391 if (entryno != 0 && (entryno + 1) * sizeof(MultiXactOffset) != BLCKSZ)
2392 MemSet(offptr + 1, 0, BLCKSZ - (entryno + 1) * sizeof(MultiXactOffset));
2393
2394 MultiXactOffsetCtl->shared->page_dirty[slotno] = true;
2395 LWLockRelease(lock);
2396 }
2397
2398 /*
2399 * And the same for members.
2400 *
2401 * (Re-)Initialize our idea of the latest page number for members.
2402 */
2403 pageno = MXOffsetToMemberPage(offset);
2404 pg_atomic_write_u64(&MultiXactMemberCtl->shared->latest_page_number,
2405 pageno);
2406
2407 /*
2408 * Zero out the remainder of the current members page. See notes in
2409 * TrimCLOG() for motivation.
2410 */
2411 flagsoff = MXOffsetToFlagsOffset(offset);
2412 if (flagsoff != 0)
2413 {
2414 int slotno;
2415 TransactionId *xidptr;
2416 int memberoff;
2417 LWLock *lock = SimpleLruGetBankLock(MultiXactMemberCtl, pageno);
2418
2419 LWLockAcquire(lock, LW_EXCLUSIVE);
2420 memberoff = MXOffsetToMemberOffset(offset);
2421 slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, offset);
2422 xidptr = (TransactionId *)
2423 (MultiXactMemberCtl->shared->page_buffer[slotno] + memberoff);
2424
2425 MemSet(xidptr, 0, BLCKSZ - memberoff);
2426
2427 /*
2428 * Note: we don't need to zero out the flag bits in the remaining
2429 * members of the current group, because they are always reset before
2430 * writing.
2431 */
2432
2433 MultiXactMemberCtl->shared->page_dirty[slotno] = true;
2434 LWLockRelease(lock);
2435 }
2436
2437 /* signal that we're officially up */
2438 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2439 MultiXactState->finishedStartup = true;
2440 LWLockRelease(MultiXactGenLock);
2441
2442 /* Now compute how far away the next members wraparound is. */
2443 SetMultiXactIdLimit(oldestMXact, oldestMXactDB, true);
2444 }
2445
2446 /*
2447 * Get the MultiXact data to save in a checkpoint record
2448 */
2449 void
2450 MultiXactGetCheckptMulti(bool is_shutdown,
2451 MultiXactId *nextMulti,
2452 MultiXactOffset *nextMultiOffset,
2453 MultiXactId *oldestMulti,
2454 Oid *oldestMultiDB)
2455 {
2456 LWLockAcquire(MultiXactGenLock, LW_SHARED);
2457 *nextMulti = MultiXactState->nextMXact;
2458 *nextMultiOffset = MultiXactState->nextOffset;
2459 *oldestMulti = MultiXactState->oldestMultiXactId;
2460 *oldestMultiDB = MultiXactState->oldestMultiXactDB;
2461 LWLockRelease(MultiXactGenLock);
2462
2463 debug_elog6(DEBUG2,
2464 "MultiXact: checkpoint is nextMulti %u, nextOffset %u, oldestMulti %u in DB %u",
2465 *nextMulti, *nextMultiOffset, *oldestMulti, *oldestMultiDB);
2466 }
2467
2468 /*
2469 * Perform a checkpoint --- either during shutdown, or on-the-fly
2470 */
2471 void
2472 CheckPointMultiXact(void)
2473 {
2474 TRACE_POSTGRESQL_MULTIXACT_CHECKPOINT_START(true);
2475
2476 /*
2477 * Write dirty MultiXact pages to disk. This may result in sync requests
2478 * queued for later handling by ProcessSyncRequests(), as part of the
2479 * checkpoint.
2480 */
2481 SimpleLruWriteAll(MultiXactOffsetCtl, true);
2482 SimpleLruWriteAll(MultiXactMemberCtl, true);
2483
2484 TRACE_POSTGRESQL_MULTIXACT_CHECKPOINT_DONE(true);
2485 }
2486
2487 /*
2488 * Set the next-to-be-assigned MultiXactId and offset
2489 *
2490 * This is used when we can determine the correct next ID/offset exactly
2491 * from a checkpoint record. Although this is only called during bootstrap
2492 * and XLog replay, we take the lock in case any hot-standby backends are
2493 * examining the values.
2494 */
2495 void
2496 MultiXactSetNextMXact(MultiXactId nextMulti,
2497 MultiXactOffset nextMultiOffset)
2498 {
2499 debug_elog4(DEBUG2, "MultiXact: setting next multi to %u offset %u",
2500 nextMulti, nextMultiOffset);
2501 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2502 MultiXactState->nextMXact = nextMulti;
2503 MultiXactState->nextOffset = nextMultiOffset;
2504 LWLockRelease(MultiXactGenLock);
2505
2506 /*
2507 * During a binary upgrade, make sure that the offsets SLRU is large
2508 * enough to contain the next value that would be created.
2509 *
2510 * We need to do this pretty early during the first startup in binary
2511 * upgrade mode: before StartupMultiXact() in fact, because this routine
2512 * is called even before that by StartupXLOG(). And we can't do it
2513 * earlier than at this point, because during that first call of this
2514 * routine we determine the MultiXactState->nextMXact value that
2515 * MaybeExtendOffsetSlru needs.
2516 */
2517 if (IsBinaryUpgrade)
2518 MaybeExtendOffsetSlru();
2519 }
2520
2521 /*
2522 * Determine the last safe MultiXactId to allocate given the currently oldest
2523 * datminmxid (ie, the oldest MultiXactId that might exist in any database
2524 * of our cluster), and the OID of the (or a) database with that value.
2525 *
2526 * is_startup is true when we are just starting the cluster, false when we
2527 * are updating state in a running cluster. This only affects log messages.
2528 */
2529 void
2530 SetMultiXactIdLimit(MultiXactId oldest_datminmxid, Oid oldest_datoid,
2531 bool is_startup)
2532 {
2533 MultiXactId multiVacLimit;
2534 MultiXactId multiWarnLimit;
2535 MultiXactId multiStopLimit;
2536 MultiXactId multiWrapLimit;
2537 MultiXactId curMulti;
2538 bool needs_offset_vacuum;
2539
2540 Assert(MultiXactIdIsValid(oldest_datminmxid));
2541
2542 /*
2543 * We pretend that a wrap will happen halfway through the multixact ID
2544 * space, but that's not really true, because multixacts wrap differently
2545 * from transaction IDs. Note that, separately from any concern about
2546 * multixact IDs wrapping, we must ensure that multixact members do not
2547 * wrap. Limits for that are set in SetOffsetVacuumLimit, not here.
2548 */
2549 multiWrapLimit = oldest_datminmxid + (MaxMultiXactId >> 1);
2550 if (multiWrapLimit < FirstMultiXactId)
2551 multiWrapLimit += FirstMultiXactId;
2552
2553 /*
2554 * We'll refuse to continue assigning MultiXactIds once we get within 3M
2555 * multi of data loss. See SetTransactionIdLimit.
2556 */
2557 multiStopLimit = multiWrapLimit - 3000000;
2558 if (multiStopLimit < FirstMultiXactId)
2559 multiStopLimit -= FirstMultiXactId;
2560
2561 /*
2562 * We'll start complaining loudly when we get within 40M multis of data
2563 * loss. This is kind of arbitrary, but if you let your gas gauge get
2564 * down to 2% of full, would you be looking for the next gas station? We
2565 * need to be fairly liberal about this number because there are lots of
2566 * scenarios where most transactions are done by automatic clients that
2567 * won't pay attention to warnings. (No, we're not gonna make this
2568 * configurable. If you know enough to configure it, you know enough to
2569 * not get in this kind of trouble in the first place.)
2570 */
2571 multiWarnLimit = multiWrapLimit - 40000000;
2572 if (multiWarnLimit < FirstMultiXactId)
2573 multiWarnLimit -= FirstMultiXactId;
2574
2575 /*
2576 * We'll start trying to force autovacuums when oldest_datminmxid gets to
2577 * be more than autovacuum_multixact_freeze_max_age mxids old.
2578 *
2579 * Note: autovacuum_multixact_freeze_max_age is a PGC_POSTMASTER parameter
2580 * so that we don't have to worry about dealing with on-the-fly changes in
2581 * its value. See SetTransactionIdLimit.
2582 */
2583 multiVacLimit = oldest_datminmxid + autovacuum_multixact_freeze_max_age;
2584 if (multiVacLimit < FirstMultiXactId)
2585 multiVacLimit += FirstMultiXactId;
2586
2587 /* Grab lock for just long enough to set the new limit values */
2588 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2589 MultiXactState->oldestMultiXactId = oldest_datminmxid;
2590 MultiXactState->oldestMultiXactDB = oldest_datoid;
2591 MultiXactState->multiVacLimit = multiVacLimit;
2592 MultiXactState->multiWarnLimit = multiWarnLimit;
2593 MultiXactState->multiStopLimit = multiStopLimit;
2594 MultiXactState->multiWrapLimit = multiWrapLimit;
2595 curMulti = MultiXactState->nextMXact;
2596 LWLockRelease(MultiXactGenLock);
2597
2598 /* Log the info */
2599 ereport(DEBUG1,
2600 (errmsg_internal("MultiXactId wrap limit is %u, limited by database with OID %u",
2601 multiWrapLimit, oldest_datoid)));
2602
2603 /*
2604 * Computing the actual limits is only possible once the data directory is
2605 * in a consistent state. There's no need to compute the limits while
2606 * still replaying WAL - no decisions about new multis are made even
2607 * though multixact creations might be replayed. So we'll only do further
2608 * checks after TrimMultiXact() has been called.
2609 */
2610 if (!MultiXactState->finishedStartup)
2611 return;
2612
2613 Assert(!InRecovery);
2614
2615 /* Set limits for offset vacuum. */
2616 needs_offset_vacuum = SetOffsetVacuumLimit(is_startup);
2617
2618 /*
2619 * If past the autovacuum force point, immediately signal an autovac
2620 * request. The reason for this is that autovac only processes one
2621 * database per invocation. Once it's finished cleaning up the oldest
2622 * database, it'll call here, and we'll signal the postmaster to start
2623 * another iteration immediately if there are still any old databases.
2624 */
2625 if ((MultiXactIdPrecedes(multiVacLimit, curMulti) ||
2626 needs_offset_vacuum) && IsUnderPostmaster)
2627 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
2628
2629 /* Give an immediate warning if past the wrap warn point */
2630 if (MultiXactIdPrecedes(multiWarnLimit, curMulti))
2631 {
2632 char *oldest_datname;
2633
2634 /*
2635 * We can be called when not inside a transaction, for example during
2636 * StartupXLOG(). In such a case we cannot do database access, so we
2637 * must just report the oldest DB's OID.
2638 *
2639 * Note: it's also possible that get_database_name fails and returns
2640 * NULL, for example because the database just got dropped. We'll
2641 * still warn, even though the warning might now be unnecessary.
2642 */
2643 if (IsTransactionState())
2644 oldest_datname = get_database_name(oldest_datoid);
2645 else
2646 oldest_datname = NULL;
2647
2648 if (oldest_datname)
2649 ereport(WARNING,
2650 (errmsg_plural("database \"%s\" must be vacuumed before %u more MultiXactId is used",
2651 "database \"%s\" must be vacuumed before %u more MultiXactIds are used",
2652 multiWrapLimit - curMulti,
2653 oldest_datname,
2654 multiWrapLimit - curMulti),
2655 errhint("To avoid MultiXactId assignment failures, execute a database-wide VACUUM in that database.\n"
2656 "You might also need to commit or roll back old prepared transactions.")));
2657 else
2658 ereport(WARNING,
2659 (errmsg_plural("database with OID %u must be vacuumed before %u more MultiXactId is used",
2660 "database with OID %u must be vacuumed before %u more MultiXactIds are used",
2661 multiWrapLimit - curMulti,
2662 oldest_datoid,
2663 multiWrapLimit - curMulti),
2664 errhint("To avoid MultiXactId assignment failures, execute a database-wide VACUUM in that database.\n"
2665 "You might also need to commit or roll back old prepared transactions.")));
2666 }
2667 }
2668
2669 /*
2670 * Ensure the next-to-be-assigned MultiXactId is at least minMulti,
2671 * and similarly nextOffset is at least minMultiOffset.
2672 *
2673 * This is used when we can determine minimum safe values from an XLog
2674 * record (either an on-line checkpoint or an mxact creation log entry).
2675 * Although this is only called during XLog replay, we take the lock in case
2676 * any hot-standby backends are examining the values.
2677 */
2678 void
2679 MultiXactAdvanceNextMXact(MultiXactId minMulti,
2680 MultiXactOffset minMultiOffset)
2681 {
2682 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2683 if (MultiXactIdPrecedes(MultiXactState->nextMXact, minMulti))
2684 {
2685 debug_elog3(DEBUG2, "MultiXact: setting next multi to %u", minMulti);
2686 MultiXactState->nextMXact = minMulti;
2687 }
2688 if (MultiXactOffsetPrecedes(MultiXactState->nextOffset, minMultiOffset))
2689 {
2690 debug_elog3(DEBUG2, "MultiXact: setting next offset to %u",
2691 minMultiOffset);
2692 MultiXactState->nextOffset = minMultiOffset;
2693 }
2694 LWLockRelease(MultiXactGenLock);
2695 }
2696
2697 /*
2698 * Update our oldestMultiXactId value, but only if it's more recent than what
2699 * we had.
2700 *
2701 * This may only be called during WAL replay.
2702 */
2703 void
2704 MultiXactAdvanceOldest(MultiXactId oldestMulti, Oid oldestMultiDB)
2705 {
2706 Assert(InRecovery);
2707
2708 if (MultiXactIdPrecedes(MultiXactState->oldestMultiXactId, oldestMulti))
2709 SetMultiXactIdLimit(oldestMulti, oldestMultiDB, false);
2710 }
2711
2712 /*
2713 * Make sure that MultiXactOffset has room for a newly-allocated MultiXactId.
2714 *
2715 * NB: this is called while holding MultiXactGenLock. We want it to be very
2716 * fast most of the time; even when it's not so fast, no actual I/O need
2717 * happen unless we're forced to write out a dirty log or xlog page to make
2718 * room in shared memory.
2719 */
2720 static void
2721 ExtendMultiXactOffset(MultiXactId multi)
2722 {
2723 int64 pageno;
2724 LWLock *lock;
2725
2726 /*
2727 * No work except at first MultiXactId of a page. But beware: just after
2728 * wraparound, the first MultiXactId of page zero is FirstMultiXactId.
2729 */
2730 if (MultiXactIdToOffsetEntry(multi) != 0 &&
2731 multi != FirstMultiXactId)
2732 return;
2733
2734 pageno = MultiXactIdToOffsetPage(multi);
2735 lock = SimpleLruGetBankLock(MultiXactOffsetCtl, pageno);
2736
2737 LWLockAcquire(lock, LW_EXCLUSIVE);
2738
2739 /* Zero the page and make an XLOG entry about it */
2740 ZeroMultiXactOffsetPage(pageno, true);
2741
2742 LWLockRelease(lock);
2743 }
2744
2745 /*
2746 * Make sure that MultiXactMember has room for the members of a newly-
2747 * allocated MultiXactId.
2748 *
2749 * Like the above routine, this is called while holding MultiXactGenLock;
2750 * same comments apply.
2751 */
2752 static void
2753 ExtendMultiXactMember(MultiXactOffset offset, int nmembers)
2754 {
2755 /*
2756 * It's possible that the members span more than one page of the members
2757 * file, so we loop to ensure we consider each page. The coding is not
2758 * optimal if the members span several pages, but that seems unusual
2759 * enough to not worry much about.
2760 */
2761 while (nmembers > 0)
2762 {
2763 int flagsoff;
2764 int flagsbit;
2765 uint32 difference;
2766
2767 /*
2768 * Only zero when at first entry of a page.
2769 */
2770 flagsoff = MXOffsetToFlagsOffset(offset);
2771 flagsbit = MXOffsetToFlagsBitShift(offset);
2772 if (flagsoff == 0 && flagsbit == 0)
2773 {
2774 int64 pageno;
2775 LWLock *lock;
2776
2777 pageno = MXOffsetToMemberPage(offset);
2778 lock = SimpleLruGetBankLock(MultiXactMemberCtl, pageno);
2779
2780 LWLockAcquire(lock, LW_EXCLUSIVE);
2781
2782 /* Zero the page and make an XLOG entry about it */
2783 ZeroMultiXactMemberPage(pageno, true);
2784
2785 LWLockRelease(lock);
2786 }
2787
2788 /*
2789 * Compute the number of items till end of current page. Careful: if
2790 * addition of unsigned ints wraps around, we're at the last page of
2791 * the last segment; since that page holds a different number of items
2792 * than other pages, we need to do it differently.
2793 */
2794 if (offset + MAX_MEMBERS_IN_LAST_MEMBERS_PAGE < offset)
2795 {
2796 /*
2797 * This is the last page of the last segment; we can compute the
2798 * number of items left to allocate in it without modulo
2799 * arithmetic.
2800 */
2801 difference = MaxMultiXactOffset - offset + 1;
2802 }
2803 else
2804 difference = MULTIXACT_MEMBERS_PER_PAGE - offset % MULTIXACT_MEMBERS_PER_PAGE;
2805
2806 /*
2807 * Advance to next page, taking care to properly handle the wraparound
2808 * case. OK if nmembers goes negative.
2809 */
2810 nmembers -= difference;
2811 offset += difference;
2812 }
2813 }
2814
2815 /*
2816 * GetOldestMultiXactId
2817 *
2818 * Return the oldest MultiXactId that's still possibly still seen as live by
2819 * any running transaction. Older ones might still exist on disk, but they no
2820 * longer have any running member transaction.
2821 *
2822 * It's not safe to truncate MultiXact SLRU segments on the value returned by
2823 * this function; however, it can be set as the new relminmxid for any table
2824 * that VACUUM knows has no remaining MXIDs < the same value. It is only safe
2825 * to truncate SLRUs when no table can possibly still have a referencing MXID.
2826 */
2827 MultiXactId
2828 GetOldestMultiXactId(void)
2829 {
2830 MultiXactId oldestMXact;
2831 MultiXactId nextMXact;
2832
2833 /*
2834 * This is the oldest valid value among all the OldestMemberMXactId[] and
2835 * OldestVisibleMXactId[] entries, or nextMXact if none are valid.
2836 */
2837 LWLockAcquire(MultiXactGenLock, LW_SHARED);
2838
2839 /*
2840 * We have to beware of the possibility that nextMXact is in the
2841 * wrapped-around state. We don't fix the counter itself here, but we
2842 * must be sure to use a valid value in our calculation.
2843 */
2844 nextMXact = MultiXactState->nextMXact;
2845 if (nextMXact < FirstMultiXactId)
2846 nextMXact = FirstMultiXactId;
2847
2848 oldestMXact = nextMXact;
2849 for (int i = 0; i < NumMemberSlots; i++)
2850 {
2851 MultiXactId thisoldest;
2852
2853 thisoldest = OldestMemberMXactId[i];
2854 if (MultiXactIdIsValid(thisoldest) &&
2855 MultiXactIdPrecedes(thisoldest, oldestMXact))
2856 oldestMXact = thisoldest;
2857 }
2858 for (int i = 0; i < NumVisibleSlots; i++)
2859 {
2860 MultiXactId thisoldest;
2861
2862 thisoldest = OldestVisibleMXactId[i];
2863 if (MultiXactIdIsValid(thisoldest) &&
2864 MultiXactIdPrecedes(thisoldest, oldestMXact))
2865 oldestMXact = thisoldest;
2866 }
2867
2868 LWLockRelease(MultiXactGenLock);
2869
2870 return oldestMXact;
2871 }
2872
2873 /*
2874 * Determine how aggressively we need to vacuum in order to prevent member
2875 * wraparound.
2876 *
2877 * To do so determine what's the oldest member offset and install the limit
2878 * info in MultiXactState, where it can be used to prevent overrun of old data
2879 * in the members SLRU area.
2880 *
2881 * The return value is true if emergency autovacuum is required and false
2882 * otherwise.
2883 */
2884 static bool
2885 SetOffsetVacuumLimit(bool is_startup)
2886 {
2887 MultiXactId oldestMultiXactId;
2888 MultiXactId nextMXact;
2889 MultiXactOffset oldestOffset = 0; /* placate compiler */
2890 MultiXactOffset prevOldestOffset;
2891 MultiXactOffset nextOffset;
2892 bool oldestOffsetKnown = false;
2893 bool prevOldestOffsetKnown;
2894 MultiXactOffset offsetStopLimit = 0;
2895 MultiXactOffset prevOffsetStopLimit;
2896
2897 /*
2898 * NB: Have to prevent concurrent truncation, we might otherwise try to
2899 * lookup an oldestMulti that's concurrently getting truncated away.
2900 */
2901 LWLockAcquire(MultiXactTruncationLock, LW_SHARED);
2902
2903 /* Read relevant fields from shared memory. */
2904 LWLockAcquire(MultiXactGenLock, LW_SHARED);
2905 oldestMultiXactId = MultiXactState->oldestMultiXactId;
2906 nextMXact = MultiXactState->nextMXact;
2907 nextOffset = MultiXactState->nextOffset;
2908 prevOldestOffsetKnown = MultiXactState->oldestOffsetKnown;
2909 prevOldestOffset = MultiXactState->oldestOffset;
2910 prevOffsetStopLimit = MultiXactState->offsetStopLimit;
2911 Assert(MultiXactState->finishedStartup);
2912 LWLockRelease(MultiXactGenLock);
2913
2914 /*
2915 * Determine the offset of the oldest multixact. Normally, we can read
2916 * the offset from the multixact itself, but there's an important special
2917 * case: if there are no multixacts in existence at all, oldestMXact
2918 * obviously can't point to one. It will instead point to the multixact
2919 * ID that will be assigned the next time one is needed.
2920 */
2921 if (oldestMultiXactId == nextMXact)
2922 {
2923 /*
2924 * When the next multixact gets created, it will be stored at the next
2925 * offset.
2926 */
2927 oldestOffset = nextOffset;
2928 oldestOffsetKnown = true;
2929 }
2930 else
2931 {
2932 /*
2933 * Figure out where the oldest existing multixact's offsets are
2934 * stored. Due to bugs in early release of PostgreSQL 9.3.X and 9.4.X,
2935 * the supposedly-earliest multixact might not really exist. We are
2936 * careful not to fail in that case.
2937 */
2938 oldestOffsetKnown =
2939 find_multixact_start(oldestMultiXactId, &oldestOffset);
2940
2941 if (oldestOffsetKnown)
2942 ereport(DEBUG1,
2943 (errmsg_internal("oldest MultiXactId member is at offset %u",
2944 oldestOffset)));
2945 else
2946 ereport(LOG,
2947 (errmsg("MultiXact member wraparound protections are disabled because oldest checkpointed MultiXact %u does not exist on disk",
2948 oldestMultiXactId)));
2949 }
2950
2951 LWLockRelease(MultiXactTruncationLock);
2952
2953 /*
2954 * If we can, compute limits (and install them MultiXactState) to prevent
2955 * overrun of old data in the members SLRU area. We can only do so if the
2956 * oldest offset is known though.
2957 */
2958 if (oldestOffsetKnown)
2959 {
2960 /* move back to start of the corresponding segment */
2961 offsetStopLimit = oldestOffset - (oldestOffset %
2962 (MULTIXACT_MEMBERS_PER_PAGE * SLRU_PAGES_PER_SEGMENT));
2963
2964 /* always leave one segment before the wraparound point */
2965 offsetStopLimit -= (MULTIXACT_MEMBERS_PER_PAGE * SLRU_PAGES_PER_SEGMENT);
2966
2967 if (!prevOldestOffsetKnown && !is_startup)
2968 ereport(LOG,
2969 (errmsg("MultiXact member wraparound protections are now enabled")));
2970
2971 ereport(DEBUG1,
2972 (errmsg_internal("MultiXact member stop limit is now %u based on MultiXact %u",
2973 offsetStopLimit, oldestMultiXactId)));
2974 }
2975 else if (prevOldestOffsetKnown)
2976 {
2977 /*
2978 * If we failed to get the oldest offset this time, but we have a
2979 * value from a previous pass through this function, use the old
2980 * values rather than automatically forcing an emergency autovacuum
2981 * cycle again.
2982 */
2983 oldestOffset = prevOldestOffset;
2984 oldestOffsetKnown = true;
2985 offsetStopLimit = prevOffsetStopLimit;
2986 }
2987
2988 /* Install the computed values */
2989 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2990 MultiXactState->oldestOffset = oldestOffset;
2991 MultiXactState->oldestOffsetKnown = oldestOffsetKnown;
2992 MultiXactState->offsetStopLimit = offsetStopLimit;
2993 LWLockRelease(MultiXactGenLock);
2994
2995 /*
2996 * Do we need an emergency autovacuum? If we're not sure, assume yes.
2997 */
2998 return !oldestOffsetKnown ||
2999 (nextOffset - oldestOffset > MULTIXACT_MEMBER_SAFE_THRESHOLD);
3000 }
3001
3002 /*
3003 * Return whether adding "distance" to "start" would move past "boundary".
3004 *
3005 * We use this to determine whether the addition is "wrapping around" the
3006 * boundary point, hence the name. The reason we don't want to use the regular
3007 * 2^31-modulo arithmetic here is that we want to be able to use the whole of
3008 * the 2^32-1 space here, allowing for more multixacts than would fit
3009 * otherwise.
3010 */
3011 static bool
3012 MultiXactOffsetWouldWrap(MultiXactOffset boundary, MultiXactOffset start,
3013 uint32 distance)
3014 {
3015 MultiXactOffset finish;
3016
3017 /*
3018 * Note that offset number 0 is not used (see GetMultiXactIdMembers), so
3019 * if the addition wraps around the UINT_MAX boundary, skip that value.
3020 */
3021 finish = start + distance;
3022 if (finish < start)
3023 finish++;
3024
3025 /*-----------------------------------------------------------------------
3026 * When the boundary is numerically greater than the starting point, any
3027 * value numerically between the two is not wrapped:
3028 *
3029 * <----S----B---->
3030 * [---) = F wrapped past B (and UINT_MAX)
3031 * [---) = F not wrapped
3032 * [----] = F wrapped past B
3033 *
3034 * When the boundary is numerically less than the starting point (i.e. the
3035 * UINT_MAX wraparound occurs somewhere in between) then all values in
3036 * between are wrapped:
3037 *
3038 * <----B----S---->
3039 * [---) = F not wrapped past B (but wrapped past UINT_MAX)
3040 * [---) = F wrapped past B (and UINT_MAX)
3041 * [----] = F not wrapped
3042 *-----------------------------------------------------------------------
3043 */
3044 if (start < boundary)
3045 return finish >= boundary || finish < start;
3046 else
3047 return finish >= boundary && finish < start;
3048 }
3049
3050 /*
3051 * Find the starting offset of the given MultiXactId.
3052 *
3053 * Returns false if the file containing the multi does not exist on disk.
3054 * Otherwise, returns true and sets *result to the starting member offset.
3055 *
3056 * This function does not prevent concurrent truncation, so if that's
3057 * required, the caller has to protect against that.
3058 */
3059 static bool
3060 find_multixact_start(MultiXactId multi, MultiXactOffset *result)
3061 {
3062 MultiXactOffset offset;
3063 int64 pageno;
3064 int entryno;
3065 int slotno;
3066 MultiXactOffset *offptr;
3067
3068 Assert(MultiXactState->finishedStartup);
3069
3070 pageno = MultiXactIdToOffsetPage(multi);
3071 entryno = MultiXactIdToOffsetEntry(multi);
3072
3073 /*
3074 * Write out dirty data, so PhysicalPageExists can work correctly.
3075 */
3076 SimpleLruWriteAll(MultiXactOffsetCtl, true);
3077 SimpleLruWriteAll(MultiXactMemberCtl, true);
3078
3079 if (!SimpleLruDoesPhysicalPageExist(MultiXactOffsetCtl, pageno))
3080 return false;
3081
3082 /* lock is acquired by SimpleLruReadPage_ReadOnly */
3083 slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno, multi);
3084 offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
3085 offptr += entryno;
3086 offset = *offptr;
3087 LWLockRelease(SimpleLruGetBankLock(MultiXactOffsetCtl, pageno));
3088
3089 *result = offset;
3090 return true;
3091 }
3092
3093 /*
3094 * Determine how many multixacts, and how many multixact members, currently
3095 * exist. Return false if unable to determine.
3096 */
3097 static bool
3098 ReadMultiXactCounts(uint32 *multixacts, MultiXactOffset *members)
3099 {
3100 MultiXactOffset nextOffset;
3101 MultiXactOffset oldestOffset;
3102 MultiXactId oldestMultiXactId;
3103 MultiXactId nextMultiXactId;
3104 bool oldestOffsetKnown;
3105
3106 LWLockAcquire(MultiXactGenLock, LW_SHARED);
3107 nextOffset = MultiXactState->nextOffset;
3108 oldestMultiXactId = MultiXactState->oldestMultiXactId;
3109 nextMultiXactId = MultiXactState->nextMXact;
3110 oldestOffset = MultiXactState->oldestOffset;
3111 oldestOffsetKnown = MultiXactState->oldestOffsetKnown;
3112 LWLockRelease(MultiXactGenLock);
3113
3114 if (!oldestOffsetKnown)
3115 return false;
3116
3117 *members = nextOffset - oldestOffset;
3118 *multixacts = nextMultiXactId - oldestMultiXactId;
3119 return true;
3120 }
3121
3122 /*
3123 * Multixact members can be removed once the multixacts that refer to them
3124 * are older than every datminmxid. autovacuum_multixact_freeze_max_age and
3125 * vacuum_multixact_freeze_table_age work together to make sure we never have
3126 * too many multixacts; we hope that, at least under normal circumstances,
3127 * this will also be sufficient to keep us from using too many offsets.
3128 * However, if the average multixact has many members, we might exhaust the
3129 * members space while still using few enough members that these limits fail
3130 * to trigger relminmxid advancement by VACUUM. At that point, we'd have no
3131 * choice but to start failing multixact-creating operations with an error.
3132 *
3133 * To prevent that, if more than a threshold portion of the members space is
3134 * used, we effectively reduce autovacuum_multixact_freeze_max_age and
3135 * to a value just less than the number of multixacts in use. We hope that
3136 * this will quickly trigger autovacuuming on the table or tables with the
3137 * oldest relminmxid, thus allowing datminmxid values to advance and removing
3138 * some members.
3139 *
3140 * As the fraction of the member space currently in use grows, we become
3141 * more aggressive in clamping this value. That not only causes autovacuum
3142 * to ramp up, but also makes any manual vacuums the user issues more
3143 * aggressive. This happens because vacuum_get_cutoffs() will clamp the
3144 * freeze table and the minimum freeze age cutoffs based on the effective
3145 * autovacuum_multixact_freeze_max_age this function returns. In the worst
3146 * case, we'll claim the freeze_max_age to zero, and every vacuum of any
3147 * table will freeze every multixact.
3148 */
3149 int
3150 MultiXactMemberFreezeThreshold(void)
3151 {
3152 MultiXactOffset members;
3153 uint32 multixacts;
3154 uint32 victim_multixacts;
3155 double fraction;
3156 int result;
3157
3158 /* If we can't determine member space utilization, assume the worst. */
3159 if (!ReadMultiXactCounts(&multixacts, &members))
3160 return 0;
3161
3162 /* If member space utilization is low, no special action is required. */
3163 if (members <= MULTIXACT_MEMBER_SAFE_THRESHOLD)
3164 return autovacuum_multixact_freeze_max_age;
3165
3166 /*
3167 * Compute a target for relminmxid advancement. The number of multixacts
3168 * we try to eliminate from the system is based on how far we are past
3169 * MULTIXACT_MEMBER_SAFE_THRESHOLD.
3170 */
3171 fraction = (double) (members - MULTIXACT_MEMBER_SAFE_THRESHOLD) /
3172 (MULTIXACT_MEMBER_DANGER_THRESHOLD - MULTIXACT_MEMBER_SAFE_THRESHOLD);
3173 victim_multixacts = multixacts * fraction;
3174
3175 /* fraction could be > 1.0, but lowest possible freeze age is zero */
3176 if (victim_multixacts > multixacts)
3177 return 0;
3178 result = multixacts - victim_multixacts;
3179
3180 /*
3181 * Clamp to autovacuum_multixact_freeze_max_age, so that we never make
3182 * autovacuum less aggressive than it would otherwise be.
3183 */
3184 return Min(result, autovacuum_multixact_freeze_max_age);
3185 }
3186
3187 typedef struct mxtruncinfo
3188 {
3189 int64 earliestExistingPage;
3190 } mxtruncinfo;
3191
3192 /*
3193 * SlruScanDirectory callback
3194 * This callback determines the earliest existing page number.
3195 */
3196 static bool
3197 SlruScanDirCbFindEarliest(SlruCtl ctl, char *filename, int64 segpage, void *data)
3198 {
3199 mxtruncinfo *trunc = (mxtruncinfo *) data;
3200
3201 if (trunc->earliestExistingPage == -1 ||
3202 ctl->PagePrecedes(segpage, trunc->earliestExistingPage))
3203 {
3204 trunc->earliestExistingPage = segpage;
3205 }
3206
3207 return false; /* keep going */
3208 }
3209
3210
3211 /*
3212 * Delete members segments [oldest, newOldest)
3213 *
3214 * The members SLRU can, in contrast to the offsets one, be filled to almost
3215 * the full range at once. This means SimpleLruTruncate() can't trivially be
3216 * used - instead the to-be-deleted range is computed using the offsets
3217 * SLRU. C.f. TruncateMultiXact().
3218 */
3219 static void
3220 PerformMembersTruncation(MultiXactOffset oldestOffset, MultiXactOffset newOldestOffset)
3221 {
3222 const int64 maxsegment = MXOffsetToMemberSegment(MaxMultiXactOffset);
3223 int64 startsegment = MXOffsetToMemberSegment(oldestOffset);
3224 int64 endsegment = MXOffsetToMemberSegment(newOldestOffset);
3225 int64 segment = startsegment;
3226
3227 /*
3228 * Delete all the segments but the last one. The last segment can still
3229 * contain, possibly partially, valid data.
3230 */
3231 while (segment != endsegment)
3232 {
3233 elog(DEBUG2, "truncating multixact members segment %" PRIx64,
3234 segment);
3235 SlruDeleteSegment(MultiXactMemberCtl, segment);
3236
3237 /* move to next segment, handling wraparound correctly */
3238 if (segment == maxsegment)
3239 segment = 0;
3240 else
3241 segment += 1;
3242 }
3243 }
3244
3245 /*
3246 * Delete offsets segments [oldest, newOldest)
3247 */
3248 static void
3249 PerformOffsetsTruncation(MultiXactId oldestMulti, MultiXactId newOldestMulti)
3250 {
3251 /*
3252 * We step back one multixact to avoid passing a cutoff page that hasn't
3253 * been created yet in the rare case that oldestMulti would be the first
3254 * item on a page and oldestMulti == nextMulti. In that case, if we
3255 * didn't subtract one, we'd trigger SimpleLruTruncate's wraparound
3256 * detection.
3257 */
3258 SimpleLruTruncate(MultiXactOffsetCtl,
3259 MultiXactIdToOffsetPage(PreviousMultiXactId(newOldestMulti)));
3260 }
3261
3262 /*
3263 * Remove all MultiXactOffset and MultiXactMember segments before the oldest
3264 * ones still of interest.
3265 *
3266 * This is only called on a primary as part of vacuum (via
3267 * vac_truncate_clog()). During recovery truncation is done by replaying
3268 * truncation WAL records logged here.
3269 *
3270 * newOldestMulti is the oldest currently required multixact, newOldestMultiDB
3271 * is one of the databases preventing newOldestMulti from increasing.
3272 */
3273 void
3274 TruncateMultiXact(MultiXactId newOldestMulti, Oid newOldestMultiDB)
3275 {
3276 MultiXactId oldestMulti;
3277 MultiXactId nextMulti;
3278 MultiXactOffset newOldestOffset;
3279 MultiXactOffset oldestOffset;
3280 MultiXactOffset nextOffset;
3281 mxtruncinfo trunc;
3282 MultiXactId earliest;
3283
3284 Assert(!RecoveryInProgress());
3285 Assert(MultiXactState->finishedStartup);
3286
3287 /*
3288 * We can only allow one truncation to happen at once. Otherwise parts of
3289 * members might vanish while we're doing lookups or similar. There's no
3290 * need to have an interlock with creating new multis or such, since those
3291 * are constrained by the limits (which only grow, never shrink).
3292 */
3293 LWLockAcquire(MultiXactTruncationLock, LW_EXCLUSIVE);
3294
3295 LWLockAcquire(MultiXactGenLock, LW_SHARED);
3296 nextMulti = MultiXactState->nextMXact;
3297 nextOffset = MultiXactState->nextOffset;
3298 oldestMulti = MultiXactState->oldestMultiXactId;
3299 LWLockRelease(MultiXactGenLock);
3300 Assert(MultiXactIdIsValid(oldestMulti));
3301
3302 /*
3303 * Make sure to only attempt truncation if there's values to truncate
3304 * away. In normal processing values shouldn't go backwards, but there's
3305 * some corner cases (due to bugs) where that's possible.
3306 */
3307 if (MultiXactIdPrecedesOrEquals(newOldestMulti, oldestMulti))
3308 {
3309 LWLockRelease(MultiXactTruncationLock);
3310 return;
3311 }
3312
3313 /*
3314 * Note we can't just plow ahead with the truncation; it's possible that
3315 * there are no segments to truncate, which is a problem because we are
3316 * going to attempt to read the offsets page to determine where to
3317 * truncate the members SLRU. So we first scan the directory to determine
3318 * the earliest offsets page number that we can read without error.
3319 *
3320 * When nextMXact is less than one segment away from multiWrapLimit,
3321 * SlruScanDirCbFindEarliest can find some early segment other than the
3322 * actual earliest. (MultiXactOffsetPagePrecedes(EARLIEST, LATEST)
3323 * returns false, because not all pairs of entries have the same answer.)
3324 * That can also arise when an earlier truncation attempt failed unlink()
3325 * or returned early from this function. The only consequence is
3326 * returning early, which wastes space that we could have liberated.
3327 *
3328 * NB: It's also possible that the page that oldestMulti is on has already
3329 * been truncated away, and we crashed before updating oldestMulti.
3330 */
3331 trunc.earliestExistingPage = -1;
3332 SlruScanDirectory(MultiXactOffsetCtl, SlruScanDirCbFindEarliest, &trunc);
3333 earliest = trunc.earliestExistingPage * MULTIXACT_OFFSETS_PER_PAGE;
3334 if (earliest < FirstMultiXactId)
3335 earliest = FirstMultiXactId;
3336
3337 /* If there's nothing to remove, we can bail out early. */
3338 if (MultiXactIdPrecedes(oldestMulti, earliest))
3339 {
3340 LWLockRelease(MultiXactTruncationLock);
3341 return;
3342 }
3343
3344 /*
3345 * First, compute the safe truncation point for MultiXactMember. This is
3346 * the starting offset of the oldest multixact.
3347 *
3348 * Hopefully, find_multixact_start will always work here, because we've
3349 * already checked that it doesn't precede the earliest MultiXact on disk.
3350 * But if it fails, don't truncate anything, and log a message.
3351 */
3352 if (oldestMulti == nextMulti)
3353 {
3354 /* there are NO MultiXacts */
3355 oldestOffset = nextOffset;
3356 }
3357 else if (!find_multixact_start(oldestMulti, &oldestOffset))
3358 {
3359 ereport(LOG,
3360 (errmsg("oldest MultiXact %u not found, earliest MultiXact %u, skipping truncation",
3361 oldestMulti, earliest)));
3362 LWLockRelease(MultiXactTruncationLock);
3363 return;
3364 }
3365
3366 /*
3367 * Secondly compute up to where to truncate. Lookup the corresponding
3368 * member offset for newOldestMulti for that.
3369 */
3370 if (newOldestMulti == nextMulti)
3371 {
3372 /* there are NO MultiXacts */
3373 newOldestOffset = nextOffset;
3374 }
3375 else if (!find_multixact_start(newOldestMulti, &newOldestOffset))
3376 {
3377 ereport(LOG,
3378 (errmsg("cannot truncate up to MultiXact %u because it does not exist on disk, skipping truncation",
3379 newOldestMulti)));
3380 LWLockRelease(MultiXactTruncationLock);
3381 return;
3382 }
3383
3384 /*
3385 * On crash, MultiXactIdCreateFromMembers() can leave behind multixids
3386 * that were not yet written out and hence have zero offset on disk. If
3387 * such a multixid becomes oldestMulti, we won't be able to look up its
3388 * offset. That should be rare, so we don't try to do anything smart about
3389 * it. Just skip the truncation, and hope that by the next truncation
3390 * attempt, oldestMulti has advanced to a valid multixid.
3391 */
3392 if (newOldestOffset == 0)
3393 {
3394 ereport(LOG,
3395 (errmsg("cannot truncate up to MultiXact %u because it has invalid offset, skipping truncation",
3396 newOldestMulti)));
3397 LWLockRelease(MultiXactTruncationLock);
3398 return;
3399 }
3400
3401 elog(DEBUG1, "performing multixact truncation: "
3402 "offsets [%u, %u), offsets segments [%" PRIx64 ", %" PRIx64 "), "
3403 "members [%u, %u), members segments [%" PRIx64 ", %" PRIx64 ")",
3404 oldestMulti, newOldestMulti,
3405 MultiXactIdToOffsetSegment(oldestMulti),
3406 MultiXactIdToOffsetSegment(newOldestMulti),
3407 oldestOffset, newOldestOffset,
3408 MXOffsetToMemberSegment(oldestOffset),
3409 MXOffsetToMemberSegment(newOldestOffset));
3410
3411 /*
3412 * Do truncation, and the WAL logging of the truncation, in a critical
3413 * section. That way offsets/members cannot get out of sync anymore, i.e.
3414 * once consistent the newOldestMulti will always exist in members, even
3415 * if we crashed in the wrong moment.
3416 */
3417 START_CRIT_SECTION();
3418
3419 /*
3420 * Prevent checkpoints from being scheduled concurrently. This is critical
3421 * because otherwise a truncation record might not be replayed after a
3422 * crash/basebackup, even though the state of the data directory would
3423 * require it.
3424 */
3425 Assert((MyProc->delayChkptFlags & DELAY_CHKPT_START) == 0);
3426 MyProc->delayChkptFlags |= DELAY_CHKPT_START;
3427
3428 /* WAL log truncation */
3429 WriteMTruncateXlogRec(newOldestMultiDB,
3430 oldestMulti, newOldestMulti,
3431 oldestOffset, newOldestOffset);
3432
3433 /*
3434 * Update in-memory limits before performing the truncation, while inside
3435 * the critical section: Have to do it before truncation, to prevent
3436 * concurrent lookups of those values. Has to be inside the critical
3437 * section as otherwise a future call to this function would error out,
3438 * while looking up the oldest member in offsets, if our caller crashes
3439 * before updating the limits.
3440 */
3441 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
3442 MultiXactState->oldestMultiXactId = newOldestMulti;
3443 MultiXactState->oldestMultiXactDB = newOldestMultiDB;
3444 LWLockRelease(MultiXactGenLock);
3445
3446 /* First truncate members */
3447 PerformMembersTruncation(oldestOffset, newOldestOffset);
3448
3449 /* Then offsets */
3450 PerformOffsetsTruncation(oldestMulti, newOldestMulti);
3451
3452 MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
3453
3454 END_CRIT_SECTION();
3455 LWLockRelease(MultiXactTruncationLock);
3456 }
3457
3458 /*
3459 * Decide whether a MultiXactOffset page number is "older" for truncation
3460 * purposes. Analogous to CLOGPagePrecedes().
3461 *
3462 * Offsetting the values is optional, because MultiXactIdPrecedes() has
3463 * translational symmetry.
3464 */
3465 static bool
3466 MultiXactOffsetPagePrecedes(int64 page1, int64 page2)
3467 {
3468 MultiXactId multi1;
3469 MultiXactId multi2;
3470
3471 multi1 = ((MultiXactId) page1) * MULTIXACT_OFFSETS_PER_PAGE;
3472 multi1 += FirstMultiXactId + 1;
3473 multi2 = ((MultiXactId) page2) * MULTIXACT_OFFSETS_PER_PAGE;
3474 multi2 += FirstMultiXactId + 1;
3475
3476 return (MultiXactIdPrecedes(multi1, multi2) &&
3477 MultiXactIdPrecedes(multi1,
3478 multi2 + MULTIXACT_OFFSETS_PER_PAGE - 1));
3479 }
3480
3481 /*
3482 * Decide whether a MultiXactMember page number is "older" for truncation
3483 * purposes. There is no "invalid offset number" so use the numbers verbatim.
3484 */
3485 static bool
3486 MultiXactMemberPagePrecedes(int64 page1, int64 page2)
3487 {
3488 MultiXactOffset offset1;
3489 MultiXactOffset offset2;
3490
3491 offset1 = ((MultiXactOffset) page1) * MULTIXACT_MEMBERS_PER_PAGE;
3492 offset2 = ((MultiXactOffset) page2) * MULTIXACT_MEMBERS_PER_PAGE;
3493
3494 return (MultiXactOffsetPrecedes(offset1, offset2) &&
3495 MultiXactOffsetPrecedes(offset1,
3496 offset2 + MULTIXACT_MEMBERS_PER_PAGE - 1));
3497 }
3498
3499 /*
3500 * Decide which of two MultiXactIds is earlier.
3501 *
3502 * XXX do we need to do something special for InvalidMultiXactId?
3503 * (Doesn't look like it.)
3504 */
3505 bool
3506 MultiXactIdPrecedes(MultiXactId multi1, MultiXactId multi2)
3507 {
3508 int32 diff = (int32) (multi1 - multi2);
3509
3510 return (diff < 0);
3511 }
3512
3513 /*
3514 * MultiXactIdPrecedesOrEquals -- is multi1 logically <= multi2?
3515 *
3516 * XXX do we need to do something special for InvalidMultiXactId?
3517 * (Doesn't look like it.)
3518 */
3519 bool
3520 MultiXactIdPrecedesOrEquals(MultiXactId multi1, MultiXactId multi2)
3521 {
3522 int32 diff = (int32) (multi1 - multi2);
3523
3524 return (diff <= 0);
3525 }
3526
3527
3528 /*
3529 * Decide which of two offsets is earlier.
3530 */
3531 static bool
3532 MultiXactOffsetPrecedes(MultiXactOffset offset1, MultiXactOffset offset2)
3533 {
3534 int32 diff = (int32) (offset1 - offset2);
3535
3536 return (diff < 0);
3537 }
3538
3539 /*
3540 * Write an xlog record reflecting the zeroing of either a MEMBERs or
3541 * OFFSETs page (info shows which)
3542 */
3543 static void
3544 WriteMZeroPageXlogRec(int64 pageno, uint8 info)
3545 {
3546 XLogBeginInsert();
3547 XLogRegisterData(&pageno, sizeof(pageno));
3548 (void) XLogInsert(RM_MULTIXACT_ID, info);
3549 }
3550
3551 /*
3552 * Write a TRUNCATE xlog record
3553 *
3554 * We must flush the xlog record to disk before returning --- see notes in
3555 * TruncateCLOG().
3556 */
3557 static void
3558 WriteMTruncateXlogRec(Oid oldestMultiDB,
3559 MultiXactId startTruncOff, MultiXactId endTruncOff,
3560 MultiXactOffset startTruncMemb, MultiXactOffset endTruncMemb)
3561 {
3562 XLogRecPtr recptr;
3563 xl_multixact_truncate xlrec;
3564
3565 xlrec.oldestMultiDB = oldestMultiDB;
3566
3567 xlrec.startTruncOff = startTruncOff;
3568 xlrec.endTruncOff = endTruncOff;
3569
3570 xlrec.startTruncMemb = startTruncMemb;
3571 xlrec.endTruncMemb = endTruncMemb;
3572
3573 XLogBeginInsert();
3574 XLogRegisterData(&xlrec, SizeOfMultiXactTruncate);
3575 recptr = XLogInsert(RM_MULTIXACT_ID, XLOG_MULTIXACT_TRUNCATE_ID);
3576 XLogFlush(recptr);
3577 }
3578
3579 /*
3580 * MULTIXACT resource manager's routines
3581 */
3582 void
3583 multixact_redo(XLogReaderState *record)
3584 {
3585 uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
3586
3587 /* Backup blocks are not used in multixact records */
3588 Assert(!XLogRecHasAnyBlockRefs(record));
3589
3590 if (info == XLOG_MULTIXACT_ZERO_OFF_PAGE)
3591 {
3592 int64 pageno;
3593 int slotno;
3594 LWLock *lock;
3595
3596 memcpy(&pageno, XLogRecGetData(record), sizeof(pageno));
3597
3598 /*
3599 * Skip the record if we already initialized the page at the previous
3600 * XLOG_MULTIXACT_CREATE_ID record. See RecordNewMultiXact().
3601 */
3602 if (pre_initialized_offsets_page != pageno)
3603 {
3604 lock = SimpleLruGetBankLock(MultiXactOffsetCtl, pageno);
3605 LWLockAcquire(lock, LW_EXCLUSIVE);
3606
3607 slotno = ZeroMultiXactOffsetPage(pageno, false);
3608 SimpleLruWritePage(MultiXactOffsetCtl, slotno);
3609 Assert(!MultiXactOffsetCtl->shared->page_dirty[slotno]);
3610
3611 LWLockRelease(lock);
3612
3613 last_initialized_offsets_page = pageno;
3614 }
3615 else
3616 elog(DEBUG1, "skipping initialization of offsets page " INT64_FORMAT " because it was already initialized on multixid creation", pageno);
3617 pre_initialized_offsets_page = -1;
3618 }
3619 else if (info == XLOG_MULTIXACT_ZERO_MEM_PAGE)
3620 {
3621 int64 pageno;
3622 int slotno;
3623 LWLock *lock;
3624
3625 memcpy(&pageno, XLogRecGetData(record), sizeof(pageno));
3626
3627 lock = SimpleLruGetBankLock(MultiXactMemberCtl, pageno);
3628 LWLockAcquire(lock, LW_EXCLUSIVE);
3629
3630 slotno = ZeroMultiXactMemberPage(pageno, false);
3631 SimpleLruWritePage(MultiXactMemberCtl, slotno);
3632 Assert(!MultiXactMemberCtl->shared->page_dirty[slotno]);
3633
3634 LWLockRelease(lock);
3635 }
3636 else if (info == XLOG_MULTIXACT_CREATE_ID)
3637 {
3638 xl_multixact_create *xlrec =
3639 (xl_multixact_create *) XLogRecGetData(record);
3640 TransactionId max_xid;
3641 int i;
3642
3643 if (pre_initialized_offsets_page != -1)
3644 {
3645 /*
3646 * If we implicitly initialized the next offsets page while
3647 * replaying an XLOG_MULTIXACT_CREATE_ID record that was generated
3648 * with an older minor version, we still expect to see an
3649 * XLOG_MULTIXACT_ZERO_OFF_PAGE record for it before any other
3650 * XLOG_MULTIXACT_CREATE_ID records. Therefore this case should
3651 * not happen. If it does, we'll continue with the replay, but
3652 * log a message to note that something's funny.
3653 */
3654 elog(LOG, "expected to see an XLOG_MULTIXACT_ZERO_OFF_PAGE record for page " INT64_FORMAT " that was implicitly initialized earlier",
3655 pre_initialized_offsets_page);
3656 pre_initialized_offsets_page = -1;
3657 }
3658
3659 /* Store the data back into the SLRU files */
3660 RecordNewMultiXact(xlrec->mid, xlrec->moff, xlrec->nmembers,
3661 xlrec->members);
3662
3663 /* Make sure nextMXact/nextOffset are beyond what this record has */
3664 MultiXactAdvanceNextMXact(xlrec->mid + 1,
3665 xlrec->moff + xlrec->nmembers);
3666
3667 /*
3668 * Make sure nextXid is beyond any XID mentioned in the record. This
3669 * should be unnecessary, since any XID found here ought to have other
3670 * evidence in the XLOG, but let's be safe.
3671 */
3672 max_xid = XLogRecGetXid(record);
3673 for (i = 0; i < xlrec->nmembers; i++)
3674 {
3675 if (TransactionIdPrecedes(max_xid, xlrec->members[i].xid))
3676 max_xid = xlrec->members[i].xid;
3677 }
3678
3679 AdvanceNextFullTransactionIdPastXid(max_xid);
3680 }
3681 else if (info == XLOG_MULTIXACT_TRUNCATE_ID)
3682 {
3683 xl_multixact_truncate xlrec;
3684
3685 memcpy(&xlrec, XLogRecGetData(record),
3686 SizeOfMultiXactTruncate);
3687
3688 elog(DEBUG1, "replaying multixact truncation: "
3689 "offsets [%u, %u), offsets segments [%" PRIx64 ", %" PRIx64 "), "
3690 "members [%u, %u), members segments [%" PRIx64 ", %" PRIx64 ")",
3691 xlrec.startTruncOff, xlrec.endTruncOff,
3692 MultiXactIdToOffsetSegment(xlrec.startTruncOff),
3693 MultiXactIdToOffsetSegment(xlrec.endTruncOff),
3694 xlrec.startTruncMemb, xlrec.endTruncMemb,
3695 MXOffsetToMemberSegment(xlrec.startTruncMemb),
3696 MXOffsetToMemberSegment(xlrec.endTruncMemb));
3697
3698 /* should not be required, but more than cheap enough */
3699 LWLockAcquire(MultiXactTruncationLock, LW_EXCLUSIVE);
3700
3701 /*
3702 * Advance the horizon values, so they're current at the end of
3703 * recovery.
3704 */
3705 SetMultiXactIdLimit(xlrec.endTruncOff, xlrec.oldestMultiDB, false);
3706
3707 PerformMembersTruncation(xlrec.startTruncMemb, xlrec.endTruncMemb);
3708 PerformOffsetsTruncation(xlrec.startTruncOff, xlrec.endTruncOff);
3709
3710 LWLockRelease(MultiXactTruncationLock);
3711 }
3712 else
3713 elog(PANIC, "multixact_redo: unknown op code %u", info);
3714 }
3715
3716 Datum
3717 pg_get_multixact_members(PG_FUNCTION_ARGS)
3718 {
3719 typedef struct
3720 {
3721 MultiXactMember *members;
3722 int nmembers;
3723 int iter;
3724 } mxact;
3725 MultiXactId mxid = PG_GETARG_TRANSACTIONID(0);
3726 mxact *multi;
3727 FuncCallContext *funccxt;
3728
3729 if (mxid < FirstMultiXactId)
3730 ereport(ERROR,
3731 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
3732 errmsg("invalid MultiXactId: %u", mxid)));
3733
3734 if (SRF_IS_FIRSTCALL())
3735 {
3736 MemoryContext oldcxt;
3737 TupleDesc tupdesc;
3738
3739 funccxt = SRF_FIRSTCALL_INIT();
3740 oldcxt = MemoryContextSwitchTo(funccxt->multi_call_memory_ctx);
3741
3742 multi = palloc(sizeof(mxact));
3743 /* no need to allow for old values here */
3744 multi->nmembers = GetMultiXactIdMembers(mxid, &multi->members, false,
3745 false);
3746 multi->iter = 0;
3747
3748 if (get_call_result_type(fcinfo, NULL, &tupdesc) != TYPEFUNC_COMPOSITE)
3749 elog(ERROR, "return type must be a row type");
3750 funccxt->tuple_desc = tupdesc;
3751 funccxt->attinmeta = TupleDescGetAttInMetadata(tupdesc);
3752 funccxt->user_fctx = multi;
3753
3754 MemoryContextSwitchTo(oldcxt);
3755 }
3756
3757 funccxt = SRF_PERCALL_SETUP();
3758 multi = (mxact *) funccxt->user_fctx;
3759
3760 while (multi->iter < multi->nmembers)
3761 {
3762 HeapTuple tuple;
3763 char *values[2];
3764
3765 values[0] = psprintf("%u", multi->members[multi->iter].xid);
3766 values[1] = mxstatus_to_string(multi->members[multi->iter].status);
3767
3768 tuple = BuildTupleFromCStrings(funccxt->attinmeta, values);
3769
3770 multi->iter++;
3771 pfree(values[0]);
3772 SRF_RETURN_NEXT(funccxt, HeapTupleGetDatum(tuple));
3773 }
3774
3775 SRF_RETURN_DONE(funccxt);
3776 }
3777
3778 /*
3779 * Entrypoint for sync.c to sync offsets files.
3780 */
3781 int
3782 multixactoffsetssyncfiletag(const FileTag *ftag, char *path)
3783 {
3784 return SlruSyncFileTag(MultiXactOffsetCtl, ftag, path);
3785 }
3786
3787 /*
3788 * Entrypoint for sync.c to sync members files.
3789 */
3790 int
3791 multixactmemberssyncfiletag(const FileTag *ftag, char *path)
3792 {
3793 return SlruSyncFileTag(MultiXactMemberCtl, ftag, path);
3794 }